diff options
Diffstat (limited to 'src')
121 files changed, 10938 insertions, 4868 deletions
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 4ed5f67..b95409f 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -377,14 +377,6 @@ set (H5HL_HDRS IDE_GENERATED_PROPERTIES ("H5HL" "${H5HL_HDRS}" "${H5HL_SOURCES}" ) -set (H5HP_SOURCES - ${HDF5_SRC_DIR}/H5HP.c -) -set (H5HP_HDRS -) -IDE_GENERATED_PROPERTIES ("H5HP" "${H5HP_HDRS}" "${H5HP_SOURCES}" ) - - set (H5I_SOURCES ${HDF5_SRC_DIR}/H5I.c ${HDF5_SRC_DIR}/H5Idbg.c @@ -441,16 +433,6 @@ set (H5MM_HDRS IDE_GENERATED_PROPERTIES ("H5MM" "${H5MM_HDRS}" "${H5MM_SOURCES}" ) -set (H5MP_SOURCES - ${HDF5_SRC_DIR}/H5MP.c - ${HDF5_SRC_DIR}/H5MPtest.c -) - -set (H5MP_HDRS -) -IDE_GENERATED_PROPERTIES ("H5MP" "${H5MP_HDRS}" "${H5MP_SOURCES}" ) - - set (H5O_SOURCES ${HDF5_SRC_DIR}/H5O.c ${HDF5_SRC_DIR}/H5Oainfo.c @@ -749,7 +731,6 @@ set (H5_MODULE_HEADERS ${HDF5_SRC_DIR}/H5Lmodule.h ${HDF5_SRC_DIR}/H5Mmodule.h ${HDF5_SRC_DIR}/H5MFmodule.h - ${HDF5_SRC_DIR}/H5MPmodule.h ${HDF5_SRC_DIR}/H5Omodule.h ${HDF5_SRC_DIR}/H5Pmodule.h ${HDF5_SRC_DIR}/H5PBmodule.h @@ -787,13 +768,11 @@ set (common_SRCS ${H5HF_SOURCES} ${H5HG_SOURCES} ${H5HL_SOURCES} - ${H5HP_SOURCES} ${H5I_SOURCES} ${H5L_SOURCES} ${H5M_SOURCES} ${H5MF_SOURCES} ${H5MM_SOURCES} - ${H5MP_SOURCES} ${H5O_SOURCES} ${H5P_SOURCES} ${H5PB_SOURCES} @@ -836,7 +815,6 @@ set (H5_PUBLIC_HEADERS ${H5M_HDRS} ${H5MF_HDRS} ${H5MM_HDRS} - ${H5MP_HDRS} ${H5O_HDRS} ${H5P_HDRS} ${H5PB_HDRS} @@ -917,8 +895,6 @@ set (H5_PRIVATE_HEADERS ${HDF5_SRC_DIR}/H5HLpkg.h ${HDF5_SRC_DIR}/H5HLprivate.h - ${HDF5_SRC_DIR}/H5HPprivate.h - ${HDF5_SRC_DIR}/H5Ipkg.h ${HDF5_SRC_DIR}/H5Iprivate.h @@ -933,9 +909,6 @@ set (H5_PRIVATE_HEADERS ${HDF5_SRC_DIR}/H5MMprivate.h - ${HDF5_SRC_DIR}/H5MPpkg.h - ${HDF5_SRC_DIR}/H5MPprivate.h - ${HDF5_SRC_DIR}/H5Opkg.h ${HDF5_SRC_DIR}/H5Oprivate.h ${HDF5_SRC_DIR}/H5Oshared.h @@ -1056,20 +1029,23 @@ if (LOCAL_BATCH_TEST) endif () endif () +#### make the H5detect program set (lib_prog_deps) -if (NOT EXISTS "${HDF5_GENERATED_SOURCE_DIR}/H5Tinit.c") - add_executable (H5detect ${HDF5_SRC_DIR}/H5detect.c) - target_include_directories (H5detect PRIVATE "${HDF5_SRC_DIR};${HDF5_SRC_BINARY_DIR};$<$<BOOL:${HDF5_ENABLE_PARALLEL}>:${MPI_C_INCLUDE_DIRS}>") - target_compile_definitions(H5detect PUBLIC ${HDF_EXTRA_C_FLAGS} ${HDF_EXTRA_FLAGS}) - TARGET_C_PROPERTIES (H5detect STATIC) - target_link_libraries (H5detect - PRIVATE "$<$<BOOL:${HDF5_ENABLE_PARALLEL}>:${MPI_C_LIBRARIES}>" $<$<OR:$<PLATFORM_ID:Windows>,$<PLATFORM_ID:MinGW>>:ws2_32.lib> - ) - target_compile_options(H5detect - PRIVATE "$<$<PLATFORM_ID:Emscripten>:-O0>" - ) - set (lib_prog_deps ${lib_prog_deps} H5detect) +add_executable (H5detect ${HDF5_SRC_DIR}/H5detect.c) +target_include_directories (H5detect PRIVATE "${HDF5_SRC_DIR};${HDF5_SRC_BINARY_DIR};$<$<BOOL:${HDF5_ENABLE_PARALLEL}>:${MPI_C_INCLUDE_DIRS}>") +target_compile_definitions(H5detect PUBLIC ${HDF_EXTRA_C_FLAGS} ${HDF_EXTRA_FLAGS}) +TARGET_C_PROPERTIES (H5detect STATIC) +target_link_libraries (H5detect + PRIVATE "$<$<BOOL:${HDF5_ENABLE_PARALLEL}>:${MPI_C_LIBRARIES}>" $<$<OR:$<PLATFORM_ID:Windows>,$<PLATFORM_ID:MinGW>>:ws2_32.lib> +) +target_compile_options(H5detect + PRIVATE "$<$<PLATFORM_ID:Emscripten>:-O0>" +) +set (lib_prog_deps ${lib_prog_deps} H5detect) +# check if a pregenerated H5Tinit.c file is present +if (NOT EXISTS "${HDF5_GENERATED_SOURCE_DIR}/H5Tinit.c") + # execute the H5detect program if (HDF5_BATCH_H5DETECT) configure_file ( ${HDF5_SOURCE_DIR}/bin/batch/${HDF5_BATCH_H5DETECT_SCRIPT}.in.cmake @@ -1077,9 +1053,9 @@ if (NOT EXISTS "${HDF5_GENERATED_SOURCE_DIR}/H5Tinit.c") ) add_custom_command ( OUTPUT gen_SRCS.stamp1 + BYPRODUCTS H5Tinit.c COMMAND ${HDF5_BATCH_CMD} ARGS ${HDF5_BINARY_DIR}/${HDF5_BATCH_H5DETECT_SCRIPT} - BYPRODUCTS H5Tinit.c gen_SRCS.stamp1 COMMAND ${CMAKE_COMMAND} ARGS -E echo "Executed batch command to create H5Tinit.c" COMMAND ${CMAKE_COMMAND} @@ -1090,31 +1066,30 @@ if (NOT EXISTS "${HDF5_GENERATED_SOURCE_DIR}/H5Tinit.c") add_custom_target (gen_H5Tinit COMMAND ${CMAKE_COMMAND} -P ${HDF5_SOURCE_DIR}/config/cmake/wait_H5Tinit.cmake ) - set_source_files_properties (${HDF5_GENERATED_SOURCE_DIR}/H5Tinit.c PROPERTIES GENERATED TRUE) else () - add_custom_command (TARGET H5detect POST_BUILD + add_custom_command ( + OUTPUT gen_SRCS.stamp1 + BYPRODUCTS H5Tinit.c COMMAND ${CMAKE_CROSSCOMPILING_EMULATOR} $<TARGET_FILE:H5detect> ARGS H5Tinit.c - BYPRODUCTS H5Tinit.c gen_SRCS.stamp1 COMMAND ${CMAKE_COMMAND} ARGS -E touch gen_SRCS.stamp1 DEPENDS H5detect WORKING_DIRECTORY ${HDF5_GENERATED_SOURCE_DIR} COMMENT "Create H5Tinit.c" ) - set_source_files_properties (${HDF5_GENERATED_SOURCE_DIR}/H5Tinit.c PROPERTIES GENERATED TRUE) if (BUILD_SHARED_LIBS) - add_custom_command (TARGET H5detect POST_BUILD + add_custom_command ( + OUTPUT shared/shared_gen_SRCS.stamp1 + BYPRODUCTS shared/H5Tinit.c COMMAND ${CMAKE_COMMAND} ARGS -E copy_if_different H5Tinit.c shared/H5Tinit.c - BYPRODUCTS shared/H5Tinit.c shared/shared_gen_SRCS.stamp1 COMMAND ${CMAKE_COMMAND} ARGS -E touch shared/shared_gen_SRCS.stamp1 - DEPENDS H5detect H5Tinit.c + DEPENDS H5detect gen_SRCS.stamp1 WORKING_DIRECTORY ${HDF5_GENERATED_SOURCE_DIR} COMMENT "Copy H5Tinit.c to shared folder" ) - set_source_files_properties (${HDF5_GENERATED_SOURCE_DIR}/shared/H5Tinit.c PROPERTIES GENERATED TRUE) endif () endif () else () @@ -1130,16 +1105,15 @@ else () if (BUILD_SHARED_LIBS) add_custom_command ( OUTPUT shared/shared_gen_SRCS.stamp1 + BYPRODUCTS shared/H5Tinit.c COMMAND ${CMAKE_COMMAND} ARGS -E copy_if_different H5Tinit.c shared/H5Tinit.c - BYPRODUCTS shared/H5Tinit.c shared/shared_gen_SRCS.stamp1 COMMAND ${CMAKE_COMMAND} ARGS -E touch shared/shared_gen_SRCS.stamp1 - DEPENDS H5Tinit.c + DEPENDS H5Tinit.c gen_SRCS.stamp1 WORKING_DIRECTORY ${HDF5_GENERATED_SOURCE_DIR} COMMENT "Copy existing H5Tinit.c to shared folder" ) - set_source_files_properties (${HDF5_GENERATED_SOURCE_DIR}/shared/H5Tinit.c PROPERTIES GENERATED TRUE) endif () endif () @@ -1150,6 +1124,7 @@ if (HDF5_ENABLE_FORMATTERS) clang_format (HDF5_SRC_DETECT_FORMAT ${HDF5_SRC_DIR}/H5detect.c) endif () +# make the H5make_libsettings program add_executable (H5make_libsettings ${HDF5_SRC_DIR}/H5make_libsettings.c) target_include_directories (H5make_libsettings PRIVATE "${HDF5_SRC_DIR};${HDF5_SRC_BINARY_DIR};$<$<BOOL:${HDF5_ENABLE_PARALLEL}>:${MPI_C_INCLUDE_DIRS}>") target_compile_definitions(H5make_libsettings PUBLIC ${HDF_EXTRA_C_FLAGS} ${HDF_EXTRA_FLAGS}) @@ -1169,10 +1144,12 @@ if (HDF5_ENABLE_FORMATTERS) clang_format (HDF5_SRC_LIBSETTINGS_FORMAT H5make_libsettings) endif () -add_custom_command (TARGET H5make_libsettings POST_BUILD +# execute the H5make_libsettings program +add_custom_command ( + OUTPUT gen_SRCS.stamp2 + BYPRODUCTS H5lib_settings.c COMMAND ${CMAKE_CROSSCOMPILING_EMULATOR} $<TARGET_FILE:H5make_libsettings> ARGS H5lib_settings.c - BYPRODUCTS H5lib_settings.c gen_SRCS.stamp2 COMMAND ${CMAKE_COMMAND} ARGS -E touch gen_SRCS.stamp2 DEPENDS H5make_libsettings @@ -1181,17 +1158,17 @@ add_custom_command (TARGET H5make_libsettings POST_BUILD ) set_source_files_properties (${HDF5_SRC_BINARY_DIR}/H5lib_settings.c PROPERTIES GENERATED TRUE) if (BUILD_SHARED_LIBS) - add_custom_command (TARGET H5make_libsettings POST_BUILD + add_custom_command ( + OUTPUT shared/shared_gen_SRCS.stamp2 + BYPRODUCTS shared/H5lib_settings.c COMMAND ${CMAKE_COMMAND} ARGS -E copy_if_different H5lib_settings.c shared/H5lib_settings.c - BYPRODUCTS shared/H5lib_settings.c shared/shared_gen_SRCS.stamp2 COMMAND ${CMAKE_COMMAND} ARGS -E touch shared/shared_gen_SRCS.stamp2 - DEPENDS H5make_libsettings H5lib_settings.c + DEPENDS H5make_libsettings gen_SRCS.stamp2 WORKING_DIRECTORY ${HDF5_SRC_BINARY_DIR} COMMENT "Copy H5lib_settings.c to shared folder" ) - set_source_files_properties (${HDF5_SRC_BINARY_DIR}/shared/H5lib_settings.c PROPERTIES GENERATED TRUE) endif () ## all_packages="AC,B,B2,D,F,FA,FL,FS,HL,I,O,S,ST,T,Z" @@ -1204,7 +1181,7 @@ option (HDF5_ENABLE_DEBUG_APIS "Turn on extra debug output in all packages" OFF) if (NOT ONLY_SHARED_LIBS) set (gen_SRCS ${HDF5_GENERATED_SOURCE_DIR}/H5Tinit.c ${HDF5_SRC_BINARY_DIR}/H5lib_settings.c) add_custom_target (gen_${HDF5_LIB_TARGET} ALL - DEPENDS ${lib_prog_deps} ${gen_SRCS} ${HDF5_GENERATED_SOURCE_DIR}/gen_SRCS.stamp1 ${HDF5_SRC_BINARY_DIR}/gen_SRCS.stamp2 + DEPENDS ${lib_prog_deps} ${HDF5_GENERATED_SOURCE_DIR}/gen_SRCS.stamp1 ${HDF5_SRC_BINARY_DIR}/gen_SRCS.stamp2 COMMENT "Generation target files" ) @@ -1243,7 +1220,7 @@ endif () if (BUILD_SHARED_LIBS) set (shared_gen_SRCS ${HDF5_GENERATED_SOURCE_DIR}/shared/H5Tinit.c ${HDF5_SRC_BINARY_DIR}/shared/H5lib_settings.c) add_custom_target (gen_${HDF5_LIBSH_TARGET} ALL - DEPENDS ${lib_prog_deps} ${shared_gen_SRCS} ${HDF5_GENERATED_SOURCE_DIR}/shared/shared_gen_SRCS.stamp1 ${HDF5_SRC_BINARY_DIR}/shared/shared_gen_SRCS.stamp2 + DEPENDS ${lib_prog_deps} ${HDF5_GENERATED_SOURCE_DIR}/shared/shared_gen_SRCS.stamp1 ${HDF5_SRC_BINARY_DIR}/shared/shared_gen_SRCS.stamp2 COMMENT "Shared generation target files" ) @@ -70,9 +70,9 @@ static int H5__mpi_delete_cb(MPI_Comm comm, int keyval, void *attr_val, int *fla /* Library Private Variables */ /*****************************/ -/* Library incompatible release versions */ -const unsigned VERS_RELEASE_EXCEPTIONS[] = {0}; -const unsigned VERS_RELEASE_EXCEPTIONS_SIZE = 0; +/* Library incompatible release versions, develop releases are incompatible by design */ +const unsigned VERS_RELEASE_EXCEPTIONS[] = {0, 1, 2}; +const unsigned VERS_RELEASE_EXCEPTIONS_SIZE = 3; /* statically initialize block for pthread_once call used in initializing */ /* the first global mutex */ @@ -83,6 +83,8 @@ hbool_t H5_libinit_g = FALSE; /* Library hasn't been initialized */ hbool_t H5_libterm_g = FALSE; /* Library isn't being shutdown */ #endif +hbool_t H5_use_selection_io_g = FALSE; + #ifdef H5_HAVE_MPE hbool_t H5_MPEinit_g = FALSE; /* MPE Library hasn't been initialized */ #endif @@ -144,7 +146,8 @@ herr_t H5_init_library(void) { size_t i; - herr_t ret_value = SUCCEED; + char * env_use_select_io = NULL; + herr_t ret_value = SUCCEED; FUNC_ENTER_NOAPI(FAIL) @@ -288,6 +291,14 @@ H5_init_library(void) } /* clang-format on */ + /* Check for HDF5_USE_SELECTION_IO env variable */ + env_use_select_io = HDgetenv("HDF5_USE_SELECTION_IO"); + if (NULL != env_use_select_io && HDstrcmp(env_use_select_io, "") && HDstrcmp(env_use_select_io, "0") && + HDstrcmp(env_use_select_io, "no") && HDstrcmp(env_use_select_io, "No") && + HDstrcmp(env_use_select_io, "NO") && HDstrcmp(env_use_select_io, "false") && + HDstrcmp(env_use_select_io, "False") && HDstrcmp(env_use_select_io, "FALSE")) + H5_use_selection_io_g = TRUE; + /* Debugging? */ H5__debug_mask("-all"); H5__debug_mask(HDgetenv("HDF5_DEBUG")); @@ -954,6 +965,7 @@ H5check_version(unsigned majnum, unsigned minnum, unsigned relnum) static int checked = 0; /* If we've already checked the version info */ static unsigned int disable_version_check = 0; /* Set if the version check should be disabled */ static const char * version_mismatch_warning = VERSION_MISMATCH_WARNING; + static const char * release_mismatch_warning = RELEASE_MISMATCH_WARNING; herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_API_NOINIT_NOERR_NOFS @@ -974,10 +986,7 @@ H5check_version(unsigned majnum, unsigned minnum, unsigned relnum) } /* H5_VERS_MAJOR and H5_VERS_MINOR must match */ - /* Cast relnum to int to avoid warning for unsigned < 0 comparison - * in first release versions */ - if (H5_VERS_MAJOR != majnum || H5_VERS_MINOR != minnum || H5_VERS_RELEASE > (int)relnum) { - + if (H5_VERS_MAJOR != majnum || H5_VERS_MINOR != minnum) { switch (disable_version_check) { case 0: HDfprintf(stderr, "%s%s", version_mismatch_warning, @@ -1012,9 +1021,10 @@ H5check_version(unsigned majnum, unsigned minnum, unsigned relnum) break; } /* end switch */ - } /* end if (H5_VERS_MAJOR != majnum || H5_VERS_MINOR != minnum || H5_VERS_RELEASE > relnum) */ + } /* end if (H5_VERS_MAJOR != majnum || H5_VERS_MINOR != minnum) */ /* H5_VERS_RELEASE should be compatible, we will only add checks for exceptions */ + /* Library develop release versions are incompatible by design */ if (H5_VERS_RELEASE != relnum) { for (unsigned i = 0; i < VERS_RELEASE_EXCEPTIONS_SIZE; i++) { /* Check for incompatible headers or incompatible library */ @@ -1022,7 +1032,7 @@ H5check_version(unsigned majnum, unsigned minnum, unsigned relnum) switch (disable_version_check) { case 0: HDfprintf( - stderr, "%s%s", version_mismatch_warning, + stderr, "%s%s", release_mismatch_warning, "You can, at your own risk, disable this warning by setting the environment\n" "variable 'HDF5_DISABLE_VERSION_CHECK' to a value of '1'.\n" "Setting it to 2 or higher will suppress the warning messages totally.\n"); @@ -1041,7 +1051,7 @@ H5check_version(unsigned majnum, unsigned minnum, unsigned relnum) "%s'HDF5_DISABLE_VERSION_CHECK' " "environment variable is set to %d, application will\n" "continue at your own risk.\n", - version_mismatch_warning, disable_version_check); + release_mismatch_warning, disable_version_check); /* Mention the versions we are referring to */ HDfprintf(stderr, "Headers are %u.%u.%u, library is %u.%u.%u\n", majnum, minnum, relnum, (unsigned)H5_VERS_MAJOR, (unsigned)H5_VERS_MINOR, @@ -303,7 +303,7 @@ H5AC_create(const H5F_t *f, H5AC_cache_config_t *config_ptr, H5AC_cache_image_co aux_ptr->sync_point_done = NULL; aux_ptr->p0_image_len = 0; - HDsprintf(prefix, "%d:", mpi_rank); + HDsnprintf(prefix, sizeof(prefix), "%d:", mpi_rank); if (mpi_rank == 0) { if (NULL == (aux_ptr->d_slist_ptr = H5SL_create(H5SL_TYPE_HADDR, NULL))) @@ -1440,21 +1440,82 @@ H5AC_resize_entry(void *thing, size_t new_size) cache_ptr = entry_ptr->cache_ptr; HDassert(cache_ptr); - /* Resize the entry */ - if (H5C_resize_entry(thing, new_size) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, "can't resize entry") - #ifdef H5_HAVE_PARALLEL - { + /* Log the generation of dirty bytes of metadata iff: + * + * 1) The entry is clean on entry, and this resize will dirty it + * (i.e. the current and new sizes are different), and + * + * 2) This is a parallel computation -- which it is if the aux_ptr + * is non-null. + * + * A few points to note about this section of the code: + * + * 1) This call must occur before the call to H5C_resize_entry() since + * H5AC__log_dirtied_entry() expects the target entry to be clean + * on entry. + * + * 2) This code has some basic issues in terms of the number of bytes + * added to the dirty bytes count. + * + * First, it adds the initial entry size to aux_ptr->dirty_bytes, + * not the final size. Note that this code used to use the final + * size, but code to support this has been removed from + * H5AC__log_dirtied_entry() for reasons unknown since I wrote this + * code. + * + * As long as all ranks do the same thing here, this probably doesn't + * matter much, although it will delay initiation of sync points. + * + * A more interesting point is that this code will not increment + * aux_ptr->dirty_bytes if a dirty entry is resized. At first glance + * this seems major, as particularly with the older file formats, + * resizes can be quite large. However, this is probably not an + * issue either, since such resizes will be accompanied by large + * amounts of dirty metadata creation in other areas -- which will + * cause aux_ptr->dirty_bytes to be incremented. + * + * The bottom line is that this code is probably OK, but the above + * points should be kept in mind. + * + * One final observation: This comment is occasioned by a bug caused + * by moving the call to H5AC__log_dirtied_entry() after the call to + * H5C_resize_entry(), and then only calling H5AC__log_dirtied_entry() + * if entry_ptr->is_dirty was false. + * + * Since H5C_resize_entry() marks the target entry dirty unless there + * is not change in size, this had the effect of not calling + * H5AC__log_dirtied_entry() when it should be, and corrupting + * the cleaned and dirtied lists used by rank 0 in the parallel + * version of the metadata cache. + * + * The point here is that you should be very careful when working with + * this code, and not modify it unless you fully understand it. + * + * JRM -- 2/28/22 + */ + + if ((!entry_ptr->is_dirty) && (entry_ptr->size != new_size)) { + + /* the entry is clean, and will be marked dirty in the resize + * operation. + */ H5AC_aux_t *aux_ptr; aux_ptr = (H5AC_aux_t *)H5C_get_aux_ptr(cache_ptr); - if ((!entry_ptr->is_dirty) && (NULL != aux_ptr)) + + if (NULL != aux_ptr) { + if (H5AC__log_dirtied_entry(entry_ptr) < 0) HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, "can't log dirtied entry") + } } #endif /* H5_HAVE_PARALLEL */ + /* Resize the entry */ + if (H5C_resize_entry(thing, new_size) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, "can't resize entry") + done: /* If currently logging, generate a message */ if (cache_ptr != NULL && cache_ptr->log_info != NULL) @@ -1636,9 +1697,14 @@ H5AC_unprotect(H5F_t *f, const H5AC_class_t *type, haddr_t addr, void *thing, un if (H5AC__log_dirtied_entry((H5AC_info_t *)thing) < 0) HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, "can't log dirtied entry") - if (deleted && aux_ptr->mpi_rank == 0) - if (H5AC__log_deleted_entry((H5AC_info_t *)thing) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, "H5AC__log_deleted_entry() failed") + if (deleted && aux_ptr->mpi_rank == 0) { + if (H5AC__log_deleted_entry((H5AC_info_t *)thing) < 0) { + /* If we fail to log the deleted entry, push an error but still + * participate in a possible sync point ahead + */ + HDONE_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, "H5AC__log_deleted_entry() failed") + } + } } /* end if */ #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5ACmpio.c b/src/H5ACmpio.c index 500a05a..7eaf751 100644 --- a/src/H5ACmpio.c +++ b/src/H5ACmpio.c @@ -304,8 +304,10 @@ H5AC__broadcast_candidate_list(H5AC_t *cache_ptr, unsigned *num_entries_ptr, had * are used to receiving from process 0, and also load it * into a buffer for transmission. */ - if (H5AC__copy_candidate_list_to_buffer(cache_ptr, &chk_num_entries, &haddr_buf_ptr) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate buffer.") + if (H5AC__copy_candidate_list_to_buffer(cache_ptr, &chk_num_entries, &haddr_buf_ptr) < 0) { + /* Push an error, but still participate in following MPI_Bcast */ + HDONE_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate buffer.") + } HDassert(chk_num_entries == num_entries); HDassert(haddr_buf_ptr != NULL); @@ -428,18 +430,23 @@ H5AC__broadcast_clean_list(H5AC_t *cache_ptr) /* allocate a buffer to store the list of entry base addresses in */ buf_size = sizeof(haddr_t) * num_entries; - if (NULL == (addr_buf_ptr = (haddr_t *)H5MM_malloc(buf_size))) - HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "memory allocation failed for addr buffer") - - /* Set up user data for callback */ - udata.aux_ptr = aux_ptr; - udata.addr_buf_ptr = addr_buf_ptr; - udata.u = 0; - - /* Free all the clean list entries, building the address list in the callback */ - /* (Callback also removes the matching entries from the dirtied list) */ - if (H5SL_free(aux_ptr->c_slist_ptr, H5AC__broadcast_clean_list_cb, &udata) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFREE, FAIL, "Can't build address list for clean entries") + if (NULL == (addr_buf_ptr = (haddr_t *)H5MM_malloc(buf_size))) { + /* Push an error, but still participate in following MPI_Bcast */ + HDONE_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "memory allocation failed for addr buffer") + } + else { + /* Set up user data for callback */ + udata.aux_ptr = aux_ptr; + udata.addr_buf_ptr = addr_buf_ptr; + udata.u = 0; + + /* Free all the clean list entries, building the address list in the callback */ + /* (Callback also removes the matching entries from the dirtied list) */ + if (H5SL_free(aux_ptr->c_slist_ptr, H5AC__broadcast_clean_list_cb, &udata) < 0) { + /* Push an error, but still participate in following MPI_Bcast */ + HDONE_ERROR(H5E_CACHE, H5E_CANTFREE, FAIL, "Can't build address list for clean entries") + } + } /* Now broadcast the list of cleaned entries */ if (MPI_SUCCESS != @@ -1448,8 +1455,10 @@ H5AC__receive_haddr_list(MPI_Comm mpi_comm, unsigned *num_entries_ptr, haddr_t * /* allocate buffers to store the list of entry base addresses in */ buf_size = sizeof(haddr_t) * num_entries; - if (NULL == (haddr_buf_ptr = (haddr_t *)H5MM_malloc(buf_size))) - HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "memory allocation failed for haddr buffer") + if (NULL == (haddr_buf_ptr = (haddr_t *)H5MM_malloc(buf_size))) { + /* Push an error, but still participate in following MPI_Bcast */ + HDONE_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "memory allocation failed for haddr buffer") + } /* Now receive the list of candidate entries */ if (MPI_SUCCESS != @@ -1800,10 +1809,14 @@ H5AC__rsp__dist_md_write__flush_to_min_clean(H5F_t *f) if (evictions_enabled) { /* construct candidate list -- process 0 only */ - if (aux_ptr->mpi_rank == 0) + if (aux_ptr->mpi_rank == 0) { + /* If constructing candidate list fails, push an error but still participate + * in collective operations during following candidate list propagation + */ if (H5AC__construct_candidate_list(cache_ptr, aux_ptr, H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate list.") + HDONE_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate list.") + } /* propagate and apply candidate list -- all processes */ if (H5AC__propagate_and_apply_candidate_list(f) < 0) @@ -1899,15 +1912,21 @@ H5AC__rsp__p0_only__flush(H5F_t *f) aux_ptr->write_permitted = FALSE; /* Check for error on the write operation */ - if (result < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't flush.") - - /* this code exists primarily for the test bed -- it allows us to - * enforce POSIX semantics on the server that pretends to be a - * file system in our parallel tests. - */ - if (aux_ptr->write_done) - (aux_ptr->write_done)(); + if (result < 0) { + /* If write operation fails, push an error but still participate + * in collective operations during following cache entry + * propagation + */ + HDONE_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't flush.") + } + else { + /* this code exists primarily for the test bed -- it allows us to + * enforce POSIX semantics on the server that pretends to be a + * file system in our parallel tests. + */ + if (aux_ptr->write_done) + (aux_ptr->write_done)(); + } } /* end if */ /* Propagate cleaned entries to other ranks. */ @@ -2019,15 +2038,21 @@ H5AC__rsp__p0_only__flush_to_min_clean(H5F_t *f) aux_ptr->write_permitted = FALSE; /* Check for error on the write operation */ - if (result < 0) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5C_flush_to_min_clean() failed.") - - /* this call exists primarily for the test code -- it is used - * to enforce POSIX semantics on the process used to simulate - * reads and writes in t_cache.c. - */ - if (aux_ptr->write_done) - (aux_ptr->write_done)(); + if (result < 0) { + /* If write operation fails, push an error but still participate + * in collective operations during following cache entry + * propagation + */ + HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5C_flush_to_min_clean() failed.") + } + else { + /* this call exists primarily for the test code -- it is used + * to enforce POSIX semantics on the process used to simulate + * reads and writes in t_cache.c. + */ + if (aux_ptr->write_done) + (aux_ptr->write_done)(); + } } /* end if */ if (H5AC__propagate_flushed_and_still_clean_entries_list(f) < 0) @@ -2093,11 +2118,11 @@ H5AC__run_sync_point(H5F_t *f, int sync_point_op) (sync_point_op == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED)); #if H5AC_DEBUG_DIRTY_BYTES_CREATION - HDfprintf(stdout, "%d:H5AC_propagate...:%u: (u/uu/i/iu/r/ru) = %zu/%u/%zu/%u/%zu/%u\n", aux_ptr->mpi_rank, + HDfprintf(stdout, "%d:H5AC_propagate...:%u: (u/uu/i/iu/m/mu) = %zu/%u/%zu/%u/%zu/%u\n", aux_ptr->mpi_rank, aux_ptr->dirty_bytes_propagations, aux_ptr->unprotect_dirty_bytes, aux_ptr->unprotect_dirty_bytes_updates, aux_ptr->insert_dirty_bytes, - aux_ptr->insert_dirty_bytes_updates, aux_ptr->rename_dirty_bytes, - aux_ptr->rename_dirty_bytes_updates); + aux_ptr->insert_dirty_bytes_updates, aux_ptr->move_dirty_bytes, + aux_ptr->move_dirty_bytes_updates); #endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ /* clear collective access flag on half of the entries in the @@ -2161,8 +2186,8 @@ H5AC__run_sync_point(H5F_t *f, int sync_point_op) aux_ptr->unprotect_dirty_bytes_updates = 0; aux_ptr->insert_dirty_bytes = 0; aux_ptr->insert_dirty_bytes_updates = 0; - aux_ptr->rename_dirty_bytes = 0; - aux_ptr->rename_dirty_bytes_updates = 0; + aux_ptr->move_dirty_bytes = 0; + aux_ptr->move_dirty_bytes_updates = 0; #endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ done: diff --git a/src/H5ACpublic.h b/src/H5ACpublic.h index c853794..42bc090 100644 --- a/src/H5ACpublic.h +++ b/src/H5ACpublic.h @@ -28,10 +28,6 @@ #include "H5public.h" #include "H5Cpublic.h" -#ifdef __cplusplus -extern "C" { -#endif - /**************************************************************************** * * structure H5AC_cache_config_t @@ -783,7 +779,4 @@ typedef struct H5AC_cache_image_config_t { //! <!-- [H5AC_cache_image_config_t_snip] --> -#ifdef __cplusplus -} -#endif #endif diff --git a/src/H5B2internal.c b/src/H5B2internal.c index c00f555..a8192df 100644 --- a/src/H5B2internal.c +++ b/src/H5B2internal.c @@ -17,7 +17,7 @@ * Dec 01 2016 * Quincey Koziol * - * Purpose: Routines for managing v2 B-tree internal ndoes. + * Purpose: Routines for managing v2 B-tree internal nodes. * *------------------------------------------------------------------------- */ diff --git a/src/H5B2leaf.c b/src/H5B2leaf.c index 20ace84..f48cf5b 100644 --- a/src/H5B2leaf.c +++ b/src/H5B2leaf.c @@ -17,7 +17,7 @@ * Dec 01 2016 * Quincey Koziol * - * Purpose: Routines for managing v2 B-tree leaf ndoes. + * Purpose: Routines for managing v2 B-tree leaf nodes. * *------------------------------------------------------------------------- */ @@ -138,16 +138,6 @@ static herr_t H5C__generate_image(H5F_t *f, H5C_t *cache_ptr, H5C_cache_entry_t static herr_t H5C__verify_len_eoa(H5F_t *f, const H5C_class_t *type, haddr_t addr, size_t *len, hbool_t actual); -#if H5C_DO_SLIST_SANITY_CHECKS -static hbool_t H5C__entry_in_skip_list(H5C_t *cache_ptr, H5C_cache_entry_t *target_ptr); -#endif /* H5C_DO_SLIST_SANITY_CHECKS */ - -#if H5C_DO_EXTREME_SANITY_CHECKS -static herr_t H5C__validate_lru_list(H5C_t *cache_ptr); -static herr_t H5C__validate_pinned_entry_list(H5C_t *cache_ptr); -static herr_t H5C__validate_protected_entry_list(H5C_t *cache_ptr); -#endif /* H5C_DO_EXTREME_SANITY_CHECKS */ - #ifndef NDEBUG static void H5C__assert_flush_dep_nocycle(const H5C_cache_entry_t *entry, const H5C_cache_entry_t *base_entry); @@ -996,7 +986,7 @@ H5C_expunge_entry(H5F_t *f, const H5C_class_t *type, haddr_t addr, unsigned flag HDassert(H5F_addr_defined(addr)); #if H5C_DO_EXTREME_SANITY_CHECKS - if (H5C__validate_lru_list(cache_ptr) < 0) + if (H5C_validate_lru_list(cache_ptr) < 0) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "LRU extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -1031,7 +1021,7 @@ H5C_expunge_entry(H5F_t *f, const H5C_class_t *type, haddr_t addr, unsigned flag done: #if H5C_DO_EXTREME_SANITY_CHECKS - if (H5C__validate_lru_list(cache_ptr) < 0) + if (H5C_validate_lru_list(cache_ptr) < 0) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "LRU extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -1138,8 +1128,8 @@ H5C_flush_cache(H5F_t *f, unsigned flags) #endif /* H5C_DO_SANITY_CHECKS */ #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -1314,8 +1304,8 @@ H5C_insert_entry(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *thing, u #if H5C_DO_EXTREME_SANITY_CHECKS /* no need to verify that entry is not already in the index as */ /* we already make that check below. */ - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -1424,6 +1414,7 @@ H5C_insert_entry(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *thing, u entry_ptr->serialization_count = 0; #endif /* NDEBUG */ + /* initialize tag list fields */ entry_ptr->tl_next = NULL; entry_ptr->tl_prev = NULL; entry_ptr->tag_info = NULL; @@ -1503,8 +1494,8 @@ H5C_insert_entry(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *thing, u H5C__UPDATE_RP_FOR_INSERTION(cache_ptr, entry_ptr, FAIL) #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed just before done") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -1518,23 +1509,32 @@ H5C_insert_entry(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *thing, u #ifdef H5_HAVE_PARALLEL if (H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) - coll_access = H5CX_get_coll_metadata_read(); + coll_access = H5F_get_coll_metadata_reads(f); entry_ptr->coll_access = coll_access; if (coll_access) { H5C__INSERT_IN_COLL_LIST(cache_ptr, entry_ptr, FAIL) /* Make sure the size of the collective entries in the cache remain in check */ - if (cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) - if (H5C_clear_coll_entries(cache_ptr, TRUE) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries") - } /* end if */ + if (H5P_USER_TRUE == H5F_COLL_MD_READ(f)) { + if (cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) { + if (H5C_clear_coll_entries(cache_ptr, TRUE) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries") + } /* end if */ + } /* end if */ + else { + if (cache_ptr->max_cache_size * 40 < cache_ptr->coll_list_size * 100) { + if (H5C_clear_coll_entries(cache_ptr, TRUE) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries") + } /* end if */ + } /* end else */ + } /* end if */ #endif done: #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -1858,8 +1858,8 @@ H5C_move_entry(H5C_t *cache_ptr, const H5C_class_t *type, haddr_t old_addr, hadd HDassert(H5F_addr_ne(old_addr, new_addr)); #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -1964,8 +1964,8 @@ H5C_move_entry(H5C_t *cache_ptr, const H5C_class_t *type, haddr_t old_addr, hadd done: #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -2011,8 +2011,7 @@ H5C_resize_entry(void *thing, size_t new_size) HGOTO_ERROR(H5E_CACHE, H5E_BADTYPE, FAIL, "Entry isn't pinned or protected??") #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || (H5C_validate_pinned_entry_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -2108,8 +2107,7 @@ H5C_resize_entry(void *thing, size_t new_size) done: #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || (H5C_validate_pinned_entry_list(cache_ptr) < 0)) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -2149,8 +2147,8 @@ H5C_pin_protected_entry(void *thing) HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -2164,8 +2162,8 @@ H5C_pin_protected_entry(void *thing) done: #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -2228,8 +2226,8 @@ H5C_protect(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *udata, unsign HDassert(H5F_addr_defined(addr)); #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, NULL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -2248,7 +2246,7 @@ H5C_protect(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *udata, unsign #ifdef H5_HAVE_PARALLEL if (H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) - coll_access = H5CX_get_coll_metadata_read(); + coll_access = H5F_get_coll_metadata_reads(f); #endif /* H5_HAVE_PARALLEL */ /* first check to see if the target is in cache */ @@ -2307,9 +2305,14 @@ H5C_protect(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *udata, unsign H5MM_memcpy(((uint8_t *)entry_ptr->image_ptr) + entry_ptr->size, H5C_IMAGE_SANITY_VALUE, H5C_IMAGE_EXTRA_SPACE); #endif /* H5C_DO_MEMORY_SANITY_CHECKS */ - if (0 == mpi_rank) - if (H5C__generate_image(f, cache_ptr, entry_ptr) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, NULL, "can't generate entry's image") + if (0 == mpi_rank) { + if (H5C__generate_image(f, cache_ptr, entry_ptr) < 0) { + /* If image generation fails, push an error but + * still participate in the following MPI_Bcast + */ + HDONE_ERROR(H5E_CACHE, H5E_CANTGET, NULL, "can't generate entry's image") + } + } } /* end if */ HDassert(entry_ptr->image_ptr); @@ -2595,16 +2598,24 @@ H5C_protect(H5F_t *f, const H5C_class_t *type, haddr_t addr, void *udata, unsign #ifdef H5_HAVE_PARALLEL /* Make sure the size of the collective entries in the cache remain in check */ - if (coll_access) - if (cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) - if (H5C_clear_coll_entries(cache_ptr, TRUE) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries") -#endif /* H5_HAVE_PARALLEL */ + if (coll_access) { + if (H5P_USER_TRUE == H5F_COLL_MD_READ(f)) { + if (cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) + if (H5C_clear_coll_entries(cache_ptr, TRUE) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries") + } /* end if */ + else { + if (cache_ptr->max_cache_size * 40 < cache_ptr->coll_list_size * 100) + if (H5C_clear_coll_entries(cache_ptr, TRUE) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries") + } /* end else */ + } /* end if */ +#endif /* H5_HAVE_PARALLEL */ done: #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, NULL, "an extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -3077,8 +3088,8 @@ H5C_unpin_entry(void *_entry_ptr) HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -3088,8 +3099,8 @@ H5C_unpin_entry(void *_entry_ptr) done: #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -3256,8 +3267,8 @@ H5C_unprotect(H5F_t *f, haddr_t addr, void *thing, unsigned flags) was_clean = !(entry_ptr->is_dirty); #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -3523,8 +3534,8 @@ H5C_unprotect(H5F_t *f, haddr_t addr, void *thing, unsigned flags) done: #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HDONE_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on exit") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -6058,8 +6069,8 @@ H5C__flush_ring(H5F_t *f, H5C_ring_t ring, unsigned flags) HDassert(ring < H5C_RING_NTYPES); #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ @@ -7182,8 +7193,20 @@ H5C__load_entry(H5F_t *f, #ifdef H5_HAVE_PARALLEL if (!coll_access || 0 == mpi_rank) { #endif /* H5_HAVE_PARALLEL */ - if (H5F_block_read(f, type->mem_type, addr, len, image) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_READERROR, NULL, "Can't read image*") + + if (H5F_block_read(f, type->mem_type, addr, len, image) < 0) { + +#ifdef H5_HAVE_PARALLEL + if (coll_access) { + /* Push an error, but still participate in following MPI_Bcast */ + HDmemset(image, 0, len); + HDONE_ERROR(H5E_CACHE, H5E_READERROR, NULL, "Can't read image*") + } + else +#endif + HGOTO_ERROR(H5E_CACHE, H5E_READERROR, NULL, "Can't read image*") + } + #ifdef H5_HAVE_PARALLEL } /* end if */ /* if the collective metadata read optimization is turned on, @@ -7230,8 +7253,19 @@ H5C__load_entry(H5F_t *f, * loaded thing, go get the on-disk image again (the extra portion). */ if (H5F_block_read(f, type->mem_type, addr + len, actual_len - len, image + len) < - 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "can't read image") + 0) { + +#ifdef H5_HAVE_PARALLEL + if (coll_access) { + /* Push an error, but still participate in following MPI_Bcast */ + HDmemset(image + len, 0, actual_len - len); + HDONE_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "can't read image") + } + else +#endif + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "can't read image") + } + #ifdef H5_HAVE_PARALLEL } /* If the collective metadata read optimization is turned on, @@ -7383,6 +7417,7 @@ H5C__load_entry(H5F_t *f, entry->serialization_count = 0; #endif /* NDEBUG */ + /* initialize tag list fields */ entry->tl_next = NULL; entry->tl_prev = NULL; entry->tag_info = NULL; @@ -7711,7 +7746,7 @@ done: /*------------------------------------------------------------------------- * - * Function: H5C__validate_lru_list + * Function: H5C_validate_lru_list * * Purpose: Debugging function that scans the LRU list for errors. * @@ -7726,15 +7761,15 @@ done: *------------------------------------------------------------------------- */ #if H5C_DO_EXTREME_SANITY_CHECKS -static herr_t -H5C__validate_lru_list(H5C_t *cache_ptr) +herr_t +H5C_validate_lru_list(H5C_t *cache_ptr) { int32_t len = 0; size_t size = 0; H5C_cache_entry_t *entry_ptr = NULL; herr_t ret_value = SUCCEED; /* Return value */ - FUNC_ENTER_STATIC + FUNC_ENTER_NOAPI(FAIL) HDassert(cache_ptr); HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); @@ -7743,51 +7778,48 @@ H5C__validate_lru_list(H5C_t *cache_ptr) (cache_ptr->LRU_head_ptr != cache_ptr->LRU_tail_ptr)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 1 failed") - if (cache_ptr->LRU_list_len < 0) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 2 failed") - if ((cache_ptr->LRU_list_len == 1) && ((cache_ptr->LRU_head_ptr != cache_ptr->LRU_tail_ptr) || (cache_ptr->LRU_head_ptr == NULL) || (cache_ptr->LRU_head_ptr->size != cache_ptr->LRU_list_size))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 3 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 2 failed") if ((cache_ptr->LRU_list_len >= 1) && ((cache_ptr->LRU_head_ptr == NULL) || (cache_ptr->LRU_head_ptr->prev != NULL) || (cache_ptr->LRU_tail_ptr == NULL) || (cache_ptr->LRU_tail_ptr->next != NULL))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 4 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 3 failed") entry_ptr = cache_ptr->LRU_head_ptr; while (entry_ptr != NULL) { if ((entry_ptr != cache_ptr->LRU_head_ptr) && ((entry_ptr->prev == NULL) || (entry_ptr->prev->next != entry_ptr))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 5 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 4 failed") if ((entry_ptr != cache_ptr->LRU_tail_ptr) && ((entry_ptr->next == NULL) || (entry_ptr->next->prev != entry_ptr))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 6 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 5 failed") if ((entry_ptr->is_pinned) || (entry_ptr->pinned_from_client) || (entry_ptr->pinned_from_cache)) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 7 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 6 failed") len++; size += entry_ptr->size; entry_ptr = entry_ptr->next; } - if ((cache_ptr->LRU_list_len != len) || (cache_ptr->LRU_list_size != size)) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 8 failed") + if ((cache_ptr->LRU_list_len != (uint32_t)len) || (cache_ptr->LRU_list_size != size)) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 7 failed") done: if (ret_value != SUCCEED) HDassert(0); FUNC_LEAVE_NOAPI(ret_value) -} /* H5C__validate_lru_list() */ +} /* H5C_validate_lru_list() */ #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ /*------------------------------------------------------------------------- * - * Function: H5C__validate_pinned_entry_list + * Function: H5C_validate_pinned_entry_list * * Purpose: Debugging function that scans the pinned entry list for * errors. @@ -7803,15 +7835,15 @@ done: *------------------------------------------------------------------------- */ #if H5C_DO_EXTREME_SANITY_CHECKS -static herr_t -H5C__validate_pinned_entry_list(H5C_t *cache_ptr) +herr_t +H5C_validate_pinned_entry_list(H5C_t *cache_ptr) { int32_t len = 0; size_t size = 0; H5C_cache_entry_t *entry_ptr = NULL; herr_t ret_value = SUCCEED; /* Return value */ - FUNC_ENTER_STATIC + FUNC_ENTER_NOAPI(FAIL) HDassert(cache_ptr); HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); @@ -7820,54 +7852,51 @@ H5C__validate_pinned_entry_list(H5C_t *cache_ptr) (cache_ptr->pel_head_ptr != cache_ptr->pel_tail_ptr)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 1 failed") - if (cache_ptr->pel_len < 0) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 2 failed") - if ((cache_ptr->pel_len == 1) && ((cache_ptr->pel_head_ptr != cache_ptr->pel_tail_ptr) || (cache_ptr->pel_head_ptr == NULL) || (cache_ptr->pel_head_ptr->size != cache_ptr->pel_size))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 3 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 2 failed") if ((cache_ptr->pel_len >= 1) && ((cache_ptr->pel_head_ptr == NULL) || (cache_ptr->pel_head_ptr->prev != NULL) || (cache_ptr->pel_tail_ptr == NULL) || (cache_ptr->pel_tail_ptr->next != NULL))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 4 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 3 failed") entry_ptr = cache_ptr->pel_head_ptr; while (entry_ptr != NULL) { if ((entry_ptr != cache_ptr->pel_head_ptr) && ((entry_ptr->prev == NULL) || (entry_ptr->prev->next != entry_ptr))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 5 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 4 failed") if ((entry_ptr != cache_ptr->pel_tail_ptr) && ((entry_ptr->next == NULL) || (entry_ptr->next->prev != entry_ptr))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 6 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 5 failed") if (!entry_ptr->is_pinned) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 7 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 6 failed") if (!(entry_ptr->pinned_from_client || entry_ptr->pinned_from_cache)) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 8 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 7 failed") len++; size += entry_ptr->size; entry_ptr = entry_ptr->next; } - if ((cache_ptr->pel_len != len) || (cache_ptr->pel_size != size)) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 9 failed") + if ((cache_ptr->pel_len != (uint32_t)len) || (cache_ptr->pel_size != size)) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 8 failed") done: if (ret_value != SUCCEED) HDassert(0); FUNC_LEAVE_NOAPI(ret_value) -} /* H5C__validate_pinned_entry_list() */ +} /* H5C_validate_pinned_entry_list() */ #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ /*------------------------------------------------------------------------- * - * Function: H5C__validate_protected_entry_list + * Function: H5C_validate_protected_entry_list * * Purpose: Debugging function that scans the protected entry list for * errors. @@ -7883,15 +7912,15 @@ done: *------------------------------------------------------------------------- */ #if H5C_DO_EXTREME_SANITY_CHECKS -static herr_t -H5C__validate_protected_entry_list(H5C_t *cache_ptr) +herr_t +H5C_validate_protected_entry_list(H5C_t *cache_ptr) { int32_t len = 0; size_t size = 0; H5C_cache_entry_t *entry_ptr = NULL; herr_t ret_value = SUCCEED; /* Return value */ - FUNC_ENTER_STATIC + FUNC_ENTER_NOAPI(FAIL) HDassert(cache_ptr); HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); @@ -7900,54 +7929,51 @@ H5C__validate_protected_entry_list(H5C_t *cache_ptr) (cache_ptr->pl_head_ptr != cache_ptr->pl_tail_ptr)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 1 failed") - if (cache_ptr->pl_len < 0) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 2 failed") - if ((cache_ptr->pl_len == 1) && ((cache_ptr->pl_head_ptr != cache_ptr->pl_tail_ptr) || (cache_ptr->pl_head_ptr == NULL) || (cache_ptr->pl_head_ptr->size != cache_ptr->pl_size))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 3 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 2 failed") if ((cache_ptr->pl_len >= 1) && ((cache_ptr->pl_head_ptr == NULL) || (cache_ptr->pl_head_ptr->prev != NULL) || (cache_ptr->pl_tail_ptr == NULL) || (cache_ptr->pl_tail_ptr->next != NULL))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 4 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 3 failed") entry_ptr = cache_ptr->pl_head_ptr; while (entry_ptr != NULL) { if ((entry_ptr != cache_ptr->pl_head_ptr) && ((entry_ptr->prev == NULL) || (entry_ptr->prev->next != entry_ptr))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 5 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 4 failed") if ((entry_ptr != cache_ptr->pl_tail_ptr) && ((entry_ptr->next == NULL) || (entry_ptr->next->prev != entry_ptr))) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 6 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 5 failed") if (!entry_ptr->is_protected) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 7 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 6 failed") if (entry_ptr->is_read_only && (entry_ptr->ro_ref_count <= 0)) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 8 failed") + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 7 failed") len++; size += entry_ptr->size; entry_ptr = entry_ptr->next; } - if ((cache_ptr->pl_len != len) || (cache_ptr->pl_size != size)) - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 9 failed") + if ((cache_ptr->pl_len != (uint32_t)len) || (cache_ptr->pl_size != size)) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 8 failed") done: if (ret_value != SUCCEED) HDassert(0); FUNC_LEAVE_NOAPI(ret_value) -} /* H5C__validate_protected_entry_list() */ +} /* H5C_validate_protected_entry_list() */ #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ /*------------------------------------------------------------------------- * - * Function: H5C__entry_in_skip_list + * Function: H5C_entry_in_skip_list * * Purpose: Debugging function that scans skip list to see if it * is in present. We need this, as it is possible for @@ -7961,8 +7987,8 @@ done: *------------------------------------------------------------------------- */ #if H5C_DO_SLIST_SANITY_CHECKS -static hbool_t -H5C__entry_in_skip_list(H5C_t *cache_ptr, H5C_cache_entry_t *target_ptr) +hbool_t +H5C_entry_in_skip_list(H5C_t *cache_ptr, H5C_cache_entry_t *target_ptr) { H5SL_node_t *node_ptr; hbool_t in_slist; @@ -7990,7 +8016,7 @@ H5C__entry_in_skip_list(H5C_t *cache_ptr, H5C_cache_entry_t *target_ptr) } return (in_slist); -} /* H5C__entry_in_skip_list() */ +} /* H5C_entry_in_skip_list() */ #endif /* H5C_DO_SLIST_SANITY_CHECKS */ /*------------------------------------------------------------------------- @@ -8480,8 +8506,8 @@ H5C__serialize_cache(H5F_t *f) #endif /* H5C_DO_SANITY_CHECKS */ #if H5C_DO_EXTREME_SANITY_CHECKS - if ((H5C__validate_protected_entry_list(cache_ptr) < 0) || - (H5C__validate_pinned_entry_list(cache_ptr) < 0) || (H5C__validate_lru_list(cache_ptr) < 0)) + if ((H5C_validate_protected_entry_list(cache_ptr) < 0) || + (H5C_validate_pinned_entry_list(cache_ptr) < 0) || (H5C_validate_lru_list(cache_ptr) < 0)) HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "an extreme sanity check failed on entry") #endif /* H5C_DO_EXTREME_SANITY_CHECKS */ diff --git a/src/H5CSprivate.h b/src/H5CSprivate.h index a238ec7..2dc28f5 100644 --- a/src/H5CSprivate.h +++ b/src/H5CSprivate.h @@ -17,10 +17,6 @@ #ifndef H5CSprivate_H #define H5CSprivate_H -#ifdef NOT_YET -#include "H5CSpublic.h" -#endif /* NOT_YET */ - /* Private headers needed by this file */ #include "H5private.h" @@ -1397,9 +1397,7 @@ H5CX_set_apl(hid_t *acspl_id, const H5P_libclass_t *libclass, /* If parallel is enabled and the file driver used is the MPI-IO * VFD, issue an MPI barrier for easier debugging if the API function - * calling this is supposed to be called collectively. Note that this - * happens only when the environment variable H5_COLL_BARRIER is set - * to non 0. + * calling this is supposed to be called collectively. */ if (H5_coll_api_sanity_check_g) { MPI_Comm mpi_comm; /* File communicator */ @@ -1425,7 +1423,7 @@ done: * Purpose: Sanity checks and sets up collective operations. * * Note: Should be called for all API routines that modify file - * file metadata but don't pass in an access property list. + * metadata but don't pass in an access property list. * * Return: Non-negative on success / Negative on failure * @@ -1456,9 +1454,7 @@ H5CX_set_loc(hid_t /* If parallel is enabled and the file driver used is the MPI-IO * VFD, issue an MPI barrier for easier debugging if the API function - * calling this is supposed to be called collectively. Note that this - * happens only when the environment variable H5_COLL_BARRIER is set - * to non 0. + * calling this is supposed to be called collectively. */ if (H5_coll_api_sanity_check_g) { MPI_Comm mpi_comm; /* File communicator */ diff --git a/src/H5CXprivate.h b/src/H5CXprivate.h index 8ec1c59..878bcf6 100644 --- a/src/H5CXprivate.h +++ b/src/H5CXprivate.h @@ -16,11 +16,6 @@ #ifndef H5CXprivate_H #define H5CXprivate_H -/* Include package's public header */ -#ifdef NOT_YET -#include "H5CXpublic.h" -#endif /* NOT_YET */ - /* Private headers needed by this file */ #include "H5private.h" /* Generic Functions */ #include "H5ACprivate.h" /* Metadata cache */ diff --git a/src/H5Cepoch.c b/src/H5Cepoch.c index 3434fed..8655881 100644 --- a/src/H5Cepoch.c +++ b/src/H5Cepoch.c @@ -78,22 +78,21 @@ static herr_t H5C__epoch_marker_fsf_size(const void H5_ATTR_UNUSED *thing, /* Local Variables */ /*******************/ -const H5AC_class_t H5AC_EPOCH_MARKER[1] = {{ - /* id = */ H5AC_EPOCH_MARKER_ID, - /* name = */ "epoch marker", - /* mem_type = */ H5FD_MEM_DEFAULT, /* value doesn't matter */ - /* flags = */ H5AC__CLASS_NO_FLAGS_SET, - /* get_initial_load_size = */ H5C__epoch_marker_get_initial_load_size, - /* get_final_load_size = */ H5C__epoch_marker_get_final_load_size, - /* verify_chksum = */ H5C__epoch_marker_verify_chksum, - /* deserialize = */ H5C__epoch_marker_deserialize, - /* image_len = */ H5C__epoch_marker_image_len, - /* pre_serialize = */ H5C__epoch_marker_pre_serialize, - /* serialize = */ H5C__epoch_marker_serialize, - /* notify = */ H5C__epoch_marker_notify, - /* free_icr = */ H5C__epoch_marker_free_icr, - /* fsf_size = */ H5C__epoch_marker_fsf_size, -}}; +const H5AC_class_t H5AC_EPOCH_MARKER[1] = { + {/* id = */ H5AC_EPOCH_MARKER_ID, + /* name = */ "epoch marker", + /* mem_type = */ H5FD_MEM_DEFAULT, /* value doesn't matter */ + /* flags = */ H5AC__CLASS_NO_FLAGS_SET, + /* get_initial_load_size = */ H5C__epoch_marker_get_initial_load_size, + /* get_final_load_size = */ H5C__epoch_marker_get_final_load_size, + /* verify_chksum = */ H5C__epoch_marker_verify_chksum, + /* deserialize = */ H5C__epoch_marker_deserialize, + /* image_len = */ H5C__epoch_marker_image_len, + /* pre_serialize = */ H5C__epoch_marker_pre_serialize, + /* serialize = */ H5C__epoch_marker_serialize, + /* notify = */ H5C__epoch_marker_notify, + /* free_icr = */ H5C__epoch_marker_free_icr, + /* fsf_size = */ H5C__epoch_marker_fsf_size}}; /*************************************************************************** * Class functions for H5C__EPOCH_MAKER_TYPE: diff --git a/src/H5Cimage.c b/src/H5Cimage.c index 7421c90..98c1291 100644 --- a/src/H5Cimage.c +++ b/src/H5Cimage.c @@ -997,6 +997,9 @@ H5C__read_cache_image(H5F_t *f, H5C_t *cache_ptr) #endif /* H5_HAVE_PARALLEL */ /* Read the buffer (if serial access, or rank 0 of parallel access) */ + /* NOTE: if this block read is being performed on rank 0 only, throwing + * an error here will cause other ranks to hang in the following MPI_Bcast. + */ if (H5F_block_read(f, H5FD_MEM_SUPER, cache_ptr->image_addr, cache_ptr->image_len, cache_ptr->image_buffer) < 0) HGOTO_ERROR(H5E_CACHE, H5E_READERROR, FAIL, "Can't read metadata cache image block") diff --git a/src/H5Cpkg.h b/src/H5Cpkg.h index 30b86b9..61c3afc 100644 --- a/src/H5Cpkg.h +++ b/src/H5Cpkg.h @@ -1011,7 +1011,7 @@ if ( ( (cache_ptr) == NULL ) || \ ( H5C__HASH_FCN((entry_ptr)->addr) >= H5C__HASH_TABLE_LEN ) || \ ( (cache_ptr)->index_size != \ ((cache_ptr)->clean_index_size + \ - (cache_ptr)->dirty_index_size) ) || \ + (cache_ptr)->dirty_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->clean_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->dirty_index_size) ) || \ ( (entry_ptr)->ring <= H5C_RING_UNDEFINED ) || \ @@ -1034,7 +1034,7 @@ if ( ( (cache_ptr) == NULL ) || \ ( (cache_ptr)->magic != H5C__H5C_T_MAGIC ) || \ ( (cache_ptr)->index_size != \ ((cache_ptr)->clean_index_size + \ - (cache_ptr)->dirty_index_size) ) || \ + (cache_ptr)->dirty_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->clean_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->dirty_index_size) ) || \ ( (cache_ptr)->index_ring_len[(entry_ptr)->ring] == 0 ) || \ @@ -1071,7 +1071,7 @@ if ( ( (cache_ptr) == NULL ) || \ ( (entry_ptr)->ht_prev != NULL ) ) || \ ( (cache_ptr)->index_size != \ ((cache_ptr)->clean_index_size + \ - (cache_ptr)->dirty_index_size) ) || \ + (cache_ptr)->dirty_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->clean_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->dirty_index_size) ) || \ ( (entry_ptr)->ring <= H5C_RING_UNDEFINED ) || \ @@ -1102,7 +1102,7 @@ if ( ( (cache_ptr) == NULL ) || \ ( (entry_ptr)->ht_prev != NULL ) || \ ( (cache_ptr)->index_size != \ ((cache_ptr)->clean_index_size + \ - (cache_ptr)->dirty_index_size) ) || \ + (cache_ptr)->dirty_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->clean_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->dirty_index_size) ) || \ ( (cache_ptr)->index_ring_len[(entry_ptr)->ring] > \ @@ -1161,7 +1161,7 @@ if ( ( (cache_ptr) == NULL ) || \ } #define H5C__PRE_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size, \ - entry_ptr, was_clean) \ + entry_ptr, was_clean) \ if ( ( (cache_ptr) == NULL ) || \ ( (cache_ptr)->index_len <= 0 ) || \ ( (cache_ptr)->index_size <= 0 ) || \ @@ -1175,9 +1175,9 @@ if ( ( (cache_ptr) == NULL ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->clean_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->dirty_index_size) ) || \ ( ( !( was_clean ) || \ - ( (cache_ptr)->clean_index_size < (old_size) ) ) && \ - ( ( (was_clean) ) || \ - ( (cache_ptr)->dirty_index_size < (old_size) ) ) ) || \ + ( (cache_ptr)->clean_index_size < (old_size) ) ) && \ + ( ( (was_clean) ) || \ + ( (cache_ptr)->dirty_index_size < (old_size) ) ) ) || \ ( (entry_ptr) == NULL ) || \ ( (entry_ptr)->ring <= H5C_RING_UNDEFINED ) || \ ( (entry_ptr)->ring >= H5C_RING_NTYPES ) || \ @@ -1196,20 +1196,20 @@ if ( ( (cache_ptr) == NULL ) || \ } #define H5C__POST_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size, \ - entry_ptr) \ + entry_ptr) \ if ( ( (cache_ptr) == NULL ) || \ ( (cache_ptr)->index_len <= 0 ) || \ ( (cache_ptr)->index_size <= 0 ) || \ ( (new_size) > (cache_ptr)->index_size ) || \ ( (cache_ptr)->index_size != \ - ((cache_ptr)->clean_index_size + \ + ((cache_ptr)->clean_index_size + \ (cache_ptr)->dirty_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->clean_index_size) ) || \ ( (cache_ptr)->index_size < ((cache_ptr)->dirty_index_size) ) || \ ( ( !((entry_ptr)->is_dirty ) || \ - ( (cache_ptr)->dirty_index_size < (new_size) ) ) && \ - ( ( ((entry_ptr)->is_dirty) ) || \ - ( (cache_ptr)->clean_index_size < (new_size) ) ) ) || \ + ( (cache_ptr)->dirty_index_size < (new_size) ) ) && \ + ( ( ((entry_ptr)->is_dirty) ) || \ + ( (cache_ptr)->clean_index_size < (new_size) ) ) ) || \ ( ( (cache_ptr)->index_len == 1 ) && \ ( (cache_ptr)->index_size != (new_size) ) ) || \ ( (cache_ptr)->index_ring_len[(entry_ptr)->ring] > \ @@ -1465,10 +1465,10 @@ if ( ( (cache_ptr)->index_size != \ H5C__PRE_HT_UPDATE_FOR_ENTRY_CLEAN_SC(cache_ptr, entry_ptr); \ (cache_ptr)->dirty_index_size -= (entry_ptr)->size; \ ((cache_ptr)->dirty_index_ring_size[entry_ptr->ring]) \ - -= (entry_ptr)->size; \ + -= (entry_ptr)->size; \ (cache_ptr)->clean_index_size += (entry_ptr)->size; \ ((cache_ptr)->clean_index_ring_size[entry_ptr->ring]) \ - += (entry_ptr)->size; \ + += (entry_ptr)->size; \ H5C__POST_HT_UPDATE_FOR_ENTRY_CLEAN_SC(cache_ptr, entry_ptr); \ } @@ -1477,18 +1477,18 @@ if ( ( (cache_ptr)->index_size != \ H5C__PRE_HT_UPDATE_FOR_ENTRY_DIRTY_SC(cache_ptr, entry_ptr); \ (cache_ptr)->clean_index_size -= (entry_ptr)->size; \ ((cache_ptr)->clean_index_ring_size[entry_ptr->ring]) \ - -= (entry_ptr)->size; \ + -= (entry_ptr)->size; \ (cache_ptr)->dirty_index_size += (entry_ptr)->size; \ ((cache_ptr)->dirty_index_ring_size[entry_ptr->ring]) \ - += (entry_ptr)->size; \ + += (entry_ptr)->size; \ H5C__POST_HT_UPDATE_FOR_ENTRY_DIRTY_SC(cache_ptr, entry_ptr); \ } #define H5C__UPDATE_INDEX_FOR_SIZE_CHANGE(cache_ptr, old_size, new_size, \ - entry_ptr, was_clean) \ + entry_ptr, was_clean) \ { \ H5C__PRE_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size, \ - entry_ptr, was_clean) \ + entry_ptr, was_clean) \ (cache_ptr)->index_size -= (old_size); \ (cache_ptr)->index_size += (new_size); \ ((cache_ptr)->index_ring_size[entry_ptr->ring]) -= (old_size); \ @@ -1497,14 +1497,14 @@ if ( ( (cache_ptr)->index_size != \ (cache_ptr)->clean_index_size -= (old_size); \ ((cache_ptr)->clean_index_ring_size[entry_ptr->ring])-= (old_size); \ } else { \ - (cache_ptr)->dirty_index_size -= (old_size); \ + (cache_ptr)->dirty_index_size -= (old_size); \ ((cache_ptr)->dirty_index_ring_size[entry_ptr->ring])-= (old_size); \ } \ if((entry_ptr)->is_dirty) { \ (cache_ptr)->dirty_index_size += (new_size); \ ((cache_ptr)->dirty_index_ring_size[entry_ptr->ring])+= (new_size); \ } else { \ - (cache_ptr)->clean_index_size += (new_size); \ + (cache_ptr)->clean_index_size += (new_size); \ ((cache_ptr)->clean_index_ring_size[entry_ptr->ring])+= (new_size); \ } \ H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->il_len, \ @@ -1791,7 +1791,7 @@ if ( ( (cache_ptr)->index_size != \ } else { /* slist disabled */ \ \ HDassert( (cache_ptr)->slist_len == 0 ); \ - HDassert( (cache_ptr)->slist_size == 0 ); \ + HDassert( (cache_ptr)->slist_size == 0 ); \ } \ } /* H5C__REMOVE_ENTRY_FROM_SLIST */ @@ -2033,16 +2033,16 @@ if ( ( (cache_ptr)->index_size != \ /* modified LRU specific code */ \ \ /* remove the entry from the LRU list, and re-insert it at the head.\ - */ \ + */ \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* Use the dirty flag to infer whether the entry is on the clean or \ @@ -2096,16 +2096,16 @@ if ( ( (cache_ptr)->index_size != \ /* modified LRU specific code */ \ \ /* remove the entry from the LRU list, and re-insert it at the head \ - */ \ + */ \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* End modified LRU specific code. */ \ @@ -2288,28 +2288,28 @@ if ( ( (cache_ptr)->index_size != \ /* modified LRU specific code */ \ \ /* remove the entry from the LRU list, and re-insert it at the \ - * head. \ - */ \ + * head. \ + */ \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* since the entry is being flushed or cleared, one would think \ - * that it must be dirty -- but that need not be the case. Use the \ - * dirty flag to infer whether the entry is on the clean or dirty \ - * LRU list, and remove it. Then insert it at the head of the \ - * clean LRU list. \ + * that it must be dirty -- but that need not be the case. Use the \ + * dirty flag to infer whether the entry is on the clean or dirty \ + * LRU list, and remove it. Then insert it at the head of the \ + * clean LRU list. \ * \ * The function presumes that a dirty entry will be either cleared \ - * or flushed shortly, so it is OK if we put a dirty entry on the \ - * clean LRU list. \ + * or flushed shortly, so it is OK if we put a dirty entry on the \ + * clean LRU list. \ */ \ \ if ( (entry_ptr)->is_dirty ) { \ @@ -2350,17 +2350,17 @@ if ( ( (cache_ptr)->index_size != \ /* modified LRU specific code */ \ \ /* remove the entry from the LRU list, and re-insert it at the \ - * head. \ - */ \ + * head. \ + */ \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* End modified LRU specific code. */ \ @@ -2424,7 +2424,7 @@ if ( ( (cache_ptr)->index_size != \ \ H5C__DLL_APPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* insert the entry at the tail of the clean or dirty LRU list as \ @@ -2465,7 +2465,7 @@ if ( ( (cache_ptr)->index_size != \ (cache_ptr)->pel_tail_ptr, \ (cache_ptr)->pel_len, \ (cache_ptr)->pel_size, (fail_val)) \ - \ + \ } else { \ \ /* modified LRU specific code */ \ @@ -2474,7 +2474,7 @@ if ( ( (cache_ptr)->index_size != \ \ H5C__DLL_APPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* End modified LRU specific code. */ \ @@ -2558,7 +2558,7 @@ if ( ( (cache_ptr)->index_size != \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* insert the entry at the head of the clean or dirty LRU list as \ @@ -2599,7 +2599,7 @@ if ( ( (cache_ptr)->index_size != \ (cache_ptr)->pel_tail_ptr, \ (cache_ptr)->pel_len, \ (cache_ptr)->pel_size, (fail_val)) \ - \ + \ } else { \ \ /* modified LRU specific code */ \ @@ -2608,7 +2608,7 @@ if ( ( (cache_ptr)->index_size != \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* End modified LRU specific code. */ \ @@ -2677,12 +2677,12 @@ if ( ( (cache_ptr)->index_size != \ HDassert( !((entry_ptr)->is_read_only) ); \ HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ HDassert( (entry_ptr)->size > 0 ); \ - \ + \ if ( (entry_ptr)->is_pinned ) { \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->pel_head_ptr, \ - (cache_ptr)->pel_tail_ptr, \ - (cache_ptr)->pel_len, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ (cache_ptr)->pel_size, (fail_val)) \ \ } else { \ @@ -2693,7 +2693,7 @@ if ( ( (cache_ptr)->index_size != \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* Similarly, remove the entry from the clean or dirty LRU list \ @@ -2739,12 +2739,12 @@ if ( ( (cache_ptr)->index_size != \ HDassert( !((entry_ptr)->is_read_only) ); \ HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ HDassert( (entry_ptr)->size > 0 ); \ - \ + \ if ( (entry_ptr)->is_pinned ) { \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->pel_head_ptr, \ - (cache_ptr)->pel_tail_ptr, \ - (cache_ptr)->pel_len, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ (cache_ptr)->pel_size, (fail_val)) \ \ } else { \ @@ -2755,7 +2755,7 @@ if ( ( (cache_ptr)->index_size != \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* End modified LRU specific code. */ \ @@ -2804,21 +2804,21 @@ if ( ( (cache_ptr)->index_size != \ HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ HDassert( (entry_ptr)->size > 0 ); \ \ - if ( ! ( (entry_ptr)->is_pinned ) && ! ( (entry_ptr->is_protected ) ) ) { \ - \ + if ( ! ( (entry_ptr)->is_pinned ) && ! ( (entry_ptr->is_protected ) ) ) {\ + \ /* modified LRU specific code */ \ \ /* remove the entry from the LRU list, and re-insert it at the head. \ - */ \ + */ \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* remove the entry from either the clean or dirty LUR list as \ @@ -2827,7 +2827,7 @@ if ( ( (cache_ptr)->index_size != \ if ( was_dirty ) { \ \ H5C__AUX_DLL_REMOVE((entry_ptr), \ - (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_head_ptr, \ (cache_ptr)->dLRU_tail_ptr, \ (cache_ptr)->dLRU_list_len, \ (cache_ptr)->dLRU_list_size, \ @@ -2836,34 +2836,34 @@ if ( ( (cache_ptr)->index_size != \ } else { \ \ H5C__AUX_DLL_REMOVE((entry_ptr), \ - (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_head_ptr, \ (cache_ptr)->cLRU_tail_ptr, \ (cache_ptr)->cLRU_list_len, \ (cache_ptr)->cLRU_list_size, \ - (fail_val)) \ + (fail_val)) \ } \ \ /* insert the entry at the head of either the clean or dirty \ - * LRU list as appropriate. \ + * LRU list as appropriate. \ */ \ \ if ( (entry_ptr)->is_dirty ) { \ \ H5C__AUX_DLL_PREPEND((entry_ptr), \ - (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_head_ptr, \ (cache_ptr)->dLRU_tail_ptr, \ (cache_ptr)->dLRU_list_len, \ (cache_ptr)->dLRU_list_size, \ - (fail_val)) \ + (fail_val)) \ \ } else { \ \ H5C__AUX_DLL_PREPEND((entry_ptr), \ - (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_head_ptr, \ (cache_ptr)->cLRU_tail_ptr, \ (cache_ptr)->cLRU_list_len, \ (cache_ptr)->cLRU_list_size, \ - (fail_val)) \ + (fail_val)) \ } \ \ /* End modified LRU specific code. */ \ @@ -2872,7 +2872,7 @@ if ( ( (cache_ptr)->index_size != \ #else /* H5C_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ -#define H5C__UPDATE_RP_FOR_MOVE(cache_ptr, entry_ptr, was_dirty, fail_val) \ +#define H5C__UPDATE_RP_FOR_MOVE(cache_ptr, entry_ptr, was_dirty, fail_val) \ { \ HDassert( (cache_ptr) ); \ HDassert( (cache_ptr)->magic == H5C__H5C_T_MAGIC ); \ @@ -2881,21 +2881,21 @@ if ( ( (cache_ptr)->index_size != \ HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ HDassert( (entry_ptr)->size > 0 ); \ \ - if ( ! ( (entry_ptr)->is_pinned ) && ! ( (entry_ptr->is_protected ) ) ) { \ - \ + if ( ! ( (entry_ptr)->is_pinned ) && ! ( (entry_ptr->is_protected ) ) ) {\ + \ /* modified LRU specific code */ \ \ /* remove the entry from the LRU list, and re-insert it at the head. \ - */ \ + */ \ \ H5C__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ H5C__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ (cache_ptr)->LRU_tail_ptr, \ - (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_len, \ (cache_ptr)->LRU_list_size, (fail_val)) \ \ /* End modified LRU specific code. */ \ @@ -2952,49 +2952,49 @@ if ( ( (cache_ptr)->index_size != \ \ if ( (entry_ptr)->coll_access ) { \ \ - H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->coll_list_len, \ - (cache_ptr)->coll_list_size, \ - (entry_ptr)->size, \ - (new_size)); \ - \ + H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->coll_list_len, \ + (cache_ptr)->coll_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ + \ } \ \ if ( (entry_ptr)->is_pinned ) { \ \ - H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->pel_len, \ - (cache_ptr)->pel_size, \ - (entry_ptr)->size, \ - (new_size)); \ - \ + H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, \ + (entry_ptr)->size, \ + (new_size)); \ + \ } else { \ \ /* modified LRU specific code */ \ \ - /* Update the size of the LRU list */ \ + /* Update the size of the LRU list */ \ \ - H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->LRU_list_len, \ - (cache_ptr)->LRU_list_size, \ - (entry_ptr)->size, \ - (new_size)); \ + H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ \ /* Similarly, update the size of the clean or dirty LRU list as \ - * appropriate. At present, the entry must be clean, but that \ - * could change. \ + * appropriate. At present, the entry must be clean, but that \ + * could change. \ */ \ \ if ( (entry_ptr)->is_dirty ) { \ \ - H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->dLRU_list_len, \ - (cache_ptr)->dLRU_list_size, \ - (entry_ptr)->size, \ - (new_size)); \ + H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ \ } else { \ \ - H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->cLRU_list_len, \ - (cache_ptr)->cLRU_list_size, \ - (entry_ptr)->size, \ - (new_size)); \ + H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ } \ \ /* End modified LRU specific code. */ \ @@ -3017,21 +3017,21 @@ if ( ( (cache_ptr)->index_size != \ \ if ( (entry_ptr)->is_pinned ) { \ \ - H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->pel_len, \ - (cache_ptr)->pel_size, \ - (entry_ptr)->size, \ - (new_size)); \ + H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, \ + (entry_ptr)->size, \ + (new_size)); \ \ } else { \ \ /* modified LRU specific code */ \ \ - /* Update the size of the LRU list */ \ + /* Update the size of the LRU list */ \ \ - H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->LRU_list_len, \ - (cache_ptr)->LRU_list_size, \ - (entry_ptr)->size, \ - (new_size)); \ + H5C__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ \ /* End modified LRU specific code. */ \ } \ @@ -3318,7 +3318,7 @@ if ( ( (hd_ptr) == NULL ) || \ ( (Size) < (entry_ptr)->size ) || \ ( ( (Size) == (entry_ptr)->size ) && ( ! ( (len) == 1 ) ) ) || \ ( ( (entry_ptr)->coll_prev == NULL ) && ( (hd_ptr) != (entry_ptr) ) ) || \ - ( ( (entry_ptr)->coll_next == NULL ) && ( (tail_ptr) != (entry_ptr) ) ) || \ + ( ( (entry_ptr)->coll_next == NULL ) && ( (tail_ptr) != (entry_ptr) ) ) ||\ ( ( (len) == 1 ) && \ ( ! ( ( (hd_ptr) == (entry_ptr) ) && ( (tail_ptr) == (entry_ptr) ) && \ ( (entry_ptr)->coll_next == NULL ) && \ @@ -3350,10 +3350,10 @@ if ( ( ( ( (head_ptr) == NULL ) || ( (tail_ptr) == NULL ) ) && \ ) \ ) { \ HDassert(0 && "COLL DLL sanity check failed"); \ - HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "COLL DLL sanity check failed") \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "COLL DLL sanity check failed")\ } -#define H5C__COLL_DLL_PRE_INSERT_SC(entry_ptr, hd_ptr, tail_ptr, len, Size, fv) \ +#define H5C__COLL_DLL_PRE_INSERT_SC(entry_ptr, hd_ptr, tail_ptr, len, Size, fv)\ if ( ( (entry_ptr) == NULL ) || \ ( (entry_ptr)->coll_next != NULL ) || \ ( (entry_ptr)->coll_prev != NULL ) || \ @@ -5074,7 +5074,7 @@ H5_DLL herr_t H5C__generate_cache_image(H5F_t *f, H5C_t *cache_ptr); H5_DLL herr_t H5C__load_cache_image(H5F_t *f); H5_DLL herr_t H5C__mark_flush_dep_serialized(H5C_cache_entry_t * entry_ptr); H5_DLL herr_t H5C__mark_flush_dep_unserialized(H5C_cache_entry_t * entry_ptr); -H5_DLL herr_t H5C__make_space_in_cache(H5F_t * f, size_t space_needed, +H5_DLL herr_t H5C__make_space_in_cache(H5F_t * f, size_t space_needed, hbool_t write_permitted); H5_DLL herr_t H5C__flush_marked_entries(H5F_t * f); H5_DLL herr_t H5C__serialize_cache(H5F_t *f); diff --git a/src/H5Cprivate.h b/src/H5Cprivate.h index 8a1043e..9514443 100644 --- a/src/H5Cprivate.h +++ b/src/H5Cprivate.h @@ -2292,6 +2292,16 @@ H5_DLL herr_t H5C_cache_image_status(H5F_t *f, hbool_t *load_ci_ptr, hbool_t * H5_DLL hbool_t H5C_cache_image_pending(const H5C_t *cache_ptr); H5_DLL herr_t H5C_get_mdc_image_info(const H5C_t *cache_ptr, haddr_t *image_addr, hsize_t *image_len); +#if H5C_DO_SLIST_SANITY_CHECKS +H5_DLL hbool_t H5C_entry_in_skip_list(H5C_t *cache_ptr, H5C_cache_entry_t *target_ptr); +#endif + +#if H5C_DO_EXTREME_SANITY_CHECKS +H5_DLL herr_t H5C_validate_lru_list(H5C_t *cache_ptr); +H5_DLL herr_t H5C_validate_pinned_entry_list(H5C_t *cache_ptr); +H5_DLL herr_t H5C_validate_protected_entry_list(H5C_t *cache_ptr); +#endif /* H5C_DO_EXTREME_SANITY_CHECKS */ + /* Logging functions */ H5_DLL herr_t H5C_start_logging(H5C_t *cache); H5_DLL herr_t H5C_stop_logging(H5C_t *cache); diff --git a/src/H5Cpublic.h b/src/H5Cpublic.h index 79ece10..c65dc7c 100644 --- a/src/H5Cpublic.h +++ b/src/H5Cpublic.h @@ -27,10 +27,6 @@ /* Public headers needed by this file */ #include "H5public.h" -#ifdef __cplusplus -extern "C" { -#endif - enum H5C_cache_incr_mode { H5C_incr__off, /**<Automatic cache size increase is disabled, and the remaining increment fields are ignored.*/ @@ -61,7 +57,4 @@ enum H5C_cache_decr_mode { /**<Automatic cache size decrease is enabled using the ageout with hit rate threshold algorithm.*/ }; -#ifdef __cplusplus -} -#endif #endif diff --git a/src/H5Dchunk.c b/src/H5Dchunk.c index b85b194..4445911 100644 --- a/src/H5Dchunk.c +++ b/src/H5Dchunk.c @@ -59,6 +59,7 @@ #include "H5Iprivate.h" /* IDs */ #include "H5MMprivate.h" /* Memory management */ #include "H5MFprivate.h" /* File memory management */ +#include "H5PBprivate.h" /* Page Buffer */ #include "H5VMprivate.h" /* Vector and array functions */ /****************/ @@ -70,6 +71,7 @@ #define H5D_CHUNK_GET_NODE_INFO(map, node) \ (map->use_single ? map->single_chunk_info : (H5D_chunk_info_t *)H5SL_item(node)) #define H5D_CHUNK_GET_NEXT_NODE(map, node) (map->use_single ? (H5SL_node_t *)NULL : H5SL_next(node)) +#define H5D_CHUNK_GET_NODE_COUNT(map) (map->use_single ? (size_t)1 : H5SL_count(map->sel_chunks)) /* Sanity check on chunk index types: commonly used by a lot of routines in this file */ #define H5D_CHUNK_STORAGE_INDEX_CHK(storage) \ @@ -239,10 +241,14 @@ typedef struct H5D_chunk_file_iter_ud_t { #ifdef H5_HAVE_PARALLEL /* information to construct a collective I/O operation for filling chunks */ -typedef struct H5D_chunk_coll_info_t { - size_t num_io; /* Number of write operations */ - haddr_t *addr; /* array of the file addresses of the write operation */ -} H5D_chunk_coll_info_t; +typedef struct H5D_chunk_coll_fill_info_t { + size_t num_chunks; /* Number of chunks in the write operation */ + struct chunk_coll_fill_info { + haddr_t addr; /* File address of the chunk */ + size_t chunk_size; /* Size of the chunk in the file */ + hbool_t unfiltered_partial_chunk; + } * chunk_info; +} H5D_chunk_coll_fill_info_t; #endif /* H5_HAVE_PARALLEL */ typedef struct H5D_chunk_iter_ud_t { @@ -257,8 +263,8 @@ typedef struct H5D_chunk_iter_ud_t { /* Chunked layout operation callbacks */ static herr_t H5D__chunk_construct(H5F_t *f, H5D_t *dset); static herr_t H5D__chunk_init(H5F_t *f, const H5D_t *dset, hid_t dapl_id); -static herr_t H5D__chunk_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *fm); +static herr_t H5D__chunk_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, + H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *fm); static herr_t H5D__chunk_io_init_selections(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm); static herr_t H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, @@ -287,9 +293,6 @@ static int H5D__chunk_format_convert_cb(const H5D_chunk_rec_t *chunk_rec, void * /* Helper routines */ static herr_t H5D__chunk_set_info_real(H5O_layout_chunk_t *layout, unsigned ndims, const hsize_t *curr_dims, const hsize_t *max_dims); -static void * H5D__chunk_mem_alloc(size_t size, const H5O_pline_t *pline); -static void * H5D__chunk_mem_xfree(void *chk, const void *pline); -static void * H5D__chunk_mem_realloc(void *chk, size_t size, const H5O_pline_t *pline); static herr_t H5D__chunk_cinfo_cache_reset(H5D_chunk_cached_t *last); static herr_t H5D__chunk_cinfo_cache_update(H5D_chunk_cached_t *last, const H5D_chunk_ud_t *udata); static hbool_t H5D__chunk_cinfo_cache_found(const H5D_chunk_cached_t *last, H5D_chunk_ud_t *udata); @@ -303,11 +306,10 @@ static herr_t H5D__chunk_file_cb(void *elem, const H5T_t *type, unsigned ndims void *fm); static herr_t H5D__chunk_mem_cb(void *elem, const H5T_t *type, unsigned ndims, const hsize_t *coords, void *fm); +static htri_t H5D__chunk_may_use_select_io(const H5D_io_info_t *io_info); static unsigned H5D__chunk_hash_val(const H5D_shared_t *shared, const hsize_t *scaled); static herr_t H5D__chunk_flush_entry(const H5D_t *dset, H5D_rdcc_ent_t *ent, hbool_t reset); static herr_t H5D__chunk_cache_evict(const H5D_t *dset, H5D_rdcc_ent_t *ent, hbool_t flush); -static hbool_t H5D__chunk_is_partial_edge_chunk(unsigned dset_ndims, const uint32_t *chunk_dims, - const hsize_t *chunk_scaled, const hsize_t *dset_dims); static void * H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, hbool_t relax, hbool_t prev_unfilt_chunk); static herr_t H5D__chunk_unlock(const H5D_io_info_t *io_info, const H5D_chunk_ud_t *udata, hbool_t dirty, @@ -315,9 +317,9 @@ static herr_t H5D__chunk_unlock(const H5D_io_info_t *io_info, const H5D_chunk_ static herr_t H5D__chunk_cache_prune(const H5D_t *dset, size_t size); static herr_t H5D__chunk_prune_fill(H5D_chunk_it_ud1_t *udata, hbool_t new_unfilt_chunk); #ifdef H5_HAVE_PARALLEL -static herr_t H5D__chunk_collective_fill(const H5D_t *dset, H5D_chunk_coll_info_t *chunk_info, - size_t chunk_size, const void *fill_buf); -static int H5D__chunk_cmp_addr(const void *addr1, const void *addr2); +static herr_t H5D__chunk_collective_fill(const H5D_t *dset, H5D_chunk_coll_fill_info_t *chunk_fill_info, + const void *fill_buf, const void *partial_chunk_fill_buf); +static int H5D__chunk_cmp_coll_fill_info(const void *_entry1, const void *_entry2); #endif /* H5_HAVE_PARALLEL */ /* Debugging helper routine callback */ @@ -1066,16 +1068,17 @@ H5D__chunk_is_data_cached(const H5D_shared_t *shared_dset) *------------------------------------------------------------------------- */ static herr_t -H5D__chunk_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, +H5D__chunk_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *fm) { const H5D_t *dataset = io_info->dset; /* Local pointer to dataset info */ hssize_t old_offset[H5O_LAYOUT_NDIMS]; /* Old selection offset */ htri_t file_space_normalized = FALSE; /* File dataspace was normalized */ unsigned f_ndims; /* The number of dimensions of the file's dataspace */ - int sm_ndims; /* The number of dimensions of the memory buffer's dataspace (signed) */ - unsigned u; /* Local index variable */ - herr_t ret_value = SUCCEED; /* Return value */ + int sm_ndims; /* The number of dimensions of the memory buffer's dataspace (signed) */ + htri_t use_selection_io = FALSE; /* Whether to use selection I/O */ + unsigned u; /* Local index variable */ + herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_STATIC @@ -1129,6 +1132,11 @@ H5D__chunk_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_inf if (H5D__chunk_io_init_selections(io_info, type_info, fm) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to create file and memory chunk selections") + /* Check if we're performing selection I/O and save the result */ + if ((use_selection_io = H5D__chunk_may_use_select_io(io_info)) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible") + io_info->use_select_io = (hbool_t)use_selection_io; + done: /* Reset the global dataspace info */ fm->file_space = NULL; @@ -1362,7 +1370,7 @@ done: * *------------------------------------------------------------------------- */ -static void * +void * H5D__chunk_mem_alloc(size_t size, const H5O_pline_t *pline) { void *ret_value = NULL; /* Return value */ @@ -1393,7 +1401,7 @@ H5D__chunk_mem_alloc(size_t size, const H5O_pline_t *pline) * *------------------------------------------------------------------------- */ -static void * +void * H5D__chunk_mem_xfree(void *chk, const void *_pline) { const H5O_pline_t *pline = (const H5O_pline_t *)_pline; @@ -1417,7 +1425,7 @@ H5D__chunk_mem_xfree(void *chk, const void *_pline) * calls H5D__chunk_mem_xfree and discards the return value. *------------------------------------------------------------------------- */ -static void +void H5D__chunk_mem_free(void *chk, const void *_pline) { (void)H5D__chunk_mem_xfree(chk, _pline); @@ -1437,7 +1445,7 @@ H5D__chunk_mem_free(void *chk, const void *_pline) * *------------------------------------------------------------------------- */ -static void * +void * H5D__chunk_mem_realloc(void *chk, size_t size, const H5O_pline_t *pline) { void *ret_value = NULL; /* Return value */ @@ -2460,6 +2468,78 @@ done: } /* end H5D__chunk_cacheable() */ /*------------------------------------------------------------------------- + * Function: H5D__chunk_may_use_select_io + * + * Purpose: A small internal function to if it may be possible to use + * selection I/O. + * + * Return: TRUE or FALSE + * + * Programmer: Neil Fortner + * 4 May 2021 + * + *------------------------------------------------------------------------- + */ +static htri_t +H5D__chunk_may_use_select_io(const H5D_io_info_t *io_info) +{ + const H5D_t *dataset = NULL; /* Local pointer to dataset info */ + htri_t ret_value = FAIL; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity check */ + HDassert(io_info); + + dataset = io_info->dset; + HDassert(dataset); + + /* Don't use selection I/O if it's globally disabled, there is a type + * conversion, or if there are filters on the dataset (for now) */ + if (!H5_use_selection_io_g || io_info->io_ops.single_read != H5D__select_read || + dataset->shared->dcpl_cache.pline.nused > 0) + ret_value = FALSE; + else { + hbool_t page_buf_enabled; + + HDassert(io_info->io_ops.single_write == H5D__select_write); + + /* Check if the page buffer is enabled */ + if (H5PB_enabled(io_info->f_sh, H5FD_MEM_DRAW, &page_buf_enabled) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if page buffer is enabled") + if (page_buf_enabled) + ret_value = FALSE; + else { + /* Check if chunks in this dataset may be cached, if so don't use + * selection I/O (for now). Note that chunks temporarily cached for + * the purpose of writing the fill value don't count, since they are + * immediately evicted. */ +#ifdef H5_HAVE_PARALLEL + /* If MPI based VFD is used and the file is opened for write access, + * must bypass the chunk-cache scheme because other MPI processes + * could be writing to other elements in the same chunk. + */ + if (io_info->using_mpi_vfd && (H5F_ACC_RDWR & H5F_INTENT(dataset->oloc.file))) + ret_value = TRUE; + else { +#endif /* H5_HAVE_PARALLEL */ + /* Check if the chunk is too large to keep in the cache */ + H5_CHECK_OVERFLOW(dataset->shared->layout.u.chunk.size, uint32_t, size_t); + if ((size_t)dataset->shared->layout.u.chunk.size > dataset->shared->cache.chunk.nbytes_max) + ret_value = TRUE; + else + ret_value = FALSE; +#ifdef H5_HAVE_PARALLEL + } /* end else */ +#endif /* H5_HAVE_PARALLEL */ + } /* end else */ + } /* end else */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__chunk_may_use_select_io() */ + +/*------------------------------------------------------------------------- * Function: H5D__chunk_read * * Purpose: Read from a chunked dataset. @@ -2475,16 +2555,17 @@ static herr_t H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t H5_ATTR_UNUSED nelmts, H5S_t H5_ATTR_UNUSED *file_space, H5S_t H5_ATTR_UNUSED *mem_space, H5D_chunk_map_t *fm) { - H5SL_node_t * chunk_node; /* Current node in chunk skip list */ - H5D_io_info_t nonexistent_io_info; /* "nonexistent" I/O info object */ - H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */ - H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ - H5D_io_info_t cpt_io_info; /* Compact I/O info object */ - H5D_storage_t cpt_store; /* Chunk storage information as compact dataset */ - hbool_t cpt_dirty; /* Temporary placeholder for compact storage "dirty" flag */ - uint32_t src_accessed_bytes = 0; /* Total accessed size in a chunk */ - hbool_t skip_missing_chunks = FALSE; /* Whether to skip missing chunks */ - herr_t ret_value = SUCCEED; /*return value */ + H5SL_node_t * chunk_node; /* Current node in chunk skip list */ + H5D_io_info_t nonexistent_io_info; /* "nonexistent" I/O info object */ + uint32_t src_accessed_bytes = 0; /* Total accessed size in a chunk */ + hbool_t skip_missing_chunks = FALSE; /* Whether to skip missing chunks */ + H5S_t ** chunk_mem_spaces = NULL; /* Array of chunk memory spaces */ + H5S_t * chunk_mem_spaces_static[8]; /* Static buffer for chunk_mem_spaces */ + H5S_t ** chunk_file_spaces = NULL; /* Array of chunk file spaces */ + H5S_t * chunk_file_spaces_static[8]; /* Static buffer for chunk_file_spaces */ + haddr_t * chunk_addrs = NULL; /* Array of chunk addresses */ + haddr_t chunk_addrs_static[8]; /* Static buffer for chunk_addrs */ + herr_t ret_value = SUCCEED; /*return value */ FUNC_ENTER_STATIC @@ -2498,23 +2579,6 @@ H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_ H5MM_memcpy(&nonexistent_io_info, io_info, sizeof(nonexistent_io_info)); nonexistent_io_info.layout_ops = *H5D_LOPS_NONEXISTENT; - /* Set up contiguous I/O info object */ - H5MM_memcpy(&ctg_io_info, io_info, sizeof(ctg_io_info)); - ctg_io_info.store = &ctg_store; - ctg_io_info.layout_ops = *H5D_LOPS_CONTIG; - - /* Initialize temporary contiguous storage info */ - H5_CHECKED_ASSIGN(ctg_store.contig.dset_size, hsize_t, io_info->dset->shared->layout.u.chunk.size, - uint32_t); - - /* Set up compact I/O info object */ - H5MM_memcpy(&cpt_io_info, io_info, sizeof(cpt_io_info)); - cpt_io_info.store = &cpt_store; - cpt_io_info.layout_ops = *H5D_LOPS_COMPACT; - - /* Initialize temporary compact storage info */ - cpt_store.compact.dirty = &cpt_dirty; - { const H5O_fill_t *fill = &(io_info->dset->shared->dcpl_cache.fill); /* Fill value info */ H5D_fill_value_t fill_status; /* Fill value status */ @@ -2532,80 +2596,215 @@ H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_ skip_missing_chunks = TRUE; } - /* Iterate through nodes in chunk skip list */ - chunk_node = H5D_CHUNK_GET_FIRST_NODE(fm); - while (chunk_node) { - H5D_chunk_info_t *chunk_info; /* Chunk information */ - H5D_chunk_ud_t udata; /* Chunk index pass-through */ + /* Different blocks depending on whether we're using selection I/O */ + if (io_info->use_select_io) { + size_t num_chunks; + size_t element_sizes[2] = {type_info->dst_type_size, 0}; + void * bufs[2] = {io_info->u.rbuf, NULL}; + + /* Cache number of chunks */ + num_chunks = H5D_CHUNK_GET_NODE_COUNT(fm); + + /* Allocate arrays of dataspaces and offsets for use with selection I/O, + * or point to static buffers */ + HDassert(sizeof(chunk_mem_spaces_static) / sizeof(chunk_mem_spaces_static[0]) == + sizeof(chunk_file_spaces_static) / sizeof(chunk_file_spaces_static[0])); + HDassert(sizeof(chunk_mem_spaces_static) / sizeof(chunk_mem_spaces_static[0]) == + sizeof(chunk_addrs_static) / sizeof(chunk_addrs_static[0])); + if (num_chunks > (sizeof(chunk_mem_spaces_static) / sizeof(chunk_mem_spaces_static[0]))) { + if (NULL == (chunk_mem_spaces = H5MM_malloc(num_chunks * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for memory space list") + if (NULL == (chunk_file_spaces = H5MM_malloc(num_chunks * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for file space list") + if (NULL == (chunk_addrs = H5MM_malloc(num_chunks * sizeof(haddr_t)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for chunk address list") + } /* end if */ + else { + chunk_mem_spaces = chunk_mem_spaces_static; + chunk_file_spaces = chunk_file_spaces_static; + chunk_addrs = chunk_addrs_static; + } /* end else */ - /* Get the actual chunk information from the skip list node */ - chunk_info = H5D_CHUNK_GET_NODE_INFO(fm, chunk_node); + /* Reset num_chunks */ + num_chunks = 0; - /* Get the info for the chunk in the file */ - if (H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") + /* Iterate through nodes in chunk skip list */ + chunk_node = H5D_CHUNK_GET_FIRST_NODE(fm); + while (chunk_node) { + H5D_chunk_info_t *chunk_info; /* Chunk information */ + H5D_chunk_ud_t udata; /* Chunk index pass-through */ - /* Sanity check */ - HDassert((H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length > 0) || - (!H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length == 0)); + /* Get the actual chunk information from the skip list node */ + chunk_info = H5D_CHUNK_GET_NODE_INFO(fm, chunk_node); - /* Check for non-existent chunk & skip it if appropriate */ - if (H5F_addr_defined(udata.chunk_block.offset) || UINT_MAX != udata.idx_hint || - !skip_missing_chunks) { - H5D_io_info_t *chk_io_info; /* Pointer to I/O info object for this chunk */ - void * chunk = NULL; /* Pointer to locked chunk buffer */ - htri_t cacheable; /* Whether the chunk is cacheable */ + /* Get the info for the chunk in the file */ + if (H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") - /* Set chunk's [scaled] coordinates */ - io_info->store->chunk.scaled = chunk_info->scaled; + /* There should be no chunks cached */ + HDassert(UINT_MAX == udata.idx_hint); - /* Determine if we should use the chunk cache */ - if ((cacheable = H5D__chunk_cacheable(io_info, udata.chunk_block.offset, FALSE)) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't tell if chunk is cacheable") - if (cacheable) { - /* Load the chunk into cache and lock it. */ + /* Sanity check */ + HDassert((H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length > 0) || + (!H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length == 0)); + + /* Check for non-existent chunk & skip it if appropriate */ + if (H5F_addr_defined(udata.chunk_block.offset)) { + /* Add chunk to list for selection I/O */ + chunk_mem_spaces[num_chunks] = chunk_info->mspace; + chunk_file_spaces[num_chunks] = chunk_info->fspace; + chunk_addrs[num_chunks] = udata.chunk_block.offset; + num_chunks++; + } /* end if */ + else if (!skip_missing_chunks) { + /* Perform the actual read operation from the nonexistent chunk + */ + if ((io_info->io_ops.single_read)(&nonexistent_io_info, type_info, + (hsize_t)chunk_info->chunk_points, chunk_info->fspace, + chunk_info->mspace) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked read failed") + } /* end if */ - /* Compute # of bytes accessed in chunk */ - H5_CHECK_OVERFLOW(type_info->src_type_size, /*From:*/ size_t, /*To:*/ uint32_t); - src_accessed_bytes = chunk_info->chunk_points * (uint32_t)type_info->src_type_size; + /* Advance to next chunk in list */ + chunk_node = H5D_CHUNK_GET_NEXT_NODE(fm, chunk_node); + } /* end while */ - /* Lock the chunk into the cache */ - if (NULL == (chunk = H5D__chunk_lock(io_info, &udata, FALSE, FALSE))) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") + /* Issue selection I/O call (we can skip the page buffer because we've + * already verified it won't be used, and the metadata accumulator + * because this is raw data) */ + if (H5F_shared_select_read(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, (uint32_t)num_chunks, + chunk_mem_spaces, chunk_file_spaces, chunk_addrs, element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunk selection read failed") + + /* Clean up memory */ + if (chunk_mem_spaces != chunk_mem_spaces_static) { + HDassert(chunk_mem_spaces); + HDassert(chunk_file_spaces != chunk_file_spaces_static); + HDassert(chunk_addrs != chunk_addrs_static); + H5MM_free(chunk_mem_spaces); + chunk_mem_spaces = NULL; + H5MM_free(chunk_file_spaces); + chunk_file_spaces = NULL; + H5MM_free(chunk_addrs); + chunk_addrs = NULL; + } /* end if */ + } /* end if */ + else { + H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */ + H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ + H5D_io_info_t cpt_io_info; /* Compact I/O info object */ + H5D_storage_t cpt_store; /* Chunk storage information as compact dataset */ + hbool_t cpt_dirty; /* Temporary placeholder for compact storage "dirty" flag */ + + /* Set up contiguous I/O info object */ + H5MM_memcpy(&ctg_io_info, io_info, sizeof(ctg_io_info)); + ctg_io_info.store = &ctg_store; + ctg_io_info.layout_ops = *H5D_LOPS_CONTIG; + + /* Initialize temporary contiguous storage info */ + H5_CHECKED_ASSIGN(ctg_store.contig.dset_size, hsize_t, io_info->dset->shared->layout.u.chunk.size, + uint32_t); + + /* Set up compact I/O info object */ + H5MM_memcpy(&cpt_io_info, io_info, sizeof(cpt_io_info)); + cpt_io_info.store = &cpt_store; + cpt_io_info.layout_ops = *H5D_LOPS_COMPACT; + + /* Initialize temporary compact storage info */ + cpt_store.compact.dirty = &cpt_dirty; + + /* Iterate through nodes in chunk skip list */ + chunk_node = H5D_CHUNK_GET_FIRST_NODE(fm); + while (chunk_node) { + H5D_chunk_info_t *chunk_info; /* Chunk information */ + H5D_chunk_ud_t udata; /* Chunk index pass-through */ + htri_t cacheable; /* Whether the chunk is cacheable */ + + /* Get the actual chunk information from the skip list node */ + chunk_info = H5D_CHUNK_GET_NODE_INFO(fm, chunk_node); + + /* Get the info for the chunk in the file */ + if (H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") - /* Set up the storage buffer information for this chunk */ - cpt_store.compact.buf = chunk; + /* Sanity check */ + HDassert((H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length > 0) || + (!H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length == 0)); - /* Point I/O info at contiguous I/O info for this chunk */ - chk_io_info = &cpt_io_info; - } /* end if */ - else if (H5F_addr_defined(udata.chunk_block.offset)) { - /* Set up the storage address information for this chunk */ - ctg_store.contig.dset_addr = udata.chunk_block.offset; + /* Check for non-existent chunk & skip it if appropriate */ + if (H5F_addr_defined(udata.chunk_block.offset) || UINT_MAX != udata.idx_hint || + !skip_missing_chunks) { + H5D_io_info_t *chk_io_info; /* Pointer to I/O info object for this chunk */ + void * chunk = NULL; /* Pointer to locked chunk buffer */ - /* Point I/O info at temporary I/O info for this chunk */ - chk_io_info = &ctg_io_info; - } /* end else if */ - else { - /* Point I/O info at "nonexistent" I/O info for this chunk */ - chk_io_info = &nonexistent_io_info; - } /* end else */ + /* Set chunk's [scaled] coordinates */ + io_info->store->chunk.scaled = chunk_info->scaled; - /* Perform the actual read operation */ - if ((io_info->io_ops.single_read)(chk_io_info, type_info, (hsize_t)chunk_info->chunk_points, - chunk_info->fspace, chunk_info->mspace) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked read failed") + /* Determine if we should use the chunk cache */ + if ((cacheable = H5D__chunk_cacheable(io_info, udata.chunk_block.offset, FALSE)) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't tell if chunk is cacheable") + if (cacheable) { + /* Load the chunk into cache and lock it. */ - /* Release the cache lock on the chunk. */ - if (chunk && H5D__chunk_unlock(io_info, &udata, FALSE, chunk, src_accessed_bytes) < 0) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk") - } /* end if */ + /* Compute # of bytes accessed in chunk */ + H5_CHECK_OVERFLOW(type_info->src_type_size, /*From:*/ size_t, /*To:*/ uint32_t); + src_accessed_bytes = chunk_info->chunk_points * (uint32_t)type_info->src_type_size; - /* Advance to next chunk in list */ - chunk_node = H5D_CHUNK_GET_NEXT_NODE(fm, chunk_node); - } /* end while */ + /* Lock the chunk into the cache */ + if (NULL == (chunk = H5D__chunk_lock(io_info, &udata, FALSE, FALSE))) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") + + /* Set up the storage buffer information for this chunk */ + cpt_store.compact.buf = chunk; + + /* Point I/O info at contiguous I/O info for this chunk */ + chk_io_info = &cpt_io_info; + } /* end if */ + else if (H5F_addr_defined(udata.chunk_block.offset)) { + /* Set up the storage address information for this chunk */ + ctg_store.contig.dset_addr = udata.chunk_block.offset; + + /* Point I/O info at temporary I/O info for this chunk */ + chk_io_info = &ctg_io_info; + } /* end else if */ + else { + /* Point I/O info at "nonexistent" I/O info for this chunk */ + chk_io_info = &nonexistent_io_info; + } /* end else */ + + /* Perform the actual read operation */ + if ((io_info->io_ops.single_read)(chk_io_info, type_info, (hsize_t)chunk_info->chunk_points, + chunk_info->fspace, chunk_info->mspace) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked read failed") + + /* Release the cache lock on the chunk. */ + if (chunk && H5D__chunk_unlock(io_info, &udata, FALSE, chunk, src_accessed_bytes) < 0) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk") + } /* end if */ + + /* Advance to next chunk in list */ + chunk_node = H5D_CHUNK_GET_NEXT_NODE(fm, chunk_node); + } /* end while */ + } /* end else */ done: + /* Cleanup on failure */ + if (ret_value < 0) { + if (chunk_mem_spaces != chunk_mem_spaces_static) + chunk_mem_spaces = H5MM_xfree(chunk_mem_spaces); + if (chunk_file_spaces != chunk_file_spaces_static) + chunk_file_spaces = H5MM_xfree(chunk_file_spaces); + if (chunk_addrs != chunk_addrs_static) + chunk_addrs = H5MM_xfree(chunk_addrs); + } /* end if */ + + /* Make sure we cleaned up */ + HDassert(!chunk_mem_spaces || chunk_mem_spaces == chunk_mem_spaces_static); + HDassert(!chunk_file_spaces || chunk_file_spaces == chunk_file_spaces_static); + HDassert(!chunk_addrs || chunk_addrs == chunk_addrs_static); + FUNC_LEAVE_NOAPI(ret_value) } /* H5D__chunk_read() */ @@ -2625,14 +2824,20 @@ static herr_t H5D__chunk_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t H5_ATTR_UNUSED nelmts, H5S_t H5_ATTR_UNUSED *file_space, H5S_t H5_ATTR_UNUSED *mem_space, H5D_chunk_map_t *fm) { - H5SL_node_t * chunk_node; /* Current node in chunk skip list */ - H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */ - H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ - H5D_io_info_t cpt_io_info; /* Compact I/O info object */ - H5D_storage_t cpt_store; /* Chunk storage information as compact dataset */ - hbool_t cpt_dirty; /* Temporary placeholder for compact storage "dirty" flag */ - uint32_t dst_accessed_bytes = 0; /* Total accessed size in a chunk */ - herr_t ret_value = SUCCEED; /* Return value */ + H5SL_node_t * chunk_node; /* Current node in chunk skip list */ + H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */ + H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ + H5D_io_info_t cpt_io_info; /* Compact I/O info object */ + H5D_storage_t cpt_store; /* Chunk storage information as compact dataset */ + hbool_t cpt_dirty; /* Temporary placeholder for compact storage "dirty" flag */ + uint32_t dst_accessed_bytes = 0; /* Total accessed size in a chunk */ + H5S_t ** chunk_mem_spaces = NULL; /* Array of chunk memory spaces */ + H5S_t * chunk_mem_spaces_static[8]; /* Static buffer for chunk_mem_spaces */ + H5S_t ** chunk_file_spaces = NULL; /* Array of chunk file spaces */ + H5S_t * chunk_file_spaces_static[8]; /* Static buffer for chunk_file_spaces */ + haddr_t * chunk_addrs = NULL; /* Array of chunk addresses */ + haddr_t chunk_addrs_static[8]; /* Static buffer for chunk_addrs */ + herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_STATIC @@ -2659,116 +2864,295 @@ H5D__chunk_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize /* Initialize temporary compact storage info */ cpt_store.compact.dirty = &cpt_dirty; - /* Iterate through nodes in chunk skip list */ - chunk_node = H5D_CHUNK_GET_FIRST_NODE(fm); - while (chunk_node) { - H5D_chunk_info_t * chunk_info; /* Chunk information */ - H5D_chk_idx_info_t idx_info; /* Chunked index info */ - H5D_io_info_t * chk_io_info; /* Pointer to I/O info object for this chunk */ - void * chunk; /* Pointer to locked chunk buffer */ - H5D_chunk_ud_t udata; /* Index pass-through */ - htri_t cacheable; /* Whether the chunk is cacheable */ - hbool_t need_insert = FALSE; /* Whether the chunk needs to be inserted into the index */ + /* Different blocks depending on whether we're using selection I/O */ + if (io_info->use_select_io) { + size_t num_chunks; + size_t element_sizes[2] = {type_info->dst_type_size, 0}; + const void *bufs[2] = {io_info->u.wbuf, NULL}; + + /* Cache number of chunks */ + num_chunks = H5D_CHUNK_GET_NODE_COUNT(fm); + + /* Allocate arrays of dataspaces and offsets for use with selection I/O, + * or point to static buffers */ + HDassert(sizeof(chunk_mem_spaces_static) / sizeof(chunk_mem_spaces_static[0]) == + sizeof(chunk_file_spaces_static) / sizeof(chunk_file_spaces_static[0])); + HDassert(sizeof(chunk_mem_spaces_static) / sizeof(chunk_mem_spaces_static[0]) == + sizeof(chunk_addrs_static) / sizeof(chunk_addrs_static[0])); + if (num_chunks > (sizeof(chunk_mem_spaces_static) / sizeof(chunk_mem_spaces_static[0]))) { + if (NULL == (chunk_mem_spaces = H5MM_malloc(num_chunks * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for memory space list") + if (NULL == (chunk_file_spaces = H5MM_malloc(num_chunks * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for file space list") + if (NULL == (chunk_addrs = H5MM_malloc(num_chunks * sizeof(haddr_t)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for chunk address list") + } /* end if */ + else { + chunk_mem_spaces = chunk_mem_spaces_static; + chunk_file_spaces = chunk_file_spaces_static; + chunk_addrs = chunk_addrs_static; + } /* end else */ - /* Get the actual chunk information from the skip list node */ - chunk_info = H5D_CHUNK_GET_NODE_INFO(fm, chunk_node); + /* Reset num_chunks */ + num_chunks = 0; - /* Look up the chunk */ - if (H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") + /* Iterate through nodes in chunk skip list */ + chunk_node = H5D_CHUNK_GET_FIRST_NODE(fm); + while (chunk_node) { + H5D_chunk_info_t * chunk_info; /* Chunk information */ + H5D_chk_idx_info_t idx_info; /* Chunked index info */ + H5D_chunk_ud_t udata; /* Index pass-through */ + htri_t cacheable; /* Whether the chunk is cacheable */ + hbool_t need_insert = FALSE; /* Whether the chunk needs to be inserted into the index */ - /* Sanity check */ - HDassert((H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length > 0) || - (!H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length == 0)); - - /* Set chunk's [scaled] coordinates */ - io_info->store->chunk.scaled = chunk_info->scaled; - - /* Determine if we should use the chunk cache */ - if ((cacheable = H5D__chunk_cacheable(io_info, udata.chunk_block.offset, TRUE)) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't tell if chunk is cacheable") - if (cacheable) { - /* Load the chunk into cache. But if the whole chunk is written, - * simply allocate space instead of load the chunk. */ - hbool_t entire_chunk = TRUE; /* Whether whole chunk is selected */ - - /* Compute # of bytes accessed in chunk */ - H5_CHECK_OVERFLOW(type_info->dst_type_size, /*From:*/ size_t, /*To:*/ uint32_t); - dst_accessed_bytes = chunk_info->chunk_points * (uint32_t)type_info->dst_type_size; - - /* Determine if we will access all the data in the chunk */ - if (dst_accessed_bytes != ctg_store.contig.dset_size || - (chunk_info->chunk_points * type_info->src_type_size) != ctg_store.contig.dset_size || - fm->fsel_type == H5S_SEL_POINTS) - entire_chunk = FALSE; - - /* Lock the chunk into the cache */ - if (NULL == (chunk = H5D__chunk_lock(io_info, &udata, entire_chunk, FALSE))) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") - - /* Set up the storage buffer information for this chunk */ - cpt_store.compact.buf = chunk; - - /* Point I/O info at main I/O info for this chunk */ - chk_io_info = &cpt_io_info; + /* Get the actual chunk information from the skip list node */ + chunk_info = H5D_CHUNK_GET_NODE_INFO(fm, chunk_node); + + /* Get the info for the chunk in the file */ + if (H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") + + /* There should be no chunks cached */ + HDassert(UINT_MAX == udata.idx_hint); + + /* Sanity check */ + HDassert((H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length > 0) || + (!H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length == 0)); + + /* Set chunk's [scaled] coordinates */ + io_info->store->chunk.scaled = chunk_info->scaled; + + /* Determine if we should use the chunk cache */ + if ((cacheable = H5D__chunk_cacheable(io_info, udata.chunk_block.offset, TRUE)) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't tell if chunk is cacheable") + if (cacheable) { + /* Load the chunk into cache. But if the whole chunk is written, + * simply allocate space instead of load the chunk. */ + void * chunk; /* Pointer to locked chunk buffer */ + hbool_t entire_chunk = TRUE; /* Whether whole chunk is selected */ + + /* Compute # of bytes accessed in chunk */ + H5_CHECK_OVERFLOW(type_info->dst_type_size, /*From:*/ size_t, /*To:*/ uint32_t); + dst_accessed_bytes = chunk_info->chunk_points * (uint32_t)type_info->dst_type_size; + + /* Determine if we will access all the data in the chunk */ + if (dst_accessed_bytes != ctg_store.contig.dset_size || + (chunk_info->chunk_points * type_info->src_type_size) != ctg_store.contig.dset_size || + fm->fsel_type == H5S_SEL_POINTS) + entire_chunk = FALSE; + + /* Lock the chunk into the cache */ + if (NULL == (chunk = H5D__chunk_lock(io_info, &udata, entire_chunk, FALSE))) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") + + /* Set up the storage buffer information for this chunk */ + cpt_store.compact.buf = chunk; + + /* Perform the actual write operation */ + if ((io_info->io_ops.single_write)(&cpt_io_info, type_info, (hsize_t)chunk_info->chunk_points, + chunk_info->fspace, chunk_info->mspace) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked write failed") + + /* Release the cache lock on the chunk */ + if (H5D__chunk_unlock(io_info, &udata, TRUE, chunk, dst_accessed_bytes) < 0) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk") + } /* end if */ + else { + /* If the chunk hasn't been allocated on disk, do so now. */ + if (!H5F_addr_defined(udata.chunk_block.offset)) { + /* Compose chunked index info struct */ + idx_info.f = io_info->dset->oloc.file; + idx_info.pline = &(io_info->dset->shared->dcpl_cache.pline); + idx_info.layout = &(io_info->dset->shared->layout.u.chunk); + idx_info.storage = &(io_info->dset->shared->layout.storage.u.chunk); + + /* Set up the size of chunk for user data */ + udata.chunk_block.length = io_info->dset->shared->layout.u.chunk.size; + + /* Allocate the chunk */ + if (H5D__chunk_file_alloc(&idx_info, NULL, &udata.chunk_block, &need_insert, + chunk_info->scaled) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, + "unable to insert/resize chunk on chunk level") + + /* Make sure the address of the chunk is returned. */ + if (!H5F_addr_defined(udata.chunk_block.offset)) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "chunk address isn't defined") + + /* Cache the new chunk information */ + H5D__chunk_cinfo_cache_update(&io_info->dset->shared->cache.chunk.last, &udata); + + /* Insert chunk into index */ + if (need_insert && io_info->dset->shared->layout.storage.u.chunk.ops->insert) + if ((io_info->dset->shared->layout.storage.u.chunk.ops->insert)(&idx_info, &udata, + NULL) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, + "unable to insert chunk addr into index") + } /* end if */ + + /* Add chunk to list for selection I/O */ + chunk_mem_spaces[num_chunks] = chunk_info->mspace; + chunk_file_spaces[num_chunks] = chunk_info->fspace; + chunk_addrs[num_chunks] = udata.chunk_block.offset; + num_chunks++; + } /* end else */ + + /* Advance to next chunk in list */ + chunk_node = H5D_CHUNK_GET_NEXT_NODE(fm, chunk_node); + } /* end while */ + + /* Issue selection I/O call (we can skip the page buffer because we've + * already verified it won't be used, and the metadata accumulator + * because this is raw data) */ + if (H5F_shared_select_write(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, (uint32_t)num_chunks, + chunk_mem_spaces, chunk_file_spaces, chunk_addrs, element_sizes, + bufs) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunk selection read failed") + + /* Clean up memory */ + if (chunk_mem_spaces != chunk_mem_spaces_static) { + HDassert(chunk_mem_spaces); + HDassert(chunk_file_spaces != chunk_file_spaces_static); + HDassert(chunk_addrs != chunk_addrs_static); + H5MM_free(chunk_mem_spaces); + chunk_mem_spaces = NULL; + H5MM_free(chunk_file_spaces); + chunk_file_spaces = NULL; + H5MM_free(chunk_addrs); + chunk_addrs = NULL; } /* end if */ - else { - /* If the chunk hasn't been allocated on disk, do so now. */ - if (!H5F_addr_defined(udata.chunk_block.offset)) { - /* Compose chunked index info struct */ - idx_info.f = io_info->dset->oloc.file; - idx_info.pline = &(io_info->dset->shared->dcpl_cache.pline); - idx_info.layout = &(io_info->dset->shared->layout.u.chunk); - idx_info.storage = &(io_info->dset->shared->layout.storage.u.chunk); - - /* Set up the size of chunk for user data */ - udata.chunk_block.length = io_info->dset->shared->layout.u.chunk.size; - - /* Allocate the chunk */ - if (H5D__chunk_file_alloc(&idx_info, NULL, &udata.chunk_block, &need_insert, - chunk_info->scaled) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, - "unable to insert/resize chunk on chunk level") - - /* Make sure the address of the chunk is returned. */ - if (!H5F_addr_defined(udata.chunk_block.offset)) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "chunk address isn't defined") - - /* Cache the new chunk information */ - H5D__chunk_cinfo_cache_update(&io_info->dset->shared->cache.chunk.last, &udata); + } /* end if */ + else { + /* Iterate through nodes in chunk skip list */ + chunk_node = H5D_CHUNK_GET_FIRST_NODE(fm); + while (chunk_node) { + H5D_chunk_info_t * chunk_info; /* Chunk information */ + H5D_chk_idx_info_t idx_info; /* Chunked index info */ + H5D_io_info_t * chk_io_info; /* Pointer to I/O info object for this chunk */ + void * chunk; /* Pointer to locked chunk buffer */ + H5D_chunk_ud_t udata; /* Index pass-through */ + htri_t cacheable; /* Whether the chunk is cacheable */ + hbool_t need_insert = FALSE; /* Whether the chunk needs to be inserted into the index */ + + /* Get the actual chunk information from the skip list node */ + chunk_info = H5D_CHUNK_GET_NODE_INFO(fm, chunk_node); + + /* Look up the chunk */ + if (H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") + + /* Sanity check */ + HDassert((H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length > 0) || + (!H5F_addr_defined(udata.chunk_block.offset) && udata.chunk_block.length == 0)); + + /* Set chunk's [scaled] coordinates */ + io_info->store->chunk.scaled = chunk_info->scaled; + + /* Determine if we should use the chunk cache */ + if ((cacheable = H5D__chunk_cacheable(io_info, udata.chunk_block.offset, TRUE)) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't tell if chunk is cacheable") + if (cacheable) { + /* Load the chunk into cache. But if the whole chunk is written, + * simply allocate space instead of load the chunk. */ + hbool_t entire_chunk = TRUE; /* Whether whole chunk is selected */ + + /* Compute # of bytes accessed in chunk */ + H5_CHECK_OVERFLOW(type_info->dst_type_size, /*From:*/ size_t, /*To:*/ uint32_t); + dst_accessed_bytes = chunk_info->chunk_points * (uint32_t)type_info->dst_type_size; + + /* Determine if we will access all the data in the chunk */ + if (dst_accessed_bytes != ctg_store.contig.dset_size || + (chunk_info->chunk_points * type_info->src_type_size) != ctg_store.contig.dset_size || + fm->fsel_type == H5S_SEL_POINTS) + entire_chunk = FALSE; + + /* Lock the chunk into the cache */ + if (NULL == (chunk = H5D__chunk_lock(io_info, &udata, entire_chunk, FALSE))) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") + + /* Set up the storage buffer information for this chunk */ + cpt_store.compact.buf = chunk; + + /* Point I/O info at main I/O info for this chunk */ + chk_io_info = &cpt_io_info; } /* end if */ + else { + /* If the chunk hasn't been allocated on disk, do so now. */ + if (!H5F_addr_defined(udata.chunk_block.offset)) { + /* Compose chunked index info struct */ + idx_info.f = io_info->dset->oloc.file; + idx_info.pline = &(io_info->dset->shared->dcpl_cache.pline); + idx_info.layout = &(io_info->dset->shared->layout.u.chunk); + idx_info.storage = &(io_info->dset->shared->layout.storage.u.chunk); + + /* Set up the size of chunk for user data */ + udata.chunk_block.length = io_info->dset->shared->layout.u.chunk.size; + + /* Allocate the chunk */ + if (H5D__chunk_file_alloc(&idx_info, NULL, &udata.chunk_block, &need_insert, + chunk_info->scaled) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, + "unable to insert/resize chunk on chunk level") + + /* Make sure the address of the chunk is returned. */ + if (!H5F_addr_defined(udata.chunk_block.offset)) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "chunk address isn't defined") + + /* Cache the new chunk information */ + H5D__chunk_cinfo_cache_update(&io_info->dset->shared->cache.chunk.last, &udata); + } /* end if */ - /* Set up the storage address information for this chunk */ - ctg_store.contig.dset_addr = udata.chunk_block.offset; + /* Set up the storage address information for this chunk */ + ctg_store.contig.dset_addr = udata.chunk_block.offset; - /* No chunk cached */ - chunk = NULL; + /* No chunk cached */ + chunk = NULL; - /* Point I/O info at temporary I/O info for this chunk */ - chk_io_info = &ctg_io_info; - } /* end else */ + /* Point I/O info at temporary I/O info for this chunk */ + chk_io_info = &ctg_io_info; + } /* end else */ - /* Perform the actual write operation */ - if ((io_info->io_ops.single_write)(chk_io_info, type_info, (hsize_t)chunk_info->chunk_points, - chunk_info->fspace, chunk_info->mspace) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked write failed") + /* Perform the actual write operation */ + if ((io_info->io_ops.single_write)(chk_io_info, type_info, (hsize_t)chunk_info->chunk_points, + chunk_info->fspace, chunk_info->mspace) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked write failed") - /* Release the cache lock on the chunk, or insert chunk into index. */ - if (chunk) { - if (H5D__chunk_unlock(io_info, &udata, TRUE, chunk, dst_accessed_bytes) < 0) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk") - } /* end if */ - else { - if (need_insert && io_info->dset->shared->layout.storage.u.chunk.ops->insert) - if ((io_info->dset->shared->layout.storage.u.chunk.ops->insert)(&idx_info, &udata, NULL) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert chunk addr into index") - } /* end else */ + /* Release the cache lock on the chunk, or insert chunk into index. */ + if (chunk) { + if (H5D__chunk_unlock(io_info, &udata, TRUE, chunk, dst_accessed_bytes) < 0) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to unlock raw data chunk") + } /* end if */ + else { + if (need_insert && io_info->dset->shared->layout.storage.u.chunk.ops->insert) + if ((io_info->dset->shared->layout.storage.u.chunk.ops->insert)(&idx_info, &udata, NULL) < + 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, + "unable to insert chunk addr into index") + } /* end else */ - /* Advance to next chunk in list */ - chunk_node = H5D_CHUNK_GET_NEXT_NODE(fm, chunk_node); - } /* end while */ + /* Advance to next chunk in list */ + chunk_node = H5D_CHUNK_GET_NEXT_NODE(fm, chunk_node); + } /* end while */ + } /* end else */ done: + /* Cleanup on failure */ + if (ret_value < 0) { + if (chunk_mem_spaces != chunk_mem_spaces_static) + chunk_mem_spaces = H5MM_xfree(chunk_mem_spaces); + if (chunk_file_spaces != chunk_file_spaces_static) + chunk_file_spaces = H5MM_xfree(chunk_file_spaces); + if (chunk_addrs != chunk_addrs_static) + chunk_addrs = H5MM_xfree(chunk_addrs); + } /* end if */ + + /* Make sure we cleaned up */ + HDassert(!chunk_mem_spaces || chunk_mem_spaces == chunk_mem_spaces_static); + HDassert(!chunk_file_spaces || chunk_file_spaces == chunk_file_spaces_static); + HDassert(!chunk_addrs || chunk_addrs == chunk_addrs_static); + FUNC_LEAVE_NOAPI(ret_value) } /* H5D__chunk_write() */ @@ -3178,7 +3562,9 @@ H5D__chunk_lookup(const H5D_t *dset, const hsize_t *scaled, H5D_chunk_ud_t *udat unsigned idx = 0; /* Index of chunk in cache, if present */ hbool_t found = FALSE; /* In cache? */ #ifdef H5_HAVE_PARALLEL - hbool_t reenable_coll_md_reads = FALSE; + H5P_coll_md_read_flag_t md_reads_file_flag; + hbool_t md_reads_context_flag; + hbool_t restore_md_reads_state = FALSE; #endif herr_t ret_value = SUCCEED; /* Return value */ @@ -3252,11 +3638,10 @@ H5D__chunk_lookup(const H5D_t *dset, const hsize_t *scaled, H5D_chunk_ud_t *udat * processes. */ if (H5F_HAS_FEATURE(idx_info.f, H5FD_FEAT_HAS_MPI)) { - hbool_t do_coll_md_reads = H5CX_get_coll_metadata_read(); - if (do_coll_md_reads) { - H5CX_set_coll_metadata_read(FALSE); - reenable_coll_md_reads = TRUE; - } + md_reads_file_flag = H5P_FORCE_FALSE; + md_reads_context_flag = FALSE; + H5F_set_coll_metadata_reads(idx_info.f, &md_reads_file_flag, &md_reads_context_flag); + restore_md_reads_state = TRUE; } #endif /* H5_HAVE_PARALLEL */ @@ -3302,8 +3687,8 @@ H5D__chunk_lookup(const H5D_t *dset, const hsize_t *scaled, H5D_chunk_ud_t *udat done: #ifdef H5_HAVE_PARALLEL /* Re-enable collective metadata reads if we disabled them */ - if (reenable_coll_md_reads) - H5CX_set_coll_metadata_read(TRUE); + if (restore_md_reads_state) + H5F_set_coll_metadata_reads(dset->oloc.file, &md_reads_file_flag, &md_reads_context_flag); #endif /* H5_HAVE_PARALLEL */ FUNC_LEAVE_NOAPI(ret_value) @@ -4319,8 +4704,8 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, const hbool_t blocks_written = FALSE; /* Flag to indicate that chunk was actually written */ hbool_t using_mpi = FALSE; /* Flag to indicate that the file is being accessed with an MPI-capable file driver */ - H5D_chunk_coll_info_t chunk_info; /* chunk address information for doing I/O */ -#endif /* H5_HAVE_PARALLEL */ + H5D_chunk_coll_fill_info_t chunk_fill_info; /* chunk address information for doing I/O */ +#endif /* H5_HAVE_PARALLEL */ hbool_t carry; /* Flag to indicate that chunk increment carrys to higher dimension (sorta) */ unsigned space_ndims; /* Dataset's space rank */ const hsize_t * space_dim; /* Dataset's dataspace dimensions */ @@ -4367,8 +4752,8 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, const using_mpi = TRUE; /* init chunk info stuff for collective I/O */ - chunk_info.num_io = 0; - chunk_info.addr = NULL; + chunk_fill_info.num_chunks = 0; + chunk_fill_info.chunk_info = NULL; } /* end if */ #endif /* H5_HAVE_PARALLEL */ @@ -4640,19 +5025,26 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, const if (using_mpi) { /* collect all chunk addresses to be written to write collectively at the end */ - /* allocate/resize address array if no more space left */ - /* Note that if we add support for parallel filters we must - * also store an array of chunk sizes and pass it to the - * apporpriate collective write function */ - if (0 == chunk_info.num_io % 1024) - if (NULL == (chunk_info.addr = (haddr_t *)H5MM_realloc( - chunk_info.addr, (chunk_info.num_io + 1024) * sizeof(haddr_t)))) + + /* allocate/resize chunk info array if no more space left */ + if (0 == chunk_fill_info.num_chunks % 1024) { + void *tmp_realloc; + + if (NULL == (tmp_realloc = H5MM_realloc(chunk_fill_info.chunk_info, + (chunk_fill_info.num_chunks + 1024) * + sizeof(struct chunk_coll_fill_info)))) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, - "memory allocation failed for chunk addresses") + "memory allocation failed for chunk fill info") + + chunk_fill_info.chunk_info = tmp_realloc; + } - /* Store the chunk's address for later */ - chunk_info.addr[chunk_info.num_io] = udata.chunk_block.offset; - chunk_info.num_io++; + /* Store info about the chunk for later */ + chunk_fill_info.chunk_info[chunk_fill_info.num_chunks].addr = udata.chunk_block.offset; + chunk_fill_info.chunk_info[chunk_fill_info.num_chunks].chunk_size = chunk_size; + chunk_fill_info.chunk_info[chunk_fill_info.num_chunks].unfiltered_partial_chunk = + (*fill_buf == unfilt_fill_buf); + chunk_fill_info.num_chunks++; /* Indicate that blocks will be written */ blocks_written = TRUE; @@ -4725,7 +5117,7 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, const #ifdef H5_HAVE_PARALLEL /* do final collective I/O */ if (using_mpi && blocks_written) - if (H5D__chunk_collective_fill(dset, &chunk_info, chunk_size, fb_info.fill_buf) < 0) + if (H5D__chunk_collective_fill(dset, &chunk_fill_info, fb_info.fill_buf, unfilt_fill_buf) < 0) HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "unable to write raw data to file") #endif /* H5_HAVE_PARALLEL */ @@ -4741,8 +5133,8 @@ done: unfilt_fill_buf = H5D__chunk_mem_xfree(unfilt_fill_buf, &def_pline); #ifdef H5_HAVE_PARALLEL - if (using_mpi && chunk_info.addr) - H5MM_free(chunk_info.addr); + if (using_mpi && chunk_fill_info.chunk_info) + H5MM_free(chunk_fill_info.chunk_info); #endif FUNC_LEAVE_NOAPI(ret_value) @@ -4936,27 +5328,35 @@ done: *------------------------------------------------------------------------- */ static herr_t -H5D__chunk_collective_fill(const H5D_t *dset, H5D_chunk_coll_info_t *chunk_info, size_t chunk_size, - const void *fill_buf) +H5D__chunk_collective_fill(const H5D_t *dset, H5D_chunk_coll_fill_info_t *chunk_fill_info, + const void *fill_buf, const void *partial_chunk_fill_buf) { - MPI_Comm mpi_comm = MPI_COMM_NULL; /* MPI communicator for file */ - int mpi_rank = (-1); /* This process's rank */ - int mpi_size = (-1); /* MPI Comm size */ - int mpi_code; /* MPI return code */ - size_t num_blocks; /* Number of blocks between processes. */ - size_t leftover_blocks; /* Number of leftover blocks to handle */ - int blocks, leftover, block_len; /* converted to int for MPI */ + MPI_Comm mpi_comm = MPI_COMM_NULL; /* MPI communicator for file */ + int mpi_rank = (-1); /* This process's rank */ + int mpi_size = (-1); /* MPI Comm size */ + int mpi_code; /* MPI return code */ + size_t num_blocks; /* Number of blocks between processes. */ + size_t leftover_blocks; /* Number of leftover blocks to handle */ + int blocks, leftover; /* converted to int for MPI */ MPI_Aint * chunk_disp_array = NULL; + MPI_Aint * block_disps = NULL; int * block_lens = NULL; MPI_Datatype mem_type = MPI_BYTE, file_type = MPI_BYTE; H5FD_mpio_xfer_t prev_xfer_mode; /* Previous data xfer mode */ hbool_t have_xfer_mode = FALSE; /* Whether the previous xffer mode has been retrieved */ - hbool_t need_addr_sort = FALSE; - int i; /* Local index variable */ + hbool_t need_sort = FALSE; + size_t i; /* Local index variable */ herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_STATIC + /* + * If a separate fill buffer is provided for partial chunks, ensure + * that the "don't filter partial edge chunks" flag is set. + */ + if (partial_chunk_fill_buf) + HDassert(dset->shared->layout.u.chunk.flags & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + /* Get the MPI communicator */ if (MPI_COMM_NULL == (mpi_comm = H5F_mpi_get_comm(dset->oloc.file))) HGOTO_ERROR(H5E_INTERNAL, H5E_MPI, FAIL, "Can't retrieve MPI communicator") @@ -4972,39 +5372,89 @@ H5D__chunk_collective_fill(const H5D_t *dset, H5D_chunk_coll_info_t *chunk_info, /* Distribute evenly the number of blocks between processes. */ if (mpi_size == 0) HGOTO_ERROR(H5E_DATASET, H5E_BADVALUE, FAIL, "Resulted in division by zero") - num_blocks = (size_t)(chunk_info->num_io / (size_t)mpi_size); /* value should be the same on all procs */ + num_blocks = + (size_t)(chunk_fill_info->num_chunks / (size_t)mpi_size); /* value should be the same on all procs */ /* After evenly distributing the blocks between processes, are there any * leftover blocks for each individual process (round-robin)? */ - leftover_blocks = (size_t)(chunk_info->num_io % (size_t)mpi_size); + leftover_blocks = (size_t)(chunk_fill_info->num_chunks % (size_t)mpi_size); /* Cast values to types needed by MPI */ H5_CHECKED_ASSIGN(blocks, int, num_blocks, size_t); H5_CHECKED_ASSIGN(leftover, int, leftover_blocks, size_t); - H5_CHECKED_ASSIGN(block_len, int, chunk_size, size_t); /* Check if we have any chunks to write on this rank */ if (num_blocks > 0 || (leftover && leftover > mpi_rank)) { + MPI_Aint partial_fill_buf_disp = 0; + hbool_t all_same_block_len = TRUE; + /* Allocate buffers */ - /* (MSC - should not need block_lens if MPI_type_create_hindexed_block is working) */ - if (NULL == (block_lens = (int *)H5MM_malloc((size_t)(blocks + 1) * sizeof(int)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk lengths buffer") if (NULL == (chunk_disp_array = (MPI_Aint *)H5MM_malloc((size_t)(blocks + 1) * sizeof(MPI_Aint)))) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk file displacement buffer") - for (i = 0; i < blocks; i++) { - /* store the chunk address as an MPI_Aint */ - chunk_disp_array[i] = (MPI_Aint)(chunk_info->addr[i + (mpi_rank * blocks)]); + if (partial_chunk_fill_buf) { + MPI_Aint fill_buf_addr; + MPI_Aint partial_fill_buf_addr; + + /* Calculate the displacement between the fill buffer and partial chunk fill buffer */ + if (MPI_SUCCESS != (mpi_code = MPI_Get_address(fill_buf, &fill_buf_addr))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code) + if (MPI_SUCCESS != (mpi_code = MPI_Get_address(partial_chunk_fill_buf, &partial_fill_buf_addr))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code) - /* MSC - should not need this if MPI_type_create_hindexed_block is working */ - block_lens[i] = block_len; +#if H5_CHECK_MPI_VERSION(3, 1) + partial_fill_buf_disp = MPI_Aint_diff(partial_fill_buf_addr, fill_buf_addr); +#else + partial_fill_buf_disp = partial_fill_buf_addr - fill_buf_addr; +#endif - /* Make sure that the addresses in the datatype are - * monotonically non-decreasing + /* + * Allocate all-zero block displacements array. If a block's displacement + * is left as zero, that block will be written to from the regular fill + * buffer. If a block represents an unfiltered partial edge chunk, its + * displacement will be set so that the block is written to from the + * unfiltered fill buffer. */ - if (i && (chunk_disp_array[i] < chunk_disp_array[i - 1])) - need_addr_sort = TRUE; + if (NULL == (block_disps = (MPI_Aint *)H5MM_calloc((size_t)(blocks + 1) * sizeof(MPI_Aint)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate block displacements buffer") + } + + /* + * Perform initial scan of chunk info list to: + * - make sure that chunk addresses are monotonically non-decreasing + * - check if all blocks have the same length + */ + for (i = 1; i < chunk_fill_info->num_chunks; i++) { + if (chunk_fill_info->chunk_info[i].addr < chunk_fill_info->chunk_info[i - 1].addr) + need_sort = TRUE; + + if (chunk_fill_info->chunk_info[i].chunk_size != chunk_fill_info->chunk_info[i - 1].chunk_size) + all_same_block_len = FALSE; + } + + if (need_sort) + HDqsort(chunk_fill_info->chunk_info, chunk_fill_info->num_chunks, + sizeof(struct chunk_coll_fill_info), H5D__chunk_cmp_coll_fill_info); + + /* Allocate buffer for block lengths if necessary */ + if (!all_same_block_len) + if (NULL == (block_lens = (int *)H5MM_malloc((size_t)(blocks + 1) * sizeof(int)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk lengths buffer") + + for (i = 0; i < (size_t)blocks; i++) { + size_t idx = i + (size_t)(mpi_rank * blocks); + + /* store the chunk address as an MPI_Aint */ + chunk_disp_array[i] = (MPI_Aint)(chunk_fill_info->chunk_info[idx].addr); + + if (!all_same_block_len) + H5_CHECKED_ASSIGN(block_lens[i], int, chunk_fill_info->chunk_info[idx].chunk_size, size_t); + + if (chunk_fill_info->chunk_info[idx].unfiltered_partial_chunk) { + HDassert(partial_chunk_fill_buf); + block_disps[i] = partial_fill_buf_disp; + } } /* end for */ /* Calculate if there are any leftover blocks after evenly @@ -5012,32 +5462,71 @@ H5D__chunk_collective_fill(const H5D_t *dset, H5D_chunk_coll_info_t *chunk_info, * to processes 0 -> leftover. */ if (leftover && leftover > mpi_rank) { - chunk_disp_array[blocks] = (MPI_Aint)chunk_info->addr[(blocks * mpi_size) + mpi_rank]; - if (blocks && (chunk_disp_array[blocks] < chunk_disp_array[blocks - 1])) - need_addr_sort = TRUE; - block_lens[blocks] = block_len; + chunk_disp_array[blocks] = + (MPI_Aint)chunk_fill_info->chunk_info[(blocks * mpi_size) + mpi_rank].addr; + + if (!all_same_block_len) + H5_CHECKED_ASSIGN(block_lens[blocks], int, + chunk_fill_info->chunk_info[(blocks * mpi_size) + mpi_rank].chunk_size, + size_t); + + if (chunk_fill_info->chunk_info[(blocks * mpi_size) + mpi_rank].unfiltered_partial_chunk) { + HDassert(partial_chunk_fill_buf); + block_disps[blocks] = partial_fill_buf_disp; + } + blocks++; } - /* Ensure that the blocks are sorted in monotonically non-decreasing - * order of offset in the file. - */ - if (need_addr_sort) - HDqsort(chunk_disp_array, (size_t)blocks, sizeof(MPI_Aint), H5D__chunk_cmp_addr); + /* Create file and memory types for the write operation */ + if (all_same_block_len) { + int block_len; + + H5_CHECKED_ASSIGN(block_len, int, chunk_fill_info->chunk_info[0].chunk_size, size_t); + + mpi_code = + MPI_Type_create_hindexed_block(blocks, block_len, chunk_disp_array, MPI_BYTE, &file_type); + if (mpi_code != MPI_SUCCESS) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed_block failed", mpi_code) + + if (partial_chunk_fill_buf) { + /* + * If filters are disabled for partial edge chunks, those chunks could + * potentially have the same block length as the other chunks, but still + * need to be written to using the unfiltered fill buffer. Use an hindexed + * block type rather than an hvector. + */ + mpi_code = + MPI_Type_create_hindexed_block(blocks, block_len, block_disps, MPI_BYTE, &mem_type); + if (mpi_code != MPI_SUCCESS) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed_block failed", mpi_code) + } + else { + mpi_code = MPI_Type_create_hvector(blocks, block_len, 0, MPI_BYTE, &mem_type); + if (mpi_code != MPI_SUCCESS) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hvector failed", mpi_code) + } + } + else { + /* + * Currently, different block lengths implies that there are partial + * edge chunks and the "don't filter partial edge chunks" flag is set. + */ + HDassert(partial_chunk_fill_buf); + HDassert(block_lens); + HDassert(block_disps); + + mpi_code = MPI_Type_create_hindexed(blocks, block_lens, chunk_disp_array, MPI_BYTE, &file_type); + if (mpi_code != MPI_SUCCESS) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code) + + mpi_code = MPI_Type_create_hindexed(blocks, block_lens, block_disps, MPI_BYTE, &mem_type); + if (mpi_code != MPI_SUCCESS) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code) + } - /* MSC - should use this if MPI_type_create_hindexed block is working: - * mpi_code = MPI_Type_create_hindexed_block(blocks, block_len, chunk_disp_array, MPI_BYTE, - * &file_type); - */ - mpi_code = MPI_Type_create_hindexed(blocks, block_lens, chunk_disp_array, MPI_BYTE, &file_type); - if (mpi_code != MPI_SUCCESS) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code) if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(&file_type))) HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) - - mpi_code = MPI_Type_create_hvector(blocks, block_len, 0, MPI_BYTE, &mem_type); - if (mpi_code != MPI_SUCCESS) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hvector failed", mpi_code) if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(&mem_type))) HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) } /* end if */ @@ -5080,39 +5569,25 @@ done: if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type))) HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) H5MM_xfree(chunk_disp_array); + H5MM_xfree(block_disps); H5MM_xfree(block_lens); FUNC_LEAVE_NOAPI(ret_value) } /* end H5D__chunk_collective_fill() */ static int -H5D__chunk_cmp_addr(const void *addr1, const void *addr2) +H5D__chunk_cmp_coll_fill_info(const void *_entry1, const void *_entry2) { - MPI_Aint _addr1 = (MPI_Aint)0, _addr2 = (MPI_Aint)0; - int ret_value = 0; + const struct chunk_coll_fill_info *entry1; + const struct chunk_coll_fill_info *entry2; FUNC_ENTER_STATIC_NOERR - _addr1 = *((const MPI_Aint *)addr1); - _addr2 = *((const MPI_Aint *)addr2); - -#if MPI_VERSION >= 3 && MPI_SUBVERSION >= 1 - { - MPI_Aint diff = MPI_Aint_diff(_addr1, _addr2); - - if (diff < (MPI_Aint)0) - ret_value = -1; - else if (diff > (MPI_Aint)0) - ret_value = 1; - else - ret_value = 0; - } -#else - ret_value = (_addr1 > _addr2) - (_addr1 < _addr2); -#endif + entry1 = (const struct chunk_coll_fill_info *)_entry1; + entry2 = (const struct chunk_coll_fill_info *)_entry2; - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5D__chunk_cmp_addr() */ + FUNC_LEAVE_NOAPI(H5F_addr_cmp(entry1->addr, entry2->addr)) +} /* end H5D__chunk_cmp_coll_fill_info() */ #endif /* H5_HAVE_PARALLEL */ /*------------------------------------------------------------------------- @@ -6696,10 +7171,10 @@ H5D__chunk_stats(const H5D_t *dset, hbool_t headers) miss_rate = 0.0; } if (miss_rate > 100) { - HDsprintf(ascii, "%7d%%", (int)(miss_rate + 0.5)); + HDsnprintf(ascii, sizeof(ascii), "%7d%%", (int)(miss_rate + 0.5)); } else { - HDsprintf(ascii, "%7.2f%%", miss_rate); + HDsnprintf(ascii, sizeof(ascii), "%7.2f%%", miss_rate); } HDfprintf(H5DEBUG(AC), " %-18s %8u %8u %7s %8d+%-9ld\n", "raw data chunks", rdcc->stats.nhits, @@ -6826,7 +7301,7 @@ done: * *------------------------------------------------------------------------- */ -static hbool_t +hbool_t H5D__chunk_is_partial_edge_chunk(unsigned dset_ndims, const uint32_t *chunk_dims, const hsize_t scaled[], const hsize_t *dset_dims) { @@ -7121,6 +7596,89 @@ done: } /* end H5D__chunk_format_convert() */ /*------------------------------------------------------------------------- + * Function: H5D__chunk_index_empty_cb + * + * Purpose: Callback function that simply stops iteration and sets the + * `empty` parameter to FALSE if called. If this callback is + * entered, it means that the chunk index contains at least + * one chunk, so is not empty. + * + * Return: H5_ITER_STOP + * + *------------------------------------------------------------------------- + */ +static int +H5D__chunk_index_empty_cb(const H5D_chunk_rec_t H5_ATTR_UNUSED *chunk_rec, void *_udata) +{ + hbool_t *empty = (hbool_t *)_udata; + int ret_value = H5_ITER_STOP; + + FUNC_ENTER_STATIC_NOERR + + *empty = FALSE; + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__chunk_index_empty_cb() */ + +/*------------------------------------------------------------------------- + * Function: H5D__chunk_index_empty + * + * Purpose: Determines whether a chunk index is empty (has no chunks + * inserted into it yet). + * + * Note: This routine is meant to be a little more performant than + * just counting the number of chunks in the index. In the + * future, this is probably a callback that the chunk index + * ops structure should provide. + * + * Return: Non-negative on Success/Negative on failure + * + *------------------------------------------------------------------------- + */ +herr_t +H5D__chunk_index_empty(const H5D_t *dset, hbool_t *empty) +{ + H5D_chk_idx_info_t idx_info; /* Chunked index info */ + H5D_rdcc_ent_t * ent; /* Cache entry */ + const H5D_rdcc_t * rdcc = NULL; /* Raw data chunk cache */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE_TAG(dset->oloc.addr) + + HDassert(dset); + HDassert(dset->shared); + HDassert(empty); + + rdcc = &(dset->shared->cache.chunk); /* raw data chunk cache */ + HDassert(rdcc); + + /* Search for cached chunks that haven't been written out */ + for (ent = rdcc->head; ent; ent = ent->next) + /* Flush the chunk out to disk, to make certain the size is correct later */ + if (H5D__chunk_flush_entry(dset, ent, FALSE) < 0) + HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "cannot flush indexed storage buffer") + + /* Compose chunked index info struct */ + idx_info.f = dset->oloc.file; + idx_info.pline = &dset->shared->dcpl_cache.pline; + idx_info.layout = &dset->shared->layout.u.chunk; + idx_info.storage = &dset->shared->layout.storage.u.chunk; + + *empty = TRUE; + + if (H5F_addr_defined(idx_info.storage->idx_addr)) { + /* Iterate over the allocated chunks */ + if ((dset->shared->layout.storage.u.chunk.ops->iterate)(&idx_info, H5D__chunk_index_empty_cb, empty) < + 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "unable to retrieve allocated chunk information from index") + } + +done: + FUNC_LEAVE_NOAPI_TAG(ret_value) +} /* end H5D__chunk_index_empty() */ + +/*------------------------------------------------------------------------- * Function: H5D__get_num_chunks_cb * * Purpose: Callback function that increments the number of written diff --git a/src/H5Dcompact.c b/src/H5Dcompact.c index 356a54e..1ac1267 100644 --- a/src/H5Dcompact.c +++ b/src/H5Dcompact.c @@ -63,8 +63,8 @@ typedef struct H5D_compact_iovv_memmanage_ud_t { /* Layout operation callbacks */ static herr_t H5D__compact_construct(H5F_t *f, H5D_t *dset); static hbool_t H5D__compact_is_space_alloc(const H5O_storage_t *storage); -static herr_t H5D__compact_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *cm); +static herr_t H5D__compact_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, + H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *cm); static herr_t H5D__compact_iovv_memmanage_cb(hsize_t dst_off, hsize_t src_off, size_t len, void *_udata); static ssize_t H5D__compact_readvv(const H5D_io_info_t *io_info, size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_size_arr[], hsize_t dset_offset_arr[], size_t mem_max_nseq, @@ -247,7 +247,7 @@ H5D__compact_is_space_alloc(const H5O_storage_t H5_ATTR_UNUSED *storage) *------------------------------------------------------------------------- */ static herr_t -H5D__compact_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info, +H5D__compact_io_init(H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info, hsize_t H5_ATTR_UNUSED nelmts, H5S_t H5_ATTR_UNUSED *file_space, H5S_t H5_ATTR_UNUSED *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *cm) { diff --git a/src/H5Dcontig.c b/src/H5Dcontig.c index 2ace14b..840c7ec 100644 --- a/src/H5Dcontig.c +++ b/src/H5Dcontig.c @@ -43,6 +43,7 @@ #include "H5FOprivate.h" /* File objects */ #include "H5Oprivate.h" /* Object headers */ #include "H5Pprivate.h" /* Property lists */ +#include "H5PBprivate.h" /* Page Buffer */ #include "H5VMprivate.h" /* Vector and array functions */ /****************/ @@ -90,8 +91,8 @@ typedef struct H5D_contig_writevv_ud_t { /* Layout operation callbacks */ static herr_t H5D__contig_construct(H5F_t *f, H5D_t *dset); static herr_t H5D__contig_init(H5F_t *f, const H5D_t *dset, hid_t dapl_id); -static herr_t H5D__contig_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *cm); +static herr_t H5D__contig_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, + H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *cm); static ssize_t H5D__contig_readvv(const H5D_io_info_t *io_info, size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[], size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[]); @@ -102,6 +103,7 @@ static herr_t H5D__contig_flush(H5D_t *dset); /* Helper routines */ static herr_t H5D__contig_write_one(H5D_io_info_t *io_info, hsize_t offset, size_t size); +static htri_t H5D__contig_may_use_select_io(const H5D_io_info_t *io_info, H5D_io_op_type_t op_type); /*********************/ /* Package Variables */ @@ -278,9 +280,16 @@ H5D__contig_fill(const H5D_io_info_t *io_info) if (using_mpi) { /* Write the chunks out from only one process */ /* !! Use the internal "independent" DXPL!! -QAK */ - if (H5_PAR_META_WRITE == mpi_rank) - if (H5D__contig_write_one(&ioinfo, offset, size) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to write fill value to dataset") + if (H5_PAR_META_WRITE == mpi_rank) { + if (H5D__contig_write_one(&ioinfo, offset, size) < 0) { + /* If writing fails, push an error and stop writing, but + * still participate in following MPI_Barrier. + */ + blocks_written = TRUE; + HDONE_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to write fill value to dataset") + break; + } + } /* Indicate that blocks are being written */ blocks_written = TRUE; @@ -559,19 +568,81 @@ H5D__contig_is_data_cached(const H5D_shared_t *shared_dset) *------------------------------------------------------------------------- */ static herr_t -H5D__contig_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info, +H5D__contig_io_init(H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info, hsize_t H5_ATTR_UNUSED nelmts, H5S_t H5_ATTR_UNUSED *file_space, H5S_t H5_ATTR_UNUSED *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *cm) { - FUNC_ENTER_STATIC_NOERR + htri_t use_selection_io = FALSE; /* Whether to use selection I/O */ + htri_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC io_info->store->contig.dset_addr = io_info->dset->shared->layout.storage.u.contig.addr; io_info->store->contig.dset_size = io_info->dset->shared->layout.storage.u.contig.size; - FUNC_LEAVE_NOAPI(SUCCEED) + /* Check if we're performing selection I/O */ + if ((use_selection_io = H5D__contig_may_use_select_io(io_info, H5D_IO_OP_READ)) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible") + io_info->use_select_io = (hbool_t)use_selection_io; + +done: + FUNC_LEAVE_NOAPI(ret_value) } /* end H5D__contig_io_init() */ /*------------------------------------------------------------------------- + * Function: H5D__contig_may_use_select_io + * + * Purpose: A small internal function to if it may be possible to use + * selection I/O. + * + * Return: TRUE/FALSE/FAIL + * + * Programmer: Neil Fortner + * 3 August 2021 + * + *------------------------------------------------------------------------- + */ +static htri_t +H5D__contig_may_use_select_io(const H5D_io_info_t *io_info, H5D_io_op_type_t op_type) +{ + const H5D_t *dataset = io_info->dset; /* Local pointer to dataset info */ + htri_t ret_value = FAIL; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity check */ + HDassert(io_info); + HDassert(dataset); + HDassert(op_type == H5D_IO_OP_READ || op_type == H5D_IO_OP_WRITE); + + /* Don't use selection I/O if it's globally disabled, if there is a type + * conversion, or if it's not a contiguous dataset, or if the sieve buffer + * exists (write) or is dirty (read) */ + if (!H5_use_selection_io_g || io_info->io_ops.single_read != H5D__select_read || + io_info->layout_ops.readvv != H5D__contig_readvv || + (op_type == H5D_IO_OP_READ && io_info->dset->shared->cache.contig.sieve_dirty) || + (op_type == H5D_IO_OP_WRITE && io_info->dset->shared->cache.contig.sieve_buf)) + ret_value = FALSE; + else { + hbool_t page_buf_enabled; + + HDassert(io_info->io_ops.single_write == H5D__select_write); + HDassert(io_info->layout_ops.writevv == H5D__contig_writevv); + + /* Check if the page buffer is enabled */ + if (H5PB_enabled(io_info->f_sh, H5FD_MEM_DRAW, &page_buf_enabled) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if page buffer is enabled") + if (page_buf_enabled) + ret_value = FALSE; + else + ret_value = TRUE; + } /* end else */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__contig_may_use_select_io() */ + +/*------------------------------------------------------------------------- * Function: H5D__contig_read * * Purpose: Read from a contiguous dataset. @@ -587,7 +658,7 @@ herr_t H5D__contig_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *fm) { - herr_t ret_value = SUCCEED; /*return value */ + herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_PACKAGE @@ -598,8 +669,20 @@ H5D__contig_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize HDassert(mem_space); HDassert(file_space); - /* Read data */ - if ((io_info->io_ops.single_read)(io_info, type_info, nelmts, file_space, mem_space) < 0) + if (io_info->use_select_io) { + size_t dst_type_size = type_info->dst_type_size; + + /* Issue selection I/O call (we can skip the page buffer because we've + * already verified it won't be used, and the metadata accumulator + * because this is raw data) */ + if (H5F_shared_select_read(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, nelmts > 0 ? 1 : 0, + &mem_space, &file_space, &(io_info->store->contig.dset_addr), + &dst_type_size, &(io_info->u.rbuf)) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "contiguous selection read failed") + } /* end if */ + else + /* Read data through legacy (non-selection I/O) pathway */ + if ((io_info->io_ops.single_read)(io_info, type_info, nelmts, file_space, mem_space) < 0) HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "contiguous read failed") done: @@ -622,7 +705,7 @@ herr_t H5D__contig_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *fm) { - herr_t ret_value = SUCCEED; /*return value */ + herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_PACKAGE @@ -633,8 +716,20 @@ H5D__contig_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsiz HDassert(mem_space); HDassert(file_space); - /* Write data */ - if ((io_info->io_ops.single_write)(io_info, type_info, nelmts, file_space, mem_space) < 0) + if (io_info->use_select_io) { + size_t dst_type_size = type_info->dst_type_size; + + /* Issue selection I/O call (we can skip the page buffer because we've + * already verified it won't be used, and the metadata accumulator + * because this is raw data) */ + if (H5F_shared_select_write(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, nelmts > 0 ? 1 : 0, + &mem_space, &file_space, &(io_info->store->contig.dset_addr), + &dst_type_size, &(io_info->u.wbuf)) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "contiguous selection write failed") + } /* end if */ + else + /* Write data through legacy (non-selection I/O) pathway */ + if ((io_info->io_ops.single_write)(io_info, type_info, nelmts, file_space, mem_space) < 0) HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "contiguous write failed") done: diff --git a/src/H5Dearray.c b/src/H5Dearray.c index abce233..cd52b66 100644 --- a/src/H5Dearray.c +++ b/src/H5Dearray.c @@ -417,7 +417,7 @@ H5D__earray_debug(FILE *stream, int indent, int fwidth, hsize_t idx, const void HDassert(elmt); /* Print element */ - HDsprintf(temp_str, "Element #%" PRIuHSIZE ":", idx); + HDsnprintf(temp_str, sizeof(temp_str), "Element #%" PRIuHSIZE ":", idx); HDfprintf(stream, "%*s%-*s %" PRIuHADDR "\n", indent, "", fwidth, temp_str, *(const haddr_t *)elmt); FUNC_LEAVE_NOAPI(SUCCEED) @@ -573,7 +573,7 @@ H5D__earray_filt_debug(FILE *stream, int indent, int fwidth, hsize_t idx, const HDassert(elmt); /* Print element */ - HDsprintf(temp_str, "Element #%" PRIuHSIZE ":", idx); + HDsnprintf(temp_str, sizeof(temp_str), "Element #%" PRIuHSIZE ":", idx); HDfprintf(stream, "%*s%-*s {%" PRIuHADDR ", %u, %0x}\n", indent, "", fwidth, temp_str, elmt->addr, elmt->nbytes, elmt->filter_mask); diff --git a/src/H5Defl.c b/src/H5Defl.c index a30955b..b22c6de 100644 --- a/src/H5Defl.c +++ b/src/H5Defl.c @@ -60,9 +60,9 @@ typedef struct H5D_efl_writevv_ud_t { /********************/ /* Layout operation callbacks */ -static herr_t H5D__efl_construct(H5F_t *f, H5D_t *dset); -static herr_t H5D__efl_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, - H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *cm); +static herr_t H5D__efl_construct(H5F_t *f, H5D_t *dset); +static herr_t H5D__efl_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, + H5S_t *file_space, H5S_t *mem_space, H5D_chunk_map_t *cm); static ssize_t H5D__efl_readvv(const H5D_io_info_t *io_info, size_t dset_max_nseq, size_t *dset_curr_seq, size_t dset_len_arr[], hsize_t dset_offset_arr[], size_t mem_max_nseq, size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[]); @@ -209,7 +209,7 @@ H5D__efl_is_space_alloc(const H5O_storage_t H5_ATTR_UNUSED *storage) *------------------------------------------------------------------------- */ static herr_t -H5D__efl_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info, +H5D__efl_io_init(H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info, hsize_t H5_ATTR_UNUSED nelmts, H5S_t H5_ATTR_UNUSED *file_space, H5S_t H5_ATTR_UNUSED *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *cm) { diff --git a/src/H5Dfarray.c b/src/H5Dfarray.c index 0741e8f..ab0f0f8 100644 --- a/src/H5Dfarray.c +++ b/src/H5Dfarray.c @@ -415,7 +415,7 @@ H5D__farray_debug(FILE *stream, int indent, int fwidth, hsize_t idx, const void HDassert(elmt); /* Print element */ - HDsprintf(temp_str, "Element #%" PRIuHSIZE ":", idx); + HDsnprintf(temp_str, sizeof(temp_str), "Element #%" PRIuHSIZE ":", idx); HDfprintf(stream, "%*s%-*s %" PRIuHADDR "\n", indent, "", fwidth, temp_str, *(const haddr_t *)elmt); FUNC_LEAVE_NOAPI(SUCCEED) @@ -675,7 +675,7 @@ H5D__farray_filt_debug(FILE *stream, int indent, int fwidth, hsize_t idx, const HDassert(elmt); /* Print element */ - HDsprintf(temp_str, "Element #%" PRIuHSIZE ":", idx); + HDsnprintf(temp_str, sizeof(temp_str), "Element #%" PRIuHSIZE ":", idx); HDfprintf(stream, "%*s%-*s {%" PRIuHADDR ", %u, %0x}\n", indent, "", fwidth, temp_str, elmt->addr, elmt->nbytes, elmt->filter_mask); diff --git a/src/H5Dint.c b/src/H5Dint.c index c9ea6bd..cc17265 100644 --- a/src/H5Dint.c +++ b/src/H5Dint.c @@ -378,40 +378,18 @@ H5D__get_space_status(const H5D_t *dset, H5D_space_status_t *allocation) /* Check for chunked layout */ if (dset->shared->layout.type == H5D_CHUNKED) { - hsize_t space_allocated; /* The number of bytes allocated for chunks */ - hssize_t snelmts; /* Temporary holder for number of elements in dataspace */ - hsize_t nelmts; /* Number of elements in dataspace */ - size_t dt_size; /* Size of datatype */ - hsize_t full_size; /* The number of bytes in the dataset when fully populated */ - - /* For chunked layout set the space status by the storage size */ - /* Get the dataset's dataspace */ - HDassert(dset->shared->space); - - /* Get the total number of elements in dataset's dataspace */ - if ((snelmts = H5S_GET_EXTENT_NPOINTS(dset->shared->space)) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "unable to retrieve number of elements in dataspace") - nelmts = (hsize_t)snelmts; - - /* Get the size of the dataset's datatype */ - if (0 == (dt_size = H5T_GET_SIZE(dset->shared->type))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "unable to retrieve size of datatype") - - /* Compute the maximum size of the dataset in bytes */ - full_size = nelmts * dt_size; - - /* Check for overflow during multiplication */ - if (nelmts != (full_size / dt_size)) - HGOTO_ERROR(H5E_DATASET, H5E_OVERFLOW, FAIL, "size of dataset's storage overflowed") - - /* Difficult to error check, since the error value is 0 and 0 is a valid value... :-/ */ - if (H5D__get_storage_size(dset, &space_allocated) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get size of dataset's storage") - - /* Decide on how much of the space is allocated */ - if (space_allocated == 0) + hsize_t n_chunks_total = dset->shared->layout.u.chunk.nchunks; + hsize_t n_chunks_alloc = 0; + + if (H5D__get_num_chunks(dset, dset->shared->space, &n_chunks_alloc) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "unable to retrieve number of allocated chunks in dataset") + + HDassert(n_chunks_alloc <= n_chunks_total); + + if (n_chunks_alloc == 0) *allocation = H5D_SPACE_STATUS_NOT_ALLOCATED; - else if (space_allocated == full_size) + else if (n_chunks_alloc == n_chunks_total) *allocation = H5D_SPACE_STATUS_ALLOCATED; else *allocation = H5D_SPACE_STATUS_PART_ALLOCATED; @@ -1301,10 +1279,19 @@ H5D__create(H5F_t *file, hid_t type_id, const H5S_t *space, hid_t dcpl_id, hid_t HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, NULL, "can't set latest indexing") } /* end if */ - /* Check if this dataset is going into a parallel file and set space allocation time */ + /* Check if the file driver would like to force early space allocation */ if (H5F_HAS_FEATURE(file, H5FD_FEAT_ALLOCATE_EARLY)) new_dset->shared->dcpl_cache.fill.alloc_time = H5D_ALLOC_TIME_EARLY; + /* + * Check if this dataset is going into a parallel file and set space allocation time. + * If the dataset has filters applied to it, writes to the dataset must be collective, + * so we don't need to force early space allocation. Otherwise, we force early space + * allocation to facilitate independent raw data operations. + */ + if (H5F_HAS_FEATURE(file, H5FD_FEAT_HAS_MPI) && (new_dset->shared->dcpl_cache.pline.nused == 0)) + new_dset->shared->dcpl_cache.fill.alloc_time = H5D_ALLOC_TIME_EARLY; + /* Set the dataset's I/O operations */ if (H5D__layout_set_io_ops(new_dset) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, NULL, "unable to initialize I/O operations") diff --git a/src/H5Dio.c b/src/H5Dio.c index 1a71ce2..cb61b71 100644 --- a/src/H5Dio.c +++ b/src/H5Dio.c @@ -165,7 +165,7 @@ H5D__read(H5D_t *dataset, hid_t mem_type_id, H5S_t *mem_space, H5S_t *file_space * difficulties with the notion. * * To solve this, we check to see if H5S_select_shape_same() returns true, - * and if the ranks of the mem and file spaces are different. If the are, + * and if the ranks of the mem and file spaces are different. If they are, * construct a new mem space that is equivalent to the old mem space, and * use that instead. * @@ -300,6 +300,7 @@ H5D__write(H5D_t *dataset, hid_t mem_type_id, H5S_t *mem_space, H5S_t *file_spac H5D_io_info_t io_info; /* Dataset I/O info */ H5D_type_info_t type_info; /* Datatype info for operation */ hbool_t type_info_init = FALSE; /* Whether the datatype info has been initialized */ + hbool_t should_alloc_space = FALSE; /* Whether or not to initialize dataset's storage */ H5S_t * projected_mem_space = NULL; /* If not NULL, ptr to dataspace containing a */ /* projection of the supplied mem_space to a new */ /* dataspace with rank equal to that of */ @@ -432,8 +433,20 @@ H5D__write(H5D_t *dataset, hid_t mem_type_id, H5S_t *mem_space, H5S_t *file_spac HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to set up I/O operation") /* Allocate dataspace and initialize it if it hasn't been. */ - if (nelmts > 0 && dataset->shared->dcpl_cache.efl.nused == 0 && - !(*dataset->shared->layout.ops->is_space_alloc)(&dataset->shared->layout.storage)) { + should_alloc_space = dataset->shared->dcpl_cache.efl.nused == 0 && + !(*dataset->shared->layout.ops->is_space_alloc)(&dataset->shared->layout.storage); + + /* + * If not using an MPI-based VFD, we only need to allocate + * and initialize storage if there's a selection in the + * dataset's dataspace. Otherwise, we always need to participate + * in the storage allocation since this may use collective + * operations and we will hang if we don't participate. + */ + if (!H5F_HAS_FEATURE(dataset->oloc.file, H5FD_FEAT_HAS_MPI)) + should_alloc_space = should_alloc_space && (nelmts > 0); + + if (should_alloc_space) { hssize_t file_nelmts; /* Number of elements in file dataset's dataspace */ hbool_t full_overwrite; /* Whether we are over-writing all the elements */ @@ -563,6 +576,10 @@ H5D__ioinfo_init(H5D_t *dset, const H5D_type_info_t *type_info, H5D_storage_t *s io_info->io_ops.single_write = H5D__scatgath_write; } /* end else */ + /* Start with selection I/O off, layout callback will turn it on if + * appropriate */ + io_info->use_select_io = FALSE; + #ifdef H5_HAVE_PARALLEL /* Determine if the file was opened with an MPI VFD */ io_info->using_mpi_vfd = H5F_HAS_FEATURE(dset->oloc.file, H5FD_FEAT_HAS_MPI); @@ -801,105 +818,47 @@ H5D__ioinfo_adjust(H5D_io_info_t *io_info, const H5D_t *dset, const H5S_t *file_ /* Check if we can use the optimized parallel I/O routines */ if (opt == TRUE) { - /* Override the I/O op pointers to the MPI-specific routines */ - io_info->io_ops.multi_read = dset->shared->layout.ops->par_read; - io_info->io_ops.multi_write = dset->shared->layout.ops->par_write; - io_info->io_ops.single_read = H5D__mpio_select_read; - io_info->io_ops.single_write = H5D__mpio_select_write; - } /* end if */ + /* Override the I/O op pointers to the MPI-specific routines, unless + * selection I/O is to be used - in this case the file driver will + * handle collective I/O */ + /* Check for selection/vector support in file driver? -NAF */ + if (!io_info->use_select_io) { + io_info->io_ops.multi_read = dset->shared->layout.ops->par_read; + io_info->io_ops.multi_write = dset->shared->layout.ops->par_write; + io_info->io_ops.single_read = H5D__mpio_select_read; + io_info->io_ops.single_write = H5D__mpio_select_write; + } /* end if */ + } /* end if */ else { - int comm_size = 0; - - /* Retrieve size of MPI communicator used for file */ - if ((comm_size = H5F_shared_mpi_get_size(io_info->f_sh)) < 0) - HGOTO_ERROR(H5E_FILE, H5E_CANTGET, FAIL, "can't get MPI communicator size") - /* Check if there are any filters in the pipeline. If there are, * we cannot break to independent I/O if this is a write operation * with multiple ranks involved; otherwise, there will be metadata * inconsistencies in the file. */ - if (comm_size > 1 && io_info->op_type == H5D_IO_OP_WRITE && - io_info->dset->shared->dcpl_cache.pline.nused > 0) { - H5D_mpio_no_collective_cause_t cause; - uint32_t local_no_collective_cause; - uint32_t global_no_collective_cause; - hbool_t local_error_message_previously_written = FALSE; - hbool_t global_error_message_previously_written = FALSE; - size_t idx; - size_t cause_strings_len; - char local_no_collective_cause_string[512] = ""; - char global_no_collective_cause_string[512] = ""; - const char * cause_strings[] = { - "independent I/O was requested", - "datatype conversions were required", - "data transforms needed to be applied", - "optimized MPI types flag wasn't set", - "one of the dataspaces was neither simple nor scalar", - "dataset was not contiguous or chunked", - "parallel writes to filtered datasets are disabled", - "an error occurred while checking if collective I/O was possible"}; - - cause_strings_len = sizeof(cause_strings) / sizeof(cause_strings[0]); - - if (H5CX_get_mpio_local_no_coll_cause(&local_no_collective_cause) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, - "unable to get local no collective cause value") - if (H5CX_get_mpio_global_no_coll_cause(&global_no_collective_cause) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, - "unable to get global no collective cause value") - - /* Append each of the "reason for breaking collective I/O" error messages to the - * local and global no collective cause strings */ - for (cause = 1, idx = 0; - (cause < H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE) && (idx < cause_strings_len); - cause <<= 1, idx++) { - if (cause & local_no_collective_cause) { - size_t local_buffer_space = sizeof(local_no_collective_cause_string) - - HDstrlen(local_no_collective_cause_string) - 1; - - /* Check if there were any previous error messages included. If so, prepend a - * semicolon to separate the messages. - */ - if (local_buffer_space && local_error_message_previously_written) { - HDstrncat(local_no_collective_cause_string, "; ", local_buffer_space); - local_buffer_space -= MIN(local_buffer_space, 2); - } - - if (local_buffer_space) - HDstrncat(local_no_collective_cause_string, cause_strings[idx], - local_buffer_space); - - local_error_message_previously_written = TRUE; - } /* end if */ - - if (cause & global_no_collective_cause) { - size_t global_buffer_space = sizeof(global_no_collective_cause_string) - - HDstrlen(global_no_collective_cause_string) - 1; - - /* Check if there were any previous error messages included. If so, prepend a - * semicolon to separate the messages. - */ - if (global_buffer_space && global_error_message_previously_written) { - HDstrncat(global_no_collective_cause_string, "; ", global_buffer_space); - global_buffer_space -= MIN(global_buffer_space, 2); - } - - if (global_buffer_space) - HDstrncat(global_no_collective_cause_string, cause_strings[idx], - global_buffer_space); - - global_error_message_previously_written = TRUE; - } /* end if */ - } /* end for */ - - HGOTO_ERROR(H5E_IO, H5E_NO_INDEPENDENT, FAIL, - "Can't perform independent write with filters in pipeline.\n" - " The following caused a break from collective I/O:\n" - " Local causes: %s\n" - " Global causes: %s", - local_no_collective_cause_string, global_no_collective_cause_string); - } /* end if */ + if (io_info->op_type == H5D_IO_OP_WRITE && io_info->dset->shared->dcpl_cache.pline.nused > 0) { + int comm_size = 0; + + /* Retrieve size of MPI communicator used for file */ + if ((comm_size = H5F_shared_mpi_get_size(io_info->f_sh)) < 0) + HGOTO_ERROR(H5E_FILE, H5E_CANTGET, FAIL, "can't get MPI communicator size") + + if (comm_size > 1) { + char local_no_coll_cause_string[512]; + char global_no_coll_cause_string[512]; + + if (H5D__mpio_get_no_coll_cause_strings(local_no_coll_cause_string, 512, + global_no_coll_cause_string, 512) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "can't get reasons for breaking collective I/O") + + HGOTO_ERROR(H5E_IO, H5E_NO_INDEPENDENT, FAIL, + "Can't perform independent write with filters in pipeline.\n" + " The following caused a break from collective I/O:\n" + " Local causes: %s\n" + " Global causes: %s", + local_no_coll_cause_string, global_no_coll_cause_string); + } + } /* If we won't be doing collective I/O, but the user asked for * collective I/O, change the request to use independent I/O diff --git a/src/H5Dlayout.c b/src/H5Dlayout.c index 6c4fc12..6fdec05 100644 --- a/src/H5Dlayout.c +++ b/src/H5Dlayout.c @@ -213,7 +213,7 @@ H5D__layout_meta_size(const H5F_t *f, const H5O_layout_t *layout, hbool_t includ ret_value++; /* Dimension sizes */ - ret_value += layout->u.chunk.ndims * layout->u.chunk.enc_bytes_per_dim; + ret_value += layout->u.chunk.ndims * (size_t)layout->u.chunk.enc_bytes_per_dim; /* Type of chunk index */ ret_value++; diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c index f8bce33..2cde4d3 100644 --- a/src/H5Dmpio.c +++ b/src/H5Dmpio.c @@ -36,6 +36,7 @@ #include "H5Eprivate.h" /* Error handling */ #include "H5Fprivate.h" /* File access */ #include "H5FDprivate.h" /* File drivers */ +#include "H5FLprivate.h" /* Free Lists */ #include "H5Iprivate.h" /* IDs */ #include "H5MMprivate.h" /* Memory management */ #include "H5Oprivate.h" /* Object headers */ @@ -43,6 +44,15 @@ #include "H5Sprivate.h" /* Dataspaces */ #include "H5VMprivate.h" /* Vector */ +/* uthash is an external, header-only hash table implementation. + * + * We include the file directly in src/ and #define a few functions + * to use our internal memory calls. + */ +#define uthash_malloc(sz) H5MM_malloc(sz) +#define uthash_free(ptr, sz) H5MM_free(ptr) /* Ignoring sz is intentional */ +#include "uthash.h" + #ifdef H5_HAVE_PARALLEL /****************/ @@ -81,9 +91,54 @@ /* Macros to represent the regularity of the selection for multiple chunk IO case. */ #define H5D_CHUNK_SELECT_REG 1 +/* + * Threshold value for redistributing shared filtered chunks + * on all MPI ranks, or just MPI rank 0 + */ +#define H5D_CHUNK_REDISTRIBUTE_THRES ((size_t)((25 * H5_MB) / sizeof(H5D_chunk_redistribute_info_t))) + +/* + * Initial allocation size for the arrays that hold + * buffers for chunk modification data that is sent + * to other ranks and the MPI_Request objects for + * those send operations + */ +#define H5D_CHUNK_NUM_SEND_MSGS_INIT 64 + +/* + * Define a tag value for the MPI messages sent/received for + * chunk modification data + */ +#define H5D_CHUNK_MOD_DATA_TAG 64 + +/* + * Macro to initialize a H5D_chk_idx_info_t + * structure, given a pointer to a H5D_io_info_t + * structure + */ +#define H5D_MPIO_INIT_CHUNK_IDX_INFO(index_info, io_info_ptr) \ + do { \ + index_info.f = (io_info_ptr)->dset->oloc.file; \ + index_info.pline = &((io_info_ptr)->dset->shared->dcpl_cache.pline); \ + index_info.layout = &((io_info_ptr)->dset->shared->layout.u.chunk); \ + index_info.storage = &((io_info_ptr)->dset->shared->layout.storage.u.chunk); \ + } while (0) + +/* + * Macro to initialize a H5D_chunk_ud_t structure + * given a pointer to a H5D_chk_idx_info_t structure + */ +#define H5D_MPIO_INIT_CHUNK_UD_INFO(chunk_ud, index_info_ptr) \ + do { \ + HDmemset(&chunk_ud, 0, sizeof(H5D_chunk_ud_t)); \ + chunk_ud.common.layout = (index_info_ptr)->layout; \ + chunk_ud.common.storage = (index_info_ptr)->storage; \ + } while (0) + /******************/ /* Local Typedefs */ /******************/ + /* Combine chunk address and chunk info into a struct for better performance. */ typedef struct H5D_chunk_addr_info_t { haddr_t chunk_addr; @@ -100,115 +155,137 @@ typedef enum H5D_mpio_no_rank0_bcast_cause_t { } H5D_mpio_no_rank0_bcast_cause_t; /* + * Information necessary for re-allocating file space for a chunk + * during a parallel write of a chunked dataset with filters + * applied. + */ +typedef struct H5D_chunk_alloc_info_t { + H5F_block_t chunk_current; + H5F_block_t chunk_new; + hsize_t chunk_idx; +} H5D_chunk_alloc_info_t; + +/* + * Information for a chunk pertaining to the dataset's chunk + * index entry for the chunk + */ +typedef struct H5D_chunk_index_info_t { + hsize_t chunk_idx; + unsigned filter_mask; + hbool_t need_insert; +} H5D_chunk_index_info_t; + +/* * Information about a single chunk when performing collective filtered I/O. All * of the fields of one of these structs are initialized at the start of collective - * filtered I/O in the function H5D__construct_filtered_io_info_list(). - * - * This struct's fields are as follows: - * - * index - The "Index" of the chunk in the dataset. The index of a chunk is used during - * the collective re-insertion of chunks into the chunk index after the collective - * I/O has been performed. + * filtered I/O in the function H5D__mpio_collective_filtered_chunk_io_setup(). This + * struct's fields are as follows: * - * scaled - The scaled coordinates of the chunk in the dataset's file dataspace. The - * coordinates are used in both the collective re-allocation of space in the file - * and the collective re-insertion of chunks into the chunk index after the collective - * I/O has been performed. + * index_info - A structure containing the information needed when collectively + * re-inserting the chunk into the dataset's chunk index. The structure + * is distributed to all ranks during the re-insertion operation. Its fields + * are as follows: * - * full_overwrite - A flag which determines whether or not a chunk needs to be read from the - * file when being updated. If a chunk is being fully overwritten (the entire - * extent is selected in its file dataspace), then it is not necessary to - * read the chunk from the file. However, if the chunk is not being fully - * overwritten, it has to be read from the file in order to update the chunk - * without trashing the parts of the chunk that are not selected. + * chunk_idx - The index of the chunk in the dataset's chunk index. * - * num_writers - The total number of processors writing to this chunk. This field is used - * when the new owner of a chunk is receiving messages, which contain selections in - * the chunk and data to update the chunk with, from other processors which have this - * chunk selected in the I/O operation. The new owner must know how many processors it - * should expect messages from so that it can post an equal number of receive calls. + * filter_mask - A bit-mask that indicates which filters are to be applied to the + * chunk. Each filter in a chunk's filter pipeline has a bit position + * that can be masked to disable that particular filter for the chunk. + * This filter mask is saved alongside the chunk in the file. * - * io_size - The total size of I/O to this chunk. This field is an accumulation of the size of - * I/O to the chunk from each processor which has the chunk selected and is used to - * determine the value for the previous full_overwrite flag. + * need_insert - A flag which determines whether or not a chunk needs to be re-inserted into + * the chunk index after the write operation. * - * buf - A pointer which serves the dual purpose of holding either the chunk data which is to be - * written to the file or the chunk data which has been read from the file. + * chunk_info - A pointer to the chunk's H5D_chunk_info_t structure, which contains useful + * information like the dataspaces containing the selection in the chunk. * - * chunk_states - In the case of dataset writes only, this struct is used to track a chunk's size and - * address in the file before and after the filtering operation has occurred. + * chunk_current - The address in the file and size of this chunk before the filtering + * operation. When reading a chunk from the file, this field is used to + * read the correct amount of bytes. It is also used when redistributing + * shared chunks among MPI ranks and as a parameter to the chunk file + * space reallocation function. * - * Its fields are as follows: + * chunk_new - The address in the file and size of this chunk after the filtering + * operation. This field is relevant when collectively re-allocating space + * in the file for all of the chunks written to in the I/O operation, as + * their sizes may have changed after their data has been filtered. * - * chunk_current - The address in the file and size of this chunk before the filtering - * operation. When reading a chunk from the file, this field is used to - * read the correct amount of bytes. It is also used when redistributing - * shared chunks among processors and as a parameter to the chunk file - * space reallocation function. + * need_read - A flag which determines whether or not a chunk needs to be read from the + * file. During writes, if a chunk is being fully overwritten (the entire extent + * is selected in its file dataspace), then it is not necessary to read the chunk + * from the file. However, if the chunk is not being fully overwritten, it has to + * be read from the file in order to update the chunk without trashing the parts + * of the chunk that are not selected. During reads, this field should generally + * be true, but may be false if the chunk isn't allocated, for example. * - * new_chunk - The address in the file and size of this chunk after the filtering - * operation. This field is relevant when collectively re-allocating space - * in the file for all of the chunks written to in the I/O operation, as - * their sizes may have changed after their data has been filtered. + * skip_filter_pline - A flag which determines whether to skip calls to the filter pipeline + * for this chunk. This flag is mostly useful for correct handling of + * partial edge chunks when the "don't filter partial edge chunks" flag + * is set on the dataset's DCPL. * - * owners - In the case of dataset writes only, this struct is used to manage which single processor - * will ultimately write data out to the chunk. It allows the other processors to act according - * to the decision and send their selection in the chunk, as well as the data they wish - * to update the chunk with, to the processor which is writing to the chunk. + * io_size - The total size of I/O to this chunk. This field is an accumulation of the size of + * I/O to the chunk from each MPI rank which has the chunk selected and is used to + * determine the value for the previous `full_overwrite` flag. * - * Its fields are as follows: + * chunk_buf_size - The size in bytes of the data buffer allocated for the chunk * - * original_owner - The processor which originally had this chunk selected at the beginning of - * the collective filtered I/O operation. This field is currently used when - * redistributing shared chunks among processors. + * orig_owner - The MPI rank which originally had this chunk selected at the beginning of + * the collective filtered I/O operation. This field is currently used when + * redistributing shared chunks among MPI ranks. * - * new_owner - The processor which has been selected to perform the write to this chunk. + * new_owner - The MPI rank which has been selected to perform the modifications to this chunk. * - * async_info - In the case of dataset writes only, this struct is used by the owning processor of the - * chunk in order to manage the MPI send and receive calls made between it and all of - * the other processors which have this chunk selected in the I/O operation. + * num_writers - The total number of MPI ranks writing to this chunk. This field is used when + * the new owner of a chunk is receiving messages from other MPI ranks that + * contain their selections in the chunk and the data to update the chunk with. + * The new owner must know how many MPI ranks it should expect messages from so + * that it can post an equal number of receive calls. * - * Its fields are as follows: + * buf - A pointer which serves the dual purpose of holding either the chunk data which is to be + * written to the file or the chunk data which has been read from the file. * - * receive_requests_array - An array containing one MPI_Request for each of the - * asynchronous MPI receive calls the owning processor of this - * chunk makes to another processor in order to receive that - * processor's chunk modification data and selection in the chunk. + * hh - A handle for hash tables provided by the uthash.h header * - * receive_buffer_array - An array of buffers into which the owning processor of this chunk - * will store chunk modification data and the selection in the chunk - * received from another processor. - * - * num_receive_requests - The number of entries in the receive_request_array and - * receive_buffer_array fields. */ typedef struct H5D_filtered_collective_io_info_t { - hsize_t index; - hsize_t scaled[H5O_LAYOUT_NDIMS]; - hbool_t full_overwrite; - size_t num_writers; - size_t io_size; - void * buf; - - struct { - H5F_block_t chunk_current; - H5F_block_t new_chunk; - } chunk_states; - - struct { - int original_owner; - int new_owner; - } owners; - - struct { - MPI_Request * receive_requests_array; - unsigned char **receive_buffer_array; - int num_receive_requests; - } async_info; + H5D_chunk_index_info_t index_info; + + H5D_chunk_info_t *chunk_info; + H5F_block_t chunk_current; + H5F_block_t chunk_new; + hbool_t need_read; + hbool_t skip_filter_pline; + size_t io_size; + size_t chunk_buf_size; + int orig_owner; + int new_owner; + int num_writers; + void * buf; + + UT_hash_handle hh; } H5D_filtered_collective_io_info_t; -/* Function pointer typedef for sort function */ -typedef int (*H5D_mpio_sort_func_cb_t)(const void *, const void *); +/* + * Information necessary for redistributing shared chunks during + * a parallel write of a chunked dataset with filters applied. + */ +typedef struct H5D_chunk_redistribute_info_t { + H5F_block_t chunk_block; + hsize_t chunk_idx; + int orig_owner; + int new_owner; + int num_writers; +} H5D_chunk_redistribute_info_t; + +/* + * Information used when re-inserting a chunk into a dataset's + * chunk index during a parallel write of a chunked dataset with + * filters applied. + */ +typedef struct H5D_chunk_insert_info_t { + H5F_block_t chunk_block; + H5D_chunk_index_info_t index_info; +} H5D_chunk_insert_info_t; /********************/ /* Local Prototypes */ @@ -216,53 +293,98 @@ typedef int (*H5D_mpio_sort_func_cb_t)(const void *, const void *); static herr_t H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm); static herr_t H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - H5D_chunk_map_t *fm); + H5D_chunk_map_t *fm, int mpi_rank, int mpi_size); static herr_t H5D__multi_chunk_filtered_collective_io(H5D_io_info_t * io_info, - const H5D_type_info_t *type_info, H5D_chunk_map_t *fm); + const H5D_type_info_t *type_info, H5D_chunk_map_t *fm, + int mpi_rank, int mpi_size); static herr_t H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - H5D_chunk_map_t *fm, int sum_chunk); + H5D_chunk_map_t *fm, int sum_chunk, int mpi_rank, int mpi_size); static herr_t H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - H5D_chunk_map_t *fm); + H5D_chunk_map_t *fm, int mpi_rank, int mpi_size); static herr_t H5D__inter_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, const H5S_t *file_space, const H5S_t *mem_space); static herr_t H5D__final_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, MPI_Datatype mpi_file_type, MPI_Datatype mpi_buf_type); static herr_t H5D__sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, - H5D_chunk_addr_info_t chunk_addr_info_array[], int many_chunk_opt); + H5D_chunk_addr_info_t chunk_addr_info_array[], int many_chunk_opt, int mpi_rank, + int mpi_size); static herr_t H5D__obtain_mpio_mode(H5D_io_info_t *io_info, H5D_chunk_map_t *fm, uint8_t assign_io_mode[], - haddr_t chunk_addr[]); + haddr_t chunk_addr[], int mpi_rank, int mpi_size); static herr_t H5D__mpio_get_sum_chunk(const H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, int *sum_chunkf); -static herr_t H5D__construct_filtered_io_info_list(const H5D_io_info_t * io_info, - const H5D_type_info_t * type_info, - const H5D_chunk_map_t * fm, - H5D_filtered_collective_io_info_t **chunk_list, - size_t * num_entries); -#if MPI_VERSION >= 3 -static herr_t H5D__chunk_redistribute_shared_chunks(const H5D_io_info_t * io_info, - const H5D_type_info_t * type_info, - const H5D_chunk_map_t * fm, - H5D_filtered_collective_io_info_t *local_chunk_array, - size_t *local_chunk_array_num_entries); -#endif -static herr_t H5D__mpio_array_gatherv(void *local_array, size_t local_array_num_entries, - size_t array_entry_size, void **gathered_array, - size_t *gathered_array_num_entries, hbool_t allgather, int root, - MPI_Comm comm, int (*sort_func)(const void *, const void *)); -static herr_t H5D__mpio_filtered_collective_write_type(H5D_filtered_collective_io_info_t *chunk_list, - size_t num_entries, MPI_Datatype *new_mem_type, - hbool_t *mem_type_derived, MPI_Datatype *new_file_type, - hbool_t *file_type_derived); -static herr_t H5D__filtered_collective_chunk_entry_io(H5D_filtered_collective_io_info_t *chunk_entry, - const H5D_io_info_t * io_info, - const H5D_type_info_t * type_info, - const H5D_chunk_map_t * fm); +static herr_t H5D__mpio_collective_filtered_chunk_io_setup(const H5D_io_info_t * io_info, + const H5D_type_info_t * type_info, + const H5D_chunk_map_t * fm, + H5D_filtered_collective_io_info_t **chunk_list, + size_t *num_entries, int mpi_rank); +static herr_t H5D__mpio_redistribute_shared_chunks(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, + const H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, + int mpi_rank, int mpi_size, + size_t **rank_chunks_assigned_map); +static herr_t H5D__mpio_redistribute_shared_chunks_int(H5D_filtered_collective_io_info_t *chunk_list, + size_t * num_chunks_assigned_map, + hbool_t all_ranks_involved, + const H5D_io_info_t * io_info, + const H5D_chunk_map_t *fm, int mpi_rank, int mpi_size); +static herr_t H5D__mpio_share_chunk_modification_data(H5D_filtered_collective_io_info_t *chunk_list, + size_t *chunk_list_num_entries, H5D_io_info_t *io_info, + const H5D_type_info_t *type_info, int mpi_rank, + int mpi_size, + H5D_filtered_collective_io_info_t **chunk_hash_table, + unsigned char *** chunk_msg_bufs, + int * chunk_msg_bufs_len); +static herr_t H5D__mpio_collective_filtered_chunk_common_io(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, + const H5D_io_info_t * io_info, + const H5D_type_info_t *type_info, int mpi_size); +static herr_t H5D__mpio_collective_filtered_chunk_read(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, + const H5D_io_info_t * io_info, + const H5D_type_info_t *type_info, int mpi_rank, + int mpi_size); +static herr_t H5D__mpio_collective_filtered_chunk_update(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, + H5D_filtered_collective_io_info_t *chunk_hash_table, + unsigned char ** chunk_msg_bufs, + int chunk_msg_bufs_len, const H5D_io_info_t *io_info, + const H5D_type_info_t *type_info, int mpi_rank, + int mpi_size); +static herr_t H5D__mpio_collective_filtered_chunk_reallocate(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, + size_t * num_chunks_assigned_map, + H5D_io_info_t * io_info, + H5D_chk_idx_info_t *idx_info, int mpi_rank, + int mpi_size); +static herr_t H5D__mpio_collective_filtered_chunk_reinsert(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, + size_t * num_chunks_assigned_map, + H5D_io_info_t * io_info, + H5D_chk_idx_info_t *idx_info, int mpi_rank, + int mpi_size); +static herr_t H5D__mpio_get_chunk_redistribute_info_types(MPI_Datatype *contig_type, + hbool_t * contig_type_derived, + MPI_Datatype *resized_type, + hbool_t * resized_type_derived); +static herr_t H5D__mpio_get_chunk_alloc_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived, + MPI_Datatype *resized_type, hbool_t *resized_type_derived); +static herr_t H5D__mpio_get_chunk_insert_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived, + MPI_Datatype *resized_type, + hbool_t * resized_type_derived); +static herr_t H5D__mpio_collective_filtered_io_type(H5D_filtered_collective_io_info_t *chunk_list, + size_t num_entries, H5D_io_op_type_t op_type, + MPI_Datatype *new_mem_type, hbool_t *mem_type_derived, + MPI_Datatype *new_file_type, hbool_t *file_type_derived); static int H5D__cmp_chunk_addr(const void *chunk_addr_info1, const void *chunk_addr_info2); static int H5D__cmp_filtered_collective_io_info_entry(const void *filtered_collective_io_info_entry1, const void *filtered_collective_io_info_entry2); -#if MPI_VERSION >= 3 -static int H5D__cmp_filtered_collective_io_info_entry_owner(const void *filtered_collective_io_info_entry1, - const void *filtered_collective_io_info_entry2); +static int H5D__cmp_chunk_redistribute_info(const void *entry1, const void *entry2); +static int H5D__cmp_chunk_redistribute_info_orig_owner(const void *entry1, const void *entry2); + +#ifdef H5Dmpio_DEBUG +static herr_t H5D__mpio_debug_init(void); +static herr_t H5D__mpio_dump_collective_filtered_chunk_list(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, int mpi_rank); #endif /*********************/ @@ -273,6 +395,188 @@ static int H5D__cmp_filtered_collective_io_info_entry_owner(const void *filtered /* Local Variables */ /*******************/ +/* Declare extern free list to manage the H5S_sel_iter_t struct */ +H5FL_EXTERN(H5S_sel_iter_t); + +#ifdef H5Dmpio_DEBUG + +/* Flags to control debug actions in this file. + * (Meant to be indexed by characters) + * + * These flags can be set with either (or both) the environment variable + * "H5D_mpio_Debug" set to a string containing one or more characters + * (flags) or by setting them as a string value for the + * "H5D_mpio_debug_key" MPI Info key. + * + * Supported characters in 'H5D_mpio_Debug' string: + * 't' trace function entry and exit + * 'f' log to file rather than debugging stream + * 'm' show (rough) memory usage statistics + * 'c' show critical timing information + * + * To only show output from a particular MPI rank, specify its rank + * number as a character, e.g.: + * + * '0' only show output from rank 0 + * + * To only show output from a particular range (up to 8 ranks supported + * between 0-9) of MPI ranks, specify the start and end ranks separated + * by a hyphen, e.g.: + * + * '0-7' only show output from ranks 0 through 7 + * + */ +static int H5D_mpio_debug_flags_s[256]; +static int H5D_mpio_debug_rank_s[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; +static hbool_t H5D_mpio_debug_inited = FALSE; +static const char *const trace_in_pre = "-> "; +static const char *const trace_out_pre = "<- "; +static int debug_indent = 0; +static FILE * debug_stream = NULL; + +/* Determine if this rank should output debugging info */ +#define H5D_MPIO_DEBUG_THIS_RANK(rank) \ + (H5D_mpio_debug_rank_s[0] < 0 || rank == H5D_mpio_debug_rank_s[0] || rank == H5D_mpio_debug_rank_s[1] || \ + rank == H5D_mpio_debug_rank_s[2] || rank == H5D_mpio_debug_rank_s[3] || \ + rank == H5D_mpio_debug_rank_s[4] || rank == H5D_mpio_debug_rank_s[5] || \ + rank == H5D_mpio_debug_rank_s[6] || rank == H5D_mpio_debug_rank_s[7]) + +/* Print some debugging string */ +#define H5D_MPIO_DEBUG(rank, string) \ + do { \ + if (debug_stream && H5D_MPIO_DEBUG_THIS_RANK(rank)) { \ + HDfprintf(debug_stream, "%*s(Rank %d) " string "\n", debug_indent, "", rank); \ + HDfflush(debug_stream); \ + } \ + } while (0) + +/* Print some debugging string with printf-style arguments */ +#define H5D_MPIO_DEBUG_VA(rank, string, ...) \ + do { \ + if (debug_stream && H5D_MPIO_DEBUG_THIS_RANK(rank)) { \ + HDfprintf(debug_stream, "%*s(Rank %d) " string "\n", debug_indent, "", rank, __VA_ARGS__); \ + HDfflush(debug_stream); \ + } \ + } while (0) + +#define H5D_MPIO_TRACE_ENTER(rank) \ + do { \ + hbool_t trace_flag = H5D_mpio_debug_flags_s[(int)'t']; \ + \ + if (trace_flag) { \ + H5D_MPIO_DEBUG_VA(rank, "%s%s", trace_in_pre, __func__); \ + debug_indent += (int)HDstrlen(trace_in_pre); \ + } \ + } while (0) + +#define H5D_MPIO_TRACE_EXIT(rank) \ + do { \ + hbool_t trace_flag = H5D_mpio_debug_flags_s[(int)'t']; \ + \ + if (trace_flag) { \ + debug_indent -= (int)HDstrlen(trace_out_pre); \ + H5D_MPIO_DEBUG_VA(rank, "%s%s", trace_out_pre, __func__); \ + } \ + } while (0) + +#define H5D_MPIO_TIME_START(rank, op_name) \ + { \ + hbool_t time_flag = H5D_mpio_debug_flags_s[(int)'c']; \ + double start_time = 0.0, end_time = 0.0; \ + const char *const op = op_name; \ + \ + if (time_flag) { \ + start_time = MPI_Wtime(); \ + } + +#define H5D_MPIO_TIME_STOP(rank) \ + if (time_flag) { \ + end_time = MPI_Wtime(); \ + H5D_MPIO_DEBUG_VA(rank, "'%s' took %f seconds", op, (end_time - start_time)); \ + } \ + } + +/*--------------------------------------------------------------------------- + * Function: H5D__mpio_parse_debug_str + * + * Purpose: Parse a string for H5Dmpio-related debugging flags + * + * Returns: N/A + * + *--------------------------------------------------------------------------- + */ +static void +H5D__mpio_parse_debug_str(const char *s) +{ + FUNC_ENTER_STATIC_NOERR + + HDassert(s); + + while (*s) { + int c = (int)(*s); + + if (c >= (int)'0' && c <= (int)'9') { + hbool_t range = FALSE; + + if (*(s + 1) && *(s + 2)) + range = (int)*(s + 1) == '-' && (int)*(s + 2) >= (int)'0' && (int)*(s + 2) <= (int)'9'; + + if (range) { + int start_rank = c - (int)'0'; + int end_rank = (int)*(s + 2) - '0'; + int num_ranks = end_rank - start_rank + 1; + int i; + + if (num_ranks > 8) { + end_rank = start_rank + 7; + num_ranks = 8; + } + + for (i = 0; i < num_ranks; i++) + H5D_mpio_debug_rank_s[i] = start_rank++; + + s += 3; + } + else + H5D_mpio_debug_rank_s[0] = c - (int)'0'; + } + else + H5D_mpio_debug_flags_s[c]++; + + s++; + } + + FUNC_LEAVE_NOAPI_VOID +} + +static herr_t +H5D__mpio_debug_init(void) +{ + const char *debug_str; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC_NOERR + + HDassert(!H5D_mpio_debug_inited); + + /* Clear the debug flag buffer */ + HDmemset(H5D_mpio_debug_flags_s, 0, sizeof(H5D_mpio_debug_flags_s)); + + /* Retrieve and parse the H5Dmpio debug string */ + debug_str = HDgetenv("H5D_mpio_Debug"); + if (debug_str) + H5D__mpio_parse_debug_str(debug_str); + + if (H5DEBUG(D)) + debug_stream = H5DEBUG(D); + + H5D_mpio_debug_inited = TRUE; + + FUNC_LEAVE_NOAPI(ret_value) +} + +#endif + /*------------------------------------------------------------------------- * Function: H5D__mpio_opt_possible * @@ -347,14 +651,9 @@ H5D__mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space, co * use collective IO will defer until each chunk IO is reached. */ -#if MPI_VERSION < 3 - /* - * Don't allow parallel writes to filtered datasets if the MPI version - * is less than 3. The functions needed (MPI_Mprobe and MPI_Imrecv) will - * not be available. - */ - if (io_info->op_type == H5D_IO_OP_WRITE && io_info->dset->shared->layout.type == H5D_CHUNKED && - io_info->dset->shared->dcpl_cache.pline.nused > 0) +#ifndef H5_HAVE_PARALLEL_FILTERED_WRITES + /* Don't allow writes to filtered datasets if the functionality is disabled */ + if (io_info->op_type == H5D_IO_OP_WRITE && io_info->dset->shared->dcpl_cache.pline.nused > 0) local_cause[0] |= H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED; #endif @@ -437,6 +736,150 @@ done: } /* H5D__mpio_opt_possible() */ /*------------------------------------------------------------------------- + * Function: H5D__mpio_get_no_coll_cause_strings + * + * Purpose: When collective I/O is broken internally, it can be useful + * for users to see a representative string for the reason(s) + * why it was broken. This routine inspects the current + * "cause" flags from the API context and prints strings into + * the caller's buffers for the local and global reasons that + * collective I/O was broken. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +herr_t +H5D__mpio_get_no_coll_cause_strings(char *local_cause, size_t local_cause_len, char *global_cause, + size_t global_cause_len) +{ + uint32_t local_no_coll_cause; + uint32_t global_no_coll_cause; + size_t local_cause_bytes_written = 0; + size_t global_cause_bytes_written = 0; + int nbits; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_PACKAGE + + HDassert((local_cause && local_cause_len > 0) || (global_cause && global_cause_len > 0)); + + /* + * Use compile-time assertion so this routine is updated + * when any new "no collective cause" values are added + */ + HDcompile_assert(H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE == (H5D_mpio_no_collective_cause_t)256); + + /* Initialize output buffers */ + if (local_cause) + *local_cause = '\0'; + if (global_cause) + *global_cause = '\0'; + + /* Retrieve the local and global cause flags from the API context */ + if (H5CX_get_mpio_local_no_coll_cause(&local_no_coll_cause) < 0) + HGOTO_ERROR(H5E_CONTEXT, H5E_CANTGET, FAIL, "unable to get local no collective cause value") + if (H5CX_get_mpio_global_no_coll_cause(&global_no_coll_cause) < 0) + HGOTO_ERROR(H5E_CONTEXT, H5E_CANTGET, FAIL, "unable to get global no collective cause value") + + /* + * Append each of the "reason for breaking collective I/O" + * error messages to the local and global cause string buffers + */ + nbits = 8 * sizeof(local_no_coll_cause); + for (int bit_pos = 0; bit_pos < nbits; bit_pos++) { + H5D_mpio_no_collective_cause_t cur_cause; + const char * cause_str; + size_t buf_space_left; + + cur_cause = (H5D_mpio_no_collective_cause_t)(1 << bit_pos); + if (cur_cause == H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE) + break; + + switch (cur_cause) { + case H5D_MPIO_SET_INDEPENDENT: + cause_str = "independent I/O was requested"; + break; + case H5D_MPIO_DATATYPE_CONVERSION: + cause_str = "datatype conversions were required"; + break; + case H5D_MPIO_DATA_TRANSFORMS: + cause_str = "data transforms needed to be applied"; + break; + case H5D_MPIO_MPI_OPT_TYPES_ENV_VAR_DISABLED: + cause_str = "optimized MPI types flag wasn't set"; + break; + case H5D_MPIO_NOT_SIMPLE_OR_SCALAR_DATASPACES: + cause_str = "one of the dataspaces was neither simple nor scalar"; + break; + case H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET: + cause_str = "dataset was not contiguous or chunked"; + break; + case H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED: + cause_str = "parallel writes to filtered datasets are disabled"; + break; + case H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE: + cause_str = "an error occurred while checking if collective I/O was possible"; + break; + case H5D_MPIO_COLLECTIVE: + case H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE: + default: + HDassert(0 && "invalid no collective cause reason"); + break; + } + + /* + * Determine if the local reasons for breaking collective I/O + * included the current cause + */ + if (local_cause && (cur_cause & local_no_coll_cause)) { + buf_space_left = local_cause_len - local_cause_bytes_written; + + /* + * Check if there were any previous error messages included. If + * so, prepend a semicolon to separate the messages. + */ + if (buf_space_left && local_cause_bytes_written) { + HDstrncat(local_cause, "; ", buf_space_left); + local_cause_bytes_written += MIN(buf_space_left, 2); + buf_space_left -= MIN(buf_space_left, 2); + } + + if (buf_space_left) { + HDstrncat(local_cause, cause_str, buf_space_left); + local_cause_bytes_written += MIN(buf_space_left, HDstrlen(cause_str)); + } + } + + /* + * Determine if the global reasons for breaking collective I/O + * included the current cause + */ + if (global_cause && (cur_cause & global_no_coll_cause)) { + buf_space_left = global_cause_len - global_cause_bytes_written; + + /* + * Check if there were any previous error messages included. If + * so, prepend a semicolon to separate the messages. + */ + if (buf_space_left && global_cause_bytes_written) { + HDstrncat(global_cause, "; ", buf_space_left); + global_cause_bytes_written += MIN(buf_space_left, 2); + buf_space_left -= MIN(buf_space_left, 2); + } + + if (buf_space_left) { + HDstrncat(global_cause, cause_str, buf_space_left); + global_cause_bytes_written += MIN(buf_space_left, HDstrlen(cause_str)); + } + } + } + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_get_no_coll_cause_strings() */ + +/*------------------------------------------------------------------------- * Function: H5D__mpio_select_read * * Purpose: MPI-IO function to read directly from app buffer to file. @@ -500,145 +943,6 @@ done: } /* end H5D__mpio_select_write() */ /*------------------------------------------------------------------------- - * Function: H5D__mpio_array_gatherv - * - * Purpose: Given an array, specified in local_array, by each processor - * calling this function, collects each array into a single - * array which is then either gathered to the processor - * specified by root, when allgather is false, or is - * distributed back to all processors when allgather is true. - * - * The number of entries in the array contributed by an - * individual processor and the size of each entry should be - * specified in local_array_num_entries and array_entry_size, - * respectively. - * - * The MPI communicator to use should be specified for comm. - * - * If the sort_func argument is supplied, the array is sorted - * before the function returns. - * - * Note: if allgather is specified as true, root is ignored. - * - * Return: Non-negative on success/Negative on failure - * - * Programmer: Jordan Henderson - * Sunday, April 9th, 2017 - * - *------------------------------------------------------------------------- - */ -static herr_t -H5D__mpio_array_gatherv(void *local_array, size_t local_array_num_entries, size_t array_entry_size, - void **_gathered_array, size_t *_gathered_array_num_entries, hbool_t allgather, - int root, MPI_Comm comm, H5D_mpio_sort_func_cb_t sort_func) -{ - size_t gathered_array_num_entries = 0; /* The size of the newly-constructed array */ - void * gathered_array = NULL; /* The newly-constructed array returned to the caller */ - int *receive_counts_array = NULL; /* Array containing number of entries each processor is contributing */ - int *displacements_array = - NULL; /* Array of displacements where each processor places its data in the final array */ - int mpi_code, mpi_rank, mpi_size; - int sendcount; - herr_t ret_value = SUCCEED; - - FUNC_ENTER_STATIC - - HDassert(_gathered_array); - HDassert(_gathered_array_num_entries); - - MPI_Comm_size(comm, &mpi_size); - MPI_Comm_rank(comm, &mpi_rank); - - /* Determine the size of the end result array by collecting the number - * of entries contributed by each processor into a single total. - */ - if (MPI_SUCCESS != (mpi_code = MPI_Allreduce(&local_array_num_entries, &gathered_array_num_entries, 1, - MPI_INT, MPI_SUM, comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code) - - /* If 0 entries resulted from the collective operation, no processor is contributing anything and there is - * nothing to do */ - if (gathered_array_num_entries > 0) { - /* - * If gathering to all processors, all processors need to allocate space for the resulting array, as - * well as the receive counts and displacements arrays for the collective MPI_Allgatherv call. - * Otherwise, only the root processor needs to allocate the space for an MPI_Gatherv call. - */ - if (allgather || (mpi_rank == root)) { - if (NULL == (gathered_array = H5MM_malloc(gathered_array_num_entries * array_entry_size))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate gathered array") - - if (NULL == (receive_counts_array = (int *)H5MM_malloc((size_t)mpi_size * sizeof(int)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate receive counts array") - - if (NULL == (displacements_array = (int *)H5MM_malloc((size_t)mpi_size * sizeof(int)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate receive displacements array") - } /* end if */ - - /* - * If gathering to all processors, inform each processor of how many entries each other processor is - * contributing to the resulting array by collecting the counts into each processor's "receive counts" - * array. Otherwise, inform only the root processor of how many entries each other processor is - * contributing. - */ - if (allgather) { - if (MPI_SUCCESS != (mpi_code = MPI_Allgather(&local_array_num_entries, 1, MPI_INT, - receive_counts_array, 1, MPI_INT, comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code) - } /* end if */ - else { - if (MPI_SUCCESS != (mpi_code = MPI_Gather(&local_array_num_entries, 1, MPI_INT, - receive_counts_array, 1, MPI_INT, root, comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Gather failed", mpi_code) - } /* end else */ - - if (allgather || (mpi_rank == root)) { - size_t i; - - /* Multiply each receive count by the size of the array entry, since the data is sent as bytes. */ - for (i = 0; i < (size_t)mpi_size; i++) - H5_CHECKED_ASSIGN(receive_counts_array[i], int, - (size_t)receive_counts_array[i] * array_entry_size, size_t); - - /* Set receive buffer offsets for the collective MPI_Allgatherv/MPI_Gatherv call. */ - displacements_array[0] = 0; - for (i = 1; i < (size_t)mpi_size; i++) - displacements_array[i] = displacements_array[i - 1] + receive_counts_array[i - 1]; - } /* end if */ - - /* As the data is sent as bytes, calculate the true sendcount for the data. */ - H5_CHECKED_ASSIGN(sendcount, int, local_array_num_entries *array_entry_size, size_t); - - if (allgather) { - if (MPI_SUCCESS != - (mpi_code = MPI_Allgatherv(local_array, sendcount, MPI_BYTE, gathered_array, - receive_counts_array, displacements_array, MPI_BYTE, comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Allgatherv failed", mpi_code) - } /* end if */ - else { - if (MPI_SUCCESS != - (mpi_code = MPI_Gatherv(local_array, sendcount, MPI_BYTE, gathered_array, - receive_counts_array, displacements_array, MPI_BYTE, root, comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Gatherv failed", mpi_code) - } /* end else */ - - if (sort_func && (allgather || (mpi_rank == root))) - HDqsort(gathered_array, gathered_array_num_entries, array_entry_size, sort_func); - } /* end if */ - - *_gathered_array = gathered_array; - *_gathered_array_num_entries = gathered_array_num_entries; - -done: - if (receive_counts_array) - H5MM_free(receive_counts_array); - if (displacements_array) - H5MM_free(displacements_array); - - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5D__mpio_array_gatherv() */ - -/*------------------------------------------------------------------------- * Function: H5D__mpio_get_sum_chunk * * Purpose: Routine for obtaining total number of chunks to cover @@ -793,11 +1097,17 @@ static herr_t H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm) { H5FD_mpio_chunk_opt_t chunk_opt_mode; - int io_option = H5D_MULTI_CHUNK_IO_MORE_OPT; - int sum_chunk = -1; +#ifdef H5Dmpio_DEBUG + hbool_t log_file_flag = FALSE; + FILE * debug_log_file = NULL; +#endif #ifdef H5_HAVE_INSTRUMENTED_LIBRARY htri_t temp_not_link_io = FALSE; #endif + int io_option = H5D_MULTI_CHUNK_IO_MORE_OPT; + int sum_chunk = -1; + int mpi_rank; + int mpi_size; herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC @@ -808,9 +1118,35 @@ H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf HDassert(type_info); HDassert(fm); - /* Disable collective metadata reads for chunked dataset I/O operations - * in order to prevent potential hangs */ - H5CX_set_coll_metadata_read(FALSE); + /* Obtain the current rank of the process and the number of ranks */ + if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) + HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain MPI rank") + if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0) + HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain MPI size") + +#ifdef H5Dmpio_DEBUG + /* Initialize file-level debugging if not initialized */ + if (!H5D_mpio_debug_inited && H5D__mpio_debug_init() < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "can't initialize H5Dmpio debugging") + + /* Open file for debugging if necessary */ + log_file_flag = H5D_mpio_debug_flags_s[(int)'f']; + if (log_file_flag) { + char debug_log_filename[1024]; + time_t time_now; + + HDsnprintf(debug_log_filename, 1024, "H5Dmpio_debug.rank%d", mpi_rank); + + if (NULL == (debug_log_file = HDfopen(debug_log_filename, "a"))) + HGOTO_ERROR(H5E_IO, H5E_OPENERROR, FAIL, "couldn't open debugging log file") + + /* Print a short header for this I/O operation */ + time_now = HDtime(NULL); + HDfprintf(debug_log_file, "##### %s", HDasctime(HDlocaltime(&time_now))); + + debug_stream = debug_log_file; + } +#endif /* Check the optional property list for the collective chunk IO optimization option */ if (H5CX_get_mpio_chunk_opt_mode(&chunk_opt_mode) < 0) @@ -824,13 +1160,10 @@ H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf /* via default path. branch by num threshold */ else { unsigned one_link_chunk_io_threshold; /* Threshold to use single collective I/O for all chunks */ - int mpi_size; /* Number of processes in MPI job */ if (H5D__mpio_get_sum_chunk(io_info, fm, &sum_chunk) < 0) HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to obtain the total chunk number of all processes"); - if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size") /* Get the chunk optimization option threshold */ if (H5CX_get_mpio_chunk_opt_num(&one_link_chunk_io_threshold) < 0) @@ -876,22 +1209,12 @@ H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf case H5D_ONE_LINK_CHUNK_IO_MORE_OPT: /* Check if there are any filters in the pipeline */ if (io_info->dset->shared->dcpl_cache.pline.nused > 0) { - /* For now, Multi-chunk IO must be forced for parallel filtered read, - * so that data can be unfiltered as it is received. There is significant - * complexity in unfiltering the data when it is read all at once into a - * single buffer. - */ - if (io_info->op_type == H5D_IO_OP_READ) { - if (H5D__multi_chunk_filtered_collective_io(io_info, type_info, fm) < 0) - HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, - "couldn't finish optimized multiple filtered chunk MPI-IO") - } /* end if */ - else if (H5D__link_chunk_filtered_collective_io(io_info, type_info, fm) < 0) + if (H5D__link_chunk_filtered_collective_io(io_info, type_info, fm, mpi_rank, mpi_size) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish filtered linked chunk MPI-IO") } /* end if */ else /* Perform unfiltered link chunk collective IO */ - if (H5D__link_chunk_collective_io(io_info, type_info, fm, sum_chunk) < 0) + if (H5D__link_chunk_collective_io(io_info, type_info, fm, sum_chunk, mpi_rank, mpi_size) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish linked chunk MPI-IO") break; @@ -899,18 +1222,28 @@ H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf default: /* multiple chunk IO via threshold */ /* Check if there are any filters in the pipeline */ if (io_info->dset->shared->dcpl_cache.pline.nused > 0) { - if (H5D__multi_chunk_filtered_collective_io(io_info, type_info, fm) < 0) + if (H5D__multi_chunk_filtered_collective_io(io_info, type_info, fm, mpi_rank, mpi_size) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish optimized multiple filtered chunk MPI-IO") } /* end if */ else /* Perform unfiltered multi chunk collective IO */ - if (H5D__multi_chunk_collective_io(io_info, type_info, fm) < 0) + if (H5D__multi_chunk_collective_io(io_info, type_info, fm, mpi_rank, mpi_size) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish optimized multiple chunk MPI-IO") break; } /* end switch */ done: +#ifdef H5Dmpio_DEBUG + /* Close debugging log file */ + if (debug_log_file) { + HDfprintf(debug_log_file, "##############\n\n"); + if (EOF == HDfclose(debug_log_file)) + HDONE_ERROR(H5E_IO, H5E_CLOSEERROR, FAIL, "couldn't close debugging log file") + debug_stream = H5DEBUG(D); + } +#endif + FUNC_LEAVE_NOAPI(ret_value) } /* end H5D__chunk_collective_io */ @@ -993,7 +1326,7 @@ done: */ static herr_t H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm, - int sum_chunk) + int sum_chunk, int mpi_rank, int mpi_size) { H5D_chunk_addr_info_t *chunk_addr_info_array = NULL; MPI_Datatype chunk_final_mtype; /* Final memory MPI datatype for all chunks with selection */ @@ -1074,9 +1407,8 @@ H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *typ /* Set up the base storage address for this chunk */ io_info->store = &ctg_store; -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "before inter_collective_io for total chunk = 1 \n"); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "before inter_collective_io for total chunk = 1"); #endif /* Perform I/O */ @@ -1092,9 +1424,8 @@ H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *typ num_chunk = H5SL_count(fm->sel_chunks); H5_CHECK_OVERFLOW(num_chunk, size_t, int); -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "total_chunks = %zu, num_chunk = %zu\n", total_chunks, num_chunk); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG_VA(mpi_rank, "total_chunks = %zu, num_chunk = %zu", total_chunks, num_chunk); #endif /* Set up MPI datatype for chunks selected */ @@ -1125,18 +1456,17 @@ H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *typ HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk file is derived datatype flags buffer") -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "before sorting the chunk address \n"); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "before sorting chunk addresses"); #endif + /* Sort the chunk address */ - if (H5D__sort_chunk(io_info, fm, chunk_addr_info_array, sum_chunk) < 0) + if (H5D__sort_chunk(io_info, fm, chunk_addr_info_array, sum_chunk, mpi_rank, mpi_size) < 0) HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to sort chunk address") ctg_store.contig.dset_addr = chunk_addr_info_array[0].chunk_addr; -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "after sorting the chunk address \n"); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "after sorting chunk addresses"); #endif /* Obtain MPI derived datatype from all individual chunks */ @@ -1241,9 +1571,9 @@ H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *typ /* No chunks selected for this process */ mpi_buf_count = (hsize_t)0; } /* end else */ -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "before coming to final collective IO\n"); + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "before coming to final collective I/O"); #endif /* Set up the base storage address for this chunk */ @@ -1256,11 +1586,11 @@ H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *typ } /* end else */ done: -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "before freeing memory inside H5D_link_collective_io ret_value = %d\n", - ret_value); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG_VA(mpi_rank, "before freeing memory inside H5D_link_collective_io ret_value = %d", + ret_value); #endif + /* Release resources */ if (chunk_addr_info_array) H5MM_xfree(chunk_addr_info_array); @@ -1293,68 +1623,89 @@ done: /*------------------------------------------------------------------------- * Function: H5D__link_chunk_filtered_collective_io * - * Purpose: Routine for one collective IO with one MPI derived datatype - * to link with all filtered chunks - * - * 1. Construct a list of selected chunks in the collective IO - * operation - * A. If any chunk is being written to by more than 1 - * process, the process writing to the chunk which - * currently has the least amount of chunks assigned - * to it becomes the new owner (in the case of ties, - * the lowest MPI rank becomes the new owner) - * 2. If the operation is a write operation - * A. Loop through each chunk in the operation - * I. If this is not a full overwrite of the chunk - * a) Read the chunk from file and pass the chunk - * through the filter pipeline in reverse order - * (Unfilter the chunk) + * Purpose: Performs collective I/O on filtered chunks by creating a + * single MPI derived datatype to link with all filtered + * chunks. The general algorithm is as follows: + * + * 1. Construct a list of selected chunks in the collective + * I/O operation + * 2. If the operation is a read operation + * A. Ensure that the list of chunks is sorted in + * monotonically non-decreasing order of chunk offset + * in the file + * B. Participate in a collective read of chunks from + * the file + * C. Loop through each selected chunk, unfiltering it and + * scattering the data to the application's read buffer + * 3. If the operation is a write operation + * A. Redistribute any chunks being written by more than 1 + * MPI rank, such that the chunk is only owned by 1 MPI + * rank. The rank writing to the chunk which currently + * has the least amount of chunks assigned to it becomes + * the new owner (in the case of ties, the lowest MPI + * rank becomes the new owner) + * B. Participate in a collective read of chunks from the + * file + * C. Loop through each chunk selected in the operation + * and for each chunk: + * I. If we actually read the chunk from the file (if + * a chunk is being fully overwritten, we skip + * reading it), pass the chunk through the filter + * pipeline in reverse order (unfilter the chunk) * II. Update the chunk data with the modifications from - * the owning process + * the owning MPI rank * III. Receive any modification data from other - * processes and update the chunk data with these + * ranks and update the chunk data with those * modifications * IV. Filter the chunk - * B. Contribute the modified chunks to an array gathered - * by all processes which contains the new sizes of - * every chunk modified in the collective IO operation - * C. All processes collectively re-allocate each chunk - * from the gathered array with their new sizes after - * the filter operation - * D. If this process has any chunks selected in the IO - * operation, create an MPI derived type for memory and - * file to write out the process' selected chunks to the - * file - * E. Perform the collective write - * F. All processes collectively re-insert each modified + * D. Contribute the modified chunks to an array gathered + * by all ranks which contains information for + * re-allocating space in the file for every chunk + * modified. Then, each rank collectively re-allocates + * each chunk from the gathered array with their new + * sizes after the filter operation + * E. Proceed with the collective write operation for all + * the modified chunks + * F. Contribute the modified chunks to an array gathered + * by all ranks which contains information for + * re-inserting every chunk modified into the chunk + * index. Then, each rank collectively re-inserts each * chunk from the gathered array into the chunk index * + * TODO: Note that steps D. and F. here are both collective + * operations that partially share data from the + * H5D_filtered_collective_io_info_t structure. To + * try to conserve on memory a bit, the distributed + * arrays these operations create are discarded after + * each operation is performed. If memory consumption + * here proves to not be an issue, the necessary data + * for both operations could be combined into a single + * structure so that only one collective MPI operation + * is needed to carry out both operations, rather than + * two. * * Return: Non-negative on success/Negative on failure * - * Programmer: Jordan Henderson - * Friday, Nov. 4th, 2016 - * *------------------------------------------------------------------------- */ static herr_t H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - H5D_chunk_map_t *fm) + H5D_chunk_map_t *fm, int mpi_rank, int mpi_size) { - H5D_filtered_collective_io_info_t *chunk_list = NULL; /* The list of chunks being read/written */ - H5D_filtered_collective_io_info_t *collective_chunk_list = - NULL; /* The list of chunks used during collective operations */ - H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ - MPI_Datatype mem_type = MPI_BYTE; - MPI_Datatype file_type = MPI_BYTE; - hbool_t mem_type_is_derived = FALSE; - hbool_t file_type_is_derived = FALSE; - size_t chunk_list_num_entries; - size_t collective_chunk_list_num_entries; - size_t * num_chunks_selected_array = NULL; /* Array of number of chunks selected on each process */ - size_t i; /* Local index variable */ - int mpi_rank, mpi_size, mpi_code; - herr_t ret_value = SUCCEED; + H5D_filtered_collective_io_info_t *chunk_list = NULL; /* The list of chunks being read/written */ + H5D_filtered_collective_io_info_t *chunk_hash_table = NULL; + unsigned char ** chunk_msg_bufs = NULL; + H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ + MPI_Datatype mem_type = MPI_BYTE; + MPI_Datatype file_type = MPI_BYTE; + hbool_t mem_type_is_derived = FALSE; + hbool_t file_type_is_derived = FALSE; + size_t * rank_chunks_assigned_map = NULL; + size_t chunk_list_num_entries; + size_t i; + int chunk_msg_bufs_len = 0; + int mpi_code; + herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC @@ -1362,11 +1713,12 @@ H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_in HDassert(type_info); HDassert(fm); - /* Obtain the current rank of the process and the number of processes */ - if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank") - if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size") +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_DEBUG_VA(mpi_rank, "Performing Linked-chunk I/O (%s) with MPI Comm size of %d", + io_info->op_type == H5D_IO_OP_WRITE ? "write" : "read", mpi_size); + H5D_MPIO_TIME_START(mpi_rank, "Linked-chunk I/O"); +#endif /* Set the actual-chunk-opt-mode property. */ H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_LINK_CHUNK); @@ -1377,123 +1729,127 @@ H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_in H5CX_set_mpio_actual_io_mode(H5D_MPIO_CHUNK_COLLECTIVE); /* Build a list of selected chunks in the collective io operation */ - if (H5D__construct_filtered_io_info_list(io_info, type_info, fm, &chunk_list, &chunk_list_num_entries) < - 0) + if (H5D__mpio_collective_filtered_chunk_io_setup(io_info, type_info, fm, &chunk_list, + &chunk_list_num_entries, mpi_rank) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't construct filtered I/O info list") - if (io_info->op_type == H5D_IO_OP_WRITE) { /* Filtered collective write */ + if (io_info->op_type == H5D_IO_OP_READ) { /* Filtered collective read */ + if (H5D__mpio_collective_filtered_chunk_read(chunk_list, chunk_list_num_entries, io_info, type_info, + mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't read filtered chunks") + } + else { /* Filtered collective write */ H5D_chk_idx_info_t index_info; - H5D_chunk_ud_t udata; hsize_t mpi_buf_count; - /* Construct chunked index info */ - index_info.f = io_info->dset->oloc.file; - index_info.pline = &(io_info->dset->shared->dcpl_cache.pline); - index_info.layout = &(io_info->dset->shared->layout.u.chunk); - index_info.storage = &(io_info->dset->shared->layout.storage.u.chunk); - - /* Set up chunk information for insertion to chunk index */ - udata.common.layout = index_info.layout; - udata.common.storage = index_info.storage; - udata.filter_mask = 0; - - /* Iterate through all the chunks in the collective write operation, - * updating each chunk with the data modifications from other processes, - * then re-filtering the chunk. + H5D_MPIO_INIT_CHUNK_IDX_INFO(index_info, io_info); + + if (mpi_size > 1) { + /* Redistribute shared chunks being written to */ + if (H5D__mpio_redistribute_shared_chunks(chunk_list, chunk_list_num_entries, io_info, fm, + mpi_rank, mpi_size, &rank_chunks_assigned_map) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to redistribute shared chunks") + + /* Send any chunk modification messages for chunks this rank no longer owns */ + if (H5D__mpio_share_chunk_modification_data(chunk_list, &chunk_list_num_entries, io_info, + type_info, mpi_rank, mpi_size, &chunk_hash_table, + &chunk_msg_bufs, &chunk_msg_bufs_len) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, + "unable to send chunk modification data between MPI ranks") + + /* Make sure the local chunk list was updated correctly */ + HDassert(chunk_list_num_entries == rank_chunks_assigned_map[mpi_rank]); + } + + /* Proceed to update all the chunks this rank owns with its own + * modification data and data from other ranks, before re-filtering + * the chunks. As chunk reads are done collectively here, all ranks + * must participate. */ - for (i = 0; i < chunk_list_num_entries; i++) - if (mpi_rank == chunk_list[i].owners.new_owner) - if (H5D__filtered_collective_chunk_entry_io(&chunk_list[i], io_info, type_info, fm) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't process chunk entry") - - /* Gather the new chunk sizes to all processes for a collective reallocation - * of the chunks in the file. - */ - if (H5D__mpio_array_gatherv(chunk_list, chunk_list_num_entries, - sizeof(H5D_filtered_collective_io_info_t), - (void **)&collective_chunk_list, &collective_chunk_list_num_entries, true, - 0, io_info->comm, NULL) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "couldn't gather new chunk sizes") - - /* Collectively re-allocate the modified chunks (from each process) in the file */ - for (i = 0; i < collective_chunk_list_num_entries; i++) { - hbool_t insert; - - if (H5D__chunk_file_alloc(&index_info, &collective_chunk_list[i].chunk_states.chunk_current, - &collective_chunk_list[i].chunk_states.new_chunk, &insert, - collective_chunk_list[i].scaled) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate chunk") - } /* end for */ - - if (NULL == (num_chunks_selected_array = (size_t *)H5MM_malloc((size_t)mpi_size * sizeof(size_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate num chunks selected array") - - if (MPI_SUCCESS != - (mpi_code = MPI_Allgather(&chunk_list_num_entries, 1, MPI_UNSIGNED_LONG_LONG, - num_chunks_selected_array, 1, MPI_UNSIGNED_LONG_LONG, io_info->comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code) - - /* If this process has any chunks selected, create a MPI type for collectively - * writing out the chunks to file. Otherwise, the process contributes to the + if (H5D__mpio_collective_filtered_chunk_update(chunk_list, chunk_list_num_entries, chunk_hash_table, + chunk_msg_bufs, chunk_msg_bufs_len, io_info, type_info, + mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't update modified chunks") + + /* Free up resources used by chunk hash table now that we're done updating chunks */ + HASH_CLEAR(hh, chunk_hash_table); + + /* All ranks now collectively re-allocate file space for all chunks */ + if (H5D__mpio_collective_filtered_chunk_reallocate(chunk_list, chunk_list_num_entries, + rank_chunks_assigned_map, io_info, &index_info, + mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, + "couldn't collectively re-allocate file space for chunks") + + /* If this rank has any chunks selected, create a MPI type for collectively + * writing out the chunks to file. Otherwise, the rank contributes to the * collective write with a none type. */ - if (chunk_list_num_entries) { - size_t offset; - - /* During the collective re-allocation of chunks in the file, the record for each - * chunk is only updated in the collective array, not in the local copy of chunks on each - * process. However, each process needs the updated chunk records so that they can create - * a MPI type for the collective write that will write to the chunk's possible new locations - * in the file instead of the old ones. This ugly hack seems to be the best solution to - * copy the information back to the local array and avoid having to modify the collective - * write type function in an ugly way so that it will accept the collective array instead - * of the local array. This works correctly because the array gather function guarantees - * that the chunk data in the collective array is ordered in blocks by rank. - */ - for (i = 0, offset = 0; i < (size_t)mpi_rank; i++) - offset += num_chunks_selected_array[i]; - - H5MM_memcpy(chunk_list, &collective_chunk_list[offset], - num_chunks_selected_array[mpi_rank] * sizeof(H5D_filtered_collective_io_info_t)); + if (H5D__mpio_collective_filtered_io_type(chunk_list, chunk_list_num_entries, io_info->op_type, + &mem_type, &mem_type_is_derived, &file_type, + &file_type_is_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "couldn't create MPI type for writing filtered chunks") - /* Create single MPI type encompassing each selection in the dataspace */ - if (H5D__mpio_filtered_collective_write_type(chunk_list, chunk_list_num_entries, &mem_type, - &mem_type_is_derived, &file_type, - &file_type_is_derived) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_BADTYPE, FAIL, "couldn't create MPI link chunk I/O type") + mpi_buf_count = (file_type_is_derived || mem_type_is_derived) ? 1 : 0; - /* Override the write buffer to point to the address of the first - * chunk data buffer + /* Setup contig storage info for I/O operation */ + if (chunk_list_num_entries) { + /* + * Override the write buffer to point to the first + * chunk's data buffer */ io_info->u.wbuf = chunk_list[0].buf; - } /* end if */ - /* We have a single, complicated MPI datatype for both memory & file */ - mpi_buf_count = (mem_type_is_derived && file_type_is_derived) ? (hsize_t)1 : (hsize_t)0; - - /* Set up the base storage address for this operation */ - ctg_store.contig.dset_addr = 0; /* Write address must be set to address 0 */ - io_info->store = &ctg_store; + /* + * Setup the base storage address for this operation + * to be the first chunk's file address + */ + ctg_store.contig.dset_addr = chunk_list[0].chunk_new.offset; + } + else + ctg_store.contig.dset_addr = 0; /* Perform I/O */ + io_info->store = &ctg_store; if (H5D__final_collective_io(io_info, type_info, mpi_buf_count, file_type, mem_type) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish MPI-IO") + /* Free up resources in anticipation of following collective operation */ + for (i = 0; i < chunk_list_num_entries; i++) { + if (chunk_list[i].buf) { + H5MM_free(chunk_list[i].buf); + chunk_list[i].buf = NULL; + } + } + /* Participate in the collective re-insertion of all chunks modified - * in this iteration into the chunk index + * into the chunk index */ - for (i = 0; i < collective_chunk_list_num_entries; i++) { - udata.chunk_block = collective_chunk_list[i].chunk_states.new_chunk; - udata.common.scaled = collective_chunk_list[i].scaled; - udata.chunk_idx = collective_chunk_list[i].index; - - if ((index_info.storage->ops->insert)(&index_info, &udata, io_info->dset) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert chunk address into index") - } /* end for */ - } /* end if */ + if (H5D__mpio_collective_filtered_chunk_reinsert(chunk_list, chunk_list_num_entries, + rank_chunks_assigned_map, io_info, &index_info, + mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, + "couldn't collectively re-insert modified chunks into chunk index") + } done: - /* Free resources used by a process which had some selection */ + /* Free the MPI buf and file types, if they were derived */ + if (mem_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + if (file_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + + if (chunk_msg_bufs) { + for (i = 0; i < (size_t)chunk_msg_bufs_len; i++) + H5MM_free(chunk_msg_bufs[i]); + + H5MM_free(chunk_msg_bufs); + } + + HASH_CLEAR(hh, chunk_hash_table); + + /* Free resources used by a rank which had some selection */ if (chunk_list) { for (i = 0; i < chunk_list_num_entries; i++) if (chunk_list[i].buf) @@ -1502,16 +1858,13 @@ done: H5MM_free(chunk_list); } /* end if */ - if (num_chunks_selected_array) - H5MM_free(num_chunks_selected_array); - if (collective_chunk_list) - H5MM_free(collective_chunk_list); + if (rank_chunks_assigned_map) + H5MM_free(rank_chunks_assigned_map); - /* Free the MPI buf and file types, if they were derived */ - if (mem_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type))) - HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) - if (file_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type))) - HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif FUNC_LEAVE_NOAPI(ret_value) } /* end H5D__link_chunk_filtered_collective_io() */ @@ -1534,7 +1887,8 @@ done: *------------------------------------------------------------------------- */ static herr_t -H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm) +H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, H5D_chunk_map_t *fm, + int mpi_rank, int mpi_size) { H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */ H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ @@ -1547,11 +1901,8 @@ H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *ty H5FD_mpio_collective_opt_t last_coll_opt_mode = H5FD_MPIO_COLLECTIVE_IO; /* Last parallel transfer with independent IO or collective IO with this mode */ - size_t total_chunk; /* Total # of chunks in dataset */ -#ifdef H5Dmpio_DEBUG - int mpi_rank; -#endif - size_t u; /* Local index variable */ + size_t total_chunk; /* Total # of chunks in dataset */ + size_t u; /* Local index variable */ H5D_mpio_actual_io_mode_t actual_io_mode = H5D_MPIO_NO_COLLECTIVE; /* Local variable for tracking the I/O mode used. */ herr_t ret_value = SUCCEED; @@ -1561,10 +1912,6 @@ H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *ty /* Set the actual chunk opt mode property */ H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_MULTI_CHUNK); -#ifdef H5Dmpio_DEBUG - mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file); -#endif - /* Retrieve total # of chunks in dataset */ H5_CHECKED_ASSIGN(total_chunk, size_t, fm->layout->u.chunk.nchunks, hsize_t); HDassert(total_chunk != 0); @@ -1572,13 +1919,13 @@ H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *ty /* Allocate memories */ chunk_io_option = (uint8_t *)H5MM_calloc(total_chunk); chunk_addr = (haddr_t *)H5MM_calloc(total_chunk * sizeof(haddr_t)); -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "total_chunk %zu\n", total_chunk); + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG_VA(mpi_rank, "total_chunk %zu", total_chunk); #endif /* Obtain IO option for each chunk */ - if (H5D__obtain_mpio_mode(io_info, fm, chunk_io_option, chunk_addr) < 0) + if (H5D__obtain_mpio_mode(io_info, fm, chunk_io_option, chunk_addr, mpi_rank, mpi_size) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTRECV, FAIL, "unable to obtain MPIO mode") /* Set up contiguous I/O info object */ @@ -1606,9 +1953,8 @@ H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *ty H5S_t * fspace; /* Dataspace describing chunk & selection in it */ H5S_t * mspace; /* Dataspace describing selection in memory corresponding to this chunk */ -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "mpi_rank = %d, chunk index = %zu\n", mpi_rank, u); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG_VA(mpi_rank, "mpi_rank = %d, chunk index = %zu", mpi_rank, u); #endif /* Get the chunk info for this chunk, if there are elements selected */ chunk_info = fm->select_chunk[u]; @@ -1626,10 +1972,9 @@ H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *ty * needs to contribute MPI NONE TYPE. */ if (chunk_io_option[u] == H5D_CHUNK_IO_MODE_COL) { -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "inside collective chunk IO mpi_rank = %d, chunk index = %zu\n", - mpi_rank, u); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG_VA(mpi_rank, "inside collective chunk IO mpi_rank = %d, chunk index = %zu", + mpi_rank, u); #endif /* Set the file & memory dataspaces */ @@ -1665,10 +2010,9 @@ H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *ty HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish shared collective MPI-IO") } /* end if */ else { /* possible independent IO for this chunk */ -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "inside independent IO mpi_rank = %d, chunk index = %zu\n", mpi_rank, - u); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG_VA(mpi_rank, "inside independent IO mpi_rank = %d, chunk index = %zu", mpi_rank, + u); #endif HDassert(chunk_io_option[u] == 0); @@ -1698,9 +2042,8 @@ H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *ty /* Perform the I/O */ if (H5D__inter_collective_io(&ctg_io_info, type_info, fspace, mspace) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish shared collective MPI-IO") -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "after inter collective IO\n"); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "after inter collective IO"); #endif } /* end else */ } /* end for */ @@ -1720,80 +2063,101 @@ done: /*------------------------------------------------------------------------- * Function: H5D__multi_chunk_filtered_collective_io * - * Purpose: To do filtered collective IO iteratively to save on memory. - * While link_chunk_filtered_collective_io will construct and - * work on a list of all of the chunks selected in the IO - * operation at once, this function works iteratively on a set - * of chunks at a time; at most one chunk per rank per - * iteration. - * - * 1. Construct a list of selected chunks in the collective IO - * operation - * A. If any chunk is being written to by more than 1 - * process, the process writing to the chunk which - * currently has the least amount of chunks assigned - * to it becomes the new owner (in the case of ties, - * the lowest MPI rank becomes the new owner) - * 2. If the operation is a read operation - * A. Loop through each chunk in the operation - * I. Read the chunk from the file - * II. Unfilter the chunk - * III. Scatter the read chunk data to the user's buffer - * 3. If the operation is a write operation - * A. Loop through each chunk in the operation - * I. If this is not a full overwrite of the chunk - * a) Read the chunk from file and pass the chunk - * through the filter pipeline in reverse order - * (Unfilter the chunk) - * II. Update the chunk data with the modifications from - * the owning process - * III. Receive any modification data from other - * processes and update the chunk data with these - * modifications - * IV. Filter the chunk - * V. Contribute the chunk to an array gathered by - * all processes which contains every chunk - * modified in this iteration (up to one chunk - * per process, some processes may not have a - * selection/may have less chunks to work on than - * other processes) - * VI. All processes collectively re-allocate each - * chunk from the gathered array with their new - * sizes after the filter operation - * VII. Proceed with the collective write operation - * for the chunks modified on this iteration - * VIII. All processes collectively re-insert each - * chunk from the gathered array into the chunk - * index + * Purpose: Performs collective I/O on filtered chunks iteratively to + * save on memory and potentially get better performance + * depending on the average number of chunks per rank. While + * linked-chunk I/O will construct and work on a list of all + * of the chunks selected in the I/O operation at once, this + * function works iteratively on a set of chunks at a time; at + * most one chunk per rank per iteration. The general + * algorithm is as follows: + * + * 1. Construct a list of selected chunks in the collective + * I/O operation + * 2. If the operation is a read operation, loop an amount of + * times equal to the maximum number of chunks selected on + * any particular rank and on each iteration: + * A. Participate in a collective read of chunks from + * the file (ranks that run out of chunks still need + * to participate) + * B. Unfilter the chunk that was read (if any) + * C. Scatter the read chunk's data to the application's + * read buffer + * 3. If the operation is a write operation, redistribute any + * chunks being written to by more than 1 MPI rank, such + * that the chunk is only owned by 1 MPI rank. The rank + * writing to the chunk which currently has the least + * amount of chunks assigned to it becomes the new owner + * (in the case of ties, the lowest MPI rank becomes the + * new owner). Then, loop an amount of times equal to the + * maximum number of chunks selected on any particular + * rank and on each iteration: + * A. Participate in a collective read of chunks from + * the file (ranks that run out of chunks still need + * to participate) + * I. If we actually read a chunk from the file (if + * a chunk is being fully overwritten, we skip + * reading it), pass the chunk through the filter + * pipeline in reverse order (unfilter the chunk) + * B. Update the chunk data with the modifications from + * the owning rank + * C. Receive any modification data from other ranks and + * update the chunk data with those modifications + * D. Filter the chunk + * E. Contribute the chunk to an array gathered by + * all ranks which contains information for + * re-allocating space in the file for every chunk + * modified in this iteration (up to one chunk per + * rank; some ranks may not have a selection/may have + * less chunks to work on than other ranks). Then, + * each rank collectively re-allocates each chunk + * from the gathered array with their new sizes + * after the filter operation + * F. Proceed with the collective write operation + * for the chunks modified on this iteration + * G. Contribute the chunk to an array gathered by + * all ranks which contains information for + * re-inserting every chunk modified on this + * iteration into the chunk index. Then, each rank + * collectively re-inserts each chunk from the + * gathered array into the chunk index + * + * TODO: Note that steps E. and G. here are both collective + * operations that partially share data from the + * H5D_filtered_collective_io_info_t structure. To + * try to conserve on memory a bit, the distributed + * arrays these operations create are discarded after + * each operation is performed. If memory consumption + * here proves to not be an issue, the necessary data + * for both operations could be combined into a single + * structure so that only one collective MPI operation + * is needed to carry out both operations, rather than + * two. * * Return: Non-negative on success/Negative on failure * - * Programmer: Jordan Henderson - * Friday, Dec. 2nd, 2016 - * *------------------------------------------------------------------------- */ static herr_t H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - H5D_chunk_map_t *fm) + H5D_chunk_map_t *fm, int mpi_rank, int mpi_size) { - H5D_filtered_collective_io_info_t *chunk_list = NULL; /* The list of chunks being read/written */ - H5D_filtered_collective_io_info_t *collective_chunk_list = - NULL; /* The list of chunks used during collective operations */ - H5D_storage_t store; /* union of EFL and chunk pointer in file space */ - H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */ - H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ - MPI_Datatype *file_type_array = NULL; - MPI_Datatype *mem_type_array = NULL; - hbool_t * file_type_is_derived_array = NULL; - hbool_t * mem_type_is_derived_array = NULL; - hbool_t * has_chunk_selected_array = - NULL; /* Array of whether or not each process is contributing a chunk to each iteration */ - size_t chunk_list_num_entries; - size_t collective_chunk_list_num_entries; - size_t i, j; /* Local index variable */ - int mpi_rank, mpi_size, mpi_code; - herr_t ret_value = SUCCEED; + H5D_filtered_collective_io_info_t *chunk_list = NULL; /* The list of chunks being read/written */ + H5D_filtered_collective_io_info_t *chunk_hash_table = NULL; + unsigned char ** chunk_msg_bufs = NULL; + H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */ + H5D_storage_t ctg_store; /* Chunk storage information as contiguous dataset */ + MPI_Datatype mem_type = MPI_BYTE; + MPI_Datatype file_type = MPI_BYTE; + hbool_t mem_type_is_derived = FALSE; + hbool_t file_type_is_derived = FALSE; + hbool_t have_chunk_to_process; + size_t chunk_list_num_entries; + size_t i; + size_t max_num_chunks; + int chunk_msg_bufs_len = 0; + int mpi_code; + herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC @@ -1801,11 +2165,12 @@ H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_i HDassert(type_info); HDassert(fm); - /* Obtain the current rank of the process and the number of processes */ - if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank") - if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size") +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_DEBUG_VA(mpi_rank, "Performing Multi-chunk I/O (%s) with MPI Comm size of %d", + io_info->op_type == H5D_IO_OP_WRITE ? "write" : "read", mpi_size); + H5D_MPIO_TIME_START(mpi_rank, "Multi-chunk I/O"); +#endif /* Set the actual chunk opt mode property */ H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_MULTI_CHUNK); @@ -1816,10 +2181,19 @@ H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_i H5CX_set_mpio_actual_io_mode(H5D_MPIO_CHUNK_COLLECTIVE); /* Build a list of selected chunks in the collective IO operation */ - if (H5D__construct_filtered_io_info_list(io_info, type_info, fm, &chunk_list, &chunk_list_num_entries) < - 0) + if (H5D__mpio_collective_filtered_chunk_io_setup(io_info, type_info, fm, &chunk_list, + &chunk_list_num_entries, mpi_rank) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't construct filtered I/O info list") + /* Retrieve the maximum number of chunks selected for any rank */ + if (MPI_SUCCESS != (mpi_code = MPI_Allreduce(&chunk_list_num_entries, &max_num_chunks, 1, + MPI_UNSIGNED_LONG_LONG, MPI_MAX, io_info->comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code) + + /* If no one has anything selected at all, end the operation */ + if (0 == max_num_chunks) + HGOTO_DONE(SUCCEED); + /* Set up contiguous I/O info object */ H5MM_memcpy(&ctg_io_info, io_info, sizeof(ctg_io_info)); ctg_io_info.store = &ctg_store; @@ -1827,190 +2201,147 @@ H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_i /* Initialize temporary contiguous storage info */ ctg_store.contig.dset_size = (hsize_t)io_info->dset->shared->layout.u.chunk.size; - ctg_store.contig.dset_addr = 0; - - /* Set dataset storage for I/O info */ - io_info->store = &store; if (io_info->op_type == H5D_IO_OP_READ) { /* Filtered collective read */ - for (i = 0; i < chunk_list_num_entries; i++) - if (H5D__filtered_collective_chunk_entry_io(&chunk_list[i], io_info, type_info, fm) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't process chunk entry") - } /* end if */ + for (i = 0; i < max_num_chunks; i++) { + /* Check if this rank has a chunk to work on for this iteration */ + have_chunk_to_process = (i < chunk_list_num_entries); + + if (H5D__mpio_collective_filtered_chunk_read(have_chunk_to_process ? &chunk_list[i] : NULL, + have_chunk_to_process ? 1 : 0, io_info, type_info, + mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't read filtered chunks") + + if (have_chunk_to_process && chunk_list[i].buf) { + H5MM_free(chunk_list[i].buf); + chunk_list[i].buf = NULL; + } + } + } else { /* Filtered collective write */ H5D_chk_idx_info_t index_info; - H5D_chunk_ud_t udata; - size_t max_num_chunks; hsize_t mpi_buf_count; /* Construct chunked index info */ - index_info.f = io_info->dset->oloc.file; - index_info.pline = &(io_info->dset->shared->dcpl_cache.pline); - index_info.layout = &(io_info->dset->shared->layout.u.chunk); - index_info.storage = &(io_info->dset->shared->layout.storage.u.chunk); - - /* Set up chunk information for insertion to chunk index */ - udata.common.layout = index_info.layout; - udata.common.storage = index_info.storage; - udata.filter_mask = 0; - - /* Retrieve the maximum number of chunks being written among all processes */ - if (MPI_SUCCESS != (mpi_code = MPI_Allreduce(&chunk_list_num_entries, &max_num_chunks, 1, - MPI_UNSIGNED_LONG_LONG, MPI_MAX, io_info->comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code) - - /* If no one is writing anything at all, end the operation */ - if (!(max_num_chunks > 0)) - HGOTO_DONE(SUCCEED); - - /* Allocate arrays for storing MPI file and mem types and whether or not the - * types were derived. - */ - if (NULL == (file_type_array = (MPI_Datatype *)H5MM_malloc(max_num_chunks * sizeof(MPI_Datatype)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate file type array") - - if (NULL == (file_type_is_derived_array = (hbool_t *)H5MM_calloc(max_num_chunks * sizeof(hbool_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate file type is derived array") - - if (NULL == (mem_type_array = (MPI_Datatype *)H5MM_malloc(max_num_chunks * sizeof(MPI_Datatype)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate mem type array") - - if (NULL == (mem_type_is_derived_array = (hbool_t *)H5MM_calloc(max_num_chunks * sizeof(hbool_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate mem type is derived array") - - /* Iterate over the max number of chunks among all processes, as this process could - * have no chunks left to work on, but it still needs to participate in the collective - * re-allocation and re-insertion of chunks modified by other processes. + H5D_MPIO_INIT_CHUNK_IDX_INFO(index_info, io_info); + + if (mpi_size > 1) { + /* Redistribute shared chunks being written to */ + if (H5D__mpio_redistribute_shared_chunks(chunk_list, chunk_list_num_entries, io_info, fm, + mpi_rank, mpi_size, NULL) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to redistribute shared chunks") + + /* Send any chunk modification messages for chunks this rank no longer owns */ + if (H5D__mpio_share_chunk_modification_data(chunk_list, &chunk_list_num_entries, io_info, + type_info, mpi_rank, mpi_size, &chunk_hash_table, + &chunk_msg_bufs, &chunk_msg_bufs_len) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, + "unable to send chunk modification data between MPI ranks") + } + + /* Iterate over the max number of chunks among all ranks, as this rank could + * have no chunks left to work on, but it still needs to participate in the + * collective re-allocation and re-insertion of chunks modified by other ranks. */ for (i = 0; i < max_num_chunks; i++) { - /* Check if this process has a chunk to work on for this iteration */ - hbool_t have_chunk_to_process = - (i < chunk_list_num_entries) && (mpi_rank == chunk_list[i].owners.new_owner); - - if (have_chunk_to_process) - if (H5D__filtered_collective_chunk_entry_io(&chunk_list[i], io_info, type_info, fm) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't process chunk entry") + /* Check if this rank has a chunk to work on for this iteration */ + have_chunk_to_process = (i < chunk_list_num_entries) && (mpi_rank == chunk_list[i].new_owner); - /* Gather the new chunk sizes to all processes for a collective re-allocation - * of the chunks in the file + /* Proceed to update the chunk this rank owns (if any left) with its + * own modification data and data from other ranks, before re-filtering + * the chunks. As chunk reads are done collectively here, all ranks + * must participate. */ - if (H5D__mpio_array_gatherv(&chunk_list[i], have_chunk_to_process ? 1 : 0, - sizeof(H5D_filtered_collective_io_info_t), - (void **)&collective_chunk_list, &collective_chunk_list_num_entries, - true, 0, io_info->comm, NULL) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "couldn't gather new chunk sizes") - - /* Participate in the collective re-allocation of all chunks modified - * in this iteration. + if (H5D__mpio_collective_filtered_chunk_update(have_chunk_to_process ? &chunk_list[i] : NULL, + have_chunk_to_process ? 1 : 0, chunk_hash_table, + chunk_msg_bufs, chunk_msg_bufs_len, io_info, + type_info, mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't update modified chunks") + + /* All ranks now collectively re-allocate file space for all chunks */ + if (H5D__mpio_collective_filtered_chunk_reallocate(have_chunk_to_process ? &chunk_list[i] : NULL, + have_chunk_to_process ? 1 : 0, NULL, io_info, + &index_info, mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, + "couldn't collectively re-allocate file space for chunks") + + /* + * If this rank has a chunk to work on, create a MPI type + * for writing out the chunk. Otherwise, the rank will + * use MPI_BYTE for the file and memory type and specify + * a count of 0. */ - for (j = 0; j < collective_chunk_list_num_entries; j++) { - hbool_t insert = FALSE; + if (H5D__mpio_collective_filtered_io_type( + have_chunk_to_process ? &chunk_list[i] : NULL, have_chunk_to_process ? 1 : 0, + io_info->op_type, &mem_type, &mem_type_is_derived, &file_type, &file_type_is_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "couldn't create MPI type for writing filtered chunks") - if (H5D__chunk_file_alloc(&index_info, &collective_chunk_list[j].chunk_states.chunk_current, - &collective_chunk_list[j].chunk_states.new_chunk, &insert, - chunk_list[j].scaled) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate chunk") - } /* end for */ - - if (NULL == - (has_chunk_selected_array = (hbool_t *)H5MM_malloc((size_t)mpi_size * sizeof(hbool_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate num chunks selected array") - - if (MPI_SUCCESS != - (mpi_code = MPI_Allgather(&have_chunk_to_process, 1, MPI_C_BOOL, has_chunk_selected_array, 1, - MPI_C_BOOL, io_info->comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code) + mpi_buf_count = (file_type_is_derived || mem_type_is_derived) ? 1 : 0; - /* If this process has a chunk to work on, create a MPI type for the - * memory and file for writing out the chunk - */ + /* Override the write buffer to point to the chunk data buffer */ if (have_chunk_to_process) { - size_t offset; - int mpi_type_count; - - for (j = 0, offset = 0; j < (size_t)mpi_rank; j++) - offset += has_chunk_selected_array[j]; - - /* Collect the new chunk info back to the local copy, since only the record in the - * collective array gets updated by the chunk re-allocation */ - H5MM_memcpy(&chunk_list[i].chunk_states.new_chunk, - &collective_chunk_list[offset].chunk_states.new_chunk, - sizeof(chunk_list[i].chunk_states.new_chunk)); - - H5_CHECKED_ASSIGN(mpi_type_count, int, chunk_list[i].chunk_states.new_chunk.length, hsize_t); - - /* Create MPI memory type for writing to chunk */ - if (MPI_SUCCESS != - (mpi_code = MPI_Type_contiguous(mpi_type_count, MPI_BYTE, &mem_type_array[i]))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code) - if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(&mem_type_array[i]))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) - mem_type_is_derived_array[i] = TRUE; - - /* Create MPI file type for writing to chunk */ - if (MPI_SUCCESS != - (mpi_code = MPI_Type_contiguous(mpi_type_count, MPI_BYTE, &file_type_array[i]))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code) - if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(&file_type_array[i]))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) - file_type_is_derived_array[i] = TRUE; - - mpi_buf_count = 1; - - /* Set up the base storage address for this operation */ - ctg_store.contig.dset_addr = chunk_list[i].chunk_states.new_chunk.offset; - - /* Override the write buffer to point to the address of the - * chunk data buffer + /* + * Override the write buffer to point to the + * chunk's data buffer */ ctg_io_info.u.wbuf = chunk_list[i].buf; - } /* end if */ - else { - mem_type_array[i] = file_type_array[i] = MPI_BYTE; - mpi_buf_count = 0; - } /* end else */ + + /* + * Setup the base storage address for this + * operation to be the chunk's file address + */ + ctg_store.contig.dset_addr = chunk_list[i].chunk_new.offset; + } + else + ctg_store.contig.dset_addr = 0; /* Perform the I/O */ - if (H5D__final_collective_io(&ctg_io_info, type_info, mpi_buf_count, file_type_array[i], - mem_type_array[i]) < 0) + if (H5D__final_collective_io(&ctg_io_info, type_info, mpi_buf_count, file_type, mem_type) < 0) HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish MPI-IO") + /* Free up resources in anticipation of following collective operation */ + if (have_chunk_to_process && chunk_list[i].buf) { + H5MM_free(chunk_list[i].buf); + chunk_list[i].buf = NULL; + } + /* Participate in the collective re-insertion of all chunks modified * in this iteration into the chunk index */ - for (j = 0; j < collective_chunk_list_num_entries; j++) { - udata.chunk_block = collective_chunk_list[j].chunk_states.new_chunk; - udata.common.scaled = collective_chunk_list[j].scaled; - udata.chunk_idx = collective_chunk_list[j].index; - - if ((index_info.storage->ops->insert)(&index_info, &udata, io_info->dset) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, - "unable to insert chunk address into index") - } /* end for */ + if (H5D__mpio_collective_filtered_chunk_reinsert(have_chunk_to_process ? &chunk_list[i] : NULL, + have_chunk_to_process ? 1 : 0, NULL, io_info, + &index_info, mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, + "couldn't collectively re-insert modified chunks into chunk index") + + /* Free the MPI types, if they were derived */ + if (mem_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + mem_type_is_derived = FALSE; + if (file_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + file_type_is_derived = FALSE; + } /* end for */ + } - if (collective_chunk_list) { - H5MM_free(collective_chunk_list); - collective_chunk_list = NULL; - } /* end if */ - if (has_chunk_selected_array) { - H5MM_free(has_chunk_selected_array); - has_chunk_selected_array = NULL; - } /* end if */ - } /* end for */ +done: + /* Free the MPI buf and file types, if they were derived */ + if (mem_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + if (file_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) - /* Free the MPI file and memory types, if they were derived */ - for (i = 0; i < max_num_chunks; i++) { - if (file_type_is_derived_array[i]) - if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type_array[i]))) - HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + if (chunk_msg_bufs) { + for (i = 0; i < (size_t)chunk_msg_bufs_len; i++) + H5MM_free(chunk_msg_bufs[i]); - if (mem_type_is_derived_array[i]) - if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type_array[i]))) - HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) - } /* end for */ - } /* end else */ + H5MM_free(chunk_msg_bufs); + } -done: + HASH_CLEAR(hh, chunk_hash_table); + + /* Free resources used by a rank which had some selection */ if (chunk_list) { for (i = 0; i < chunk_list_num_entries; i++) if (chunk_list[i].buf) @@ -2019,16 +2350,10 @@ done: H5MM_free(chunk_list); } /* end if */ - if (collective_chunk_list) - H5MM_free(collective_chunk_list); - if (file_type_array) - H5MM_free(file_type_array); - if (mem_type_array) - H5MM_free(mem_type_array); - if (file_type_is_derived_array) - H5MM_free(file_type_is_derived_array); - if (mem_type_is_derived_array) - H5MM_free(mem_type_is_derived_array); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif FUNC_LEAVE_NOAPI(ret_value) } /* end H5D__multi_chunk_filtered_collective_io() */ @@ -2054,11 +2379,22 @@ H5D__inter_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf hbool_t mbt_is_derived = FALSE; hbool_t mft_is_derived = FALSE; MPI_Datatype mpi_file_type, mpi_buf_type; - int mpi_code; /* MPI return code */ - herr_t ret_value = SUCCEED; /* return value */ + int mpi_code; /* MPI return code */ +#ifdef H5Dmpio_DEBUG + int mpi_rank; +#endif + herr_t ret_value = SUCCEED; /* return value */ FUNC_ENTER_STATIC +#ifdef H5Dmpio_DEBUG + mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file); + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Inter collective I/O"); + if (mpi_rank < 0) + HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain MPI rank") +#endif + if ((file_space != NULL) && (mem_space != NULL)) { int mpi_file_count; /* Number of file "objects" to transfer */ hsize_t *permute_map = NULL; /* array that holds the mapping from the old, @@ -2117,9 +2453,8 @@ H5D__inter_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf mft_is_derived = FALSE; } /* end else */ -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "before final collective IO \n"); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "before final collective I/O"); #endif /* Perform final collective I/O operation */ @@ -2133,9 +2468,10 @@ done: if (mft_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mpi_file_type))) HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "before leaving inter_collective_io ret_value = %d\n", ret_value); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_DEBUG_VA(mpi_rank, "before leaving inter_collective_io ret_value = %d", ret_value); + H5D_MPIO_TRACE_EXIT(mpi_rank); #endif FUNC_LEAVE_NOAPI(ret_value) @@ -2157,10 +2493,21 @@ static herr_t H5D__final_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t mpi_buf_count, MPI_Datatype mpi_file_type, MPI_Datatype mpi_buf_type) { +#ifdef H5Dmpio_DEBUG + int mpi_rank; +#endif herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC +#ifdef H5Dmpio_DEBUG + mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file); + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Final collective I/O"); + if (mpi_rank < 0) + HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain MPI rank") +#endif + /* Pass buf type, file type to the file driver. */ if (H5CX_set_mpi_coll_datatypes(mpi_buf_type, mpi_file_type) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't set MPI-I/O collective I/O datatypes") @@ -2175,10 +2522,12 @@ H5D__final_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf } /* end else */ done: -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "ret_value before leaving final_collective_io=%d\n", ret_value); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_DEBUG_VA(mpi_rank, "ret_value before leaving final_collective_io=%d", ret_value); + H5D_MPIO_TRACE_EXIT(mpi_rank); #endif + FUNC_LEAVE_NOAPI(ret_value) } /* end H5D__final_collective_io */ @@ -2220,62 +2569,149 @@ H5D__cmp_chunk_addr(const void *chunk_addr_info1, const void *chunk_addr_info2) * * Return: -1, 0, 1 * - * Programmer: Jordan Henderson - * Wednesday, Nov. 30th, 2016 - * *------------------------------------------------------------------------- */ static int H5D__cmp_filtered_collective_io_info_entry(const void *filtered_collective_io_info_entry1, const void *filtered_collective_io_info_entry2) { - haddr_t addr1 = HADDR_UNDEF, addr2 = HADDR_UNDEF; + const H5D_filtered_collective_io_info_t *entry1; + const H5D_filtered_collective_io_info_t *entry2; + haddr_t addr1 = HADDR_UNDEF; + haddr_t addr2 = HADDR_UNDEF; + int ret_value; FUNC_ENTER_STATIC_NOERR - addr1 = ((const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry1) - ->chunk_states.new_chunk.offset; - addr2 = ((const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry2) - ->chunk_states.new_chunk.offset; + entry1 = (const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry1; + entry2 = (const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry2; - FUNC_LEAVE_NOAPI(H5F_addr_cmp(addr1, addr2)) -} /* end H5D__cmp_filtered_collective_io_info_entry() */ + addr1 = entry1->chunk_new.offset; + addr2 = entry2->chunk_new.offset; + + /* + * If both chunk addresses are defined, H5F_addr_cmp is safe to use. + * Otherwise, if both addresses aren't defined, compared chunk + * entries based on their chunk index. Finally, if only one chunk + * address is defined, return the appropriate value based on which + * is defined. + */ + if (H5F_addr_defined(addr1) && H5F_addr_defined(addr2)) { + ret_value = H5F_addr_cmp(addr1, addr2); + } + else if (!H5F_addr_defined(addr1) && !H5F_addr_defined(addr2)) { + hsize_t chunk_idx1 = entry1->index_info.chunk_idx; + hsize_t chunk_idx2 = entry2->index_info.chunk_idx; -#if MPI_VERSION >= 3 + ret_value = (chunk_idx1 > chunk_idx2) - (chunk_idx1 < chunk_idx2); + } + else + ret_value = H5F_addr_defined(addr1) ? 1 : -1; + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__cmp_filtered_collective_io_info_entry() */ /*------------------------------------------------------------------------- - * Function: H5D__cmp_filtered_collective_io_info_entry_owner + * Function: H5D__cmp_chunk_redistribute_info * - * Purpose: Routine to compare filtered collective chunk io info - * entries's original owner fields + * Purpose: Routine to compare two H5D_chunk_redistribute_info_t + * structures * - * Description: Callback for qsort() to compare filtered collective chunk - * io info entries's original owner fields + * Description: Callback for qsort() to compare two + * H5D_chunk_redistribute_info_t structures + * + * Return: -1, 0, 1 + * + *------------------------------------------------------------------------- + */ +static int +H5D__cmp_chunk_redistribute_info(const void *_entry1, const void *_entry2) +{ + const H5D_chunk_redistribute_info_t *entry1; + const H5D_chunk_redistribute_info_t *entry2; + hsize_t chunk_index1; + hsize_t chunk_index2; + int ret_value; + + FUNC_ENTER_STATIC_NOERR + + entry1 = (const H5D_chunk_redistribute_info_t *)_entry1; + entry2 = (const H5D_chunk_redistribute_info_t *)_entry2; + + chunk_index1 = entry1->chunk_idx; + chunk_index2 = entry2->chunk_idx; + + if (chunk_index1 == chunk_index2) { + int orig_owner1 = entry1->orig_owner; + int orig_owner2 = entry2->orig_owner; + + ret_value = (orig_owner1 > orig_owner2) - (orig_owner1 < orig_owner2); + } + else + ret_value = (chunk_index1 > chunk_index2) - (chunk_index1 < chunk_index2); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__cmp_chunk_redistribute_info() */ + +/*------------------------------------------------------------------------- + * Function: H5D__cmp_chunk_redistribute_info_orig_owner * - * Return: The difference between the two - * H5D_filtered_collective_io_info_t's original owner fields + * Purpose: Routine to compare the original owning MPI rank for two + * H5D_chunk_redistribute_info_t structures * - * Programmer: Jordan Henderson - * Monday, Apr. 10th, 2017 + * Description: Callback for qsort() to compare the original owning MPI + * rank for two H5D_chunk_redistribute_info_t + * structures + * + * Return: -1, 0, 1 * *------------------------------------------------------------------------- */ static int -H5D__cmp_filtered_collective_io_info_entry_owner(const void *filtered_collective_io_info_entry1, - const void *filtered_collective_io_info_entry2) +H5D__cmp_chunk_redistribute_info_orig_owner(const void *_entry1, const void *_entry2) { - int owner1 = -1, owner2 = -1; + const H5D_chunk_redistribute_info_t *entry1; + const H5D_chunk_redistribute_info_t *entry2; + int owner1 = -1; + int owner2 = -1; + int ret_value; FUNC_ENTER_STATIC_NOERR - owner1 = ((const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry1) - ->owners.original_owner; - owner2 = ((const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry2) - ->owners.original_owner; + entry1 = (const H5D_chunk_redistribute_info_t *)_entry1; + entry2 = (const H5D_chunk_redistribute_info_t *)_entry2; - FUNC_LEAVE_NOAPI(owner1 - owner2) -} /* end H5D__cmp_filtered_collective_io_info_entry_owner() */ -#endif + owner1 = entry1->orig_owner; + owner2 = entry2->orig_owner; + + if (owner1 == owner2) { + haddr_t addr1 = entry1->chunk_block.offset; + haddr_t addr2 = entry2->chunk_block.offset; + + /* + * If both chunk addresses are defined, H5F_addr_cmp is safe to use. + * Otherwise, if both addresses aren't defined, compared chunk + * entries based on their chunk index. Finally, if only one chunk + * address is defined, return the appropriate value based on which + * is defined. + */ + if (H5F_addr_defined(addr1) && H5F_addr_defined(addr2)) { + ret_value = H5F_addr_cmp(addr1, addr2); + } + else if (!H5F_addr_defined(addr1) && !H5F_addr_defined(addr2)) { + hsize_t chunk_idx1 = entry1->chunk_idx; + hsize_t chunk_idx2 = entry2->chunk_idx; + + ret_value = (chunk_idx1 > chunk_idx2) - (chunk_idx1 < chunk_idx2); + } + else + ret_value = H5F_addr_defined(addr1) ? 1 : -1; + } + else + ret_value = (owner1 > owner2) - (owner1 < owner2); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__cmp_chunk_redistribute_info_orig_owner() */ /*------------------------------------------------------------------------- * Function: H5D__sort_chunk @@ -2304,26 +2740,24 @@ H5D__cmp_filtered_collective_io_info_entry_owner(const void *filtered_collective */ static herr_t H5D__sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, - H5D_chunk_addr_info_t chunk_addr_info_array[], int sum_chunk) + H5D_chunk_addr_info_t chunk_addr_info_array[], int sum_chunk, int mpi_rank, int mpi_size) { - H5SL_node_t * chunk_node; /* Current node in chunk skip list */ - H5D_chunk_info_t *chunk_info; /* Current chunking info. of this node. */ - haddr_t chunk_addr; /* Current chunking address of this node */ - haddr_t *total_chunk_addr_array = NULL; /* The array of chunk address for the total number of chunk */ - hbool_t do_sort = FALSE; /* Whether the addresses need to be sorted */ - int bsearch_coll_chunk_threshold; - int many_chunk_opt = H5D_OBTAIN_ONE_CHUNK_ADDR_IND; - int mpi_size; /* Number of MPI processes */ - int mpi_code; /* MPI return code */ - int i; /* Local index variable */ - herr_t ret_value = SUCCEED; /* Return value */ + H5SL_node_t * chunk_node; /* Current node in chunk skip list */ + H5D_chunk_info_t *chunk_info; /* Current chunking info. of this node. */ + haddr_t chunk_addr; /* Current chunking address of this node */ + haddr_t *total_chunk_addr_array = NULL; /* The array of chunk address for the total number of chunk */ + H5P_coll_md_read_flag_t md_reads_file_flag; + hbool_t md_reads_context_flag; + hbool_t restore_md_reads_state = FALSE; + hbool_t do_sort = FALSE; /* Whether the addresses need to be sorted */ + int bsearch_coll_chunk_threshold; + int many_chunk_opt = H5D_OBTAIN_ONE_CHUNK_ADDR_IND; + int mpi_code; /* MPI return code */ + int i; /* Local index variable */ + herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_STATIC - /* Retrieve # of MPI processes */ - if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size") - /* Calculate the actual threshold to obtain all chunk addresses collectively * The bigger this number is, the more possible the use of obtaining chunk * address collectively. @@ -2337,31 +2771,56 @@ H5D__sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, ((sum_chunk / mpi_size) >= H5D_ALL_CHUNK_ADDR_THRES_COL_NUM)) many_chunk_opt = H5D_OBTAIN_ALL_CHUNK_ADDR_COL; -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "many_chunk_opt= %d\n", many_chunk_opt); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG_VA(mpi_rank, "many_chunk_opt = %d", many_chunk_opt); #endif /* If we need to optimize the way to obtain the chunk address */ if (many_chunk_opt != H5D_OBTAIN_ONE_CHUNK_ADDR_IND) { - int mpi_rank; - -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "Coming inside H5D_OBTAIN_ALL_CHUNK_ADDR_COL\n"); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "Coming inside H5D_OBTAIN_ALL_CHUNK_ADDR_COL"); #endif /* Allocate array for chunk addresses */ if (NULL == (total_chunk_addr_array = (haddr_t *)H5MM_malloc(sizeof(haddr_t) * (size_t)fm->layout->u.chunk.nchunks))) HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "unable to allocate memory chunk address array") - /* Retrieve all the chunk addresses with process 0 */ - if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank") - if (mpi_rank == 0) { - if (H5D__chunk_addrmap(io_info, total_chunk_addr_array) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address") + herr_t result; + + /* + * If enabled, disable collective metadata reads here. + * Since the chunk address mapping is done on rank 0 + * only here, it will cause problems if collective + * metadata reads are enabled. + */ + if (H5F_get_coll_metadata_reads(io_info->dset->oloc.file)) { + md_reads_file_flag = H5P_FORCE_FALSE; + md_reads_context_flag = FALSE; + H5F_set_coll_metadata_reads(io_info->dset->oloc.file, &md_reads_file_flag, + &md_reads_context_flag); + restore_md_reads_state = TRUE; + } + + result = H5D__chunk_addrmap(io_info, total_chunk_addr_array); + + /* Ensure that we restore the old collective metadata reads state */ + if (restore_md_reads_state) { + H5F_set_coll_metadata_reads(io_info->dset->oloc.file, &md_reads_file_flag, + &md_reads_context_flag); + restore_md_reads_state = FALSE; + } + + if (result < 0) { + size_t u; + + /* Clear total chunk address array */ + for (u = 0; u < (size_t)fm->layout->u.chunk.nchunks; u++) + total_chunk_addr_array[u] = HADDR_UNDEF; + + /* Push error, but still participate in following MPI_Bcast */ + HDONE_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address") + } } /* end if */ /* Broadcasting the MPI_IO option info. and chunk address info. */ @@ -2405,10 +2864,10 @@ H5D__sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, chunk_node = H5SL_next(chunk_node); } /* end while */ -#ifdef H5D_DEBUG - if (H5DEBUG(D)) - HDfprintf(H5DEBUG(D), "before Qsort\n"); +#ifdef H5Dmpio_DEBUG + H5D_MPIO_DEBUG(mpi_rank, "before Qsort"); #endif + if (do_sort) { size_t num_chunks = H5SL_count(fm->sel_chunks); @@ -2416,6 +2875,10 @@ H5D__sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, } /* end if */ done: + /* Re-enable collective metadata reads if we disabled them */ + if (restore_md_reads_state) + H5F_set_coll_metadata_reads(io_info->dset->oloc.file, &md_reads_file_flag, &md_reads_context_flag); + if (total_chunk_addr_array) H5MM_xfree(total_chunk_addr_array); @@ -2461,22 +2924,24 @@ done: */ static herr_t H5D__obtain_mpio_mode(H5D_io_info_t *io_info, H5D_chunk_map_t *fm, uint8_t assign_io_mode[], - haddr_t chunk_addr[]) + haddr_t chunk_addr[], int mpi_rank, int mpi_size) { - size_t total_chunks; - unsigned percent_nproc_per_chunk, threshold_nproc_per_chunk; - uint8_t * io_mode_info = NULL; - uint8_t * recv_io_mode_info = NULL; - uint8_t * mergebuf = NULL; - uint8_t * tempbuf; - H5SL_node_t * chunk_node; - H5D_chunk_info_t *chunk_info; - int mpi_size, mpi_rank; - MPI_Comm comm; - int root; - size_t ic; - int mpi_code; - herr_t ret_value = SUCCEED; + size_t total_chunks; + unsigned percent_nproc_per_chunk, threshold_nproc_per_chunk; + uint8_t * io_mode_info = NULL; + uint8_t * recv_io_mode_info = NULL; + uint8_t * mergebuf = NULL; + uint8_t * tempbuf; + H5SL_node_t * chunk_node; + H5D_chunk_info_t * chunk_info; + H5P_coll_md_read_flag_t md_reads_file_flag; + hbool_t md_reads_context_flag; + hbool_t restore_md_reads_state = FALSE; + MPI_Comm comm; + int root; + size_t ic; + int mpi_code; + herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC @@ -2484,12 +2949,6 @@ H5D__obtain_mpio_mode(H5D_io_info_t *io_info, H5D_chunk_map_t *fm, uint8_t assig root = 0; comm = io_info->comm; - /* Obtain the number of process and the current rank of the process */ - if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank") - if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size") - /* Setup parameters */ H5_CHECKED_ASSIGN(total_chunks, size_t, fm->layout->u.chunk.nchunks, hsize_t); if (H5CX_get_mpio_chunk_opt_ratio(&percent_nproc_per_chunk) < 0) @@ -2536,6 +2995,20 @@ H5D__obtain_mpio_mode(H5D_io_info_t *io_info, H5D_chunk_map_t *fm, uint8_t assig size_t nproc; unsigned *nproc_per_chunk; + /* + * If enabled, disable collective metadata reads here. + * Since the chunk address mapping is done on rank 0 + * only here, it will cause problems if collective + * metadata reads are enabled. + */ + if (H5F_get_coll_metadata_reads(io_info->dset->oloc.file)) { + md_reads_file_flag = H5P_FORCE_FALSE; + md_reads_context_flag = FALSE; + H5F_set_coll_metadata_reads(io_info->dset->oloc.file, &md_reads_file_flag, + &md_reads_context_flag); + restore_md_reads_state = TRUE; + } + /* pre-computing: calculate number of processes and regularity of the selection occupied in each chunk */ if (NULL == (nproc_per_chunk = (unsigned *)H5MM_calloc(total_chunks * sizeof(unsigned)))) @@ -2602,6 +3075,10 @@ H5D__obtain_mpio_mode(H5D_io_info_t *io_info, H5D_chunk_map_t *fm, uint8_t assig #endif done: + /* Re-enable collective metadata reads if we disabled them */ + if (restore_md_reads_state) + H5F_set_coll_metadata_reads(io_info->dset->oloc.file, &md_reads_file_flag, &md_reads_context_flag); + if (io_mode_info) H5MM_free(io_mode_info); if (mergebuf) @@ -2615,34 +3092,32 @@ done: } /* end H5D__obtain_mpio_mode() */ /*------------------------------------------------------------------------- - * Function: H5D__construct_filtered_io_info_list + * Function: H5D__mpio_collective_filtered_chunk_io_setup * * Purpose: Constructs a list of entries which contain the necessary * information for inter-process communication when performing * collective io on filtered chunks. This list is used by - * each process when performing I/O on locally selected chunks - * and also in operations that must be collectively done - * on every chunk, such as chunk re-allocation, insertion of - * chunks into the chunk index, etc. + * each MPI rank when performing I/O on locally selected + * chunks and also in operations that must be collectively + * done on every chunk, such as chunk re-allocation, insertion + * of chunks into the chunk index, etc. * * Return: Non-negative on success/Negative on failure * - * Programmer: Jordan Henderson - * Tuesday, January 10th, 2017 - * *------------------------------------------------------------------------- */ static herr_t -H5D__construct_filtered_io_info_list(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - const H5D_chunk_map_t * fm, - H5D_filtered_collective_io_info_t **chunk_list, size_t *num_entries) +H5D__mpio_collective_filtered_chunk_io_setup(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, + const H5D_chunk_map_t * fm, + H5D_filtered_collective_io_info_t **chunk_list, + size_t *num_entries, int mpi_rank) { - H5D_filtered_collective_io_info_t *local_info_array = - NULL; /* The list of initially selected chunks for this process */ - size_t num_chunks_selected; - size_t i; - int mpi_rank; - herr_t ret_value = SUCCEED; + H5D_filtered_collective_io_info_t *local_info_array = NULL; + H5D_chunk_ud_t udata; + hbool_t filter_partial_edge_chunks; + size_t num_chunks_selected; + size_t i; + herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC @@ -2652,19 +3127,23 @@ H5D__construct_filtered_io_info_list(const H5D_io_info_t *io_info, const H5D_typ HDassert(chunk_list); HDassert(num_entries); - if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank") +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Filtered Collective I/O Setup"); +#endif - /* Each process builds a local list of the chunks they have selected */ + /* Each rank builds a local list of the chunks they have selected */ if ((num_chunks_selected = H5SL_count(fm->sel_chunks))) { H5D_chunk_info_t *chunk_info; - H5D_chunk_ud_t udata; H5SL_node_t * chunk_node; hsize_t select_npoints; - hssize_t chunk_npoints; + hbool_t need_sort = FALSE; - if (NULL == (local_info_array = (H5D_filtered_collective_io_info_t *)H5MM_malloc( - num_chunks_selected * sizeof(H5D_filtered_collective_io_info_t)))) + /* Determine whether partial edge chunks should be filtered */ + filter_partial_edge_chunks = !(io_info->dset->shared->layout.u.chunk.flags & + H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + + if (NULL == (local_info_array = H5MM_malloc(num_chunks_selected * sizeof(*local_info_array)))) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate local io info array buffer") chunk_node = H5SL_first(fm->sel_chunks); @@ -2675,275 +3154,787 @@ H5D__construct_filtered_io_info_list(const H5D_io_info_t *io_info, const H5D_typ if (H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") - local_info_array[i].index = chunk_info->index; - local_info_array[i].chunk_states.chunk_current = local_info_array[i].chunk_states.new_chunk = - udata.chunk_block; - local_info_array[i].num_writers = 0; - local_info_array[i].owners.original_owner = local_info_array[i].owners.new_owner = mpi_rank; - local_info_array[i].buf = NULL; - - local_info_array[i].async_info.num_receive_requests = 0; - local_info_array[i].async_info.receive_buffer_array = NULL; - local_info_array[i].async_info.receive_requests_array = NULL; - - H5MM_memcpy(local_info_array[i].scaled, chunk_info->scaled, sizeof(chunk_info->scaled)); - - select_npoints = H5S_GET_SELECT_NPOINTS(chunk_info->mspace); - local_info_array[i].io_size = (size_t)select_npoints * type_info->src_type_size; - - /* Currently the full overwrite status of a chunk is only obtained on a per-process - * basis. This means that if the total selection in the chunk, as determined by the combination - * of selections of all of the processes interested in the chunk, covers the entire chunk, - * the performance optimization of not reading the chunk from the file is still valid, but - * is not applied in the current implementation. Something like an appropriately placed - * MPI_Allreduce or a running total of the number of chunk points selected during chunk - * redistribution should suffice for implementing this case - JTH. + /* Initialize rank-local chunk info */ + local_info_array[i].chunk_info = chunk_info; + local_info_array[i].chunk_buf_size = 0; + local_info_array[i].num_writers = 0; + local_info_array[i].orig_owner = mpi_rank; + local_info_array[i].new_owner = mpi_rank; + local_info_array[i].buf = NULL; + + select_npoints = H5S_GET_SELECT_NPOINTS(chunk_info->fspace); + local_info_array[i].io_size = (size_t)select_npoints * type_info->dst_type_size; + + /* + * Determine whether this chunk will need to be read from the file. If this is + * a read operation, the chunk will be read. If this is a write operation, we + * generally need to read a filtered chunk from the file before modifying it, + * unless the chunk is being fully overwritten. + * + * TODO: Currently the full overwrite status of a chunk is only obtained on a + * per-rank basis. This means that if the total selection in the chunk, as + * determined by the combination of selections of all of the ranks interested in + * the chunk, covers the entire chunk, the performance optimization of not reading + * the chunk from the file is still valid, but is not applied in the current + * implementation. + * + * To implement this case, a few approaches were considered: + * + * - Keep a running total (distributed to each rank) of the number of chunk + * elements selected during chunk redistribution and compare that to the total + * number of elements in the chunk once redistribution is finished + * + * - Process all incoming chunk messages before doing I/O (these are currently + * processed AFTER doing I/O), combine the owning rank's selection in a chunk + * with the selections received from other ranks and check to see whether that + * combined selection covers the entire chunk + * + * The first approach will be dangerous if the application performs an overlapping + * write to a chunk, as the number of selected elements can equal or exceed the + * number of elements in the chunk without the whole chunk selection being covered. + * While it might be considered erroneous for an application to do an overlapping + * write, we don't explicitly disallow it. + * + * The second approach contains a bit of complexity in that part of the chunk + * messages will be needed before doing I/O and part will be needed after doing I/O. + * Since modification data from chunk messages can't be applied until after any I/O + * is performed (otherwise, we'll overwrite any applied modification data), chunk + * messages are currently entirely processed after I/O. However, in order to determine + * if a chunk is being fully overwritten, we need the dataspace portion of the chunk + * messages before doing I/O. The naive way to do this is to process chunk messages + * twice, using just the relevant information from the message before and after I/O. + * The better way would be to avoid processing chunk messages twice by extracting (and + * keeping around) the dataspace portion of the message before I/O and processing the + * rest of the chunk message after I/O. Note that the dataspace portion of each chunk + * message is used to correctly apply chunk modification data from the message, so + * must be kept around both before and after I/O in this case. + */ + if (io_info->op_type == H5D_IO_OP_READ) + local_info_array[i].need_read = TRUE; + else { + local_info_array[i].need_read = + local_info_array[i].io_size < (size_t)io_info->dset->shared->layout.u.chunk.size; + } + + local_info_array[i].skip_filter_pline = FALSE; + if (!filter_partial_edge_chunks) { + /* + * If this is a partial edge chunk and the "don't filter partial edge + * chunks" flag is set, make sure not to apply filters to the chunk. + */ + if (H5D__chunk_is_partial_edge_chunk(io_info->dset->shared->ndims, + io_info->dset->shared->layout.u.chunk.dim, + chunk_info->scaled, io_info->dset->shared->curr_dims)) + local_info_array[i].skip_filter_pline = TRUE; + } + + /* Initialize the chunk's shared info */ + local_info_array[i].chunk_current = udata.chunk_block; + local_info_array[i].chunk_new = udata.chunk_block; + + /* + * Check if the list is not in ascending order of offset in the file + * or has unallocated chunks. In either case, the list should get + * sorted. + */ + if (i) { + haddr_t curr_chunk_offset = local_info_array[i].chunk_current.offset; + haddr_t prev_chunk_offset = local_info_array[i - 1].chunk_current.offset; + + if (!H5F_addr_defined(prev_chunk_offset) || !H5F_addr_defined(curr_chunk_offset) || + (curr_chunk_offset < prev_chunk_offset)) + need_sort = TRUE; + } + + /* + * Extensible arrays may calculate a chunk's index a little differently + * than normal when the dataset's unlimited dimension is not the + * slowest-changing dimension, so set the index here based on what the + * extensible array code calculated instead of what was calculated + * in the chunk file mapping. */ - if ((chunk_npoints = H5S_GET_EXTENT_NPOINTS(chunk_info->fspace)) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid") - local_info_array[i].full_overwrite = - (local_info_array[i].io_size >= (hsize_t)chunk_npoints * type_info->dst_type_size) ? TRUE - : FALSE; + if (io_info->dset->shared->layout.u.chunk.idx_type == H5D_CHUNK_IDX_EARRAY) + local_info_array[i].index_info.chunk_idx = udata.chunk_idx; + else + local_info_array[i].index_info.chunk_idx = chunk_info->index; + + local_info_array[i].index_info.filter_mask = udata.filter_mask; + local_info_array[i].index_info.need_insert = FALSE; chunk_node = H5SL_next(chunk_node); - } /* end for */ - } /* end if */ - - /* Redistribute shared chunks to new owners as necessary */ - if (io_info->op_type == H5D_IO_OP_WRITE) -#if MPI_VERSION >= 3 - if (H5D__chunk_redistribute_shared_chunks(io_info, type_info, fm, local_info_array, - &num_chunks_selected) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to redistribute shared chunks") -#else - HGOTO_ERROR( - H5E_DATASET, H5E_WRITEERROR, FAIL, - "unable to redistribute shared chunks - MPI version < 3 (MPI_Mprobe and MPI_Imrecv missing)") + } + + /* Ensure the chunk list is sorted in ascending order of offset in the file */ + if (need_sort) + HDqsort(local_info_array, num_chunks_selected, sizeof(H5D_filtered_collective_io_info_t), + H5D__cmp_filtered_collective_io_info_entry); + +#ifdef H5Dmpio_DEBUG + H5D__mpio_dump_collective_filtered_chunk_list(local_info_array, num_chunks_selected, mpi_rank); #endif + } + else if (H5F_get_coll_metadata_reads(io_info->dset->oloc.file)) { + hsize_t scaled[H5O_LAYOUT_NDIMS] = {0}; + + /* + * If this rank has no selection in the dataset and collective + * metadata reads are enabled, do a fake lookup of a chunk to + * ensure that this rank has the chunk index opened. Otherwise, + * only the ranks that had a selection will have opened the + * chunk index and they will have done so independently. Therefore, + * when ranks with no selection participate in later collective + * metadata reads, they will try to open the chunk index collectively + * and issues will occur since other ranks won't participate. + * + * In the future, we should consider having a chunk index "open" + * callback that can be used to ensure collectivity between ranks + * in a more natural way, but this hack should suffice for now. + */ + if (H5D__chunk_lookup(io_info->dset, scaled, &udata) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") + } *chunk_list = local_info_array; *num_entries = num_chunks_selected; done: - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5D__construct_filtered_io_info_list() */ +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif -#if MPI_VERSION >= 3 + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_collective_filtered_chunk_io_setup() */ /*------------------------------------------------------------------------- - * Function: H5D__chunk_redistribute_shared_chunks - * - * Purpose: When performing a collective write on a Dataset with - * filters applied, this function is used to redistribute any - * chunks which are selected by more than one process, so as - * to preserve file integrity after the write by ensuring - * that any shared chunks are only modified by one process. - * - * The current implementation follows this 3-phase process: - * - * - Collect everyone's list of chunks into one large list, - * sort the list in increasing order of chunk offset in the - * file and hand the list off to rank 0 - * - * - Rank 0 scans the list looking for matching runs of chunk - * offset in the file (corresponding to a shared chunk which - * has been selected by more than one rank in the I/O - * operation) and for each shared chunk, it redistributes - * the chunk to the process writing to the chunk which - * currently has the least amount of chunks assigned to it - * by modifying the "new_owner" field in each of the list - * entries corresponding to that chunk - * - * - After the chunks have been redistributed, rank 0 re-sorts - * the list in order of previous owner so that each rank - * will get back exactly the array that they contributed to - * the redistribution operation, with the "new_owner" field - * of each chunk they are modifying having possibly been - * modified. Rank 0 then scatters each segment of the list - * back to its corresponding rank + * Function: H5D__mpio_redistribute_shared_chunks + * + * Purpose: When performing a parallel write on a chunked Dataset with + * filters applied, we must ensure that any particular chunk + * is only written to by a single MPI rank in order to avoid + * potential data races on the chunk. This function is used to + * redistribute (by assigning ownership to a single rank) any + * chunks which are selected by more than one MPI rank. + * + * An initial Allgather is performed to determine how many + * chunks each rank has selected in the write operation and + * then that number is compared against a threshold value to + * determine whether chunk redistribution should be done on + * MPI rank 0 only, or on all MPI ranks. * * Return: Non-negative on success/Negative on failure * - * Programmer: Jordan Henderson - * Monday, May 1, 2017 + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_redistribute_shared_chunks(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, const H5D_io_info_t *io_info, + const H5D_chunk_map_t *fm, int mpi_rank, int mpi_size, + size_t **rank_chunks_assigned_map) +{ + hbool_t redistribute_on_all_ranks; + size_t *num_chunks_map = NULL; + size_t coll_chunk_list_size = 0; + size_t i; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(chunk_list || 0 == chunk_list_num_entries); + HDassert(io_info); + HDassert(fm); + HDassert(mpi_size > 1); /* No chunk sharing is possible for MPI Comm size of 1 */ + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Redistribute shared chunks"); +#endif + + /* + * Allocate an array for each rank to keep track of the number of + * chunks assigned to any other rank in order to cut down on future + * MPI communication. + */ + if (NULL == (num_chunks_map = H5MM_malloc((size_t)mpi_size * sizeof(*num_chunks_map)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "couldn't allocate assigned chunks array") + + /* Perform initial Allgather to determine the collective chunk list size */ + if (MPI_SUCCESS != (mpi_code = MPI_Allgather(&chunk_list_num_entries, 1, H5_SIZE_T_AS_MPI_TYPE, + num_chunks_map, 1, H5_SIZE_T_AS_MPI_TYPE, io_info->comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code) + + for (i = 0; i < (size_t)mpi_size; i++) + coll_chunk_list_size += num_chunks_map[i]; + + /* + * Determine whether we should perform chunk redistribution on all + * ranks or just rank 0. For a relatively small number of chunks, + * we redistribute on all ranks to cut down on MPI communication + * overhead. For a larger number of chunks, we redistribute on + * rank 0 only to cut down on memory usage. + */ + redistribute_on_all_ranks = coll_chunk_list_size < H5D_CHUNK_REDISTRIBUTE_THRES; + + if (H5D__mpio_redistribute_shared_chunks_int(chunk_list, num_chunks_map, redistribute_on_all_ranks, + io_info, fm, mpi_rank, mpi_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTREDISTRIBUTE, FAIL, "can't redistribute shared chunks") + + /* + * If the caller provided a pointer for the mapping from + * rank value -> number of chunks assigned, return that + * mapping here. + */ + if (rank_chunks_assigned_map) { + /* + * If we performed chunk redistribution on rank 0 only, distribute + * the rank value -> number of chunks assigned mapping back to all + * ranks. + */ + if (!redistribute_on_all_ranks) { + if (MPI_SUCCESS != + (mpi_code = MPI_Bcast(num_chunks_map, mpi_size, H5_SIZE_T_AS_MPI_TYPE, 0, io_info->comm))) + HMPI_GOTO_ERROR(FAIL, "couldn't broadcast chunk mapping to other ranks", mpi_code) + } + + *rank_chunks_assigned_map = num_chunks_map; + } + +done: + if (!rank_chunks_assigned_map || (ret_value < 0)) { + num_chunks_map = H5MM_xfree(num_chunks_map); + } + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_redistribute_shared_chunks() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_redistribute_shared_chunks_int + * + * Purpose: Routine to perform redistribution of shared chunks during + * parallel writes to datasets with filters applied. + * + * If `all_ranks_involved` is TRUE, chunk redistribution + * occurs on all MPI ranks. This is usually done when there + * is a relatively small number of chunks involved in order to + * cut down on MPI communication overhead while increasing + * total memory usage a bit. + * + * If `all_ranks_involved` is FALSE, only rank 0 will perform + * chunk redistribution. This is usually done when there is + * a relatively large number of chunks involved in order to + * cut down on total memory usage at the cost of increased + * overhead from MPI communication. + * + * This implementation is as follows: + * + * - All MPI ranks send their list of selected chunks to the + * ranks involved in chunk redistribution. Then, the + * involved ranks sort this new list in order of chunk + * index. + * + * - The involved ranks scan the list looking for matching + * runs of chunk index values (corresponding to a shared + * chunk which has been selected by more than one rank in + * the I/O operation) and for each shared chunk, + * redistribute the chunk to the MPI rank writing to the + * chunk which currently has the least amount of chunks + * assigned to it. This is done by modifying the "new_owner" + * field in each of the list entries corresponding to that + * chunk. The involved ranks then re-sort the list in order + * of original chunk owner so that each rank's section of + * contributed chunks is contiguous in the collective chunk + * list. + * + * - If chunk redistribution occurred on all ranks, each rank + * scans through the collective chunk list to find their + * contributed section of chunks and uses that to update + * their local chunk list with the newly-updated "new_owner" + * and "num_writers" fields. If chunk redistribution + * occurred only on rank 0, an MPI_Scatterv operation will + * be used to scatter the segments of the collective chunk + * list from rank 0 back to the corresponding ranks. + * + * Return: Non-negative on success/Negative on failure * *------------------------------------------------------------------------- */ static herr_t -H5D__chunk_redistribute_shared_chunks(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - const H5D_chunk_map_t * fm, - H5D_filtered_collective_io_info_t *local_chunk_array, - size_t * local_chunk_array_num_entries) +H5D__mpio_redistribute_shared_chunks_int(H5D_filtered_collective_io_info_t *chunk_list, + size_t *num_chunks_assigned_map, hbool_t all_ranks_involved, + const H5D_io_info_t *io_info, const H5D_chunk_map_t *fm, + int mpi_rank, int mpi_size) { - H5D_filtered_collective_io_info_t *shared_chunks_info_array = - NULL; /* The list of all chunks selected in the operation by all processes */ - H5S_sel_iter_t *mem_iter = NULL; /* Memory iterator for H5D__gather_mem */ - unsigned char **mod_data = - NULL; /* Array of chunk modification data buffers sent by a process to new chunk owners */ - MPI_Request *send_requests = NULL; /* Array of MPI_Isend chunk modification data send requests */ - MPI_Status * send_statuses = NULL; /* Array of MPI_Isend chunk modification send statuses */ - hbool_t mem_iter_init = FALSE; - size_t shared_chunks_info_array_num_entries = 0; - size_t num_send_requests = 0; - size_t * num_assigned_chunks_array = NULL; - size_t i, last_assigned_idx; - int * send_counts = NULL; - int * send_displacements = NULL; - int scatter_recvcount_int; - int mpi_rank, mpi_size, mpi_code; + MPI_Datatype struct_type; + MPI_Datatype packed_type; + hbool_t struct_type_derived = FALSE; + hbool_t packed_type_derived = FALSE; + size_t i; + size_t coll_chunk_list_num_entries = 0; + void * coll_chunk_list = NULL; + int * counts_disps_array = NULL; + int * counts_ptr = NULL; + int * displacements_ptr = NULL; + int num_chunks_int; + int mpi_code; herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC + HDassert(num_chunks_assigned_map); + HDassert(chunk_list || 0 == num_chunks_assigned_map[mpi_rank]); HDassert(io_info); - HDassert(type_info); HDassert(fm); - HDassert(local_chunk_array_num_entries); + HDassert(mpi_size > 1); - if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank") - if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0) - HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size") +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Redistribute shared chunks (internal)"); +#endif - /* Set to latest format for encoding dataspace */ - H5CX_set_libver_bounds(NULL); + /* + * Make sure it's safe to cast this rank's number + * of chunks to be sent into an int for MPI + */ + H5_CHECKED_ASSIGN(num_chunks_int, int, num_chunks_assigned_map[mpi_rank], size_t); - if (*local_chunk_array_num_entries) - if (NULL == (send_requests = - (MPI_Request *)H5MM_malloc(*local_chunk_array_num_entries * sizeof(MPI_Request)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate send requests buffer") + /* + * Phase 1 - Participate in collective gathering of every rank's + * list of chunks to the ranks which are performing the redistribution + * operation. + */ - if (NULL == (mem_iter = (H5S_sel_iter_t *)H5MM_malloc(sizeof(H5S_sel_iter_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate memory iterator") + if (all_ranks_involved || (mpi_rank == 0)) { + /* + * Allocate array to store the receive counts of each rank, as well as + * the displacements into the final array where each rank will place + * their data. The first half of the array contains the receive counts + * (in rank order), while the latter half contains the displacements + * (also in rank order). + */ + if (NULL == (counts_disps_array = H5MM_malloc(2 * (size_t)mpi_size * sizeof(*counts_disps_array)))) { + /* Push an error, but still participate in collective gather operation */ + HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "couldn't allocate receive counts and displacements array") + } + else { + /* Set the receive counts from the assigned chunks map */ + counts_ptr = counts_disps_array; + + for (i = 0; i < (size_t)mpi_size; i++) + H5_CHECKED_ASSIGN(counts_ptr[i], int, num_chunks_assigned_map[i], size_t); + + /* Set the displacements into the receive buffer for the gather operation */ + displacements_ptr = &counts_disps_array[mpi_size]; + + *displacements_ptr = 0; + for (i = 1; i < (size_t)mpi_size; i++) + displacements_ptr[i] = displacements_ptr[i - 1] + counts_ptr[i - 1]; + } + } - /* Gather every rank's list of chunks to rank 0 to allow it to perform the redistribution operation. After - * this call, the gathered list will initially be sorted in increasing order of chunk offset in the file. + /* + * Construct MPI derived types for extracting information + * necessary for MPI communication */ - if (H5D__mpio_array_gatherv(local_chunk_array, *local_chunk_array_num_entries, - sizeof(H5D_filtered_collective_io_info_t), (void **)&shared_chunks_info_array, - &shared_chunks_info_array_num_entries, false, 0, io_info->comm, - H5D__cmp_filtered_collective_io_info_entry) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "couldn't gather array") + if (H5D__mpio_get_chunk_redistribute_info_types(&packed_type, &packed_type_derived, &struct_type, + &struct_type_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "can't create derived datatypes for chunk redistribution info") + + /* Perform gather operation */ + if (H5_mpio_gatherv_alloc(chunk_list, num_chunks_int, struct_type, counts_ptr, displacements_ptr, + packed_type, all_ranks_involved, 0, io_info->comm, mpi_rank, mpi_size, + &coll_chunk_list, &coll_chunk_list_num_entries) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, + "can't gather chunk redistribution info to involved ranks") - /* Rank 0 redistributes any shared chunks to new owners as necessary */ - if (mpi_rank == 0) { - if (NULL == (send_counts = (int *)H5MM_calloc((size_t)mpi_size * sizeof(int)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate send counts buffer") + /* + * If all ranks are redistributing shared chunks, we no + * longer need the receive counts and displacements array + */ + if (all_ranks_involved) { + counts_disps_array = H5MM_xfree(counts_disps_array); + } - if (NULL == (send_displacements = (int *)H5MM_malloc((size_t)mpi_size * sizeof(int)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate send displacements buffer") + /* + * Phase 2 - Involved ranks now redistribute any shared chunks to new + * owners as necessary. + */ - if (NULL == (num_assigned_chunks_array = (size_t *)H5MM_calloc((size_t)mpi_size * sizeof(size_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, - "unable to allocate number of assigned chunks array") + if (all_ranks_involved || (mpi_rank == 0)) { + H5D_chunk_redistribute_info_t *chunk_entry; + hsize_t curr_chunk_idx; + size_t set_begin_index; + int num_writers; + int new_chunk_owner; - for (i = 0; i < shared_chunks_info_array_num_entries;) { - H5D_filtered_collective_io_info_t *chunk_entry; - haddr_t last_seen_addr = shared_chunks_info_array[i].chunk_states.chunk_current.offset; - size_t set_begin_index = i; - size_t num_writers = 0; - int new_chunk_owner = shared_chunks_info_array[i].owners.original_owner; + /* Clear the mapping from rank value -> number of assigned chunks */ + HDmemset(num_chunks_assigned_map, 0, (size_t)mpi_size * sizeof(*num_chunks_assigned_map)); - /* Process each set of duplicate entries caused by another process writing to the same chunk */ - do { - chunk_entry = &shared_chunks_info_array[i]; + /* Sort collective chunk list according to chunk index */ + HDqsort(coll_chunk_list, coll_chunk_list_num_entries, sizeof(H5D_chunk_redistribute_info_t), + H5D__cmp_chunk_redistribute_info); - send_counts[chunk_entry->owners.original_owner] += (int)sizeof(*chunk_entry); + /* + * Process all chunks in the collective chunk list. + * Note that the loop counter is incremented by both + * the outer loop (while processing each entry in + * the collective chunk list) and the inner loop + * (while processing duplicate entries for shared + * chunks). + */ + chunk_entry = &((H5D_chunk_redistribute_info_t *)coll_chunk_list)[0]; + for (i = 0; i < coll_chunk_list_num_entries;) { + /* Set chunk's initial new owner to its original owner */ + new_chunk_owner = chunk_entry->orig_owner; + + /* + * Set the current chunk index so we know when we've processed + * all duplicate entries for a particular shared chunk + */ + curr_chunk_idx = chunk_entry->chunk_idx; + + /* Reset the initial number of writers to this chunk */ + num_writers = 0; + + /* Set index for the beginning of this section of duplicate chunk entries */ + set_begin_index = i; - /* The new owner of the chunk is determined by the process + /* + * Process each chunk entry in the set for the current + * (possibly shared) chunk and increment the loop counter + * while doing so. + */ + do { + /* + * The new owner of the chunk is determined by the rank * writing to the chunk which currently has the least amount * of chunks assigned to it */ - if (num_assigned_chunks_array[chunk_entry->owners.original_owner] < - num_assigned_chunks_array[new_chunk_owner]) - new_chunk_owner = chunk_entry->owners.original_owner; + if (num_chunks_assigned_map[chunk_entry->orig_owner] < + num_chunks_assigned_map[new_chunk_owner]) + new_chunk_owner = chunk_entry->orig_owner; + /* Update the number of writers to this particular chunk */ num_writers++; - } while (++i < shared_chunks_info_array_num_entries && - shared_chunks_info_array[i].chunk_states.chunk_current.offset == last_seen_addr); - /* Set all of the chunk entries' "new_owner" fields */ + chunk_entry++; + } while (++i < coll_chunk_list_num_entries && chunk_entry->chunk_idx == curr_chunk_idx); + + /* We should never have more writers to a chunk than the number of MPI ranks */ + HDassert(num_writers <= mpi_size); + + /* Set all processed chunk entries' "new_owner" and "num_writers" fields */ for (; set_begin_index < i; set_begin_index++) { - shared_chunks_info_array[set_begin_index].owners.new_owner = new_chunk_owner; - shared_chunks_info_array[set_begin_index].num_writers = num_writers; - } /* end for */ + H5D_chunk_redistribute_info_t *entry; - num_assigned_chunks_array[new_chunk_owner]++; - } /* end for */ + entry = &((H5D_chunk_redistribute_info_t *)coll_chunk_list)[set_begin_index]; - /* Sort the new list in order of previous owner so that each original owner of a chunk - * entry gets that entry back, with the possibly newly-modified "new_owner" field + entry->new_owner = new_chunk_owner; + entry->num_writers = num_writers; + } + + /* Update the number of chunks assigned to the MPI rank that now owns this chunk */ + num_chunks_assigned_map[new_chunk_owner]++; + } + + /* + * Re-sort the collective chunk list in order of original chunk owner + * so that each rank's section of contributed chunks is contiguous in + * the collective chunk list. + * + * NOTE: this re-sort is frail in that it needs to sort the collective + * chunk list so that each rank's section of contributed chunks + * is in the exact order it was contributed in, or things will + * be scrambled when each rank's local chunk list is updated. + * Therefore, the sorting algorithm here is tied to the one + * used during the I/O setup operation. Specifically, chunks + * are first sorted by ascending order of offset in the file and + * then by chunk index. In the future, a better redistribution + * algorithm may be devised that doesn't rely on frail sorting, + * but the current implementation is a quick and naive approach. */ - if (shared_chunks_info_array_num_entries > 1) - HDqsort(shared_chunks_info_array, shared_chunks_info_array_num_entries, - sizeof(H5D_filtered_collective_io_info_t), - H5D__cmp_filtered_collective_io_info_entry_owner); - - send_displacements[0] = 0; - for (i = 1; i < (size_t)mpi_size; i++) - send_displacements[i] = send_displacements[i - 1] + send_counts[i - 1]; - } /* end if */ + HDqsort(coll_chunk_list, coll_chunk_list_num_entries, sizeof(H5D_chunk_redistribute_info_t), + H5D__cmp_chunk_redistribute_info_orig_owner); + } - /* Scatter the segments of the list back to each process */ - H5_CHECKED_ASSIGN(scatter_recvcount_int, int, - *local_chunk_array_num_entries * sizeof(H5D_filtered_collective_io_info_t), size_t); - if (MPI_SUCCESS != - (mpi_code = MPI_Scatterv(shared_chunks_info_array, send_counts, send_displacements, MPI_BYTE, - local_chunk_array, scatter_recvcount_int, MPI_BYTE, 0, io_info->comm))) - HMPI_GOTO_ERROR(FAIL, "unable to scatter shared chunks info buffer", mpi_code) + if (all_ranks_involved) { + /* + * If redistribution occurred on all ranks, search for the section + * in the collective chunk list corresponding to this rank's locally + * selected chunks and update the local list after redistribution. + */ + for (i = 0; i < coll_chunk_list_num_entries; i++) + if (mpi_rank == ((H5D_chunk_redistribute_info_t *)coll_chunk_list)[i].orig_owner) + break; - if (shared_chunks_info_array) { - H5MM_free(shared_chunks_info_array); - shared_chunks_info_array = NULL; - } /* end if */ + for (size_t j = 0; j < (size_t)num_chunks_int; j++) { + H5D_chunk_redistribute_info_t *coll_entry; - /* Now that the chunks have been redistributed, each process must send its modification data - * to the new owners of any of the chunks it previously possessed. Accordingly, each process - * must also issue asynchronous receives for any messages it may receive for each of the - * chunks it is assigned, in order to avoid potential deadlocking issues. + coll_entry = &((H5D_chunk_redistribute_info_t *)coll_chunk_list)[i++]; + + chunk_list[j].new_owner = coll_entry->new_owner; + chunk_list[j].num_writers = coll_entry->num_writers; + } + } + else { + /* + * If redistribution occurred only on rank 0, scatter the segments + * of the collective chunk list back to each rank so that their + * local chunk lists get updated + */ + if (MPI_SUCCESS != + (mpi_code = MPI_Scatterv(coll_chunk_list, counts_ptr, displacements_ptr, packed_type, chunk_list, + num_chunks_int, struct_type, 0, io_info->comm))) + HMPI_GOTO_ERROR(FAIL, "unable to scatter shared chunks info buffer", mpi_code) + } + +#ifdef H5Dmpio_DEBUG + H5D__mpio_dump_collective_filtered_chunk_list(chunk_list, num_chunks_assigned_map[mpi_rank], mpi_rank); +#endif + +done: + H5MM_free(coll_chunk_list); + + if (struct_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + if (packed_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&packed_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + + H5MM_free(counts_disps_array); + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_redistribute_shared_chunks_int() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_share_chunk_modification_data + * + * Purpose: When performing a parallel write on a chunked dataset with + * filters applied, we must first ensure that any particular + * chunk is only written to by a single MPI rank in order to + * avoid potential data races on the chunk. Once dataset + * chunks have been redistributed in a suitable manner, each + * MPI rank must send its chunk data to other ranks for each + * chunk it no longer owns. + * + * The current implementation here follows the Nonblocking + * Consensus algorithm described in: + * http://unixer.de/publications/img/hoefler-dsde-protocols.pdf + * + * First, each MPI rank scans through its list of selected + * chunks and does the following for each chunk: + * + * * If a chunk in the MPI rank's chunk list is still owned + * by that rank, the rank checks how many messages are + * incoming for that chunk and adds that to its running + * total. Then, the rank updates its local chunk list so + * that any previous chunk entries for chunks that are no + * longer owned by the rank get overwritten by chunk + * entries for chunks the rank still owns. Since the data + * for the chunks no longer owned will have already been + * sent, those chunks can effectively be discarded. + * * If a chunk in the MPI rank's chunk list is no longer + * owned by that rank, the rank sends the data it wishes to + * update the chunk with to the MPI rank that now has + * ownership of that chunk. To do this, it encodes the + * chunk's index, its selection in the chunk and its + * modification data into a buffer and then posts a + * non-blocking MPI_Issend to the owning rank. + * + * Once this step is complete, all MPI ranks allocate arrays + * to hold chunk message receive buffers and MPI request + * objects for each non-blocking receive they will post for + * incoming chunk modification messages. Then, all MPI ranks + * enter a loop that alternates between non-blocking + * MPI_Iprobe calls to probe for incoming messages and + * MPI_Testall calls to see if all send requests have + * completed. As chunk modification messages arrive, + * non-blocking MPI_Irecv calls will be posted for each + * message. + * + * Once all send requests have completed, an MPI_Ibarrier is + * posted and the loop then alternates between MPI_Iprobe + * calls and MPI_Test calls to check if all ranks have reached + * the non-blocking barrier. Once all ranks have reached the + * barrier, processing can move on to updating the selected + * chunks that are owned in the operation. + * + * Any chunk messages that were received from other ranks + * will be returned through the `chunk_msg_bufs` array and + * `chunk_msg_bufs_len` will be set appropriately. + * + * NOTE: The use of non-blocking sends and receives of chunk + * data here may contribute to large amounts of memory + * usage/MPI request overhead if the number of shared + * chunks is high. If this becomes a problem, it may be + * useful to split the message receiving loop away so + * that chunk modification messages can be received and + * processed immediately (MPI_Recv) using a single chunk + * message buffer. However, it's possible this may + * degrade performance since the chunk message sends + * are synchronous (MPI_Issend) in the Nonblocking + * Consensus algorithm. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_share_chunk_modification_data(H5D_filtered_collective_io_info_t *chunk_list, + size_t *chunk_list_num_entries, H5D_io_info_t *io_info, + const H5D_type_info_t *type_info, int mpi_rank, int mpi_size, + H5D_filtered_collective_io_info_t **chunk_hash_table, + unsigned char ***chunk_msg_bufs, int *chunk_msg_bufs_len) +{ +#if H5_CHECK_MPI_VERSION(3, 0) + H5D_filtered_collective_io_info_t *chunk_table = NULL; + H5S_sel_iter_t * mem_iter = NULL; + unsigned char ** msg_send_bufs = NULL; + unsigned char ** msg_recv_bufs = NULL; + MPI_Request * send_requests = NULL; + MPI_Request * recv_requests = NULL; + MPI_Request ibarrier = MPI_REQUEST_NULL; + hbool_t mem_iter_init = FALSE; + hbool_t ibarrier_posted = FALSE; + size_t send_bufs_nalloc = 0; + size_t num_send_requests = 0; + size_t num_recv_requests = 0; + size_t num_msgs_incoming = 0; + size_t last_assigned_idx; + size_t i; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(chunk_list_num_entries); + HDassert(chunk_list || 0 == *chunk_list_num_entries); + HDassert(io_info); + HDassert(type_info); + HDassert(mpi_size > 1); + HDassert(chunk_msg_bufs); + HDassert(chunk_msg_bufs_len); + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Share chunk modification data"); +#endif + + /* Set to latest format for encoding dataspace */ + H5CX_set_libver_bounds(NULL); + + if (*chunk_list_num_entries) { + /* Allocate a selection iterator for iterating over chunk dataspaces */ + if (NULL == (mem_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate dataspace selection iterator") + + /* + * Allocate send buffer and MPI_Request arrays for non-blocking + * sends of outgoing chunk messages + */ + send_bufs_nalloc = H5D_CHUNK_NUM_SEND_MSGS_INIT; + if (NULL == (msg_send_bufs = H5MM_malloc(send_bufs_nalloc * sizeof(*msg_send_bufs)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, + "couldn't allocate chunk modification message buffer array") + + if (NULL == (send_requests = H5MM_malloc(send_bufs_nalloc * sizeof(*send_requests)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate send requests array") + } + + /* + * For each chunk this rank owns, add to the total number of + * incoming MPI messages, then update the local chunk list to + * overwrite any previous chunks no longer owned by this rank. + * Since the data for those chunks will have already been sent, + * this rank should no longer be interested in them and they + * can effectively be discarded. This bookkeeping also makes + * the code for the collective file space re-allocation and + * chunk re-insertion operations a bit simpler. + * + * For each chunk this rank doesn't own, use non-blocking + * synchronous sends to send the data this rank is writing to + * the rank that does own the chunk. */ - if (*local_chunk_array_num_entries) - if (NULL == (mod_data = (unsigned char **)H5MM_malloc(*local_chunk_array_num_entries * - sizeof(unsigned char *)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate modification data buffer array") - - /* Perform all the sends on the chunks that this rank doesn't own */ - /* (Sends and recvs must be two separate loops, to avoid deadlock) */ - for (i = 0, last_assigned_idx = 0; i < *local_chunk_array_num_entries; i++) { - H5D_filtered_collective_io_info_t *chunk_entry = &local_chunk_array[i]; - - if (mpi_rank != chunk_entry->owners.new_owner) { - H5D_chunk_info_t *chunk_info = NULL; + for (i = 0, last_assigned_idx = 0; i < *chunk_list_num_entries; i++) { + H5D_filtered_collective_io_info_t *chunk_entry = &chunk_list[i]; + + if (mpi_rank == chunk_entry->new_owner) { + num_msgs_incoming += (size_t)(chunk_entry->num_writers - 1); + + /* + * Overwrite chunk entries this rank doesn't own with entries that it + * does own, since it has sent the necessary data and is no longer + * interested in the chunks it doesn't own. + */ + chunk_list[last_assigned_idx] = chunk_list[i]; + + /* + * Since, at large scale, a chunk's index value may be larger than + * the maximum value that can be stored in an int, we cannot rely + * on using a chunk's index value as the tag for the MPI messages + * sent/received for a chunk. Therefore, add this chunk to a hash + * table with the chunk's index as a key so that we can quickly find + * the chunk when processing chunk messages that were received. The + * message itself will contain the chunk's index so we can update + * the correct chunk with the received data. + */ + HASH_ADD(hh, chunk_table, index_info.chunk_idx, sizeof(hsize_t), &chunk_list[last_assigned_idx]); + + last_assigned_idx++; + } + else { + H5D_chunk_info_t *chunk_info = chunk_entry->chunk_info; unsigned char * mod_data_p = NULL; hsize_t iter_nelmts; - size_t mod_data_size; + size_t mod_data_size = 0; + size_t space_size = 0; - /* Look up the chunk and get its file and memory dataspaces */ - if (NULL == (chunk_info = (H5D_chunk_info_t *)H5SL_search(fm->sel_chunks, &chunk_entry->index))) - HGOTO_ERROR(H5E_DATASPACE, H5E_NOTFOUND, FAIL, "can't locate chunk in skip list") + /* Add the size of the chunk index to the encoded size */ + mod_data_size += sizeof(hsize_t); - /* Determine size of serialized chunk file dataspace, plus the size of - * the data being written - */ - if (H5S_encode(chunk_info->fspace, &mod_data_p, &mod_data_size) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTENCODE, FAIL, "unable to get encoded dataspace size") + /* Determine size of serialized chunk file dataspace */ + if (H5S_encode(chunk_info->fspace, &mod_data_p, &space_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "unable to get encoded dataspace size") + mod_data_size += space_size; + /* Determine size of data being written */ iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace); - H5_CHECK_OVERFLOW(iter_nelmts, hsize_t, size_t); + mod_data_size += (size_t)iter_nelmts * type_info->src_type_size; - if (NULL == (mod_data[num_send_requests] = (unsigned char *)H5MM_malloc(mod_data_size))) + if (NULL == (msg_send_bufs[num_send_requests] = H5MM_malloc(mod_data_size))) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, - "couldn't allocate chunk modification send buffer") + "couldn't allocate chunk modification message buffer") + + mod_data_p = msg_send_bufs[num_send_requests]; + + /* Store the chunk's index into the buffer */ + HDmemcpy(mod_data_p, &chunk_entry->index_info.chunk_idx, sizeof(hsize_t)); + mod_data_p += sizeof(hsize_t); /* Serialize the chunk's file dataspace into the buffer */ - mod_data_p = mod_data[num_send_requests]; if (H5S_encode(chunk_info->fspace, &mod_data_p, &mod_data_size) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTENCODE, FAIL, "unable to encode dataspace") /* Initialize iterator for memory selection */ - if (H5S_select_iter_init(mem_iter, chunk_info->mspace, type_info->src_type_size, 0) < 0) + if (H5S_select_iter_init(mem_iter, chunk_info->mspace, type_info->src_type_size, + H5S_SEL_ITER_SHARE_WITH_DATASPACE) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize memory selection information") mem_iter_init = TRUE; @@ -2952,466 +3943,2057 @@ H5D__chunk_redistribute_shared_chunks(const H5D_io_info_t *io_info, const H5D_ty if (0 == H5D__gather_mem(io_info->u.wbuf, mem_iter, (size_t)iter_nelmts, mod_data_p)) HGOTO_ERROR(H5E_IO, H5E_CANTGATHER, FAIL, "couldn't gather from write buffer") - /* Send modification data to new owner */ + /* + * Ensure that the size of the chunk data being sent can be + * safely cast to an int for MPI. Note that this should + * generally be OK for now (unless a rank is sending a + * whole 32-bit-sized chunk of data + its encoded selection), + * but if we allow larger than 32-bit-sized chunks in the + * future, this may become a problem and derived datatypes + * will need to be used. + */ H5_CHECK_OVERFLOW(mod_data_size, size_t, int) - H5_CHECK_OVERFLOW(chunk_entry->index, hsize_t, int) + + /* Send modification data to new owner */ if (MPI_SUCCESS != - (mpi_code = MPI_Isend(mod_data[num_send_requests], (int)mod_data_size, MPI_BYTE, - chunk_entry->owners.new_owner, (int)chunk_entry->index, io_info->comm, - &send_requests[num_send_requests]))) - HMPI_GOTO_ERROR(FAIL, "MPI_Isend failed", mpi_code) + (mpi_code = MPI_Issend(msg_send_bufs[num_send_requests], (int)mod_data_size, MPI_BYTE, + chunk_entry->new_owner, H5D_CHUNK_MOD_DATA_TAG, io_info->comm, + &send_requests[num_send_requests]))) + HMPI_GOTO_ERROR(FAIL, "MPI_Issend failed", mpi_code) + + num_send_requests++; + + /* Resize send buffer and send request arrays if necessary */ + if (num_send_requests == send_bufs_nalloc) { + void *tmp_alloc; + + send_bufs_nalloc = (size_t)((double)send_bufs_nalloc * 1.5); + + if (NULL == + (tmp_alloc = H5MM_realloc(msg_send_bufs, send_bufs_nalloc * sizeof(*msg_send_bufs)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, + "couldn't resize chunk modification message buffer array") + msg_send_bufs = tmp_alloc; + + if (NULL == + (tmp_alloc = H5MM_realloc(send_requests, send_bufs_nalloc * sizeof(*send_requests)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't resize send requests array") + send_requests = tmp_alloc; + } - if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + if (H5S_SELECT_ITER_RELEASE(mem_iter) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release memory selection iterator") mem_iter_init = FALSE; + } + } - num_send_requests++; - } /* end if */ - } /* end for */ + /* Check if the number of send or receive requests will overflow an int (MPI requirement) */ + if (num_send_requests > INT_MAX || num_msgs_incoming > INT_MAX) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, + "too many shared chunks in parallel filtered write operation") + + H5_CHECK_OVERFLOW(num_send_requests, size_t, int) + H5_CHECK_OVERFLOW(num_msgs_incoming, size_t, int) + + /* + * Allocate receive buffer and MPI_Request arrays for non-blocking + * receives of incoming chunk messages + */ + if (num_msgs_incoming) { + if (NULL == (msg_recv_bufs = H5MM_malloc(num_msgs_incoming * sizeof(*msg_recv_bufs)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, + "couldn't allocate chunk modification message buffer array") - /* Perform all the recvs on the chunks this rank owns */ - for (i = 0, last_assigned_idx = 0; i < *local_chunk_array_num_entries; i++) { - H5D_filtered_collective_io_info_t *chunk_entry = &local_chunk_array[i]; + if (NULL == (recv_requests = H5MM_malloc(num_msgs_incoming * sizeof(*recv_requests)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate receive requests array") + } - if (mpi_rank == chunk_entry->owners.new_owner) { - /* Allocate all necessary buffers for an asynchronous receive operation */ - if (chunk_entry->num_writers > 1) { - MPI_Message message; - MPI_Status status; - size_t j; + /* Process any incoming messages until everyone is done */ + do { + MPI_Status status; + int msg_flag; - chunk_entry->async_info.num_receive_requests = (int)chunk_entry->num_writers - 1; - if (NULL == (chunk_entry->async_info.receive_requests_array = (MPI_Request *)H5MM_malloc( - (size_t)chunk_entry->async_info.num_receive_requests * sizeof(MPI_Request)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate async requests array") + /* Probe for an incoming message from any rank */ + if (MPI_SUCCESS != (mpi_code = MPI_Iprobe(MPI_ANY_SOURCE, H5D_CHUNK_MOD_DATA_TAG, io_info->comm, + &msg_flag, &status))) + HMPI_GOTO_ERROR(FAIL, "MPI_Iprobe failed", mpi_code) - if (NULL == - (chunk_entry->async_info.receive_buffer_array = (unsigned char **)H5MM_malloc( - (size_t)chunk_entry->async_info.num_receive_requests * sizeof(unsigned char *)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate async receive buffers") + /* + * If a message was found, allocate a buffer for the message and + * post a non-blocking receive to receive it + */ + if (msg_flag) { +#if H5_CHECK_MPI_VERSION(3, 0) + MPI_Count msg_size = 0; - for (j = 0; j < chunk_entry->num_writers - 1; j++) { - int count = 0; + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&status, MPI_BYTE, &msg_size))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements_x failed", mpi_code) - /* Probe for a particular message from any process, removing that message - * from the receive queue in the process and allocating that much memory - * for the asynchronous receive - */ - if (MPI_SUCCESS != (mpi_code = MPI_Mprobe(MPI_ANY_SOURCE, (int)chunk_entry->index, - io_info->comm, &message, &status))) - HMPI_GOTO_ERROR(FAIL, "MPI_Mprobe failed", mpi_code) - - if (MPI_SUCCESS != (mpi_code = MPI_Get_count(&status, MPI_BYTE, &count))) - HMPI_GOTO_ERROR(FAIL, "MPI_Get_count failed", mpi_code) - - HDassert(count >= 0); - if (NULL == (chunk_entry->async_info.receive_buffer_array[j] = - (unsigned char *)H5MM_malloc((size_t)count * sizeof(char *)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, - "unable to allocate modification data receive buffer") - - if (MPI_SUCCESS != (mpi_code = MPI_Imrecv( - chunk_entry->async_info.receive_buffer_array[j], count, MPI_BYTE, - &message, &chunk_entry->async_info.receive_requests_array[j]))) - HMPI_GOTO_ERROR(FAIL, "MPI_Imrecv failed", mpi_code) - } /* end for */ - } /* end if */ - - local_chunk_array[last_assigned_idx++] = local_chunk_array[i]; - } /* end else */ - } /* end for */ + H5_CHECK_OVERFLOW(msg_size, MPI_Count, int) +#else + int msg_size = 0; - *local_chunk_array_num_entries = last_assigned_idx; + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&status, MPI_BYTE, &msg_size))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) +#endif - /* Wait for all async send requests to complete before returning */ - if (num_send_requests) { - if (NULL == (send_statuses = (MPI_Status *)H5MM_malloc(num_send_requests * sizeof(MPI_Status)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate send statuses buffer") + if (msg_size <= 0) + HGOTO_ERROR(H5E_DATASET, H5E_BADVALUE, FAIL, "invalid chunk modification message size") - H5_CHECK_OVERFLOW(num_send_requests, size_t, int); - if (MPI_SUCCESS != (mpi_code = MPI_Waitall((int)num_send_requests, send_requests, send_statuses))) - HMPI_GOTO_ERROR(FAIL, "MPI_Waitall failed", mpi_code) - } /* end if */ + HDassert((num_recv_requests + 1) <= num_msgs_incoming); + if (NULL == + (msg_recv_bufs[num_recv_requests] = H5MM_malloc((size_t)msg_size * sizeof(unsigned char)))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, + "couldn't allocate chunk modification message receive buffer") -done: - /* Now that all async send requests have completed, free up the send - * buffers used in the async operations + if (MPI_SUCCESS != (mpi_code = MPI_Irecv(msg_recv_bufs[num_recv_requests], (int)msg_size, + MPI_BYTE, status.MPI_SOURCE, H5D_CHUNK_MOD_DATA_TAG, + io_info->comm, &recv_requests[num_recv_requests]))) + HMPI_GOTO_ERROR(FAIL, "MPI_Irecv failed", mpi_code) + + num_recv_requests++; + } + + if (ibarrier_posted) { + int ibarrier_completed; + + if (MPI_SUCCESS != (mpi_code = MPI_Test(&ibarrier, &ibarrier_completed, MPI_STATUS_IGNORE))) + HMPI_GOTO_ERROR(FAIL, "MPI_Test failed", mpi_code) + + if (ibarrier_completed) + break; + } + else { + int all_sends_completed; + + /* Determine if all send requests have completed */ + if (MPI_SUCCESS != (mpi_code = MPI_Testall((int)num_send_requests, send_requests, + &all_sends_completed, MPI_STATUSES_IGNORE))) + HMPI_GOTO_ERROR(FAIL, "MPI_Testall failed", mpi_code) + + if (all_sends_completed) { + /* Post non-blocking barrier */ + if (MPI_SUCCESS != (mpi_code = MPI_Ibarrier(io_info->comm, &ibarrier))) + HMPI_GOTO_ERROR(FAIL, "MPI_Ibarrier failed", mpi_code) + ibarrier_posted = TRUE; + + /* + * Now that all send requests have completed, free up the + * send buffers used in the non-blocking operations + */ + if (msg_send_bufs) { + for (i = 0; i < num_send_requests; i++) { + if (msg_send_bufs[i]) + H5MM_free(msg_send_bufs[i]); + } + + msg_send_bufs = H5MM_xfree(msg_send_bufs); + } + } + } + } while (1); + + /* + * Ensure all receive requests have completed before moving on. + * For linked-chunk I/O, more overlap with computation could + * theoretically be achieved by returning the receive requests + * array and postponing this wait until during chunk updating + * when the data is really needed. However, multi-chunk I/O + * only updates a chunk at a time and the messages may not come + * in the order that chunks are processed. So, the safest way to + * support both I/O modes is to simply make sure all messages + * are available. */ - for (i = 0; i < num_send_requests; i++) { - if (mod_data[i]) - H5MM_free(mod_data[i]); - } /* end for */ + if (MPI_SUCCESS != (mpi_code = MPI_Waitall((int)num_recv_requests, recv_requests, MPI_STATUSES_IGNORE))) + HMPI_GOTO_ERROR(FAIL, "MPI_Waitall failed", mpi_code) + + /* Set the new number of locally-selected chunks */ + *chunk_list_num_entries = last_assigned_idx; + + /* Return chunk message buffers if any were received */ + *chunk_hash_table = chunk_table; + *chunk_msg_bufs = msg_recv_bufs; + *chunk_msg_bufs_len = (int)num_recv_requests; +done: + if (ret_value < 0) { + /* If this rank failed, make sure to participate in collective barrier */ + if (!ibarrier_posted) { + if (MPI_SUCCESS != (mpi_code = MPI_Ibarrier(io_info->comm, &ibarrier))) + HMPI_GOTO_ERROR(FAIL, "MPI_Ibarrier failed", mpi_code) + } + + if (num_send_requests) { + for (i = 0; i < num_send_requests; i++) { + MPI_Cancel(&send_requests[i]); + } + } + + if (recv_requests) { + for (i = 0; i < num_recv_requests; i++) { + MPI_Cancel(&recv_requests[i]); + } + } + + if (msg_recv_bufs) { + for (i = 0; i < num_recv_requests; i++) { + H5MM_free(msg_recv_bufs[i]); + } + + H5MM_free(msg_recv_bufs); + } + + HASH_CLEAR(hh, chunk_table); + } + + if (recv_requests) + H5MM_free(recv_requests); if (send_requests) H5MM_free(send_requests); - if (send_statuses) - H5MM_free(send_statuses); - if (send_counts) - H5MM_free(send_counts); - if (send_displacements) - H5MM_free(send_displacements); - if (mod_data) - H5MM_free(mod_data); - if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0) - HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator") - if (mem_iter) - H5MM_free(mem_iter); - if (num_assigned_chunks_array) - H5MM_free(num_assigned_chunks_array); - if (shared_chunks_info_array) - H5MM_free(shared_chunks_info_array); + + if (msg_send_bufs) { + for (i = 0; i < num_send_requests; i++) { + if (msg_send_bufs[i]) + H5MM_free(msg_send_bufs[i]); + } + + H5MM_free(msg_send_bufs); + } + + if (mem_iter) { + if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release dataspace selection iterator") + mem_iter = H5FL_FREE(H5S_sel_iter_t, mem_iter); + } + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif FUNC_LEAVE_NOAPI(ret_value) -} /* end H5D__chunk_redistribute_shared_chunks() */ +#else + FUNC_ENTER_STATIC + HERROR( + H5E_DATASET, H5E_WRITEERROR, + "unable to send chunk modification data between MPI ranks - MPI version < 3 (MPI_Ibarrier missing)") + FUNC_LEAVE_NOAPI(FAIL) #endif +} /* end H5D__mpio_share_chunk_modification_data() */ /*------------------------------------------------------------------------- - * Function: H5D__mpio_filtered_collective_write_type + * Function: H5D__mpio_collective_filtered_chunk_common_io * - * Purpose: Constructs a MPI derived datatype for both the memory and - * the file for a collective write of filtered chunks. The - * datatype contains the offsets in the file and the locations - * of the filtered chunk data buffers. + * Purpose: This routine performs the common part of collective I/O + * when reading or writing filtered chunks collectively. * * Return: Non-negative on success/Negative on failure * - * Programmer: Jordan Henderson - * Tuesday, November 22, 2016 - * *------------------------------------------------------------------------- */ static herr_t -H5D__mpio_filtered_collective_write_type(H5D_filtered_collective_io_info_t *chunk_list, size_t num_entries, - MPI_Datatype *new_mem_type, hbool_t *mem_type_derived, - MPI_Datatype *new_file_type, hbool_t *file_type_derived) +H5D__mpio_collective_filtered_chunk_common_io(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, const H5D_io_info_t *io_info, + const H5D_type_info_t *type_info, int mpi_size) { - MPI_Aint *write_buf_array = NULL; /* Relative displacements of filtered chunk data buffers */ - MPI_Aint *file_offset_array = NULL; /* Chunk offsets in the file */ - int * length_array = NULL; /* Filtered Chunk lengths */ - herr_t ret_value = SUCCEED; + H5D_io_info_t coll_io_info; + H5D_storage_t ctg_store; + MPI_Datatype file_type = MPI_DATATYPE_NULL; + MPI_Datatype mem_type = MPI_DATATYPE_NULL; + hbool_t mem_type_is_derived = FALSE; + hbool_t file_type_is_derived = FALSE; + hsize_t mpi_buf_count; + haddr_t base_read_offset = HADDR_UNDEF; + size_t num_chunks; + size_t i; + char fake_buf; /* Used as a fake buffer for ranks with no chunks, thus a NULL buf pointer */ + int mpi_code; + herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC - HDassert(chunk_list); - HDassert(new_mem_type); - HDassert(mem_type_derived); - HDassert(new_file_type); - HDassert(file_type_derived); + HDassert(chunk_list || 0 == chunk_list_num_entries); + HDassert(io_info); + HDassert(type_info); - if (num_entries > 0) { - size_t i; - int mpi_code; - void * base_buf; - - H5_CHECK_OVERFLOW(num_entries, size_t, int); - - /* Allocate arrays */ - if (NULL == (length_array = (int *)H5MM_malloc((size_t)num_entries * sizeof(int)))) - HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, - "memory allocation failed for filtered collective write length array") - if (NULL == (write_buf_array = (MPI_Aint *)H5MM_malloc((size_t)num_entries * sizeof(MPI_Aint)))) - HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, - "memory allocation failed for filtered collective write buf length array") - if (NULL == (file_offset_array = (MPI_Aint *)H5MM_malloc((size_t)num_entries * sizeof(MPI_Aint)))) - HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, - "memory allocation failed for collective write offset array") - - /* Ensure the list is sorted in ascending order of offset in the file */ - HDqsort(chunk_list, num_entries, sizeof(H5D_filtered_collective_io_info_t), - H5D__cmp_filtered_collective_io_info_entry); + /* Initialize temporary I/O info */ + coll_io_info = *io_info; - base_buf = chunk_list[0].buf; - for (i = 0; i < num_entries; i++) { - /* Set up the offset in the file, the length of the chunk data, and the relative - * displacement of the chunk data write buffer - */ - file_offset_array[i] = (MPI_Aint)chunk_list[i].chunk_states.new_chunk.offset; - length_array[i] = (int)chunk_list[i].chunk_states.new_chunk.length; - write_buf_array[i] = (MPI_Aint)chunk_list[i].buf - (MPI_Aint)base_buf; - } /* end for */ + /* + * Construct MPI derived datatype for collective I/O on chunks + */ + if (H5D__mpio_collective_filtered_io_type(chunk_list, chunk_list_num_entries, io_info->op_type, &mem_type, + &mem_type_is_derived, &file_type, &file_type_is_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_BADTYPE, FAIL, "couldn't create MPI I/O type for chunk I/O") - /* Create memory MPI type */ - if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed((int)num_entries, length_array, - write_buf_array, MPI_BYTE, new_mem_type))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code) - *mem_type_derived = TRUE; - if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_mem_type))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) - - /* Create file MPI type */ - if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed((int)num_entries, length_array, - file_offset_array, MPI_BYTE, new_file_type))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code) - *file_type_derived = TRUE; - if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_file_type))) - HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) - } /* end if */ + /* + * For reads, determine how many chunks are actually being read. + * Note that if this is a read during a write operation + * (read chunk -> unfilter -> modify -> write back), some + * chunks may not need to be read if they're being fully + * overwritten during a write operation. + */ + if (io_info->op_type == H5D_IO_OP_READ) { + for (i = 0, num_chunks = 0; i < chunk_list_num_entries; i++) { + HDassert(chunk_list[i].buf); + + if (chunk_list[i].need_read) { + if (!H5F_addr_defined(base_read_offset)) + base_read_offset = chunk_list[i].chunk_current.offset; + + num_chunks++; + } + } + } + else + num_chunks = chunk_list_num_entries; + + /* + * If this rank doesn't have a selection, it can + * skip I/O if independent I/O was requested at + * the low level, or if the MPI communicator size + * is 1. + * + * Otherwise, this rank has to participate in + * collective I/O, but probably has a NULL buf + * pointer, so override to a fake buffer since our + * write/read function expects one. + */ + if (num_chunks == 0) { + H5FD_mpio_collective_opt_t coll_opt_mode; + + /* Get the collective_opt property to check whether the application wants to do IO individually. */ + if (H5CX_get_mpio_coll_opt(&coll_opt_mode) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O collective_opt property") + + if ((mpi_size == 1) || (H5FD_MPIO_INDIVIDUAL_IO == coll_opt_mode)) { + HGOTO_DONE(SUCCEED) + } + else { + if (io_info->op_type == H5D_IO_OP_WRITE) + coll_io_info.u.wbuf = &fake_buf; + else + coll_io_info.u.rbuf = &fake_buf; + } + } + + /* + * Setup for I/O operation + */ + + mpi_buf_count = (num_chunks) ? 1 : 0; + + if (num_chunks) { + /* + * Setup the base storage address for this operation + * to be the first chunk's file address + */ + if (io_info->op_type == H5D_IO_OP_WRITE) + ctg_store.contig.dset_addr = chunk_list[0].chunk_new.offset; + else + ctg_store.contig.dset_addr = base_read_offset; + } + else + ctg_store.contig.dset_addr = 0; + + ctg_store.contig.dset_size = (hsize_t)io_info->dset->shared->layout.u.chunk.size; + coll_io_info.store = &ctg_store; + + /* Perform I/O */ + if (H5D__final_collective_io(&coll_io_info, type_info, mpi_buf_count, file_type, mem_type) < 0) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish MPI I/O") done: - if (write_buf_array) - H5MM_free(write_buf_array); - if (file_offset_array) - H5MM_free(file_offset_array); - if (length_array) - H5MM_free(length_array); + /* Free the MPI buf and file types, if they were derived */ + if (mem_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + if (file_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) FUNC_LEAVE_NOAPI(ret_value) -} /* end H5D__mpio_filtered_collective_write_type() */ +} /* end H5D__mpio_collective_filtered_chunk_common_io() */ /*------------------------------------------------------------------------- - * Function: H5D__filtered_collective_chunk_entry_io + * Function: H5D__mpio_collective_filtered_chunk_read * - * Purpose: Given an entry for a filtered chunk, performs the necessary - * steps for updating the chunk data during a collective - * write, or for reading the chunk from file during a - * collective read. + * Purpose: This routine coordinates a collective read across all ranks + * of the chunks they have selected. Each rank will then go + * and * * Return: Non-negative on success/Negative on failure * - * Programmer: Jordan Henderson - * Wednesday, January 18, 2017 - * *------------------------------------------------------------------------- */ static herr_t -H5D__filtered_collective_chunk_entry_io(H5D_filtered_collective_io_info_t *chunk_entry, - const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, - const H5D_chunk_map_t *fm) +H5D__mpio_collective_filtered_chunk_read(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, const H5D_io_info_t *io_info, + const H5D_type_info_t *type_info, int mpi_rank, int mpi_size) { - H5D_chunk_info_t *chunk_info = NULL; - H5S_sel_iter_t * mem_iter = NULL; /* Memory iterator for H5D__scatter_mem/H5D__gather_mem */ - H5S_sel_iter_t * file_iter = NULL; - H5Z_EDC_t err_detect; /* Error detection info */ - H5Z_cb_t filter_cb; /* I/O filter callback function */ - unsigned filter_mask = 0; - hsize_t iter_nelmts; /* Number of points to iterate over for the chunk IO operation */ - hssize_t extent_npoints; - hsize_t true_chunk_size; - hbool_t mem_iter_init = FALSE; - hbool_t file_iter_init = FALSE; - size_t buf_size; - size_t i; - H5S_t * dataspace = NULL; /* Other process' dataspace for the chunk */ - void * tmp_gath_buf = NULL; /* Temporary gather buffer to gather into from application buffer - before scattering out to the chunk data buffer (when writing data), - or vice versa (when reading data) */ - int mpi_code; - herr_t ret_value = SUCCEED; + H5D_fill_buf_info_t fb_info; + H5D_chunk_info_t * chunk_info = NULL; + H5D_io_info_t coll_io_info; + H5Z_EDC_t err_detect; /* Error detection info */ + H5Z_cb_t filter_cb; /* I/O filter callback function */ + hsize_t file_chunk_size = 0; + hsize_t iter_nelmts; /* Number of points to iterate over for the chunk IO operation */ + hbool_t should_fill = FALSE; + hbool_t fb_info_init = FALSE; + hbool_t index_empty = FALSE; + size_t i; + H5S_t * fill_space = NULL; + void * base_read_buf = NULL; + herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC - HDassert(chunk_entry); + HDassert(chunk_list || 0 == chunk_list_num_entries); HDassert(io_info); HDassert(type_info); - HDassert(fm); - /* Retrieve filter settings from API context */ - if (H5CX_get_err_detect(&err_detect) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get error detection info") - if (H5CX_get_filter_cb(&filter_cb) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get I/O filter callback function") +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Filtered collective chunk read"); +#else + (void)mpi_rank; +#endif + + /* Initialize temporary I/O info */ + coll_io_info = *io_info; + coll_io_info.u.rbuf = NULL; - /* Look up the chunk and get its file and memory dataspaces */ - if (NULL == (chunk_info = (H5D_chunk_info_t *)H5SL_search(fm->sel_chunks, &chunk_entry->index))) - HGOTO_ERROR(H5E_DATASPACE, H5E_NOTFOUND, FAIL, "can't locate chunk in skip list") + if (chunk_list_num_entries) { + /* Retrieve filter settings from API context */ + if (H5CX_get_err_detect(&err_detect) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get error detection info") + if (H5CX_get_filter_cb(&filter_cb) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get I/O filter callback function") - if ((extent_npoints = H5S_GET_EXTENT_NPOINTS(chunk_info->fspace)) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid") - true_chunk_size = (hsize_t)extent_npoints * type_info->src_type_size; + /* Set size of full chunks in dataset */ + file_chunk_size = io_info->dset->shared->layout.u.chunk.size; + + /* Determine if fill values should be "read" for unallocated chunks */ + should_fill = (io_info->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_ALLOC) || + ((io_info->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_IFSET) && + io_info->dset->shared->dcpl_cache.fill.fill_defined); + } - /* If the size of the filtered chunk is larger than the number of points in the - * chunk file space extent times the datatype size, allocate enough space to hold the - * whole filtered chunk. Otherwise, allocate a buffer equal to the size of the - * chunk so that the unfiltering operation doesn't have to grow the buffer. + /* + * Allocate memory buffers for all chunks being read. Chunk data buffers are of + * the largest size between the chunk's current filtered size and the chunk's true + * size, as calculated by the number of elements in the chunk's file space extent + * multiplied by the datatype size. This tries to ensure that: + * + * * If we're reading the chunk and the filter normally reduces the chunk size, + * the unfiltering operation won't need to grow the buffer. + * * If we're reading the chunk and the filter normally grows the chunk size, + * we make sure to read into a buffer of size equal to the filtered chunk's + * size; reading into a (smaller) buffer of size equal to the unfiltered + * chunk size would of course be bad. */ - buf_size = MAX(chunk_entry->chunk_states.chunk_current.length, true_chunk_size); + for (i = 0; i < chunk_list_num_entries; i++) { + HDassert(chunk_list[i].need_read); + + chunk_list[i].chunk_buf_size = MAX(chunk_list[i].chunk_current.length, file_chunk_size); + + if (NULL == (chunk_list[i].buf = H5MM_malloc(chunk_list[i].chunk_buf_size))) { + /* Push an error, but participate in collective read */ + HDONE_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk data buffer") + break; + } + + /* + * Check if chunk is currently allocated. If not, don't try to + * read it from the file. Instead, just fill the chunk buffer + * with the fill value if necessary. + */ + if (H5F_addr_defined(chunk_list[i].chunk_current.offset)) { + /* Set first read buffer */ + if (!base_read_buf) + base_read_buf = chunk_list[i].buf; + + /* Set chunk's new length for eventual filter pipeline calls */ + if (chunk_list[i].skip_filter_pline) + chunk_list[i].chunk_new.length = file_chunk_size; + else + chunk_list[i].chunk_new.length = chunk_list[i].chunk_current.length; + } + else { + chunk_list[i].need_read = FALSE; + + /* Set chunk's new length for eventual filter pipeline calls */ + chunk_list[i].chunk_new.length = file_chunk_size; + + if (should_fill) { + /* Initialize fill value buffer if not already initialized */ + if (!fb_info_init) { + hsize_t chunk_dims[H5S_MAX_RANK]; + + HDassert(io_info->dset->shared->ndims == io_info->dset->shared->layout.u.chunk.ndims - 1); + for (size_t j = 0; j < io_info->dset->shared->layout.u.chunk.ndims - 1; j++) + chunk_dims[j] = (hsize_t)io_info->dset->shared->layout.u.chunk.dim[j]; + + /* Get a dataspace for filling chunk memory buffers */ + if (NULL == (fill_space = H5S_create_simple( + io_info->dset->shared->layout.u.chunk.ndims - 1, chunk_dims, NULL))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to create chunk fill dataspace") + + /* Initialize fill value buffer */ + if (H5D__fill_init(&fb_info, NULL, (H5MM_allocate_t)H5D__chunk_mem_alloc, + (void *)&io_info->dset->shared->dcpl_cache.pline, + (H5MM_free_t)H5D__chunk_mem_free, + (void *)&io_info->dset->shared->dcpl_cache.pline, + &io_info->dset->shared->dcpl_cache.fill, io_info->dset->shared->type, + io_info->dset->shared->type_id, 0, file_chunk_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "can't initialize fill value buffer") + + fb_info_init = TRUE; + } - if (NULL == (chunk_entry->buf = H5MM_malloc(buf_size))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk data buffer") + /* Write fill value to memory buffer */ + HDassert(fb_info.fill_buf); + if (H5D__fill(fb_info.fill_buf, io_info->dset->shared->type, chunk_list[i].buf, + type_info->mem_type, fill_space) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't fill chunk buffer with fill value") + } + } + } - /* If this is not a full chunk overwrite or this is a read operation, the chunk must be - * read from the file and unfiltered. + /* + * If dataset is incrementally allocated and hasn't been written to + * yet, the chunk index should be empty. In this case, a collective + * read of chunks is essentially a no-op, so avoid it here. */ - if (!chunk_entry->full_overwrite || io_info->op_type == H5D_IO_OP_READ) { - H5FD_mpio_xfer_t xfer_mode; /* Parallel transfer for this request */ + index_empty = FALSE; + if (io_info->dset->shared->dcpl_cache.fill.alloc_time == H5D_ALLOC_TIME_INCR) + if (H5D__chunk_index_empty(io_info->dset, &index_empty) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't determine if chunk index is empty") - chunk_entry->chunk_states.new_chunk.length = chunk_entry->chunk_states.chunk_current.length; + if (!index_empty) { + /* + * Override the read buffer to point to the address of + * the first chunk data buffer being read into + */ + if (base_read_buf) + coll_io_info.u.rbuf = base_read_buf; - /* Currently, these chunk reads are done independently and will likely - * cause issues with collective metadata reads enabled. In the future, - * this should be refactored to use collective chunk reads - JTH */ + /* Perform collective chunk read */ + if (H5D__mpio_collective_filtered_chunk_common_io(chunk_list, chunk_list_num_entries, &coll_io_info, + type_info, mpi_size) < 0) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish collective filtered chunk read") + } - /* Get the original state of parallel I/O transfer mode */ - if (H5CX_get_io_xfer_mode(&xfer_mode) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode") + /* + * Iterate through all the read chunks, unfiltering them and scattering their + * data out to the application's read buffer. + */ + for (i = 0; i < chunk_list_num_entries; i++) { + chunk_info = chunk_list[i].chunk_info; + + /* Unfilter the chunk, unless we didn't read it from the file */ + if (chunk_list[i].need_read && !chunk_list[i].skip_filter_pline) { + if (H5Z_pipeline(&io_info->dset->shared->dcpl_cache.pline, H5Z_FLAG_REVERSE, + &(chunk_list[i].index_info.filter_mask), err_detect, filter_cb, + (size_t *)&chunk_list[i].chunk_new.length, &chunk_list[i].chunk_buf_size, + &chunk_list[i].buf) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTFILTER, FAIL, "couldn't unfilter chunk for modifying") + } + + /* Scatter the chunk data to the read buffer */ + iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->fspace); + + if (H5D_select_io_mem(io_info->u.rbuf, chunk_info->mspace, chunk_list[i].buf, chunk_info->fspace, + type_info->src_type_size, (size_t)iter_nelmts) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't copy chunk data to read buffer") + } - /* Change the xfer_mode to independent for handling the I/O */ - if (H5CX_set_io_xfer_mode(H5FD_MPIO_INDEPENDENT) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't set MPI-I/O transfer mode") +done: + /* Free all resources used by entries in the chunk list */ + for (i = 0; i < chunk_list_num_entries; i++) { + if (chunk_list[i].buf) { + H5MM_free(chunk_list[i].buf); + chunk_list[i].buf = NULL; + } + } - if (H5F_shared_block_read(io_info->f_sh, H5FD_MEM_DRAW, - chunk_entry->chunk_states.chunk_current.offset, - chunk_entry->chunk_states.new_chunk.length, chunk_entry->buf) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "unable to read raw data chunk") + /* Release the fill buffer info, if it's been initialized */ + if (fb_info_init && H5D__fill_term(&fb_info) < 0) + HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "Can't release fill buffer info") + if (fill_space && (H5S_close(fill_space) < 0)) + HDONE_ERROR(H5E_DATASET, H5E_CLOSEERROR, FAIL, "can't close fill space") - /* Return to the original I/O transfer mode setting */ - if (H5CX_set_io_xfer_mode(xfer_mode) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't set MPI-I/O transfer mode") +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif - if (H5Z_pipeline(&io_info->dset->shared->dcpl_cache.pline, H5Z_FLAG_REVERSE, &filter_mask, err_detect, - filter_cb, (size_t *)&chunk_entry->chunk_states.new_chunk.length, &buf_size, - &chunk_entry->buf) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTFILTER, FAIL, "couldn't unfilter chunk for modifying") - } /* end if */ - else { - chunk_entry->chunk_states.new_chunk.length = true_chunk_size; - } /* end else */ + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_collective_filtered_chunk_read() */ - /* Initialize iterator for memory selection */ - if (NULL == (mem_iter = (H5S_sel_iter_t *)H5MM_malloc(sizeof(H5S_sel_iter_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate memory iterator") +/*------------------------------------------------------------------------- + * Function: H5D__mpio_collective_filtered_chunk_update + * + * Purpose: When performing a parallel write on a chunked dataset with + * filters applied, all ranks must update their owned chunks + * with their own modification data and data from other ranks. + * This routine is responsible for coordinating that process. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_collective_filtered_chunk_update(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, + H5D_filtered_collective_io_info_t *chunk_hash_table, + unsigned char **chunk_msg_bufs, int chunk_msg_bufs_len, + const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, + int mpi_rank, int mpi_size) +{ + H5D_fill_buf_info_t fb_info; + H5D_chunk_info_t * chunk_info = NULL; + H5S_sel_iter_t * sel_iter = NULL; /* Dataspace selection iterator for H5D__scatter_mem */ + H5D_io_info_t coll_io_info; + H5Z_EDC_t err_detect; /* Error detection info */ + H5Z_cb_t filter_cb; /* I/O filter callback function */ + hsize_t file_chunk_size = 0; + hsize_t iter_nelmts; /* Number of points to iterate over for the chunk IO operation */ + hbool_t should_fill = FALSE; + hbool_t fb_info_init = FALSE; + hbool_t sel_iter_init = FALSE; + hbool_t index_empty = FALSE; + size_t i; + H5S_t * dataspace = NULL; + H5S_t * fill_space = NULL; + void * base_read_buf = NULL; + herr_t ret_value = SUCCEED; - if (H5S_select_iter_init(mem_iter, chunk_info->mspace, type_info->src_type_size, 0) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize memory selection information") - mem_iter_init = TRUE; + FUNC_ENTER_STATIC - /* If this is a read operation, scatter the read chunk data to the user's buffer. - * - * If this is a write operation, update the chunk data buffer with the modifications - * from the current process, then apply any modifications from other processes. Finally, - * filter the newly-updated chunk. - */ - switch (io_info->op_type) { - case H5D_IO_OP_READ: - if (NULL == (file_iter = (H5S_sel_iter_t *)H5MM_malloc(sizeof(H5S_sel_iter_t)))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate file iterator") + HDassert(chunk_list || 0 == chunk_list_num_entries); + HDassert((chunk_msg_bufs && chunk_hash_table) || 0 == chunk_msg_bufs_len); + HDassert(io_info); + HDassert(type_info); - if (H5S_select_iter_init(file_iter, chunk_info->fspace, type_info->src_type_size, 0) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, - "unable to initialize memory selection information") - file_iter_init = TRUE; +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Filtered collective chunk update"); +#endif + + if (chunk_list_num_entries) { + /* Retrieve filter settings from API context */ + if (H5CX_get_err_detect(&err_detect) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get error detection info") + if (H5CX_get_filter_cb(&filter_cb) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get I/O filter callback function") - iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->fspace); + /* Set size of full chunks in dataset */ + file_chunk_size = io_info->dset->shared->layout.u.chunk.size; - if (NULL == (tmp_gath_buf = H5MM_malloc(iter_nelmts * type_info->src_type_size))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate temporary gather buffer") + /* Determine if fill values should be written to chunks */ + should_fill = (io_info->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_ALLOC) || + ((io_info->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_IFSET) && + io_info->dset->shared->dcpl_cache.fill.fill_defined); + } - if (!H5D__gather_mem(chunk_entry->buf, file_iter, (size_t)iter_nelmts, tmp_gath_buf)) - HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't gather from chunk buffer") + /* + * Allocate memory buffers for all owned chunks. Chunk data buffers are of the + * largest size between the chunk's current filtered size and the chunk's true + * size, as calculated by the number of elements in the chunk's file space extent + * multiplied by the datatype size. This tries to ensure that: + * + * * If we're fully overwriting the chunk and the filter normally reduces the + * chunk size, we simply have the exact buffer size required to hold the + * unfiltered chunk data. + * * If we're fully overwriting the chunk and the filter normally grows the + * chunk size (e.g., fletcher32 filter), the final filtering operation + * (hopefully) won't need to grow the buffer. + * * If we're reading the chunk and the filter normally reduces the chunk size, + * the unfiltering operation won't need to grow the buffer. + * * If we're reading the chunk and the filter normally grows the chunk size, + * we make sure to read into a buffer of size equal to the filtered chunk's + * size; reading into a (smaller) buffer of size equal to the unfiltered + * chunk size would of course be bad. + */ + for (i = 0; i < chunk_list_num_entries; i++) { + HDassert(mpi_rank == chunk_list[i].new_owner); - iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace); + chunk_list[i].chunk_buf_size = MAX(chunk_list[i].chunk_current.length, file_chunk_size); - if (H5D__scatter_mem(tmp_gath_buf, mem_iter, (size_t)iter_nelmts, io_info->u.rbuf) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't scatter to read buffer") + /* + * If this chunk hasn't been allocated yet and we aren't writing + * out fill values to it, make sure to 0-fill its memory buffer + * so we don't use uninitialized memory. + */ + if (!H5F_addr_defined(chunk_list[i].chunk_current.offset) && !should_fill) + chunk_list[i].buf = H5MM_calloc(chunk_list[i].chunk_buf_size); + else + chunk_list[i].buf = H5MM_malloc(chunk_list[i].chunk_buf_size); + if (NULL == chunk_list[i].buf) { + /* Push an error, but participate in collective read */ + HDONE_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk data buffer") break; + } + + /* Set chunk's new length for eventual filter pipeline calls */ + if (chunk_list[i].need_read) { + /* + * Check if chunk is currently allocated. If not, don't try to + * read it from the file. Instead, just fill the chunk buffer + * with the fill value if fill values are to be written. + */ + if (H5F_addr_defined(chunk_list[i].chunk_current.offset)) { + /* Set first read buffer */ + if (!base_read_buf) + base_read_buf = chunk_list[i].buf; + + /* Set chunk's new length for eventual filter pipeline calls */ + if (chunk_list[i].skip_filter_pline) + chunk_list[i].chunk_new.length = file_chunk_size; + else + chunk_list[i].chunk_new.length = chunk_list[i].chunk_current.length; + } + else { + chunk_list[i].need_read = FALSE; + + /* Set chunk's new length for eventual filter pipeline calls */ + chunk_list[i].chunk_new.length = file_chunk_size; + + if (should_fill) { + /* Initialize fill value buffer if not already initialized */ + if (!fb_info_init) { + hsize_t chunk_dims[H5S_MAX_RANK]; + + HDassert(io_info->dset->shared->ndims == + io_info->dset->shared->layout.u.chunk.ndims - 1); + for (size_t j = 0; j < io_info->dset->shared->layout.u.chunk.ndims - 1; j++) + chunk_dims[j] = (hsize_t)io_info->dset->shared->layout.u.chunk.dim[j]; + + /* Get a dataspace for filling chunk memory buffers */ + if (NULL == (fill_space = H5S_create_simple( + io_info->dset->shared->layout.u.chunk.ndims - 1, chunk_dims, NULL))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, + "unable to create chunk fill dataspace") + + /* Initialize fill value buffer */ + if (H5D__fill_init(&fb_info, NULL, (H5MM_allocate_t)H5D__chunk_mem_alloc, + (void *)&io_info->dset->shared->dcpl_cache.pline, + (H5MM_free_t)H5D__chunk_mem_free, + (void *)&io_info->dset->shared->dcpl_cache.pline, + &io_info->dset->shared->dcpl_cache.fill, + io_info->dset->shared->type, io_info->dset->shared->type_id, 0, + file_chunk_size) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "can't initialize fill value buffer") + + fb_info_init = TRUE; + } + + /* Write fill value to memory buffer */ + HDassert(fb_info.fill_buf); + if (H5D__fill(fb_info.fill_buf, io_info->dset->shared->type, chunk_list[i].buf, + type_info->mem_type, fill_space) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, + "couldn't fill chunk buffer with fill value") + } + } + } + else + chunk_list[i].chunk_new.length = file_chunk_size; + } - case H5D_IO_OP_WRITE: - iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace); + /* + * If dataset is incrementally allocated and hasn't been written to + * yet, the chunk index should be empty. In this case, a collective + * read of chunks is essentially a no-op, so avoid it here. + */ + index_empty = FALSE; + if (io_info->dset->shared->dcpl_cache.fill.alloc_time == H5D_ALLOC_TIME_INCR) + if (H5D__chunk_index_empty(io_info->dset, &index_empty) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't determine if chunk index is empty") - if (NULL == (tmp_gath_buf = H5MM_malloc(iter_nelmts * type_info->src_type_size))) - HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate temporary gather buffer") + if (!index_empty) { + /* + * Setup for I/O operation + */ - /* Gather modification data from the application write buffer into a temporary buffer */ - if (0 == H5D__gather_mem(io_info->u.wbuf, mem_iter, (size_t)iter_nelmts, tmp_gath_buf)) - HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "couldn't gather from write buffer") + /* Initialize temporary I/O info */ + coll_io_info = *io_info; + coll_io_info.op_type = H5D_IO_OP_READ; - if (H5S_SELECT_ITER_RELEASE(mem_iter) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator") - mem_iter_init = FALSE; + /* Override the read buffer to point to the address of the first + * chunk data buffer being read into + */ + if (base_read_buf) + coll_io_info.u.rbuf = base_read_buf; - /* Initialize iterator for file selection */ - if (H5S_select_iter_init(mem_iter, chunk_info->fspace, type_info->dst_type_size, 0) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, - "unable to initialize file selection information") - mem_iter_init = TRUE; + /* Read all chunks that need to be read from the file */ + if (H5D__mpio_collective_filtered_chunk_common_io(chunk_list, chunk_list_num_entries, &coll_io_info, + type_info, mpi_size) < 0) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish collective filtered chunk read") + } - iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->fspace); + /* + * Now that all owned chunks have been read, update the chunks + * with modification data from the owning rank and other ranks. + */ - /* Scatter the owner's modification data into the chunk data buffer according to - * the file space. - */ - if (H5D__scatter_mem(tmp_gath_buf, mem_iter, (size_t)iter_nelmts, chunk_entry->buf) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't scatter to chunk data buffer") + /* Process all chunks with data from the owning rank first */ + for (i = 0; i < chunk_list_num_entries; i++) { + HDassert(mpi_rank == chunk_list[i].new_owner); - if (H5S_SELECT_ITER_RELEASE(mem_iter) < 0) - HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator") - mem_iter_init = FALSE; + chunk_info = chunk_list[i].chunk_info; - if (MPI_SUCCESS != - (mpi_code = MPI_Waitall(chunk_entry->async_info.num_receive_requests, - chunk_entry->async_info.receive_requests_array, MPI_STATUSES_IGNORE))) - HMPI_GOTO_ERROR(FAIL, "MPI_Waitall failed", mpi_code) + /* + * If this chunk wasn't being fully overwritten, we read it from + * the file, so we need to unfilter it + */ + if (chunk_list[i].need_read && !chunk_list[i].skip_filter_pline) { + if (H5Z_pipeline(&io_info->dset->shared->dcpl_cache.pline, H5Z_FLAG_REVERSE, + &(chunk_list[i].index_info.filter_mask), err_detect, filter_cb, + (size_t *)&chunk_list[i].chunk_new.length, &chunk_list[i].chunk_buf_size, + &chunk_list[i].buf) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTFILTER, FAIL, "couldn't unfilter chunk for modifying") + } + + iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace); + + if (H5D_select_io_mem(chunk_list[i].buf, chunk_info->fspace, io_info->u.wbuf, chunk_info->mspace, + type_info->dst_type_size, (size_t)iter_nelmts) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't copy chunk data to write buffer") + } - /* For each asynchronous receive call previously posted, receive the chunk modification - * buffer from another rank and update the chunk data - */ - for (i = 0; i < (size_t)chunk_entry->async_info.num_receive_requests; i++) { - const unsigned char *mod_data_p; + /* Allocate iterator for memory selection */ + if (NULL == (sel_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate memory iterator") - /* Decode the process' chunk file dataspace */ - mod_data_p = chunk_entry->async_info.receive_buffer_array[i]; - if (NULL == (dataspace = H5S_decode(&mod_data_p))) + /* Now process all received chunk message buffers */ + for (i = 0; i < (size_t)chunk_msg_bufs_len; i++) { + H5D_filtered_collective_io_info_t *chunk_entry = NULL; + const unsigned char * msg_ptr = chunk_msg_bufs[i]; + hsize_t chunk_idx; + + if (msg_ptr) { + /* Retrieve the chunk's index value */ + HDmemcpy(&chunk_idx, msg_ptr, sizeof(hsize_t)); + msg_ptr += sizeof(hsize_t); + + /* Find the chunk entry according to its chunk index */ + HASH_FIND(hh, chunk_hash_table, &chunk_idx, sizeof(hsize_t), chunk_entry); + HDassert(chunk_entry); + HDassert(mpi_rank == chunk_entry->new_owner); + + /* + * Only process the chunk if its data buffer is allocated. + * In the case of multi-chunk I/O, we're only working on + * a chunk at a time, so we need to skip over messages + * that aren't for the chunk we're currently working on. + */ + if (!chunk_entry->buf) + continue; + else { + /* Decode the chunk file dataspace from the message */ + if (NULL == (dataspace = H5S_decode(&msg_ptr))) HGOTO_ERROR(H5E_DATASET, H5E_CANTDECODE, FAIL, "unable to decode dataspace") - if (H5S_select_iter_init(mem_iter, dataspace, type_info->dst_type_size, 0) < 0) + if (H5S_select_iter_init(sel_iter, dataspace, type_info->dst_type_size, + H5S_SEL_ITER_SHARE_WITH_DATASPACE) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize memory selection information") - mem_iter_init = TRUE; + sel_iter_init = TRUE; iter_nelmts = H5S_GET_SELECT_NPOINTS(dataspace); /* Update the chunk data with the received modification data */ - if (H5D__scatter_mem(mod_data_p, mem_iter, (size_t)iter_nelmts, chunk_entry->buf) < 0) + if (H5D__scatter_mem(msg_ptr, sel_iter, (size_t)iter_nelmts, chunk_entry->buf) < 0) HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't scatter to write buffer") - if (H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + if (H5S_SELECT_ITER_RELEASE(sel_iter) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator") - mem_iter_init = FALSE; + sel_iter_init = FALSE; + if (dataspace) { if (H5S_close(dataspace) < 0) HGOTO_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "can't close dataspace") dataspace = NULL; } - H5MM_free(chunk_entry->async_info.receive_buffer_array[i]); - } /* end for */ - /* Filter the chunk */ - if (H5Z_pipeline(&io_info->dset->shared->dcpl_cache.pline, 0, &filter_mask, err_detect, filter_cb, - (size_t *)&chunk_entry->chunk_states.new_chunk.length, &buf_size, - &chunk_entry->buf) < 0) + H5MM_free(chunk_msg_bufs[i]); + chunk_msg_bufs[i] = NULL; + } + } + } + + /* Finally, filter all the chunks */ + for (i = 0; i < chunk_list_num_entries; i++) { + if (!chunk_list[i].skip_filter_pline) { + if (H5Z_pipeline(&io_info->dset->shared->dcpl_cache.pline, 0, + &(chunk_list[i].index_info.filter_mask), err_detect, filter_cb, + (size_t *)&chunk_list[i].chunk_new.length, &chunk_list[i].chunk_buf_size, + &chunk_list[i].buf) < 0) HGOTO_ERROR(H5E_PLINE, H5E_CANTFILTER, FAIL, "output pipeline failed") + } #if H5_SIZEOF_SIZE_T > 4 - /* Check for the chunk expanding too much to encode in a 32-bit value */ - if (chunk_entry->chunk_states.new_chunk.length > ((size_t)0xffffffff)) - HGOTO_ERROR(H5E_DATASET, H5E_BADRANGE, FAIL, "chunk too large for 32-bit length") + /* Check for the chunk expanding too much to encode in a 32-bit value */ + if (chunk_list[i].chunk_new.length > ((size_t)0xffffffff)) + HGOTO_ERROR(H5E_DATASET, H5E_BADRANGE, FAIL, "chunk too large for 32-bit length") #endif - break; + } - default: - HGOTO_ERROR(H5E_DATASET, H5E_BADVALUE, FAIL, "invalid I/O operation") - } /* end switch */ +done: + if (sel_iter) { + if (sel_iter_init && H5S_SELECT_ITER_RELEASE(sel_iter) < 0) + HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator") + sel_iter = H5FL_FREE(H5S_sel_iter_t, sel_iter); + } + if (dataspace && (H5S_close(dataspace) < 0)) + HDONE_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "can't close dataspace") + if (fill_space && (H5S_close(fill_space) < 0)) + HDONE_ERROR(H5E_DATASET, H5E_CLOSEERROR, FAIL, "can't close fill space") + + /* Release the fill buffer info, if it's been initialized */ + if (fb_info_init && H5D__fill_term(&fb_info) < 0) + HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "Can't release fill buffer info") + + /* On failure, try to free all resources used by entries in the chunk list */ + if (ret_value < 0) { + for (i = 0; i < chunk_list_num_entries; i++) { + if (chunk_list[i].buf) { + H5MM_free(chunk_list[i].buf); + chunk_list[i].buf = NULL; + } + } + } + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_collective_filtered_chunk_update() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_collective_filtered_chunk_reallocate + * + * Purpose: When performing a parallel write on a chunked dataset with + * filters applied, all ranks must eventually get together and + * perform a collective reallocation of space in the file for + * all chunks that were modified on all ranks. This routine is + * responsible for coordinating that process. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_collective_filtered_chunk_reallocate(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, size_t *num_chunks_assigned_map, + H5D_io_info_t *io_info, H5D_chk_idx_info_t *idx_info, + int mpi_rank, int mpi_size) +{ + H5D_chunk_alloc_info_t *collective_list = NULL; + MPI_Datatype send_type; + MPI_Datatype recv_type; + hbool_t send_type_derived = FALSE; + hbool_t recv_type_derived = FALSE; + hbool_t need_sort = FALSE; + size_t collective_num_entries = 0; + size_t num_local_chunks_processed = 0; + size_t i; + void * gathered_array = NULL; + int * counts_disps_array = NULL; + int * counts_ptr = NULL; + int * displacements_ptr = NULL; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(chunk_list || 0 == chunk_list_num_entries); + HDassert(io_info); + HDassert(idx_info); + HDassert(idx_info->storage->idx_type != H5D_CHUNK_IDX_NONE); + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Reallocation of chunk file space"); +#endif + + /* + * Make sure it's safe to cast this rank's number + * of chunks to be sent into an int for MPI + */ + H5_CHECK_OVERFLOW(chunk_list_num_entries, size_t, int); + + /* Create derived datatypes for the chunk file space info needed */ + if (H5D__mpio_get_chunk_alloc_info_types(&recv_type, &recv_type_derived, &send_type, &send_type_derived) < + 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "can't create derived datatypes for chunk file space info") + + /* + * Gather the new chunk sizes to all ranks for a collective reallocation + * of the chunks in the file. + */ + if (num_chunks_assigned_map) { + /* + * If a mapping between rank value -> number of assigned chunks has + * been provided (usually during linked-chunk I/O), we can use this + * to optimize MPI overhead a bit since MPI ranks won't need to + * first inform each other about how many chunks they're contributing. + */ + if (NULL == (counts_disps_array = H5MM_malloc(2 * (size_t)mpi_size * sizeof(*counts_disps_array)))) { + /* Push an error, but still participate in collective gather operation */ + HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "couldn't allocate receive counts and displacements array") + } + else { + /* Set the receive counts from the assigned chunks map */ + counts_ptr = counts_disps_array; + + for (i = 0; i < (size_t)mpi_size; i++) + H5_CHECKED_ASSIGN(counts_ptr[i], int, num_chunks_assigned_map[i], size_t); + + /* Set the displacements into the receive buffer for the gather operation */ + displacements_ptr = &counts_disps_array[mpi_size]; + + *displacements_ptr = 0; + for (i = 1; i < (size_t)mpi_size; i++) + displacements_ptr[i] = displacements_ptr[i - 1] + counts_ptr[i - 1]; + } + + /* Perform gather operation */ + if (H5_mpio_gatherv_alloc(chunk_list, (int)chunk_list_num_entries, send_type, counts_ptr, + displacements_ptr, recv_type, TRUE, 0, io_info->comm, mpi_rank, mpi_size, + &gathered_array, &collective_num_entries) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "can't gather chunk file space info to/from ranks") + } + else { + /* + * If no mapping between rank value -> number of assigned chunks has + * been provided (usually during multi-chunk I/O), all MPI ranks will + * need to first inform other ranks about how many chunks they're + * contributing before performing the actual gather operation. Use + * the 'simple' MPI_Allgatherv wrapper for this. + */ + if (H5_mpio_gatherv_alloc_simple(chunk_list, (int)chunk_list_num_entries, send_type, recv_type, TRUE, + 0, io_info->comm, mpi_rank, mpi_size, &gathered_array, + &collective_num_entries) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "can't gather chunk file space info to/from ranks") + } + + /* Collectively re-allocate the modified chunks (from each rank) in the file */ + collective_list = (H5D_chunk_alloc_info_t *)gathered_array; + for (i = 0, num_local_chunks_processed = 0; i < collective_num_entries; i++) { + H5D_chunk_alloc_info_t *coll_entry = &collective_list[i]; + hbool_t need_insert; + hbool_t update_local_chunk; + + if (H5D__chunk_file_alloc(idx_info, &coll_entry->chunk_current, &coll_entry->chunk_new, &need_insert, + NULL) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate chunk") + + /* + * If we just re-allocated a chunk that is local to this + * rank, make sure to update the chunk entry in the local + * chunk list + */ + update_local_chunk = + (num_local_chunks_processed < chunk_list_num_entries) && + (coll_entry->chunk_idx == chunk_list[num_local_chunks_processed].index_info.chunk_idx); + + if (update_local_chunk) { + H5D_filtered_collective_io_info_t *local_chunk; + + local_chunk = &chunk_list[num_local_chunks_processed]; + + /* Sanity check that this chunk is actually local */ + HDassert(mpi_rank == local_chunk->orig_owner); + HDassert(mpi_rank == local_chunk->new_owner); + + local_chunk->chunk_new = coll_entry->chunk_new; + local_chunk->index_info.need_insert = need_insert; + + /* + * Since chunk reallocation can move chunks around, check if + * the local chunk list is still in ascending offset of order + * in the file + */ + if (num_local_chunks_processed) { + haddr_t curr_chunk_offset = local_chunk->chunk_new.offset; + haddr_t prev_chunk_offset = chunk_list[num_local_chunks_processed - 1].chunk_new.offset; + + HDassert(H5F_addr_defined(prev_chunk_offset) && H5F_addr_defined(curr_chunk_offset)); + if (curr_chunk_offset < prev_chunk_offset) + need_sort = TRUE; + } + + num_local_chunks_processed++; + } + } + + HDassert(chunk_list_num_entries == num_local_chunks_processed); + + /* + * Ensure this rank's local chunk list is sorted in + * ascending order of offset in the file + */ + if (need_sort) + HDqsort(chunk_list, chunk_list_num_entries, sizeof(H5D_filtered_collective_io_info_t), + H5D__cmp_filtered_collective_io_info_entry); + +done: + H5MM_free(gathered_array); + H5MM_free(counts_disps_array); + + if (send_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&send_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + if (recv_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&recv_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5D__mpio_collective_filtered_chunk_reallocate() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_collective_filtered_chunk_reinsert + * + * Purpose: When performing a parallel write on a chunked dataset with + * filters applied, all ranks must eventually get together and + * perform a collective reinsertion into the dataset's chunk + * index of chunks that were modified. This routine is + * responsible for coordinating that process. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_collective_filtered_chunk_reinsert(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, size_t *num_chunks_assigned_map, + H5D_io_info_t *io_info, H5D_chk_idx_info_t *idx_info, + int mpi_rank, int mpi_size) +{ + H5D_chunk_ud_t chunk_ud; + MPI_Datatype send_type; + MPI_Datatype recv_type; + hbool_t send_type_derived = FALSE; + hbool_t recv_type_derived = FALSE; + hsize_t scaled_coords[H5O_LAYOUT_NDIMS]; + size_t collective_num_entries = 0; + size_t i; + void * gathered_array = NULL; + int * counts_disps_array = NULL; + int * counts_ptr = NULL; + int * displacements_ptr = NULL; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(chunk_list || 0 == chunk_list_num_entries); + HDassert(io_info); + HDassert(idx_info); + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TRACE_ENTER(mpi_rank); + H5D_MPIO_TIME_START(mpi_rank, "Reinsertion of modified chunks into chunk index"); +#endif + + /* Only re-insert chunks if index has an insert method */ + if (!idx_info->storage->ops->insert) + HGOTO_DONE(SUCCEED); + + /* + * Make sure it's safe to cast this rank's number + * of chunks to be sent into an int for MPI + */ + H5_CHECK_OVERFLOW(chunk_list_num_entries, size_t, int); + + /* Create derived datatypes for the chunk re-insertion info needed */ + if (H5D__mpio_get_chunk_insert_info_types(&recv_type, &recv_type_derived, &send_type, + &send_type_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, + "can't create derived datatypes for chunk re-insertion info") + + /* + * Gather information to all ranks for a collective re-insertion + * of the modified chunks into the chunk index + */ + if (num_chunks_assigned_map) { + /* + * If a mapping between rank value -> number of assigned chunks has + * been provided (usually during linked-chunk I/O), we can use this + * to optimize MPI overhead a bit since MPI ranks won't need to + * first inform each other about how many chunks they're contributing. + */ + if (NULL == (counts_disps_array = H5MM_malloc(2 * (size_t)mpi_size * sizeof(*counts_disps_array)))) { + /* Push an error, but still participate in collective gather operation */ + HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "couldn't allocate receive counts and displacements array") + } + else { + /* Set the receive counts from the assigned chunks map */ + counts_ptr = counts_disps_array; + + for (i = 0; i < (size_t)mpi_size; i++) + H5_CHECKED_ASSIGN(counts_ptr[i], int, num_chunks_assigned_map[i], size_t); + + /* Set the displacements into the receive buffer for the gather operation */ + displacements_ptr = &counts_disps_array[mpi_size]; + + *displacements_ptr = 0; + for (i = 1; i < (size_t)mpi_size; i++) + displacements_ptr[i] = displacements_ptr[i - 1] + counts_ptr[i - 1]; + } + + /* Perform gather operation */ + if (H5_mpio_gatherv_alloc(chunk_list, (int)chunk_list_num_entries, send_type, counts_ptr, + displacements_ptr, recv_type, TRUE, 0, io_info->comm, mpi_rank, mpi_size, + &gathered_array, &collective_num_entries) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, + "can't gather chunk index re-insertion info to/from ranks") + } + else { + /* + * If no mapping between rank value -> number of assigned chunks has + * been provided (usually during multi-chunk I/O), all MPI ranks will + * need to first inform other ranks about how many chunks they're + * contributing before performing the actual gather operation. Use + * the 'simple' MPI_Allgatherv wrapper for this. + */ + if (H5_mpio_gatherv_alloc_simple(chunk_list, (int)chunk_list_num_entries, send_type, recv_type, TRUE, + 0, io_info->comm, mpi_rank, mpi_size, &gathered_array, + &collective_num_entries) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, + "can't gather chunk index re-insertion info to/from ranks") + } + + /* Initialize static chunk udata fields from chunk index info */ + H5D_MPIO_INIT_CHUNK_UD_INFO(chunk_ud, idx_info); + + for (i = 0; i < collective_num_entries; i++) { + H5D_chunk_insert_info_t *coll_entry = &((H5D_chunk_insert_info_t *)gathered_array)[i]; + + /* + * We only need to reinsert this chunk if we had to actually + * allocate or reallocate space in the file for it + */ + if (!coll_entry->index_info.need_insert) + continue; + + chunk_ud.chunk_block = coll_entry->chunk_block; + chunk_ud.chunk_idx = coll_entry->index_info.chunk_idx; + chunk_ud.filter_mask = coll_entry->index_info.filter_mask; + chunk_ud.common.scaled = scaled_coords; + + /* Calculate scaled coordinates for the chunk */ + if (idx_info->layout->idx_type == H5D_CHUNK_IDX_EARRAY && idx_info->layout->u.earray.unlim_dim > 0) { + /* + * Extensible arrays where the unlimited dimension is not + * the slowest-changing dimension "swizzle" the coordinates + * to move the unlimited dimension value to offset 0. Therefore, + * we use the "swizzled" down chunks to calculate the "swizzled" + * scaled coordinates and then we undo the "swizzle" operation. + * + * TODO: In the future, this is something that should be handled + * by the particular chunk index rather than manually + * here. Likely, the chunk index ops should get a new + * callback that accepts a chunk index and provides the + * caller with the scaled coordinates for that chunk. + */ + H5VM_array_calc_pre(chunk_ud.chunk_idx, io_info->dset->shared->ndims, + idx_info->layout->u.earray.swizzled_down_chunks, scaled_coords); + + H5VM_unswizzle_coords(hsize_t, scaled_coords, idx_info->layout->u.earray.unlim_dim); + } + else { + H5VM_array_calc_pre(chunk_ud.chunk_idx, io_info->dset->shared->ndims, + io_info->dset->shared->layout.u.chunk.down_chunks, scaled_coords); + } + + scaled_coords[io_info->dset->shared->ndims] = 0; + +#ifndef NDEBUG + /* + * If a matching local chunk entry is found, the + * `chunk_info` structure (which contains the chunk's + * pre-computed scaled coordinates) will be valid + * for this rank. Compare those coordinates against + * the calculated coordinates above to make sure + * they match. + */ + for (size_t dbg_idx = 0; dbg_idx < chunk_list_num_entries; dbg_idx++) { + if (coll_entry->index_info.chunk_idx == chunk_list[dbg_idx].index_info.chunk_idx) { + hbool_t coords_match = !HDmemcmp(scaled_coords, chunk_list[dbg_idx].chunk_info->scaled, + io_info->dset->shared->ndims * sizeof(hsize_t)); + + HDassert(coords_match && "Calculated scaled coordinates for chunk didn't match " + "chunk's actual scaled coordinates!"); + break; + } + } +#endif + + if ((idx_info->storage->ops->insert)(idx_info, &chunk_ud, io_info->dset) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert chunk address into index") + } + +done: + H5MM_free(gathered_array); + H5MM_free(counts_disps_array); + + if (send_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&send_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + if (recv_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&recv_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + +#ifdef H5Dmpio_DEBUG + H5D_MPIO_TIME_STOP(mpi_rank); + H5D_MPIO_TRACE_EXIT(mpi_rank); +#endif + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_collective_filtered_chunk_reinsert() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_get_chunk_redistribute_info_types + * + * Purpose: Constructs MPI derived datatypes for communicating the + * info from a H5D_filtered_collective_io_info_t structure + * that is necessary for redistributing shared chunks during a + * collective write of filtered chunks. + * + * The datatype returned through `contig_type` has an extent + * equal to the size of an H5D_chunk_redistribute_info_t + * structure and is suitable for communicating that structure + * type. + * + * The datatype returned through `resized_type` has an extent + * equal to the size of an H5D_filtered_collective_io_info_t + * structure. This makes it suitable for sending an array of + * those structures, while extracting out just the info + * necessary for the chunk redistribution operation during + * communication. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_get_chunk_redistribute_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived, + MPI_Datatype *resized_type, hbool_t *resized_type_derived) +{ + MPI_Datatype struct_type = MPI_DATATYPE_NULL; + hbool_t struct_type_derived = FALSE; + MPI_Datatype chunk_block_type = MPI_DATATYPE_NULL; + hbool_t chunk_block_type_derived = FALSE; + MPI_Datatype types[5]; + MPI_Aint displacements[5]; + int block_lengths[5]; + int field_count; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(contig_type); + HDassert(contig_type_derived); + HDassert(resized_type); + HDassert(resized_type_derived); + + *contig_type_derived = FALSE; + *resized_type_derived = FALSE; + + /* Create struct type for the inner H5F_block_t structure */ + if (H5F_mpi_get_file_block_type(FALSE, &chunk_block_type, &chunk_block_type_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't create derived type for chunk file description") + + field_count = 5; + HDassert(field_count == (sizeof(types) / sizeof(MPI_Datatype))); + + /* + * Create structure type to pack chunk H5F_block_t structure + * next to chunk_idx, orig_owner, new_owner and num_writers + * fields + */ + block_lengths[0] = 1; + block_lengths[1] = 1; + block_lengths[2] = 1; + block_lengths[3] = 1; + block_lengths[4] = 1; + displacements[0] = offsetof(H5D_chunk_redistribute_info_t, chunk_block); + displacements[1] = offsetof(H5D_chunk_redistribute_info_t, chunk_idx); + displacements[2] = offsetof(H5D_chunk_redistribute_info_t, orig_owner); + displacements[3] = offsetof(H5D_chunk_redistribute_info_t, new_owner); + displacements[4] = offsetof(H5D_chunk_redistribute_info_t, num_writers); + types[0] = chunk_block_type; + types[1] = HSIZE_AS_MPI_TYPE; + types[2] = MPI_INT; + types[3] = MPI_INT; + types[4] = MPI_INT; + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, contig_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code) + *contig_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(contig_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + + /* Create struct type to extract the chunk_current, chunk_idx, orig_owner, + * new_owner and num_writers fields from a H5D_filtered_collective_io_info_t + * structure + */ + block_lengths[0] = 1; + block_lengths[1] = 1; + block_lengths[2] = 1; + block_lengths[3] = 1; + block_lengths[4] = 1; + displacements[0] = offsetof(H5D_filtered_collective_io_info_t, chunk_current); + displacements[1] = offsetof(H5D_filtered_collective_io_info_t, index_info.chunk_idx); + displacements[2] = offsetof(H5D_filtered_collective_io_info_t, orig_owner); + displacements[3] = offsetof(H5D_filtered_collective_io_info_t, new_owner); + displacements[4] = offsetof(H5D_filtered_collective_io_info_t, num_writers); + types[0] = chunk_block_type; + types[1] = HSIZE_AS_MPI_TYPE; + types[2] = MPI_INT; + types[3] = MPI_INT; + types[4] = MPI_INT; + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code) + struct_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized( + struct_type, 0, sizeof(H5D_filtered_collective_io_info_t), resized_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code) + *resized_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(resized_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + +done: + if (struct_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + if (chunk_block_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_block_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + + if (ret_value < 0) { + if (*resized_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(resized_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *resized_type_derived = FALSE; + } + if (*contig_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(contig_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *contig_type_derived = FALSE; + } + } + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_get_chunk_redistribute_info_types() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_get_chunk_alloc_info_types + * + * Purpose: Constructs MPI derived datatypes for communicating the info + * from a H5D_filtered_collective_io_info_t structure that is + * necessary for re-allocating file space during a collective + * write of filtered chunks. + * + * The datatype returned through `contig_type` has an extent + * equal to the size of an H5D_chunk_alloc_info_t structure + * and is suitable for communicating that structure type. + * + * The datatype returned through `resized_type` has an extent + * equal to the size of an H5D_filtered_collective_io_info_t + * structure. This makes it suitable for sending an array of + * those structures, while extracting out just the info + * necessary for the chunk file space reallocation operation + * during communication. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_get_chunk_alloc_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived, + MPI_Datatype *resized_type, hbool_t *resized_type_derived) +{ + MPI_Datatype struct_type = MPI_DATATYPE_NULL; + hbool_t struct_type_derived = FALSE; + MPI_Datatype chunk_block_type = MPI_DATATYPE_NULL; + hbool_t chunk_block_type_derived = FALSE; + MPI_Datatype types[3]; + MPI_Aint displacements[3]; + int block_lengths[3]; + int field_count; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(contig_type); + HDassert(contig_type_derived); + HDassert(resized_type); + HDassert(resized_type_derived); + + *contig_type_derived = FALSE; + *resized_type_derived = FALSE; + + /* Create struct type for the inner H5F_block_t structure */ + if (H5F_mpi_get_file_block_type(FALSE, &chunk_block_type, &chunk_block_type_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't create derived type for chunk file description") + + field_count = 3; + HDassert(field_count == (sizeof(types) / sizeof(MPI_Datatype))); + + /* + * Create structure type to pack both chunk H5F_block_t structures + * next to chunk_idx field + */ + block_lengths[0] = 1; + block_lengths[1] = 1; + block_lengths[2] = 1; + displacements[0] = offsetof(H5D_chunk_alloc_info_t, chunk_current); + displacements[1] = offsetof(H5D_chunk_alloc_info_t, chunk_new); + displacements[2] = offsetof(H5D_chunk_alloc_info_t, chunk_idx); + types[0] = chunk_block_type; + types[1] = chunk_block_type; + types[2] = HSIZE_AS_MPI_TYPE; + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, contig_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code) + *contig_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(contig_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + + /* + * Create struct type to extract the chunk_current, chunk_new and chunk_idx + * fields from a H5D_filtered_collective_io_info_t structure + */ + block_lengths[0] = 1; + block_lengths[1] = 1; + block_lengths[2] = 1; + displacements[0] = offsetof(H5D_filtered_collective_io_info_t, chunk_current); + displacements[1] = offsetof(H5D_filtered_collective_io_info_t, chunk_new); + displacements[2] = offsetof(H5D_filtered_collective_io_info_t, index_info.chunk_idx); + types[0] = chunk_block_type; + types[1] = chunk_block_type; + types[2] = HSIZE_AS_MPI_TYPE; + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code) + struct_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized( + struct_type, 0, sizeof(H5D_filtered_collective_io_info_t), resized_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code) + *resized_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(resized_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + +done: + if (struct_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + if (chunk_block_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_block_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + + if (ret_value < 0) { + if (*resized_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(resized_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *resized_type_derived = FALSE; + } + if (*contig_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(contig_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *contig_type_derived = FALSE; + } + } + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_get_chunk_alloc_info_types() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_get_chunk_insert_info_types + * + * Purpose: Constructs MPI derived datatypes for communicating the + * information necessary when reinserting chunks into a + * dataset's chunk index. This includes the chunk's new offset + * and size (H5F_block_t) and the inner `index_info` structure + * of a H5D_filtered_collective_io_info_t structure. + * + * The datatype returned through `contig_type` has an extent + * equal to the size of an H5D_chunk_insert_info_t structure + * and is suitable for communicating that structure type. + * + * The datatype returned through `resized_type` has an extent + * equal to the size of the encompassing + * H5D_filtered_collective_io_info_t structure. This makes it + * suitable for sending an array of + * H5D_filtered_collective_io_info_t structures, while + * extracting out just the information needed during + * communication. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_get_chunk_insert_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived, + MPI_Datatype *resized_type, hbool_t *resized_type_derived) +{ + MPI_Datatype struct_type = MPI_DATATYPE_NULL; + hbool_t struct_type_derived = FALSE; + MPI_Datatype chunk_block_type = MPI_DATATYPE_NULL; + hbool_t chunk_block_type_derived = FALSE; + MPI_Aint contig_type_extent; + MPI_Datatype types[4]; + MPI_Aint displacements[4]; + int block_lengths[4]; + int field_count; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(contig_type); + HDassert(contig_type_derived); + HDassert(resized_type); + HDassert(resized_type_derived); + + *contig_type_derived = FALSE; + *resized_type_derived = FALSE; + + /* Create struct type for an H5F_block_t structure */ + if (H5F_mpi_get_file_block_type(FALSE, &chunk_block_type, &chunk_block_type_derived) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't create derived type for chunk file description") + + field_count = 4; + HDassert(field_count == (sizeof(types) / sizeof(MPI_Datatype))); + + /* + * Create struct type to pack information into memory as follows: + * + * Chunk's new Offset/Size (H5F_block_t) -> + * Chunk Index Info (H5D_chunk_index_info_t) + */ + block_lengths[0] = 1; + block_lengths[1] = 1; + block_lengths[2] = 1; + block_lengths[3] = 1; + displacements[0] = offsetof(H5D_chunk_insert_info_t, chunk_block); + displacements[1] = offsetof(H5D_chunk_insert_info_t, index_info.chunk_idx); + displacements[2] = offsetof(H5D_chunk_insert_info_t, index_info.filter_mask); + displacements[3] = offsetof(H5D_chunk_insert_info_t, index_info.need_insert); + types[0] = chunk_block_type; + types[1] = HSIZE_AS_MPI_TYPE; + types[2] = MPI_UNSIGNED; + types[3] = MPI_C_BOOL; + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code) + struct_type_derived = TRUE; + + contig_type_extent = (MPI_Aint)(sizeof(H5F_block_t) + sizeof(H5D_chunk_index_info_t)); + + if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized(struct_type, 0, contig_type_extent, contig_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code) + *contig_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(contig_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + + struct_type_derived = FALSE; + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + + /* + * Create struct type to correctly extract all needed + * information from a H5D_filtered_collective_io_info_t + * structure. + */ + displacements[0] = offsetof(H5D_filtered_collective_io_info_t, chunk_new); + displacements[1] = offsetof(H5D_filtered_collective_io_info_t, index_info.chunk_idx); + displacements[2] = offsetof(H5D_filtered_collective_io_info_t, index_info.filter_mask); + displacements[3] = offsetof(H5D_filtered_collective_io_info_t, index_info.need_insert); + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code) + struct_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized( + struct_type, 0, sizeof(H5D_filtered_collective_io_info_t), resized_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code) + *resized_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(resized_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) done: - if (chunk_entry->async_info.receive_buffer_array) - H5MM_free(chunk_entry->async_info.receive_buffer_array); - if (chunk_entry->async_info.receive_requests_array) - H5MM_free(chunk_entry->async_info.receive_requests_array); - if (tmp_gath_buf) - H5MM_free(tmp_gath_buf); - if (file_iter_init && H5S_SELECT_ITER_RELEASE(file_iter) < 0) - HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator") - if (file_iter) - H5MM_free(file_iter); - if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0) - HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator") - if (mem_iter) - H5MM_free(mem_iter); - if (dataspace) - if (H5S_close(dataspace) < 0) - HDONE_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "can't close dataspace") + if (struct_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + if (chunk_block_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_block_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + } + + if (ret_value < 0) { + if (*resized_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(resized_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *resized_type_derived = FALSE; + } + if (*contig_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(contig_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *contig_type_derived = FALSE; + } + } FUNC_LEAVE_NOAPI(ret_value) -} /* end H5D__filtered_collective_chunk_entry_io() */ +} /* end H5D__mpio_get_chunk_insert_info_types() */ + +/*------------------------------------------------------------------------- + * Function: H5D__mpio_collective_filtered_io_type + * + * Purpose: Constructs a MPI derived datatype for both the memory and + * the file for a collective I/O operation on filtered chunks. + * The datatype contains the chunk offsets and lengths in the + * file and the locations of the chunk data buffers to read + * into/write from. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +static herr_t +H5D__mpio_collective_filtered_io_type(H5D_filtered_collective_io_info_t *chunk_list, size_t num_entries, + H5D_io_op_type_t op_type, MPI_Datatype *new_mem_type, + hbool_t *mem_type_derived, MPI_Datatype *new_file_type, + hbool_t *file_type_derived) +{ + MPI_Aint *io_buf_array = NULL; /* Relative displacements of filtered chunk data buffers */ + MPI_Aint *file_offset_array = NULL; /* Chunk offsets in the file */ + int * length_array = NULL; /* Filtered Chunk lengths */ + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + HDassert(chunk_list || 0 == num_entries); + HDassert(new_mem_type); + HDassert(mem_type_derived); + HDassert(new_file_type); + HDassert(file_type_derived); + + *mem_type_derived = FALSE; + *file_type_derived = FALSE; + *new_mem_type = MPI_BYTE; + *new_file_type = MPI_BYTE; + + if (num_entries > 0) { + H5F_block_t *chunk_block; + size_t last_valid_idx = 0; + size_t i; + int chunk_count; + + /* + * Determine number of chunks for I/O operation and + * setup for derived datatype creation if I/O operation + * includes multiple chunks + */ + if (num_entries == 1) { + /* Set last valid index to 0 for contiguous datatype creation */ + last_valid_idx = 0; + + if (op_type == H5D_IO_OP_WRITE) + chunk_count = 1; + else + chunk_count = chunk_list[0].need_read ? 1 : 0; + } + else { + MPI_Aint chunk_buf; + MPI_Aint base_buf; + haddr_t base_offset = HADDR_UNDEF; + + H5_CHECK_OVERFLOW(num_entries, size_t, int); + + /* Allocate arrays */ + if (NULL == (length_array = H5MM_malloc((size_t)num_entries * sizeof(int)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for filtered collective I/O length array") + if (NULL == (io_buf_array = H5MM_malloc((size_t)num_entries * sizeof(MPI_Aint)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for filtered collective I/O buf length array") + if (NULL == (file_offset_array = H5MM_malloc((size_t)num_entries * sizeof(MPI_Aint)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for filtered collective I/O offset array") + + /* + * If doing a write, we can set the base chunk offset + * and base chunk data buffer right away. + * + * If doing a read, some chunks may be skipped over + * for reading if they aren't yet allocated in the + * file. Therefore, we have to find the first chunk + * actually being read in order to set the base chunk + * offset and base chunk data buffer. + */ + if (op_type == H5D_IO_OP_WRITE) { +#if H5_CHECK_MPI_VERSION(3, 0) + if (MPI_SUCCESS != (mpi_code = MPI_Get_address(chunk_list[0].buf, &base_buf))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code) +#else + base_buf = (MPI_Aint)chunk_list[0].buf; +#endif + + base_offset = chunk_list[0].chunk_new.offset; + } + + for (i = 0, chunk_count = 0; i < num_entries; i++) { + if (op_type == H5D_IO_OP_READ) { + /* + * If this chunk isn't being read, don't add it + * to the MPI type we're building up for I/O + */ + if (!chunk_list[i].need_read) + continue; + + /* + * If this chunk is being read, go ahead and + * set the base chunk offset and base chunk + * data buffer if we haven't already + */ + if (!H5F_addr_defined(base_offset)) { +#if H5_CHECK_MPI_VERSION(3, 0) + if (MPI_SUCCESS != (mpi_code = MPI_Get_address(chunk_list[i].buf, &base_buf))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code) +#else + base_buf = (MPI_Aint)chunk_list[i].buf; +#endif + + base_offset = chunk_list[i].chunk_current.offset; + } + } + + /* Set convenience pointer for current chunk block */ + chunk_block = + (op_type == H5D_IO_OP_READ) ? &chunk_list[i].chunk_current : &chunk_list[i].chunk_new; + + /* + * Set the current chunk entry's offset in the file, relative to + * the first chunk entry + */ + HDassert(H5F_addr_defined(chunk_block->offset)); + file_offset_array[chunk_count] = (MPI_Aint)(chunk_block->offset - base_offset); + + /* + * Ensure the chunk list is sorted in ascending ordering of + * offset in the file + */ + if (chunk_count) + HDassert(file_offset_array[chunk_count] > file_offset_array[chunk_count - 1]); + + /* Set the current chunk entry's size for the I/O operation */ + H5_CHECK_OVERFLOW(chunk_block->length, hsize_t, int); + length_array[chunk_count] = (int)chunk_block->length; + + /* + * Set the displacement of the chunk entry's chunk data buffer, + * relative to the first entry's data buffer + */ +#if H5_CHECK_MPI_VERSION(3, 1) + if (MPI_SUCCESS != (mpi_code = MPI_Get_address(chunk_list[i].buf, &chunk_buf))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code) + + io_buf_array[chunk_count] = MPI_Aint_diff(chunk_buf, base_buf); +#else + chunk_buf = (MPI_Aint)chunk_list[i].buf; + io_buf_array[chunk_count] = chunk_buf - base_buf; +#endif + + /* + * Set last valid index in case only a single chunk will + * be involved in the I/O operation + */ + last_valid_idx = i; + + chunk_count++; + } /* end for */ + } + + /* + * Create derived datatypes for the chunk list if this + * rank has any chunks to work on + */ + if (chunk_count > 0) { + if (chunk_count == 1) { + int chunk_len; + + /* Single chunk - use a contiguous type for both memory and file */ + + /* Ensure that we can cast chunk size to an int for MPI */ + chunk_block = (op_type == H5D_IO_OP_READ) ? &chunk_list[last_valid_idx].chunk_current + : &chunk_list[last_valid_idx].chunk_new; + H5_CHECKED_ASSIGN(chunk_len, int, chunk_block->length, hsize_t); + + if (MPI_SUCCESS != (mpi_code = MPI_Type_contiguous(chunk_len, MPI_BYTE, new_file_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code) + *new_mem_type = *new_file_type; + + /* + * Since we use the same datatype for both memory and file, only + * mark the file type as derived so the caller doesn't try to + * free the same type twice + */ + *mem_type_derived = FALSE; + *file_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_file_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + } + else { + HDassert(file_offset_array); + HDassert(length_array); + HDassert(io_buf_array); + + /* Multiple chunks - use an hindexed type for both memory and file */ + + /* Create memory MPI type */ + if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed( + chunk_count, length_array, io_buf_array, MPI_BYTE, new_mem_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code) + *mem_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_mem_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + + /* Create file MPI type */ + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_hindexed(chunk_count, length_array, file_offset_array, + MPI_BYTE, new_file_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code) + *file_type_derived = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_file_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + } + } + } /* end if */ + +done: + if (file_offset_array) + H5MM_free(file_offset_array); + if (io_buf_array) + H5MM_free(io_buf_array); + if (length_array) + H5MM_free(length_array); + + if (ret_value < 0) { + if (*file_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(new_file_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *file_type_derived = FALSE; + } + if (*mem_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(new_mem_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *mem_type_derived = FALSE; + } + } + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_collective_filtered_io_type() */ + +#ifdef H5Dmpio_DEBUG + +static herr_t +H5D__mpio_dump_collective_filtered_chunk_list(H5D_filtered_collective_io_info_t *chunk_list, + size_t chunk_list_num_entries, int mpi_rank) +{ + H5D_filtered_collective_io_info_t *chunk_entry; + size_t i; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC_NOERR + + H5D_MPIO_DEBUG(mpi_rank, "CHUNK LIST: ["); + for (i = 0; i < chunk_list_num_entries; i++) { + unsigned chunk_rank; + + chunk_entry = &chunk_list[i]; + + HDassert(chunk_entry->chunk_info); + chunk_rank = (unsigned)H5S_GET_EXTENT_NDIMS(chunk_entry->chunk_info->fspace); + + H5D_MPIO_DEBUG(mpi_rank, " {"); + H5D_MPIO_DEBUG_VA(mpi_rank, " - Entry %zu -", i); + + H5D_MPIO_DEBUG(mpi_rank, " - Chunk Fspace Info -"); + H5D_MPIO_DEBUG_VA(mpi_rank, + " Chunk Current Info: { Offset: %" PRIuHADDR ", Length: %" PRIuHADDR " }", + chunk_entry->chunk_current.offset, chunk_entry->chunk_current.length); + H5D_MPIO_DEBUG_VA(mpi_rank, " Chunk New Info: { Offset: %" PRIuHADDR ", Length: %" PRIuHADDR " }", + chunk_entry->chunk_new.offset, chunk_entry->chunk_new.length); + + H5D_MPIO_DEBUG(mpi_rank, " - Chunk Insert Info -"); + H5D_MPIO_DEBUG_VA(mpi_rank, + " Chunk Scaled Coords (4-d): { %" PRIuHSIZE ", %" PRIuHSIZE ", %" PRIuHSIZE + ", %" PRIuHSIZE " }", + chunk_rank < 1 ? 0 : chunk_entry->chunk_info->scaled[0], + chunk_rank < 2 ? 0 : chunk_entry->chunk_info->scaled[1], + chunk_rank < 3 ? 0 : chunk_entry->chunk_info->scaled[2], + chunk_rank < 4 ? 0 : chunk_entry->chunk_info->scaled[3]); + H5D_MPIO_DEBUG_VA(mpi_rank, " Chunk Index: %" PRIuHSIZE, chunk_entry->index_info.chunk_idx); + H5D_MPIO_DEBUG_VA(mpi_rank, " Filter Mask: %u", chunk_entry->index_info.filter_mask); + H5D_MPIO_DEBUG_VA(mpi_rank, " Need Insert: %s", + chunk_entry->index_info.need_insert ? "YES" : "NO"); + + H5D_MPIO_DEBUG(mpi_rank, " - Other Info -"); + H5D_MPIO_DEBUG_VA(mpi_rank, " Chunk Info Ptr: %p", (void *)chunk_entry->chunk_info); + H5D_MPIO_DEBUG_VA(mpi_rank, " Need Read: %s", chunk_entry->need_read ? "YES" : "NO"); + H5D_MPIO_DEBUG_VA(mpi_rank, " Chunk I/O Size: %zu", chunk_entry->io_size); + H5D_MPIO_DEBUG_VA(mpi_rank, " Chunk Buffer Size: %zu", chunk_entry->chunk_buf_size); + H5D_MPIO_DEBUG_VA(mpi_rank, " Original Owner: %d", chunk_entry->orig_owner); + H5D_MPIO_DEBUG_VA(mpi_rank, " New Owner: %d", chunk_entry->new_owner); + H5D_MPIO_DEBUG_VA(mpi_rank, " # of Writers: %d", chunk_entry->num_writers); + H5D_MPIO_DEBUG_VA(mpi_rank, " Chunk Data Buffer Ptr: %p", (void *)chunk_entry->buf); + + H5D_MPIO_DEBUG(mpi_rank, " }"); + } + H5D_MPIO_DEBUG(mpi_rank, "]"); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__mpio_dump_collective_filtered_chunk_list() */ + +#endif + #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5Dpkg.h b/src/H5Dpkg.h index 49c95a5..0e0eb08 100644 --- a/src/H5Dpkg.h +++ b/src/H5Dpkg.h @@ -121,9 +121,9 @@ typedef herr_t (*H5D_layout_construct_func_t)(H5F_t *f, H5D_t *dset); typedef herr_t (*H5D_layout_init_func_t)(H5F_t *f, const H5D_t *dset, hid_t dapl_id); typedef hbool_t (*H5D_layout_is_space_alloc_func_t)(const H5O_storage_t *storage); typedef hbool_t (*H5D_layout_is_data_cached_func_t)(const H5D_shared_t *shared_dset); -typedef herr_t (*H5D_layout_io_init_func_t)(const struct H5D_io_info_t *io_info, - const H5D_type_info_t *type_info, hsize_t nelmts, - H5S_t *file_space, H5S_t *mem_space, struct H5D_chunk_map_t *cm); +typedef herr_t (*H5D_layout_io_init_func_t)(struct H5D_io_info_t *io_info, const H5D_type_info_t *type_info, + hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, + struct H5D_chunk_map_t *cm); typedef herr_t (*H5D_layout_read_func_t)(struct H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space, struct H5D_chunk_map_t *fm); @@ -222,6 +222,7 @@ typedef struct H5D_io_info_t { H5D_layout_ops_t layout_ops; /* Dataset layout I/O operation function pointers */ H5D_io_ops_t io_ops; /* I/O operation function pointers */ H5D_io_op_type_t op_type; + hbool_t use_select_io; /* Whether to use selection I/O */ union { void * rbuf; /* Pointer to buffer for read */ const void *wbuf; /* Pointer to buffer to write */ @@ -559,6 +560,7 @@ H5_DLL herr_t H5D__alloc_storage(const H5D_io_info_t *io_info, H5D_time_alloc_t hbool_t full_overwrite, hsize_t old_dim[]); H5_DLL herr_t H5D__get_storage_size(const H5D_t *dset, hsize_t *storage_size); H5_DLL herr_t H5D__get_chunk_storage_size(H5D_t *dset, const hsize_t *offset, hsize_t *storage_size); +H5_DLL herr_t H5D__chunk_index_empty(const H5D_t *dset, hbool_t *empty); H5_DLL herr_t H5D__get_num_chunks(const H5D_t *dset, const H5S_t *space, hsize_t *nchunks); H5_DLL herr_t H5D__get_chunk_info(const H5D_t *dset, const H5S_t *space, hsize_t chk_idx, hsize_t *coord, unsigned *filter_mask, haddr_t *offset, hsize_t *size); @@ -591,6 +593,10 @@ H5_DLL herr_t H5D__select_read(const H5D_io_info_t *io_info, const H5D_type_info H5_DLL herr_t H5D__select_write(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space); +/* Functions that perform direct copying between memory buffers */ +H5_DLL herr_t H5D_select_io_mem(void *dst_buf, const H5S_t *dst_space, const void *src_buf, + const H5S_t *src_space, size_t elmt_size, size_t nelmts); + /* Functions that perform scatter-gather serial I/O operations */ H5_DLL herr_t H5D__scatter_mem(const void *_tscat_buf, H5S_sel_iter_t *iter, size_t nelmts, void *_buf); H5_DLL size_t H5D__gather_mem(const void *_buf, H5S_sel_iter_t *iter, size_t nelmts, @@ -635,7 +641,13 @@ H5_DLL herr_t H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_ov const hsize_t old_dim[]); H5_DLL herr_t H5D__chunk_file_alloc(const H5D_chk_idx_info_t *idx_info, const H5F_block_t *old_chunk, H5F_block_t *new_chunk, hbool_t *need_insert, const hsize_t *scaled); +H5_DLL void * H5D__chunk_mem_alloc(size_t size, const H5O_pline_t *pline); +H5_DLL void H5D__chunk_mem_free(void *chk, const void *_pline); +H5_DLL void * H5D__chunk_mem_xfree(void *chk, const void *pline); +H5_DLL void * H5D__chunk_mem_realloc(void *chk, size_t size, const H5O_pline_t *pline); H5_DLL herr_t H5D__chunk_update_old_edge_chunks(H5D_t *dset, hsize_t old_dim[]); +H5_DLL hbool_t H5D__chunk_is_partial_edge_chunk(unsigned dset_ndims, const uint32_t *chunk_dims, + const hsize_t *chunk_scaled, const hsize_t *dset_dims); H5_DLL herr_t H5D__chunk_prune_by_extent(H5D_t *dset, const hsize_t *old_dim); H5_DLL herr_t H5D__chunk_set_sizes(H5D_t *dset); #ifdef H5_HAVE_PARALLEL @@ -694,11 +706,11 @@ H5_DLL herr_t H5D__fill_term(H5D_fill_buf_info_t *fb_info); #ifdef H5_HAVE_PARALLEL -#ifdef H5S_DEBUG +#ifdef H5D_DEBUG #ifndef H5Dmpio_DEBUG #define H5Dmpio_DEBUG #endif /*H5Dmpio_DEBUG*/ -#endif /*H5S_DEBUG*/ +#endif /*H5D_DEBUG*/ /* MPI-IO function to read, it will select either regular or irregular read */ H5_DLL herr_t H5D__mpio_select_read(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts, H5S_t *file_space, H5S_t *mem_space); @@ -727,6 +739,8 @@ H5_DLL herr_t H5D__chunk_collective_write(H5D_io_info_t *io_info, const H5D_type * memory and the file */ H5_DLL htri_t H5D__mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space, const H5S_t *mem_space, const H5D_type_info_t *type_info); +H5_DLL herr_t H5D__mpio_get_no_coll_cause_strings(char *local_cause, size_t local_cause_len, + char *global_cause, size_t global_cause_len); #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5Dpublic.h b/src/H5Dpublic.h index 75f4b95..02644ed 100644 --- a/src/H5Dpublic.h +++ b/src/H5Dpublic.h @@ -77,7 +77,7 @@ typedef enum H5D_chunk_index_t { */ typedef enum H5D_alloc_time_t { H5D_ALLOC_TIME_ERROR = -1, /**< Error */ - H5D_ALLOC_TIME_DEFAULT = 0, /**< \todo Define this! */ + H5D_ALLOC_TIME_DEFAULT = 0, /**< Default (layout dependent) */ H5D_ALLOC_TIME_EARLY = 1, /**< Allocate on creation */ H5D_ALLOC_TIME_LATE = 2, /**< Allocate on first write */ H5D_ALLOC_TIME_INCR = 3 /**< Allocate incrementally (by chunk) */ @@ -91,9 +91,9 @@ typedef enum H5D_alloc_time_t { typedef enum H5D_space_status_t { H5D_SPACE_STATUS_ERROR = -1, /**< Error */ H5D_SPACE_STATUS_NOT_ALLOCATED = 0, /**< Space has not been allocated for this dataset. */ - H5D_SPACE_STATUS_PART_ALLOCATED = 1, /**< Space has been allocated for this dataset. */ - H5D_SPACE_STATUS_ALLOCATED = 2 /**< Space has been partially allocated for this dataset. (Used only for - datasets with chunked storage.) */ + H5D_SPACE_STATUS_PART_ALLOCATED = 1, /**< Space has been partially allocated for this dataset. + (Used only for datasets with chunked storage.) */ + H5D_SPACE_STATUS_ALLOCATED = 2 /**< Space has been allocated for this dataset. */ } H5D_space_status_t; //! <!-- [H5D_space_status_t_snip] --> @@ -127,8 +127,8 @@ typedef enum H5D_fill_value_t { */ typedef enum H5D_vds_view_t { H5D_VDS_ERROR = -1, /**< Error */ - H5D_VDS_FIRST_MISSING = 0, /**< \todo Define this! */ - H5D_VDS_LAST_AVAILABLE = 1 /**< \todo Define this! */ + H5D_VDS_FIRST_MISSING = 0, /**< Include all data before the first missing mapped data */ + H5D_VDS_LAST_AVAILABLE = 1 /**< Include all available mapped data */ } H5D_vds_view_t; //! <!-- [H5D_vds_view_t_snip] --> @@ -682,8 +682,7 @@ H5_DLL herr_t H5Dget_chunk_info_by_coord(hid_t dset_id, const hsize_t *offset, u * Iterate over all chunked datasets and chunks in a file. * \snippet H5D_examples.c H5Ovisit_cb * - * \version 1.?.? - * \todo When was this function introduced? + * \since 1.13.0 * */ H5_DLL herr_t H5Dchunk_iter(hid_t dset_id, hid_t dxpl_id, H5D_chunk_iter_op_t cb, void *op_data); diff --git a/src/H5Dselect.c b/src/H5Dselect.c index e64d657..f464ca5 100644 --- a/src/H5Dselect.c +++ b/src/H5Dselect.c @@ -105,6 +105,9 @@ H5D__select_io(const H5D_io_info_t *io_info, size_t elmt_size, size_t nelmts, H5 HDassert(io_info->store); HDassert(io_info->u.rbuf); + if (elmt_size == 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_BADVALUE, FAIL, "invalid elmt_size of 0") + /* Check for only one element in selection */ if (nelmts == 1) { hsize_t single_mem_off; /* Offset in memory */ @@ -226,8 +229,6 @@ H5D__select_io(const H5D_io_info_t *io_info, size_t elmt_size, size_t nelmts, H5 /* Decrement number of elements left to process */ HDassert(((size_t)tmp_file_len % elmt_size) == 0); - if (elmt_size == 0) - HGOTO_ERROR(H5E_DATASPACE, H5E_BADVALUE, FAIL, "Resulted in division by zero") nelmts -= ((size_t)tmp_file_len / elmt_size); } /* end while */ } /* end else */ @@ -257,6 +258,188 @@ done: } /* end H5D__select_io() */ /*------------------------------------------------------------------------- + * Function: H5D_select_io_mem + * + * Purpose: Perform memory copies directly between two memory buffers + * according to the selections in the `dst_space` and + * `src_space` dataspaces. + * + * Note: This routine is [basically] the same as H5D__select_io, + * with the only difference being that the readvv/writevv + * calls are exchanged for H5VM_memcpyvv calls. Changes should + * be made to both routines. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +herr_t +H5D_select_io_mem(void *dst_buf, const H5S_t *dst_space, const void *src_buf, const H5S_t *src_space, + size_t elmt_size, size_t nelmts) +{ + H5S_sel_iter_t *dst_sel_iter = NULL; /* Destination dataspace iteration info */ + H5S_sel_iter_t *src_sel_iter = NULL; /* Source dataspace iteration info */ + hbool_t dst_sel_iter_init = FALSE; /* Destination dataspace selection iterator initialized? */ + hbool_t src_sel_iter_init = FALSE; /* Source dataspace selection iterator initialized? */ + hsize_t * dst_off = NULL; /* Pointer to sequence offsets in destination buffer */ + hsize_t * src_off = NULL; /* Pointer to sequence offsets in source buffer */ + size_t * dst_len = NULL; /* Pointer to sequence lengths in destination buffer */ + size_t * src_len = NULL; /* Pointer to sequence lengths in source buffer */ + size_t curr_dst_seq; /* Current destination buffer sequence to operate on */ + size_t curr_src_seq; /* Current source buffer sequence to operate on */ + size_t dst_nseq; /* Number of sequences generated for destination buffer */ + size_t src_nseq; /* Number of sequences generated for source buffer */ + size_t dxpl_vec_size; /* Vector length from API context's DXPL */ + size_t vec_size; /* Vector length */ + ssize_t bytes_copied; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_NOAPI(FAIL) + + HDassert(dst_buf); + HDassert(dst_space); + HDassert(src_buf); + HDassert(src_space); + + if (elmt_size == 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_BADVALUE, FAIL, "invalid elmt_size of 0") + + /* Check for only one element in selection */ + if (nelmts == 1) { + hsize_t single_dst_off; /* Offset in dst_space */ + hsize_t single_src_off; /* Offset in src_space */ + size_t single_dst_len; /* Length in dst_space */ + size_t single_src_len; /* Length in src_space */ + + /* Get offset of first element in selections */ + if (H5S_SELECT_OFFSET(dst_space, &single_dst_off) < 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "can't retrieve destination selection offset") + if (H5S_SELECT_OFFSET(src_space, &single_src_off) < 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "can't retrieve source selection offset") + + /* Set up necessary information for I/O operation */ + dst_nseq = src_nseq = 1; + curr_dst_seq = curr_src_seq = 0; + single_dst_off *= elmt_size; + single_src_off *= elmt_size; + single_dst_len = single_src_len = elmt_size; + + /* Perform vectorized memcpy from src_buf to dst_buf */ + if ((bytes_copied = + H5VM_memcpyvv(dst_buf, dst_nseq, &curr_dst_seq, &single_dst_len, &single_dst_off, src_buf, + src_nseq, &curr_src_seq, &single_src_len, &single_src_off)) < 0) + HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "vectorized memcpy failed") + + HDassert(((size_t)bytes_copied % elmt_size) == 0); + } + else { + unsigned sel_iter_flags = H5S_SEL_ITER_GET_SEQ_LIST_SORTED | H5S_SEL_ITER_SHARE_WITH_DATASPACE; + size_t dst_nelem; /* Number of elements used in destination buffer sequences */ + size_t src_nelem; /* Number of elements used in source buffer sequences */ + + /* Get info from API context */ + if (H5CX_get_vec_size(&dxpl_vec_size) < 0) + HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "can't retrieve I/O vector size") + + /* Allocate the vector I/O arrays */ + if (dxpl_vec_size > H5D_IO_VECTOR_SIZE) + vec_size = dxpl_vec_size; + else + vec_size = H5D_IO_VECTOR_SIZE; + + if (NULL == (dst_len = H5FL_SEQ_MALLOC(size_t, vec_size))) + HGOTO_ERROR(H5E_IO, H5E_CANTALLOC, FAIL, "can't allocate I/O length vector array") + if (NULL == (dst_off = H5FL_SEQ_MALLOC(hsize_t, vec_size))) + HGOTO_ERROR(H5E_IO, H5E_CANTALLOC, FAIL, "can't allocate I/O offset vector array") + if (NULL == (src_len = H5FL_SEQ_MALLOC(size_t, vec_size))) + HGOTO_ERROR(H5E_IO, H5E_CANTALLOC, FAIL, "can't allocate I/O length vector array") + if (NULL == (src_off = H5FL_SEQ_MALLOC(hsize_t, vec_size))) + HGOTO_ERROR(H5E_IO, H5E_CANTALLOC, FAIL, "can't allocate I/O offset vector array") + + /* Allocate the dataspace selection iterators */ + if (NULL == (dst_sel_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTALLOC, FAIL, "can't allocate destination selection iterator") + if (NULL == (src_sel_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTALLOC, FAIL, "can't allocate source selection iterator") + + /* Initialize destination selection iterator */ + if (H5S_select_iter_init(dst_sel_iter, dst_space, elmt_size, sel_iter_flags) < 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator") + dst_sel_iter_init = TRUE; /* Destination selection iteration info has been initialized */ + + /* Initialize source selection iterator */ + if (H5S_select_iter_init(src_sel_iter, src_space, elmt_size, H5S_SEL_ITER_SHARE_WITH_DATASPACE) < 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTINIT, FAIL, "unable to initialize selection iterator") + src_sel_iter_init = TRUE; /* Source selection iteration info has been initialized */ + + /* Initialize sequence counts */ + curr_dst_seq = curr_src_seq = 0; + dst_nseq = src_nseq = 0; + + /* Loop, until all bytes are processed */ + while (nelmts > 0) { + /* Check if more destination buffer sequences are needed */ + if (curr_dst_seq >= dst_nseq) { + /* Get sequences for destination selection */ + if (H5S_SELECT_ITER_GET_SEQ_LIST(dst_sel_iter, vec_size, nelmts, &dst_nseq, &dst_nelem, + dst_off, dst_len) < 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "sequence length generation failed") + + /* Start at the beginning of the sequences again */ + curr_dst_seq = 0; + } + + /* Check if more source buffer sequences are needed */ + if (curr_src_seq >= src_nseq) { + /* Get sequences for source selection */ + if (H5S_SELECT_ITER_GET_SEQ_LIST(src_sel_iter, vec_size, nelmts, &src_nseq, &src_nelem, + src_off, src_len) < 0) + HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "sequence length generation failed") + + /* Start at the beginning of the sequences again */ + curr_src_seq = 0; + } /* end if */ + + /* Perform vectorized memcpy from src_buf to dst_buf */ + if ((bytes_copied = H5VM_memcpyvv(dst_buf, dst_nseq, &curr_dst_seq, dst_len, dst_off, src_buf, + src_nseq, &curr_src_seq, src_len, src_off)) < 0) + HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "vectorized memcpy failed") + + /* Decrement number of elements left to process */ + HDassert(((size_t)bytes_copied % elmt_size) == 0); + nelmts -= ((size_t)bytes_copied / elmt_size); + } + } + +done: + /* Release selection iterators */ + if (src_sel_iter) { + if (src_sel_iter_init && H5S_SELECT_ITER_RELEASE(src_sel_iter) < 0) + HDONE_ERROR(H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator") + + src_sel_iter = H5FL_FREE(H5S_sel_iter_t, src_sel_iter); + } + if (dst_sel_iter) { + if (dst_sel_iter_init && H5S_SELECT_ITER_RELEASE(dst_sel_iter) < 0) + HDONE_ERROR(H5E_DATASPACE, H5E_CANTRELEASE, FAIL, "unable to release selection iterator") + + dst_sel_iter = H5FL_FREE(H5S_sel_iter_t, dst_sel_iter); + } + + /* Release vector arrays, if allocated */ + if (src_off) + src_off = H5FL_SEQ_FREE(hsize_t, src_off); + if (src_len) + src_len = H5FL_SEQ_FREE(size_t, src_len); + if (dst_off) + dst_off = H5FL_SEQ_FREE(hsize_t, dst_off); + if (dst_len) + dst_len = H5FL_SEQ_FREE(size_t, dst_len); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D_select_io_mem() */ + +/*------------------------------------------------------------------------- * Function: H5D__select_read * * Purpose: Reads directly from file into application memory. diff --git a/src/H5EAdbg.c b/src/H5EAdbg.c index b0e564c..b377422 100644 --- a/src/H5EAdbg.c +++ b/src/H5EAdbg.c @@ -237,7 +237,7 @@ H5EA__iblock_debug(H5F_t *f, haddr_t H5_ATTR_UNUSED addr, FILE *stream, int inde HDfprintf(stream, "%*sData Block Addresses in Index Block:\n", indent, ""); for (u = 0; u < iblock->ndblk_addrs; u++) { /* Print address */ - HDsprintf(temp_str, "Address #%u:", u); + HDsnprintf(temp_str, sizeof(temp_str), "Address #%u:", u); HDfprintf(stream, "%*s%-*s %" PRIuHADDR "\n", (indent + 3), "", MAX(0, (fwidth - 3)), temp_str, iblock->dblk_addrs[u]); } /* end for */ @@ -252,7 +252,7 @@ H5EA__iblock_debug(H5F_t *f, haddr_t H5_ATTR_UNUSED addr, FILE *stream, int inde HDfprintf(stream, "%*sSuper Block Addresses in Index Block:\n", indent, ""); for (u = 0; u < iblock->nsblk_addrs; u++) { /* Print address */ - HDsprintf(temp_str, "Address #%u:", u); + HDsnprintf(temp_str, sizeof(temp_str), "Address #%u:", u); HDfprintf(stream, "%*s%-*s %" PRIuHADDR "\n", (indent + 3), "", MAX(0, (fwidth - 3)), temp_str, iblock->sblk_addrs[u]); } /* end for */ @@ -341,7 +341,7 @@ H5EA__sblock_debug(H5F_t *f, haddr_t addr, FILE *stream, int indent, int fwidth, HDfprintf(stream, "%*sData Block Addresses in Super Block:\n", indent, ""); for (u = 0; u < sblock->ndblks; u++) { /* Print address */ - HDsprintf(temp_str, "Address #%u:", u); + HDsnprintf(temp_str, sizeof(temp_str), "Address #%u:", u); HDfprintf(stream, "%*s%-*s %" PRIuHADDR "\n", (indent + 3), "", MAX(0, (fwidth - 3)), temp_str, sblock->dblk_addrs[u]); } /* end for */ diff --git a/src/H5EAprivate.h b/src/H5EAprivate.h index 19dabd9..9481559f 100644 --- a/src/H5EAprivate.h +++ b/src/H5EAprivate.h @@ -26,11 +26,6 @@ #ifndef H5EAprivate_H #define H5EAprivate_H -/* Include package's public header */ -#ifdef NOT_YET -#include "H5EApublic.h" -#endif /* NOT_YET */ - /* Private headers needed by this file */ #include "H5ACprivate.h" /* Metadata cache */ #include "H5Fprivate.h" /* File access */ diff --git a/src/H5EAtest.c b/src/H5EAtest.c index 7924eaa..24efbc2 100644 --- a/src/H5EAtest.c +++ b/src/H5EAtest.c @@ -322,7 +322,7 @@ H5EA__test_debug(FILE *stream, int indent, int fwidth, hsize_t idx, const void * HDassert(elmt); /* Print element */ - HDsprintf(temp_str, "Element #%llu:", (unsigned long long)idx); + HDsnprintf(temp_str, sizeof(temp_str), "Element #%llu:", (unsigned long long)idx); HDfprintf(stream, "%*s%-*s %llu\n", indent, "", fwidth, temp_str, (unsigned long long)*(const uint64_t *)elmt); @@ -236,6 +236,61 @@ done: } /* end H5ESget_op_counter() */ /*------------------------------------------------------------------------- + * Function: H5ESget_requests + * + * Purpose: Retrieve the requests in an event set. Up to *count + * requests are stored in the provided requests array, and + * the connector ids corresponding to these requests are + * stored in the provided connector_ids array. Either or + * both of these arrays may be NULL, in which case this + * information is not returned. If these arrays are + * non-NULL, they must be large enough to contain *count + * entries. On exit, *count is set to the total number of + * events in the event set. + * + * Events are returned in the order they were added to the + * event set. If order is H5_ITER_INC or H5_ITER_NATIVE, + * events will be returned starting from the oldest. If order + * is H5_ITER_DEC, events will be returned starting with the + * newest/most recent. + * + * Return: SUCCEED / FAIL + * + * Programmer: Neil Fortner + * Tuesday, November 23, 2021 + * + *------------------------------------------------------------------------- + */ +herr_t +H5ESget_requests(hid_t es_id, H5_iter_order_t order, hid_t *connector_ids, void **requests, size_t array_len, + size_t *count /*out*/) +{ + H5ES_t *es; /* Event set */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_API(FAIL) + H5TRACE6("e", "iIo*i**xzx", es_id, order, connector_ids, requests, array_len, count); + + /* Check arguments */ + if (NULL == (es = H5I_object_verify(es_id, H5I_EVENTSET))) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "invalid event set identifier") + if (order <= H5_ITER_UNKNOWN || order >= H5_ITER_N) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "invalid iteration order specified") + + /* Call internal routine */ + if (array_len > 0 && (requests || connector_ids)) + if (H5ES__get_requests(es, order, connector_ids, requests, array_len) < 0) + HGOTO_ERROR(H5E_EVENTSET, H5E_CANTGET, FAIL, "can't get requests") + + /* Retrieve the count, if non-NULL */ + if (count) + *count = H5ES__list_count(&es->active); + +done: + FUNC_LEAVE_API(ret_value) +} /* end H5ESget_requests() */ + +/*------------------------------------------------------------------------- * Function: H5ESwait * * Purpose: Wait (with timeout) for operations in event set to complete diff --git a/src/H5ESdevelop.h b/src/H5ESdevelop.h index 5a0f2b4..2fb9aeb 100644 --- a/src/H5ESdevelop.h +++ b/src/H5ESdevelop.h @@ -42,6 +42,8 @@ extern "C" { #endif H5_DLL herr_t H5ESinsert_request(hid_t es_id, hid_t connector_id, void *request); +H5_DLL herr_t H5ESget_requests(hid_t es_id, H5_iter_order_t order, hid_t *connector_ids, void **requests, + size_t array_len, size_t *count); #ifdef __cplusplus } diff --git a/src/H5ESint.c b/src/H5ESint.c index c66be16..7eb5909 100644 --- a/src/H5ESint.c +++ b/src/H5ESint.c @@ -50,6 +50,14 @@ /* Local Typedefs */ /******************/ +/* Callback context for get events operations */ +typedef struct H5ES_get_requests_ctx_t { + hid_t *connector_ids; /* Output buffer for list of connector IDs that match the above requests */ + void **requests; /* Output buffer for list of requests in event set */ + size_t array_len; /* Length of the above output buffers */ + size_t i; /* Number of elements filled in output buffers */ +} H5ES_get_requests_ctx_t; + /* Callback context for wait operations */ typedef struct H5ES_wait_ctx_t { H5ES_t * es; /* Event set being operated on */ @@ -84,6 +92,7 @@ static herr_t H5ES__close(H5ES_t *es); static herr_t H5ES__close_cb(void *es, void **request_token); static herr_t H5ES__insert(H5ES_t *es, H5VL_t *connector, void *request_token, const char *app_file, const char *app_func, unsigned app_line, const char *caller, const char *api_args); +static int H5ES__get_requests_cb(H5ES_event_t *ev, void *_ctx); static herr_t H5ES__handle_fail(H5ES_t *es, H5ES_event_t *ev); static herr_t H5ES__op_complete(H5ES_t *es, H5ES_event_t *ev, H5VL_request_status_t ev_status); static int H5ES__wait_cb(H5ES_event_t *ev, void *_ctx); @@ -282,7 +291,8 @@ H5ES__insert(H5ES_t *es, H5VL_t *connector, void *request_token, const char *app * there's no need to duplicate it. */ ev->op_info.api_name = caller; - if (NULL == (ev->op_info.api_args = H5MM_xstrdup(api_args))) + HDassert(ev->op_info.api_args == NULL); + if (api_args && NULL == (ev->op_info.api_args = H5MM_xstrdup(api_args))) HGOTO_ERROR(H5E_EVENTSET, H5E_CANTALLOC, FAIL, "can't copy API routine arguments") /* Append fully initialized event onto the event set's 'active' list */ @@ -419,6 +429,86 @@ done: } /* end H5ES__insert_request() */ /*------------------------------------------------------------------------- + * Function: H5ES__get_requests_cb + * + * Purpose: Iterator callback for H5ES__get_events - adds the event to + * the list. + * + * Return: SUCCEED / FAIL + * + * Programmer: Neil Fortner + * Tuesday, November 23, 2021 + * + *------------------------------------------------------------------------- + */ +static int +H5ES__get_requests_cb(H5ES_event_t *ev, void *_ctx) +{ + H5ES_get_requests_ctx_t *ctx = (H5ES_get_requests_ctx_t *)_ctx; /* Callback context */ + int ret_value = H5_ITER_CONT; /* Return value */ + + FUNC_ENTER_STATIC_NOERR + + /* Sanity check */ + HDassert(ev); + HDassert(ctx); + HDassert(ctx->i < ctx->array_len); + + /* Get the connector ID for the event */ + if (ctx->connector_ids) + ctx->connector_ids[ctx->i] = ev->request->connector->id; + + /* Get the request for the event */ + if (ctx->requests) + ctx->requests[ctx->i] = ev->request->data; + + /* Check if we've run out of room in the arrays */ + if (++ctx->i == ctx->array_len) + ret_value = H5_ITER_STOP; + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5ES__get_requests_cb() */ + +/*------------------------------------------------------------------------- + * Function: H5ES__get_requests + * + * Purpose: Get all requests in an event set. + * + * Return: SUCCEED / FAIL + * + * Programmer: Neil Fortner + * Tuesday, November 23, 2021 + * + *------------------------------------------------------------------------- + */ +herr_t +H5ES__get_requests(H5ES_t *es, H5_iter_order_t order, hid_t *connector_ids, void **requests, size_t array_len) +{ + H5ES_get_requests_ctx_t ctx; /* Callback context */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Sanity check */ + HDassert(es); + HDassert(array_len > 0); + HDassert(requests || connector_ids); + + /* Set up context for iterator callbacks */ + ctx.connector_ids = connector_ids; + ctx.requests = requests; + ctx.array_len = array_len; + ctx.i = 0; + + /* Iterate over the events in the set */ + if (H5ES__list_iterate(&es->active, order, H5ES__get_requests_cb, &ctx) < 0) + HGOTO_ERROR(H5E_EVENTSET, H5E_BADITER, FAIL, "iteration failed") + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5ES__get_requests() */ + +/*------------------------------------------------------------------------- * Function: H5ES__handle_fail * * Purpose: Handle a failed event @@ -661,7 +751,7 @@ H5ES__wait(H5ES_t *es, uint64_t timeout, size_t *num_in_progress, hbool_t *op_fa ctx.op_failed = op_failed; /* Iterate over the events in the set, waiting for them to complete */ - if (H5ES__list_iterate(&es->active, H5ES__wait_cb, &ctx) < 0) + if (H5ES__list_iterate(&es->active, H5_ITER_NATIVE, H5ES__wait_cb, &ctx) < 0) HGOTO_ERROR(H5E_EVENTSET, H5E_BADITER, FAIL, "iteration failed") done: @@ -769,7 +859,7 @@ H5ES__cancel(H5ES_t *es, size_t *num_not_canceled, hbool_t *op_failed) ctx.op_failed = op_failed; /* Iterate over the events in the set, attempting to cancel them */ - if (H5ES__list_iterate(&es->active, H5ES__cancel_cb, &ctx) < 0) + if (H5ES__list_iterate(&es->active, H5_ITER_NATIVE, H5ES__cancel_cb, &ctx) < 0) HGOTO_ERROR(H5E_EVENTSET, H5E_BADITER, FAIL, "iteration failed") done: @@ -806,13 +896,13 @@ H5ES__get_err_info_cb(H5ES_event_t *ev, void *_ctx) * so there's no need to duplicate them internally, but they are duplicated * here, when they are given back to the user. */ - if (NULL == (ctx->curr_err_info->api_name = H5MM_strdup(ev->op_info.api_name))) + if (NULL == (ctx->curr_err_info->api_name = H5MM_xstrdup(ev->op_info.api_name))) HGOTO_ERROR(H5E_EVENTSET, H5E_CANTALLOC, H5_ITER_ERROR, "can't copy HDF5 API routine name") - if (NULL == (ctx->curr_err_info->api_args = H5MM_strdup(ev->op_info.api_args))) + if (NULL == (ctx->curr_err_info->api_args = H5MM_xstrdup(ev->op_info.api_args))) HGOTO_ERROR(H5E_EVENTSET, H5E_CANTALLOC, H5_ITER_ERROR, "can't copy HDF5 API routine arguments") - if (NULL == (ctx->curr_err_info->app_file_name = H5MM_strdup(ev->op_info.app_file_name))) + if (NULL == (ctx->curr_err_info->app_file_name = H5MM_xstrdup(ev->op_info.app_file_name))) HGOTO_ERROR(H5E_EVENTSET, H5E_CANTALLOC, H5_ITER_ERROR, "can't copy HDF5 application file name") - if (NULL == (ctx->curr_err_info->app_func_name = H5MM_strdup(ev->op_info.app_func_name))) + if (NULL == (ctx->curr_err_info->app_func_name = H5MM_xstrdup(ev->op_info.app_func_name))) HGOTO_ERROR(H5E_EVENTSET, H5E_CANTALLOC, H5_ITER_ERROR, "can't copy HDF5 application function name") ctx->curr_err_info->app_line_num = ev->op_info.app_line_num; ctx->curr_err_info->op_ins_count = ev->op_info.op_ins_count; @@ -883,7 +973,7 @@ H5ES__get_err_info(H5ES_t *es, size_t num_err_info, H5ES_err_info_t err_info[], ctx.curr_err_info = &err_info[0]; /* Iterate over the failed events in the set, copying their error info */ - if (H5ES__list_iterate(&es->failed, H5ES__get_err_info_cb, &ctx) < 0) + if (H5ES__list_iterate(&es->failed, H5_ITER_NATIVE, H5ES__get_err_info_cb, &ctx) < 0) HGOTO_ERROR(H5E_EVENTSET, H5E_BADITER, FAIL, "iteration failed") /* Set # of failed events cleared from event set's failed list */ @@ -957,7 +1047,7 @@ H5ES__close(H5ES_t *es) "can't close event set while unfinished operations are present (i.e. wait on event set first)") /* Iterate over the failed events in the set, releasing them */ - if (H5ES__list_iterate(&es->failed, H5ES__close_failed_cb, (void *)es) < 0) + if (H5ES__list_iterate(&es->failed, H5_ITER_NATIVE, H5ES__close_failed_cb, (void *)es) < 0) HGOTO_ERROR(H5E_EVENTSET, H5E_BADITER, FAIL, "iteration failed") /* Release the event set */ diff --git a/src/H5ESlist.c b/src/H5ESlist.c index 3180322..61a9dd1 100644 --- a/src/H5ESlist.c +++ b/src/H5ESlist.c @@ -135,7 +135,10 @@ H5ES__list_count(const H5ES_event_list_t *el) * each event. * * Note: Iteration is safe for deleting the current event. Modifying - * the list in other ways is likely unsafe. + * the list in other ways is likely unsafe. If order is + * H5_ITER_INC or H5_ITER_NATIVE events are visited starting + * with the oldest, otherwise they are visited starting with + * the newest. * * Return: SUCCEED / FAIL * @@ -145,7 +148,7 @@ H5ES__list_count(const H5ES_event_list_t *el) *------------------------------------------------------------------------- */ int -H5ES__list_iterate(H5ES_event_list_t *el, H5ES_list_iter_func_t cb, void *ctx) +H5ES__list_iterate(H5ES_event_list_t *el, H5_iter_order_t order, H5ES_list_iter_func_t cb, void *ctx) { H5ES_event_t *ev; /* Event in list */ int ret_value = H5_ITER_CONT; /* Return value */ @@ -157,12 +160,12 @@ H5ES__list_iterate(H5ES_event_list_t *el, H5ES_list_iter_func_t cb, void *ctx) HDassert(cb); /* Iterate over events in list */ - ev = el->head; + ev = (order == H5_ITER_DEC) ? el->tail : el->head; while (ev) { H5ES_event_t *tmp; /* Temporary event */ /* Get pointer to next node, so it's safe if this one is removed */ - tmp = ev->next; + tmp = (order == H5_ITER_DEC) ? ev->prev : ev->next; /* Perform iterator callback */ if ((ret_value = (*cb)(ev, ctx)) != H5_ITER_CONT) { diff --git a/src/H5ESpkg.h b/src/H5ESpkg.h index a7a8e20..6ee50fa 100644 --- a/src/H5ESpkg.h +++ b/src/H5ESpkg.h @@ -81,6 +81,8 @@ typedef int (*H5ES_list_iter_func_t)(H5ES_event_t *ev, void *ctx); H5_DLL H5ES_t *H5ES__create(void); H5_DLL herr_t H5ES__insert_request(H5ES_t *es, H5VL_t *connector, void *token); H5_DLL herr_t H5ES__wait(H5ES_t *es, uint64_t timeout, size_t *num_in_progress, hbool_t *op_failed); +H5_DLL herr_t H5ES__get_requests(H5ES_t *es, H5_iter_order_t order, hid_t *connector_ids, void **requests, + size_t array_len); H5_DLL herr_t H5ES__cancel(H5ES_t *es, size_t *num_not_canceled, hbool_t *op_failed); H5_DLL herr_t H5ES__get_err_info(H5ES_t *es, size_t num_err_info, H5ES_err_info_t err_info[], size_t *num_cleared); @@ -88,7 +90,8 @@ H5_DLL herr_t H5ES__get_err_info(H5ES_t *es, size_t num_err_info, H5ES_err_info /* Event list operations */ H5_DLL void H5ES__list_append(H5ES_event_list_t *el, H5ES_event_t *ev); H5_DLL size_t H5ES__list_count(const H5ES_event_list_t *el); -H5_DLL int H5ES__list_iterate(H5ES_event_list_t *el, H5ES_list_iter_func_t cb, void *ctx); +H5_DLL int H5ES__list_iterate(H5ES_event_list_t *el, H5_iter_order_t order, H5ES_list_iter_func_t cb, + void *ctx); H5_DLL void H5ES__list_remove(H5ES_event_list_t *el, const H5ES_event_t *ev); /* Event operations */ diff --git a/src/H5ESpublic.h b/src/H5ESpublic.h index c8696b3..c8d1c7b 100644 --- a/src/H5ESpublic.h +++ b/src/H5ESpublic.h @@ -200,7 +200,18 @@ H5_DLL herr_t H5ESget_count(hid_t es_id, size_t *count); /** * \ingroup H5ES * - * \todo Fill in the blanks! + * \brief Retrieves the next operation counter to be assigned in an event set + * + * \es_id + * \param[out] counter The next counter value to be assigned to an event + * \returns \herr_t + * + * \details H5ESget_op_counter() retrieves the \p counter that will be assigned + * to the next operation inserted into the event set \p es_id. + * + * \note This is designed for wrapper libraries mainly, to use as a mechanism + * for matching operations inserted into the event set with possible + * errors that occur. * * \since 1.13.0 * diff --git a/src/H5FAdblkpage.c b/src/H5FAdblkpage.c index 713bd67..f6a5aef 100644 --- a/src/H5FAdblkpage.c +++ b/src/H5FAdblkpage.c @@ -147,7 +147,7 @@ H5FA__dblk_page_create(H5FA_hdr_t *hdr, haddr_t addr, size_t nelmts) FUNC_ENTER_PACKAGE #ifdef H5FA_DEBUG - HDfprintf(stderr, "%s: Called, addr = %a\n", __func__, addr); + HDfprintf(stderr, "%s: Called, addr = %" PRIuHADDR "\n", __func__, addr); #endif /* H5FA_DEBUG */ /* Sanity check */ diff --git a/src/H5FAprivate.h b/src/H5FAprivate.h index 26057bf..745c129 100644 --- a/src/H5FAprivate.h +++ b/src/H5FAprivate.h @@ -24,11 +24,6 @@ #ifndef H5FAprivate_H #define H5FAprivate_H -/* Include package's public header */ -#ifdef NOT_YET -#include "H5FApublic.h" -#endif /* NOT_YET */ - /* Private headers needed by this file */ #include "H5ACprivate.h" /* Metadata cache */ #include "H5Fprivate.h" /* File access */ @@ -134,7 +129,7 @@ H5_DLL herr_t H5FA_patch_file(H5FA_t *fa, H5F_t *f); H5_DLL herr_t H5FA_get_stats(const H5FA_t *ea, H5FA_stat_t *stats); /* Debugging routines */ -#ifdef H5FA_DEBUGGING -#endif /* H5FA_DEBUGGING */ +#ifdef H5FA_DEBUG +#endif /* H5FA_DEBUG */ #endif /* H5FAprivate_H */ diff --git a/src/H5FAtest.c b/src/H5FAtest.c index 384a657..b57f562 100644 --- a/src/H5FAtest.c +++ b/src/H5FAtest.c @@ -303,7 +303,7 @@ H5FA__test_debug(FILE *stream, int indent, int fwidth, hsize_t idx, const void * HDassert(elmt); /* Print element */ - HDsprintf(temp_str, "Element #%llu:", (unsigned long long)idx); + HDsnprintf(temp_str, sizeof(temp_str), "Element #%llu:", (unsigned long long)idx); HDfprintf(stream, "%*s%-*s %llu\n", indent, "", fwidth, temp_str, (unsigned long long)*(const uint64_t *)elmt); @@ -214,6 +214,8 @@ H5FDregister(const H5FD_class_t *cls) /* Check arguments */ if (!cls) HGOTO_ERROR(H5E_ARGS, H5E_UNINITIALIZED, H5I_INVALID_HID, "null class pointer is disallowed") + if (cls->version != H5FD_CLASS_VERSION) + HGOTO_ERROR(H5E_ARGS, H5E_VERSION, H5I_INVALID_HID, "wrong file driver version #") if (!cls->open || !cls->close) HGOTO_ERROR(H5E_ARGS, H5E_UNINITIALIZED, H5I_INVALID_HID, "'open' and/or 'close' methods are not defined") @@ -928,9 +930,10 @@ H5FD_cmp(const H5FD_t *f1, const H5FD_t *f2) { int ret_value = -1; /* Return value */ - FUNC_ENTER_NOAPI_NOERR /* return value is arbitrary */ + FUNC_ENTER_NOAPI_NOERR; /* return value is arbitrary */ - if ((!f1 || !f1->cls) && (!f2 || !f2->cls)) HGOTO_DONE(0) + if ((!f1 || !f1->cls) && (!f2 || !f2->cls)) + HGOTO_DONE(0) if (!f1 || !f1->cls) HGOTO_DONE(-1) if (!f2 || !f2->cls) @@ -1479,6 +1482,370 @@ done: } /* end H5FDwrite() */ /*------------------------------------------------------------------------- + * Function: H5FDread_vector + * + * Purpose: Perform count reads from the specified file at the offsets + * provided in the addrs array, with the lengths and memory + * types provided in the sizes and types arrays. Data read + * is returned in the buffers provided in the bufs array. + * + * All reads are done according to the data transfer property + * list dxpl_id (which may be the constant H5P_DEFAULT). + * + * Return: Success: SUCCEED + * All reads have completed successfully, and + * the results havce been into the supplied + * buffers. + * + * Failure: FAIL + * The contents of supplied buffers are undefined. + * + * Programmer: JRM -- 6/10/20 + * + * Changes: None. + * + *------------------------------------------------------------------------- + */ +herr_t +H5FDread_vector(H5FD_t *file, hid_t dxpl_id, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], + size_t sizes[], void *bufs[] /* out */) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_API(FAIL) + H5TRACE7("e", "*#iIu*Mt*a*zx", file, dxpl_id, count, types, addrs, sizes, bufs); + + /* Check arguments */ + if (!file) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file pointer cannot be NULL") + + if (!file->cls) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file class pointer cannot be NULL") + + if ((!types) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "types parameter can't be NULL if count is positive") + + if ((!addrs) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "addrs parameter can't be NULL if count is positive") + + if ((!sizes) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "sizes parameter can't be NULL if count is positive") + + if ((!bufs) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "bufs parameter can't be NULL if count is positive") + + if ((count > 0) && (sizes[0] == 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "sizes[0] can't be 0") + + if ((count > 0) && (types[0] == H5FD_MEM_NOLIST)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "count[0] can't be H5FD_MEM_NOLIST") + + /* Get the default dataset transfer property list if the user + * didn't provide one + */ + if (H5P_DEFAULT == dxpl_id) { + dxpl_id = H5P_DATASET_XFER_DEFAULT; + } + else { + if (TRUE != H5P_isa_class(dxpl_id, H5P_DATASET_XFER)) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a data transfer property list") + } + + /* Set DXPL for operation */ + H5CX_set_dxpl(dxpl_id); + + /* Call private function */ + /* (Note compensating for base addresses addition in internal routine) */ + if (H5FD_read_vector(file, count, types, addrs, sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "file vector read request failed") + +done: + FUNC_LEAVE_API(ret_value) +} /* end H5FDread_vector() */ + +/*------------------------------------------------------------------------- + * Function: H5FDwrite_vector + * + * Purpose: Perform count writes to the specified file at the offsets + * provided in the addrs array, with the lengths and memory + * types provided in the sizes and types arrays. Data to be + * written is in the buffers provided in the bufs array. + * + * All writes are done according to the data transfer property + * list dxpl_id (which may be the constant H5P_DEFAULT). + * + * Return: Success: SUCCEED + * All writes have completed successfully + * + * Failure: FAIL + * One or more of the writes failed. + * + * Programmer: JRM -- 6/10/20 + * + * Changes: None. + * + *------------------------------------------------------------------------- + */ +herr_t +H5FDwrite_vector(H5FD_t *file, hid_t dxpl_id, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], + size_t sizes[], const void *bufs[] /* in */) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_API(FAIL) + H5TRACE7("e", "*#iIu*Mt*a*z**x", file, dxpl_id, count, types, addrs, sizes, bufs); + + /* Check arguments */ + if (!file) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file pointer cannot be NULL") + + if (!file->cls) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file class pointer cannot be NULL") + + if ((!types) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "types parameter can't be NULL if count is positive") + + if ((!addrs) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "addrs parameter can't be NULL if count is positive") + + if ((!sizes) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "sizes parameter can't be NULL if count is positive") + + if ((!bufs) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "bufs parameter can't be NULL if count is positive") + + if ((count > 0) && (sizes[0] == 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "sizes[0] can't be 0") + + if ((count > 0) && (types[0] == H5FD_MEM_NOLIST)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "count[0] can't be H5FD_MEM_NOLIST") + + /* Get the default dataset transfer property list if the user didn't provide one */ + if (H5P_DEFAULT == dxpl_id) { + dxpl_id = H5P_DATASET_XFER_DEFAULT; + } + else { + if (TRUE != H5P_isa_class(dxpl_id, H5P_DATASET_XFER)) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a data transfer property list") + } + + /* Set DXPL for operation */ + H5CX_set_dxpl(dxpl_id); + + /* Call private function */ + /* (Note compensating for base address addition in internal routine) */ + if (H5FD_write_vector(file, count, types, addrs, sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "file vector write request failed") + +done: + FUNC_LEAVE_API(ret_value) +} /* end H5FDwrite_vector() */ + +/*------------------------------------------------------------------------- + * Function: H5FDread_selection + * + * Purpose: Perform count reads from the specified file at the + * locations selected in the dataspaces in the file_spaces + * array, with each of those dataspaces starting at the file + * address specified by the corresponding element of the + * offsets array, and with the size of each element in the + * dataspace specified by the corresponding element of the + * element_sizes array. The memory type provided by type is + * the same for all selections. Data read is returned in + * the locations selected in the dataspaces in the + * mem_spaces array, within the buffers provided in the + * corresponding elements of the bufs array. + * + * If i > 0 and element_sizes[i] == 0, presume + * element_sizes[n] = element_sizes[i-1] for all n >= i and + * < count. + * + * If the underlying VFD supports selection reads, pass the + * call through directly. + * + * If it doesn't, convert the selection read into a sequence + * of individual reads. + * + * All reads are done according to the data transfer property + * list dxpl_id (which may be the constant H5P_DEFAULT). + * + * Return: Success: SUCCEED + * All reads have completed successfully, and + * the results havce been into the supplied + * buffers. + * + * Failure: FAIL + * The contents of supplied buffers are undefined. + * + * Programmer: NAF -- 5/19/21 + * + * Changes: None. + * + *------------------------------------------------------------------------- + */ +herr_t +H5FDread_selection(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, hid_t mem_space_ids[], + hid_t file_space_ids[], haddr_t offsets[], size_t element_sizes[], void *bufs[] /* out */) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_API(FAIL) + H5TRACE9("e", "*#MtiIu*i*i*a*zx", file, type, dxpl_id, count, mem_space_ids, file_space_ids, offsets, + element_sizes, bufs); + + /* Check arguments */ + if (!file) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file pointer cannot be NULL") + + if (!file->cls) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file class pointer cannot be NULL") + + if ((!mem_space_ids) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "mem_spaces parameter can't be NULL if count is positive") + + if ((!file_space_ids) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file_spaces parameter can't be NULL if count is positive") + + if ((!offsets) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "offsets parameter can't be NULL if count is positive") + + if ((!element_sizes) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, + "element_sizes parameter can't be NULL if count is positive") + + if ((!bufs) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "bufs parameter can't be NULL if count is positive") + + if ((count > 0) && (element_sizes[0] == 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "sizes[0] can't be 0") + + if ((count > 0) && (bufs[0] == NULL)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "bufs[0] can't be NULL") + + /* Get the default dataset transfer property list if the user didn't provide one */ + if (H5P_DEFAULT == dxpl_id) { + dxpl_id = H5P_DATASET_XFER_DEFAULT; + } + else { + if (TRUE != H5P_isa_class(dxpl_id, H5P_DATASET_XFER)) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a data transfer property list") + } + + /* Set DXPL for operation */ + H5CX_set_dxpl(dxpl_id); + + /* Call private function */ + /* (Note compensating for base address addition in internal routine) */ + if (H5FD_read_selection_id(file, type, count, mem_space_ids, file_space_ids, offsets, element_sizes, + bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "file selection read request failed") + +done: + FUNC_LEAVE_API(ret_value) +} /* end H5FDread_selection() */ + +/*------------------------------------------------------------------------- + * Function: H5FDwrite_selection + * + * Purpose: Perform count writes to the specified file at the + * locations selected in the dataspaces in the file_spaces + * array, with each of those dataspaces starting at the file + * address specified by the corresponding element of the + * offsets array, and with the size of each element in the + * dataspace specified by the corresponding element of the + * element_sizes array. The memory type provided by type is + * the same for all selections. Data write is from + * the locations selected in the dataspaces in the + * mem_spaces array, within the buffers provided in the + * corresponding elements of the bufs array. + * + * If i > 0 and element_sizes[i] == 0, presume + * element_sizes[n] = element_sizes[i-1] for all n >= i and + * < count. + * + * If the underlying VFD supports selection writes, pass the + * call through directly. + * + * If it doesn't, convert the selection write into a sequence + * of individual writes. + * + * All writes are done according to the data transfer property + * list dxpl_id (which may be the constant H5P_DEFAULT). + * + * Return: Success: SUCCEED + * All writes have completed successfully + * + * Failure: FAIL + * One or more of the writes failed. + * + * Programmer: NAF -- 5/14/21 + * + * Changes: None. + * + *------------------------------------------------------------------------- + */ +herr_t +H5FDwrite_selection(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, hid_t mem_space_ids[], + hid_t file_space_ids[], haddr_t offsets[], size_t element_sizes[], const void *bufs[]) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_API(FAIL) + H5TRACE9("e", "*#MtiIu*i*i*a*z**x", file, type, dxpl_id, count, mem_space_ids, file_space_ids, offsets, + element_sizes, bufs); + + /* Check arguments */ + if (!file) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file pointer cannot be NULL") + + if (!file->cls) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file class pointer cannot be NULL") + + if ((!mem_space_ids) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "mem_spaces parameter can't be NULL if count is positive") + + if ((!file_space_ids) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "file_spaces parameter can't be NULL if count is positive") + + if ((!offsets) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "offsets parameter can't be NULL if count is positive") + + if ((!element_sizes) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, + "element_sizes parameter can't be NULL if count is positive") + + if ((!bufs) && (count > 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "bufs parameter can't be NULL if count is positive") + + if ((count > 0) && (element_sizes[0] == 0)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "sizes[0] can't be 0") + + if ((count > 0) && (bufs[0] == NULL)) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "bufs[0] can't be NULL") + + /* Get the default dataset transfer property list if the user didn't provide one */ + if (H5P_DEFAULT == dxpl_id) { + dxpl_id = H5P_DATASET_XFER_DEFAULT; + } + else { + if (TRUE != H5P_isa_class(dxpl_id, H5P_DATASET_XFER)) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a data transfer property list") + } + + /* Set DXPL for operation */ + H5CX_set_dxpl(dxpl_id); + + /* Call private function */ + /* (Note compensating for base address addition in internal routine) */ + if (H5FD_write_selection_id(file, type, count, mem_space_ids, file_space_ids, offsets, element_sizes, + bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "file selection write request failed") + +done: + FUNC_LEAVE_API(ret_value) +} /* end H5FDwrite_selection() */ + +/*------------------------------------------------------------------------- * Function: H5FDflush * * Purpose: Notify driver to flush all cached data. If the driver has no @@ -1826,7 +2193,7 @@ H5FD_ctl(H5FD_t *file, uint64_t op_code, uint64_t flags, const void *input, void else if (flags & H5FD_CTL__FAIL_IF_UNKNOWN_FLAG) { HGOTO_ERROR(H5E_VFL, H5E_FCNTL, FAIL, - "VFD ctl request failed (no ctl callback and fail if unknown flag is set)") + "VFD ctl request failed (no ctl and fail if unknown flag is set)") } done: diff --git a/src/H5FDcore.c b/src/H5FDcore.c index a1750ee..0604316 100644 --- a/src/H5FDcore.c +++ b/src/H5FDcore.c @@ -152,6 +152,7 @@ static herr_t H5FD__core_delete(const char *filename, hid_t fapl_id); static inline const H5FD_core_fapl_t *H5FD__core_get_default_config(void); static const H5FD_class_t H5FD_core_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_CORE_VALUE, /* value */ "core", /* name */ MAXADDR, /* maxaddr */ @@ -180,6 +181,10 @@ static const H5FD_class_t H5FD_core_g = { H5FD__core_get_handle, /* get_handle */ H5FD__core_read, /* read */ H5FD__core_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ H5FD__core_flush, /* flush */ H5FD__core_truncate, /* truncate */ H5FD__core_lock, /* lock */ diff --git a/src/H5FDdevelop.h b/src/H5FDdevelop.h index 938f7f6..f5b32ed 100644 --- a/src/H5FDdevelop.h +++ b/src/H5FDdevelop.h @@ -25,6 +25,9 @@ /* Public Macros */ /*****************/ +/* H5FD_class_t struct version */ +#define H5FD_CLASS_VERSION 0x01 /* File driver struct version */ + /* Map "fractal heap" header blocks to 'ohdr' type file memory, since its * a fair amount of work to add a new kind of file memory and they are similar * enough to object headers and probably too minor to deserve their own type. @@ -160,6 +163,7 @@ typedef struct H5FD_t H5FD_t; /* Class information for each file driver */ typedef struct H5FD_class_t { + unsigned version; /**< File driver class struct version # */ H5FD_class_value_t value; const char * name; haddr_t maxaddr; @@ -188,6 +192,16 @@ typedef struct H5FD_class_t { herr_t (*get_handle)(H5FD_t *file, hid_t fapl, void **file_handle); herr_t (*read)(H5FD_t *file, H5FD_mem_t type, hid_t dxpl, haddr_t addr, size_t size, void *buffer); herr_t (*write)(H5FD_t *file, H5FD_mem_t type, hid_t dxpl, haddr_t addr, size_t size, const void *buffer); + herr_t (*read_vector)(H5FD_t *file, hid_t dxpl, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], + size_t sizes[], void *bufs[]); + herr_t (*write_vector)(H5FD_t *file, hid_t dxpl, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], + size_t sizes[], const void *bufs[]); + herr_t (*read_selection)(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, size_t count, hid_t mem_spaces[], + hid_t file_spaces[], haddr_t offsets[], size_t element_sizes[], + void *bufs[] /*out*/); + herr_t (*write_selection)(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, size_t count, hid_t mem_spaces[], + hid_t file_spaces[], haddr_t offsets[], size_t element_sizes[], + const void *bufs[] /*in*/); herr_t (*flush)(H5FD_t *file, hid_t dxpl_id, hbool_t closing); herr_t (*truncate)(H5FD_t *file, hid_t dxpl_id, hbool_t closing); herr_t (*lock)(H5FD_t *file, hbool_t rw); @@ -223,6 +237,9 @@ struct H5FD_t { hbool_t paged_aggr; /* Paged aggregation for file space is enabled or not */ }; +/* VFD initialization function */ +typedef hid_t (*H5FD_init_t)(void); + /********************/ /* Public Variables */ /********************/ @@ -235,7 +252,7 @@ struct H5FD_t { extern "C" { #endif -H5_DLL hid_t H5FDperform_init(hid_t (*)(void)); +H5_DLL hid_t H5FDperform_init(H5FD_init_t op); H5_DLL hid_t H5FDregister(const H5FD_class_t *cls); H5_DLL htri_t H5FDis_driver_registered_by_name(const char *driver_name); H5_DLL htri_t H5FDis_driver_registered_by_value(H5FD_class_value_t driver_value); @@ -254,6 +271,16 @@ H5_DLL herr_t H5FDread(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, haddr_t ad void *buf /*out*/); H5_DLL herr_t H5FDwrite(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr, size_t size, const void *buf); +H5_DLL herr_t H5FDread_vector(H5FD_t *file, hid_t dxpl_id, uint32_t count, H5FD_mem_t types[], + haddr_t addrs[], size_t sizes[], void *bufs[] /* out */); +H5_DLL herr_t H5FDwrite_vector(H5FD_t *file, hid_t dxpl_id, uint32_t count, H5FD_mem_t types[], + haddr_t addrs[], size_t sizes[], const void *bufs[] /* in */); +H5_DLL herr_t H5FDread_selection(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, + hid_t mem_spaces[], hid_t file_spaces[], haddr_t offsets[], + size_t element_sizes[], void *bufs[] /* out */); +H5_DLL herr_t H5FDwrite_selection(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, + hid_t mem_spaces[], hid_t file_spaces[], haddr_t offsets[], + size_t element_sizes[], const void *bufs[]); H5_DLL herr_t H5FDflush(H5FD_t *file, hid_t dxpl_id, hbool_t closing); H5_DLL herr_t H5FDtruncate(H5FD_t *file, hid_t dxpl_id, hbool_t closing); H5_DLL herr_t H5FDlock(H5FD_t *file, hbool_t rw); diff --git a/src/H5FDdirect.c b/src/H5FDdirect.c index d07d909..25ee970 100644 --- a/src/H5FDdirect.c +++ b/src/H5FDdirect.c @@ -142,6 +142,7 @@ static herr_t H5FD__direct_unlock(H5FD_t *_file); static herr_t H5FD__direct_delete(const char *filename, hid_t fapl_id); static const H5FD_class_t H5FD_direct_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_DIRECT_VALUE, /* value */ "direct", /* name */ MAXADDR, /* maxaddr */ @@ -170,6 +171,10 @@ static const H5FD_class_t H5FD_direct_g = { H5FD__direct_get_handle, /* get_handle */ H5FD__direct_read, /* read */ H5FD__direct_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ NULL, /* flush */ H5FD__direct_truncate, /* truncate */ H5FD__direct_lock, /* lock */ @@ -213,8 +218,11 @@ H5FD_direct_init(void) else ignore_disabled_file_locks_s = FAIL; /* Environment variable not set, or not set correctly */ - if (H5I_VFL != H5I_get_type(H5FD_DIRECT_g)) + if (H5I_VFL != H5I_get_type(H5FD_DIRECT_g)) { H5FD_DIRECT_g = H5FD_register(&H5FD_direct_g, sizeof(H5FD_class_t), FALSE); + if (H5I_INVALID_HID == H5FD_DIRECT_g) + HGOTO_ERROR(H5E_ID, H5E_CANTREGISTER, H5I_INVALID_HID, "unable to register direct"); + } /* Set return value */ ret_value = H5FD_DIRECT_g; diff --git a/src/H5FDfamily.c b/src/H5FDfamily.c index 4e54197..66a1a68 100644 --- a/src/H5FDfamily.c +++ b/src/H5FDfamily.c @@ -112,6 +112,7 @@ static herr_t H5FD__family_delete(const char *filename, hid_t fapl_id); /* The class struct */ static const H5FD_class_t H5FD_family_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_FAMILY_VALUE, /* value */ "family", /* name */ HADDR_MAX, /* maxaddr */ @@ -140,6 +141,10 @@ static const H5FD_class_t H5FD_family_g = { H5FD__family_get_handle, /* get_handle */ H5FD__family_read, /* read */ H5FD__family_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ H5FD__family_flush, /* flush */ H5FD__family_truncate, /* truncate */ H5FD__family_lock, /* lock */ @@ -233,20 +238,20 @@ H5FD__family_get_default_printf_filename(const char *old_filename) HGOTO_ERROR(H5E_VFL, H5E_CANTALLOC, NULL, "can't allocate new filename buffer") /* Determine if filename contains a ".h5" extension. */ - if ((file_extension = strstr(old_filename, ".h5"))) { + if ((file_extension = HDstrstr(old_filename, ".h5"))) { /* Insert the printf format between the filename and ".h5" extension. */ HDstrcpy(tmp_buffer, old_filename); - file_extension = strstr(tmp_buffer, ".h5"); + file_extension = HDstrstr(tmp_buffer, ".h5"); HDsprintf(file_extension, "%s%s", suffix, ".h5"); } - else if ((file_extension = strrchr(old_filename, '.'))) { + else if ((file_extension = HDstrrchr(old_filename, '.'))) { char *new_extension_loc = NULL; /* If the filename doesn't contain a ".h5" extension, but contains * AN extension, just insert the printf format before that extension. */ HDstrcpy(tmp_buffer, old_filename); - new_extension_loc = strrchr(tmp_buffer, '.'); + new_extension_loc = HDstrrchr(tmp_buffer, '.'); HDsprintf(new_extension_loc, "%s%s", suffix, file_extension); } else { diff --git a/src/H5FDhdfs.c b/src/H5FDhdfs.c index 4927a40..f0ffb62 100644 --- a/src/H5FDhdfs.c +++ b/src/H5FDhdfs.c @@ -278,6 +278,7 @@ static herr_t H5FD__hdfs_truncate(H5FD_t *_file, hid_t dxpl_id, hbool_t closing static herr_t H5FD__hdfs_validate_config(const H5FD_hdfs_fapl_t *fa); static const H5FD_class_t H5FD_hdfs_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_HDFS_VALUE, /* value */ "hdfs", /* name */ MAXADDR, /* maxaddr */ @@ -306,6 +307,10 @@ static const H5FD_class_t H5FD_hdfs_g = { H5FD__hdfs_get_handle, /* get_handle */ H5FD__hdfs_read, /* read */ H5FD__hdfs_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ NULL, /* flush */ H5FD__hdfs_truncate, /* truncate */ NULL, /* lock */ diff --git a/src/H5FDint.c b/src/H5FDint.c index d7fe33c..0c3fe9e 100644 --- a/src/H5FDint.c +++ b/src/H5FDint.c @@ -34,6 +34,7 @@ #include "H5Eprivate.h" /* Error handling */ #include "H5Fprivate.h" /* File access */ #include "H5FDpkg.h" /* File Drivers */ +#include "H5FLprivate.h" /* Free Lists */ #include "H5Iprivate.h" /* IDs */ #include "H5PLprivate.h" /* Plugins */ @@ -41,10 +42,51 @@ /* Local Macros */ /****************/ +/* Length of sequence lists requested from dataspace selections */ +#define H5FD_SEQ_LIST_LEN 128 + +/* Length of stack allocated arrays for building vector I/O operations. + * Corresponds to the number of contiguous blocks in a selection I/O operation. + * If more space is needed dynamic allocation will be used instead. */ +#define H5FD_LOCAL_VECTOR_LEN 8 + +/* Length of stack allocated arrays for dataspace IDs/structs for selection I/O + * operations. Corresponds to the number of file selection/memory selection + * pairs (along with addresses, etc.) in a selection I/O operation. If more + * space is needed dynamic allocation will be used instead */ +#define H5FD_LOCAL_SEL_ARR_LEN 8 + /******************/ /* Local Typedefs */ /******************/ +/************************************************************************* + * + * H5FD_vsrt_tmp_t + * + * Structure used to store vector I/O request addresses and the associated + * indexes in the addrs[] array for the purpose of determine the sorted + * order. + * + * This is done by allocating an array of H5FD_vsrt_tmp_t of length + * count, loading it with the contents of the addrs[] array and the + * associated indices, and then sorting it. + * + * This sorted array of H5FD_vsrt_tmp_t is then used to populate sorted + * versions of the types[], addrs[], sizes[] and bufs[] vectors. + * + * addr: haddr_t containing the value of addrs[i], + * + * index: integer containing the value of i used to obtain the + * value of the addr field from the addrs[] vector. + * + *************************************************************************/ + +typedef struct H5FD_vsrt_tmp_t { + haddr_t addr; + int index; +} H5FD_vsrt_tmp_t; + /* Information needed for iterating over the registered VFD hid_t IDs. * The name or value of the new VFD that is being registered is stored * in the name (or value) field and the found_id field is initialized to @@ -66,7 +108,13 @@ typedef struct H5FD_get_driver_ud_t { /********************/ /* Local Prototypes */ /********************/ -static int H5FD__get_driver_cb(void *obj, hid_t id, void *_op_data); +static int H5FD__get_driver_cb(void *obj, hid_t id, void *_op_data); +static herr_t H5FD__read_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, + H5S_t **mem_spaces, H5S_t **file_spaces, haddr_t offsets[], + size_t element_sizes[], void *bufs[] /* out */); +static herr_t H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, + H5S_t **mem_spaces, H5S_t **file_spaces, haddr_t offsets[], + size_t element_sizes[], const void *bufs[]); /*********************/ /* Package Variables */ @@ -80,6 +128,9 @@ static int H5FD__get_driver_cb(void *obj, hid_t id, void *_op_data); /* Local Variables */ /*******************/ +/* Declare extern free list to manage the H5S_sel_iter_t struct */ +H5FL_EXTERN(H5S_sel_iter_t); + /*------------------------------------------------------------------------- * Function: H5FD_locate_signature * @@ -260,6 +311,1689 @@ done: } /* end H5FD_write() */ /*------------------------------------------------------------------------- + * Function: H5FD_read_vector + * + * Purpose: Private version of H5FDread_vector() + * + * Perform count reads from the specified file at the offsets + * provided in the addrs array, with the lengths and memory + * types provided in the sizes and types arrays. Data read + * is returned in the buffers provided in the bufs array. + * + * If i > 0 and sizes[i] == 0, presume sizes[n] = sizes[i-1] + * for all n >= i and < count. + * + * Similarly, if i > 0 and types[i] == H5FD_MEM_NOLIST, + * presume types[n] = types[i-1] for all n >= i and < count. + * + * If the underlying VFD supports vector reads, pass the + * call through directly. + * + * If it doesn't, convert the vector read into a sequence + * of individual reads. + * + * Note that it is not in general possible to convert a + * vector read into a selection read, because each element + * in the vector read may have a different memory type. + * In contrast, selection reads are of a single type. + * + * Return: Success: SUCCEED + * All reads have completed successfully, and + * the results havce been into the supplied + * buffers. + * + * Failure: FAIL + * The contents of supplied buffers are undefined. + * + * Programmer: JRM -- 6/10/20 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +herr_t +H5FD_read_vector(H5FD_t *file, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], size_t sizes[], + void *bufs[] /* out */) +{ + hbool_t addrs_cooked = FALSE; + hbool_t extend_sizes = FALSE; + hbool_t extend_types = FALSE; + uint32_t i; + size_t size; + H5FD_mem_t type; + hid_t dxpl_id = H5I_INVALID_HID; /* DXPL for operation */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert((types) || (count == 0)); + HDassert((addrs) || (count == 0)); + HDassert((sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* verify that the first elements of the sizes and types arrays are + * valid. + */ + HDassert((count == 0) || (sizes[0] != 0)); + HDassert((count == 0) || (types[0] != H5FD_MEM_NOLIST)); + + /* Get proper DXPL for I/O */ + dxpl_id = H5CX_get_dxpl(); + +#ifndef H5_HAVE_PARALLEL + /* The no-op case + * + * Do not return early for Parallel mode since the I/O could be a + * collective transfer. + */ + if (0 == count) { + HGOTO_DONE(SUCCEED) + } +#endif /* H5_HAVE_PARALLEL */ + + if (file->base_addr > 0) { + + /* apply the base_addr offset to the addrs array. Must undo before + * we return. + */ + for (i = 0; i < count; i++) { + + addrs[i] += file->base_addr; + } + addrs_cooked = TRUE; + } + + /* If the file is open for SWMR read access, allow access to data past + * the end of the allocated space (the 'eoa'). This is done because the + * eoa stored in the file's superblock might be out of sync with the + * objects being written within the file by the application performing + * SWMR write operations. + */ + if ((!(file->access_flags & H5F_ACC_SWMR_READ)) && (count > 0)) { + haddr_t eoa; + + extend_sizes = FALSE; + extend_types = FALSE; + + for (i = 0; i < count; i++) { + + if (!extend_sizes) { + + if (sizes[i] == 0) { + + extend_sizes = TRUE; + size = sizes[i - 1]; + } + else { + + size = sizes[i]; + } + } + + if (!extend_types) { + + if (types[i] == H5FD_MEM_NOLIST) { + + extend_types = TRUE; + type = types[i - 1]; + } + else { + + type = types[i]; + } + } + + if (HADDR_UNDEF == (eoa = (file->cls->get_eoa)(file, type))) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "driver get_eoa request failed") + + if ((addrs[i] + size) > eoa) + + HGOTO_ERROR(H5E_ARGS, H5E_OVERFLOW, FAIL, + "addr overflow, addrs[%d] = %llu, sizes[%d] = %llu, eoa = %llu", (int)i, + (unsigned long long)(addrs[i]), (int)i, (unsigned long long)size, + (unsigned long long)eoa) + } + } + + /* if the underlying VFD supports vector read, make the call */ + if (file->cls->read_vector) { + + if ((file->cls->read_vector)(file, dxpl_id, count, types, addrs, sizes, bufs) < 0) + + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "driver read vector request failed") + } + else { + + /* otherwise, implement the vector read as a sequence of regular + * read calls. + */ + extend_sizes = FALSE; + extend_types = FALSE; + + for (i = 0; i < count; i++) { + + /* we have already verified that sizes[0] != 0 and + * types[0] != H5FD_MEM_NOLIST + */ + + if (!extend_sizes) { + + if (sizes[i] == 0) { + + extend_sizes = TRUE; + size = sizes[i - 1]; + } + else { + + size = sizes[i]; + } + } + + if (!extend_types) { + + if (types[i] == H5FD_MEM_NOLIST) { + + extend_types = TRUE; + type = types[i - 1]; + } + else { + + type = types[i]; + } + } + + if ((file->cls->read)(file, type, dxpl_id, addrs[i], size, bufs[i]) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "driver read request failed") + } + } + +done: + /* undo the base addr offset to the addrs array if necessary */ + if (addrs_cooked) { + + HDassert(file->base_addr > 0); + + for (i = 0; i < count; i++) { + + addrs[i] -= file->base_addr; + } + } + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD_read_vector() */ + +/*------------------------------------------------------------------------- + * Function: H5FD_write_vector + * + * Purpose: Private version of H5FDwrite_vector() + * + * Perform count writes to the specified file at the offsets + * provided in the addrs array, with the lengths and memory + * types provided in the sizes and types arrays. Data written + * is taken from the buffers provided in the bufs array. + * + * If i > 0 and sizes[i] == 0, presume sizes[n] = sizes[i-1] + * for all n >= i and < count. + * + * Similarly, if i > 0 and types[i] == H5FD_MEM_NOLIST, + * presume types[n] = types[i-1] for all n >= i and < count. + * + * If the underlying VFD supports vector writes, pass the + * call through directly. + * + * If it doesn't, convert the vector write into a sequence + * of individual writes. + * + * Note that it is not in general possible to convert a + * vector write into a selection write, because each element + * in the vector write may have a different memory type. + * In contrast, selection writes are of a single type. + * + * Return: Success: SUCCEED + * All writes have completed successfully. + * + * Failure: FAIL + * One or more writes failed. + * + * Programmer: JRM -- 6/10/20 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +herr_t +H5FD_write_vector(H5FD_t *file, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], size_t sizes[], + const void *bufs[]) +{ + hbool_t addrs_cooked = FALSE; + hbool_t extend_sizes = FALSE; + hbool_t extend_types = FALSE; + uint32_t i; + size_t size; + H5FD_mem_t type; + hid_t dxpl_id; /* DXPL for operation */ + haddr_t eoa = HADDR_UNDEF; /* EOA for file */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert((types) || (count == 0)); + HDassert((addrs) || (count == 0)); + HDassert((sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* verify that the first elements of the sizes and types arrays are + * valid. + */ + HDassert((count == 0) || (sizes[0] != 0)); + HDassert((count == 0) || (types[0] != H5FD_MEM_NOLIST)); + + /* Get proper DXPL for I/O */ + dxpl_id = H5CX_get_dxpl(); + +#ifndef H5_HAVE_PARALLEL + /* The no-op case + * + * Do not return early for Parallel mode since the I/O could be a + * collective transfer. + */ + if (0 == count) + HGOTO_DONE(SUCCEED) +#endif /* H5_HAVE_PARALLEL */ + + if (file->base_addr > 0) { + + /* apply the base_addr offset to the addrs array. Must undo before + * we return. + */ + for (i = 0; i < count; i++) { + + addrs[i] += file->base_addr; + } + addrs_cooked = TRUE; + } + + extend_sizes = FALSE; + extend_types = FALSE; + + for (i = 0; i < count; i++) { + + if (!extend_sizes) { + + if (sizes[i] == 0) { + + extend_sizes = TRUE; + size = sizes[i - 1]; + } + else { + + size = sizes[i]; + } + } + + if (!extend_types) { + + if (types[i] == H5FD_MEM_NOLIST) { + + extend_types = TRUE; + type = types[i - 1]; + } + else { + + type = types[i]; + } + } + + if (HADDR_UNDEF == (eoa = (file->cls->get_eoa)(file, type))) + + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "driver get_eoa request failed") + + if ((addrs[i] + size) > eoa) + + HGOTO_ERROR(H5E_ARGS, H5E_OVERFLOW, FAIL, "addr overflow, addrs[%d] = %llu, sizes[%d] = %llu, \ + eoa = %llu", + (int)i, (unsigned long long)(addrs[i]), (int)i, (unsigned long long)size, + (unsigned long long)eoa) + } + + /* if the underlying VFD supports vector write, make the call */ + if (file->cls->write_vector) { + + if ((file->cls->write_vector)(file, dxpl_id, count, types, addrs, sizes, bufs) < 0) + + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "driver write vector request failed") + } + else { + /* otherwise, implement the vector write as a sequence of regular + * write calls. + */ + extend_sizes = FALSE; + extend_types = FALSE; + + for (i = 0; i < count; i++) { + + /* we have already verified that sizes[0] != 0 and + * types[0] != H5FD_MEM_NOLIST + */ + + if (!extend_sizes) { + + if (sizes[i] == 0) { + + extend_sizes = TRUE; + size = sizes[i - 1]; + } + else { + + size = sizes[i]; + } + } + + if (!extend_types) { + + if (types[i] == H5FD_MEM_NOLIST) { + + extend_types = TRUE; + type = types[i - 1]; + } + else { + + type = types[i]; + } + } + + if ((file->cls->write)(file, type, dxpl_id, addrs[i], size, bufs[i]) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "driver write request failed") + } + } + +done: + /* undo the base addr offset to the addrs array if necessary */ + if (addrs_cooked) { + + HDassert(file->base_addr > 0); + + for (i = 0; i < count; i++) { + + addrs[i] -= file->base_addr; + } + } + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD_write_vector() */ + +/*------------------------------------------------------------------------- + * Function: H5FD__read_selection_translate + * + * Purpose: Translates a selection read call to a vector read call if + * vector reads are supported, or a series of scalar read + * calls otherwise. + * + * Return: Success: SUCCEED + * All reads have completed successfully, and + * the results havce been into the supplied + * buffers. + * + * Failure: FAIL + * The contents of supplied buffers are undefined. + * + * Programmer: NAF -- 5/13/21 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +static herr_t +H5FD__read_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, + H5S_t **mem_spaces, H5S_t **file_spaces, haddr_t offsets[], + size_t element_sizes[], void *bufs[] /* out */) +{ + hbool_t extend_sizes = FALSE; + hbool_t extend_bufs = FALSE; + uint32_t i; + size_t element_size; + void * buf; + hbool_t use_vector = FALSE; + haddr_t addrs_local[H5FD_LOCAL_VECTOR_LEN]; + haddr_t * addrs = addrs_local; + size_t sizes_local[H5FD_LOCAL_VECTOR_LEN]; + size_t * sizes = sizes_local; + void * vec_bufs_local[H5FD_LOCAL_VECTOR_LEN]; + void ** vec_bufs = vec_bufs_local; + hsize_t file_off[H5FD_SEQ_LIST_LEN]; + size_t file_len[H5FD_SEQ_LIST_LEN]; + hsize_t mem_off[H5FD_SEQ_LIST_LEN]; + size_t mem_len[H5FD_SEQ_LIST_LEN]; + size_t file_seq_i; + size_t mem_seq_i; + size_t file_nseq; + size_t mem_nseq; + size_t io_len; + size_t nelmts; + hssize_t hss_nelmts; + size_t seq_nelem; + H5S_sel_iter_t *file_iter = NULL; + H5S_sel_iter_t *mem_iter = NULL; + hbool_t file_iter_init = FALSE; + hbool_t mem_iter_init = FALSE; + H5FD_mem_t types[2] = {type, H5FD_MEM_NOLIST}; + size_t vec_arr_nalloc = H5FD_LOCAL_VECTOR_LEN; + size_t vec_arr_nused = 0; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert(mem_spaces); + HDassert(file_spaces); + HDassert(offsets); + HDassert(element_sizes); + HDassert(bufs); + + /* Verify that the first elements of the element_sizes and bufs arrays are + * valid. */ + HDassert(element_sizes[0] != 0); + HDassert(bufs[0] != NULL); + + /* Check if we're using vector I/O */ + use_vector = file->cls->read_vector != NULL; + + /* Allocate sequence lists for memory and file spaces */ + if (NULL == (file_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_VFL, H5E_CANTALLOC, FAIL, "couldn't allocate file selection iterator") + if (NULL == (mem_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_VFL, H5E_CANTALLOC, FAIL, "couldn't allocate memory selection iterator") + + /* Loop over dataspaces */ + for (i = 0; i < count; i++) { + + /* we have already verified that element_sizes[0] != 0 and bufs[0] + * != NULL */ + + if (!extend_sizes) { + + if (element_sizes[i] == 0) { + + extend_sizes = TRUE; + element_size = element_sizes[i - 1]; + } + else { + + element_size = element_sizes[i]; + } + } + + if (!extend_bufs) { + + if (bufs[i] == NULL) { + + extend_bufs = TRUE; + buf = bufs[i - 1]; + } + else { + + buf = bufs[i]; + } + } + + /* Initialize sequence lists for memory and file spaces */ + if (H5S_select_iter_init(file_iter, file_spaces[i], element_size, 0) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "can't initialize sequence list for file space") + file_iter_init = TRUE; + if (H5S_select_iter_init(mem_iter, mem_spaces[i], element_size, 0) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "can't initialize sequence list for memory space") + mem_iter_init = TRUE; + + /* Get the number of elements in selection */ + if ((hss_nelmts = (hssize_t)H5S_GET_SELECT_NPOINTS(file_spaces[i])) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTCOUNT, FAIL, "can't get number of elements selected") + H5_CHECKED_ASSIGN(nelmts, size_t, hss_nelmts, hssize_t); + +#ifndef NDEBUG + /* Verify mem space has the same number of elements */ + { + if ((hss_nelmts = (hssize_t)H5S_GET_SELECT_NPOINTS(mem_spaces[i])) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTCOUNT, FAIL, "can't get number of elements selected") + HDassert((hssize_t)nelmts == hss_nelmts); + } +#endif /* NDEBUG */ + + /* Initialize values so sequence lists are retrieved on the first + * iteration */ + file_seq_i = H5FD_SEQ_LIST_LEN; + mem_seq_i = H5FD_SEQ_LIST_LEN; + file_nseq = 0; + mem_nseq = 0; + + /* Loop until all elements are processed */ + while (file_seq_i < file_nseq || nelmts > 0) { + /* Fill/refill file sequence list if necessary */ + if (file_seq_i == H5FD_SEQ_LIST_LEN) { + if (H5S_SELECT_ITER_GET_SEQ_LIST(file_iter, H5FD_SEQ_LIST_LEN, SIZE_MAX, &file_nseq, + &seq_nelem, file_off, file_len) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed") + HDassert(file_nseq > 0); + + nelmts -= seq_nelem; + file_seq_i = 0; + } + HDassert(file_seq_i < file_nseq); + + /* Fill/refill memory sequence list if necessary */ + if (mem_seq_i == H5FD_SEQ_LIST_LEN) { + if (H5S_SELECT_ITER_GET_SEQ_LIST(mem_iter, H5FD_SEQ_LIST_LEN, SIZE_MAX, &mem_nseq, &seq_nelem, + mem_off, mem_len) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed") + HDassert(mem_nseq > 0); + + mem_seq_i = 0; + } + HDassert(mem_seq_i < mem_nseq); + + /* Calculate length of this IO */ + io_len = MIN(file_len[file_seq_i], mem_len[mem_seq_i]); + + /* Check if we're using vector I/O */ + if (use_vector) { + /* Check if we need to extend the arrays */ + if (vec_arr_nused == vec_arr_nalloc) { + /* Check if we're using the static arrays */ + if (addrs == addrs_local) { + HDassert(sizes == sizes_local); + HDassert(vec_bufs == vec_bufs_local); + + /* Allocate dynamic arrays */ + if (NULL == (addrs = H5MM_malloc(sizeof(addrs_local) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for address list") + if (NULL == (sizes = H5MM_malloc(sizeof(sizes_local) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for size list") + if (NULL == (vec_bufs = H5MM_malloc(sizeof(vec_bufs_local) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for buffer list") + + /* Copy the existing data */ + (void)H5MM_memcpy(addrs, addrs_local, sizeof(addrs_local)); + (void)H5MM_memcpy(sizes, sizes_local, sizeof(sizes_local)); + (void)H5MM_memcpy(vec_bufs, vec_bufs_local, sizeof(vec_bufs_local)); + } + else { + void *tmp_ptr; + + /* Reallocate arrays */ + if (NULL == (tmp_ptr = H5MM_realloc(addrs, vec_arr_nalloc * sizeof(*addrs) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory reallocation failed for address list") + addrs = tmp_ptr; + if (NULL == (tmp_ptr = H5MM_realloc(sizes, vec_arr_nalloc * sizeof(*sizes) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory reallocation failed for size list") + sizes = tmp_ptr; + if (NULL == + (tmp_ptr = H5MM_realloc(vec_bufs, vec_arr_nalloc * sizeof(*vec_bufs) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory reallocation failed for buffer list") + vec_bufs = tmp_ptr; + } + + /* Record that we've doubled the array sizes */ + vec_arr_nalloc *= 2; + } + + /* Add this segment to vector read list */ + addrs[vec_arr_nused] = offsets[i] + file_off[file_seq_i]; + sizes[vec_arr_nused] = io_len; + vec_bufs[vec_arr_nused] = (void *)((uint8_t *)buf + mem_off[mem_seq_i]); + vec_arr_nused++; + } + else + /* Issue scalar read call */ + if ((file->cls->read)(file, type, dxpl_id, offsets[i] + file_off[file_seq_i], io_len, + (void *)((uint8_t *)buf + mem_off[mem_seq_i])) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "driver read request failed") + + /* Update file sequence */ + if (io_len == file_len[file_seq_i]) + file_seq_i++; + else { + file_off[file_seq_i] += io_len; + file_len[file_seq_i] -= io_len; + } + + /* Update memory sequence */ + if (io_len == mem_len[mem_seq_i]) + mem_seq_i++; + else { + mem_off[mem_seq_i] += io_len; + mem_len[mem_seq_i] -= io_len; + } + } + + /* Make sure both memory and file sequences terminated at the same time */ + if (mem_seq_i < mem_nseq) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADVALUE, FAIL, "file selection terminated before memory selection") + + /* Terminate iterators */ + if (H5S_SELECT_ITER_RELEASE(file_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release file selection iterator") + file_iter_init = FALSE; + if (H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release memory selection iterator") + mem_iter_init = FALSE; + } + + /* Issue vector read call if appropriate */ + if (use_vector) { + H5_CHECK_OVERFLOW(vec_arr_nused, size_t, uint32_t) + if ((file->cls->read_vector)(file, dxpl_id, (uint32_t)vec_arr_nused, types, addrs, sizes, vec_bufs) < + 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "driver read vector request failed") + } + +done: + /* Terminate and free iterators */ + if (file_iter) { + if (file_iter_init && H5S_SELECT_ITER_RELEASE(file_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release file selection iterator") + file_iter = H5FL_FREE(H5S_sel_iter_t, file_iter); + } + if (mem_iter) { + if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release memory selection iterator") + mem_iter = H5FL_FREE(H5S_sel_iter_t, mem_iter); + } + + /* Cleanup vector arrays */ + if (use_vector) { + if (addrs != addrs_local) + addrs = H5MM_xfree(addrs); + if (sizes != sizes_local) + sizes = H5MM_xfree(sizes); + if (vec_bufs != vec_bufs_local) + vec_bufs = H5MM_xfree(vec_bufs); + } + + /* Make sure we cleaned up */ + HDassert(!addrs || addrs == addrs_local); + HDassert(!sizes || sizes == sizes_local); + HDassert(!vec_bufs || vec_bufs == vec_bufs_local); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD__read_selection_translate() */ + +/*------------------------------------------------------------------------- + * Function: H5FD_read_selection + * + * Purpose: Private version of H5FDread_selection() + * + * Perform count reads from the specified file at the + * locations selected in the dataspaces in the file_spaces + * array, with each of those dataspaces starting at the file + * address specified by the corresponding element of the + * offsets array, and with the size of each element in the + * dataspace specified by the corresponding element of the + * element_sizes array. The memory type provided by type is + * the same for all selections. Data read is returned in + * the locations selected in the dataspaces in the + * mem_spaces array, within the buffers provided in the + * corresponding elements of the bufs array. + * + * If i > 0 and element_sizes[i] == 0, presume + * element_sizes[n] = element_sizes[i-1] for all n >= i and + * < count. + * + * If the underlying VFD supports selection reads, pass the + * call through directly. + * + * If it doesn't, convert the vector read into a sequence + * of individual reads. + * + * Return: Success: SUCCEED + * All reads have completed successfully, and + * the results havce been into the supplied + * buffers. + * + * Failure: FAIL + * The contents of supplied buffers are undefined. + * + * Programmer: NAF -- 3/29/21 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +herr_t +H5FD_read_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, H5S_t **mem_spaces, H5S_t **file_spaces, + haddr_t offsets[], size_t element_sizes[], void *bufs[] /* out */) +{ + hbool_t offsets_cooked = FALSE; + hid_t mem_space_ids_local[H5FD_LOCAL_SEL_ARR_LEN]; + hid_t * mem_space_ids = mem_space_ids_local; + hid_t file_space_ids_local[H5FD_LOCAL_SEL_ARR_LEN]; + hid_t * file_space_ids = file_space_ids_local; + uint32_t num_spaces = 0; + hid_t dxpl_id = H5I_INVALID_HID; /* DXPL for operation */ + uint32_t i; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert((mem_spaces) || (count == 0)); + HDassert((file_spaces) || (count == 0)); + HDassert((offsets) || (count == 0)); + HDassert((element_sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* Verify that the first elements of the element_sizes and bufs arrays are + * valid. */ + HDassert((count == 0) || (element_sizes[0] != 0)); + HDassert((count == 0) || (bufs[0] != NULL)); + + /* Get proper DXPL for I/O */ + dxpl_id = H5CX_get_dxpl(); + +#ifndef H5_HAVE_PARALLEL + /* The no-op case + * + * Do not return early for Parallel mode since the I/O could be a + * collective transfer. + */ + if (0 == count) { + HGOTO_DONE(SUCCEED) + } +#endif /* H5_HAVE_PARALLEL */ + + if (file->base_addr > 0) { + + /* apply the base_addr offset to the offsets array. Must undo before + * we return. + */ + for (i = 0; i < count; i++) { + + offsets[i] += file->base_addr; + } + offsets_cooked = TRUE; + } + + /* If the file is open for SWMR read access, allow access to data past + * the end of the allocated space (the 'eoa'). This is done because the + * eoa stored in the file's superblock might be out of sync with the + * objects being written within the file by the application performing + * SWMR write operations. + */ + /* For now at least, only check that the offset is not past the eoa, since + * looking into the highest offset in the selection (different from the + * bounds) is potentially expensive. + */ + if (!(file->access_flags & H5F_ACC_SWMR_READ)) { + haddr_t eoa; + + if (HADDR_UNDEF == (eoa = (file->cls->get_eoa)(file, type))) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "driver get_eoa request failed") + + for (i = 0; i < count; i++) { + + if ((offsets[i]) > eoa) + + HGOTO_ERROR(H5E_ARGS, H5E_OVERFLOW, FAIL, "addr overflow, offsets[%d] = %llu, eoa = %llu", + (int)i, (unsigned long long)(offsets[i]), (unsigned long long)eoa) + } + } + + /* if the underlying VFD supports selection read, make the call */ + if (file->cls->read_selection) { + /* Allocate array of space IDs if necessary, otherwise use local + * buffers */ + if (count > sizeof(mem_space_ids_local) / sizeof(mem_space_ids_local[0])) { + if (NULL == (mem_space_ids = H5MM_malloc(count * sizeof(hid_t)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + if (NULL == (file_space_ids = H5MM_malloc(count * sizeof(hid_t)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + } + + /* Create IDs for all dataspaces */ + for (; num_spaces < count; num_spaces++) { + if ((mem_space_ids[num_spaces] = H5I_register(H5I_DATASPACE, mem_spaces[num_spaces], TRUE)) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTREGISTER, FAIL, "unable to register dataspace ID") + + if ((file_space_ids[num_spaces] = H5I_register(H5I_DATASPACE, file_spaces[num_spaces], TRUE)) < + 0) { + if (H5I_dec_app_ref(mem_space_ids[num_spaces]) < 0) + HDONE_ERROR(H5E_VFL, H5E_CANTDEC, FAIL, "problem freeing id") + HGOTO_ERROR(H5E_VFL, H5E_CANTREGISTER, FAIL, "unable to register dataspace ID") + } + } + + if ((file->cls->read_selection)(file, type, dxpl_id, count, mem_space_ids, file_space_ids, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "driver read selection request failed") + } + else + /* Otherwise, implement the selection read as a sequence of regular + * or vector read calls. + */ + if (H5FD__read_selection_translate(file, type, dxpl_id, count, mem_spaces, file_spaces, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "translation to vector or scalar read failed") + +done: + /* undo the base addr offset to the offsets array if necessary */ + if (offsets_cooked) { + + HDassert(file->base_addr > 0); + + for (i = 0; i < count; i++) { + + offsets[i] -= file->base_addr; + } + } + + /* Cleanup dataspace arrays */ + for (i = 0; i < num_spaces; i++) { + if (H5I_dec_app_ref(mem_space_ids[i]) < 0) + HDONE_ERROR(H5E_VFL, H5E_CANTDEC, FAIL, "problem freeing id") + if (H5I_dec_app_ref(file_space_ids[i]) < 0) + HDONE_ERROR(H5E_VFL, H5E_CANTDEC, FAIL, "problem freeing id") + } + if (mem_space_ids != mem_space_ids_local) + mem_space_ids = H5MM_xfree(mem_space_ids); + if (file_space_ids != file_space_ids_local) + file_space_ids = H5MM_xfree(file_space_ids); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD_read_selection() */ + +/*------------------------------------------------------------------------- + * Function: H5FD_read_selection_id + * + * Purpose: Like H5FD_read_selection(), but takes hid_t arrays instead + * of H5S_t * arrays for the dataspaces. + * + * Return: Success: SUCCEED + * All reads have completed successfully, and + * the results havce been into the supplied + * buffers. + * + * Failure: FAIL + * The contents of supplied buffers are undefined. + * + * Programmer: NAF -- 5/19/21 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +herr_t +H5FD_read_selection_id(H5FD_t *file, H5FD_mem_t type, uint32_t count, hid_t mem_space_ids[], + hid_t file_space_ids[], haddr_t offsets[], size_t element_sizes[], + void *bufs[] /* out */) +{ + hbool_t offsets_cooked = FALSE; + H5S_t * mem_spaces_local[H5FD_LOCAL_SEL_ARR_LEN]; + H5S_t ** mem_spaces = mem_spaces_local; + H5S_t * file_spaces_local[H5FD_LOCAL_SEL_ARR_LEN]; + H5S_t ** file_spaces = file_spaces_local; + hid_t dxpl_id = H5I_INVALID_HID; /* DXPL for operation */ + uint32_t i; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert((mem_space_ids) || (count == 0)); + HDassert((file_space_ids) || (count == 0)); + HDassert((offsets) || (count == 0)); + HDassert((element_sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* Verify that the first elements of the element_sizes and bufs arrays are + * valid. */ + HDassert((count == 0) || (element_sizes[0] != 0)); + HDassert((count == 0) || (bufs[0] != NULL)); + + /* Get proper DXPL for I/O */ + dxpl_id = H5CX_get_dxpl(); + +#ifndef H5_HAVE_PARALLEL + /* The no-op case + * + * Do not return early for Parallel mode since the I/O could be a + * collective transfer. + */ + if (0 == count) { + HGOTO_DONE(SUCCEED) + } +#endif /* H5_HAVE_PARALLEL */ + + if (file->base_addr > 0) { + + /* apply the base_addr offset to the offsets array. Must undo before + * we return. + */ + for (i = 0; i < count; i++) { + + offsets[i] += file->base_addr; + } + offsets_cooked = TRUE; + } + + /* If the file is open for SWMR read access, allow access to data past + * the end of the allocated space (the 'eoa'). This is done because the + * eoa stored in the file's superblock might be out of sync with the + * objects being written within the file by the application performing + * SWMR write operations. + */ + /* For now at least, only check that the offset is not past the eoa, since + * looking into the highest offset in the selection (different from the + * bounds) is potentially expensive. + */ + if (!(file->access_flags & H5F_ACC_SWMR_READ)) { + haddr_t eoa; + + if (HADDR_UNDEF == (eoa = (file->cls->get_eoa)(file, type))) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "driver get_eoa request failed") + + for (i = 0; i < count; i++) { + + if ((offsets[i]) > eoa) + + HGOTO_ERROR(H5E_ARGS, H5E_OVERFLOW, FAIL, "addr overflow, offsets[%d] = %llu, eoa = %llu", + (int)i, (unsigned long long)(offsets[i]), (unsigned long long)eoa) + } + } + + /* if the underlying VFD supports selection read, make the call */ + if (file->cls->read_selection) { + if ((file->cls->read_selection)(file, type, dxpl_id, count, mem_space_ids, file_space_ids, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "driver read selection request failed") + } + else { + /* Otherwise, implement the selection read as a sequence of regular + * or vector read calls. + */ + + /* Allocate arrays of space objects if necessary, otherwise use local + * buffers */ + if (count > sizeof(mem_spaces_local) / sizeof(mem_spaces_local[0])) { + if (NULL == (mem_spaces = H5MM_malloc(count * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + if (NULL == (file_spaces = H5MM_malloc(count * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + } + + /* Get object pointers for all dataspaces */ + for (i = 0; i < count; i++) { + if (NULL == (mem_spaces[i] = (H5S_t *)H5I_object_verify(mem_space_ids[i], H5I_DATASPACE))) + HGOTO_ERROR(H5E_VFL, H5E_BADTYPE, H5I_INVALID_HID, "can't retrieve memory dataspace from ID") + if (NULL == (file_spaces[i] = (H5S_t *)H5I_object_verify(file_space_ids[i], H5I_DATASPACE))) + HGOTO_ERROR(H5E_VFL, H5E_BADTYPE, H5I_INVALID_HID, "can't retrieve file dataspace from ID") + } + + /* Translate to vector or scalar I/O */ + if (H5FD__read_selection_translate(file, type, dxpl_id, count, mem_spaces, file_spaces, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "translation to vector or scalar read failed") + } + +done: + /* undo the base addr offset to the offsets array if necessary */ + if (offsets_cooked) { + + HDassert(file->base_addr > 0); + + for (i = 0; i < count; i++) { + + offsets[i] -= file->base_addr; + } + } + + /* Cleanup dataspace arrays */ + if (mem_spaces != mem_spaces_local) + mem_spaces = H5MM_xfree(mem_spaces); + if (file_spaces != file_spaces_local) + file_spaces = H5MM_xfree(file_spaces); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD_read_selection_id() */ + +/*------------------------------------------------------------------------- + * Function: H5FD__write_selection_translate + * + * Purpose: Translates a selection write call to a vector write call + * if vector writes are supported, or a series of scalar + * write calls otherwise. + * + * Return: Success: SUCCEED + * All writes have completed successfully. + * + * Failure: FAIL + * One or more writes failed. + * + * Programmer: NAF -- 5/13/21 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +static herr_t +H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count, + H5S_t **mem_spaces, H5S_t **file_spaces, haddr_t offsets[], + size_t element_sizes[], const void *bufs[]) +{ + hbool_t extend_sizes = FALSE; + hbool_t extend_bufs = FALSE; + uint32_t i; + size_t element_size; + const void * buf; + hbool_t use_vector = FALSE; + haddr_t addrs_local[H5FD_LOCAL_VECTOR_LEN]; + haddr_t * addrs = addrs_local; + size_t sizes_local[H5FD_LOCAL_VECTOR_LEN]; + size_t * sizes = sizes_local; + const void * vec_bufs_local[H5FD_LOCAL_VECTOR_LEN]; + const void ** vec_bufs = vec_bufs_local; + hsize_t file_off[H5FD_SEQ_LIST_LEN]; + size_t file_len[H5FD_SEQ_LIST_LEN]; + hsize_t mem_off[H5FD_SEQ_LIST_LEN]; + size_t mem_len[H5FD_SEQ_LIST_LEN]; + size_t file_seq_i; + size_t mem_seq_i; + size_t file_nseq; + size_t mem_nseq; + size_t io_len; + size_t nelmts; + hssize_t hss_nelmts; + size_t seq_nelem; + H5S_sel_iter_t *file_iter = NULL; + H5S_sel_iter_t *mem_iter = NULL; + hbool_t file_iter_init = FALSE; + hbool_t mem_iter_init = FALSE; + H5FD_mem_t types[2] = {type, H5FD_MEM_NOLIST}; + size_t vec_arr_nalloc = H5FD_LOCAL_VECTOR_LEN; + size_t vec_arr_nused = 0; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert(mem_spaces); + HDassert(file_spaces); + HDassert(offsets); + HDassert(element_sizes); + HDassert(bufs); + + /* Verify that the first elements of the element_sizes and bufs arrays are + * valid. */ + HDassert(element_sizes[0] != 0); + HDassert(bufs[0] != NULL); + + /* Check if we're using vector I/O */ + use_vector = file->cls->write_vector != NULL; + + /* Allocate sequence lists for memory and file spaces */ + if (NULL == (file_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_VFL, H5E_CANTALLOC, FAIL, "couldn't allocate file selection iterator") + if (NULL == (mem_iter = H5FL_MALLOC(H5S_sel_iter_t))) + HGOTO_ERROR(H5E_VFL, H5E_CANTALLOC, FAIL, "couldn't allocate memory selection iterator") + + /* Loop over dataspaces */ + for (i = 0; i < count; i++) { + + /* we have already verified that element_sizes[0] != 0 and bufs[0] + * != NULL */ + + if (!extend_sizes) { + + if (element_sizes[i] == 0) { + + extend_sizes = TRUE; + element_size = element_sizes[i - 1]; + } + else { + + element_size = element_sizes[i]; + } + } + + if (!extend_bufs) { + + if (bufs[i] == NULL) { + + extend_bufs = TRUE; + buf = bufs[i - 1]; + } + else { + + buf = bufs[i]; + } + } + + /* Initialize sequence lists for memory and file spaces */ + if (H5S_select_iter_init(file_iter, file_spaces[i], element_size, 0) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "can't initialize sequence list for file space") + file_iter_init = TRUE; + if (H5S_select_iter_init(mem_iter, mem_spaces[i], element_size, 0) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "can't initialize sequence list for memory space") + mem_iter_init = TRUE; + + /* Get the number of elements in selection */ + if ((hss_nelmts = (hssize_t)H5S_GET_SELECT_NPOINTS(file_spaces[i])) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTCOUNT, FAIL, "can't get number of elements selected") + H5_CHECKED_ASSIGN(nelmts, size_t, hss_nelmts, hssize_t); + +#ifndef NDEBUG + /* Verify mem space has the same number of elements */ + { + if ((hss_nelmts = (hssize_t)H5S_GET_SELECT_NPOINTS(mem_spaces[i])) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTCOUNT, FAIL, "can't get number of elements selected") + HDassert((hssize_t)nelmts == hss_nelmts); + } +#endif /* NDEBUG */ + + /* Initialize values so sequence lists are retrieved on the first + * iteration */ + file_seq_i = H5FD_SEQ_LIST_LEN; + mem_seq_i = H5FD_SEQ_LIST_LEN; + file_nseq = 0; + mem_nseq = 0; + + /* Loop until all elements are processed */ + while (file_seq_i < file_nseq || nelmts > 0) { + /* Fill/refill file sequence list if necessary */ + if (file_seq_i == H5FD_SEQ_LIST_LEN) { + if (H5S_SELECT_ITER_GET_SEQ_LIST(file_iter, H5FD_SEQ_LIST_LEN, SIZE_MAX, &file_nseq, + &seq_nelem, file_off, file_len) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed") + HDassert(file_nseq > 0); + + nelmts -= seq_nelem; + file_seq_i = 0; + } + HDassert(file_seq_i < file_nseq); + + /* Fill/refill memory sequence list if necessary */ + if (mem_seq_i == H5FD_SEQ_LIST_LEN) { + if (H5S_SELECT_ITER_GET_SEQ_LIST(mem_iter, H5FD_SEQ_LIST_LEN, SIZE_MAX, &mem_nseq, &seq_nelem, + mem_off, mem_len) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_UNSUPPORTED, FAIL, "sequence length generation failed") + HDassert(mem_nseq > 0); + + mem_seq_i = 0; + } + HDassert(mem_seq_i < mem_nseq); + + /* Calculate length of this IO */ + io_len = MIN(file_len[file_seq_i], mem_len[mem_seq_i]); + + /* Check if we're using vector I/O */ + if (use_vector) { + /* Check if we need to extend the arrays */ + if (vec_arr_nused == vec_arr_nalloc) { + /* Check if we're using the static arrays */ + if (addrs == addrs_local) { + HDassert(sizes == sizes_local); + HDassert(vec_bufs == vec_bufs_local); + + /* Allocate dynamic arrays */ + if (NULL == (addrs = H5MM_malloc(sizeof(addrs_local) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for address list") + if (NULL == (sizes = H5MM_malloc(sizeof(sizes_local) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for size list") + if (NULL == (vec_bufs = H5MM_malloc(sizeof(vec_bufs_local) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory allocation failed for buffer list") + + /* Copy the existing data */ + (void)H5MM_memcpy(addrs, addrs_local, sizeof(addrs_local)); + (void)H5MM_memcpy(sizes, sizes_local, sizeof(sizes_local)); + (void)H5MM_memcpy(vec_bufs, vec_bufs_local, sizeof(vec_bufs_local)); + } + else { + void *tmp_ptr; + + /* Reallocate arrays */ + if (NULL == (tmp_ptr = H5MM_realloc(addrs, vec_arr_nalloc * sizeof(*addrs) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory reallocation failed for address list") + addrs = tmp_ptr; + if (NULL == (tmp_ptr = H5MM_realloc(sizes, vec_arr_nalloc * sizeof(*sizes) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory reallocation failed for size list") + sizes = tmp_ptr; + if (NULL == + (tmp_ptr = H5MM_realloc(vec_bufs, vec_arr_nalloc * sizeof(*vec_bufs) * 2))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "memory reallocation failed for buffer list") + vec_bufs = tmp_ptr; + } + + /* Record that we've doubled the array sizes */ + vec_arr_nalloc *= 2; + } + + /* Add this segment to vector write list */ + addrs[vec_arr_nused] = offsets[i] + file_off[file_seq_i]; + sizes[vec_arr_nused] = io_len; + vec_bufs[vec_arr_nused] = (const void *)((const uint8_t *)buf + mem_off[mem_seq_i]); + vec_arr_nused++; + } + else + /* Issue scalar write call */ + if ((file->cls->write)(file, type, dxpl_id, offsets[i] + file_off[file_seq_i], io_len, + (const void *)((const uint8_t *)buf + mem_off[mem_seq_i])) < 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "driver write request failed") + + /* Update file sequence */ + if (io_len == file_len[file_seq_i]) + file_seq_i++; + else { + file_off[file_seq_i] += io_len; + file_len[file_seq_i] -= io_len; + } + + /* Update memory sequence */ + if (io_len == mem_len[mem_seq_i]) + mem_seq_i++; + else { + mem_off[mem_seq_i] += io_len; + mem_len[mem_seq_i] -= io_len; + } + } + + /* Make sure both memory and file sequences terminated at the same time */ + if (mem_seq_i < mem_nseq) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADVALUE, FAIL, "file selection terminated before memory selection") + + /* Terminate iterators */ + if (H5S_SELECT_ITER_RELEASE(file_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release file selection iterator") + file_iter_init = FALSE; + if (H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release memory selection iterator") + mem_iter_init = FALSE; + } + + /* Issue vector write call if appropriate */ + if (use_vector) { + H5_CHECK_OVERFLOW(vec_arr_nused, size_t, uint32_t) + if ((file->cls->write_vector)(file, dxpl_id, (uint32_t)vec_arr_nused, types, addrs, sizes, vec_bufs) < + 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "driver write vector request failed") + } + +done: + /* Terminate and free iterators */ + if (file_iter) { + if (file_iter_init && H5S_SELECT_ITER_RELEASE(file_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release file selection iterator") + file_iter = H5FL_FREE(H5S_sel_iter_t, file_iter); + } + if (mem_iter) { + if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTFREE, FAIL, "can't release memory selection iterator") + mem_iter = H5FL_FREE(H5S_sel_iter_t, mem_iter); + } + + /* Cleanup vector arrays */ + if (use_vector) { + if (addrs != addrs_local) + addrs = H5MM_xfree(addrs); + if (sizes != sizes_local) + sizes = H5MM_xfree(sizes); + if (vec_bufs != vec_bufs_local) + vec_bufs = H5MM_xfree(vec_bufs); + } + + /* Make sure we cleaned up */ + HDassert(!addrs || addrs == addrs_local); + HDassert(!sizes || sizes == sizes_local); + HDassert(!vec_bufs || vec_bufs == vec_bufs_local); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD__write_selection_translate() */ + +/*------------------------------------------------------------------------- + * Function: H5FD_write_selection + * + * Purpose: Private version of H5FDwrite_selection() + * + * Perform count writes to the specified file at the + * locations selected in the dataspaces in the file_spaces + * array, with each of those dataspaces starting at the file + * address specified by the corresponding element of the + * offsets array, and with the size of each element in the + * dataspace specified by the corresponding element of the + * element_sizes array. The memory type provided by type is + * the same for all selections. Data write is from + * the locations selected in the dataspaces in the + * mem_spaces array, within the buffers provided in the + * corresponding elements of the bufs array. + * + * If i > 0 and element_sizes[i] == 0, presume + * element_sizes[n] = element_sizes[i-1] for all n >= i and + * < count. + * + * If the underlying VFD supports selection writes, pass the + * call through directly. + * + * If it doesn't, convert the vector write into a sequence + * of individual writes. + * + * Return: Success: SUCCEED + * All writes have completed successfully. + * + * Failure: FAIL + * One or more writes failed. + * + * Programmer: NAF -- 3/29/21 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +herr_t +H5FD_write_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, H5S_t **mem_spaces, H5S_t **file_spaces, + haddr_t offsets[], size_t element_sizes[], const void *bufs[]) +{ + hbool_t offsets_cooked = FALSE; + hid_t mem_space_ids_local[H5FD_LOCAL_SEL_ARR_LEN]; + hid_t * mem_space_ids = mem_space_ids_local; + hid_t file_space_ids_local[H5FD_LOCAL_SEL_ARR_LEN]; + hid_t * file_space_ids = file_space_ids_local; + uint32_t num_spaces = 0; + hid_t dxpl_id = H5I_INVALID_HID; /* DXPL for operation */ + uint32_t i; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert((mem_spaces) || (count == 0)); + HDassert((file_spaces) || (count == 0)); + HDassert((offsets) || (count == 0)); + HDassert((element_sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* Verify that the first elements of the element_sizes and bufs arrays are + * valid. */ + HDassert((count == 0) || (element_sizes[0] != 0)); + HDassert((count == 0) || (bufs[0] != NULL)); + + /* Get proper DXPL for I/O */ + dxpl_id = H5CX_get_dxpl(); + +#ifndef H5_HAVE_PARALLEL + /* The no-op case + * + * Do not return early for Parallel mode since the I/O could be a + * collective transfer. + */ + if (0 == count) { + HGOTO_DONE(SUCCEED) + } +#endif /* H5_HAVE_PARALLEL */ + + if (file->base_addr > 0) { + + /* apply the base_addr offset to the offsets array. Must undo before + * we return. + */ + for (i = 0; i < count; i++) { + + offsets[i] += file->base_addr; + } + offsets_cooked = TRUE; + } + + /* For now at least, only check that the offset is not past the eoa, since + * looking into the highest offset in the selection (different from the + * bounds) is potentially expensive. + */ + { + haddr_t eoa; + + if (HADDR_UNDEF == (eoa = (file->cls->get_eoa)(file, type))) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "driver get_eoa request failed") + + for (i = 0; i < count; i++) { + + if ((offsets[i]) > eoa) + + HGOTO_ERROR(H5E_ARGS, H5E_OVERFLOW, FAIL, "addr overflow, offsets[%d] = %llu, eoa = %llu", + (int)i, (unsigned long long)(offsets[i]), (unsigned long long)eoa) + } + } + + /* if the underlying VFD supports selection write, make the call */ + if (file->cls->write_selection) { + /* Allocate array of space IDs if necessary, otherwise use local + * buffers */ + if (count > sizeof(mem_space_ids_local) / sizeof(mem_space_ids_local[0])) { + if (NULL == (mem_space_ids = H5MM_malloc(count * sizeof(hid_t)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + if (NULL == (file_space_ids = H5MM_malloc(count * sizeof(hid_t)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + } + + /* Create IDs for all dataspaces */ + for (; num_spaces < count; num_spaces++) { + if ((mem_space_ids[num_spaces] = H5I_register(H5I_DATASPACE, mem_spaces[num_spaces], TRUE)) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTREGISTER, FAIL, "unable to register dataspace ID") + + if ((file_space_ids[num_spaces] = H5I_register(H5I_DATASPACE, file_spaces[num_spaces], TRUE)) < + 0) { + if (H5I_dec_app_ref(mem_space_ids[num_spaces]) < 0) + HDONE_ERROR(H5E_VFL, H5E_CANTDEC, FAIL, "problem freeing id") + HGOTO_ERROR(H5E_VFL, H5E_CANTREGISTER, FAIL, "unable to register dataspace ID") + } + } + + if ((file->cls->write_selection)(file, type, dxpl_id, count, mem_space_ids, file_space_ids, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "driver write selection request failed") + } + else + /* Otherwise, implement the selection write as a sequence of regular + * or vector write calls. + */ + if (H5FD__write_selection_translate(file, type, dxpl_id, count, mem_spaces, file_spaces, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "translation to vector or scalar write failed") + +done: + /* undo the base addr offset to the offsets array if necessary */ + if (offsets_cooked) { + + HDassert(file->base_addr > 0); + + for (i = 0; i < count; i++) { + + offsets[i] -= file->base_addr; + } + } + + /* Cleanup dataspace arrays */ + for (i = 0; i < num_spaces; i++) { + if (H5I_dec_app_ref(mem_space_ids[i]) < 0) + HDONE_ERROR(H5E_VFL, H5E_CANTDEC, FAIL, "problem freeing id") + if (H5I_dec_app_ref(file_space_ids[i]) < 0) + HDONE_ERROR(H5E_VFL, H5E_CANTDEC, FAIL, "problem freeing id") + } + if (mem_space_ids != mem_space_ids_local) + mem_space_ids = H5MM_xfree(mem_space_ids); + if (file_space_ids != file_space_ids_local) + file_space_ids = H5MM_xfree(file_space_ids); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD_write_selection() */ + +/*------------------------------------------------------------------------- + * Function: H5FD_write_selection_id + * + * Purpose: Like H5FD_write_selection(), but takes hid_t arrays + * instead of H5S_t * arrays for the dataspaces. + * + * Return: Success: SUCCEED + * All writes have completed successfully. + * + * Failure: FAIL + * One or more writes failed. + * + * Programmer: NAF -- 5/19/21 + * + * Changes: None + * + *------------------------------------------------------------------------- + */ +herr_t +H5FD_write_selection_id(H5FD_t *file, H5FD_mem_t type, uint32_t count, hid_t mem_space_ids[], + hid_t file_space_ids[], haddr_t offsets[], size_t element_sizes[], const void *bufs[]) +{ + hbool_t offsets_cooked = FALSE; + H5S_t * mem_spaces_local[H5FD_LOCAL_SEL_ARR_LEN]; + H5S_t ** mem_spaces = mem_spaces_local; + H5S_t * file_spaces_local[H5FD_LOCAL_SEL_ARR_LEN]; + H5S_t ** file_spaces = file_spaces_local; + hid_t dxpl_id = H5I_INVALID_HID; /* DXPL for operation */ + uint32_t i; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(file); + HDassert(file->cls); + HDassert((mem_space_ids) || (count == 0)); + HDassert((file_space_ids) || (count == 0)); + HDassert((offsets) || (count == 0)); + HDassert((element_sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* Verify that the first elements of the element_sizes and bufs arrays are + * valid. */ + HDassert((count == 0) || (element_sizes[0] != 0)); + HDassert((count == 0) || (bufs[0] != NULL)); + + /* Get proper DXPL for I/O */ + dxpl_id = H5CX_get_dxpl(); + +#ifndef H5_HAVE_PARALLEL + /* The no-op case + * + * Do not return early for Parallel mode since the I/O could be a + * collective transfer. + */ + if (0 == count) { + HGOTO_DONE(SUCCEED) + } +#endif /* H5_HAVE_PARALLEL */ + + if (file->base_addr > 0) { + + /* apply the base_addr offset to the offsets array. Must undo before + * we return. + */ + for (i = 0; i < count; i++) { + + offsets[i] += file->base_addr; + } + offsets_cooked = TRUE; + } + + /* For now at least, only check that the offset is not past the eoa, since + * looking into the highest offset in the selection (different from the + * bounds) is potentially expensive. + */ + { + haddr_t eoa; + + if (HADDR_UNDEF == (eoa = (file->cls->get_eoa)(file, type))) + HGOTO_ERROR(H5E_VFL, H5E_CANTINIT, FAIL, "driver get_eoa request failed") + + for (i = 0; i < count; i++) { + + if ((offsets[i]) > eoa) + + HGOTO_ERROR(H5E_ARGS, H5E_OVERFLOW, FAIL, "addr overflow, offsets[%d] = %llu, eoa = %llu", + (int)i, (unsigned long long)(offsets[i]), (unsigned long long)eoa) + } + } + + /* if the underlying VFD supports selection write, make the call */ + if (file->cls->write_selection) { + if ((file->cls->write_selection)(file, type, dxpl_id, count, mem_space_ids, file_space_ids, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "driver write selection request failed") + } + else { + /* Otherwise, implement the selection write as a sequence of regular + * or vector write calls. + */ + + /* Allocate arrays of space objects if necessary, otherwise use local + * buffers */ + if (count > sizeof(mem_spaces_local) / sizeof(mem_spaces_local[0])) { + if (NULL == (mem_spaces = H5MM_malloc(count * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + if (NULL == (file_spaces = H5MM_malloc(count * sizeof(H5S_t *)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for dataspace list") + } + + /* Get object pointers for all dataspaces */ + for (i = 0; i < count; i++) { + if (NULL == (mem_spaces[i] = (H5S_t *)H5I_object_verify(mem_space_ids[i], H5I_DATASPACE))) + HGOTO_ERROR(H5E_VFL, H5E_BADTYPE, H5I_INVALID_HID, "can't retrieve memory dataspace from ID") + if (NULL == (file_spaces[i] = (H5S_t *)H5I_object_verify(file_space_ids[i], H5I_DATASPACE))) + HGOTO_ERROR(H5E_VFL, H5E_BADTYPE, H5I_INVALID_HID, "can't retrieve file dataspace from ID") + } + + /* Translate to vector or scalar I/O */ + if (H5FD__write_selection_translate(file, type, dxpl_id, count, mem_spaces, file_spaces, offsets, + element_sizes, bufs) < 0) + HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "translation to vector or scalar write failed") + } + +done: + /* undo the base addr offset to the offsets array if necessary */ + if (offsets_cooked) { + + HDassert(file->base_addr > 0); + + for (i = 0; i < count; i++) { + + offsets[i] -= file->base_addr; + } + } + + /* Cleanup dataspace arrays */ + if (mem_spaces != mem_spaces_local) + mem_spaces = H5MM_xfree(mem_spaces); + if (file_spaces != file_spaces_local) + file_spaces = H5MM_xfree(file_spaces); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD_write_selection_id() */ + +/*------------------------------------------------------------------------- * Function: H5FD_set_eoa * * Purpose: Private version of H5FDset_eoa() @@ -399,6 +2133,246 @@ H5FD_driver_query(const H5FD_class_t *driver, unsigned long *flags /*out*/) } /* end H5FD_driver_query() */ /*------------------------------------------------------------------------- + * Function: H5FD_sort_vector_io_req + * + * Purpose: Determine whether the supplied vector I/O request is + * sorted. + * + * if is is, set *vector_was_sorted to TRUE, set: + * + * *s_types_ptr = types + * *s_addrs_ptr = addrs + * *s_sizes_ptr = sizes + * *s_bufs_ptr = bufs + * + * and return. + * + * If it is not sorted, duplicate the type, addrs, sizes, + * and bufs vectors, storing the base addresses of the new + * vectors in *s_types_ptr, *s_addrs_ptr, *s_sizes_ptr, and + * *s_bufs_ptr respectively. Determine the sorted order + * of the vector I/O request, and load it into the new + * vectors in sorted order. + * + * Note that in this case, it is the callers responsibility + * to free the sorted vectors. + * + * JRM -- 3/15/21 + * + * Return: SUCCEED/FAIL + * + *------------------------------------------------------------------------- + */ + +static int +H5FD__vsrt_tmp_cmp(const void *element_1, const void *element_2) +{ + haddr_t addr_1 = ((const H5FD_vsrt_tmp_t *)element_1)->addr; + haddr_t addr_2 = ((const H5FD_vsrt_tmp_t *)element_2)->addr; + int ret_value = 0; /* Return value */ + + FUNC_ENTER_STATIC_NOERR + + /* Sanity checks */ + HDassert(H5F_addr_defined(addr_1)); + HDassert(H5F_addr_defined(addr_2)); + + /* Compare the addresses */ + if (H5F_addr_gt(addr_1, addr_2)) + ret_value = 1; + else if (H5F_addr_lt(addr_1, addr_2)) + ret_value = -1; + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5FD__vsrt_tmp_cmp() */ + +herr_t +H5FD_sort_vector_io_req(hbool_t *vector_was_sorted, uint32_t _count, H5FD_mem_t types[], haddr_t addrs[], + size_t sizes[], H5_flexible_const_ptr_t bufs[], H5FD_mem_t **s_types_ptr, + haddr_t **s_addrs_ptr, size_t **s_sizes_ptr, H5_flexible_const_ptr_t **s_bufs_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + size_t count = (size_t)_count; + size_t i; + struct H5FD_vsrt_tmp_t *srt_tmp = NULL; + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + + HDassert(vector_was_sorted); + + HDassert((types) || (count == 0)); + HDassert((addrs) || (count == 0)); + HDassert((sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* verify that the first elements of the sizes and types arrays are + * valid. + */ + HDassert((count == 0) || (sizes[0] != 0)); + HDassert((count == 0) || (types[0] != H5FD_MEM_NOLIST)); + + HDassert((count == 0) || ((s_types_ptr) && (NULL == *s_types_ptr))); + HDassert((count == 0) || ((s_addrs_ptr) && (NULL == *s_addrs_ptr))); + HDassert((count == 0) || ((s_sizes_ptr) && (NULL == *s_sizes_ptr))); + HDassert((count == 0) || ((s_bufs_ptr) && (NULL == *s_bufs_ptr))); + + /* scan the addrs array to see if it is sorted */ + for (i = 1; i < count; i++) { + HDassert(H5F_addr_defined(addrs[i - 1])); + + if (H5F_addr_gt(addrs[i - 1], addrs[i])) + break; + else if (H5F_addr_eq(addrs[i - 1], addrs[i])) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "duplicate addr in vector") + } + + /* if we traversed the entire array without breaking out, then + * the array was already sorted */ + if (i >= count) + *vector_was_sorted = TRUE; + else + *vector_was_sorted = FALSE; + + if (*vector_was_sorted) { + + *s_types_ptr = types; + *s_addrs_ptr = addrs; + *s_sizes_ptr = sizes; + *s_bufs_ptr = bufs; + } + else { + + /* must sort the addrs array in increasing addr order, while + * maintaining the association between each addr, and the + * sizes[], types[], and bufs[] values at the same index. + * + * Do this by allocating an array of struct H5FD_vsrt_tmp_t, where + * each instance of H5FD_vsrt_tmp_t has two fields, addr and index. + * Load the array with the contents of the addrs array and + * the index of the associated entry. Sort the array, allocate + * the s_types_ptr, s_addrs_ptr, s_sizes_ptr, and s_bufs_ptr + * arrays and populate them using the mapping provided by + * the sorted array of H5FD_vsrt_tmp_t. + */ + int j; + size_t fixed_size_index = count; + size_t fixed_type_index = count; + size_t srt_tmp_size; + + srt_tmp_size = (count * sizeof(struct H5FD_vsrt_tmp_t)); + + if (NULL == (srt_tmp = (H5FD_vsrt_tmp_t *)HDmalloc(srt_tmp_size))) + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "can't alloc srt_tmp") + + for (i = 0; i < count; i++) { + HDassert(i == (size_t)((int)i)); + + srt_tmp[i].addr = addrs[i]; + srt_tmp[i].index = (int)i; + } + + /* sort the srt_tmp array */ + HDqsort(srt_tmp, count, sizeof(struct H5FD_vsrt_tmp_t), H5FD__vsrt_tmp_cmp); + + /* verify no duplicate entries */ + i = 1; + + for (i = 1; i < count; i++) { + HDassert(H5F_addr_lt(srt_tmp[i - 1].addr, srt_tmp[i].addr)); + + if (H5F_addr_eq(addrs[i - 1], addrs[i])) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "duplicate addr in vector") + } + + if ((NULL == (*s_types_ptr = (H5FD_mem_t *)HDmalloc(count * sizeof(H5FD_mem_t)))) || + (NULL == (*s_addrs_ptr = (haddr_t *)HDmalloc(count * sizeof(haddr_t)))) || + (NULL == (*s_sizes_ptr = (size_t *)HDmalloc(count * sizeof(size_t)))) || + (NULL == + (*s_bufs_ptr = (H5_flexible_const_ptr_t *)HDmalloc(count * sizeof(H5_flexible_const_ptr_t))))) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "can't alloc sorted vector(s)") + } + + HDassert(sizes[0] != 0); + HDassert(types[0] != H5FD_MEM_NOLIST); + + /* Scan the sizes and types vectors to determine if the fixed size / type + * optimization is in use, and if so, to determine the index of the last + * valid value on each vector. We have already verified that the first + * elements of these arrays are valid so we can start at the second + * element (if it exists). + */ + for (i = 1; i < count && ((fixed_size_index == count) || (fixed_type_index == count)); i++) { + if ((fixed_size_index == count) && (sizes[i] == 0)) + fixed_size_index = i - 1; + if ((fixed_type_index == count) && (types[i] == H5FD_MEM_NOLIST)) + fixed_type_index = i - 1; + } + + HDassert(fixed_size_index <= count); + HDassert(fixed_type_index <= count); + + /* populate the sorted vectors */ + for (i = 0; i < count; i++) { + + j = srt_tmp[i].index; + + (*s_types_ptr)[j] = types[MIN(i, fixed_type_index)]; + (*s_addrs_ptr)[j] = addrs[i]; + (*s_sizes_ptr)[j] = sizes[MIN(i, fixed_size_index)]; + (*s_bufs_ptr)[j] = bufs[i]; + } + } + +done: + if (srt_tmp) { + + HDfree(srt_tmp); + srt_tmp = NULL; + } + + /* On failure, free the sorted vectors if they were allocated. + * Note that we only allocate these vectors if the original array + * was not sorted -- thus we check both for failure, and for + * the flag indicating that the original vector was not sorted + * in increasing address order. + */ + if ((ret_value != SUCCEED) && (!(*vector_was_sorted))) { + + /* free space allocated for sorted vectors */ + if (*s_types_ptr) { + + HDfree(*s_types_ptr); + *s_types_ptr = NULL; + } + + if (*s_addrs_ptr) { + + HDfree(*s_addrs_ptr); + *s_addrs_ptr = NULL; + } + + if (*s_sizes_ptr) { + + HDfree(*s_sizes_ptr); + *s_sizes_ptr = NULL; + } + + if (*s_bufs_ptr) { + + HDfree(*s_bufs_ptr); + *s_bufs_ptr = NULL; + } + } + + FUNC_LEAVE_NOAPI(ret_value) + +} /* end H5FD_sort_vector_io_req() */ + +/*------------------------------------------------------------------------- * Function: H5FD_delete * * Purpose: Private version of H5FDdelete() @@ -418,6 +2392,7 @@ H5FD_delete(const char *filename, hid_t fapl_id) FUNC_ENTER_NOAPI(FAIL) /* Sanity checks */ + HDassert(filename); /* Get file access property list */ diff --git a/src/H5FDlog.c b/src/H5FDlog.c index af34682..4d2e705 100644 --- a/src/H5FDlog.c +++ b/src/H5FDlog.c @@ -180,41 +180,46 @@ static herr_t H5FD__log_unlock(H5FD_t *_file); static herr_t H5FD__log_delete(const char *filename, hid_t fapl_id); static const H5FD_class_t H5FD_log_g = { - H5FD_LOG_VALUE, /* value */ - "log", /* name */ - MAXADDR, /* maxaddr */ - H5F_CLOSE_WEAK, /* fc_degree */ - H5FD__log_term, /* terminate */ - NULL, /* sb_size */ - NULL, /* sb_encode */ - NULL, /* sb_decode */ - sizeof(H5FD_log_fapl_t), /* fapl_size */ - H5FD__log_fapl_get, /* fapl_get */ - H5FD__log_fapl_copy, /* fapl_copy */ - H5FD__log_fapl_free, /* fapl_free */ - 0, /* dxpl_size */ - NULL, /* dxpl_copy */ - NULL, /* dxpl_free */ - H5FD__log_open, /* open */ - H5FD__log_close, /* close */ - H5FD__log_cmp, /* cmp */ - H5FD__log_query, /* query */ - NULL, /* get_type_map */ - H5FD__log_alloc, /* alloc */ - H5FD__log_free, /* free */ - H5FD__log_get_eoa, /* get_eoa */ - H5FD__log_set_eoa, /* set_eoa */ - H5FD__log_get_eof, /* get_eof */ - H5FD__log_get_handle, /* get_handle */ - H5FD__log_read, /* read */ - H5FD__log_write, /* write */ - NULL, /* flush */ - H5FD__log_truncate, /* truncate */ - H5FD__log_lock, /* lock */ - H5FD__log_unlock, /* unlock */ - H5FD__log_delete, /* del */ - NULL, /* ctl */ - H5FD_FLMAP_DICHOTOMY /* fl_map */ + H5FD_CLASS_VERSION, /* struct version */ + H5FD_LOG_VALUE, /* value */ + "log", /* name */ + MAXADDR, /* maxaddr */ + H5F_CLOSE_WEAK, /* fc_degree */ + H5FD__log_term, /* terminate */ + NULL, /* sb_size */ + NULL, /* sb_encode */ + NULL, /* sb_decode */ + sizeof(H5FD_log_fapl_t), /* fapl_size */ + H5FD__log_fapl_get, /* fapl_get */ + H5FD__log_fapl_copy, /* fapl_copy */ + H5FD__log_fapl_free, /* fapl_free */ + 0, /* dxpl_size */ + NULL, /* dxpl_copy */ + NULL, /* dxpl_free */ + H5FD__log_open, /* open */ + H5FD__log_close, /* close */ + H5FD__log_cmp, /* cmp */ + H5FD__log_query, /* query */ + NULL, /* get_type_map */ + H5FD__log_alloc, /* alloc */ + H5FD__log_free, /* free */ + H5FD__log_get_eoa, /* get_eoa */ + H5FD__log_set_eoa, /* set_eoa */ + H5FD__log_get_eof, /* get_eof */ + H5FD__log_get_handle, /* get_handle */ + H5FD__log_read, /* read */ + H5FD__log_write, /* write */ + NULL, /* read vector */ + NULL, /* write vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ + NULL, /* flush */ + H5FD__log_truncate, /* truncate */ + H5FD__log_lock, /* lock */ + H5FD__log_unlock, /* unlock */ + H5FD__log_delete, /* del */ + NULL, /* ctl */ + H5FD_FLMAP_DICHOTOMY /* fl_map */ }; /* Default configuration, if none provided */ diff --git a/src/H5FDmirror.c b/src/H5FDmirror.c index 84c744c..0ab5345 100644 --- a/src/H5FDmirror.c +++ b/src/H5FDmirror.c @@ -160,6 +160,7 @@ static herr_t H5FD__mirror_unlock(H5FD_t *_file); static herr_t H5FD__mirror_verify_reply(H5FD_mirror_t *file); static const H5FD_class_t H5FD_mirror_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_MIRROR_VALUE, /* value */ "mirror", /* name */ MAXADDR, /* maxaddr */ @@ -188,6 +189,10 @@ static const H5FD_class_t H5FD_mirror_g = { NULL, /* get_handle */ H5FD__mirror_read, /* read */ H5FD__mirror_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ NULL, /* flush */ H5FD__mirror_truncate, /* truncate */ H5FD__mirror_lock, /* lock */ @@ -225,9 +230,11 @@ H5FD_mirror_init(void) LOG_OP_CALL(__func__); - if (H5I_VFL != H5I_get_type(H5FD_MIRROR_g)) + if (H5I_VFL != H5I_get_type(H5FD_MIRROR_g)) { H5FD_MIRROR_g = H5FD_register(&H5FD_mirror_g, sizeof(H5FD_class_t), FALSE); - + if (H5I_INVALID_HID == H5FD_MIRROR_g) + HGOTO_ERROR(H5E_ID, H5E_CANTREGISTER, H5I_INVALID_HID, "unable to register mirror"); + } ret_value = H5FD_MIRROR_g; done: diff --git a/src/H5FDmirror_priv.h b/src/H5FDmirror_priv.h index 6a7b13e..f647c21 100644 --- a/src/H5FDmirror_priv.h +++ b/src/H5FDmirror_priv.h @@ -28,10 +28,10 @@ extern "C" { * = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = */ -/* The maximum allowed size for a receiving buffer when accepting bytes to +/* Define the maximum allowed size for a receiving buffer when accepting bytes to * write. Writes larger than this size are performed by multiple accept-write * steps by the Writer. */ -#define H5FD_MIRROR_DATA_BUFFER_MAX H5_GB /* 1 Gigabyte */ +#define H5FD_MIRROR_DATA_BUFFER_MAX (1024 * 1024 * 1024) /* 1 Gigabyte */ #define H5FD_MIRROR_XMIT_CURR_VERSION 1 #define H5FD_MIRROR_XMIT_MAGIC 0x87F8005B diff --git a/src/H5FDmpio.c b/src/H5FDmpio.c index c72578d..2a5e462 100644 --- a/src/H5FDmpio.c +++ b/src/H5FDmpio.c @@ -87,49 +87,66 @@ static herr_t H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, ha void *buf); static herr_t H5FD__mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr, size_t size, const void *buf); +static herr_t H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t count, + H5FD_mem_t types[], haddr_t addrs[], size_t sizes[], void *bufs[]); +static herr_t H5FD__mpio_write_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t count, + H5FD_mem_t types[], haddr_t addrs[], size_t sizes[], + const void *bufs[]); static herr_t H5FD__mpio_flush(H5FD_t *_file, hid_t dxpl_id, hbool_t closing); static herr_t H5FD__mpio_truncate(H5FD_t *_file, hid_t dxpl_id, hbool_t closing); static herr_t H5FD__mpio_delete(const char *filename, hid_t fapl_id); static herr_t H5FD__mpio_ctl(H5FD_t *_file, uint64_t op_code, uint64_t flags, const void *input, void **output); +/* Other functions */ +static herr_t H5FD__mpio_vector_build_types( + uint32_t count, H5FD_mem_t types[], haddr_t addrs[], size_t sizes[], H5_flexible_const_ptr_t bufs[], + haddr_t *s_addrs[], size_t *s_sizes[], H5_flexible_const_ptr_t *s_bufs[], hbool_t *vector_was_sorted, + MPI_Offset *mpi_off, H5_flexible_const_ptr_t *mpi_bufs_base, int *size_i, MPI_Datatype *buf_type, + hbool_t *buf_type_created, MPI_Datatype *file_type, hbool_t *file_type_created, char *unused); + /* The MPIO file driver information */ static const H5FD_class_t H5FD_mpio_g = { - H5_VFD_MPIO, /* value */ - "mpio", /* name */ - HADDR_MAX, /* maxaddr */ - H5F_CLOSE_SEMI, /* fc_degree */ - H5FD__mpio_term, /* terminate */ - NULL, /* sb_size */ - NULL, /* sb_encode */ - NULL, /* sb_decode */ - 0, /* fapl_size */ - NULL, /* fapl_get */ - NULL, /* fapl_copy */ - NULL, /* fapl_free */ - 0, /* dxpl_size */ - NULL, /* dxpl_copy */ - NULL, /* dxpl_free */ - H5FD__mpio_open, /* open */ - H5FD__mpio_close, /* close */ - NULL, /* cmp */ - H5FD__mpio_query, /* query */ - NULL, /* get_type_map */ - NULL, /* alloc */ - NULL, /* free */ - H5FD__mpio_get_eoa, /* get_eoa */ - H5FD__mpio_set_eoa, /* set_eoa */ - H5FD__mpio_get_eof, /* get_eof */ - H5FD__mpio_get_handle, /* get_handle */ - H5FD__mpio_read, /* read */ - H5FD__mpio_write, /* write */ - H5FD__mpio_flush, /* flush */ - H5FD__mpio_truncate, /* truncate */ - NULL, /* lock */ - NULL, /* unlock */ - H5FD__mpio_delete, /* del */ - H5FD__mpio_ctl, /* ctl */ - H5FD_FLMAP_DICHOTOMY /* fl_map */ + H5FD_CLASS_VERSION, /* struct version */ + H5_VFD_MPIO, /* value */ + "mpio", /* name */ + HADDR_MAX, /* maxaddr */ + H5F_CLOSE_SEMI, /* fc_degree */ + H5FD__mpio_term, /* terminate */ + NULL, /* sb_size */ + NULL, /* sb_encode */ + NULL, /* sb_decode */ + 0, /* fapl_size */ + NULL, /* fapl_get */ + NULL, /* fapl_copy */ + NULL, /* fapl_free */ + 0, /* dxpl_size */ + NULL, /* dxpl_copy */ + NULL, /* dxpl_free */ + H5FD__mpio_open, /* open */ + H5FD__mpio_close, /* close */ + NULL, /* cmp */ + H5FD__mpio_query, /* query */ + NULL, /* get_type_map */ + NULL, /* alloc */ + NULL, /* free */ + H5FD__mpio_get_eoa, /* get_eoa */ + H5FD__mpio_set_eoa, /* set_eoa */ + H5FD__mpio_get_eof, /* get_eof */ + H5FD__mpio_get_handle, /* get_handle */ + H5FD__mpio_read, /* read */ + H5FD__mpio_write, /* write */ + H5FD__mpio_read_vector, /* read_vector */ + H5FD__mpio_write_vector, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ + H5FD__mpio_flush, /* flush */ + H5FD__mpio_truncate, /* truncate */ + NULL, /* lock */ + NULL, /* unlock */ + H5FD__mpio_delete, /* del */ + H5FD__mpio_ctl, /* ctl */ + H5FD_FLMAP_DICHOTOMY /* fl_map */ }; #ifdef H5FDmpio_DEBUG @@ -188,6 +205,41 @@ H5FD__mpio_parse_debug_str(const char *s) FUNC_LEAVE_NOAPI_VOID } /* end H5FD__mpio_parse_debug_str() */ + +/*--------------------------------------------------------------------------- + * Function: H5FD__mem_t_to_str + * + * Purpose: Returns a string representing the enum value in an H5FD_mem_t + * enum + * + * Returns: H5FD_mem_t enum value string + * + *--------------------------------------------------------------------------- + */ +static const char * +H5FD__mem_t_to_str(H5FD_mem_t mem_type) +{ + switch (mem_type) { + case H5FD_MEM_NOLIST: + return "H5FD_MEM_NOLIST"; + case H5FD_MEM_DEFAULT: + return "H5FD_MEM_DEFAULT"; + case H5FD_MEM_SUPER: + return "H5FD_MEM_SUPER"; + case H5FD_MEM_BTREE: + return "H5FD_MEM_BTREE"; + case H5FD_MEM_DRAW: + return "H5FD_MEM_DRAW"; + case H5FD_MEM_GHEAP: + return "H5FD_MEM_GHEAP"; + case H5FD_MEM_LHEAP: + return "H5FD_MEM_LHEAP"; + case H5FD_MEM_OHDR: + return "H5FD_MEM_OHDR"; + default: + return "(Unknown)"; + } +} #endif /* H5FDmpio_DEBUG */ /*------------------------------------------------------------------------- @@ -864,14 +916,19 @@ H5FD__mpio_open(const char *name, unsigned flags, hid_t fapl_id, haddr_t H5_ATTR file->mpi_size = mpi_size; /* Only processor p0 will get the filesize and broadcast it. */ - if (mpi_rank == 0) + if (mpi_rank == 0) { + /* If MPI_File_get_size fails, broadcast file size as -1 to signal error */ if (MPI_SUCCESS != (mpi_code = MPI_File_get_size(fh, &file_size))) - HMPI_GOTO_ERROR(NULL, "MPI_File_get_size failed", mpi_code) + file_size = (MPI_Offset)-1; + } /* Broadcast file size */ if (MPI_SUCCESS != (mpi_code = MPI_Bcast(&file_size, (int)sizeof(MPI_Offset), MPI_BYTE, 0, comm))) HMPI_GOTO_ERROR(NULL, "MPI_Bcast failed", mpi_code) + if (file_size < 0) + HMPI_GOTO_ERROR(NULL, "MPI_File_get_size failed", mpi_code) + /* Determine if the file should be truncated */ if (file_size && (flags & H5F_ACC_TRUNC)) { /* Truncate the file */ @@ -989,7 +1046,6 @@ H5FD__mpio_query(const H5FD_t H5_ATTR_UNUSED *_file, unsigned long *flags /* out *flags |= H5FD_FEAT_AGGREGATE_METADATA; /* OK to aggregate metadata allocations */ *flags |= H5FD_FEAT_AGGREGATE_SMALLDATA; /* OK to aggregate "small" raw data allocations */ *flags |= H5FD_FEAT_HAS_MPI; /* This driver uses MPI */ - *flags |= H5FD_FEAT_ALLOCATE_EARLY; /* Allocate space early instead of late */ *flags |= H5FD_FEAT_DEFAULT_VFD_COMPATIBLE; /* VFD creates a file which can be opened with the default VFD */ } /* end if */ @@ -1155,7 +1211,7 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU MPI_Status mpi_stat; /* Status from I/O operation */ MPI_Datatype buf_type = MPI_BYTE; /* MPI description of the selection in memory */ int size_i; /* Integer copy of 'size' to read */ -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) MPI_Count bytes_read = 0; /* Number of bytes read in */ MPI_Count type_size; /* MPI datatype used for I/O's size */ MPI_Count io_size; /* Actual number of bytes requested */ @@ -1167,6 +1223,7 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU int n; #endif hbool_t use_view_this_time = FALSE; + hbool_t derived_type = FALSE; hbool_t rank0_bcast = FALSE; /* If read-with-rank0-and-bcast flag was used */ #ifdef H5FDmpio_DEBUG hbool_t H5FD_mpio_debug_t_flag = (H5FD_mpio_debug_flags_s[(int)'t'] && H5FD_MPIO_TRACE_THIS_RANK(file)); @@ -1194,8 +1251,6 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU if (H5FD_mpi_haddr_to_MPIOff(addr, &mpi_off /*out*/) < 0) HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't convert from haddr to MPI off") size_i = (int)size; - if ((hsize_t)size_i != size) - HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't convert from size to size_i") /* Only look for MPI views for raw data transfers */ if (type == H5FD_MEM_DRAW) { @@ -1262,10 +1317,14 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU rank0_bcast = TRUE; /* Read on rank 0 Bcast to other ranks */ - if (file->mpi_rank == 0) + if (file->mpi_rank == 0) { + /* If MPI_File_read_at fails, push an error, but continue + * to participate in following MPI_Bcast */ if (MPI_SUCCESS != (mpi_code = MPI_File_read_at(file->f, mpi_off, buf, size_i, buf_type, &mpi_stat))) - HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code) + HMPI_DONE_ERROR(FAIL, "MPI_File_read_at failed", mpi_code) + } + if (MPI_SUCCESS != (mpi_code = MPI_Bcast(buf, size_i, buf_type, 0, file->comm))) HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code) } /* end if */ @@ -1295,6 +1354,21 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code) } /* end if */ else { + if (size != (hsize_t)size_i) { + /* If HERE, then we need to work around the integer size limit + * of 2GB. The input size_t size variable cannot fit into an integer, + * but we can get around that limitation by creating a different datatype + * and then setting the integer size (or element count) to 1 when using + * the derived_type. + */ + + if (H5_mpio_create_large_type(size, 0, MPI_BYTE, &buf_type) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") + + derived_type = TRUE; + size_i = 1; + } + #ifdef H5FDmpio_DEBUG if (H5FD_mpio_debug_r_flag) HDfprintf(stderr, "%s: (%d) doing MPI independent IO\n", __func__, file->mpi_rank); @@ -1308,12 +1382,22 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU /* Only retrieve bytes read if this rank _actually_ participated in I/O */ if (!rank0_bcast || (rank0_bcast && file->mpi_rank == 0)) { /* How many bytes were actually read? */ -#if MPI_VERSION >= 3 - if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_read))) +#if H5_CHECK_MPI_VERSION(3, 0) + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_read))) { #else - if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read))) + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read))) { #endif - HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) + if (rank0_bcast && file->mpi_rank == 0) { + /* If MPI_Get_elements(_x) fails for a rank 0 bcast strategy, + * push an error, but continue to participate in the following + * MPI_Bcast. + */ + bytes_read = -1; + HMPI_DONE_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) + } + else + HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) + } } /* end if */ /* If the rank0-bcast feature was used, broadcast the # of bytes read to @@ -1323,7 +1407,7 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU * of the data. (QAK - 2019/1/2) */ if (rank0_bcast) -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) if (MPI_SUCCESS != MPI_Bcast(&bytes_read, 1, MPI_COUNT, 0, file->comm)) #else if (MPI_SUCCESS != MPI_Bcast(&bytes_read, 1, MPI_INT, 0, file->comm)) @@ -1331,7 +1415,7 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", 0) /* Get the type's size */ -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) if (MPI_SUCCESS != (mpi_code = MPI_Type_size_x(buf_type, &type_size))) #else if (MPI_SUCCESS != (mpi_code = MPI_Type_size(buf_type, &type_size))) @@ -1347,8 +1431,8 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU #ifdef H5FDmpio_DEBUG if (H5FD_mpio_debug_r_flag) - HDfprintf(stderr, "%s: (%d) mpi_off = %ld bytes_read = %lld\n", __func__, file->mpi_rank, - (long)mpi_off, bytes_read); + HDfprintf(stderr, "%s: (%d) mpi_off = %ld bytes_read = %lld type = %s\n", __func__, file->mpi_rank, + (long)mpi_off, (long long)bytes_read, H5FD__mem_t_to_str(type)); #endif /* @@ -1358,6 +1442,9 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU HDmemset((char *)buf + bytes_read, 0, (size_t)n); done: + if (derived_type) + MPI_Type_free(&buf_type); + #ifdef H5FDmpio_DEBUG if (H5FD_mpio_debug_t_flag) HDfprintf(stderr, "%s: (%d) Leaving\n", __func__, file->mpi_rank); @@ -1395,7 +1482,7 @@ H5FD__mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t H5_ATTR_UNUSED dxpl_id, h MPI_Offset mpi_off; MPI_Status mpi_stat; /* Status from I/O operation */ MPI_Datatype buf_type = MPI_BYTE; /* MPI description of the selection in memory */ -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) MPI_Count bytes_written; MPI_Count type_size; /* MPI datatype used for I/O's size */ MPI_Count io_size; /* Actual number of bytes requested */ @@ -1470,20 +1557,6 @@ H5FD__mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t H5_ATTR_UNUSED dxpl_id, h */ mpi_off = 0; } /* end if */ - else if (size != (hsize_t)size_i) { - /* If HERE, then we need to work around the integer size limit - * of 2GB. The input size_t size variable cannot fit into an integer, - * but we can get around that limitation by creating a different datatype - * and then setting the integer size (or element count) to 1 when using - * the derived_type. - */ - - if (H5_mpio_create_large_type(size, 0, MPI_BYTE, &buf_type) < 0) - HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") - - derived_type = TRUE; - size_i = 1; - } /* Write the data. */ if (use_view_this_time) { @@ -1529,6 +1602,21 @@ H5FD__mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t H5_ATTR_UNUSED dxpl_id, h HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code) } /* end if */ else { + if (size != (hsize_t)size_i) { + /* If HERE, then we need to work around the integer size limit + * of 2GB. The input size_t size variable cannot fit into an integer, + * but we can get around that limitation by creating a different datatype + * and then setting the integer size (or element count) to 1 when using + * the derived_type. + */ + + if (H5_mpio_create_large_type(size, 0, MPI_BYTE, &buf_type) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") + + derived_type = TRUE; + size_i = 1; + } + #ifdef H5FDmpio_DEBUG if (H5FD_mpio_debug_w_flag) HDfprintf(stderr, "%s: (%d) doing MPI independent IO\n", __func__, file->mpi_rank); @@ -1540,7 +1628,7 @@ H5FD__mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t H5_ATTR_UNUSED dxpl_id, h } /* end else */ /* How many bytes were actually written? */ -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_written))) #else if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_written))) @@ -1548,7 +1636,7 @@ H5FD__mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t H5_ATTR_UNUSED dxpl_id, h HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) /* Get the type's size */ -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) if (MPI_SUCCESS != (mpi_code = MPI_Type_size_x(buf_type, &type_size))) #else if (MPI_SUCCESS != (mpi_code = MPI_Type_size(buf_type, &type_size))) @@ -1564,8 +1652,8 @@ H5FD__mpio_write(H5FD_t *_file, H5FD_mem_t type, hid_t H5_ATTR_UNUSED dxpl_id, h #ifdef H5FDmpio_DEBUG if (H5FD_mpio_debug_w_flag) - HDfprintf(stderr, "%s: (%d) mpi_off = %ld bytes_written = %lld\n", __func__, file->mpi_rank, - (long)mpi_off, bytes_written); + HDfprintf(stderr, "%s: (%d) mpi_off = %ld bytes_written = %lld type = %s\n", __func__, + file->mpi_rank, (long)mpi_off, (long long)bytes_written, H5FD__mem_t_to_str(type)); #endif /* Each process will keep track of its perceived EOF value locally, and @@ -1592,6 +1680,1050 @@ done: } /* end H5FD__mpio_write() */ /*------------------------------------------------------------------------- + * Function: H5FD__mpio_vector_build_types + * + * Purpose: Build MPI datatypes and calculate offset, base buffer, and + * size for MPIO vector I/O. Spun off from common code in + * H5FD__mpio_vector_read() and H5FD__mpio_vector_write(). + * + * Return: Success: SUCCEED. + * Failure: FAIL. + * + * Programmer: Neil Fortner + * March 14, 2022 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5FD__mpio_vector_build_types(uint32_t count, H5FD_mem_t types[], haddr_t addrs[], size_t sizes[], + H5_flexible_const_ptr_t bufs[], haddr_t *s_addrs[], size_t *s_sizes[], + H5_flexible_const_ptr_t *s_bufs[], hbool_t *vector_was_sorted, + MPI_Offset *mpi_off, H5_flexible_const_ptr_t *mpi_bufs_base, int *size_i, + MPI_Datatype *buf_type, hbool_t *buf_type_created, MPI_Datatype *file_type, + hbool_t *file_type_created, char *unused) +{ + hsize_t bigio_count; /* Transition point to create derived type */ + hbool_t fixed_size = FALSE; + size_t size; + H5FD_mem_t * s_types = NULL; + int * mpi_block_lengths = NULL; + MPI_Aint mpi_bufs_base_Aint; + MPI_Aint * mpi_bufs = NULL; + MPI_Aint * mpi_displacements = NULL; + MPI_Datatype *sub_types = NULL; + uint8_t * sub_types_created = NULL; + int i; + int j; + int mpi_code; /* MPI return code */ + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(s_sizes); + HDassert(s_bufs); + HDassert(vector_was_sorted); + HDassert(*vector_was_sorted); + HDassert(mpi_off); + HDassert(mpi_bufs_base); + HDassert(size_i); + HDassert(buf_type); + HDassert(buf_type_created); + HDassert(!*buf_type_created); + HDassert(file_type); + HDassert(file_type_created); + HDassert(!*file_type_created); + HDassert(unused); + + /* Get bio I/O transition point (may be lower than 2G for testing) */ + bigio_count = H5_mpi_get_bigio_count(); + + if (count == 1) { + /* Single block. Just use a series of MPI_BYTEs for the file view. + */ + *size_i = (int)sizes[0]; + *buf_type = MPI_BYTE; + *file_type = MPI_BYTE; + *mpi_bufs_base = bufs[0]; + + /* Setup s_addrs, s_sizes and s_bufs (needed for incomplete read filling code and eof + * calculation code) */ + *s_addrs = addrs; + *s_sizes = sizes; + *s_bufs = bufs; + + /* some numeric conversions */ + if (H5FD_mpi_haddr_to_MPIOff(addrs[0], mpi_off) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI offset") + + /* Check for size overflow */ + if (sizes[0] > bigio_count) { + /* We need to work around the integer size limit of 2GB. The input size_t size + * variable cannot fit into an integer, but we can get around that limitation by + * creating a different datatype and then setting the integer size (or element + * count) to 1 when using the derived_type. */ + + if (H5_mpio_create_large_type(sizes[0], 0, MPI_BYTE, buf_type) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") + *buf_type_created = TRUE; + + if (H5_mpio_create_large_type(sizes[0], 0, MPI_BYTE, file_type) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") + *file_type_created = TRUE; + + *size_i = 1; + } + } + else if (count > 0) { /* create MPI derived types describing the vector write */ + + /* sort the vector I/O request into increasing address order if required + * + * If the vector is already sorted, the base addresses of types, addrs, sizes, + * and bufs will be returned in s_types, s_addrs, s_sizes, and s_bufs respectively. + * + * If the vector was not already sorted, new, sorted versions of types, addrs, sizes, and bufs + * are allocated, populated, and returned in s_types, s_addrs, s_sizes, and s_bufs respectively. + * In this case, this function must free the memory allocated for the sorted vectors. + */ + if (H5FD_sort_vector_io_req(vector_was_sorted, count, types, addrs, sizes, bufs, &s_types, s_addrs, + s_sizes, s_bufs) < 0) + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "can't sort vector I/O request") + + if ((NULL == (mpi_block_lengths = (int *)HDmalloc((size_t)count * sizeof(int)))) || + (NULL == (mpi_displacements = (MPI_Aint *)HDmalloc((size_t)count * sizeof(MPI_Aint)))) || + (NULL == (mpi_bufs = (MPI_Aint *)HDmalloc((size_t)count * sizeof(MPI_Aint))))) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "can't alloc mpi block lengths / displacement") + } + + /* when we setup mpi_bufs[] below, all addresses are offsets from + * mpi_bufs_base. + * + * Since these offsets must all be positive, we must scan through + * s_bufs[] to find the smallest value, and choose that for + * mpi_bufs_base. + */ + + j = 0; /* guess at the index of the smallest value of s_bufs[] */ + + for (i = 1; i < (int)count; i++) { + + if ((*s_bufs)[i].cvp < (*s_bufs)[j].cvp) { + + j = i; + } + } + + *mpi_bufs_base = (*s_bufs)[j]; + + if (MPI_SUCCESS != (mpi_code = MPI_Get_address(mpi_bufs_base->cvp, &mpi_bufs_base_Aint))) + + HMPI_GOTO_ERROR(FAIL, "MPI_Get_address for s_bufs[] to mpi_bufs_base failed", mpi_code) + + *size_i = 1; + + fixed_size = FALSE; + + /* load the mpi_block_lengths and mpi_displacements arrays */ + for (i = 0; i < (int)count; i++) { + /* Determine size of this vector element */ + if (!fixed_size) { + if ((*s_sizes)[i] == 0) { + HDassert(vector_was_sorted); + fixed_size = TRUE; + size = sizes[i - 1]; + } + else { + size = (*s_sizes)[i]; + } + } + + /* Add to block lengths and displacements arrays */ + mpi_block_lengths[i] = (int)size; + mpi_displacements[i] = (MPI_Aint)(*s_addrs)[i]; + + /* convert s_bufs[i] to MPI_Aint... */ + if (MPI_SUCCESS != (mpi_code = MPI_Get_address((*s_bufs)[i].cvp, &(mpi_bufs[i])))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_address for s_bufs[] - mpi_bufs_base failed", mpi_code) + + /*... and then subtract mpi_bufs_base_Aint from it. */ +#if ((MPI_VERSION > 3) || ((MPI_VERSION == 3) && (MPI_SUBVERSION >= 1))) + mpi_bufs[i] = MPI_Aint_diff(mpi_bufs[i], mpi_bufs_base_Aint); +#else + mpi_bufs[i] = mpi_bufs[i] - mpi_bufs_base_Aint; +#endif + + /* Check for size overflow */ + if (size > bigio_count) { + /* We need to work around the integer size limit of 2GB. The input size_t size + * variable cannot fit into an integer, but we can get around that limitation by + * creating a different datatype and then setting the integer size (or element + * count) to 1 when using the derived_type. */ + + /* Allocate arrays to keep track of types and whether they were created, if + * necessary */ + if (!sub_types) { + HDassert(!sub_types_created); + + if (NULL == (sub_types = (int *)HDmalloc((size_t)count * sizeof(MPI_Datatype)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "can't alloc sub types array") + if (NULL == (sub_types_created = (uint8_t *)HDcalloc((size_t)count, 1))) { + H5MM_free(sub_types); + sub_types = NULL; + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "can't alloc sub types created array") + } + + /* Initialize sub_types to all MPI_BYTE */ + for (j = 0; j < (int)count; j++) + sub_types[j] = MPI_BYTE; + } + HDassert(sub_types_created); + + /* Create type for large block */ + if (H5_mpio_create_large_type(size, 0, MPI_BYTE, &sub_types[i]) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") + sub_types_created[i] = TRUE; + + /* Only one of these large types for this vector element */ + mpi_block_lengths[i] = 1; + } + else + HDassert(size == (size_t)mpi_block_lengths[i]); + } + + /* create the memory MPI derived types */ + if (sub_types) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_create_struct((int)count, mpi_block_lengths, mpi_bufs, + sub_types, buf_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct for buf_type failed", mpi_code) + } + else if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed((int)count, mpi_block_lengths, mpi_bufs, + MPI_BYTE, buf_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed for buf_type failed", mpi_code) + + *buf_type_created = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(buf_type))) + + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit for buf_type failed", mpi_code) + + /* create the file MPI derived type */ + if (sub_types) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_create_struct((int)count, mpi_block_lengths, + mpi_displacements, sub_types, file_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct for file_type failed", mpi_code) + } + else if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed((int)count, mpi_block_lengths, + mpi_displacements, MPI_BYTE, file_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed for file_type failed", mpi_code) + + *file_type_created = TRUE; + + if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(file_type))) + + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit for file_type failed", mpi_code) + + /* Free up memory used to build types */ + HDassert(mpi_block_lengths); + HDfree(mpi_block_lengths); + mpi_block_lengths = NULL; + + HDassert(mpi_displacements); + HDfree(mpi_displacements); + mpi_displacements = NULL; + + HDassert(mpi_bufs); + HDfree(mpi_bufs); + mpi_bufs = NULL; + + if (sub_types) { + HDassert(sub_types); + + for (i = 0; i < (int)count; i++) + if (sub_types_created[i]) + MPI_Type_free(&sub_types[i]); + + HDfree(sub_types); + sub_types = NULL; + HDfree(sub_types_created); + sub_types_created = NULL; + } + + /* some numeric conversions */ + if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, mpi_off) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0") + } + else { + /* setup for null participation in the collective operation. */ + *buf_type = MPI_BYTE; + *file_type = MPI_BYTE; + + /* Set non-NULL pointer for I/O operation */ + mpi_bufs_base->vp = unused; + + /* MPI count to read */ + *size_i = 0; + + /* some numeric conversions */ + if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, mpi_off) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0") + } + +done: + /* free sorted vectors if they exist */ + if (!vector_was_sorted) + if (s_types) { + HDfree(s_types); + s_types = NULL; + } + + /* Clean up on error */ + if (ret_value < 0) { + if (mpi_block_lengths) { + HDfree(mpi_block_lengths); + mpi_block_lengths = NULL; + } + + if (mpi_displacements) { + HDfree(mpi_displacements); + mpi_displacements = NULL; + } + + if (mpi_bufs) { + HDfree(mpi_bufs); + mpi_bufs = NULL; + } + + if (sub_types) { + HDassert(sub_types_created); + + for (i = 0; i < (int)count; i++) + if (sub_types_created[i]) + MPI_Type_free(&sub_types[i]); + + HDfree(sub_types); + sub_types = NULL; + HDfree(sub_types_created); + sub_types_created = NULL; + } + } + + /* Make sure we cleaned up */ + HDassert(!mpi_block_lengths); + HDassert(!mpi_displacements); + HDassert(!mpi_bufs); + HDassert(!sub_types); + HDassert(!sub_types_created); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD__mpio_vector_build_types() */ + +/*------------------------------------------------------------------------- + * Function: H5FD__mpio_read_vector() + * + * Purpose: The behaviour of this function dependes on the value of + * the io_xfer_mode obtained from the context. + * + * If it is H5FD_MPIO_COLLECTIVE, this is a collective + * operation, which allows us to use MPI_File_set_view, and + * then perform the entire vector read in a single MPI call. + * + * Do this (if count is positive), by constructing memory + * and file derived types from the supplied vector, using + * file type to set the file view, and then reading the + * the memory type from file. Note that this read is + * either independent or collective depending on the + * value of mpio_coll_opt -- again obtained from the context. + * + * If count is zero, participate in the collective read + * (if so configured) with an empty read. + * + * Finally, set the file view back to its default state. + * + * In contrast, if io_xfer_mode is H5FD_MPIO_INDEPENDENT, + * this call is independent, and thus we cannot use + * MPI_File_set_view(). + * + * In this case, simply walk the vector, and issue an + * independent read for each entry. + * + * Return: Success: SUCCEED. + * Failure: FAIL. + * + * Programmer: John Mainzer + * March 15, 2021 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t count, H5FD_mem_t types[], + haddr_t addrs[], size_t sizes[], void *bufs[]) +{ + H5FD_mpio_t * file = (H5FD_mpio_t *)_file; + hbool_t vector_was_sorted = TRUE; + haddr_t * s_addrs = NULL; + size_t * s_sizes = NULL; + void ** s_bufs = NULL; + char unused = 0; /* Unused, except for non-NULL pointer value */ + void * mpi_bufs_base = NULL; + MPI_Datatype buf_type = MPI_BYTE; /* MPI description of the selection in memory */ + hbool_t buf_type_created = FALSE; + MPI_Datatype file_type = MPI_BYTE; /* MPI description of the selection in file */ + hbool_t file_type_created = FALSE; + int i; + int mpi_code; /* MPI return code */ + MPI_Offset mpi_off = 0; + MPI_Status mpi_stat; /* Status from I/O operation */ + H5FD_mpio_xfer_t xfer_mode; /* I/O transfer mode */ + H5FD_mpio_collective_opt_t coll_opt_mode; /* whether we are doing collective or independent I/O */ + int size_i; +#if MPI_VERSION >= 3 + MPI_Count bytes_read = 0; /* Number of bytes read in */ + MPI_Count type_size; /* MPI datatype used for I/O's size */ + MPI_Count io_size; /* Actual number of bytes requested */ + MPI_Count n; +#else + int bytes_read = 0; /* Number of bytes read in */ + int type_size; /* MPI datatype used for I/O's size */ + int io_size; /* Actual number of bytes requested */ + int n; +#endif + hbool_t rank0_bcast = FALSE; /* If read-with-rank0-and-bcast flag was used */ +#ifdef H5FDmpio_DEBUG + hbool_t H5FD_mpio_debug_t_flag = (H5FD_mpio_debug_flags_s[(int)'t'] && H5FD_MPIO_TRACE_THIS_RANK(file)); + hbool_t H5FD_mpio_debug_r_flag = (H5FD_mpio_debug_flags_s[(int)'r'] && H5FD_MPIO_TRACE_THIS_RANK(file)); +#endif + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_t_flag) + HDfprintf(stderr, "%s: (%d) Entering\n", __func__, file->mpi_rank); +#endif + + /* Sanity checks */ + HDassert(file); + HDassert(H5FD_MPIO == file->pub.driver_id); + HDassert((types) || (count == 0)); + HDassert((addrs) || (count == 0)); + HDassert((sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* verify that the first elements of the sizes and types arrays are + * valid. + */ + HDassert((count == 0) || (sizes[0] != 0)); + HDassert((count == 0) || (types[0] != H5FD_MEM_NOLIST)); + + /* Get the transfer mode from the API context + * + * This flag is set to H5FD_MPIO_COLLECTIVE if the API call is + * collective, and to H5FD_MPIO_INDEPENDENT if it is not. + * + * While this doesn't mean that we are actually about to do a collective + * read, it does mean that all ranks are here, so we can use MPI_File_set_view(). + */ + if (H5CX_get_io_xfer_mode(&xfer_mode) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode") + + if (xfer_mode == H5FD_MPIO_COLLECTIVE) { + /* Build MPI types, etc. */ + if (H5FD__mpio_vector_build_types(count, types, addrs, sizes, (H5_flexible_const_ptr_t *)bufs, + &s_addrs, &s_sizes, (H5_flexible_const_ptr_t **)&s_bufs, + &vector_was_sorted, &mpi_off, + (H5_flexible_const_ptr_t *)&mpi_bufs_base, &size_i, &buf_type, + &buf_type_created, &file_type, &file_type_created, &unused) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't build MPI datatypes for I/O") + + /* free sorted addrs vector if it exists */ + if (!vector_was_sorted) + if (s_addrs) { + HDfree(s_addrs); + s_addrs = NULL; + } + + /* Portably initialize MPI status variable */ + HDmemset(&mpi_stat, 0, sizeof(mpi_stat)); + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_r_flag) + HDfprintf(stdout, "%s: mpi_off = %ld size_i = %d\n", __func__, (long)mpi_off, size_i); +#endif + + /* Setup the file view. */ + if (MPI_SUCCESS != (mpi_code = MPI_File_set_view(file->f, mpi_off, MPI_BYTE, file_type, + H5FD_mpi_native_g, file->info))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code) + + /* Reset mpi_off to 0 since the view now starts at the data offset */ + if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, &mpi_off) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0") + + /* Get the collective_opt property to check whether the application wants to do IO individually. + */ + if (H5CX_get_mpio_coll_opt(&coll_opt_mode) < 0) + + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't get MPI-I/O collective_op property") + + /* Read the data. */ +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_r_flag) + HDfprintf(stdout, "%s: using MPIO collective mode\n", __func__); +#endif + if (coll_opt_mode == H5FD_MPIO_COLLECTIVE_IO) { +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_r_flag) + HDfprintf(stdout, "%s: doing MPI collective IO\n", __func__); +#endif + /* Check whether we should read from rank 0 and broadcast to other ranks */ + if (H5CX_get_mpio_rank0_bcast()) { +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_r_flag) + HDfprintf(stdout, "%s: doing read-rank0-and-MPI_Bcast\n", __func__); +#endif + /* Indicate path we've taken */ + rank0_bcast = TRUE; + + /* Read on rank 0 Bcast to other ranks */ + if (file->mpi_rank == 0) + if (MPI_SUCCESS != (mpi_code = MPI_File_read_at(file->f, mpi_off, mpi_bufs_base, size_i, + buf_type, &mpi_stat))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at_all failed", mpi_code) + if (MPI_SUCCESS != (mpi_code = MPI_Bcast(mpi_bufs_base, size_i, buf_type, 0, file->comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code) + } /* end if */ + else if (MPI_SUCCESS != (mpi_code = MPI_File_read_at_all(file->f, mpi_off, mpi_bufs_base, size_i, + buf_type, &mpi_stat))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at_all failed", mpi_code) + } /* end if */ + else if (size_i > 0) { +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_r_flag) + HDfprintf(stdout, "%s: doing MPI independent IO\n", __func__); +#endif + + if (MPI_SUCCESS != + (mpi_code = MPI_File_read_at(file->f, mpi_off, mpi_bufs_base, size_i, buf_type, &mpi_stat))) + + HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code) + + } /* end else */ + + /* Reset the file view */ + if (MPI_SUCCESS != (mpi_code = MPI_File_set_view(file->f, (MPI_Offset)0, MPI_BYTE, MPI_BYTE, + H5FD_mpi_native_g, file->info))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code) + + /* Only retrieve bytes read if this rank _actually_ participated in I/O */ + if (!rank0_bcast || (rank0_bcast && file->mpi_rank == 0)) { + /* How many bytes were actually read? */ +#if MPI_VERSION >= 3 + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_read))) +#else + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read))) +#endif + HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) + } /* end if */ + + /* If the rank0-bcast feature was used, broadcast the # of bytes read to + * other ranks, which didn't perform any I/O. + */ + /* NOTE: This could be optimized further to be combined with the broadcast + * of the data. (QAK - 2019/1/2) + * Or have rank 0 clear the unread parts of the buffer prior to + * the bcast. (NAF - 2021/9/15) + */ + if (rank0_bcast) +#if MPI_VERSION >= 3 + if (MPI_SUCCESS != MPI_Bcast(&bytes_read, 1, MPI_COUNT, 0, file->comm)) +#else + if (MPI_SUCCESS != MPI_Bcast(&bytes_read, 1, MPI_INT, 0, file->comm)) +#endif + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", 0) + + /* Get the type's size */ +#if MPI_VERSION >= 3 + if (MPI_SUCCESS != (mpi_code = MPI_Type_size_x(buf_type, &type_size))) +#else + if (MPI_SUCCESS != (mpi_code = MPI_Type_size(buf_type, &type_size))) +#endif + HMPI_GOTO_ERROR(FAIL, "MPI_Type_size failed", mpi_code) + + /* Compute the actual number of bytes requested */ + io_size = type_size * size_i; + + /* Check for read failure */ + if (bytes_read < 0 || bytes_read > io_size) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "file read failed") + + /* Check for incomplete read */ + n = io_size - bytes_read; + if (n > 0) { + i = (int)count - 1; + + /* Iterate over sorted array in reverse, filling in zeroes to + * sections of the buffers that were not read to */ + do { + HDassert(i >= 0); + +#if MPI_VERSION >= 3 + io_size = MIN(n, (MPI_Count)s_sizes[i]); + bytes_read = (MPI_Count)s_sizes[i] - io_size; +#else + io_size = MIN(n, (int)s_sizes[i]); + bytes_read = (int)s_sizes[i] - io_size; +#endif + HDassert(bytes_read >= 0); + + HDmemset((char *)s_bufs[i] + bytes_read, 0, (size_t)io_size); + + n -= io_size; + i--; + } while (n > 0); + } + } + else if (count > 0) { + haddr_t max_addr = HADDR_MAX; + hbool_t fixed_size = FALSE; + size_t size; + + /* The read is part of an independent operation. As a result, + * we can't use MPI_File_set_view() (since it it a collective operation), + * and thus we can't use the above code to construct the MPI datatypes. + * In the future, we could write code to detect when a contiguous slab + * in the file selection spans multiple vector elements and construct a + * memory datatype to match this larger block in the file, but for now + * just read in each element of the vector in a separate + * MPI_File_read_at() call. + * + * We could also just detect the case when the entire file selection is + * contiguous, which would allow us to use + * H5FD__mpio_vector_build_types() to construct the memory datatype. + */ + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_r_flag) + HDfprintf(stdout, "%s: doing MPI independent IO\n", __func__); +#endif + + /* Loop over vector elements */ + for (i = 0; i < (int)count; i++) { + /* Convert address to mpi offset */ + if (H5FD_mpi_haddr_to_MPIOff(addrs[i], &mpi_off) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't convert from haddr to MPI off") + + /* Calculate I/O size */ + if (!fixed_size) { + if (sizes[i] == 0) { + fixed_size = TRUE; + size = sizes[i - 1]; + } + else { + size = sizes[i]; + } + } + size_i = (int)size; + + if (size != (size_t)size_i) { + /* If HERE, then we need to work around the integer size limit + * of 2GB. The input size_t size variable cannot fit into an integer, + * but we can get around that limitation by creating a different datatype + * and then setting the integer size (or element count) to 1 when using + * the derived_type. + */ + + if (H5_mpio_create_large_type(size, 0, MPI_BYTE, &buf_type) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") + + buf_type_created = TRUE; + size_i = 1; + } + + /* Check if we actually need to do I/O */ + if (addrs[i] < max_addr) { + /* Portably initialize MPI status variable */ + HDmemset(&mpi_stat, 0, sizeof(mpi_stat)); + + /* Issue read */ + if (MPI_SUCCESS != + (mpi_code = MPI_File_read_at(file->f, mpi_off, bufs[i], size_i, buf_type, &mpi_stat))) + + HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code) + + /* How many bytes were actually read? */ +#if MPI_VERSION >= 3 + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, MPI_BYTE, &bytes_read))) +#else + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read))) +#endif + HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) + + /* Compute the actual number of bytes requested */ +#if MPI_VERSION >= 3 + io_size = (MPI_Count)size; +#else + io_size = (int)size; +#endif + + /* Check for read failure */ + if (bytes_read < 0 || bytes_read > io_size) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "file read failed") + + /* + * If we didn't read the entire I/O, fill in zeroes beyond end of + * the physical MPI file and don't issue any more reads at higher + * addresses. + */ + if ((n = (io_size - bytes_read)) > 0) { + HDmemset((char *)bufs[i] + bytes_read, 0, (size_t)n); + max_addr = addrs[i] + (haddr_t)bytes_read; + } + } + else { + /* Read is past the max address, fill in zeroes */ + HDmemset((char *)bufs[i], 0, size); + } + } + } + +done: + if (buf_type_created) { + MPI_Type_free(&buf_type); + } + + if (file_type_created) { + MPI_Type_free(&file_type); + } + + /* free sorted vectors if they exist */ + if (!vector_was_sorted) { + if (s_addrs) { + HDfree(s_addrs); + s_addrs = NULL; + } + if (s_sizes) { + HDfree(s_sizes); + s_sizes = NULL; + } + if (s_bufs) { + HDfree(s_bufs); + s_bufs = NULL; + } + } + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_t_flag) + HDfprintf(stdout, "%s: Leaving, proc %d: ret_value = %d\n", __func__, file->mpi_rank, ret_value); +#endif + + FUNC_LEAVE_NOAPI(ret_value) + +} /* end H5FD__mpio_read_vector() */ + +/*------------------------------------------------------------------------- + * Function: H5FD__mpio_write_vector + * + * Purpose: The behaviour of this function dependes on the value of + * the io_xfer_mode obtained from the context. + * + * If it is H5FD_MPIO_COLLECTIVE, this is a collective + * operation, which allows us to use MPI_File_set_view, and + * then perform the entire vector write in a single MPI call. + * + * Do this (if count is positive), by constructing memory + * and file derived types from the supplied vector, using + * file type to set the file view, and then writing the + * the memory type to file. Note that this write is + * either independent or collective depending on the + * value of mpio_coll_opt -- again obtained from the context. + * + * If count is zero, participate in the collective write + * (if so configured) with an empty write. + * + * Finally, set the file view back to its default state. + * + * In contrast, if io_xfer_mode is H5FD_MPIO_INDEPENDENT, + * this call is independent, and thus we cannot use + * MPI_File_set_view(). + * + * In this case, simply walk the vector, and issue an + * independent write for each entry. + * + * Return: Success: SUCCEED. + * Failure: FAIL. + * + * Programmer: John Mainzer + * March 15, 2021 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5FD__mpio_write_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t count, H5FD_mem_t types[], + haddr_t addrs[], size_t sizes[], const void *bufs[]) +{ + H5FD_mpio_t * file = (H5FD_mpio_t *)_file; + hbool_t vector_was_sorted = TRUE; + haddr_t * s_addrs = NULL; + size_t * s_sizes = NULL; + const void ** s_bufs = NULL; + char unused = 0; /* Unused, except for non-NULL pointer value */ + const void * mpi_bufs_base = NULL; + MPI_Datatype buf_type = MPI_BYTE; /* MPI description of the selection in memory */ + hbool_t buf_type_created = FALSE; + MPI_Datatype file_type = MPI_BYTE; /* MPI description of the selection in file */ + hbool_t file_type_created = FALSE; + int i; + int mpi_code; /* MPI return code */ + MPI_Offset mpi_off = 0; + MPI_Status mpi_stat; /* Status from I/O operation */ + H5FD_mpio_xfer_t xfer_mode; /* I/O transfer mode */ + H5FD_mpio_collective_opt_t coll_opt_mode; /* whether we are doing collective or independent I/O */ + int size_i; +#ifdef H5FDmpio_DEBUG + hbool_t H5FD_mpio_debug_t_flag = (H5FD_mpio_debug_flags_s[(int)'t'] && H5FD_MPIO_TRACE_THIS_RANK(file)); + hbool_t H5FD_mpio_debug_w_flag = (H5FD_mpio_debug_flags_s[(int)'w'] && H5FD_MPIO_TRACE_THIS_RANK(file)); +#endif + haddr_t max_addr = 0; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_STATIC + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_t_flag) + HDfprintf(stderr, "%s: (%d) Entering\n", __func__, file->mpi_rank); +#endif + + /* Sanity checks */ + HDassert(file); + HDassert(H5FD_MPIO == file->pub.driver_id); + HDassert((types) || (count == 0)); + HDassert((addrs) || (count == 0)); + HDassert((sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* verify that the first elements of the sizes and types arrays are + * valid. + */ + HDassert((count == 0) || (sizes[0] != 0)); + HDassert((count == 0) || (types[0] != H5FD_MEM_NOLIST)); + + /* Verify that no data is written when between MPI_Barrier()s during file flush */ + + HDassert(!H5CX_get_mpi_file_flushing()); + + /* Get the transfer mode from the API context + * + * This flag is set to H5FD_MPIO_COLLECTIVE if the API call is + * collective, and to H5FD_MPIO_INDEPENDENT if it is not. + * + * While this doesn't mean that we are actually about to do a collective + * write, it does mean that all ranks are here, so we can use MPI_File_set_view(). + */ + if (H5CX_get_io_xfer_mode(&xfer_mode) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode") + + if (xfer_mode == H5FD_MPIO_COLLECTIVE) { + /* Build MPI types, etc. */ + if (H5FD__mpio_vector_build_types(count, types, addrs, sizes, (H5_flexible_const_ptr_t *)bufs, + &s_addrs, &s_sizes, (H5_flexible_const_ptr_t **)&s_bufs, + &vector_was_sorted, &mpi_off, + (H5_flexible_const_ptr_t *)&mpi_bufs_base, &size_i, &buf_type, + &buf_type_created, &file_type, &file_type_created, &unused) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't build MPI datatypes for I/O") + + /* Compute max addr writted to */ + if (count > 0) + max_addr = s_addrs[count - 1] + (haddr_t)(s_sizes[count - 1]); + + /* free sorted vectors if they exist */ + if (!vector_was_sorted) { + if (s_addrs) { + HDfree(s_addrs); + s_addrs = NULL; + } + if (s_sizes) { + HDfree(s_sizes); + s_sizes = NULL; + } + if (s_bufs) { + HDfree(s_bufs); + s_bufs = NULL; + } + } + + /* Portably initialize MPI status variable */ + HDmemset(&mpi_stat, 0, sizeof(MPI_Status)); + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_w_flag) + HDfprintf(stdout, "%s: mpi_off = %ld size_i = %d\n", __func__, (long)mpi_off, size_i); +#endif + + /* Setup the file view. */ + if (MPI_SUCCESS != (mpi_code = MPI_File_set_view(file->f, mpi_off, MPI_BYTE, file_type, + H5FD_mpi_native_g, file->info))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code) + + /* Reset mpi_off to 0 since the view now starts at the data offset */ + if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, &mpi_off) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0") + + /* Get the collective_opt property to check whether the application wants to do IO individually. + */ + if (H5CX_get_mpio_coll_opt(&coll_opt_mode) < 0) + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't get MPI-I/O collective_op property") + + /* Write the data. */ +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_w_flag) + HDfprintf(stdout, "%s: using MPIO collective mode\n", __func__); +#endif + + if (coll_opt_mode == H5FD_MPIO_COLLECTIVE_IO) { +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_w_flag) + HDfprintf(stdout, "%s: doing MPI collective IO\n", __func__); +#endif + + if (MPI_SUCCESS != (mpi_code = MPI_File_write_at_all(file->f, mpi_off, mpi_bufs_base, size_i, + buf_type, &mpi_stat))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_write_at_all failed", mpi_code) + } /* end if */ + else if (size_i > 0) { +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_w_flag) + HDfprintf(stdout, "%s: doing MPI independent IO\n", __func__); +#endif + + if (MPI_SUCCESS != + (mpi_code = MPI_File_write_at(file->f, mpi_off, mpi_bufs_base, size_i, buf_type, &mpi_stat))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_write_at failed", mpi_code) + } /* end else */ + + /* Reset the file view */ + if (MPI_SUCCESS != (mpi_code = MPI_File_set_view(file->f, (MPI_Offset)0, MPI_BYTE, MPI_BYTE, + H5FD_mpi_native_g, file->info))) + HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code) + } + else if (count > 0) { + hbool_t fixed_size = FALSE; + size_t size; + + /* The read is part of an independent operation. As a result, + * we can't use MPI_File_set_view() (since it it a collective operation), + * and thus we can't use the above code to construct the MPI datatypes. + * In the future, we could write code to detect when a contiguous slab + * in the file selection spans multiple vector elements and construct a + * memory datatype to match this larger block in the file, but for now + * just read in each element of the vector in a separate + * MPI_File_read_at() call. + * + * We could also just detect the case when the entire file selection is + * contiguous, which would allow us to use + * H5FD__mpio_vector_build_types() to construct the memory datatype. + */ + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_w_flag) + HDfprintf(stdout, "%s: doing MPI independent IO\n", __func__); +#endif + + /* Loop over vector elements */ + for (i = 0; i < (int)count; i++) { + /* Convert address to mpi offset */ + if (H5FD_mpi_haddr_to_MPIOff(addrs[i], &mpi_off) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't convert from haddr to MPI off") + + /* Calculate I/O size */ + if (!fixed_size) { + if (sizes[i] == 0) { + fixed_size = TRUE; + size = sizes[i - 1]; + } + else { + size = sizes[i]; + } + } + size_i = (int)size; + + if (size != (size_t)size_i) { + /* If HERE, then we need to work around the integer size limit + * of 2GB. The input size_t size variable cannot fit into an integer, + * but we can get around that limitation by creating a different datatype + * and then setting the integer size (or element count) to 1 when using + * the derived_type. + */ + + if (H5_mpio_create_large_type(size, 0, MPI_BYTE, &buf_type) < 0) + HGOTO_ERROR(H5E_INTERNAL, H5E_CANTGET, FAIL, "can't create MPI-I/O datatype") + + buf_type_created = TRUE; + size_i = 1; + } + + /* Perform write */ + if (MPI_SUCCESS != + (mpi_code = MPI_File_write_at(file->f, mpi_off, bufs[i], size_i, buf_type, &mpi_stat))) + + HMPI_GOTO_ERROR(FAIL, "MPI_File_write_at failed", mpi_code) + + /* Check if this is the highest address written to so far */ + if (addrs[i] + size > max_addr) + max_addr = addrs[i] + size; + } + } + + /* Each process will keep track of its perceived EOF value locally, and + * ultimately we will reduce this value to the maximum amongst all + * processes, but until then keep the actual eof at HADDR_UNDEF just in + * case something bad happens before that point. (rather have a value + * we know is wrong sitting around rather than one that could only + * potentially be wrong.) + */ + file->eof = HADDR_UNDEF; + + /* check to see if the local eof has changed been extended, and update if so */ + if (max_addr > file->local_eof) + file->local_eof = max_addr; + +done: + if (buf_type_created) + MPI_Type_free(&buf_type); + + if (file_type_created) + MPI_Type_free(&file_type); + + /* Cleanup on error */ + if (ret_value < 0 && !vector_was_sorted) { + if (s_addrs) { + HDfree(s_addrs); + s_addrs = NULL; + } + if (s_sizes) { + HDfree(s_sizes); + s_sizes = NULL; + } + if (s_bufs) { + HDfree(s_bufs); + s_bufs = NULL; + } + } + + /* Make sure we cleaned up */ + HDassert(vector_was_sorted || !s_addrs); + HDassert(vector_was_sorted || !s_sizes); + HDassert(vector_was_sorted || !s_bufs); + +#ifdef H5FDmpio_DEBUG + if (H5FD_mpio_debug_t_flag) + HDfprintf(stdout, "%s: Leaving, proc %d: ret_value = %d\n", __func__, file->mpi_rank, ret_value); +#endif + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5FD__mpio_write_vector() */ + +/*------------------------------------------------------------------------- * Function: H5FD__mpio_flush * * Purpose: Makes sure that all data is on disk. This is collective. @@ -1703,17 +2835,19 @@ H5FD__mpio_truncate(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, hbool_t H5_ATTR HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_code) /* Only processor p0 will get the filesize and broadcast it. */ - /* (Note that throwing an error here will cause non-rank 0 processes - * to hang in following Bcast. -QAK, 3/17/2018) - */ - if (0 == file->mpi_rank) + if (0 == file->mpi_rank) { + /* If MPI_File_get_size fails, broadcast file size as -1 to signal error */ if (MPI_SUCCESS != (mpi_code = MPI_File_get_size(file->f, &size))) - HMPI_GOTO_ERROR(FAIL, "MPI_File_get_size failed", mpi_code) + size = (MPI_Offset)-1; + } /* Broadcast file size */ if (MPI_SUCCESS != (mpi_code = MPI_Bcast(&size, (int)sizeof(MPI_Offset), MPI_BYTE, 0, file->comm))) HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code) + if (size < 0) + HMPI_GOTO_ERROR(FAIL, "MPI_File_get_size failed", mpi_code) + if (H5FD_mpi_haddr_to_MPIOff(file->eoa, &needed_eof) < 0) HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "cannot convert from haddr_t to MPI_Offset") @@ -1796,9 +2930,13 @@ H5FD__mpio_delete(const char *filename, hid_t fapl_id) HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_code) /* Delete the file */ - if (mpi_rank == 0) + if (mpi_rank == 0) { + /* If MPI_File_delete fails, push an error but + * still participate in the following MPI_Barrier + */ if (MPI_SUCCESS != (mpi_code = MPI_File_delete(filename, info))) - HMPI_GOTO_ERROR(FAIL, "MPI_File_delete failed", mpi_code) + HMPI_DONE_ERROR(FAIL, "MPI_File_delete failed", mpi_code) + } /* Set up a barrier (don't want processes to run ahead of the delete) */ if (MPI_SUCCESS != (mpi_code = MPI_Barrier(comm))) diff --git a/src/H5FDmpio.h b/src/H5FDmpio.h index 00dea1b..5ce98ca 100644 --- a/src/H5FDmpio.h +++ b/src/H5FDmpio.h @@ -223,7 +223,7 @@ H5_DLL herr_t H5Pset_dxpl_mpio_collective_opt(hid_t dxpl_id, H5FD_mpio_collectiv * * Use of this function is optional. * - * \todo Add missing version information + * \since 1.8.0 * */ H5_DLL herr_t H5Pset_dxpl_mpio_chunk_opt(hid_t dxpl_id, H5FD_mpio_chunk_opt_t opt_mode); @@ -247,7 +247,7 @@ H5_DLL herr_t H5Pset_dxpl_mpio_chunk_opt(hid_t dxpl_id, H5FD_mpio_chunk_opt_t op * otherwise, a separate I/O process will be invoked for each chunk * (multi-chunk I/O). * - * \todo Add missing version information + * \since 1.8.0 * */ H5_DLL herr_t H5Pset_dxpl_mpio_chunk_opt_num(hid_t dxpl_id, unsigned num_chunk_per_proc); @@ -272,7 +272,7 @@ H5_DLL herr_t H5Pset_dxpl_mpio_chunk_opt_num(hid_t dxpl_id, unsigned num_chunk_p * percent_proc_per_chunk, the library will do collective I/O for this * chunk; otherwise, independent I/O will be done for the chunk. * - * \todo Add missing version information + * \since 1.8.0 * */ H5_DLL herr_t H5Pset_dxpl_mpio_chunk_opt_ratio(hid_t dxpl_id, unsigned percent_num_proc_per_chunk); diff --git a/src/H5FDmulti.c b/src/H5FDmulti.c index 3dcfa37..20c538f 100644 --- a/src/H5FDmulti.c +++ b/src/H5FDmulti.c @@ -176,6 +176,7 @@ static herr_t H5FD_multi_ctl(H5FD_t *_file, uint64_t op_code, uint64_t flags, c /* The class struct */ static const H5FD_class_t H5FD_multi_g = { + H5FD_CLASS_VERSION, /* struct version */ H5_VFD_MULTI, /* value */ "multi", /* name */ HADDR_MAX, /* maxaddr */ @@ -204,6 +205,10 @@ static const H5FD_class_t H5FD_multi_g = { H5FD_multi_get_handle, /* get_handle */ H5FD_multi_read, /* read */ H5FD_multi_write, /* write */ + NULL, /*read_vector */ + NULL, /*write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ H5FD_multi_flush, /* flush */ H5FD_multi_truncate, /* truncate */ H5FD_multi_lock, /* lock */ @@ -517,7 +522,7 @@ H5FD_split_populate_config(const char *meta_ext, hid_t meta_plist_id, const char meta_name_g[sizeof(meta_name_g) - 1] = '\0'; } else - sprintf(meta_name_g, "%%s%s", meta_ext); + snprintf(meta_name_g, sizeof(meta_name_g), "%%s%s", meta_ext); } else { strncpy(meta_name_g, "%s.meta", sizeof(meta_name_g)); @@ -535,7 +540,7 @@ H5FD_split_populate_config(const char *meta_ext, hid_t meta_plist_id, const char raw_name_g[sizeof(raw_name_g) - 1] = '\0'; } else - sprintf(raw_name_g, "%%s%s", raw_ext); + snprintf(raw_name_g, sizeof(raw_name_g), "%%s%s", raw_ext); } else { strncpy(raw_name_g, "%s.raw", sizeof(raw_name_g)); @@ -634,7 +639,7 @@ H5FD_multi_populate_config(const H5FD_mem_t *memb_map, const hid_t *memb_fapl, c if (!memb_name) { assert(strlen(letters) == H5FD_MEM_NTYPES); for (mt = H5FD_MEM_DEFAULT; mt < H5FD_MEM_NTYPES; mt = (H5FD_mem_t)(mt + 1)) { - sprintf(_memb_name_g[mt], "%%s-%c.h5", letters[mt]); + snprintf(_memb_name_g[mt], 16, "%%s-%c.h5", letters[mt]); _memb_name_ptrs[mt] = _memb_name_g[mt]; } memb_name = _memb_name_ptrs; diff --git a/src/H5FDperform.c b/src/H5FDperform.c index 096fdd6..4a68c6e 100644 --- a/src/H5FDperform.c +++ b/src/H5FDperform.c @@ -29,13 +29,14 @@ * Function: H5FDperform_init * * Purpose: Ensure that the library is initialized and then call - * the provided VFD initializer. + * the provided VFD initializer * - * Return: Success: identifier for the VFD just initialized + * Return: Success: Identifier for the VFD just initialized * Failure: H5I_INVALID_HID *------------------------------------------------------------------------- */ -hid_t H5FDperform_init(hid_t (*init)(void)) +hid_t +H5FDperform_init(H5FD_init_t op) { hid_t ret_value = H5I_INVALID_HID; /* Return value */ @@ -43,16 +44,16 @@ hid_t H5FDperform_init(hid_t (*init)(void)) /*NO TRACE*/ /* It is possible that an application will evaluate an - * `H5FD_*` symbol (`H5FD_FAMILY`, `H5FD_MULTI`, `H5FD_SEC2`, et - * cetera) before the library has had an opportunity to initialize. - * Call H5_init_library() to make sure that the library has been - * initialized before `init` is run. + * `H5FD_*` symbol (`H5FD_FAMILY`, `H5FD_MULTI`, `H5FD_SEC2`, etc. + * before the library has had an opportunity to initialize. Call + * H5_init_library() to make sure that the library has been initialized + * before `init` is run. */ - if (H5_init_library() < 0) { + if (H5_init_library() < 0) HGOTO_ERROR(H5E_FUNC, H5E_CANTINIT, H5I_INVALID_HID, "library initialization failed") - } - ret_value = init(); + ret_value = op(); + done: FUNC_LEAVE_API_NOINIT(ret_value) } diff --git a/src/H5FDprivate.h b/src/H5FDprivate.h index 61b4c60..bcbc693 100644 --- a/src/H5FDprivate.h +++ b/src/H5FDprivate.h @@ -24,6 +24,7 @@ /* Private headers needed by this file */ #include "H5Pprivate.h" /* Property lists */ +#include "H5Sprivate.h" /* Dataspaces */ /* * The MPI drivers are needed because there are @@ -94,6 +95,9 @@ typedef enum H5FD_get_driver_kind_t { H5FD_GET_DRIVER_BY_VALUE /* Value field is set */ } H5FD_get_driver_kind_t; +/* Forward declarations for prototype arguments */ +struct H5S_t; + /*****************************/ /* Library Private Variables */ /*****************************/ @@ -140,6 +144,22 @@ H5_DLL herr_t H5FD_set_feature_flags(H5FD_t *file, unsigned long feature_flags) H5_DLL herr_t H5FD_get_fs_type_map(const H5FD_t *file, H5FD_mem_t *type_map); H5_DLL herr_t H5FD_read(H5FD_t *file, H5FD_mem_t type, haddr_t addr, size_t size, void *buf /*out*/); H5_DLL herr_t H5FD_write(H5FD_t *file, H5FD_mem_t type, haddr_t addr, size_t size, const void *buf); +H5_DLL herr_t H5FD_read_vector(H5FD_t *file, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], + size_t sizes[], void *bufs[] /* out */); +H5_DLL herr_t H5FD_write_vector(H5FD_t *file, uint32_t count, H5FD_mem_t types[], haddr_t addrs[], + size_t sizes[], const void *bufs[] /* out */); +H5_DLL herr_t H5FD_read_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, struct H5S_t **mem_spaces, + struct H5S_t **file_spaces, haddr_t offsets[], size_t element_sizes[], + void *bufs[] /* out */); +H5_DLL herr_t H5FD_write_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, struct H5S_t **mem_spaces, + struct H5S_t **file_spaces, haddr_t offsets[], size_t element_sizes[], + const void *bufs[]); +H5_DLL herr_t H5FD_read_selection_id(H5FD_t *file, H5FD_mem_t type, uint32_t count, hid_t mem_space_ids[], + hid_t file_space_ids[], haddr_t offsets[], size_t element_sizes[], + void *bufs[] /* out */); +H5_DLL herr_t H5FD_write_selection_id(H5FD_t *file, H5FD_mem_t type, uint32_t count, hid_t mem_space_ids[], + hid_t file_space_ids[], haddr_t offsets[], size_t element_sizes[], + const void *bufs[]); H5_DLL herr_t H5FD_flush(H5FD_t *file, hbool_t closing); H5_DLL herr_t H5FD_truncate(H5FD_t *file, hbool_t closing); H5_DLL herr_t H5FD_lock(H5FD_t *file, hbool_t rw); @@ -152,6 +172,10 @@ H5_DLL herr_t H5FD_set_base_addr(H5FD_t *file, haddr_t base_addr); H5_DLL haddr_t H5FD_get_base_addr(const H5FD_t *file); H5_DLL herr_t H5FD_set_paged_aggr(H5FD_t *file, hbool_t paged); +H5_DLL herr_t H5FD_sort_vector_io_req(hbool_t *vector_was_sorted, uint32_t count, H5FD_mem_t types[], + haddr_t addrs[], size_t sizes[], H5_flexible_const_ptr_t bufs[], + H5FD_mem_t **s_types_ptr, haddr_t **s_addrs_ptr, size_t **s_sizes_ptr, + H5_flexible_const_ptr_t **s_bufs_ptr); H5_DLL herr_t H5FD_init(void); /* Function prototypes for MPI based VFDs*/ diff --git a/src/H5FDros3.c b/src/H5FDros3.c index df06526..fcce76d 100644 --- a/src/H5FDros3.c +++ b/src/H5FDros3.c @@ -237,6 +237,7 @@ static herr_t H5FD__ros3_truncate(H5FD_t *_file, hid_t dxpl_id, hbool_t closing static herr_t H5FD__ros3_validate_config(const H5FD_ros3_fapl_t *fa); static const H5FD_class_t H5FD_ros3_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_ROS3_VALUE, /* value */ "ros3", /* name */ MAXADDR, /* maxaddr */ @@ -265,6 +266,10 @@ static const H5FD_class_t H5FD_ros3_g = { H5FD__ros3_get_handle, /* get_handle */ H5FD__ros3_read, /* read */ H5FD__ros3_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ NULL, /* flush */ H5FD__ros3_truncate, /* truncate */ NULL, /* lock */ @@ -304,8 +309,12 @@ H5FD_ros3_init(void) HDfprintf(stdout, "H5FD_ros3_init() called.\n"); #endif - if (H5I_VFL != H5I_get_type(H5FD_ROS3_g)) + if (H5I_VFL != H5I_get_type(H5FD_ROS3_g)) { H5FD_ROS3_g = H5FD_register(&H5FD_ros3_g, sizeof(H5FD_class_t), FALSE); + if (H5I_INVALID_HID == H5FD_ROS3_g) { + HGOTO_ERROR(H5E_ID, H5E_CANTREGISTER, H5I_INVALID_HID, "unable to register ros3"); + } + } #if ROS3_STATS /* pre-compute statsbin boundaries diff --git a/src/H5FDsec2.c b/src/H5FDsec2.c index 46f5fd4..cc417070 100644 --- a/src/H5FDsec2.c +++ b/src/H5FDsec2.c @@ -143,6 +143,7 @@ static herr_t H5FD__sec2_ctl(H5FD_t *_file, uint64_t op_code, uint64_t flags, c void **output); static const H5FD_class_t H5FD_sec2_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_SEC2_VALUE, /* value */ "sec2", /* name */ MAXADDR, /* maxaddr */ @@ -171,6 +172,10 @@ static const H5FD_class_t H5FD_sec2_g = { H5FD__sec2_get_handle, /* get_handle */ H5FD__sec2_read, /* read */ H5FD__sec2_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ NULL, /* flush */ H5FD__sec2_truncate, /* truncate */ H5FD__sec2_lock, /* lock */ diff --git a/src/H5FDspace.c b/src/H5FDspace.c index de52dc3..48b06ba 100644 --- a/src/H5FDspace.c +++ b/src/H5FDspace.c @@ -148,7 +148,7 @@ H5FD__alloc_real(H5FD_t *file, H5FD_mem_t type, hsize_t size, haddr_t *frag_addr FUNC_ENTER_PACKAGE #ifdef H5FD_ALLOC_DEBUG - HDfprintf(stderr, "%s: type = %u, size = %Hu\n", __func__, (unsigned)type, size); + HDfprintf(stderr, "%s: type = %u, size = %" PRIuHSIZE "\n", __func__, (unsigned)type, size); #endif /* H5FD_ALLOC_DEBUG */ /* check args */ @@ -211,7 +211,7 @@ H5FD__alloc_real(H5FD_t *file, H5FD_mem_t type, hsize_t size, haddr_t *frag_addr done: #ifdef H5FD_ALLOC_DEBUG - HDfprintf(stderr, "%s: ret_value = %a\n", __func__, ret_value); + HDfprintf(stderr, "%s: ret_value = %" PRIuHADDR "\n", __func__, ret_value); #endif /* H5FD_ALLOC_DEBUG */ FUNC_LEAVE_NOAPI(ret_value) } /* end H5FD__alloc_real() */ @@ -287,7 +287,8 @@ H5FD__free_real(H5FD_t *file, H5FD_mem_t type, haddr_t addr, hsize_t size) HDassert(size > 0); #ifdef H5FD_ALLOC_DEBUG - HDfprintf(stderr, "%s: type = %u, addr = %a, size = %Hu\n", __func__, (unsigned)type, addr, size); + HDfprintf(stderr, "%s: type = %u, addr = %" PRIuHADDR ", size = %" PRIuHSIZE "\n", __func__, + (unsigned)type, addr, size); #endif /* H5FD_ALLOC_DEBUG */ /* Sanity checking */ @@ -317,11 +318,11 @@ H5FD__free_real(H5FD_t *file, H5FD_mem_t type, haddr_t addr, hsize_t size) eoa = file->cls->get_eoa(file, type); #ifdef H5FD_ALLOC_DEBUG - HDfprintf(stderr, "%s: eoa = %a\n", __func__, eoa); + HDfprintf(stderr, "%s: eoa = %" PRIuHADDR "\n", __func__, eoa); #endif /* H5FD_ALLOC_DEBUG */ if (eoa == (addr + size)) { #ifdef H5FD_ALLOC_DEBUG - HDfprintf(stderr, "%s: Reducing file size to = %a\n", __func__, addr); + HDfprintf(stderr, "%s: Reducing file size to = %" PRIuHADDR "\n", __func__, addr); #endif /* H5FD_ALLOC_DEBUG */ if (file->cls->set_eoa(file, type, addr) < 0) HGOTO_ERROR(H5E_VFL, H5E_CANTSET, FAIL, "set end of space allocation request failed") @@ -330,8 +331,8 @@ H5FD__free_real(H5FD_t *file, H5FD_mem_t type, haddr_t addr, hsize_t size) else { /* leak memory */ #ifdef H5FD_ALLOC_DEBUG - HDfprintf(stderr, "%s: LEAKED MEMORY!!! type = %u, addr = %a, size = %Hu\n", __func__, (unsigned)type, - addr, size); + HDfprintf(stderr, "%s: LEAKED MEMORY!!! type = %u, addr = %" PRIuHADDR ", size = %" PRIuHSIZE "\n", + __func__, (unsigned)type, addr, size); #endif /* H5FD_ALLOC_DEBUG */ } /* end else */ diff --git a/src/H5FDsplitter.c b/src/H5FDsplitter.c index 31438cd..124c54f 100644 --- a/src/H5FDsplitter.c +++ b/src/H5FDsplitter.c @@ -138,6 +138,7 @@ static herr_t H5FD__splitter_ctl(H5FD_t *_file, uint64_t op_code, uint64_t flag void **output); static const H5FD_class_t H5FD_splitter_g = { + H5FD_CLASS_VERSION, /* struct version */ H5FD_SPLITTER_VALUE, /* value */ "splitter", /* name */ MAXADDR, /* maxaddr */ @@ -166,6 +167,10 @@ static const H5FD_class_t H5FD_splitter_g = { H5FD__splitter_get_handle, /* get_handle */ H5FD__splitter_read, /* read */ H5FD__splitter_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ H5FD__splitter_flush, /* flush */ H5FD__splitter_truncate, /* truncate */ H5FD__splitter_lock, /* lock */ @@ -527,20 +532,20 @@ H5FD__splitter_get_default_wo_path(char *new_path, size_t new_path_len, const ch HGOTO_ERROR(H5E_VFL, H5E_CANTSET, FAIL, "filename exceeds max length") /* Determine if filename contains a ".h5" extension. */ - if ((file_extension = strstr(base_filename, ".h5"))) { + if ((file_extension = HDstrstr(base_filename, ".h5"))) { /* Insert the suffix between the filename and ".h5" extension. */ HDstrcpy(new_path, base_filename); - file_extension = strstr(new_path, ".h5"); + file_extension = HDstrstr(new_path, ".h5"); HDsprintf(file_extension, "%s%s", suffix, ".h5"); } - else if ((file_extension = strrchr(base_filename, '.'))) { + else if ((file_extension = HDstrrchr(base_filename, '.'))) { char *new_extension_loc = NULL; /* If the filename doesn't contain a ".h5" extension, but contains * AN extension, just insert the suffix before that extension. */ HDstrcpy(new_path, base_filename); - new_extension_loc = strrchr(new_path, '.'); + new_extension_loc = HDstrrchr(new_path, '.'); HDsprintf(new_extension_loc, "%s%s", suffix, file_extension); } else { diff --git a/src/H5FDstdio.c b/src/H5FDstdio.c index 122379a..6624685 100644 --- a/src/H5FDstdio.c +++ b/src/H5FDstdio.c @@ -183,41 +183,46 @@ static herr_t H5FD_stdio_unlock(H5FD_t *_file); static herr_t H5FD_stdio_delete(const char *filename, hid_t fapl_id); static const H5FD_class_t H5FD_stdio_g = { - H5_VFD_STDIO, /* value */ - "stdio", /* name */ - MAXADDR, /* maxaddr */ - H5F_CLOSE_WEAK, /* fc_degree */ - H5FD_stdio_term, /* terminate */ - NULL, /* sb_size */ - NULL, /* sb_encode */ - NULL, /* sb_decode */ - 0, /* fapl_size */ - NULL, /* fapl_get */ - NULL, /* fapl_copy */ - NULL, /* fapl_free */ - 0, /* dxpl_size */ - NULL, /* dxpl_copy */ - NULL, /* dxpl_free */ - H5FD_stdio_open, /* open */ - H5FD_stdio_close, /* close */ - H5FD_stdio_cmp, /* cmp */ - H5FD_stdio_query, /* query */ - NULL, /* get_type_map */ - H5FD_stdio_alloc, /* alloc */ - NULL, /* free */ - H5FD_stdio_get_eoa, /* get_eoa */ - H5FD_stdio_set_eoa, /* set_eoa */ - H5FD_stdio_get_eof, /* get_eof */ - H5FD_stdio_get_handle, /* get_handle */ - H5FD_stdio_read, /* read */ - H5FD_stdio_write, /* write */ - H5FD_stdio_flush, /* flush */ - H5FD_stdio_truncate, /* truncate */ - H5FD_stdio_lock, /* lock */ - H5FD_stdio_unlock, /* unlock */ - H5FD_stdio_delete, /* del */ - NULL, /* ctl */ - H5FD_FLMAP_DICHOTOMY /* fl_map */ + H5FD_CLASS_VERSION, /* struct version */ + H5_VFD_STDIO, /* value */ + "stdio", /* name */ + MAXADDR, /* maxaddr */ + H5F_CLOSE_WEAK, /* fc_degree */ + H5FD_stdio_term, /* terminate */ + NULL, /* sb_size */ + NULL, /* sb_encode */ + NULL, /* sb_decode */ + 0, /* fapl_size */ + NULL, /* fapl_get */ + NULL, /* fapl_copy */ + NULL, /* fapl_free */ + 0, /* dxpl_size */ + NULL, /* dxpl_copy */ + NULL, /* dxpl_free */ + H5FD_stdio_open, /* open */ + H5FD_stdio_close, /* close */ + H5FD_stdio_cmp, /* cmp */ + H5FD_stdio_query, /* query */ + NULL, /* get_type_map */ + H5FD_stdio_alloc, /* alloc */ + NULL, /* free */ + H5FD_stdio_get_eoa, /* get_eoa */ + H5FD_stdio_set_eoa, /* set_eoa */ + H5FD_stdio_get_eof, /* get_eof */ + H5FD_stdio_get_handle, /* get_handle */ + H5FD_stdio_read, /* read */ + H5FD_stdio_write, /* write */ + NULL, /* read_vector */ + NULL, /* write_vector */ + NULL, /* read_selection */ + NULL, /* write_selection */ + H5FD_stdio_flush, /* flush */ + H5FD_stdio_truncate, /* truncate */ + H5FD_stdio_lock, /* lock */ + H5FD_stdio_unlock, /* unlock */ + H5FD_stdio_delete, /* del */ + NULL, /* ctl */ + H5FD_FLMAP_DICHOTOMY /* fl_map */ }; /*------------------------------------------------------------------------- diff --git a/src/H5FLprivate.h b/src/H5FLprivate.h index 6519551..8e9d0d4 100644 --- a/src/H5FLprivate.h +++ b/src/H5FLprivate.h @@ -50,6 +50,11 @@ */ /* #define H5FL_TRACK */ #ifdef H5FL_TRACK + +#ifndef H5_HAVE_CODESTACK +#error "Free list tracking requires code stack to be enabled" +#endif + /* Macro for inclusion in the free list allocation calls */ #define H5FL_TRACK_INFO , __FILE__, __func__, __LINE__ @@ -273,16 +278,17 @@ typedef struct H5FL_arr_head_t { #define H5FL_BARR_DEFINE_STATIC(b, t, m) static H5FL_ARR_DEFINE_COMMON(sizeof(b), t, m) /* Allocate an array of type 't' */ -#define H5FL_ARR_MALLOC(t, elem) H5FL_arr_malloc(&(H5FL_ARR_NAME(t)), elem) +#define H5FL_ARR_MALLOC(t, elem) H5FL_arr_malloc(&(H5FL_ARR_NAME(t)), elem H5FL_TRACK_INFO) /* Allocate an array of type 't' and clear it to all zeros */ -#define H5FL_ARR_CALLOC(t, elem) H5FL_arr_calloc(&(H5FL_ARR_NAME(t)), elem) +#define H5FL_ARR_CALLOC(t, elem) H5FL_arr_calloc(&(H5FL_ARR_NAME(t)), elem H5FL_TRACK_INFO) /* Free an array of type 't' */ #define H5FL_ARR_FREE(t, obj) (t *)H5FL_arr_free(&(H5FL_ARR_NAME(t)), obj) /* Re-allocate an array of type 't' */ -#define H5FL_ARR_REALLOC(t, obj, new_elem) H5FL_arr_realloc(&(H5FL_ARR_NAME(t)), obj, new_elem) +#define H5FL_ARR_REALLOC(t, obj, new_elem) \ + H5FL_arr_realloc(&(H5FL_ARR_NAME(t)), obj, new_elem H5FL_TRACK_INFO) #else /* H5_NO_ARR_FREE_LISTS */ /* Common macro for H5FL_ARR_DEFINE & H5FL_ARR_DEFINE_STATIC (and H5FL_BARR variants) */ @@ -405,10 +411,10 @@ H5_DLL void *H5FL_reg_calloc(H5FL_reg_head_t *head H5FL_TRACK_PARAMS); H5_DLL void *H5FL_reg_free(H5FL_reg_head_t *head, void *obj); /* Array free lists */ -H5_DLL void *H5FL_arr_malloc(H5FL_arr_head_t *head, size_t elem); -H5_DLL void *H5FL_arr_calloc(H5FL_arr_head_t *head, size_t elem); +H5_DLL void *H5FL_arr_malloc(H5FL_arr_head_t *head, size_t elem H5FL_TRACK_PARAMS); +H5_DLL void *H5FL_arr_calloc(H5FL_arr_head_t *head, size_t elem H5FL_TRACK_PARAMS); H5_DLL void *H5FL_arr_free(H5FL_arr_head_t *head, void *obj); -H5_DLL void *H5FL_arr_realloc(H5FL_arr_head_t *head, void *obj, size_t new_elem); +H5_DLL void *H5FL_arr_realloc(H5FL_arr_head_t *head, void *obj, size_t new_elem H5FL_TRACK_PARAMS); /* Sequence free lists */ H5_DLL void *H5FL_seq_malloc(H5FL_seq_head_t *head, size_t elem H5FL_TRACK_PARAMS); @@ -143,7 +143,8 @@ H5FS_create(H5F_t *f, haddr_t *fs_addr, const H5FS_create_t *fs_create, uint16_t /* Set the return value */ ret_value = fspace; #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: fspace = %p, fspace->addr = %a\n", __func__, fspace, fspace->addr); + HDfprintf(stderr, "%s: fspace = %p, fspace->addr = %" PRIuHADDR "\n", __func__, (void *)fspace, + fspace->addr); #endif /* H5FS_DEBUG */ done: @@ -152,7 +153,7 @@ done: HDONE_ERROR(H5E_FSPACE, H5E_CANTFREE, NULL, "unable to destroy free space header") #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: Leaving, ret_value = %d\n", __func__, ret_value); + HDfprintf(stderr, "%s: Leaving, ret_value = %p\n", __func__, (void *)ret_value); #endif /* H5FS_DEBUG */ FUNC_LEAVE_NOAPI(ret_value) } /* H5FS_create() */ @@ -180,8 +181,8 @@ H5FS_open(H5F_t *f, haddr_t fs_addr, uint16_t nclasses, const H5FS_section_class FUNC_ENTER_NOAPI(NULL) #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: Opening free space manager, fs_addr = %a, nclasses = %Zu\n", __func__, fs_addr, - nclasses); + HDfprintf(stderr, "%s: Opening free space manager, fs_addr = %" PRIuHADDR ", nclasses = %Zu\n", __func__, + fs_addr, nclasses); #endif /* H5FS_DEBUG */ /* Check arguments. */ @@ -201,10 +202,10 @@ H5FS_open(H5F_t *f, haddr_t fs_addr, uint16_t nclasses, const H5FS_section_class (fspace = (H5FS_t *)H5AC_protect(f, H5AC_FSPACE_HDR, fs_addr, &cache_udata, H5AC__READ_ONLY_FLAG))) HGOTO_ERROR(H5E_FSPACE, H5E_CANTPROTECT, NULL, "unable to load free space header") #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: fspace->sect_addr = %a\n", __func__, fspace->sect_addr); - HDfprintf(stderr, "%s: fspace->sect_size = %Hu\n", __func__, fspace->sect_size); - HDfprintf(stderr, "%s: fspace->alloc_sect_size = %Hu\n", __func__, fspace->alloc_sect_size); - HDfprintf(stderr, "%s: fspace->sinfo = %p\n", __func__, fspace->sinfo); + HDfprintf(stderr, "%s: fspace->sect_addr = %" PRIuHADDR "\n", __func__, fspace->sect_addr); + HDfprintf(stderr, "%s: fspace->sect_size = %" PRIuHSIZE "\n", __func__, fspace->sect_size); + HDfprintf(stderr, "%s: fspace->alloc_sect_size = %" PRIuHSIZE "\n", __func__, fspace->alloc_sect_size); + HDfprintf(stderr, "%s: fspace->sinfo = %p\n", __func__, (void *)fspace->sinfo); HDfprintf(stderr, "%s: fspace->rc = %u\n", __func__, fspace->rc); #endif /* H5FS_DEBUG */ @@ -248,7 +249,7 @@ H5FS_delete(H5F_t *f, haddr_t fs_addr) FUNC_ENTER_NOAPI(FAIL) #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: Deleting free space manager, fs_addr = %a\n", __func__, fs_addr); + HDfprintf(stderr, "%s: Deleting free space manager, fs_addr = %" PRIuHADDR "\n", __func__, fs_addr); #endif /* H5FS_DEBUG */ /* Check arguments. */ @@ -318,7 +319,7 @@ H5FS_delete(H5F_t *f, haddr_t fs_addr) /* Delete serialized section storage, if there are any */ #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: fspace->sect_addr = %a\n", __func__, fspace->sect_addr); + HDfprintf(stderr, "%s: fspace->sect_addr = %" PRIuHADDR "\n", __func__, fspace->sect_addr); #endif /* H5FS_DEBUG */ if (fspace->serial_sect_count > 0) { unsigned sinfo_status = 0; /* Free space section info's status in the metadata cache */ @@ -404,8 +405,8 @@ H5FS_close(H5F_t *f, H5FS_t *fspace) HDassert(f); HDassert(fspace); #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: Entering, fspace = %p, fspace->addr = %a, fspace->sinfo = %p\n", __func__, fspace, - fspace->addr, fspace->sinfo); + HDfprintf(stderr, "%s: Entering, fspace = %p, fspace->addr = %" PRIuHADDR ", fspace->sinfo = %p\n", + __func__, (void *)fspace, fspace->addr, (void *)fspace->sinfo); #endif /* H5FS_DEBUG */ /* Check if section info is valid */ @@ -413,11 +414,12 @@ H5FS_close(H5F_t *f, H5FS_t *fspace) if (fspace->sinfo) { #ifdef H5FS_DEBUG HDfprintf(stderr, - "%s: fspace->tot_sect_count = %Hu, fspace->serial_sect_count = %Hu, fspace->sect_addr = " - "%a, fspace->rc = %u\n", + "%s: fspace->tot_sect_count = %" PRIuHSIZE ", fspace->serial_sect_count = %" PRIuHSIZE + ", fspace->sect_addr = %" PRIuHADDR ", fspace->rc = %u\n", __func__, fspace->tot_sect_count, fspace->serial_sect_count, fspace->sect_addr, fspace->rc); - HDfprintf(stderr, "%s: fspace->alloc_sect_size = %Hu, fspace->sect_size = %Hu\n", __func__, - fspace->alloc_sect_size, fspace->sect_size); + HDfprintf(stderr, + "%s: fspace->alloc_sect_size = %" PRIuHSIZE ", fspace->sect_size = %" PRIuHSIZE "\n", + __func__, fspace->alloc_sect_size, fspace->sect_size); #endif /* H5FS_DEBUG */ /* If there are sections to serialize, update them */ /* (if the free space manager is persistent) */ @@ -708,7 +710,7 @@ H5FS__incr(H5FS_t *fspace) FUNC_ENTER_PACKAGE #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: Entering, fpace->addr = %a, fspace->rc = %u\n", __func__, fspace->addr, + HDfprintf(stderr, "%s: Entering, fpace->addr = %" PRIuHADDR ", fspace->rc = %u\n", __func__, fspace->addr, fspace->rc); #endif /* H5FS_DEBUG */ @@ -748,7 +750,7 @@ H5FS__decr(H5FS_t *fspace) FUNC_ENTER_PACKAGE #ifdef H5FS_DEBUG - HDfprintf(stderr, "%s: Entering, fpace->addr = %a, fspace->rc = %u\n", __func__, fspace->addr, + HDfprintf(stderr, "%s: Entering, fpace->addr = %" PRIuHADDR ", fspace->rc = %u\n", __func__, fspace->addr, fspace->rc); #endif /* H5FS_DEBUG */ diff --git a/src/H5FSsection.c b/src/H5FSsection.c index 6c5a850..55f8a94 100644 --- a/src/H5FSsection.c +++ b/src/H5FSsection.c @@ -123,7 +123,7 @@ H5FS__sinfo_new(H5F_t *f, H5FS_t *fspace) HDassert(f); HDassert(fspace); #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: fspace->addr = %a\n", __func__, fspace->addr); + HDfprintf(stderr, "%s: fspace->addr = %" PRIuHADDR "\n", __func__, fspace->addr); #endif /* H5FS_SINFO_DEBUG */ /* Allocate the free space header */ @@ -136,7 +136,7 @@ H5FS__sinfo_new(H5F_t *f, H5FS_t *fspace) sinfo->sect_off_size = (fspace->max_sect_addr + 7) / 8; sinfo->sect_len_size = H5VM_limit_enc_size((uint64_t)fspace->max_sect_size); #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: fspace->max_sect_size = %Hu\n", __func__, fspace->max_sect_size); + HDfprintf(stderr, "%s: fspace->max_sect_size = %" PRIuHSIZE "\n", __func__, fspace->max_sect_size); HDfprintf(stderr, "%s: fspace->max_sect_addr = %u\n", __func__, fspace->max_sect_addr); HDfprintf(stderr, "%s: sinfo->nbins = %u\n", __func__, sinfo->nbins); HDfprintf(stderr, "%s: sinfo->sect_off_size = %u, sinfo->sect_len_size = %u\n", __func__, @@ -200,10 +200,12 @@ H5FS__sinfo_lock(H5F_t *f, H5FS_t *fspace, unsigned accmode) FUNC_ENTER_STATIC #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: Called, fspace->addr = %a, fspace->sinfo = %p, fspace->sect_addr = %a\n", __func__, - fspace->addr, fspace->sinfo, fspace->sect_addr); - HDfprintf(stderr, "%s: fspace->alloc_sect_size = %Hu, fspace->sect_size = %Hu\n", __func__, - fspace->alloc_sect_size, fspace->sect_size); + HDfprintf(stderr, + "%s: Called, fspace->addr = %" PRIuHADDR ", fspace->sinfo = %p, fspace->sect_addr = %" PRIuHADDR + "\n", + __func__, fspace->addr, (void *)fspace->sinfo, fspace->sect_addr); + HDfprintf(stderr, "%s: fspace->alloc_sect_size = %" PRIuHSIZE ", fspace->sect_size = %" PRIuHSIZE "\n", + __func__, fspace->alloc_sect_size, fspace->sect_size); #endif /* H5FS_SINFO_DEBUG */ /* Check arguments. */ @@ -251,8 +253,8 @@ H5FS__sinfo_lock(H5F_t *f, H5FS_t *fspace, unsigned accmode) HDassert(H5F_addr_defined(fspace->addr)); #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: Reading in existing sections, fspace->sect_addr = %a\n", __func__, - fspace->sect_addr); + HDfprintf(stderr, "%s: Reading in existing sections, fspace->sect_addr = %" PRIuHADDR "\n", + __func__, fspace->sect_addr); #endif /* H5FS_SINFO_DEBUG */ /* Protect the free space sections */ cache_udata.f = f; @@ -289,10 +291,12 @@ H5FS__sinfo_lock(H5F_t *f, H5FS_t *fspace, unsigned accmode) done: #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: Leaving, fspace->addr = %a, fspace->sinfo = %p, fspace->sect_addr = %a\n", - __func__, fspace->addr, fspace->sinfo, fspace->sect_addr); - HDfprintf(stderr, "%s: fspace->alloc_sect_size = %Hu, fspace->sect_size = %Hu\n", __func__, - fspace->alloc_sect_size, fspace->sect_size); + HDfprintf(stderr, + "%s: Leaving, fspace->addr = %" PRIuHADDR + ", fspace->sinfo = %p, fspace->sect_addr = %" PRIuHADDR "\n", + __func__, fspace->addr, (void *)fspace->sinfo, fspace->sect_addr); + HDfprintf(stderr, "%s: fspace->alloc_sect_size = %" PRIuHSIZE ", fspace->sect_size = %" PRIuHSIZE "\n", + __func__, fspace->alloc_sect_size, fspace->sect_size); #endif /* H5FS_SINFO_DEBUG */ FUNC_LEAVE_NOAPI(ret_value) } /* H5FS__sinfo_lock() */ @@ -331,14 +335,16 @@ H5FS__sinfo_unlock(H5F_t *f, H5FS_t *fspace, hbool_t modified) FUNC_ENTER_STATIC #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: Called, modified = %t, fspace->addr = %a, fspace->sect_addr = %a\n", __func__, - modified, fspace->addr, fspace->sect_addr); + HDfprintf(stderr, + "%s: Called, modified = %d, fspace->addr = %" PRIuHADDR ", fspace->sect_addr = %" PRIuHADDR + "\n", + __func__, modified, fspace->addr, fspace->sect_addr); HDfprintf( stderr, - "%s: fspace->sinfo_lock_count = %u, fspace->sinfo_modified = %t, fspace->sinfo_protected = %t\n", + "%s: fspace->sinfo_lock_count = %u, fspace->sinfo_modified = %d, fspace->sinfo_protected = %d\n", __func__, fspace->sinfo_lock_count, fspace->sinfo_modified, fspace->sinfo_protected); - HDfprintf(stderr, "%s: fspace->alloc_sect_size = %Hu, fspace->sect_size = %Hu\n", __func__, - fspace->alloc_sect_size, fspace->sect_size); + HDfprintf(stderr, "%s: fspace->alloc_sect_size = %" PRIuHSIZE ", fspace->sect_size = %" PRIuHSIZE "\n", + __func__, fspace->alloc_sect_size, fspace->sect_size); #endif /* H5FS_SINFO_DEBUG */ /* Check arguments. */ @@ -490,7 +496,8 @@ H5FS__sinfo_unlock(H5F_t *f, H5FS_t *fspace, hbool_t modified) #ifdef H5FS_SINFO_DEBUG HDfprintf(stderr, - "%s: Freeing section info on disk, old_sect_addr = %a, old_alloc_sect_size = %Hu\n", + "%s: Freeing section info on disk, old_sect_addr = %" PRIuHADDR + ", old_alloc_sect_size = %" PRIuHSIZE "\n", __func__, old_sect_addr, old_alloc_sect_size); #endif /* H5FS_SINFO_DEBUG */ /* Release space for section info in file */ @@ -1343,7 +1350,8 @@ H5FS_sect_add(H5F_t *f, H5FS_t *fspace, H5FS_section_info_t *sect, unsigned flag FUNC_ENTER_NOAPI(FAIL) #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: *sect = {%a, %Hu, %u, %s}\n", __func__, sect->addr, sect->size, sect->type, + HDfprintf(stderr, "%s: *sect = {%" PRIuHADDR ", %" PRIuHSIZE ", %u, %s}\n", __func__, sect->addr, + sect->size, sect->type, (sect->state == H5FS_SECT_LIVE ? "H5FS_SECT_LIVE" : "H5FS_SECT_SERIALIZED")); #endif /* H5FS_SINFO_DEBUG */ @@ -1384,7 +1392,7 @@ H5FS_sect_add(H5F_t *f, H5FS_t *fspace, H5FS_section_info_t *sect, unsigned flag HGOTO_ERROR(H5E_FSPACE, H5E_CANTINSERT, FAIL, "can't insert free space section into skip list") #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: fspace->tot_space = %Hu\n", __func__, fspace->tot_space); + HDfprintf(stderr, "%s: fspace->tot_space = %" PRIuHSIZE "\n", __func__, fspace->tot_space); #endif /* H5FS_SINFO_DEBUG */ /* Mark free space sections as changed */ /* (if adding sections while deserializing sections, don't set the flag) */ @@ -1429,8 +1437,8 @@ H5FS_sect_try_extend(H5F_t *f, H5FS_t *fspace, haddr_t addr, hsize_t size, hsize FUNC_ENTER_NOAPI(FAIL) #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: addr = %a, size = %Hu, extra_requested = %hu\n", __func__, addr, size, - extra_requested); + HDfprintf(stderr, "%s: addr = %" PRIuHADDR ", size = %" PRIuHSIZE ", extra_requested = %" PRIuHSIZE "\n", + __func__, addr, size, extra_requested); #endif /* H5FS_SINFO_DEBUG */ /* Check arguments. */ @@ -1442,9 +1450,10 @@ H5FS_sect_try_extend(H5F_t *f, H5FS_t *fspace, haddr_t addr, hsize_t size, hsize /* Check for any sections on free space list */ #ifdef H5FS_SINFO_DEBUG - HDfprintf(stderr, "%s: fspace->tot_sect_count = %Hu\n", __func__, fspace->tot_sect_count); - HDfprintf(stderr, "%s: fspace->serial_sect_count = %Hu\n", __func__, fspace->serial_sect_count); - HDfprintf(stderr, "%s: fspace->ghost_sect_count = %Hu\n", __func__, fspace->ghost_sect_count); + HDfprintf(stderr, "%s: fspace->tot_sect_count = %" PRIuHSIZE "\n", __func__, fspace->tot_sect_count); + HDfprintf(stderr, "%s: fspace->serial_sect_count = %" PRIuHSIZE "\n", __func__, + fspace->serial_sect_count); + HDfprintf(stderr, "%s: fspace->ghost_sect_count = %" PRIuHSIZE "\n", __func__, fspace->ghost_sect_count); #endif /* H5FS_SINFO_DEBUG */ if (fspace->tot_sect_count > 0) { H5FS_section_info_t *sect; /* Temporary free space section */ @@ -1592,7 +1601,7 @@ H5FS_sect_try_merge(H5F_t *f, H5FS_t *fspace, H5FS_section_info_t *sect, unsigne } /* end if */ else { /* Check if section is merged */ - if (sect->size > saved_fs_size) { + if (sect->size != saved_fs_size) { if (H5FS__sect_link(fspace, sect, flags) < 0) HGOTO_ERROR(H5E_FSPACE, H5E_CANTINSERT, FAIL, "can't insert free space section into skip list") diff --git a/src/H5Fio.c b/src/H5Fio.c index 5a9d2c1..53fec97 100644 --- a/src/H5Fio.c +++ b/src/H5Fio.c @@ -233,12 +233,101 @@ H5F_block_write(H5F_t *f, H5FD_mem_t type, haddr_t addr, size_t size, const void /* Pass through page buffer layer */ if (H5PB_write(f->shared, map_type, addr, size, buf) < 0) HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "write through page buffer failed") - done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5F_block_write() */ /*------------------------------------------------------------------------- + * Function: H5F_shared_select_read + * + * Purpose: Reads some data from a file/server/etc into a buffer. + * The location of the data is defined by the mem_spaces and + * file_spaces dataspace arrays, along with the offsets + * array. The addresses is relative to the base address for + * the file. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Neil Fortner + * May 3 2021 + * + *------------------------------------------------------------------------- + */ +herr_t +H5F_shared_select_read(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, H5S_t **mem_spaces, + H5S_t **file_spaces, haddr_t offsets[], size_t element_sizes[], void *bufs[] /* out */) +{ + H5FD_mem_t map_type; /* Mapped memory type */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(f_sh); + HDassert((mem_spaces) || (count == 0)); + HDassert((file_spaces) || (count == 0)); + HDassert((offsets) || (count == 0)); + HDassert((element_sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* Treat global heap as raw data */ + map_type = (type == H5FD_MEM_GHEAP) ? H5FD_MEM_DRAW : type; + + /* Pass down to file driver layer (bypass page buffer for now) */ + if (H5FD_read_selection(f_sh->lf, map_type, count, mem_spaces, file_spaces, offsets, element_sizes, + bufs) < 0) + HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "selection read through file driver failed") + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5F_shared_select_read() */ + +/*------------------------------------------------------------------------- + * Function: H5F_shared_select_write + * + * Purpose: Writes some data from a buffer to a file/server/etc. + * The location of the data is defined by the mem_spaces and + * file_spaces dataspace arrays, along with the offsets + * array. The addresses is relative to the base address for + * the file. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Neil Fortner + * May 4 2021 + * + *------------------------------------------------------------------------- + */ +herr_t +H5F_shared_select_write(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, H5S_t **mem_spaces, + H5S_t **file_spaces, haddr_t offsets[], size_t element_sizes[], const void *bufs[]) +{ + H5FD_mem_t map_type; /* Mapped memory type */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(f_sh); + HDassert((mem_spaces) || (count == 0)); + HDassert((file_spaces) || (count == 0)); + HDassert((offsets) || (count == 0)); + HDassert((element_sizes) || (count == 0)); + HDassert((bufs) || (count == 0)); + + /* Treat global heap as raw data */ + map_type = (type == H5FD_MEM_GHEAP) ? H5FD_MEM_DRAW : type; + + /* Pass down to file driver layer (bypass page buffer for now) */ + if (H5FD_write_selection(f_sh->lf, map_type, count, mem_spaces, file_spaces, offsets, element_sizes, + bufs) < 0) + HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "selection write through file driver failed") + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5F_shared_select_write() */ + +/*------------------------------------------------------------------------- * Function: H5F_flush_tagged_metadata * * Purpose: Flushes metadata with specified tag in the metadata cache diff --git a/src/H5Fmpi.c b/src/H5Fmpi.c index 53d2d78..02d8d52 100644 --- a/src/H5Fmpi.c +++ b/src/H5Fmpi.c @@ -31,11 +31,12 @@ /***********/ /* Headers */ /***********/ -#include "H5private.h" /* Generic Functions */ -#include "H5Eprivate.h" /* Error handling */ -#include "H5Fpkg.h" /* File access */ -#include "H5FDprivate.h" /* File drivers */ -#include "H5Iprivate.h" /* IDs */ +#include "H5private.h" /* Generic Functions */ +#include "H5CXprivate.h" /* API Contexts */ +#include "H5Eprivate.h" /* Error handling */ +#include "H5Fpkg.h" /* File access */ +#include "H5FDprivate.h" /* File drivers */ +#include "H5Iprivate.h" /* IDs */ #include "H5VLnative_private.h" /* Native VOL connector */ @@ -402,4 +403,189 @@ H5F_mpi_retrieve_comm(hid_t loc_id, hid_t acspl_id, MPI_Comm *mpi_comm) done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5F_mpi_retrieve_comm */ + +/*------------------------------------------------------------------------- + * Function: H5F_get_coll_metadata_reads + * + * Purpose: Determines whether collective metadata reads should be + * performed. This routine is meant to be the single source of + * truth for the collective metadata reads status, as it + * coordinates between the file-global flag and the flag set + * for the current operation in the current API context. + * + * Return: TRUE/FALSE (can't fail) + * + *------------------------------------------------------------------------- + */ +hbool_t +H5F_get_coll_metadata_reads(const H5F_t *file) +{ + H5P_coll_md_read_flag_t file_flag = H5P_USER_FALSE; + hbool_t ret_value = FALSE; + + FUNC_ENTER_NOAPI_NOERR + + HDassert(file && file->shared); + + /* Retrieve the file-global flag */ + file_flag = H5F_COLL_MD_READ(file); + + /* If file flag is set to H5P_FORCE_FALSE, exit early + * with FALSE, since collective metadata reads have + * been explicitly disabled somewhere in the library. + */ + if (H5P_FORCE_FALSE == file_flag) + ret_value = FALSE; + else { + /* If file flag is set to H5P_USER_TRUE, ignore + * any settings in the API context. A file-global + * setting of H5P_USER_TRUE for collective metadata + * reads should ignore any settings on an Access + * Property List for an individual operation. + */ + if (H5P_USER_TRUE == file_flag) + ret_value = TRUE; + else { + /* Get the collective metadata reads flag from + * the current API context. + */ + ret_value = H5CX_get_coll_metadata_read(); + } + } + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5F_get_coll_metadata_reads() */ + +/*------------------------------------------------------------------------- + * Function: H5F_set_coll_metadata_reads + * + * Purpose: Used to temporarily modify the collective metadata reads + * status. This is useful for cases where either: + * + * * Collective metadata reads are enabled, but need to be + * disabled for an operation about to occur that may trigger + * an independent metadata read (such as only rank 0 doing + * something) + * + * * Metadata reads are currently independent, but it is + * guaranteed that the application has maintained + * collectivity at the interface level (e.g., an operation + * that modifies metadata is being performed). In this case, + * it should be safe to enable collective metadata reads, + * barring any internal library issues that may occur + * + * After completion, the `file_flag` parameter will be set to + * the previous value of the file-global collective metadata + * reads flag. The `context_flag` parameter will be set to the + * previous value of the API context's collective metadata + * reads flag. Another call to this routine should be made to + * restore these values (see below warning). + * + * !! WARNING !! + * It is dangerous to modify the collective metadata reads + * status, as this can cause crashes, hangs and corruption in + * the HDF5 file when improperly done. Therefore, the + * `file_flag` and `context_flag` parameters are both + * mandatory, and it is assumed that the caller will guarantee + * these settings are restored with another call to this + * routine once the bracketed operation is complete. + * !! WARNING !! + * + * Return: Nothing + * + *------------------------------------------------------------------------- + */ +void +H5F_set_coll_metadata_reads(H5F_t *file, H5P_coll_md_read_flag_t *file_flag, hbool_t *context_flag) +{ + H5P_coll_md_read_flag_t prev_file_flag = H5P_USER_FALSE; + hbool_t prev_context_flag = FALSE; + + FUNC_ENTER_NOAPI_NOERR + + HDassert(file && file->shared); + HDassert(file_flag); + HDassert(context_flag); + + /* Save old state */ + prev_file_flag = H5F_COLL_MD_READ(file); + prev_context_flag = H5CX_get_coll_metadata_read(); + + /* Set new desired state */ + if (prev_file_flag != *file_flag) { + file->shared->coll_md_read = *file_flag; + *file_flag = prev_file_flag; + } + if (prev_context_flag != *context_flag) { + H5CX_set_coll_metadata_read(*context_flag); + *context_flag = prev_context_flag; + } + + FUNC_LEAVE_NOAPI_VOID +} /* end H5F_set_coll_metadata_reads() */ + +/*------------------------------------------------------------------------- + * Function: H5F_mpi_get_file_block_type + * + * Purpose: Creates an MPI derived datatype for communicating an + * H5F_block_t structure. If `commit` is specified as TRUE, + * the resulting datatype will be committed and ready for + * use in communication. Otherwise, the type is only suitable + * for building other derived types. + * + * If TRUE is returned through `new_type_derived`, this lets + * the caller know that the datatype has been derived and + * should be freed with MPI_Type_free once it is no longer + * needed. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +herr_t +H5F_mpi_get_file_block_type(hbool_t commit, MPI_Datatype *new_type, hbool_t *new_type_derived) +{ + MPI_Datatype types[2]; + MPI_Aint displacements[2]; + int block_lengths[2]; + int field_count; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_NOAPI(FAIL) + + HDassert(new_type); + HDassert(new_type_derived); + + *new_type_derived = FALSE; + + field_count = 2; + HDassert(field_count == sizeof(types) / sizeof(MPI_Datatype)); + + block_lengths[0] = 1; + block_lengths[1] = 1; + displacements[0] = offsetof(H5F_block_t, offset); + displacements[1] = offsetof(H5F_block_t, length); + types[0] = HADDR_AS_MPI_TYPE; + types[1] = HSIZE_AS_MPI_TYPE; + if (MPI_SUCCESS != + (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, new_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code) + *new_type_derived = TRUE; + + if (commit && MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_type))) + HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code) + +done: + if (ret_value < 0) { + if (*new_type_derived) { + if (MPI_SUCCESS != (mpi_code = MPI_Type_free(new_type))) + HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code) + *new_type_derived = FALSE; + } + } + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5F_mpi_get_file_block_type() */ + #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5Fprivate.h b/src/H5Fprivate.h index a5ccbab..629aee1 100644 --- a/src/H5Fprivate.h +++ b/src/H5Fprivate.h @@ -758,6 +758,7 @@ struct H5O_loc_t; struct H5HG_heap_t; struct H5VL_class_t; struct H5P_genplist_t; +struct H5S_t; /* Forward declarations for anonymous H5F objects */ @@ -923,6 +924,14 @@ H5_DLL herr_t H5F_shared_block_write(H5F_shared_t *f_sh, H5FD_mem_t type, haddr_ const void *buf); H5_DLL herr_t H5F_block_write(H5F_t *f, H5FD_mem_t type, haddr_t addr, size_t size, const void *buf); +/* Functions that operate on selections of elements in the file */ +H5_DLL herr_t H5F_shared_select_read(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, + struct H5S_t **mem_spaces, struct H5S_t **file_spaces, haddr_t offsets[], + size_t element_sizes[], void *bufs[] /* out */); +H5_DLL herr_t H5F_shared_select_write(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, + struct H5S_t **mem_spaces, struct H5S_t **file_spaces, + haddr_t offsets[], size_t element_sizes[], const void *bufs[]); + /* Functions that flush or evict */ H5_DLL herr_t H5F_flush_tagged_metadata(H5F_t *f, haddr_t tag); H5_DLL herr_t H5F_evict_tagged_metadata(H5F_t *f, haddr_t tag); @@ -962,6 +971,9 @@ H5_DLL MPI_Comm H5F_mpi_get_comm(const H5F_t *f); H5_DLL int H5F_shared_mpi_get_size(const H5F_shared_t *f_sh); H5_DLL int H5F_mpi_get_size(const H5F_t *f); H5_DLL herr_t H5F_mpi_retrieve_comm(hid_t loc_id, hid_t acspl_id, MPI_Comm *mpi_comm); +H5_DLL herr_t H5F_mpi_get_file_block_type(hbool_t commit, MPI_Datatype *new_type, hbool_t *new_type_derived); +H5_DLL hbool_t H5F_get_coll_metadata_reads(const H5F_t *f); +H5_DLL void H5F_set_coll_metadata_reads(H5F_t *f, H5P_coll_md_read_flag_t *file_flag, hbool_t *context_flag); #endif /* H5_HAVE_PARALLEL */ /* External file cache routines */ diff --git a/src/H5Fpublic.h b/src/H5Fpublic.h index c3230e1..671eec3 100644 --- a/src/H5Fpublic.h +++ b/src/H5Fpublic.h @@ -1606,7 +1606,7 @@ H5_DLL herr_t H5Fget_page_buffering_stats(hid_t file_id, unsigned accesses[2], u * \brief Obtains information about a cache image if it exists * * \file_id - * \param[out] image_addr Offset of the cache image if it exists, or \c HADDR_UNDEF if it does not + * \param[out] image_addr Offset of the cache image if it exists, or #HADDR_UNDEF if it does not * \param[out] image_size Length of the cache image if it exists, or 0 if it does not * \returns \herr_t * @@ -1878,6 +1878,7 @@ H5_DLL herr_t H5Fget_info1(hid_t obj_id, H5F_info1_t *file_info); * * \deprecated When? * + * \todo In which version was this function introduced? * \todo In which version was this function deprecated? * */ @@ -1896,6 +1897,7 @@ H5_DLL herr_t H5Fset_latest_format(hid_t file_id, hbool_t latest_format); * \details H5Fis_hdf5() determines whether a file is in the HDF5 format. * * \todo In which version was this function deprecated? + * \todo In which version was this function introduced? * */ H5_DLL htri_t H5Fis_hdf5(const char *file_name); diff --git a/src/H5Gprivate.h b/src/H5Gprivate.h index 4cf4623..d1725f6 100644 --- a/src/H5Gprivate.h +++ b/src/H5Gprivate.h @@ -130,7 +130,7 @@ typedef enum { typedef int H5G_own_loc_t; /* Structure to store information about the name an object was opened with */ -typedef struct { +typedef struct H5G_name_t { H5RS_str_t *full_path_r; /* Path to object, as seen from root of current file mounting hierarchy */ H5RS_str_t *user_path_r; /* Path to object, as opened by user */ unsigned obj_hidden; /* Whether the object is visible in group hier. */ diff --git a/src/H5HFcache.c b/src/H5HFcache.c index f409479..22ad09b 100644 --- a/src/H5HFcache.c +++ b/src/H5HFcache.c @@ -895,7 +895,7 @@ H5HF__cache_iblock_get_initial_load_size(void *_udata, size_t *image_len) * Function: H5HF__cache_iblock_verify_chksum * * Purpose: Verify the computed checksum of the data structure is the - * same as the stored chksum. + * same as the stored checksum. * * Return: Success: TRUE/FALSE * Failure: Negative diff --git a/src/H5HP.c b/src/H5HP.c deleted file mode 100644 index d164223..0000000 --- a/src/H5HP.c +++ /dev/null @@ -1,904 +0,0 @@ -/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * - * Copyright by The HDF Group. * - * Copyright by the Board of Trustees of the University of Illinois. * - * All rights reserved. * - * * - * This file is part of HDF5. The full HDF5 copyright notice, including * - * terms governing use, modification, and redistribution, is contained in * - * the COPYING file, which can be found at the root of the source code * - * distribution tree, or in https://www.hdfgroup.org/licenses. * - * If you do not have access to either file, you may request a copy from * - * help@hdfgroup.org. * - * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ - -/* - * Purpose: Provides a heap abstract data type. - * - * (See chapter 11 - "Priority Queues" of _Algorithms_, by - * Sedgewick for additional information) - * - */ - -/* Private headers needed */ -#include "H5private.h" /* Generic Functions */ -#include "H5Eprivate.h" /* Error handling */ -#include "H5HPprivate.h" /* Heap routines */ -#include "H5FLprivate.h" /* Memory management functions */ - -/* Local Macros */ -#define H5HP_START_SIZE 16 /* Initial number of entries for heaps */ - -/* Private typedefs & structs */ - -/* Data structure for entries in the internal heap array */ -typedef struct { - int val; /* Value to be used for heap condition */ - H5HP_info_t *obj; /* Pointer to object stored in heap */ -} H5HP_ent_t; - -/* Main heap data structure */ -struct H5HP_t { - H5HP_type_t type; /* Type of heap (minimum or maximum value at "top") */ - size_t nobjs; /* Number of active objects in heap array */ - size_t nalloc; /* Number of allocated locations in heap array */ - H5HP_ent_t *heap; /* Pointer to array containing heap entries */ -}; - -/* Static functions */ -static herr_t H5HP__swim_max(H5HP_t *heap, size_t loc); -static herr_t H5HP__swim_min(H5HP_t *heap, size_t loc); -static herr_t H5HP__sink_max(H5HP_t *heap, size_t loc); -static herr_t H5HP__sink_min(H5HP_t *heap, size_t loc); - -/* Declare a free list to manage the H5HP_t struct */ -H5FL_DEFINE_STATIC(H5HP_t); - -/* Declare a free list to manage sequences of H5HP_ent_t */ -H5FL_SEQ_DEFINE_STATIC(H5HP_ent_t); - -/*-------------------------------------------------------------------------- - NAME - H5HP__swim_max - PURPOSE - Restore heap condition by moving an object upward - USAGE - herr_t H5HP__swim_max(heap, loc) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - size_t loc; IN: Location to start from - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Restore the heap condition for the heap's array by "swimming" the object - at a location upward. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - This routine is for "maximum" value heaps. - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -static herr_t -H5HP__swim_max(H5HP_t *heap, size_t loc) -{ - int val; /* Temporary copy value of object to move in heap */ - H5HP_info_t *obj; /* Temporary pointer to object to move in heap */ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_STATIC_NOERR - - /* Get copies of the information about the object to move in the heap */ - val = heap->heap[loc].val; - obj = heap->heap[loc].obj; - - /* Move object up in heap until it's reached the maximum location possible */ - while (heap->heap[loc / 2].val < val) { - /* Move object "above" current location in heap down */ - heap->heap[loc].val = heap->heap[loc / 2].val; - heap->heap[loc].obj = heap->heap[loc / 2].obj; - - /* Update heap location for object which moved */ - heap->heap[loc].obj->heap_loc = loc; - - /* Move to location "above" current location */ - loc = loc / 2; - } /* end while */ - - /* Put object into heap at correct location */ - heap->heap[loc].val = val; - heap->heap[loc].obj = obj; - - /* Update heap location for object */ - heap->heap[loc].obj->heap_loc = loc; - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP__swim_max() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP__swim_min - PURPOSE - Restore heap condition by moving an object upward - USAGE - herr_t H5HP__swim_min(heap, loc) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - size_t loc; IN: Location to start from - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Restore the heap condition for the heap's array by "swimming" the object - at a location upward. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - This routine is for "minimum" value heaps. - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -static herr_t -H5HP__swim_min(H5HP_t *heap, size_t loc) -{ - int val; /* Temporary copy value of object to move in heap */ - H5HP_info_t *obj; /* Temporary pointer to object to move in heap */ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_STATIC_NOERR - - /* Get copies of the information about the object to move in the heap */ - val = heap->heap[loc].val; - obj = heap->heap[loc].obj; - - /* Move object up in heap until it's reached the minimum location possible */ - while (heap->heap[loc / 2].val > val) { - /* Move object "above" current location in heap down */ - heap->heap[loc].val = heap->heap[loc / 2].val; - heap->heap[loc].obj = heap->heap[loc / 2].obj; - - /* Update heap location for object which moved */ - heap->heap[loc].obj->heap_loc = loc; - - /* Move to location "above" current location */ - loc = loc / 2; - } /* end while */ - - /* Put object into heap at correct location */ - heap->heap[loc].val = val; - heap->heap[loc].obj = obj; - - /* Update heap location for object */ - heap->heap[loc].obj->heap_loc = loc; - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP__swim_min() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP__sink_max - PURPOSE - Restore heap condition by moving an object downward - USAGE - herr_t H5HP__sink_max(heap, loc) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - size_t loc; IN: Location to start from - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Restore the heap condition for the heap's array by "sinking" the object - at a location downward. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - This routine is for "maximum" value heaps. - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -static herr_t -H5HP__sink_max(H5HP_t *heap, size_t loc) -{ - int val; /* Temporary copy value of object to move in heap */ - void * obj; /* Temporary pointer to object to move in heap */ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_STATIC_NOERR - - /* Get copies of the information about the object to move in the heap */ - val = heap->heap[loc].val; - obj = heap->heap[loc].obj; - - /* Move object up in heap until it's reached the maximum location possible */ - while ((2 * loc) <= heap->nobjs) { - size_t new_loc = loc * 2; /* New object's potential location area */ - - /* Get the greater of the two objects below the location in heap */ - if (new_loc < heap->nobjs && (heap->heap[new_loc].val < heap->heap[new_loc + 1].val)) - new_loc++; - - /* Check if the object is smaller than the larger of the objects below it */ - /* If so, its in the correct location now, and we can get out */ - if (val >= heap->heap[new_loc].val) - break; - - /* Move the greater of the two objects below the current location up */ - heap->heap[loc].val = heap->heap[new_loc].val; - heap->heap[loc].obj = heap->heap[new_loc].obj; - - /* Update heap location for object which moved */ - heap->heap[loc].obj->heap_loc = loc; - - /* Move to location "below" current location */ - loc = new_loc; - } /* end while */ - - /* Put object into heap at correct location */ - heap->heap[loc].val = val; - heap->heap[loc].obj = (H5HP_info_t *)obj; - - /* Update heap location for object */ - heap->heap[loc].obj->heap_loc = loc; - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP__sink_max() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP__sink_min - PURPOSE - Restore heap condition by moving an object downward - USAGE - herr_t H5HP__sink_min(heap, loc) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - size_t loc; IN: Location to start from - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Restore the heap condition for the heap's array by "sinking" the object - at a location downward. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - This routine is for "minimum" value heaps. - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -static herr_t -H5HP__sink_min(H5HP_t *heap, size_t loc) -{ - int val; /* Temporary copy value of object to move in heap */ - void * obj; /* Temporary pointer to object to move in heap */ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_STATIC_NOERR - - /* Get copies of the information about the object to move in the heap */ - val = heap->heap[loc].val; - obj = heap->heap[loc].obj; - - /* Move object up in heap until it's reached the maximum location possible */ - while ((2 * loc) <= heap->nobjs) { - size_t new_loc = loc * 2; /* New object's potential location area */ - - /* Get the lesser of the two objects below the location in heap */ - if (new_loc < heap->nobjs && (heap->heap[new_loc].val > heap->heap[new_loc + 1].val)) - new_loc++; - - /* Check if the object is greater than the larger of the objects below it */ - /* If so, its in the correct location now, and we can get out */ - if (val <= heap->heap[new_loc].val) - break; - - /* Move the greater of the two objects below the current location up */ - heap->heap[loc].val = heap->heap[new_loc].val; - heap->heap[loc].obj = heap->heap[new_loc].obj; - - /* Update heap location for object which moved */ - heap->heap[loc].obj->heap_loc = loc; - - /* Move to location "below" current location */ - loc = new_loc; - } /* end while */ - - /* Put object into heap at correct location */ - heap->heap[loc].val = val; - heap->heap[loc].obj = (H5HP_info_t *)obj; - - /* Update heap location for object */ - heap->heap[loc].obj->heap_loc = loc; - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP__sink_min() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_create - PURPOSE - Create a heap - USAGE - H5HP_t *H5HP_create(heap_type) - H5HP_type_t heap_type; IN: Type of heap to create - - RETURNS - Returns a pointer to a heap on success, NULL on failure. - DESCRIPTION - Create a priority queue. The SIZE is used to set the initial number of - entries allocated. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -H5HP_t * -H5HP_create(H5HP_type_t heap_type) -{ - H5HP_t *new_heap = NULL; /* Pointer to new heap object created */ - H5HP_t *ret_value; /* Return value */ - - FUNC_ENTER_NOAPI(NULL) - - /* Check args */ - HDassert(heap_type == H5HP_MIN_HEAP || heap_type == H5HP_MAX_HEAP); - - /* Allocate ref-counted string structure */ - if ((new_heap = H5FL_MALLOC(H5HP_t)) == NULL) - HGOTO_ERROR(H5E_HEAP, H5E_NOSPACE, NULL, "memory allocation failed"); - - /* Allocate the array to store the heap entries */ - if ((new_heap->heap = H5FL_SEQ_MALLOC(H5HP_ent_t, (size_t)(H5HP_START_SIZE + 1))) == NULL) - HGOTO_ERROR(H5E_HEAP, H5E_NOSPACE, NULL, "memory allocation failed"); - - /* Set the internal fields */ - new_heap->type = heap_type; - new_heap->nobjs = 0; - new_heap->nalloc = H5HP_START_SIZE + 1; - - /* Set the information in the 0'th location based on the type of heap */ - if (heap_type == H5HP_MIN_HEAP) { - /* Set the value in the '0' location to be the minimum value, to - * simplify the algorithms - */ - new_heap->heap[0].val = INT_MIN; - new_heap->heap[0].obj = NULL; - } /* end if */ - else { - /* Set the value in the '0' location to be the maximum value, to - * simplify the algorithms - */ - new_heap->heap[0].val = INT_MAX; - new_heap->heap[0].obj = NULL; - } /* end else */ - - /* Set the return value */ - ret_value = new_heap; - -done: - /* Error cleanup */ - if (NULL == ret_value) { - if (NULL != new_heap) { - if (NULL != new_heap->heap) - new_heap->heap = H5FL_SEQ_FREE(H5HP_ent_t, new_heap->heap); - new_heap = H5FL_FREE(H5HP_t, new_heap); - } /* end if */ - } /* end if */ - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP_create() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_count - PURPOSE - Check the number of elements in a heap - USAGE - ssize_t H5HP_count(heap) - const H5HP_t *heap; IN: Pointer to heap to query - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Checks the number of elements in heap - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -ssize_t -H5HP_count(const H5HP_t *heap) -{ - ssize_t ret_value; /* Return value */ - - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check args */ - HDassert(heap); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - /* Return the number of objects in the heap */ - H5_CHECK_OVERFLOW(heap->nobjs, size_t, ssize_t); - ret_value = (ssize_t)heap->nobjs; - - /* No post-condition check necessary, since heap is constant */ - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP_count() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_insert - PURPOSE - Insert an object into a heap, with an initial value - USAGE - herr_t H5HP_insert(heap, val, obj) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - int val; IN: Initial value for object in heap - void *obj; IN: Pointer to object to insert into heap - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Inserts a OBJ into a HEAP, with an initial VALue. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -herr_t -H5HP_insert(H5HP_t *heap, int val, void *obj) -{ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_NOAPI(FAIL) - - /* Check args */ - HDassert(heap); - HDassert(obj); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - /* Increment number of objects in heap */ - heap->nobjs++; - - /* Check if we need to allocate more room for heap array */ - if (heap->nobjs >= heap->nalloc) { - size_t n = MAX(H5HP_START_SIZE, 2 * (heap->nalloc - 1)) + 1; - H5HP_ent_t *new_heap = H5FL_SEQ_REALLOC(H5HP_ent_t, heap->heap, n); - - if (!new_heap) - HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "unable to extend heap array"); - heap->heap = new_heap; - heap->nalloc = n; - } /* end if */ - - /* Insert new object at end of heap */ - heap->heap[heap->nobjs].val = val; - heap->heap[heap->nobjs].obj = (H5HP_info_t *)obj; - heap->heap[heap->nobjs].obj->heap_loc = heap->nobjs; - - /* Restore heap condition */ - if (heap->type == H5HP_MAX_HEAP) { - if (H5HP__swim_max(heap, heap->nobjs) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTINSERT, FAIL, "unable to restore heap condition"); - } /* end if */ - else { - if (H5HP__swim_min(heap, heap->nobjs) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTINSERT, FAIL, "unable to restore heap condition"); - } /* end else */ - -done: - - /* Check internal consistency */ - /* (Post-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP_insert() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_top - PURPOSE - Check the value of the top object in the heap - USAGE - herr_t H5HP_top(heap, val) - const H5HP_t *heap; IN: Pointer to heap to modify - int val; IN/OUT: Initial value for object in heap - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Checks the value of the top object in a heap - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -herr_t -H5HP_top(const H5HP_t *heap, int *val) -{ - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check args */ - HDassert(heap); - HDassert(val); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - /* Get value of the top object in the heap */ - *val = heap->heap[1].val; - - /* No post-condition check necessary, since heap is constant */ - FUNC_LEAVE_NOAPI(SUCCEED); -} /* end H5HP_top() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_remove - PURPOSE - Remove an object into a heap - USAGE - herr_t H5HP_remove(heap, val, obj) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - int *val; OUT: Pointer to value of object removed from heap - void **obj; OUT: Pointer to object removed from heap - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Removes the top object on a heap, returning its value and object pointer - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -herr_t -H5HP_remove(H5HP_t *heap, int *val, void **obj) -{ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_NOAPI(FAIL) - - /* Check args */ - HDassert(heap); - HDassert(val); - HDassert(obj); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - /* Check if there are any objects on the heap to remove */ - if (heap->nobjs == 0) - HGOTO_ERROR(H5E_HEAP, H5E_NOTFOUND, FAIL, "heap is empty"); - - /* Get the information for the top object on the heap */ - HDassert(heap->heap[1].obj->heap_loc == 1); - *val = heap->heap[1].val; - *obj = heap->heap[1].obj; - - /* Move the last element in the heap to the top */ - heap->heap[1].val = heap->heap[heap->nobjs].val; - heap->heap[1].obj = heap->heap[heap->nobjs].obj; - heap->heap[1].obj->heap_loc = 1; - - /* Decrement number of objects in heap */ - heap->nobjs--; - - /* Restore heap condition, if there are objects on the heap */ - if (heap->nobjs > 0) { - if (heap->type == H5HP_MAX_HEAP) { - if (H5HP__sink_max(heap, (size_t)1) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTDELETE, FAIL, "unable to restore heap condition"); - } /* end if */ - else { - if (H5HP__sink_min(heap, (size_t)1) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTDELETE, FAIL, "unable to restore heap condition"); - } /* end else */ - } /* end if */ - -done: - - /* Check internal consistency */ - /* (Post-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP_remove() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_change - PURPOSE - Change the priority of an object on a heap - USAGE - herr_t H5HP_change(heap, val, obj) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - int val; IN: New priority value for object - void *obj; IN: Pointer to object to modify - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Changes the priority of an object on a heap. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -herr_t -H5HP_change(H5HP_t *heap, int val, void *_obj) -{ - H5HP_info_t *obj = (H5HP_info_t *)_obj; /* Alias for object */ - size_t obj_loc; /* Location of object in heap */ - int old_val; /* Object's old priority value */ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_NOAPI(FAIL) - - /* Check args */ - HDassert(heap); - HDassert(obj); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - /* Get the location of the object in the heap */ - obj_loc = obj->heap_loc; - HDassert(obj_loc > 0 && obj_loc <= heap->nobjs); - - /* Change the heap object's priority */ - old_val = heap->heap[obj_loc].val; - heap->heap[obj_loc].val = val; - - /* Restore heap condition */ - if (val < old_val) { - if (heap->type == H5HP_MAX_HEAP) { - if (H5HP__sink_max(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition"); - } /* end if */ - else { - if (H5HP__swim_min(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition"); - } /* end else */ - } /* end if */ - else { - if (heap->type == H5HP_MAX_HEAP) { - if (H5HP__swim_max(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition"); - } /* end if */ - else { - if (H5HP__sink_min(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition"); - } /* end else */ - } /* end else */ - -done: - - /* Check internal consistency */ - /* (Post-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP_change() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_incr - PURPOSE - Increment the priority of an object on a heap - USAGE - herr_t H5HP_incr(heap, amt, obj) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - unsigned amt; IN: Amount to increase priority by - void *obj; IN: Pointer to object to modify - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Increments the priority of an object on a heap by one. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -herr_t -H5HP_incr(H5HP_t *heap, unsigned amt, void *_obj) -{ - H5HP_info_t *obj = (H5HP_info_t *)_obj; /* Alias for object */ - size_t obj_loc; /* Location of object in heap */ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_NOAPI(FAIL) - - /* Check args */ - HDassert(heap); - HDassert(obj); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - /* Get the location of the object in the heap */ - obj_loc = obj->heap_loc; - HDassert(obj_loc > 0 && obj_loc <= heap->nobjs); - - /* Change the heap object's priority */ - heap->heap[obj_loc].val += (int)amt; - - /* Restore heap condition */ - if (H5HP_MAX_HEAP == heap->type) { - if (H5HP__swim_max(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition") - } /* end if */ - else { - if (H5HP__sink_min(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition") - } /* end else */ - -done: - - /* Check internal consistency */ - /* (Post-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP_incr() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_decr - PURPOSE - Decrement the priority of an object on a heap - USAGE - herr_t H5HP_dec(heap, amt, obj) - H5HP_t *heap; IN/OUT: Pointer to heap to modify - unsigned amt; IN: Amount to decrease priority by - void *obj; IN: Pointer to object to modify - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Decrements the priority of an object on a heap by one. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -herr_t -H5HP_decr(H5HP_t *heap, unsigned amt, void *_obj) -{ - H5HP_info_t *obj = (H5HP_info_t *)_obj; /* Alias for object */ - size_t obj_loc; /* Location of object in heap */ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_NOAPI(FAIL) - - /* Check args */ - HDassert(heap); - HDassert(obj); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - /* Get the location of the object in the heap */ - obj_loc = obj->heap_loc; - HDassert(obj_loc > 0 && obj_loc <= heap->nobjs); - - /* Change the heap object's priority */ - H5_CHECK_OVERFLOW(amt, unsigned, int); - heap->heap[obj_loc].val -= (int)amt; - - /* Restore heap condition */ - if (heap->type == H5HP_MAX_HEAP) { - if (H5HP__sink_max(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition"); - } /* end if */ - else { - if (H5HP__swim_min(heap, obj_loc) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTRESTORE, FAIL, "unable to restore heap condition"); - } /* end else */ - -done: - - /* Check internal consistency */ - /* (Post-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(heap->heap[0].obj == NULL); - - FUNC_LEAVE_NOAPI(ret_value); -} /* end H5HP_decr() */ - -/*-------------------------------------------------------------------------- - NAME - H5HP_close - PURPOSE - Close a heap, deallocating it. - USAGE - herr_t H5HP_close(heap) - H5HP_t *heap; IN/OUT: Pointer to heap to close - - RETURNS - Returns non-negative on success, negative on failure. - DESCRIPTION - Close a heap, freeing all internal information. Any objects left in - the heap are not deallocated. - GLOBAL VARIABLES - COMMENTS, BUGS, ASSUMPTIONS - EXAMPLES - REVISION LOG ---------------------------------------------------------------------------*/ -herr_t -H5HP_close(H5HP_t *heap) -{ - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check args */ - HDassert(heap); - - /* Check internal consistency */ - /* (Pre-condition) */ - HDassert(heap->nobjs < heap->nalloc); - HDassert(heap->heap); - HDassert((heap->type == H5HP_MAX_HEAP && heap->heap[0].val == INT_MAX) || - (heap->type == H5HP_MIN_HEAP && heap->heap[0].val == INT_MIN)); - HDassert(NULL == heap->heap[0].obj); - - /* Free internal structures for heap */ - heap->heap = H5FL_SEQ_FREE(H5HP_ent_t, heap->heap); - - /* Free actual heap object */ - heap = H5FL_FREE(H5HP_t, heap); - - FUNC_LEAVE_NOAPI(SUCCEED) -} /* end H5HP_close() */ diff --git a/src/H5HPprivate.h b/src/H5HPprivate.h deleted file mode 100644 index 50020bc..0000000 --- a/src/H5HPprivate.h +++ /dev/null @@ -1,68 +0,0 @@ -/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * - * Copyright by The HDF Group. * - * Copyright by the Board of Trustees of the University of Illinois. * - * All rights reserved. * - * * - * This file is part of HDF5. The full HDF5 copyright notice, including * - * terms governing use, modification, and redistribution, is contained in * - * the COPYING file, which can be found at the root of the source code * - * distribution tree, or in https://www.hdfgroup.org/licenses. * - * If you do not have access to either file, you may request a copy from * - * help@hdfgroup.org. * - * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ - -/* - * This file contains private information about the H5HP module - */ -#ifndef H5HPprivate_H -#define H5HPprivate_H - -/**************************************/ -/* Public headers needed by this file */ -/**************************************/ -#ifdef LATER -#include "H5HPpublic.h" -#endif /* LATER */ - -/***************************************/ -/* Private headers needed by this file */ -/***************************************/ -#include "H5private.h" - -/************/ -/* Typedefs */ -/************/ - -/* Typedef for heap struct (defined in H5HP.c) */ -typedef struct H5HP_t H5HP_t; - -/* Typedef for objects which can be inserted into heaps */ -/* This _must_ be the first field in objects which can be inserted into heaps */ -typedef struct H5HP_info_t { - size_t heap_loc; /* Location of object in heap */ -} H5HP_info_t; - -/* Typedef for type of heap to create */ -typedef enum { - H5HP_MIN_HEAP, /* Minimum values in heap are at the "top" */ - H5HP_MAX_HEAP /* Maximum values in heap are at the "top" */ -} H5HP_type_t; - -/**********/ -/* Macros */ -/**********/ - -/********************/ -/* Private routines */ -/********************/ -H5_DLL H5HP_t *H5HP_create(H5HP_type_t heap_type); -H5_DLL herr_t H5HP_insert(H5HP_t *heap, int val, void *obj); -H5_DLL ssize_t H5HP_count(const H5HP_t *heap); -H5_DLL herr_t H5HP_top(const H5HP_t *heap, int *val); -H5_DLL herr_t H5HP_remove(H5HP_t *heap, int *val, void **ptr); -H5_DLL herr_t H5HP_change(H5HP_t *heap, int val, void *obj); -H5_DLL herr_t H5HP_incr(H5HP_t *heap, unsigned amt, void *obj); -H5_DLL herr_t H5HP_decr(H5HP_t *heap, unsigned amt, void *obj); -H5_DLL herr_t H5HP_close(H5HP_t *heap); - -#endif /* H5HPprivate_H */ diff --git a/src/H5Lpublic.h b/src/H5Lpublic.h index e5a826a..ca5f6e6 100644 --- a/src/H5Lpublic.h +++ b/src/H5Lpublic.h @@ -580,7 +580,7 @@ H5_DLL herr_t H5Lget_val_by_idx(hid_t loc_id, const char *group_name, H5_index_t * name includes either a relative path or an absolute path to the * target link, intermediate steps along the path must be verified * before the existence of the target link can be safely checked. If - * the path is not verified and an intermediate element of the path + * the path is not verified, and an intermediate element of the path * does not exist, H5Lexists() will fail. The example in the next * paragraph illustrates one step-by-step method for verifying the * existence of a link with a relative or absolute path. @@ -620,13 +620,13 @@ H5_DLL herr_t H5Lget_val_by_idx(hid_t loc_id, const char *group_name, H5_index_t * H5Lexists() with arguments \c file, \c "/", and \c lapl * returns a positive value; in other words, * \Code{H5Lexists(file, "/", lapl)} returns a positive value. - * In HDF5 version 1.8.16, this function returns 0.</li> + * In the HDF5 1.8 release, this function returns 0.</li> * <li>Let \c root denote a valid HDF5 group identifier that refers to the * root group of an HDF5 file, and let \c lapl denote a valid link * access property list identifier. A call to H5Lexists() with * arguments c root, \c "/", and \c lapl returns a positive value; * in other words, \Code{H5Lexists(root, "/", lapl)} returns a positive - * value. In HDF5 version 1.8.16, this function returns 0.</li> + * value. In the HDF5 1.8 release, this function returns 0.</li> * </ol> * Note that the function accepts link names and path names. This is * potentially misleading to callers, and we plan to separate the @@ -650,8 +650,10 @@ H5MF__add_sect(H5F_t *f, H5FD_mem_t alloc_type, H5FS_t *fspace, H5MF_free_sectio H5AC_set_ring(fsm_ring, &orig_ring); #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: adding node, node->sect_info.addr = %a, node->sect_info.size = %Hu\n", __func__, - node->sect_info.addr, node->sect_info.size); + HDfprintf(stderr, + "%s: adding node, node->sect_info.addr = %" PRIuHADDR ", node->sect_info.size = %" PRIuHSIZE + "\n", + __func__, node->sect_info.addr, node->sect_info.size); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* Add the section */ if (H5FS_sect_add(f, fspace, (H5FS_section_info_t *)node, H5FS_ADD_RETURNED_SPACE, &udata) < 0) @@ -703,7 +705,7 @@ H5MF__find_sect(H5F_t *f, H5FD_mem_t alloc_type, hsize_t size, H5FS_t *fspace, h HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "error locating free space in file") #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: section found = %t\n", __func__, ret_value); + HDfprintf(stderr, "%s: section found = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* Check for actually finding section */ @@ -731,7 +733,7 @@ H5MF__find_sect(H5F_t *f, H5FD_mem_t alloc_type, hsize_t size, H5FS_t *fspace, h node->sect_info.size -= size; #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: re-adding node, node->sect_info.size = %Hu\n", __func__, + HDfprintf(stderr, "%s: re-adding node, node->sect_info.size = %" PRIuHSIZE "\n", __func__, node->sect_info.size); #endif /* H5MF_ALLOC_DEBUG_MORE */ @@ -775,7 +777,7 @@ H5MF_alloc(H5F_t *f, H5FD_mem_t alloc_type, hsize_t size) FUNC_ENTER_NOAPI_TAG(H5AC__FREESPACE_TAG, HADDR_UNDEF) #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: alloc_type = %u, size = %Hu\n", __func__, (unsigned)alloc_type, size); + HDfprintf(stderr, "%s: alloc_type = %u, size = %" PRIuHSIZE "\n", __func__, (unsigned)alloc_type, size); #endif /* H5MF_ALLOC_DEBUG */ /* check arguments */ @@ -848,7 +850,8 @@ done: H5AC_set_ring(orig_ring, NULL); #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: Leaving: ret_value = %a, size = %Hu\n", __func__, ret_value, size); + HDfprintf(stderr, "%s: Leaving: ret_value = %" PRIuHADDR ", size = %" PRIuHSIZE "\n", __func__, ret_value, + size); #endif /* H5MF_ALLOC_DEBUG */ #ifdef H5MF_ALLOC_DEBUG_DUMP H5MF__sects_dump(f, stderr); @@ -888,7 +891,7 @@ H5MF__alloc_pagefs(H5F_t *f, H5FD_mem_t alloc_type, hsize_t size) FUNC_ENTER_STATIC #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: alloc_type = %u, size = %Hu\n", __func__, (unsigned)alloc_type, size); + HDfprintf(stderr, "%s: alloc_type = %u, size = %" PRIuHSIZE "\n", __func__, (unsigned)alloc_type, size); #endif /* H5MF_ALLOC_DEBUG */ H5MF__alloc_to_fs_type(f->shared, alloc_type, size, &ptype); @@ -985,7 +988,8 @@ H5MF__alloc_pagefs(H5F_t *f, H5FD_mem_t alloc_type, hsize_t size) done: #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: Leaving: ret_value = %a, size = %Hu\n", __func__, ret_value, size); + HDfprintf(stderr, "%s: Leaving: ret_value = %" PRIuHADDR ", size = %" PRIuHSIZE "\n", __func__, ret_value, + size); #endif /* H5MF_ALLOC_DEBUG */ #ifdef H5MF_ALLOC_DEBUG_DUMP H5MF__sects_dump(f, stderr); @@ -1031,7 +1035,7 @@ H5MF_alloc_tmp(H5F_t *f, hsize_t size) FUNC_ENTER_NOAPI(HADDR_UNDEF) #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: size = %Hu\n", __func__, size); + HDfprintf(stderr, "%s: size = %" PRIuHSIZE "\n", __func__, size); #endif /* H5MF_ALLOC_DEBUG */ /* check args */ @@ -1083,8 +1087,8 @@ H5MF_xfree(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size) FUNC_ENTER_NOAPI_TAG(H5AC__FREESPACE_TAG, FAIL) #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: Entering - alloc_type = %u, addr = %a, size = %Hu\n", __func__, - (unsigned)alloc_type, addr, size); + HDfprintf(stderr, "%s: Entering - alloc_type = %u, addr = %" PRIuHADDR ", size = %" PRIuHSIZE "\n", + __func__, (unsigned)alloc_type, addr, size); #endif /* H5MF_ALLOC_DEBUG */ /* check arguments */ @@ -1133,7 +1137,7 @@ H5MF_xfree(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size) * space is at the end of the file */ #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: fs_addr = %a\n", __func__, f->shared->fs_addr[fs_type]); + HDfprintf(stderr, "%s: fs_addr = %" PRIuHADDR "\n", __func__, f->shared->fs_addr[fs_type]); #endif /* H5MF_ALLOC_DEBUG_MORE */ if (!H5F_addr_defined(f->shared->fs_addr[fs_type])) { htri_t status; /* "can absorb" status for section into */ @@ -1149,8 +1153,9 @@ H5MF_xfree(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size) HGOTO_DONE(SUCCEED) else if (size < f->shared->fs_threshold) { #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: dropping addr = %a, size = %Hu, on the floor!\n", __func__, addr, - size); + HDfprintf(stderr, + "%s: dropping addr = %" PRIuHADDR ", size = %" PRIuHSIZE ", on the floor!\n", + __func__, addr, size); #endif /* H5MF_ALLOC_DEBUG_MORE */ HGOTO_DONE(SUCCEED) } /* end else-if */ @@ -1167,7 +1172,8 @@ H5MF_xfree(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size) */ if (f->shared->fs_state[fs_type] == H5F_FS_STATE_DELETING || !H5F_HAVE_FREE_SPACE_MANAGER(f)) { #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: dropping addr = %a, size = %Hu, on the floor!\n", __func__, addr, size); + HDfprintf(stderr, "%s: dropping addr = %" PRIuHADDR ", size = %" PRIuHSIZE ", on the floor!\n", + __func__, addr, size); #endif /* H5MF_ALLOC_DEBUG_MORE */ HGOTO_DONE(SUCCEED) } /* end if */ @@ -1276,7 +1282,9 @@ H5MF_try_extend(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size, hsi FUNC_ENTER_NOAPI_TAG(H5AC__FREESPACE_TAG, FAIL) #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: Entering: alloc_type = %u, addr = %a, size = %Hu, extra_requested = %Hu\n", + HDfprintf(stderr, + "%s: Entering: alloc_type = %u, addr = %" PRIuHADDR ", size = %" PRIuHSIZE + ", extra_requested = %" PRIuHSIZE "\n", __func__, (unsigned)alloc_type, addr, size, extra_requested); #endif /* H5MF_ALLOC_DEBUG */ @@ -1329,7 +1337,7 @@ H5MF_try_extend(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size, hsi if ((ret_value = H5F__try_extend(f, map_type, end, extra_requested + frag_size)) < 0) HGOTO_ERROR(H5E_RESOURCE, H5E_CANTEXTEND, FAIL, "error extending file") #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: extended = %t\n", __func__, ret_value); + HDfprintf(stderr, "%s: extended = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* If extending at EOA succeeds: */ @@ -1367,7 +1375,7 @@ H5MF_try_extend(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size, hsi HGOTO_ERROR(H5E_RESOURCE, H5E_CANTEXTEND, FAIL, "error extending aggregation block") #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: H5MF__aggr_try_extend = %t\n", __func__, ret_value); + HDfprintf(stderr, "%s: H5MF__aggr_try_extend = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG_MORE */ } /* end if */ @@ -1393,7 +1401,7 @@ H5MF_try_extend(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size, hsi HGOTO_ERROR(H5E_RESOURCE, H5E_CANTEXTEND, FAIL, "error extending block in free space manager") #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: Try to H5FS_sect_try_extend = %t\n", __func__, ret_value); + HDfprintf(stderr, "%s: Try to H5FS_sect_try_extend = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG_MORE */ } /* end if */ @@ -1404,7 +1412,7 @@ H5MF_try_extend(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size, hsi if (frag_size <= H5F_PGEND_META_THRES(f) && extra_requested <= frag_size) ret_value = TRUE; #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: Try to extend into the page end threshold = %t\n", __func__, + HDfprintf(stderr, "%s: Try to extend into the page end threshold = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG_MORE */ } /* end if */ @@ -1417,7 +1425,7 @@ done: H5AC_set_ring(orig_ring, NULL); #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: Leaving: ret_value = %t\n", __func__, ret_value); + HDfprintf(stderr, "%s: Leaving: ret_value = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG */ #ifdef H5MF_ALLOC_DEBUG_DUMP H5MF__sects_dump(f, stderr); @@ -1452,8 +1460,8 @@ H5MF_try_shrink(H5F_t *f, H5FD_mem_t alloc_type, haddr_t addr, hsize_t size) FUNC_ENTER_NOAPI_TAG(H5AC__FREESPACE_TAG, FAIL) #ifdef H5MF_ALLOC_DEBUG - HDfprintf(stderr, "%s: Entering - alloc_type = %u, addr = %a, size = %Hu\n", __func__, - (unsigned)alloc_type, addr, size); + HDfprintf(stderr, "%s: Entering - alloc_type = %u, addr = %" PRIuHADDR ", size = %" PRIuHSIZE "\n", + __func__, (unsigned)alloc_type, addr, size); #endif /* H5MF_ALLOC_DEBUG */ /* check arguments */ @@ -1589,8 +1597,9 @@ H5MF__close_delete_fstype(H5F_t *f, H5F_mem_page_t type) HDassert((H5FD_mem_t)type < H5FD_MEM_NTYPES); #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: Check 1.0 - f->shared->fs_man[%u] = %p, f->shared->fs_addr[%u] = %a\n", __func__, - (unsigned)type, f->shared->fs_man[type], (unsigned)type, f->shared->fs_addr[type]); + HDfprintf(stderr, "%s: Check 1.0 - f->shared->fs_man[%u] = %p, f->shared->fs_addr[%u] = %" PRIuHADDR "\n", + __func__, (unsigned)type, (void *)f->shared->fs_man[type], (unsigned)type, + f->shared->fs_addr[type]); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* If the free space manager for this type is open, close it */ @@ -1599,8 +1608,9 @@ H5MF__close_delete_fstype(H5F_t *f, H5F_mem_page_t type) HGOTO_ERROR(H5E_RESOURCE, H5E_CANTRELEASE, FAIL, "can't close the free space manager") #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: Check 2.0 - f->shared->fs_man[%u] = %p, f->shared->fs_addr[%u] = %a\n", __func__, - (unsigned)type, f->shared->fs_man[type], (unsigned)type, f->shared->fs_addr[type]); + HDfprintf(stderr, "%s: Check 2.0 - f->shared->fs_man[%u] = %p, f->shared->fs_addr[%u] = %" PRIuHADDR "\n", + __func__, (unsigned)type, (void *)f->shared->fs_man[type], (unsigned)type, + f->shared->fs_addr[type]); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* If there is free space manager info for this type, delete it */ diff --git a/src/H5MFaggr.c b/src/H5MFaggr.c index 8ffc5c8..78bf620 100644 --- a/src/H5MFaggr.c +++ b/src/H5MFaggr.c @@ -92,7 +92,7 @@ H5MF_aggr_vfd_alloc(H5F_t *f, H5FD_mem_t alloc_type, hsize_t size) FUNC_ENTER_NOAPI(HADDR_UNDEF) #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: alloc_type = %u, size = %Hu\n", __func__, (unsigned)alloc_type, size); + HDfprintf(stderr, "%s: alloc_type = %u, size = %" PRIuHSIZE "\n", __func__, (unsigned)alloc_type, size); #endif /* H5MF_AGGR_DEBUG */ /* check arguments */ @@ -120,7 +120,8 @@ H5MF_aggr_vfd_alloc(H5F_t *f, H5FD_mem_t alloc_type, hsize_t size) done: #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: Leaving: ret_value = %a, size = %Hu\n", __func__, ret_value, size); + HDfprintf(stderr, "%s: Leaving: ret_value = %" PRIuHADDR ", size = %" PRIuHSIZE "\n", __func__, ret_value, + size); #endif /* H5MF_AGGR_DEBUG */ FUNC_LEAVE_NOAPI(ret_value) @@ -150,7 +151,7 @@ H5MF__aggr_alloc(H5F_t *f, H5F_blk_aggr_t *aggr, H5F_blk_aggr_t *other_aggr, H5F FUNC_ENTER_STATIC #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: type = %u, size = %Hu\n", __func__, (unsigned)type, size); + HDfprintf(stderr, "%s: type = %u, size = %" PRIuHSIZE "\n", __func__, (unsigned)type, size); #endif /* H5MF_AGGR_DEBUG */ /* check args */ @@ -199,8 +200,8 @@ H5MF__aggr_alloc(H5F_t *f, H5F_blk_aggr_t *aggr, H5F_blk_aggr_t *other_aggr, H5F H5FD_mem_t alloc_type, other_alloc_type; /* Current aggregator & 'other' aggregator types */ #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: aggr = {%a, %Hu, %Hu}\n", __func__, aggr->addr, aggr->tot_size, - aggr->size); + HDfprintf(stderr, "%s: aggr = {%" PRIuHADDR ", %" PRIuHSIZE ", %" PRIuHSIZE "}\n", __func__, + aggr->addr, aggr->tot_size, aggr->size); #endif /* H5MF_AGGR_DEBUG */ /* Turn off alignment if allocation < threshold */ @@ -388,7 +389,7 @@ H5MF__aggr_alloc(H5F_t *f, H5F_blk_aggr_t *aggr, H5F_blk_aggr_t *other_aggr, H5F done: #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: ret_value = %a\n", __func__, ret_value); + HDfprintf(stderr, "%s: ret_value = %" PRIuHADDR "\n", __func__, ret_value); #endif /* H5MF_AGGR_DEBUG */ FUNC_LEAVE_NOAPI(ret_value) } /* end H5MF__aggr_alloc() */ @@ -537,8 +538,11 @@ done: if (H5F_addr_eq((sect->sect_info.addr + sect->sect_info.size), aggr->addr) || H5F_addr_eq((aggr->addr + aggr->size), sect->sect_info.addr)) { #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: section {%a, %Hu} adjoins aggr = {%a, %Hu}\n", "H5MF__aggr_can_absorb", - sect->sect_info.addr, sect->sect_info.size, aggr->addr, aggr->size); + HDfprintf(stderr, + "%s: section {%" PRIuHADDR ", %" PRIuHSIZE "} adjoins aggr = {%" PRIuHADDR + ", %" PRIuHSIZE "}\n", + "H5MF__aggr_can_absorb", sect->sect_info.addr, sect->sect_info.size, aggr->addr, + aggr->size); #endif /* H5MF_AGGR_DEBUG */ /* Check if aggregator would get too large and should be absorbed into section */ if ((aggr->size + sect->sect_info.size) >= aggr->alloc_size) @@ -587,7 +591,9 @@ done: /* Check if the section adjoins the beginning or end of the aggregator */ if (H5F_addr_eq((sect->sect_info.addr + sect->sect_info.size), aggr->addr)) { #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: aggr {%a, %Hu} adjoins front of section = {%a, %Hu}\n", + HDfprintf(stderr, + "%s: aggr {%" PRIuHADDR ", %" PRIuHSIZE "} adjoins front of section = {%" PRIuHADDR + ", %" PRIuHSIZE "}\n", "H5MF__aggr_absorb", aggr->addr, aggr->size, sect->sect_info.addr, sect->sect_info.size); #endif /* H5MF_AGGR_DEBUG */ @@ -599,7 +605,9 @@ done: HDassert(H5F_addr_eq((aggr->addr + aggr->size), sect->sect_info.addr)); #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: aggr {%a, %Hu} adjoins end of section = {%a, %Hu}\n", + HDfprintf(stderr, + "%s: aggr {%" PRIuHADDR ", %" PRIuHSIZE "} adjoins end of section = {%" PRIuHADDR + ", %" PRIuHSIZE "}\n", "H5MF__aggr_absorb", aggr->addr, aggr->size, sect->sect_info.addr, sect->sect_info.size); #endif /* H5MF_AGGR_DEBUG */ @@ -617,7 +625,9 @@ done: /* Check if the section adjoins the beginning or end of the aggregator */ if (H5F_addr_eq((sect->sect_info.addr + sect->sect_info.size), aggr->addr)) { #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: section {%a, %Hu} adjoins front of aggr = {%a, %Hu}\n", + HDfprintf(stderr, + "%s: section {%" PRIuHADDR ", %" PRIuHSIZE "} adjoins front of aggr = {%" PRIuHADDR + ", %" PRIuHSIZE "}\n", "H5MF__aggr_absorb", sect->sect_info.addr, sect->sect_info.size, aggr->addr, aggr->size); #endif /* H5MF_AGGR_DEBUG */ @@ -635,7 +645,9 @@ done: HDassert(H5F_addr_eq((aggr->addr + aggr->size), sect->sect_info.addr)); #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: section {%a, %Hu} adjoins end of aggr = {%a, %Hu}\n", + HDfprintf(stderr, + "%s: section {%" PRIuHADDR ", %" PRIuHSIZE "} adjoins end of aggr = {%" PRIuHADDR + ", %" PRIuHSIZE "}\n", "H5MF__aggr_absorb", sect->sect_info.addr, sect->sect_info.size, aggr->addr, aggr->size); #endif /* H5MF_AGGR_DEBUG */ @@ -723,7 +735,8 @@ done: tmp_addr = aggr->addr; tmp_size = aggr->size; #ifdef H5MF_AGGR_DEBUG - HDfprintf(stderr, "%s: tmp_addr = %a, tmp_size = %Hu\n", __func__, tmp_addr, tmp_size); + HDfprintf(stderr, "%s: tmp_addr = %" PRIuHADDR ", tmp_size = %" PRIuHSIZE "\n", __func__, + tmp_addr, tmp_size); #endif /* H5MF_AGGR_DEBUG */ /* Reset aggregator block information */ diff --git a/src/H5MFsection.c b/src/H5MFsection.c index 69b2ca0..13675f5 100644 --- a/src/H5MFsection.c +++ b/src/H5MFsection.c @@ -472,8 +472,9 @@ H5MF__sect_simple_can_shrink(const H5FS_section_info_t *_sect, void *_udata) /* Set the shrinking type */ udata->shrink = H5MF_SHRINK_EOA; #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: section {%a, %Hu}, shrinks file, eoa = %a\n", __func__, sect->sect_info.addr, - sect->sect_info.size, eoa); + HDfprintf(stderr, + "%s: section {%" PRIuHADDR ", %" PRIuHSIZE "}, shrinks file, eoa = %" PRIuHADDR "\n", + __func__, sect->sect_info.addr, sect->sect_info.size, eoa); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* Indicate shrinking can occur */ @@ -496,8 +497,9 @@ H5MF__sect_simple_can_shrink(const H5FS_section_info_t *_sect, void *_udata) /* Set the aggregator to operate on */ udata->aggr = &(udata->f->shared->meta_aggr); #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: section {%a, %Hu}, adjoins metadata aggregator\n", __func__, - sect->sect_info.addr, sect->sect_info.size); + HDfprintf(stderr, + "%s: section {%" PRIuHADDR ", %" PRIuHSIZE "}, adjoins metadata aggregator\n", + __func__, sect->sect_info.addr, sect->sect_info.size); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* Indicate shrinking can occur */ @@ -517,8 +519,9 @@ H5MF__sect_simple_can_shrink(const H5FS_section_info_t *_sect, void *_udata) /* Set the aggregator to operate on */ udata->aggr = &(udata->f->shared->sdata_aggr); #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: section {%a, %Hu}, adjoins small data aggregator\n", __func__, - sect->sect_info.addr, sect->sect_info.size); + HDfprintf(stderr, + "%s: section {%" PRIuHADDR ", %" PRIuHSIZE "}, adjoins small data aggregator\n", + __func__, sect->sect_info.addr, sect->sect_info.size); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* Indicate shrinking can occur */ @@ -625,8 +628,8 @@ H5MF__sect_small_add(H5FS_section_info_t **_sect, unsigned *flags, void *_udata) FUNC_ENTER_STATIC #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: Entering, section {%a, %Hu}\n", __func__, (*sect)->sect_info.addr, - (*sect)->sect_info.size); + HDfprintf(stderr, "%s: Entering, section {%" PRIuHADDR ", %" PRIuHSIZE "}\n", __func__, + (*sect)->sect_info.addr, (*sect)->sect_info.size); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* Do not adjust the section raw data or global heap data */ @@ -653,8 +656,8 @@ H5MF__sect_small_add(H5FS_section_info_t **_sect, unsigned *flags, void *_udata) else if (prem <= H5F_PGEND_META_THRES(udata->f)) { (*sect)->sect_info.size += prem; #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: section is adjusted {%a, %Hu}\n", __func__, (*sect)->sect_info.addr, - (*sect)->sect_info.size); + HDfprintf(stderr, "%s: section is adjusted {%" PRIuHADDR ", %" PRIuHSIZE "}\n", __func__, + (*sect)->sect_info.addr, (*sect)->sect_info.size); #endif /* H5MF_ALLOC_DEBUG_MORE */ } /* end if */ @@ -702,7 +705,7 @@ H5MF__sect_small_can_merge(const H5FS_section_info_t *_sect1, const H5FS_section ret_value = FALSE; #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: Leaving: ret_value = %t\n", __func__, ret_value); + HDfprintf(stderr, "%s: Leaving: ret_value = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG_MORE */ FUNC_LEAVE_NOAPI(ret_value) @@ -806,7 +809,7 @@ H5MF__sect_large_can_merge(const H5FS_section_info_t *_sect1, const H5FS_section ret_value = H5F_addr_eq(sect1->sect_info.addr + sect1->sect_info.size, sect2->sect_info.addr); #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: Leaving: ret_value = %t\n", __func__, ret_value); + HDfprintf(stderr, "%s: Leaving: ret_value = %d\n", __func__, ret_value); #endif /* H5MF_ALLOC_DEBUG_MORE */ FUNC_LEAVE_NOAPI(ret_value) @@ -894,8 +897,9 @@ H5MF__sect_large_can_shrink(const H5FS_section_info_t *_sect, void *_udata) /* Set the shrinking type */ udata->shrink = H5MF_SHRINK_EOA; #ifdef H5MF_ALLOC_DEBUG_MORE - HDfprintf(stderr, "%s: section {%a, %Hu}, shrinks file, eoa = %a\n", __func__, sect->sect_info.addr, - sect->sect_info.size, eoa); + HDfprintf(stderr, + "%s: section {%" PRIuHADDR ", %" PRIuHSIZE "}, shrinks file, eoa = %" PRIuHADDR "\n", + __func__, sect->sect_info.addr, sect->sect_info.size, eoa); #endif /* H5MF_ALLOC_DEBUG_MORE */ /* Indicate shrinking can occur */ diff --git a/src/H5MP.c b/src/H5MP.c deleted file mode 100644 index 397d26b..0000000 --- a/src/H5MP.c +++ /dev/null @@ -1,443 +0,0 @@ -/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * - * Copyright by The HDF Group. * - * Copyright by the Board of Trustees of the University of Illinois. * - * All rights reserved. * - * * - * This file is part of HDF5. The full HDF5 copyright notice, including * - * terms governing use, modification, and redistribution, is contained in * - * the COPYING file, which can be found at the root of the source code * - * distribution tree, or in https://www.hdfgroup.org/licenses. * - * If you do not have access to either file, you may request a copy from * - * help@hdfgroup.org. * - * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ - -/*------------------------------------------------------------------------- - * - * Created: H5MP.c - * May 2 2005 - * Quincey Koziol - * - * Purpose: Implements memory pools. (Similar to Apache's APR - * memory pools) - * - * Please see the documentation in: - * doc/html/TechNotes/MemoryPools.html for a full description - * of how they work, etc. - * - *------------------------------------------------------------------------- - */ - -#include "H5MPmodule.h" /* This source code file is part of the H5MP module */ - -/* Private headers */ -#include "H5private.h" /* Generic Functions */ -#include "H5Eprivate.h" /* Error handling */ -#include "H5MMprivate.h" /* Memory management */ -#include "H5MPpkg.h" /* Memory Pools */ - -/****************/ -/* Local Macros */ -/****************/ - -/* Minimum sized block */ -#define H5MP_MIN_BLOCK (H5MP_BLOCK_ALIGN(sizeof(H5MP_page_blk_t)) + H5MP_BLOCK_ALIGNMENT) - -/* First block in page */ -#define H5MP_PAGE_FIRST_BLOCK(p) \ - (H5MP_page_blk_t *)((void *)((unsigned char *)(p) + H5MP_BLOCK_ALIGN(sizeof(H5MP_page_t)))) - -/******************/ -/* Local Typedefs */ -/******************/ - -/********************/ -/* Local Prototypes */ -/********************/ - -/********************************/ -/* Package Variable Definitions */ -/********************************/ - -/********************/ -/* Static Variables */ -/********************/ - -/* Declare a free list to manage the H5MP_pool_t struct */ -H5FL_DEFINE(H5MP_pool_t); - -/*------------------------------------------------------------------------- - * Function: H5MP_create - * - * Purpose: Create a new memory pool - * - * Return: Pointer to the memory pool "header" on success/NULL on failure - * - * Programmer: Quincey Koziol - * May 2 2005 - * - *------------------------------------------------------------------------- - */ -H5MP_pool_t * -H5MP_create(size_t page_size, unsigned flags) -{ - H5MP_pool_t *mp = NULL; /* New memory pool header */ - H5MP_pool_t *ret_value = NULL; /* Return value */ - - FUNC_ENTER_NOAPI(NULL) - - /* Allocate space for the pool header */ - if (NULL == (mp = H5FL_MALLOC(H5MP_pool_t))) - HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for memory pool header") - - /* Assign information */ - mp->page_size = H5MP_BLOCK_ALIGN(page_size); - mp->flags = flags; - - /* Initialize information */ - mp->free_size = 0; - mp->first = NULL; - mp->max_size = mp->page_size - H5MP_BLOCK_ALIGN(sizeof(H5MP_page_t)); - - /* Create factory for pool pages */ - if (NULL == (mp->page_fac = H5FL_fac_init(page_size))) - HGOTO_ERROR(H5E_RESOURCE, H5E_CANTINIT, NULL, "can't create page factory") - - /* Set return value */ - ret_value = mp; - -done: - if (NULL == ret_value && mp) - if (H5MP_close(mp) < 0) - HDONE_ERROR(H5E_RESOURCE, H5E_CANTFREE, NULL, "unable to free memory pool header") - - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5MP_create() */ - -/*------------------------------------------------------------------------- - * Function: H5MP__new_page - * - * Purpose: Allocate new page for a memory pool - * - * Return: Pointer to the page allocated on success/NULL on failure - * - * Programmer: Quincey Koziol - * May 4 2005 - * - *------------------------------------------------------------------------- - */ -static H5MP_page_t * -H5MP__new_page(H5MP_pool_t *mp, size_t page_size) -{ - H5MP_page_t * new_page; /* New page created */ - H5MP_page_blk_t *first_blk; /* Pointer to first block in page */ - H5MP_page_t * ret_value = NULL; /* Return value */ - - FUNC_ENTER_STATIC - - /* Sanity check */ - HDassert(mp); - HDassert(page_size >= mp->page_size); - - /* Allocate page */ - if (page_size > mp->page_size) { - if (NULL == (new_page = (H5MP_page_t *)H5MM_malloc(page_size))) - HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for page") - new_page->free_size = page_size - H5MP_BLOCK_ALIGN(sizeof(H5MP_page_t)); - new_page->fac_alloc = FALSE; - } /* end if */ - else { - if (NULL == (new_page = (H5MP_page_t *)H5FL_FAC_MALLOC(mp->page_fac))) - HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for page") - new_page->free_size = mp->max_size; - new_page->fac_alloc = TRUE; - } /* end else */ - - /* Initialize page information */ - first_blk = H5MP_PAGE_FIRST_BLOCK(new_page); - first_blk->size = new_page->free_size; - first_blk->page = new_page; - first_blk->is_free = TRUE; - first_blk->prev = NULL; - first_blk->next = NULL; - - /* Insert into page list */ - new_page->prev = NULL; - new_page->next = mp->first; - if (mp->first) - mp->first->prev = new_page; - mp->first = new_page; - - /* Account for new free space */ - new_page->free_blk = first_blk; - mp->free_size += new_page->free_size; - - /* Assign return value */ - ret_value = new_page; - -done: - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5MP__new_page() */ - -/*------------------------------------------------------------------------- - * Function: H5MP_malloc - * - * Purpose: Allocate space in a memory pool - * - * Return: Pointer to the space allocated on success/NULL on failure - * - * Programmer: Quincey Koziol - * May 2 2005 - * - *------------------------------------------------------------------------- - */ -void * -H5MP_malloc(H5MP_pool_t *mp, size_t request) -{ - H5MP_page_t * alloc_page = NULL; /* Page to allocate space from */ - H5MP_page_blk_t *alloc_free; /* Pointer to free space in page */ - size_t needed; /* Size requested, plus block header and alignment */ - void * ret_value = NULL; /* Return value */ - - FUNC_ENTER_NOAPI(NULL) - - /* Sanity check */ - HDassert(mp); - HDassert(request > 0); - - /* Compute actual size needed */ - needed = H5MP_BLOCK_ALIGN(request) + H5MP_BLOCK_ALIGN(sizeof(H5MP_page_blk_t)); - - /* See if the request can be handled by existing free space */ - if (needed <= mp->free_size) { - size_t pool_free_avail; /* Amount of free space possibly available in pool */ - - /* Locate page with enough free space */ - alloc_page = mp->first; - pool_free_avail = mp->free_size; - while (alloc_page && pool_free_avail >= needed) { - /* If we found a page with enough free space, search for large - * enough free block on that page */ - if (alloc_page->free_size >= needed) { - size_t page_free_avail; /* Amount of free space possibly available */ - - /* Locate large enough block */ - alloc_free = alloc_page->free_blk; - page_free_avail = alloc_page->free_size; - while (alloc_free && page_free_avail >= needed) { - if (alloc_free->is_free) { - /* If we found a large enough block, leave now */ - if (alloc_free->size >= needed) - goto found; /* Needed to escape double "while" loop */ - - /* Decrement amount of potential space left */ - page_free_avail -= alloc_free->size; - } /* end if */ - - /* Go to next block */ - alloc_free = alloc_free->next; - } /* end while */ - } /* end if */ - - /* Decrement amount of potential space left */ - pool_free_avail -= alloc_page->free_size; - - /* Go to next page */ - alloc_page = alloc_page->next; - } /* end while */ - } /* end if */ - - { - size_t page_size; /* Size of page needed */ - - /* Check if the request is too large for a standard page */ - page_size = - (needed > mp->max_size) ? (needed + H5MP_BLOCK_ALIGN(sizeof(H5MP_page_t))) : mp->page_size; - - /* Allocate new page */ - if (NULL == (alloc_page = H5MP__new_page(mp, page_size))) - HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for page") - - /* Set the block to allocate from */ - alloc_free = alloc_page->free_blk; - } /* end block */ - - /* Allocate space in page */ -found: - - /* Sanity check */ - HDassert(alloc_page); - HDassert(alloc_free); - - /* Check if we can subdivide the free space */ - if (alloc_free->size > (needed + H5MP_MIN_BLOCK)) { - H5MP_page_blk_t *new_free; /* New free block created */ - - /* Carve out new free block after block to allocate */ - new_free = (H5MP_page_blk_t *)((void *)(((unsigned char *)alloc_free) + needed)); - - /* Link into existing lists */ - new_free->next = alloc_free->next; - if (alloc_free->next) - alloc_free->next->prev = new_free; - new_free->prev = alloc_free; - alloc_free->next = new_free; - - /* Set blocks' information */ - new_free->size = alloc_free->size - needed; - new_free->is_free = TRUE; - new_free->page = alloc_free->page; - alloc_free->size = needed; - alloc_free->is_free = FALSE; - } /* end if */ - else { - /* Use whole free space block for new block */ - alloc_free->is_free = FALSE; - } /* end else */ - - /* Update page & pool's free size information */ - alloc_page->free_size -= alloc_free->size; - if (alloc_page->free_blk == alloc_free) - alloc_page->free_blk = alloc_free->next; - mp->free_size -= alloc_free->size; - - /* Set new space pointer for the return value */ - ret_value = ((unsigned char *)alloc_free) + H5MP_BLOCK_ALIGN(sizeof(H5MP_page_blk_t)); - -done: - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5MP_malloc() */ - -/*------------------------------------------------------------------------- - * Function: H5MP_free - * - * Purpose: Release space in a memory pool - * - * Return: NULL on success/NULL on failure - * - * Programmer: Quincey Koziol - * May 3 2005 - * - * Note: Should we release pages that have no used blocks? - * - *------------------------------------------------------------------------- - */ -void * -H5MP_free(H5MP_pool_t *mp, void *spc) -{ - H5MP_page_blk_t *spc_blk; /* Block for space to free */ - H5MP_page_t * spc_page; /* Page containing block to free */ - void * ret_value = NULL; /* Return value */ - - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Sanity check */ - HDassert(mp); - HDassert(spc); - - /* Get block header for space to free */ - spc_blk = - (H5MP_page_blk_t *)((void *)(((unsigned char *)spc) - H5MP_BLOCK_ALIGN(sizeof(H5MP_page_blk_t)))); - - /* Mark block as free */ - HDassert(spc_blk->is_free == FALSE); - spc_blk->is_free = TRUE; - - /* Add it's space to the amount of free space in the page & pool */ - spc_page = spc_blk->page; - spc_page->free_size += spc_blk->size; - mp->free_size += spc_blk->size; - - /* Move page with newly freed space to front of list of pages in pool */ - if (spc_page != mp->first) { - /* Remove page from list */ - spc_page->prev->next = spc_page->next; - if (spc_page->next) - spc_page->next->prev = spc_page->prev; - - /* Insert page at beginning of list */ - spc_page->prev = NULL; - spc_page->next = mp->first; - mp->first->prev = spc_page; - mp->first = spc_page; - } /* end if */ - - /* Check if block can be merged with free space after it on page */ - if (spc_blk->next != NULL) { - H5MP_page_blk_t *next_blk; /* Block following space to free */ - - next_blk = spc_blk->next; - HDassert(next_blk->prev == spc_blk); - if (next_blk->is_free) { - spc_blk->size += next_blk->size; - spc_blk->next = next_blk->next; - } /* end if */ - } /* end if */ - - /* Check if block can be merged with free space before it on page */ - if (spc_blk->prev != NULL) { - H5MP_page_blk_t *prev_blk; /* Block before space to free */ - - prev_blk = spc_blk->prev; - HDassert(prev_blk->next == spc_blk); - if (prev_blk->is_free) { - prev_blk->size += spc_blk->size; - prev_blk->next = spc_blk->next; - } /* end if */ - } /* end if */ - - /* Check if the block freed becomes the first free block on the page */ - if (spc_page->free_blk == NULL || spc_blk < spc_page->free_blk) - spc_page->free_blk = spc_blk; - - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5MP_free() */ - -/*------------------------------------------------------------------------- - * Function: H5MP_close - * - * Purpose: Release all memory for a pool and destroy pool - * - * Return: Non-negative on success/negative on failure - * - * Programmer: Quincey Koziol - * May 3 2005 - * - *------------------------------------------------------------------------- - */ -herr_t -H5MP_close(H5MP_pool_t *mp) -{ - herr_t ret_value = SUCCEED; /* Return value */ - - FUNC_ENTER_NOAPI(FAIL) - - /* Release memory for pool pages */ - if (mp->first != NULL) { - H5MP_page_t *page, *next_page; /* Pointer to pages in pool */ - - /* Iterate through pages, releasing them */ - page = mp->first; - while (page) { - next_page = page->next; - - /* Free the page appropriately */ - if (page->fac_alloc) - page = (H5MP_page_t *)H5FL_FAC_FREE(mp->page_fac, page); - else - page = (H5MP_page_t *)H5MM_xfree(page); - - page = next_page; - } /* end while */ - } /* end if */ - - /* Release page factory */ - if (mp->page_fac) - if (H5FL_fac_term(mp->page_fac) < 0) - HGOTO_ERROR(H5E_RESOURCE, H5E_CANTRELEASE, FAIL, "can't destroy page factory") - -done: - /* Free the memory pool itself */ - mp = H5FL_FREE(H5MP_pool_t, mp); - - FUNC_LEAVE_NOAPI(ret_value) -} /* end H5MP_close() */ diff --git a/src/H5MPmodule.h b/src/H5MPmodule.h deleted file mode 100644 index 8e34598..0000000 --- a/src/H5MPmodule.h +++ /dev/null @@ -1,32 +0,0 @@ -/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * - * Copyright by The HDF Group. * - * All rights reserved. * - * * - * This file is part of HDF5. The full HDF5 copyright notice, including * - * terms governing use, modification, and redistribution, is contained in * - * the COPYING file, which can be found at the root of the source code * - * distribution tree, or in https://www.hdfgroup.org/licenses. * - * If you do not have access to either file, you may request a copy from * - * help@hdfgroup.org. * - * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ - -/* - * Programmer: Quincey Koziol - * Saturday, September 12, 2015 - * - * Purpose: This file contains declarations which define macros for the - * H5MP package. Including this header means that the source file - * is part of the H5MP package. - */ -#ifndef H5MPmodule_H -#define H5MPmodule_H - -/* Define the proper control macros for the generic FUNC_ENTER/LEAVE and error - * reporting macros. - */ -#define H5MP_MODULE -#define H5_MY_PKG H5MP -#define H5_MY_PKG_ERR H5E_RESOURCE -#define H5_MY_PKG_INIT NO - -#endif /* H5MPmodule_H */ diff --git a/src/H5MPpkg.h b/src/H5MPpkg.h deleted file mode 100644 index 64c5293..0000000 --- a/src/H5MPpkg.h +++ /dev/null @@ -1,99 +0,0 @@ -/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * - * Copyright by The HDF Group. * - * Copyright by the Board of Trustees of the University of Illinois. * - * All rights reserved. * - * * - * This file is part of HDF5. The full HDF5 copyright notice, including * - * terms governing use, modification, and redistribution, is contained in * - * the COPYING file, which can be found at the root of the source code * - * distribution tree, or in https://www.hdfgroup.org/licenses. * - * If you do not have access to either file, you may request a copy from * - * help@hdfgroup.org. * - * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ - -/* - * Programmer: Quincey Koziol - * Monday, May 2, 2005 - * - * Purpose: This file contains declarations which are visible only within - * the H5MP package. Source files outside the H5MP package should - * include H5MPprivate.h instead. - */ -#if !(defined H5MP_FRIEND || defined H5MP_MODULE) -#error "Do not include this file outside the H5MP package!" -#endif - -#ifndef H5MPpkg_H -#define H5MPpkg_H - -/* Get package's private header */ -#include "H5MPprivate.h" /* Memory Pools */ - -/* Other private headers needed by this file */ -#include "H5FLprivate.h" /* Free Lists */ - -/**************************/ -/* Package Private Macros */ -/**************************/ - -/* Alignment macros */ -/* (Ideas from Apache APR :-) */ - -/* Default alignment necessary */ -#define H5MP_BLOCK_ALIGNMENT 8 - -/* General alignment macro */ -/* (this only works for aligning to power of 2 boundary) */ -#define H5MP_ALIGN(x, a) (((x) + ((size_t)(a)) - 1) & ~(((size_t)(a)) - 1)) - -/* Default alignment */ -#define H5MP_BLOCK_ALIGN(x) H5MP_ALIGN(x, H5MP_BLOCK_ALIGNMENT) - -/****************************/ -/* Package Private Typedefs */ -/****************************/ - -/* Free block in pool */ -typedef struct H5MP_page_blk_t { - size_t size; /* Size of block (includes this H5MP_page_blk_t info) */ - unsigned is_free : 1; /* Flag to indicate the block is free */ - struct H5MP_page_t * page; /* Pointer to page block is located in */ - struct H5MP_page_blk_t *prev; /* Pointer to previous block in page */ - struct H5MP_page_blk_t *next; /* Pointer to next block in page */ -} H5MP_page_blk_t; - -/* Memory pool page */ -typedef struct H5MP_page_t { - size_t free_size; /* Total amount of free space in page */ - unsigned fac_alloc : 1; /* Flag to indicate the page was allocated by the pool's factory */ - H5MP_page_blk_t * free_blk; /* Pointer to first free block in page */ - struct H5MP_page_t *next; /* Pointer to next page in pool */ - struct H5MP_page_t *prev; /* Pointer to previous page in pool */ -} H5MP_page_t; - -/* Memory pool header */ -struct H5MP_pool_t { - H5FL_fac_head_t *page_fac; /* Free-list factory for pages */ - size_t page_size; /* Page size for pool */ - size_t free_size; /* Total amount of free space in pool */ - size_t max_size; /* Maximum block that will fit in a standard page */ - H5MP_page_t * first; /* Pointer to first page in pool */ - unsigned flags; /* Bit flags for pool settings */ -}; - -/*****************************************/ -/* Package Private Variable Declarations */ -/*****************************************/ - -/******************************/ -/* Package Private Prototypes */ -/******************************/ -#ifdef H5MP_TESTING -H5_DLL herr_t H5MP_get_pool_free_size(const H5MP_pool_t *mp, size_t *free_size); -H5_DLL htri_t H5MP_pool_is_free_size_correct(const H5MP_pool_t *mp); -H5_DLL herr_t H5MP_get_pool_first_page(const H5MP_pool_t *mp, H5MP_page_t **page); -H5_DLL herr_t H5MP_get_page_free_size(const H5MP_page_t *mp, size_t *page); -H5_DLL herr_t H5MP_get_page_next_page(const H5MP_page_t *page, H5MP_page_t **next_page); -#endif /* H5MP_TESTING */ - -#endif /* H5MPpkg_H */ diff --git a/src/H5MPprivate.h b/src/H5MPprivate.h deleted file mode 100644 index 2b06650..0000000 --- a/src/H5MPprivate.h +++ /dev/null @@ -1,57 +0,0 @@ -/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * - * Copyright by The HDF Group. * - * Copyright by the Board of Trustees of the University of Illinois. * - * All rights reserved. * - * * - * This file is part of HDF5. The full HDF5 copyright notice, including * - * terms governing use, modification, and redistribution, is contained in * - * the COPYING file, which can be found at the root of the source code * - * distribution tree, or in https://www.hdfgroup.org/licenses. * - * If you do not have access to either file, you may request a copy from * - * help@hdfgroup.org. * - * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ - -/*------------------------------------------------------------------------- - * - * Created: H5MPprivate.h - * May 2 2005 - * Quincey Koziol - * - * Purpose: Private header for memory pool routines. - * - *------------------------------------------------------------------------- - */ - -#ifndef H5MPprivate_H -#define H5MPprivate_H - -/* Include package's public header (not yet) */ -/* #include "H5MPpublic.h" */ - -/* Private headers needed by this file */ - -/**************************/ -/* Library Private Macros */ -/**************************/ - -/* Pool creation flags */ -/* Default settings */ -#define H5MP_FLG_DEFAULT 0 -#define H5MP_PAGE_SIZE_DEFAULT 4096 /* (bytes) */ - -/****************************/ -/* Library Private Typedefs */ -/****************************/ - -/* Memory pool header (defined in H5MPpkg.c) */ -typedef struct H5MP_pool_t H5MP_pool_t; - -/***************************************/ -/* Library-private Function Prototypes */ -/***************************************/ -H5_DLL H5MP_pool_t *H5MP_create(size_t page_size, unsigned flags); -H5_DLL void * H5MP_malloc(H5MP_pool_t *mp, size_t request); -H5_DLL void * H5MP_free(H5MP_pool_t *mp, void *spc); -H5_DLL herr_t H5MP_close(H5MP_pool_t *mp); - -#endif /* H5MPprivate_H */ diff --git a/src/H5MPtest.c b/src/H5MPtest.c deleted file mode 100644 index 27e7bbe..0000000 --- a/src/H5MPtest.c +++ /dev/null @@ -1,213 +0,0 @@ -/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * - * Copyright by The HDF Group. * - * Copyright by the Board of Trustees of the University of Illinois. * - * All rights reserved. * - * * - * This file is part of HDF5. The full HDF5 copyright notice, including * - * terms governing use, modification, and redistribution, is contained in * - * the COPYING file, which can be found at the root of the source code * - * distribution tree, or in https://www.hdfgroup.org/licenses. * - * If you do not have access to either file, you may request a copy from * - * help@hdfgroup.org. * - * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ - -/* Programmer: Quincey Koziol - * Tuesday, May 3, 2005 - * - * Purpose: Memory pool testing functions. - */ - -#include "H5MPmodule.h" /* This source code file is part of the H5MP module */ -#define H5MP_TESTING /*include H5MP testing funcs*/ - -/* Private headers */ -#include "H5private.h" /* Generic Functions */ -#include "H5MPpkg.h" /* Memory Pools */ -#include "H5Eprivate.h" /* Error handling */ - -/* Static Prototypes */ - -/* Package variables */ - -/*------------------------------------------------------------------------- - * Function: H5MP_get_pool_free_size - * - * Purpose: Retrieve the total amount of free space in entire pool - * - * Return: Success: non-negative - * - * Failure: negative - * - * Programmer: Quincey Koziol - * Tuesday, May 3, 2005 - * - *------------------------------------------------------------------------- - */ -herr_t -H5MP_get_pool_free_size(const H5MP_pool_t *mp, size_t *free_size) -{ - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check arguments. */ - HDassert(mp); - HDassert(free_size); - - /* Get memory pool's free space */ - *free_size = mp->free_size; - - FUNC_LEAVE_NOAPI(SUCCEED) -} /* H5MP_get_pool_free_size() */ - -/*------------------------------------------------------------------------- - * Function: H5MP_get_pool_first_page - * - * Purpose: Retrieve the first page in a memory pool - * - * Return: Success: non-negative - * - * Failure: negative - * - * Programmer: Quincey Koziol - * Tuesday, May 3, 2005 - * - *------------------------------------------------------------------------- - */ -herr_t -H5MP_get_pool_first_page(const H5MP_pool_t *mp, H5MP_page_t **page) -{ - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check arguments. */ - HDassert(mp); - HDassert(page); - - /* Get memory pool's first page */ - *page = mp->first; - - FUNC_LEAVE_NOAPI(SUCCEED) -} /* H5MP_get_pool_first_page() */ - -/*------------------------------------------------------------------------- - * Function: H5MP_pool_is_free_size_correct - * - * Purpose: Check that the free space reported in each page corresponds - * to the free size in each page and that the free space in the - * free blocks for a page corresponds with the free space for - * the page. - * - * Return: Success: non-negative - * - * Failure: negative - * - * Programmer: Quincey Koziol - * Wednesday, May 3, 2005 - * - *------------------------------------------------------------------------- - */ -htri_t -H5MP_pool_is_free_size_correct(const H5MP_pool_t *mp) -{ - H5MP_page_t *page; /* Pointer to current page */ - size_t pool_free; /* Size of pages' free space */ - htri_t ret_value = TRUE; /* Return value */ - - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check arguments. */ - HDassert(mp); - - /* Iterate through pages, checking the free size & accumulating the - * free space for all the pages */ - page = mp->first; - pool_free = 0; - while (page != NULL) { - H5MP_page_blk_t *blk; /* Pointer to current free block */ - size_t page_free; /* Size of blocks on free list */ - - /* Iterate through the blocks in page, accumulating free space */ - blk = (H5MP_page_blk_t *)((void *)((unsigned char *)page + H5MP_BLOCK_ALIGN(sizeof(H5MP_page_t)))); - page_free = 0; - while (blk != NULL) { - if (blk->is_free) - page_free += blk->size; - blk = blk->next; - } /* end while */ - - /* Check that the free space from the blocks on the free list - * corresponds to space in page */ - if (page_free != page->free_size) - HGOTO_DONE(FALSE) - - /* Increment the amount of free space in pool */ - pool_free += page->free_size; - - /* Advance to next page */ - page = page->next; - } /* end while */ - - /* Check that the free space from the pages - * corresponds to free space in pool */ - if (pool_free != mp->free_size) - HGOTO_DONE(FALSE) - -done: - FUNC_LEAVE_NOAPI(ret_value) -} /* H5MP_pool_is_free_size_correct() */ - -/*------------------------------------------------------------------------- - * Function: H5MP_get_page_free_size - * - * Purpose: Retrieve the amount of free space in given page - * - * Return: Success: non-negative - * - * Failure: negative - * - * Programmer: Quincey Koziol - * Tuesday, May 3, 2005 - * - *------------------------------------------------------------------------- - */ -herr_t -H5MP_get_page_free_size(const H5MP_page_t *page, size_t *free_size) -{ - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check arguments. */ - HDassert(page); - HDassert(free_size); - - /* Get memory page's free space */ - *free_size = page->free_size; - - FUNC_LEAVE_NOAPI(SUCCEED) -} /* H5MP_get_page_free_size() */ - -/*------------------------------------------------------------------------- - * Function: H5MP_get_page_next_page - * - * Purpose: Retrieve the next page in the pool - * - * Return: Success: non-negative - * - * Failure: negative - * - * Programmer: Quincey Koziol - * Tuesday, May 3, 2005 - * - *------------------------------------------------------------------------- - */ -herr_t -H5MP_get_page_next_page(const H5MP_page_t *page, H5MP_page_t **next_page) -{ - FUNC_ENTER_NOAPI_NOINIT_NOERR - - /* Check arguments. */ - HDassert(page); - HDassert(next_page); - - /* Get next memory page */ - *next_page = page->next; - - FUNC_LEAVE_NOAPI(SUCCEED) -} /* H5MP_get_page_next_page() */ diff --git a/src/H5Ocache.c b/src/H5Ocache.c index ba47da3..c7586cc 100644 --- a/src/H5Ocache.c +++ b/src/H5Ocache.c @@ -346,7 +346,7 @@ H5O__cache_deserialize(const void *image, size_t len, void *_udata, hbool_t *dir done: /* Release the [possibly partially initialized] object header on errors */ if (!ret_value && oh) - if (H5O__free(oh) < 0) + if (H5O__free(oh, FALSE) < 0) HDONE_ERROR(H5E_OHDR, H5E_CANTRELEASE, NULL, "unable to destroy object header data") FUNC_LEAVE_NOAPI(ret_value) @@ -639,7 +639,7 @@ H5O__cache_free_icr(void *_thing) HDassert(oh->cache_info.type == H5AC_OHDR); /* Destroy object header */ - if (H5O__free(oh) < 0) + if (H5O__free(oh, FALSE) < 0) HGOTO_ERROR(H5E_OHDR, H5E_CANTRELEASE, FAIL, "can't destroy object header") done: @@ -1242,7 +1242,7 @@ H5O__prefix_deserialize(const uint8_t *_image, H5O_cache_ud_t *udata) /* Save the object header for later use in 'deserialize' callback */ udata->oh = oh; - if (H5O__free(saved_oh) < 0) + if (H5O__free(saved_oh, FALSE) < 0) HGOTO_ERROR(H5E_OHDR, H5E_CANTRELEASE, FAIL, "can't destroy object header") udata->free_oh = FALSE; } @@ -1255,7 +1255,7 @@ H5O__prefix_deserialize(const uint8_t *_image, H5O_cache_ud_t *udata) done: /* Release the [possibly partially initialized] object header on errors */ if (ret_value < 0 && oh) - if (H5O__free(oh) < 0) + if (H5O__free(oh, FALSE) < 0) HDONE_ERROR(H5E_OHDR, H5E_CANTRELEASE, FAIL, "unable to destroy object header data") FUNC_LEAVE_NOAPI(ret_value) diff --git a/src/H5Ocopy.c b/src/H5Ocopy.c index 05dfc72..0b0bb55 100644 --- a/src/H5Ocopy.c +++ b/src/H5Ocopy.c @@ -33,7 +33,6 @@ #include "H5Aprivate.h" /* Attributes */ #include "H5CXprivate.h" /* API Contexts */ #include "H5Eprivate.h" /* Error handling */ -#include "H5ESprivate.h" /* Event Sets */ #include "H5FLprivate.h" /* Free lists */ #include "H5Iprivate.h" /* IDs */ #include "H5HGprivate.h" /* Global Heaps */ @@ -772,7 +771,7 @@ done: /* Free destination object header on failure */ if (ret_value < 0) { if (oh_dst && !inserted) { - if (H5O__free(oh_dst) < 0) + if (H5O__free(oh_dst, TRUE) < 0) HDONE_ERROR(H5E_OHDR, H5E_CANTFREE, FAIL, "unable to destroy object header data") if (H5O_loc_reset(oloc_dst) < 0) HDONE_ERROR(H5E_OHDR, H5E_CANTFREE, FAIL, "unable to destroy object header data") diff --git a/src/H5Ocopy_ref.c b/src/H5Ocopy_ref.c index f1f8aaf..1cda3ea 100644 --- a/src/H5Ocopy_ref.c +++ b/src/H5Ocopy_ref.c @@ -288,21 +288,22 @@ H5O__copy_expand_ref_object2(H5O_loc_t *src_oloc, hid_t tid_src, const H5T_t *dt size_t nbytes_src, H5O_loc_t *dst_oloc, H5G_loc_t *dst_root_loc, void *buf_dst, size_t ref_count, H5O_copy_t *cpy_info) { - H5T_t * dt_mem = NULL; /* Memory datatype */ - H5T_t * dt_dst = NULL; /* Destination datatype */ - hid_t tid_mem = H5I_INVALID_HID; /* Datatype ID for memory datatype */ - hid_t tid_dst = H5I_INVALID_HID; /* Datatype ID for memory datatype */ - H5T_path_t *tpath_src_mem = NULL, *tpath_mem_dst = NULL; /* Datatype conversion paths */ - size_t i; /* Local index variable */ - hbool_t reg_tid_src = (tid_src == H5I_INVALID_HID); - hid_t dst_loc_id = H5I_INVALID_HID; - void * conv_buf = NULL; /* Buffer for converting data */ - size_t conv_buf_size = 0; /* Buffer size */ - void * reclaim_buf = NULL; /* Buffer for reclaiming data */ - H5S_t * buf_space = NULL; /* Dataspace describing buffer */ - hsize_t buf_dim[1] = {ref_count}; /* Dimension for buffer */ - size_t token_size = H5F_SIZEOF_ADDR(src_oloc->file); - herr_t ret_value = SUCCEED; + H5T_t * dt_mem = NULL; /* Memory datatype */ + H5T_t * dt_dst = NULL; /* Destination datatype */ + hid_t tid_mem = H5I_INVALID_HID; /* Datatype ID for memory datatype */ + hid_t tid_dst = H5I_INVALID_HID; /* Datatype ID for memory datatype */ + H5T_path_t * tpath_src_mem = NULL, *tpath_mem_dst = NULL; /* Datatype conversion paths */ + size_t i; /* Local index variable */ + hbool_t reg_tid_src = (tid_src == H5I_INVALID_HID); + hid_t dst_loc_id = H5I_INVALID_HID; + void * conv_buf = NULL; /* Buffer for converting data */ + size_t conv_buf_size = 0; /* Buffer size */ + void * reclaim_buf = NULL; /* Buffer for reclaiming data */ + H5S_t * buf_space = NULL; /* Dataspace describing buffer */ + hsize_t buf_dim[1] = {ref_count}; /* Dimension for buffer */ + size_t token_size = H5F_SIZEOF_ADDR(src_oloc->file); + const unsigned char zeros[H5R_REF_BUF_SIZE] = {0}; + herr_t ret_value = SUCCEED; FUNC_ENTER_STATIC @@ -353,29 +354,34 @@ H5O__copy_expand_ref_object2(H5O_loc_t *src_oloc, hid_t tid_src, const H5T_t *dt /* Making equivalent references in the destination file */ for (i = 0; i < ref_count; i++) { - H5R_ref_t * ref_ptr = (H5R_ref_t *)conv_buf; - H5R_ref_priv_t *ref = (H5R_ref_priv_t *)&ref_ptr[i]; - H5O_token_t tmp_token = {0}; - - /* Get src object address */ - if (H5R__get_obj_token(ref, &tmp_token, &token_size) < 0) - HGOTO_ERROR(H5E_OHDR, H5E_CANTGET, FAIL, "unable to get object token") - if (H5VL_native_token_to_addr(src_oloc->file, H5I_FILE, tmp_token, &src_oloc->addr) < 0) - HGOTO_ERROR(H5E_OHDR, H5E_CANTUNSERIALIZE, FAIL, "can't deserialize object token into address") - - /* Attempt to copy object from source to destination file */ - if (H5O__copy_obj_by_ref(src_oloc, dst_oloc, dst_root_loc, cpy_info) < 0) - HGOTO_ERROR(H5E_OHDR, H5E_CANTCOPY, FAIL, "unable to copy object") - - /* Set dst object address */ - if (H5VL_native_addr_to_token(dst_oloc->file, H5I_FILE, dst_oloc->addr, &tmp_token) < 0) - HGOTO_ERROR(H5E_OHDR, H5E_CANTSERIALIZE, FAIL, "can't serialize address into object token") - if (H5R__set_obj_token(ref, (const H5O_token_t *)&tmp_token, token_size) < 0) - HGOTO_ERROR(H5E_OHDR, H5E_CANTSET, FAIL, "unable to set object token") - /* Do not set app_ref since references are released once the copy is done */ - if (H5R__set_loc_id(ref, dst_loc_id, TRUE, FALSE) < 0) - HGOTO_ERROR(H5E_OHDR, H5E_CANTSET, FAIL, "unable to set destination loc id") - } /* end for */ + H5R_ref_t * ref_ptr = (H5R_ref_t *)conv_buf; + H5R_ref_priv_t *ref = (H5R_ref_priv_t *)&ref_ptr[i]; + + /* Check for null reference - only expand reference if it is not null */ + if (HDmemcmp(ref, zeros, H5R_REF_BUF_SIZE)) { + H5O_token_t tmp_token = {0}; + + /* Get src object address */ + if (H5R__get_obj_token(ref, &tmp_token, &token_size) < 0) + HGOTO_ERROR(H5E_OHDR, H5E_CANTGET, FAIL, "unable to get object token") + if (H5VL_native_token_to_addr(src_oloc->file, H5I_FILE, tmp_token, &src_oloc->addr) < 0) + HGOTO_ERROR(H5E_OHDR, H5E_CANTUNSERIALIZE, FAIL, + "can't deserialize object token into address") + + /* Attempt to copy object from source to destination file */ + if (H5O__copy_obj_by_ref(src_oloc, dst_oloc, dst_root_loc, cpy_info) < 0) + HGOTO_ERROR(H5E_OHDR, H5E_CANTCOPY, FAIL, "unable to copy object") + + /* Set dst object address */ + if (H5VL_native_addr_to_token(dst_oloc->file, H5I_FILE, dst_oloc->addr, &tmp_token) < 0) + HGOTO_ERROR(H5E_OHDR, H5E_CANTSERIALIZE, FAIL, "can't serialize address into object token") + if (H5R__set_obj_token(ref, (const H5O_token_t *)&tmp_token, token_size) < 0) + HGOTO_ERROR(H5E_OHDR, H5E_CANTSET, FAIL, "unable to set object token") + /* Do not set app_ref since references are released once the copy is done */ + if (H5R__set_loc_id(ref, dst_loc_id, TRUE, FALSE) < 0) + HGOTO_ERROR(H5E_OHDR, H5E_CANTSET, FAIL, "unable to set destination loc id") + } /* end if */ + } /* end for */ /* Copy into another buffer, to reclaim memory later */ if (NULL == (reclaim_buf = H5FL_BLK_MALLOC(type_conv, conv_buf_size))) diff --git a/src/H5Odtype.c b/src/H5Odtype.c index fa49924..9af79f4 100644 --- a/src/H5Odtype.c +++ b/src/H5Odtype.c @@ -1731,7 +1731,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_NO_CLASS: case H5T_NCLASSES: default: - HDsprintf(buf, "H5T_CLASS_%d", (int)(dt->shared->type)); + HDsnprintf(buf, sizeof(buf), "H5T_CLASS_%d", (int)(dt->shared->type)); s = buf; break; } /* end switch */ @@ -1746,7 +1746,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt HDfprintf(stream, "%*s%-*s %u\n", indent, "", fwidth, "Number of members:", dt->shared->u.compnd.nmembs); for (i = 0; i < dt->shared->u.compnd.nmembs; i++) { - HDsprintf(buf, "Member %u:", i); + HDsnprintf(buf, sizeof(buf), "Member %u:", i); HDfprintf(stream, "%*s%-*s %s\n", indent, "", fwidth, buf, dt->shared->u.compnd.memb[i].name); HDfprintf(stream, "%*s%-*s %lu\n", indent + 3, "", MAX(0, fwidth - 3), "Byte offset:", (unsigned long)(dt->shared->u.compnd.memb[i].offset)); @@ -1759,7 +1759,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt HDfprintf(stream, "%*s%-*s %u\n", indent, "", fwidth, "Number of members:", dt->shared->u.enumer.nmembs); for (i = 0; i < dt->shared->u.enumer.nmembs; i++) { - HDsprintf(buf, "Member %u:", i); + HDsnprintf(buf, sizeof(buf), "Member %u:", i); HDfprintf(stream, "%*s%-*s %s\n", indent, "", fwidth, buf, dt->shared->u.enumer.name[i]); HDfprintf(stream, "%*s%-*s 0x", indent, "", fwidth, "Raw bytes of value:"); for (k = 0; k < dt->shared->parent->shared->size; k++) @@ -1799,13 +1799,14 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_CSET_RESERVED_13: case H5T_CSET_RESERVED_14: case H5T_CSET_RESERVED_15: - HDsprintf(buf, "H5T_CSET_RESERVED_%d", (int)(dt->shared->u.atomic.u.s.cset)); + HDsnprintf(buf, sizeof(buf), "H5T_CSET_RESERVED_%d", (int)(dt->shared->u.atomic.u.s.cset)); s = buf; break; case H5T_CSET_ERROR: default: - HDsprintf(buf, "Unknown character set: %d", (int)(dt->shared->u.atomic.u.s.cset)); + HDsnprintf(buf, sizeof(buf), "Unknown character set: %d", + (int)(dt->shared->u.atomic.u.s.cset)); s = buf; break; } /* end switch */ @@ -1837,13 +1838,14 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_STR_RESERVED_13: case H5T_STR_RESERVED_14: case H5T_STR_RESERVED_15: - HDsprintf(buf, "H5T_STR_RESERVED_%d", (int)(dt->shared->u.atomic.u.s.pad)); + HDsnprintf(buf, sizeof(buf), "H5T_STR_RESERVED_%d", (int)(dt->shared->u.atomic.u.s.pad)); s = buf; break; case H5T_STR_ERROR: default: - HDsprintf(buf, "Unknown string padding: %d", (int)(dt->shared->u.atomic.u.s.pad)); + HDsnprintf(buf, sizeof(buf), "Unknown string padding: %d", + (int)(dt->shared->u.atomic.u.s.pad)); s = buf; break; } /* end switch */ @@ -1862,7 +1864,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_VLEN_BADTYPE: case H5T_VLEN_MAXTYPE: default: - HDsprintf(buf, "H5T_VLEN_%d", dt->shared->u.vlen.type); + HDsnprintf(buf, sizeof(buf), "H5T_VLEN_%d", dt->shared->u.vlen.type); s = buf; break; } /* end switch */ @@ -1880,7 +1882,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_LOC_BADLOC: case H5T_LOC_MAXLOC: default: - HDsprintf(buf, "H5T_LOC_%d", (int)dt->shared->u.vlen.loc); + HDsnprintf(buf, sizeof(buf), "H5T_LOC_%d", (int)dt->shared->u.vlen.loc); s = buf; break; } /* end switch */ @@ -1911,13 +1913,13 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_CSET_RESERVED_13: case H5T_CSET_RESERVED_14: case H5T_CSET_RESERVED_15: - HDsprintf(buf, "H5T_CSET_RESERVED_%d", (int)(dt->shared->u.vlen.cset)); + HDsnprintf(buf, sizeof(buf), "H5T_CSET_RESERVED_%d", (int)(dt->shared->u.vlen.cset)); s = buf; break; case H5T_CSET_ERROR: default: - HDsprintf(buf, "Unknown character set: %d", (int)(dt->shared->u.vlen.cset)); + HDsnprintf(buf, sizeof(buf), "Unknown character set: %d", (int)(dt->shared->u.vlen.cset)); s = buf; break; } /* end switch */ @@ -1949,13 +1951,13 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_STR_RESERVED_13: case H5T_STR_RESERVED_14: case H5T_STR_RESERVED_15: - HDsprintf(buf, "H5T_STR_RESERVED_%d", (int)(dt->shared->u.vlen.pad)); + HDsnprintf(buf, sizeof(buf), "H5T_STR_RESERVED_%d", (int)(dt->shared->u.vlen.pad)); s = buf; break; case H5T_STR_ERROR: default: - HDsprintf(buf, "Unknown string padding: %d", (int)(dt->shared->u.vlen.pad)); + HDsnprintf(buf, sizeof(buf), "Unknown string padding: %d", (int)(dt->shared->u.vlen.pad)); s = buf; break; } /* end switch */ @@ -1995,7 +1997,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_ORDER_ERROR: default: - HDsprintf(buf, "H5T_ORDER_%d", dt->shared->u.atomic.order); + HDsnprintf(buf, sizeof(buf), "H5T_ORDER_%d", dt->shared->u.atomic.order); s = buf; break; } /* end switch */ @@ -2069,9 +2071,9 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_NPAD: default: if (dt->shared->u.atomic.u.f.pad < 0) - HDsprintf(buf, "H5T_PAD_%d", -(dt->shared->u.atomic.u.f.pad)); + HDsnprintf(buf, sizeof(buf), "H5T_PAD_%d", -(dt->shared->u.atomic.u.f.pad)); else - HDsprintf(buf, "bit-%d", dt->shared->u.atomic.u.f.pad); + HDsnprintf(buf, sizeof(buf), "bit-%d", dt->shared->u.atomic.u.f.pad); s = buf; break; } /* end switch */ @@ -2092,7 +2094,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_NORM_ERROR: default: - HDsprintf(buf, "H5T_NORM_%d", (int)(dt->shared->u.atomic.u.f.norm)); + HDsnprintf(buf, sizeof(buf), "H5T_NORM_%d", (int)(dt->shared->u.atomic.u.f.norm)); s = buf; } /* end switch */ HDfprintf(stream, "%*s%-*s %s\n", indent, "", fwidth, "Normalization:", s); @@ -2129,7 +2131,7 @@ H5O__dtype_debug(H5F_t *f, const void *mesg, FILE *stream, int indent, int fwidt case H5T_SGN_ERROR: case H5T_NSGN: default: - HDsprintf(buf, "H5T_SGN_%d", (int)(dt->shared->u.atomic.u.i.sign)); + HDsnprintf(buf, sizeof(buf), "H5T_SGN_%d", (int)(dt->shared->u.atomic.u.i.sign)); s = buf; break; } /* end switch */ diff --git a/src/H5Oint.c b/src/H5Oint.c index ee79b0c..2348790 100644 --- a/src/H5Oint.c +++ b/src/H5Oint.c @@ -289,7 +289,7 @@ H5O_create(H5F_t *f, size_t size_hint, size_t initial_rc, hid_t ocpl_id, H5O_loc HGOTO_ERROR(H5E_OHDR, H5E_BADVALUE, FAIL, "Can't apply object header to file") done: - if ((FAIL == ret_value) && (NULL != oh) && (H5O__free(oh) < 0)) + if ((FAIL == ret_value) && (NULL != oh) && (H5O__free(oh, TRUE) < 0)) HDONE_ERROR(H5E_OHDR, H5E_CANTFREE, FAIL, "can't delete object header") FUNC_LEAVE_NOAPI(ret_value) @@ -353,7 +353,7 @@ H5O_create_ohdr(H5F_t *f, hid_t ocpl_id) ret_value = oh; done: - if ((NULL == ret_value) && (NULL != oh) && (H5O__free(oh) < 0)) + if ((NULL == ret_value) && (NULL != oh) && (H5O__free(oh, TRUE) < 0)) HDONE_ERROR(H5E_OHDR, H5E_CANTFREE, NULL, "can't delete object header") FUNC_LEAVE_NOAPI(ret_value) @@ -3014,7 +3014,7 @@ H5O_get_proxy(const H5O_t *oh) *------------------------------------------------------------------------- */ herr_t -H5O__free(H5O_t *oh) +H5O__free(H5O_t *oh, hbool_t force) { unsigned u; /* Local index variable */ herr_t ret_value = SUCCEED; /* Return value */ @@ -3038,10 +3038,12 @@ H5O__free(H5O_t *oh) for (u = 0; u < oh->nmesgs; u++) { #ifndef NDEBUG /* Verify that message is clean, unless it could have been marked - * dirty by decoding */ + * dirty by decoding, or if this is a forced free (in case of + * failure during creation of the object some messages may be dirty) + */ if (oh->ndecode_dirtied && oh->mesg[u].dirty) oh->ndecode_dirtied--; - else + else if (!force) HDassert(oh->mesg[u].dirty == 0); #endif /* NDEBUG */ diff --git a/src/H5Opkg.h b/src/H5Opkg.h index ebfe636..1fe918d 100644 --- a/src/H5Opkg.h +++ b/src/H5Opkg.h @@ -551,7 +551,7 @@ H5_DLL herr_t H5O__visit(H5G_loc_t *loc, const char *obj_name, H5_index_t idx_ty H5O_iterate2_t op, void *op_data, unsigned fields); H5_DLL herr_t H5O__inc_rc(H5O_t *oh); H5_DLL herr_t H5O__dec_rc(H5O_t *oh); -H5_DLL herr_t H5O__free(H5O_t *oh); +H5_DLL herr_t H5O__free(H5O_t *oh, hbool_t force); /* Object header message routines */ H5_DLL herr_t H5O__msg_alloc(H5F_t *f, H5O_t *oh, const H5O_msg_class_t *type, unsigned *mesg_flags, diff --git a/src/H5Opublic.h b/src/H5Opublic.h index b05a2a8..70f451e 100644 --- a/src/H5Opublic.h +++ b/src/H5Opublic.h @@ -510,7 +510,7 @@ H5_DLL herr_t H5Oget_info3(hid_t loc_id, H5O_info2_t *oinfo, unsigned fields); * location and relative name * * \fgdta_loc_obj_id{loc_id} - * \param[in] name Name of group, relative to \p loc_id + * \param[in] name Name of object, relative to \p loc_id * \param[out] oinfo Buffer in which to return object information * \param[in] fields Flags specifying the fields to include in \p oinfo * \lapl_id @@ -1834,7 +1834,7 @@ H5_DLL herr_t H5Oget_info1(hid_t loc_id, H5O_info1_t *oinfo); * by location and relative name * * \fgdta_loc_obj_id{loc_id} - * \param[in] name Name of group, relative to \p loc_id + * \param[in] name Name of object, relative to \p loc_id * \param[out] oinfo Buffer in which to return object information * \lapl_id * @@ -1960,7 +1960,7 @@ H5_DLL herr_t H5Oget_info2(hid_t loc_id, H5O_info1_t *oinfo, unsigned fields); * by location and relative name * * \fgdta_loc_obj_id{loc_id} - * \param[in] name Name of group, relative to \p loc_id + * \param[in] name Name of object, relative to \p loc_id * \param[out] oinfo Buffer in which to return object information * \param[in] fields Flags specifying the fields to include in \p oinfo * \lapl_id @@ -1303,6 +1303,75 @@ done: } /* end H5PB_write() */ /*------------------------------------------------------------------------- + * Function: H5PB_enabled + * + * Purpose: Check if the page buffer may be enabled for the specified + * file and data access type. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Neil Fortner + * + *------------------------------------------------------------------------- + */ +herr_t +H5PB_enabled(H5F_shared_t *f_sh, H5FD_mem_t type, hbool_t *enabled) +{ + H5PB_t *page_buf; /* Page buffering info for this file */ + hbool_t bypass_pb = FALSE; /* Whether to bypass page buffering */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI_NOERR + + /* Sanity checks */ + HDassert(f_sh); + + /* Get pointer to page buffer info for this file */ + page_buf = f_sh->page_buf; + +#ifdef H5_HAVE_PARALLEL + if (H5F_SHARED_HAS_FEATURE(f_sh, H5FD_FEAT_HAS_MPI)) { +#if 1 + bypass_pb = TRUE; +#else + /* MSC - why this stopped working ? */ + int mpi_size; + + if ((mpi_size = H5F_shared_mpi_get_size(f_sh)) < 0) + HGOTO_ERROR(H5E_PAGEBUF, H5E_CANTGET, FAIL, "can't retrieve MPI communicator size") + if (1 != mpi_size) + bypass_pb = TRUE; +#endif + } /* end if */ +#endif + + /* If page buffering is disabled, or if this is a parallel raw data access, + * bypass page buffering. Note that page buffering may still be disabled for + * large metadata access or large non-parallel raw data access, but this + * function doesn't take I/O size into account so if it returns TRUE the + * page buffer may still be disabled for some I/O. If it returns FALSE it is + * always disabled for this access type. + */ + if (NULL == page_buf || (bypass_pb && H5FD_MEM_DRAW == type)) { + /* Update statistics, since wherever this function is called, if it + * returns FALSE, the calling function performs I/O avoiding the page + * buffer layer */ + if (page_buf) { + HDassert(type == H5FD_MEM_DRAW); + page_buf->bypasses[1]++; + } /* end if */ + + /* Page buffer is disabled, at least for this data access type */ + *enabled = FALSE; + } /* end if */ + else + /* Page buffer may be enabled */ + *enabled = TRUE; + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5PB_enabled() */ + +/*------------------------------------------------------------------------- * Function: H5PB__insert_entry() * * Purpose: This function was created without documentation. diff --git a/src/H5PBprivate.h b/src/H5PBprivate.h index e0197bf..0a255fc 100644 --- a/src/H5PBprivate.h +++ b/src/H5PBprivate.h @@ -23,11 +23,6 @@ #ifndef H5PBprivate_H #define H5PBprivate_H -/* Include package's public header */ -#ifdef NOT_YET -#include "H5PBpublic.h" -#endif /* NOT_YET */ - /* Private headers needed by this header */ #include "H5private.h" /* Generic Functions */ #include "H5Fprivate.h" /* File access */ @@ -91,6 +86,7 @@ H5_DLL herr_t H5PB_update_entry(H5PB_t *page_buf, haddr_t addr, size_t size, con H5_DLL herr_t H5PB_remove_entry(const H5F_shared_t *f_sh, haddr_t addr); H5_DLL herr_t H5PB_read(H5F_shared_t *f_sh, H5FD_mem_t type, haddr_t addr, size_t size, void *buf /*out*/); H5_DLL herr_t H5PB_write(H5F_shared_t *f_sh, H5FD_mem_t type, haddr_t addr, size_t size, const void *buf); +H5_DLL herr_t H5PB_enabled(H5F_shared_t *f_sh, H5FD_mem_t type, hbool_t *enabled); /* Statistics routines */ H5_DLL herr_t H5PB_reset_stats(H5PB_t *page_buf); diff --git a/src/H5PLpath.c b/src/H5PLpath.c index 87ff831..b86fd6e 100644 --- a/src/H5PLpath.c +++ b/src/H5PLpath.c @@ -709,7 +709,7 @@ H5PL__path_table_iterate_process_path(const char *plugin_path, H5PL_iterate_type /* Specify a file mask. *.* = We want everything! - * skip the path if the directory can't be opened */ - HDsprintf(service, "%s\\*.dll", plugin_path); + HDsnprintf(service, sizeof(service), "%s\\*.dll", plugin_path); if ((hFind = FindFirstFileA(service, &fdFile)) == INVALID_HANDLE_VALUE) HGOTO_DONE(H5_ITER_CONT) @@ -934,7 +934,7 @@ H5PL__find_plugin_in_path(const H5PL_search_params_t *search_params, hbool_t *fo *found = FALSE; /* Specify a file mask. *.* = We want everything! */ - HDsprintf(service, "%s\\*.dll", dir); + HDsnprintf(service, sizeof(service), "%s\\*.dll", dir); if ((hFind = FindFirstFileA(service, &fdFile)) == INVALID_HANDLE_VALUE) HGOTO_ERROR(H5E_PLUGIN, H5E_OPENERROR, FAIL, "can't open directory") diff --git a/src/H5Pfapl.c b/src/H5Pfapl.c index 47c17db..bfac42b 100644 --- a/src/H5Pfapl.c +++ b/src/H5Pfapl.c @@ -1185,7 +1185,7 @@ done: * * Purpose: Set the file driver (DRIVER_ID) for a file access * property list (PLIST_ID) and supply an optional - * struct containing the driver-specific properites + * struct containing the driver-specific properties * (DRIVER_INFO). The driver properties will be copied into the * property list and the reference count on the driver will be * incremented, allowing the caller to close the driver ID but @@ -1248,8 +1248,8 @@ H5P_set_driver_by_name(H5P_genplist_t *plist, const char *driver_name, const cha FUNC_ENTER_NOAPI(FAIL) - assert(plist); - assert(driver_name); + HDassert(plist); + HDassert(driver_name); /* Register the driver */ if ((new_driver_id = H5FD_register_driver_by_name(driver_name, app_ref)) < 0) @@ -1336,8 +1336,8 @@ H5P_set_driver_by_value(H5P_genplist_t *plist, H5FD_class_value_t driver_value, FUNC_ENTER_NOAPI(FAIL) - assert(plist); - assert(driver_value >= 0); + HDassert(plist); + HDassert(driver_value >= 0); /* Register the driver */ if ((new_driver_id = H5FD_register_driver_by_value(driver_value, app_ref)) < 0) @@ -5284,15 +5284,14 @@ H5P__decode_coll_md_read_flag_t(const void **_pp, void *_value) * Function: H5Pset_all_coll_metadata_ops * * Purpose: Tell the library whether the metadata read operations will - * be done collectively (1) or not (0). Default is independent. - * With collective mode, the library will optimize access to - * metadata operations on the file. + * be done collectively (1) or not (0). Default is independent. + * With collective mode, the library will optimize access to + * metadata operations on the file. * * Note: This routine accepts file access property lists, link - * access property lists, attribute access property lists, - * dataset access property lists, group access property lists, - * named datatype access property lists, - * and dataset transfer property lists. + * access property lists, attribute access property lists, + * dataset access property lists, group access property lists + * and named datatype access property lists. * * Return: Non-negative on success/Negative on failure * @@ -5312,7 +5311,7 @@ H5Pset_all_coll_metadata_ops(hid_t plist_id, hbool_t is_collective) H5TRACE2("e", "ib", plist_id, is_collective); /* Compare the property list's class against the other class */ - /* (Dataset, group, attribute, and named datype access property lists + /* (Dataset, group, attribute, and named datatype access property lists * are sub-classes of link access property lists -QAK) */ if (TRUE != H5P_isa_class(plist_id, H5P_LINK_ACCESS) && TRUE != H5P_isa_class(plist_id, H5P_FILE_ACCESS)) @@ -5342,10 +5341,9 @@ done: * Purpose: Gets information about collective metadata read mode. * * Note: This routine accepts file access property lists, link - * access property lists, attribute access property lists, - * dataset access property lists, group access property lists, - * named datatype access property lists, - * and dataset transfer property lists. + * access property lists, attribute access property lists, + * dataset access property lists, group access property lists, + * and named datatype access property lists. * * Return: Non-negative on success/Negative on failure * @@ -5363,7 +5361,7 @@ H5Pget_all_coll_metadata_ops(hid_t plist_id, hbool_t *is_collective /*out*/) H5TRACE2("e", "ix", plist_id, is_collective); /* Compare the property list's class against the other class */ - /* (Dataset, group, attribute, and named datype access property lists + /* (Dataset, group, attribute, and named datatype access property lists * are sub-classes of link access property lists -QAK) */ if (TRUE != H5P_isa_class(plist_id, H5P_LINK_ACCESS) && TRUE != H5P_isa_class(plist_id, H5P_FILE_ACCESS)) diff --git a/src/H5Pmodule.h b/src/H5Pmodule.h index 6e92e66..66a9574 100644 --- a/src/H5Pmodule.h +++ b/src/H5Pmodule.h @@ -111,7 +111,8 @@ * * \defgroup GAPL General Access Properties * \ingroup H5P - * \todo Should this be as standalone page? + * The functions in this section can be applied to different kinds of property + * lists. * * \defgroup GCPL Group Creation Properties * \ingroup H5P diff --git a/src/H5Ppublic.h b/src/H5Ppublic.h index 536407c..d0bc2b8 100644 --- a/src/H5Ppublic.h +++ b/src/H5Ppublic.h @@ -116,21 +116,70 @@ extern "C" { /* Define property list class callback function pointer types */ //! <!-- [H5P_cls_create_func_t_snip] --> /** - * \todo Document me! + * \brief Callback function for H5Pcreate_class() + * + * \param[in] prop_id The identifier of the property list class being created + * \param[in] create_data User pointer to any class creation data required + * \return \herr_t + * + * \details This function is called when a new property list of the class + * with which this function was registered is being created. The + * function is called after any registered parent create function is + * called for each property value. + * + * If the create function returns a negative value, the new list is not + * returned to the user and the property list creation routine returns + * an error value. + * + * \since 1.4.0 + * */ typedef herr_t (*H5P_cls_create_func_t)(hid_t prop_id, void *create_data); //! <!-- [H5P_cls_create_func_t_snip] --> //! <!-- [H5P_cls_copy_func_t_snip] --> /** - * \todo Document me! + * \brief Callback function for H5Pcreate_class() + * + * \param[in] new_prop_id The identifier of the property list copy + * \param[in] old_prop_id The identifier of the property list being copied + * \param[in] copy_data User pointer to any copy data required + * \return \herr_t + * + * \details This function is called when an existing property list of this + * class is copied. The copy callback function is called after any + * registered parent copy callback function is called for each property + * value. + * + * If the copy routine returns a negative value, the new list is not + * returned to the user and the property list copy function returns an + * error value. + * + * \since 1.4.0 + * */ typedef herr_t (*H5P_cls_copy_func_t)(hid_t new_prop_id, hid_t old_prop_id, void *copy_data); //! <!-- [H5P_cls_copy_func_t_snip] --> //! <!-- [H5P_cls_close_func_t_snip] --> /** - * \todo Document me! + * \brief Callback function for H5Pcreate_class() + * + * \param[in] prop_id The identifier of the property list class being created + * \param[in] close_data User pointer to any close data required + * \return \herr_t + * + * \details This function is called when a property list of the class + * with which this function was registered is being closed. The + * function is called after any registered parent close function is + * called for each property value. + * + * If the close function returns a negative value, the new list is not + * returned to the user and the property list close routine returns + * an error value. + * + * \since 1.4.0 + * */ typedef herr_t (*H5P_cls_close_func_t)(hid_t prop_id, void *close_data); //! <!-- [H5P_cls_close_func_t_snip] --> @@ -145,8 +194,8 @@ typedef herr_t (*H5P_cls_close_func_t)(hid_t prop_id, void *close_data); * \param[in,out] value The value for the property * \return \herr_t * - * \details The H5P_prp_cb1_t() describes the parameters used by the - * property create,copy and close callback functions. + * \details The H5P_prp_cb1_t() function describes the parameters used by the + * property create, copy and close callback functions. */ typedef herr_t (*H5P_prp_cb1_t)(const char *name, size_t size, void *value); //! <!-- [H5P_prp_cb1_t_snip] --> @@ -161,8 +210,8 @@ typedef herr_t (*H5P_prp_cb1_t)(const char *name, size_t size, void *value); * \param[in] value The value for the property * \return \herr_t * - * \details The H5P_prp_cb2_t() describes the parameters used by the - * property set ,copy and delete callback functions. + * \details The H5P_prp_cb2_t() function describes the parameters used by the + * property set, copy and delete callback functions. */ typedef herr_t (*H5P_prp_cb2_t)(hid_t prop_id, const char *name, size_t size, void *value); //! <!-- [H5P_prp_cb2_t_snip] --> @@ -172,13 +221,28 @@ typedef H5P_prp_cb2_t H5P_prp_set_func_t; typedef H5P_prp_cb2_t H5P_prp_get_func_t; //! <!-- [H5P_prp_encode_func_t_snip] --> /** - * \todo Document me! + * \brief Callback function for encoding property values + * + * \param[in] value The property value to be encoded + * \param[out] buf The encoded property value + * \param[out] size The size of \p buf + * \return \herr_t + * + * \note There is currently no public API which exposes a callback of this type. + * */ typedef herr_t (*H5P_prp_encode_func_t)(const void *value, void **buf, size_t *size); //! <!-- [H5P_prp_encode_func_t_snip] --> //! <!-- [H5P_prp_decode_func_t_snip] --> /** - * \todo Document me! + * \brief Callback function for decoding property values + * + * \param[in] buf A buffer containing an encoded property value + * \param[out] value The decoded property value + * \return \herr_t + * + * \note There is currently no public API which exposes a callback of this type. + * */ typedef herr_t (*H5P_prp_decode_func_t)(const void **buf, void *value); //! <!-- [H5P_prp_decode_func_t_snip] --> @@ -187,7 +251,16 @@ typedef H5P_prp_cb1_t H5P_prp_copy_func_t; //! <!-- [H5P_prp_compare_func_t_snip] --> /** - * \todo Document me! + * \brief Callback function for comparing property values + * + * \param[in] value1 A property value + * \param[in] value2 A property value + * \param[in] size The size of the \p value1 and \p value2 buffers + * \return Returns a positive value if \c value1 is greater than \c value2, a + * negative value if \c value2 is greater than \c value1 and zero if + * \c value1 and \c value2 are equal. + * + * \see H5Pregister(), H5Pinsert() */ typedef int (*H5P_prp_compare_func_t)(const void *value1, const void *value2, size_t size); //! <!-- [H5P_prp_compare_func_t_snip] --> @@ -197,7 +270,19 @@ typedef H5P_prp_cb1_t H5P_prp_close_func_t; /* Define property list iteration function type */ //! <!-- [H5P_iterate_t_snip] --> /** - * \todo Document me! + * \brief Callback function for H5Piterate() + * + * \param[in] id The identifier of a property list or property list class + * \param[in] name The name of the current property + * \param[in,out] iter_data The user context passed to H5Piterate() + * \return \herr_t_iter + * + * \details This function is called for each property encountered when + * iterating over a property list or property list class + * via H5Piterate(). + * + * \since 1.4.0 + * */ typedef herr_t (*H5P_iterate_t)(hid_t id, const char *name, void *iter_data); //! <!-- [H5P_iterate_t_snip] --> @@ -264,15 +349,15 @@ typedef enum H5D_mpio_no_collective_cause_t { H5D_MPIO_DATA_TRANSFORMS = 0x04, /**< Collective I/O was not performed because data transforms needed to be applied */ H5D_MPIO_MPI_OPT_TYPES_ENV_VAR_DISABLED = 0x08, - /**< \todo FIXME! */ + /**< Collective I/O was disabled by environment variable (\Code{HDF5_MPI_OPT_TYPES}) */ H5D_MPIO_NOT_SIMPLE_OR_SCALAR_DATASPACES = 0x10, /**< Collective I/O was not performed because one of the dataspaces was neither simple nor scalar */ H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET = 0x20, /**< Collective I/O was not performed because the dataset was neither contiguous nor chunked */ H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED = 0x40, - /**< \todo FIXME! */ + /**< Collective I/O was not performed because parallel filtered writes are disabled */ H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE = 0x80, - /**< \todo FIXME! */ + /**< Error */ H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE = 0x100 /**< Sentinel */ } H5D_mpio_no_collective_cause_t; @@ -577,77 +662,12 @@ H5_DLL hid_t H5Pcreate(hid_t cls_id); * those existing properties, only add or remove their own class * properties. Property list classes defined and supported in the * HDF5 library distribution are listed and briefly described in - * H5Pcreate(). The \p create routine is called when a new property - * list of this class is being created. The #H5P_cls_create_func_t - * callback function is defined as follows: - * - * \snippet this H5P_cls_create_func_t_snip + * H5Pcreate(). The \p create, \p copy, \p close functions are called + * when a property list of the new class is created, copied, or closed, + * respectively. * - * The parameters to this callback function are defined as follows: - * <table> - * <tr> - * <td>\ref hid_t \c prop_id</td> - * <td>IN: The identifier of the property list being created</td> - * </tr> - * <tr> - * <td>\Code{void * create_data}</td> - * <td>IN: User pointer to any class creation data required</td> - * </tr> - * </table> - * - * The \p create routine is called after any registered - * \p create function is called for each property value. If the - * \p create routine returns a negative value, the new list is not - * returned to the user and the property list creation routine returns - * an error value. - * - * The \p copy routine is called when an existing property - * list of this class is copied. The #H5P_cls_copy_func_t callback - * function is defined as follows: - * \snippet this H5P_cls_copy_func_t_snip - * - * The parameters to this callback function are defined as follows: - * <table> - * <tr> - * <td>\ref hid_t \c prop_id</td> - * <td>IN: The identifier of the property list created by copying</td> - * </tr> - * <tr> - * <td>\Code{void * copy_data}</td> - * <td>IN: User pointer to any class copy data required</td> - * </tr> - * </table> - * - * The \p copy routine is called after any registered \p copy function - * is called for each property value. If the \p copy routine returns a - * negative value, the new list is not returned to the user and the - * property list \p copy routine returns an error value. - * - * The \p close routine is called when a property list of this class - * is being closed. The #H5P_cls_close_func_t callback function is - * defined as follows: - * \snippet this H5P_cls_close_func_t_snip - * - * The parameters to this callback function are defined as follows: - * <table> - * <tr> - * <td>\ref hid_t \c prop_id</td> - * <td>IN: The identifier of the property list being closed</td> - * </tr> - * <tr> - * <td>\Code{void * close_data}</td> - * <td>IN: User pointer to any class close data required</td> - * </tr> - * </table> - * - * The \p close routine is called before any registered \p close - * function is called for each property value. If the \p close routine - * returns a negative value, the property list close routine returns - * an error value but the property list is still closed. - * - * H5Pclose_class() can be used to release the property list class - * identifier returned by this function so that resources leaks will - * not develop. + * H5Pclose_class() must be used to release the property list class + * identifier returned by this function. * * \since 1.4.0 * @@ -1376,35 +1396,12 @@ H5_DLL htri_t H5Pisa_class(hid_t plist_id, hid_t pclass_id); * returned in this case, the iterator cannot be restarted if * one of the calls to its operator returns non-zero. * - * The prototype for the #H5P_iterate_t operator is as follows: - * \snippet this H5P_iterate_t_snip - * - * The operation receives the property list or class + * The operation \p iter_func receives the property list or class * identifier for the object being iterated over, \p id, the * name of the current property within the object, \p name, * and the pointer to the operator data passed in to H5Piterate(), - * \p iter_data. The valid return values from an operator are - * as follows: + * \p iter_data. * - * <table> - * <tr> - * <td>Zero</td> - * <td>Causes the iterator to continue, returning zero when all - * properties have been processed</td> - * </tr> - * <tr> - * <td>Positive</td> - * <td>Causes the iterator to immediately return that positive - * value, indicating short-circuit success. The iterator - * can be restarted at the index of the next property</td> - * </tr> - * <tr> - * <td>Negative</td> - * <td>Causes the iterator to immediately return that value, - * indicating failure. The iterator can be restarted at the - * index of the next property</td> - * </tr> - * </table> * H5Piterate() assumes that the properties in the object * identified by \p id remain unchanged through the iteration. * If the membership changes during the iteration, the function's @@ -1877,9 +1874,6 @@ H5_DLL herr_t H5Pget_attr_phase_change(hid_t plist_id, unsigned *max_compact, un * * \brief Returns information about a filter in a pipeline * - * \todo Signature for H5Pget_filter2 is different in H5Pocpl.c than in - * H5Ppublic.h - * * \ocpl_id{plist_id} * \param[in] idx Sequence number within the filter pipeline of the filter * for which information is sought @@ -4205,17 +4199,14 @@ H5_DLL herr_t H5Pset_alignment(hid_t fapl_id, hsize_t threshold, hsize_t alignme * * \note Note: Raw dataset chunk caching is not currently * supported when using the MPI I/O and MPI POSIX file drivers - * in read/write mode; see H5Pset_fapl_mpio() and - * H5Pset_fapl_mpiposix(), respectively. When using one of these - * file drivers, all calls to H5Dread() and H5Dwrite() will access + * in read/write mode; see H5Pset_fapl_mpio(). When using this + * file driver, all calls to H5Dread() and H5Dwrite() will access * the disk directly, and H5Pset_cache() will have no effect on * performance. * * \note Raw dataset chunk caching is supported when these drivers are * used in read-only mode. * - * \todo Check on H5Pset_fapl_mpio() and H5Pset_fapl_mpiposix(). - * * \version 1.8.0 The use of the \p mdc_nelmts parameter was discontinued. * Metadata cache configuration is managed with * H5Pset_mdc_config() and H5Pget_mdc_config(). @@ -5483,12 +5474,38 @@ H5_DLL herr_t H5Pset_coll_metadata_write(hid_t plist_id, hbool_t is_collective); H5_DLL herr_t H5Pget_coll_metadata_write(hid_t plist_id, hbool_t *is_collective); /** - * \todo Add missing documentation + * \ingroup FAPL + * + * \brief Get the MPI communicator and info + * + * \fapl_id + * \param[out] comm MPI communicator + * \param[out] info MPI info object + * \return \herr_t + * + * \details H5Pget_mpi_params() gets the MPI communicator and info stored in + * the file access property list \p fapl_id. + * + * \todo When was this introduced? + * */ H5_DLL herr_t H5Pget_mpi_params(hid_t fapl_id, MPI_Comm *comm, MPI_Info *info); /** - * \todo Add missing documentation + * \ingroup FAPL + * + * \brief Set the MPI communicator and info + * + * \fapl_id + * \param[in] comm MPI communicator + * \param[in] info MPI info object + * \return \herr_t + * + * \details H5Pset_mpi_params() sets the MPI communicator and info stored in + * the file access property list \p fapl_id. + * + * \todo When was this introduced? + * */ H5_DLL herr_t H5Pset_mpi_params(hid_t fapl_id, MPI_Comm comm, MPI_Info info); #endif /* H5_HAVE_PARALLEL */ @@ -7100,9 +7117,6 @@ H5_DLL herr_t H5Pget_virtual_printf_gap(hid_t dapl_id, hsize_t *gap_size); * * \dapl_id * \param[out] view The flag specifying the view of the virtual dataset. - * Valid values are: - * \li #H5D_VDS_FIRST_MISSING - * \li #H5D_VDS_LAST_AVAILABLE * * \return \herr_t * @@ -7456,11 +7470,7 @@ H5_DLL herr_t H5Pset_virtual_printf_gap(hid_t dapl_id, hsize_t gap_size); * * \dapl_id * \param[in] view Flag specifying the extent of the data to be included - * in the view. Valid values are: - * \li #H5D_VDS_FIRST_MISSING: View includes all data - * before the first missing mapped data - * \li #H5D_VDS_LAST_AVAILABLE View includes all - * available mapped data + * in the view. * * \return \herr_t * @@ -7628,8 +7638,11 @@ H5_DLL herr_t H5Pget_hyper_vector_size(hid_t fapl_id, size_t *size /*out*/); * \details H5Pget_preserve() checks the status of the dataset transfer * property list. * + * \since 1.0.0 + * * \version 1.6.0 The flag parameter was changed from INTEGER to LOGICAL to * better match the C API. (Fortran 90) + * \version 1.8.2 Deprecated. * */ H5_DLL int H5Pget_preserve(hid_t plist_id); @@ -7657,6 +7670,8 @@ H5_DLL int H5Pget_preserve(hid_t plist_id); * * Please refer to the function H5Pset_type_conv_cb() for more details. * + * \since 1.8.0 + * */ H5_DLL herr_t H5Pget_type_conv_cb(hid_t dxpl_id, H5T_conv_except_func_t *op, void **operate_data); /** @@ -7680,6 +7695,8 @@ H5_DLL herr_t H5Pget_type_conv_cb(hid_t dxpl_id, H5T_conv_except_func_t *op, voi * H5Pset_vlen_mem_manager(), returning the parameters set by * that function. * + * \since 1.0.0 + * */ H5_DLL herr_t H5Pget_vlen_mem_manager(hid_t plist_id, H5MM_allocate_t *alloc_func, void **alloc_info, H5MM_free_t *free_func, void **free_info); @@ -7923,8 +7940,9 @@ H5_DLL herr_t H5Pset_hyper_vector_size(hid_t plist_id, size_t size); * I/O pipeline treats the destination datapoints as completely * uninitialized. * - * \todo Add missing version information: introduction, deprecation, etc. - * Why is the declaration not in the deprecated section? + * \since 1.0.0 + * + * \version 1.8.2 Deprecated. * */ H5_DLL herr_t H5Pset_preserve(hid_t plist_id, hbool_t status); @@ -7952,7 +7970,7 @@ H5_DLL herr_t H5Pset_preserve(hid_t plist_id, hbool_t status); * function prototype is as follows: * \snippet H5Tpublic.h H5T_conv_except_func_t_snip * - * \todo Add version information. + * \since 1.8.0 * */ H5_DLL herr_t H5Pset_type_conv_cb(hid_t dxpl_id, H5T_conv_except_func_t op, void *operate_data); @@ -8002,7 +8020,8 @@ H5_DLL herr_t H5Pset_type_conv_cb(hid_t dxpl_id, H5T_conv_except_func_t op, void * set to \c NULL and the \p alloc_info and \p free_info parameters are * ignored. * - * \todo Add version information. + * \since 1.0.0 + * */ H5_DLL herr_t H5Pset_vlen_mem_manager(hid_t plist_id, H5MM_allocate_t alloc_func, void *alloc_info, H5MM_free_t free_func, void *free_info); @@ -350,7 +350,7 @@ done: */ /* Disable warning for "format not a string literal" here -QAK */ /* - * This pragma only needs to surround the sprintf() calls with + * This pragma only needs to surround the snprintf() calls with * format_templ in the code below, but early (4.4.7, at least) gcc only * allows diagnostic pragmas to be toggled outside of functions. */ diff --git a/src/H5Shyper.c b/src/H5Shyper.c index 0c765d7..7284846 100644 --- a/src/H5Shyper.c +++ b/src/H5Shyper.c @@ -296,12 +296,12 @@ H5S__hyper_print_spans_helper(FILE *f, const H5S_hyper_span_t *span, unsigned de FUNC_ENTER_STATIC_NOERR while (span) { - HDfprintf(f, "%s: %*sdepth=%u, span=%p, (%Hu, %Hu), next=%p\n", __func__, depth * 2, "", depth, span, - span->low, span->high, span->next); + HDfprintf(f, "%s: %*sdepth=%u, span=%p, (%" PRIuHSIZE ", %" PRIuHSIZE "), next=%p\n", __func__, + depth * 2, "", depth, (void *)span, span->low, span->high, (void *)span->next); if (span->down) { - HDfprintf(f, "%s: %*sspans=%p, count=%u, bounds[0]={%Hu, %Hu}, head=%p\n", __func__, - (depth + 1) * 2, "", span->down, span->down->count, span->down->low_bounds[0], - span->down->high_bounds[0], span->down->head); + HDfprintf(f, "%s: %*sspans=%p, count=%u, bounds[0]={%" PRIuHSIZE ", %" PRIuHSIZE "}, head=%p\n", + __func__, (depth + 1) * 2, "", (void *)span->down, span->down->count, + span->down->low_bounds[0], span->down->high_bounds[0], (void *)span->down->head); H5S__hyper_print_spans_helper(f, span->down->head, depth + 1); } /* end if */ span = span->next; @@ -316,8 +316,9 @@ H5S__hyper_print_spans(FILE *f, const H5S_hyper_span_info_t *span_lst) FUNC_ENTER_STATIC_NOERR if (span_lst != NULL) { - HDfprintf(f, "%s: spans=%p, count=%u, bounds[0]={%Hu, %Hu}, head=%p\n", __func__, span_lst, - span_lst->count, span_lst->low_bounds[0], span_lst->high_bounds[0], span_lst->head); + HDfprintf(f, "%s: spans=%p, count=%u, bounds[0]={%" PRIuHSIZE ", %" PRIuHSIZE "}, head=%p\n", + __func__, (void *)span_lst, span_lst->count, span_lst->low_bounds[0], + span_lst->high_bounds[0], (void *)span_lst->head); H5S__hyper_print_spans_helper(f, span_lst->head, 0); } /* end if */ @@ -344,16 +345,16 @@ H5S__hyper_print_diminfo_helper(FILE *f, const char *field, unsigned ndims, cons if (dinfo != NULL) { HDfprintf(f, "%s: %s: start=[", __func__, field); for (u = 0; u < ndims; u++) - HDfprintf(f, "%Hd%s", dinfo[u].start, (u < (ndims - 1) ? ", " : "]\n")); + HDfprintf(f, "%" PRIuHSIZE "%s", dinfo[u].start, (u < (ndims - 1) ? ", " : "]\n")); HDfprintf(f, "%s: %s: stride=[", __func__, field); for (u = 0; u < ndims; u++) - HDfprintf(f, "%Hu%s", dinfo[u].stride, (u < (ndims - 1) ? ", " : "]\n")); + HDfprintf(f, "%" PRIuHSIZE "%s", dinfo[u].stride, (u < (ndims - 1) ? ", " : "]\n")); HDfprintf(f, "%s: %s: count=[", __func__, field); for (u = 0; u < ndims; u++) - HDfprintf(f, "%Hu%s", dinfo[u].count, (u < (ndims - 1) ? ", " : "]\n")); + HDfprintf(f, "%" PRIuHSIZE "%s", dinfo[u].count, (u < (ndims - 1) ? ", " : "]\n")); HDfprintf(f, "%s: %s: block=[", __func__, field); for (u = 0; u < ndims; u++) - HDfprintf(f, "%Hu%s", dinfo[u].block, (u < (ndims - 1) ? ", " : "]\n")); + HDfprintf(f, "%" PRIuHSIZE "%s", dinfo[u].block, (u < (ndims - 1) ? ", " : "]\n")); } /* end if */ else HDfprintf(f, "%s: %s==NULL\n", __func__, field); @@ -412,31 +413,31 @@ H5S__hyper_print_spans_dfs(FILE *f, const H5S_hyper_span_info_t *span_lst, unsig for (u = 0; u < depth; u++) HDfprintf(f, "\t"); - HDfprintf(f, "DIM[%u]: ref_count=%u, #elems=%u, head=%p, tail=%p, actual_tail=%p, matched=%t\n", depth, - span_lst->count, num_elems, span_lst->head, span_lst->tail, actual_tail, + HDfprintf(f, "DIM[%u]: ref_count=%u, #elems=%u, head=%p, tail=%p, actual_tail=%p, matched=%d\n", depth, + span_lst->count, num_elems, (void *)span_lst->head, (void *)span_lst->tail, (void *)actual_tail, (span_lst->tail == actual_tail)); for (u = 0; u < depth; u++) HDfprintf(f, "\t"); HDfprintf(f, "low_bounds=["); for (u = 0; u < dims - 1; u++) - HDfprintf(f, "%llu,", span_lst->low_bounds[u]); - HDfprintf(f, "%llu]\n", span_lst->low_bounds[dims - 1]); + HDfprintf(f, "%" PRIuHSIZE ",", span_lst->low_bounds[u]); + HDfprintf(f, "%" PRIuHSIZE "]\n", span_lst->low_bounds[dims - 1]); for (u = 0; u < depth; u++) HDfprintf(f, "\t"); HDfprintf(f, "high_bounds=["); for (u = 0; u < dims - 1; u++) - HDfprintf(f, "%llu,", span_lst->high_bounds[u]); - HDfprintf(f, "%llu]\n", span_lst->high_bounds[dims - 1]); + HDfprintf(f, "%" PRIuHSIZE ",", span_lst->high_bounds[u]); + HDfprintf(f, "%" PRIuHSIZE "]\n", span_lst->high_bounds[dims - 1]); cur_elem = span_lst->head; elem_idx = 0; while (cur_elem) { for (u = 0; u < depth; u++) HDfprintf(f, "\t"); - HDfprintf(f, "ELEM[%u]: ptr=%p, low=%Hu, high=%Hu, down=%p\n", elem_idx++, cur_elem, cur_elem->low, - cur_elem->high, cur_elem->down); + HDfprintf(f, "ELEM[%u]: ptr=%p, low=%" PRIuHSIZE ", high=%" PRIuHSIZE ", down=%p\n", elem_idx++, + (void *)cur_elem, cur_elem->low, cur_elem->high, (void *)cur_elem->down); if (cur_elem->down) H5S__hyper_print_spans_dfs(f, cur_elem->down, depth + 1, dims); cur_elem = cur_elem->next; @@ -473,7 +474,7 @@ H5S__hyper_print_space_dfs(FILE *f, const H5S_t *space) HDassert(hslab); HDfprintf(f, "=======================\n"); - HDfprintf(f, "SPACE: span_lst=%p, #dims=%u, offset_changed=%d\n", hslab->span_lst, dims, + HDfprintf(f, "SPACE: span_lst=%p, #dims=%u, offset_changed=%d\n", (void *)hslab->span_lst, dims, space->select.offset_changed); HDfprintf(f, " offset=["); @@ -484,25 +485,25 @@ H5S__hyper_print_space_dfs(FILE *f, const H5S_t *space) HDfprintf(f, " low_bounds=["); if (space->select.sel_info.hslab->diminfo_valid == H5S_DIMINFO_VALID_YES) { for (u = 0; u < dims - 1; u++) - HDfprintf(f, "%llu,", space->select.sel_info.hslab->diminfo.low_bounds[u]); - HDfprintf(f, "%llu]\n", space->select.sel_info.hslab->diminfo.low_bounds[dims - 1]); + HDfprintf(f, "%" PRIuHSIZE ",", space->select.sel_info.hslab->diminfo.low_bounds[u]); + HDfprintf(f, "%" PRIuHSIZE "]\n", space->select.sel_info.hslab->diminfo.low_bounds[dims - 1]); } /* end if */ else { for (u = 0; u < dims - 1; u++) - HDfprintf(f, "%llu,", space->select.sel_info.hslab->span_lst->low_bounds[u]); - HDfprintf(f, "%llu]\n", space->select.sel_info.hslab->span_lst->low_bounds[dims - 1]); + HDfprintf(f, "%" PRIuHSIZE ",", space->select.sel_info.hslab->span_lst->low_bounds[u]); + HDfprintf(f, "%" PRIuHSIZE "]\n", space->select.sel_info.hslab->span_lst->low_bounds[dims - 1]); } /* end else */ HDfprintf(f, " high_bounds=["); if (space->select.sel_info.hslab->diminfo_valid == H5S_DIMINFO_VALID_YES) { for (u = 0; u < dims - 1; u++) - HDfprintf(f, "%llu,", space->select.sel_info.hslab->diminfo.high_bounds[u]); - HDfprintf(f, "%llu]\n", space->select.sel_info.hslab->diminfo.high_bounds[dims - 1]); + HDfprintf(f, "%" PRIuHSIZE ",", space->select.sel_info.hslab->diminfo.high_bounds[u]); + HDfprintf(f, "%" PRIuHSIZE "]\n", space->select.sel_info.hslab->diminfo.high_bounds[dims - 1]); } /* end if */ else { for (u = 0; u < dims - 1; u++) - HDfprintf(f, "%llu,", space->select.sel_info.hslab->span_lst->high_bounds[u]); - HDfprintf(f, "%llu]\n", space->select.sel_info.hslab->span_lst->high_bounds[dims - 1]); + HDfprintf(f, "%" PRIuHSIZE ",", space->select.sel_info.hslab->span_lst->high_bounds[u]); + HDfprintf(f, "%" PRIuHSIZE "]\n", space->select.sel_info.hslab->span_lst->high_bounds[dims - 1]); } /* end else */ /* Print out diminfo, if it's valid */ diff --git a/src/H5Smpio.c b/src/H5Smpio.c index aec5560..da9b41d 100644 --- a/src/H5Smpio.c +++ b/src/H5Smpio.c @@ -223,7 +223,7 @@ H5S__mpio_create_point_datatype(size_t elmt_size, hsize_t num_points, MPI_Aint * /* Check whether standard or BIGIO processing will be employeed */ if (bigio_count >= num_points) { -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) /* Create an MPI datatype for the whole point selection */ if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed_block((int)num_points, 1, disp, elmt_type, new_type))) @@ -284,7 +284,7 @@ H5S__mpio_create_point_datatype(size_t elmt_size, hsize_t num_points, MPI_Aint * #endif for (i = 0; i < num_big_types; i++) { -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed_block((int)bigio_count, 1, &disp[(hsize_t)i * bigio_count], elmt_type, &inner_types[i]))) @@ -300,7 +300,7 @@ H5S__mpio_create_point_datatype(size_t elmt_size, hsize_t num_points, MPI_Aint * } /* end for*/ if (remaining_points) { -#if MPI_VERSION >= 3 +#if H5_CHECK_MPI_VERSION(3, 0) if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed_block( remaining_points, 1, &disp[(hsize_t)num_big_types * bigio_count], elmt_type, &inner_types[num_big_types]))) diff --git a/src/H5Spoint.c b/src/H5Spoint.c index 240b722..bc667b1 100644 --- a/src/H5Spoint.c +++ b/src/H5Spoint.c @@ -1060,7 +1060,7 @@ H5S__point_get_version_enc_size(const H5S_t *space, uint32_t *version, uint8_t * hsize_t bounds_start[H5S_MAX_RANK]; /* Starting coordinate of bounding box */ hsize_t bounds_end[H5S_MAX_RANK]; /* Opposite coordinate of bounding box */ hsize_t max_size = 0; /* Maximum selection size */ - unsigned u; /* Local index veriable */ + unsigned u; /* Local index variable */ herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_STATIC diff --git a/src/H5Spublic.h b/src/H5Spublic.h index 30ca813..536f290 100644 --- a/src/H5Spublic.h +++ b/src/H5Spublic.h @@ -171,7 +171,7 @@ H5_DLL herr_t H5Sclose(hid_t space_id); * composing the entire current extent). If either \p stride or * \p block is NULL, then it will be set to \p 1. * - * \since 1.12.0 + * \since 1.10.6 * */ H5_DLL hid_t H5Scombine_hyperslab(hid_t space_id, H5S_seloper_t op, const hsize_t start[], @@ -194,7 +194,7 @@ H5_DLL hid_t H5Scombine_hyperslab(hid_t space_id, H5S_seloper_t op, const hsize_ * from \p space1_id is copied for the dataspace extent of the * newly created dataspace. * - * \since 1.12.0 + * \since 1.10.6 * */ H5_DLL hid_t H5Scombine_select(hid_t space1_id, H5S_seloper_t op, hid_t space2_id); @@ -795,7 +795,7 @@ H5_DLL htri_t H5Sis_simple(hid_t space_id); * \p space2_id. The first selection is modified to contain the * result of \p space1_id operated on by \p space2_id. * - * \since 1.12.0 + * \since 1.10.6 * */ H5_DLL herr_t H5Smodify_select(hid_t space1_id, H5S_seloper_t op, hid_t space2_id); @@ -940,7 +940,7 @@ H5_DLL herr_t H5Ssel_iter_reset(hid_t sel_iter_id, hid_t space_id); * * \note This can be useful for VOL developers to implement chunked datasets. * - * \since 1.12.0 + * \since 1.10.6 */ H5_DLL herr_t H5Sselect_adjust(hid_t spaceid, const hssize_t *offset); /** @@ -977,7 +977,7 @@ H5_DLL herr_t H5Sselect_all(hid_t spaceid); * offset) from the source dataspace \p src_id to the destination * dataspace \p dst_id. * - * \since 1.12.0 + * \since 1.10.6 * */ H5_DLL herr_t H5Sselect_copy(hid_t dst_id, hid_t src_id); @@ -1205,7 +1205,7 @@ H5_DLL herr_t H5Sselect_hyperslab(hid_t space_id, H5S_seloper_t op, const hsize_ * \note Assumes that \p start & \p end block bounds are inclusive, so * \p start == \p end value is OK. * - * \since 1.12.0 + * \since 1.10.6 * */ H5_DLL htri_t H5Sselect_intersect_block(hid_t space_id, const hsize_t *start, const hsize_t *end); @@ -1244,7 +1244,7 @@ H5_DLL herr_t H5Sselect_none(hid_t spaceid); * into a third selection.This can be useful for VOL developers to * implement chunked or virtual datasets. * - * \since 1.12.0 + * \since 1.10.6 * */ H5_DLL hid_t H5Sselect_project_intersection(hid_t src_space_id, hid_t dst_space_id, @@ -1265,7 +1265,7 @@ H5_DLL hid_t H5Sselect_project_intersection(hid_t src_space_id, hid_t dst_space_ * This is primarily used for reading the entire selection in * one swoop. * - * \since 1.12.0 + * \since 1.10.6 * */ H5_DLL htri_t H5Sselect_shape_same(hid_t space1_id, hid_t space2_id); @@ -192,10 +192,10 @@ H5TS_tid_destructor(void *_v) return; /* TBD use an atomic CAS */ - HDpthread_mutex_lock(&H5TS_tid_mtx); + pthread_mutex_lock(&H5TS_tid_mtx); tid->next = H5TS_tid_next_free; H5TS_tid_next_free = tid; - HDpthread_mutex_unlock(&H5TS_tid_mtx); + pthread_mutex_unlock(&H5TS_tid_mtx); } /*-------------------------------------------------------------------------- @@ -215,8 +215,8 @@ H5TS_tid_destructor(void *_v) static void H5TS_tid_init(void) { - HDpthread_mutex_init(&H5TS_tid_mtx, NULL); - HDpthread_key_create(&H5TS_tid_key, H5TS_tid_destructor); + pthread_mutex_init(&H5TS_tid_mtx, NULL); + pthread_key_create(&H5TS_tid_key, H5TS_tid_destructor); } /*-------------------------------------------------------------------------- @@ -246,7 +246,7 @@ H5TS_tid_init(void) uint64_t H5TS_thread_id(void) { - H5TS_tid_t *tid = HDpthread_getspecific(H5TS_tid_key); + H5TS_tid_t *tid = pthread_getspecific(H5TS_tid_key); H5TS_tid_t proto_tid; /* An ID is already assigned. */ @@ -260,14 +260,14 @@ H5TS_thread_id(void) * point `tid` at `proto_tid` if we need to allocate some * memory. */ - HDpthread_mutex_lock(&H5TS_tid_mtx); + pthread_mutex_lock(&H5TS_tid_mtx); if ((tid = H5TS_tid_next_free) != NULL) H5TS_tid_next_free = tid->next; else if (H5TS_tid_next_id != UINT64_MAX) { tid = &proto_tid; tid->id = ++H5TS_tid_next_id; } - HDpthread_mutex_unlock(&H5TS_tid_mtx); + pthread_mutex_unlock(&H5TS_tid_mtx); /* If a prototype ID record was established, copy it to the heap. */ if (tid == &proto_tid) @@ -281,7 +281,7 @@ H5TS_thread_id(void) * to it. */ tid->next = NULL; - if (HDpthread_setspecific(H5TS_tid_key, tid) != 0) { + if (pthread_setspecific(H5TS_tid_key, tid) != 0) { H5TS_tid_destructor(tid); return 0; } @@ -323,29 +323,29 @@ H5TS_pthread_first_thread_init(void) #endif /* initialize global API mutex lock */ - HDpthread_mutex_init(&H5_g.init_lock.atomic_lock, NULL); - HDpthread_cond_init(&H5_g.init_lock.cond_var, NULL); + pthread_mutex_init(&H5_g.init_lock.atomic_lock, NULL); + pthread_cond_init(&H5_g.init_lock.cond_var, NULL); H5_g.init_lock.lock_count = 0; - HDpthread_mutex_init(&H5_g.init_lock.atomic_lock2, NULL); + pthread_mutex_init(&H5_g.init_lock.atomic_lock2, NULL); H5_g.init_lock.attempt_lock_count = 0; /* Initialize integer thread identifiers. */ H5TS_tid_init(); /* initialize key for thread-specific error stacks */ - HDpthread_key_create(&H5TS_errstk_key_g, H5TS__key_destructor); + pthread_key_create(&H5TS_errstk_key_g, H5TS__key_destructor); #ifdef H5_HAVE_CODESTACK /* initialize key for thread-specific function stacks */ - HDpthread_key_create(&H5TS_funcstk_key_g, H5TS__key_destructor); + pthread_key_create(&H5TS_funcstk_key_g, H5TS__key_destructor); #endif /* H5_HAVE_CODESTACK */ /* initialize key for thread-specific API contexts */ - HDpthread_key_create(&H5TS_apictx_key_g, H5TS__key_destructor); + pthread_key_create(&H5TS_apictx_key_g, H5TS__key_destructor); /* initialize key for thread cancellability mechanism */ - HDpthread_key_create(&H5TS_cancel_key_s, H5TS__key_destructor); + pthread_key_create(&H5TS_cancel_key_s, H5TS__key_destructor); FUNC_LEAVE_NOAPI_VOID_NAMECHECK_ONLY } /* end H5TS_pthread_first_thread_init() */ @@ -380,13 +380,13 @@ H5TS__mutex_acquire(H5TS_mutex_t *mutex, unsigned int lock_count, hbool_t *acqui *acquired = TRUE; #else /* H5_HAVE_WIN_THREADS */ /* Attempt to acquire the mutex lock */ - if (0 == HDpthread_mutex_lock(&mutex->atomic_lock)) { - pthread_t my_thread_id = HDpthread_self(); + if (0 == pthread_mutex_lock(&mutex->atomic_lock)) { + pthread_t my_thread_id = pthread_self(); /* Check if locked already */ if (mutex->lock_count) { /* Check for this thread already owning the lock */ - if (HDpthread_equal(my_thread_id, mutex->owner_thread)) { + if (pthread_equal(my_thread_id, mutex->owner_thread)) { /* Already owned by self - increment count */ mutex->lock_count += lock_count; *acquired = TRUE; @@ -401,7 +401,7 @@ H5TS__mutex_acquire(H5TS_mutex_t *mutex, unsigned int lock_count, hbool_t *acqui *acquired = TRUE; } /* end else */ - if (0 != HDpthread_mutex_unlock(&mutex->atomic_lock)) + if (0 != pthread_mutex_unlock(&mutex->atomic_lock)) ret_value = -1; } /* end if */ else @@ -463,35 +463,35 @@ herr_t H5TS_mutex_lock(H5TS_mutex_t *mutex) EnterCriticalSection(&mutex->CriticalSection); #else /* H5_HAVE_WIN_THREADS */ /* Acquire the "attempt" lock, increment the attempt lock count, release the lock */ - ret_value = HDpthread_mutex_lock(&mutex->atomic_lock2); + ret_value = pthread_mutex_lock(&mutex->atomic_lock2); if (ret_value) HGOTO_DONE(ret_value); mutex->attempt_lock_count++; - ret_value = HDpthread_mutex_unlock(&mutex->atomic_lock2); + ret_value = pthread_mutex_unlock(&mutex->atomic_lock2); if (ret_value) HGOTO_DONE(ret_value); /* Acquire the library lock */ - ret_value = HDpthread_mutex_lock(&mutex->atomic_lock); + ret_value = pthread_mutex_lock(&mutex->atomic_lock); if (ret_value) HGOTO_DONE(ret_value); /* Check if this thread already owns the lock */ - if (mutex->lock_count && HDpthread_equal(HDpthread_self(), mutex->owner_thread)) + if (mutex->lock_count && pthread_equal(pthread_self(), mutex->owner_thread)) /* already owned by self - increment count */ mutex->lock_count++; else { /* Wait until the lock is released by current owner thread */ while (mutex->lock_count) - HDpthread_cond_wait(&mutex->cond_var, &mutex->atomic_lock); + pthread_cond_wait(&mutex->cond_var, &mutex->atomic_lock); /* After we've received the signal, take ownership of the mutex */ - mutex->owner_thread = HDpthread_self(); + mutex->owner_thread = pthread_self(); mutex->lock_count = 1; } /* end else */ /* Release the library lock */ - ret_value = HDpthread_mutex_unlock(&mutex->atomic_lock); + ret_value = pthread_mutex_unlock(&mutex->atomic_lock); done: #endif /* H5_HAVE_WIN_THREADS */ @@ -530,12 +530,12 @@ H5TS__mutex_unlock(H5TS_mutex_t *mutex, unsigned int *lock_count) #else /* H5_HAVE_WIN_THREADS */ /* Reset the lock count for this thread */ - ret_value = HDpthread_mutex_lock(&mutex->atomic_lock); + ret_value = pthread_mutex_lock(&mutex->atomic_lock); if (ret_value) HGOTO_DONE(ret_value); *lock_count = mutex->lock_count; mutex->lock_count = 0; - ret_value = HDpthread_mutex_unlock(&mutex->atomic_lock); + ret_value = pthread_mutex_unlock(&mutex->atomic_lock); /* If the lock count drops to zero, signal the condition variable, to * wake another thread. @@ -543,7 +543,7 @@ H5TS__mutex_unlock(H5TS_mutex_t *mutex, unsigned int *lock_count) if (mutex->lock_count == 0) { int err; - err = HDpthread_cond_signal(&mutex->cond_var); + err = pthread_cond_signal(&mutex->cond_var); if (err != 0) ret_value = err; } /* end if */ @@ -586,11 +586,11 @@ H5TS_mutex_unlock(H5TS_mutex_t *mutex) #else /* H5_HAVE_WIN_THREADS */ /* Decrement the lock count for this thread */ - ret_value = HDpthread_mutex_lock(&mutex->atomic_lock); + ret_value = pthread_mutex_lock(&mutex->atomic_lock); if (ret_value) HGOTO_DONE(ret_value); mutex->lock_count--; - ret_value = HDpthread_mutex_unlock(&mutex->atomic_lock); + ret_value = pthread_mutex_unlock(&mutex->atomic_lock); /* If the lock count drops to zero, signal the condition variable, to * wake another thread. @@ -598,7 +598,7 @@ H5TS_mutex_unlock(H5TS_mutex_t *mutex) if (mutex->lock_count == 0) { int err; - err = HDpthread_cond_signal(&mutex->cond_var); + err = pthread_cond_signal(&mutex->cond_var); if (err != 0) ret_value = err; } /* end if */ @@ -630,13 +630,13 @@ H5TSmutex_get_attempt_count(unsigned int *count) #ifdef H5_HAVE_WIN_THREADS /* Add Win32 equivalent here when async is supported */ #else /* H5_HAVE_WIN_THREADS */ - ret_value = HDpthread_mutex_lock(&H5_g.init_lock.atomic_lock2); + ret_value = pthread_mutex_lock(&H5_g.init_lock.atomic_lock2); if (ret_value) HGOTO_DONE(ret_value); *count = H5_g.init_lock.attempt_lock_count; - ret_value = HDpthread_mutex_unlock(&H5_g.init_lock.atomic_lock2); + ret_value = pthread_mutex_unlock(&H5_g.init_lock.atomic_lock2); if (ret_value) HGOTO_DONE(ret_value); @@ -725,7 +725,7 @@ H5TS_cancel_count_inc(void) HGOTO_DONE(FAIL); /* Set the thread's cancellation counter with the new object */ - ret_value = HDpthread_setspecific(H5TS_cancel_key_s, (void *)cancel_counter); + ret_value = pthread_setspecific(H5TS_cancel_key_s, (void *)cancel_counter); if (ret_value) { HDfree(cancel_counter); HGOTO_DONE(FAIL); @@ -735,7 +735,7 @@ H5TS_cancel_count_inc(void) /* Check if thread entering library */ if (cancel_counter->cancel_count == 0) /* Set cancellation state to 'disable', and remember previous state */ - ret_value = HDpthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cancel_counter->previous_state); + ret_value = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cancel_counter->previous_state); /* Increment # of times the library API was re-entered, to avoid resetting * previous cancellation state until the final API routine is returning. @@ -788,7 +788,7 @@ H5TS_cancel_count_dec(void) /* Check for leaving last API routine */ if (cancel_counter->cancel_count == 1) /* Reset to previous thread cancellation state, if last API */ - ret_value = HDpthread_setcancelstate(cancel_counter->previous_state, NULL); + ret_value = pthread_setcancelstate(cancel_counter->previous_state, NULL); /* Decrement cancellation counter */ --cancel_counter->cancel_count; @@ -995,7 +995,7 @@ H5TS_create_thread(H5TS_thread_cb_t func, H5TS_attr_t *attr, void *udata) #else /* H5_HAVE_WIN_THREADS */ - HDpthread_create(&ret_value, attr, (void *(*)(void *))func, udata); + pthread_create(&ret_value, attr, (void *(*)(void *))func, udata); #endif /* H5_HAVE_WIN_THREADS */ diff --git a/src/H5TSprivate.h b/src/H5TSprivate.h index 3150f59..5d69854 100644 --- a/src/H5TSprivate.h +++ b/src/H5TSprivate.h @@ -13,11 +13,9 @@ /*------------------------------------------------------------------------- * - * Created: H5TSprivate.h - * May 2 2000 - * Chee Wai LEE + * Created: H5TSprivate.h * - * Purpose: Private non-prototype header. + * Purpose: Thread-safety abstractions used by the library * *------------------------------------------------------------------------- */ diff --git a/src/H5Tcommit.c b/src/H5Tcommit.c index e99494a..a3a1aa0 100644 --- a/src/H5Tcommit.c +++ b/src/H5Tcommit.c @@ -1127,7 +1127,7 @@ H5T_open(const H5G_loc_t *loc) done: if (ret_value == NULL) { if (dt) { - if (shared_fo == NULL) { /* Need to free shared of */ + if (shared_fo == NULL) { /* Need to free shared file object */ if (dt->shared->owned_vol_obj && H5VL_free_object(dt->shared->owned_vol_obj) < 0) HDONE_ERROR(H5E_DATATYPE, H5E_CANTCLOSEOBJ, NULL, "unable to close owned VOL object") dt->shared = H5FL_FREE(H5T_shared_t, dt->shared); diff --git a/src/H5Tconv.c b/src/H5Tconv.c index 94697af..5efff10 100644 --- a/src/H5Tconv.c +++ b/src/H5Tconv.c @@ -2753,7 +2753,7 @@ H5T__conv_enum_init(H5T_t *src, H5T_t *dst, H5T_cdata_t *cdata) HDassert(domain[1] >= domain[0]); length = (unsigned)(domain[1] - domain[0]) + 1; if (src->shared->u.enumer.nmembs < 2 || - (double)length / src->shared->u.enumer.nmembs < (double)(1.2f)) { + (double)length / src->shared->u.enumer.nmembs < (double)(1.2F)) { priv->base = domain[0]; priv->length = length; if (NULL == (map = (int *)H5MM_malloc(length * sizeof(int)))) diff --git a/src/H5Tnative.c b/src/H5Tnative.c index 4529e57..e6fab51 100644 --- a/src/H5Tnative.c +++ b/src/H5Tnative.c @@ -43,7 +43,7 @@ static herr_t H5T__cmp_offset(size_t *comp_size, size_t *offset, size_t elem_siz * * Purpose: High-level API to return the native type of a datatype. * The native type is chosen by matching the size and class of - * querried datatype from the following native premitive + * queried datatype from the following native primitive * datatypes: * H5T_NATIVE_CHAR H5T_NATIVE_UCHAR * H5T_NATIVE_SHORT H5T_NATIVE_USHORT @@ -56,7 +56,7 @@ static herr_t H5T__cmp_offset(size_t *comp_size, size_t *offset, size_t elem_siz * H5T_NATIVE_LDOUBLE * * Compound, array, enum, and VL types all choose among these - * types for their members. Time, Bifield, Opaque, Reference + * types for their members. Time, Bitfield, Opaque, Reference * types are only copy out. * * Return: Success: Returns the native data type if successful. @@ -696,7 +696,7 @@ H5_GCC_DIAG_OFF("duplicated-branches") /*------------------------------------------------------------------------- * Function: H5T__get_native_float * - * Purpose: Returns the native floatt type of a datatype. + * Purpose: Returns the native float type of a datatype. * * Return: Success: Returns the native data type if successful. * diff --git a/src/H5Tprivate.h b/src/H5Tprivate.h index 6624096..9731379 100644 --- a/src/H5Tprivate.h +++ b/src/H5Tprivate.h @@ -101,6 +101,8 @@ typedef struct H5T_subset_info_t { } H5T_subset_info_t; /* Forward declarations for prototype arguments */ +struct H5G_loc_t; +struct H5G_name_t; struct H5O_shared_t; /* The native endianness of the platform */ @@ -120,14 +122,14 @@ H5_DLL size_t H5T_get_size(const H5T_t *dt); H5_DLL hbool_t H5T_get_force_conv(const H5T_t *dt); H5_DLL int H5T_cmp(const H5T_t *dt1, const H5T_t *dt2, hbool_t superset); H5_DLL herr_t H5T_encode(H5T_t *obj, unsigned char *buf, size_t *nalloc); -H5_DLL H5T_t * H5T_decode(size_t buf_size, const unsigned char *buf); -H5_DLL herr_t H5T_debug(const H5T_t *dt, FILE *stream); -H5_DLL struct H5O_loc_t *H5T_oloc(H5T_t *dt); -H5_DLL H5G_name_t *H5T_nameof(const H5T_t *dt); -H5_DLL htri_t H5T_is_immutable(const H5T_t *dt); -H5_DLL htri_t H5T_is_named(const H5T_t *dt); -H5_DLL herr_t H5T_convert_committed_datatype(H5T_t *dt, H5F_t *f); -H5_DLL htri_t H5T_is_relocatable(const H5T_t *dt); +H5_DLL H5T_t * H5T_decode(size_t buf_size, const unsigned char *buf); +H5_DLL herr_t H5T_debug(const H5T_t *dt, FILE *stream); +H5_DLL struct H5O_loc_t * H5T_oloc(H5T_t *dt); +H5_DLL struct H5G_name_t *H5T_nameof(const H5T_t *dt); +H5_DLL htri_t H5T_is_immutable(const H5T_t *dt); +H5_DLL htri_t H5T_is_named(const H5T_t *dt); +H5_DLL herr_t H5T_convert_committed_datatype(H5T_t *dt, H5F_t *f); +H5_DLL htri_t H5T_is_relocatable(const H5T_t *dt); H5_DLL H5T_path_t *H5T_path_find(const H5T_t *src, const H5T_t *dst); H5_DLL hbool_t H5T_path_noop(const H5T_path_t *p); H5_DLL H5T_bkg_t H5T_path_bkg(const H5T_path_t *p); @@ -159,7 +161,7 @@ H5_DLL herr_t H5T_invoke_vol_optional(H5T_t *dt, H5VL_optional_args_t *args, hi H5_DLL H5R_type_t H5T_get_ref_type(const H5T_t *dt); /* Operations on named datatypes */ -H5_DLL H5T_t *H5T_open(const H5G_loc_t *loc); +H5_DLL H5T_t *H5T_open(const struct H5G_loc_t *loc); H5_DLL int H5T_link(const H5T_t *type, int adjust); H5_DLL herr_t H5T_update_shared(H5T_t *type); diff --git a/src/H5VLcallback.c b/src/H5VLcallback.c index 4cf4d53..0c5c73d 100644 --- a/src/H5VLcallback.c +++ b/src/H5VLcallback.c @@ -30,7 +30,7 @@ #include "H5private.h" /* Generic Functions */ #include "H5Eprivate.h" /* Error handling */ #include "H5ESprivate.h" /* Event Sets */ -#include "H5Fprivate.h" /* File access */ +#include "H5Fprivate.h" /* File access */ #include "H5Iprivate.h" /* IDs */ #include "H5MMprivate.h" /* Memory management */ #include "H5Pprivate.h" /* Property lists */ diff --git a/src/H5VLnative.h b/src/H5VLnative.h index fe8ede2..5e43c4e 100644 --- a/src/H5VLnative.h +++ b/src/H5VLnative.h @@ -74,7 +74,7 @@ typedef union H5VL_native_attr_optional_args_t { #define H5VL_NATIVE_DATASET_CHUNK_WRITE 7 /* H5Dchunk_write */ #define H5VL_NATIVE_DATASET_GET_VLEN_BUF_SIZE 8 /* H5Dvlen_get_buf_size */ #define H5VL_NATIVE_DATASET_GET_OFFSET 9 /* H5Dget_offset */ -#define H5VL_NATIVE_DATASET_CHUNK_ITER 10 /* H5Dget_offset */ +#define H5VL_NATIVE_DATASET_CHUNK_ITER 10 /* H5Dchunk_iter */ /* NOTE: If values over 1023 are added, the H5VL_RESERVED_NATIVE_OPTIONAL macro * must be updated. */ @@ -208,8 +208,8 @@ typedef union H5VL_native_dataset_optional_args_t { #ifdef H5_HAVE_PARALLEL #define H5VL_NATIVE_FILE_GET_MPI_ATOMICITY 26 /* H5Fget_mpi_atomicity */ #define H5VL_NATIVE_FILE_SET_MPI_ATOMICITY 27 /* H5Fset_mpi_atomicity */ -#endif /* H5_HAVE_PARALLEL */ -#define H5VL_NATIVE_FILE_POST_OPEN 28 /* Adjust file after open, with wrapping context */ +#endif +#define H5VL_NATIVE_FILE_POST_OPEN 28 /* Adjust file after open, with wrapping context */ /* NOTE: If values over 1023 are added, the H5VL_RESERVED_NATIVE_OPTIONAL macro * must be updated. */ diff --git a/src/H5VLnative_token.c b/src/H5VLnative_token.c index bed0164..b5bd7b8 100644 --- a/src/H5VLnative_token.c +++ b/src/H5VLnative_token.c @@ -112,7 +112,7 @@ H5VL__native_token_to_str(void *obj, H5I_type_t obj_type, const H5O_token_t *tok if (NULL == (*token_str = H5MM_malloc(addr_ndigits + 1))) HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "can't allocate buffer for token string") - HDsnprintf(*token_str, addr_ndigits + 1, H5_PRINTF_HADDR_FMT, addr); + HDsnprintf(*token_str, addr_ndigits + 1, "%" PRIuHADDR, addr); done: FUNC_LEAVE_NOAPI(ret_value) @@ -139,7 +139,7 @@ H5VL__native_str_to_token(void *obj, H5I_type_t obj_type, const char *token_str, /* Check parameters */ HDassert(token_str); - HDsscanf(token_str, H5_PRINTF_HADDR_FMT, &addr); + HDsscanf(token_str, "%" PRIuHADDR, &addr); if (H5VL_native_addr_to_token(obj, obj_type, addr, token) < 0) HGOTO_ERROR(H5E_FILE, H5E_CANTDECODE, FAIL, "can't convert address to object token") diff --git a/src/H5VLpassthru.c b/src/H5VLpassthru.c index 6b4c59f..6eda875 100644 --- a/src/H5VLpassthru.c +++ b/src/H5VLpassthru.c @@ -641,16 +641,13 @@ H5VL_pass_through_info_to_str(const void *_info, char **str) under_vol_str_len = strlen(under_vol_string); /* Allocate space for our info */ - *str = (char *)H5allocate_memory(32 + under_vol_str_len, (hbool_t)0); + size_t strSize = 32 + under_vol_str_len; + *str = (char *)H5allocate_memory(strSize, (hbool_t)0); assert(*str); - /* Encode our info - * Normally we'd use snprintf() here for a little extra safety, but that - * call had problems on Windows until recently. So, to be as platform-independent - * as we can, we're using sprintf() instead. - */ - sprintf(*str, "under_vol=%u;under_info={%s}", (unsigned)under_value, - (under_vol_string ? under_vol_string : "")); + /* Encode our info */ + snprintf(*str, strSize, "under_vol=%u;under_info={%s}", (unsigned)under_value, + (under_vol_string ? under_vol_string : "")); return 0; } /* end H5VL_pass_through_info_to_str() */ diff --git a/src/H5VMprivate.h b/src/H5VMprivate.h index 0d3bd0f..e773bae 100644 --- a/src/H5VMprivate.h +++ b/src/H5VMprivate.h @@ -393,8 +393,8 @@ static const unsigned char LogTable256[] = { static inline unsigned H5_ATTR_UNUSED H5VM_log2_gen(uint64_t n) { - unsigned r; /* r will be log2(n) */ - register unsigned int t, tt, ttt; /* temporaries */ + unsigned r; /* r will be log2(n) */ + unsigned int t, tt, ttt; /* temporaries */ if ((ttt = (unsigned)(n >> 32))) if ((tt = (unsigned)(n >> 48))) @@ -594,14 +594,9 @@ H5Z__flush_file_cb(void *obj_ptr, hid_t H5_ATTR_UNUSED obj_id, void H5_ATTR_PARA /* Do a global flush if the file is opened for write */ if (H5F_ACC_RDWR & H5F_INTENT(f)) { -/* When parallel HDF5 is defined, check for collective metadata reads on this - * file and set the flag for metadata I/O in the API context. -QAK, 2018/02/14 - */ #ifdef H5_HAVE_PARALLEL /* Check if MPIO driver is used */ if (H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) { - H5P_coll_md_read_flag_t coll_md_read; /* Do all metadata reads collectively */ - /* Sanity check for collectively calling H5Zunregister, if requested */ /* (Sanity check assumes that a barrier on one file's comm * is sufficient (i.e. that there aren't different comms for @@ -621,13 +616,8 @@ H5Z__flush_file_cb(void *obj_ptr, hid_t H5_ATTR_UNUSED obj_id, void H5_ATTR_PARA /* Set the "sanity checked" flag */ object->sanity_checked = TRUE; } /* end if */ - - /* Check whether to use the collective metadata read DXPL */ - coll_md_read = H5F_COLL_MD_READ(f); - if (H5P_USER_TRUE == coll_md_read) - H5CX_set_coll_metadata_read(TRUE); - } /* end if */ -#endif /* H5_HAVE_PARALLEL */ + } /* end if */ +#endif /* H5_HAVE_PARALLEL */ /* Call the flush routine for mounted file hierarchies */ if (H5F_flush_mounts((H5F_t *)obj_ptr) < 0) diff --git a/src/H5Znbit.c b/src/H5Znbit.c index 21363bc..905d417 100644 --- a/src/H5Znbit.c +++ b/src/H5Znbit.c @@ -975,7 +975,7 @@ H5Z__filter_nbit(unsigned flags, size_t cd_nelmts, const unsigned cd_values[], s /* input; decompress */ if (flags & H5Z_FLAG_REVERSE) { - size_out = d_nelmts * cd_values[4]; /* cd_values[4] stores datatype size */ + size_out = d_nelmts * (size_t)cd_values[4]; /* cd_values[4] stores datatype size */ /* allocate memory space for decompressed buffer */ if (NULL == (outbuf = (unsigned char *)H5MM_malloc(size_out))) @@ -1170,7 +1170,8 @@ H5Z__nbit_decompress_one_array(unsigned char *data, size_t data_offset, unsigned n = total_size / p.size; for (i = 0; i < n; i++) - H5Z__nbit_decompress_one_atomic(data, data_offset + i * p.size, buffer, j, buf_len, &p); + H5Z__nbit_decompress_one_atomic(data, data_offset + i * (size_t)p.size, buffer, j, buf_len, + &p); break; case H5Z_NBIT_ARRAY: @@ -1178,8 +1179,8 @@ H5Z__nbit_decompress_one_array(unsigned char *data, size_t data_offset, unsigned n = total_size / base_size; /* number of base_type elements inside the array datatype */ begin_index = *parms_index; for (i = 0; i < n; i++) { - if (H5Z__nbit_decompress_one_array(data, data_offset + i * base_size, buffer, j, buf_len, - parms, parms_index) < 0) + if (H5Z__nbit_decompress_one_array(data, data_offset + i * (size_t)base_size, buffer, j, + buf_len, parms, parms_index) < 0) HGOTO_ERROR(H5E_PLINE, H5E_CANTFILTER, FAIL, "can't decompress array") *parms_index = begin_index; } @@ -1190,8 +1191,8 @@ H5Z__nbit_decompress_one_array(unsigned char *data, size_t data_offset, unsigned n = total_size / base_size; /* number of base_type elements inside the array datatype */ begin_index = *parms_index; for (i = 0; i < n; i++) { - if (H5Z__nbit_decompress_one_compound(data, data_offset + i * base_size, buffer, j, buf_len, - parms, parms_index) < 0) + if (H5Z__nbit_decompress_one_compound(data, data_offset + i * (size_t)base_size, buffer, j, + buf_len, parms, parms_index) < 0) HGOTO_ERROR(H5E_PLINE, H5E_CANTFILTER, FAIL, "can't decompress compound") *parms_index = begin_index; } @@ -1291,7 +1292,7 @@ H5Z__nbit_decompress(unsigned char *data, unsigned d_nelmts, unsigned char *buff FUNC_ENTER_STATIC /* may not have to initialize to zeros */ - HDmemset(data, 0, d_nelmts * parms[4]); + HDmemset(data, 0, d_nelmts * (size_t)parms[4]); /* initialization before the loop */ j = 0; @@ -1309,7 +1310,7 @@ H5Z__nbit_decompress(unsigned char *data, unsigned d_nelmts, unsigned char *buff HGOTO_ERROR(H5E_PLINE, H5E_BADTYPE, FAIL, "invalid datatype precision/offset") for (i = 0; i < d_nelmts; i++) - H5Z__nbit_decompress_one_atomic(data, i * p.size, buffer, &j, &buf_len, &p); + H5Z__nbit_decompress_one_atomic(data, i * (size_t)p.size, buffer, &j, &buf_len, &p); break; case H5Z_NBIT_ARRAY: @@ -1468,7 +1469,7 @@ H5Z__nbit_compress_one_array(unsigned char *data, size_t data_offset, unsigned c p.offset = parms[(*parms_index)++]; n = total_size / p.size; for (i = 0; i < n; i++) - H5Z__nbit_compress_one_atomic(data, data_offset + i * p.size, buffer, j, buf_len, &p); + H5Z__nbit_compress_one_atomic(data, data_offset + i * (size_t)p.size, buffer, j, buf_len, &p); break; case H5Z_NBIT_ARRAY: @@ -1476,8 +1477,8 @@ H5Z__nbit_compress_one_array(unsigned char *data, size_t data_offset, unsigned c n = total_size / base_size; /* number of base_type elements inside the array datatype */ begin_index = *parms_index; for (i = 0; i < n; i++) { - H5Z__nbit_compress_one_array(data, data_offset + i * base_size, buffer, j, buf_len, parms, - parms_index); + H5Z__nbit_compress_one_array(data, data_offset + i * (size_t)base_size, buffer, j, buf_len, + parms, parms_index); *parms_index = begin_index; } break; @@ -1487,8 +1488,8 @@ H5Z__nbit_compress_one_array(unsigned char *data, size_t data_offset, unsigned c n = total_size / base_size; /* number of base_type elements inside the array datatype */ begin_index = *parms_index; for (i = 0; i < n; i++) { - H5Z__nbit_compress_one_compound(data, data_offset + i * base_size, buffer, j, buf_len, parms, - parms_index); + H5Z__nbit_compress_one_compound(data, data_offset + i * (size_t)base_size, buffer, j, buf_len, + parms, parms_index); *parms_index = begin_index; } break; @@ -1574,7 +1575,7 @@ H5Z__nbit_compress(unsigned char *data, unsigned d_nelmts, unsigned char *buffer p.offset = parms[7]; for (i = 0; i < d_nelmts; i++) - H5Z__nbit_compress_one_atomic(data, i * p.size, buffer, &new_size, &buf_len, &p); + H5Z__nbit_compress_one_atomic(data, i * (size_t)p.size, buffer, &new_size, &buf_len, &p); break; case H5Z_NBIT_ARRAY: diff --git a/src/H5Zscaleoffset.c b/src/H5Zscaleoffset.c index 9942cee..46c1a10 100644 --- a/src/H5Zscaleoffset.c +++ b/src/H5Zscaleoffset.c @@ -1205,7 +1205,7 @@ H5Z__filter_scaleoffset(unsigned flags, size_t cd_nelmts, const unsigned cd_valu /* prepare parameters to pass to compress/decompress functions */ p.size = cd_values[H5Z_SCALEOFFSET_PARM_SIZE]; - p.mem_order = H5T_native_order_g; + p.mem_order = (unsigned)H5T_native_order_g; /* input; decompress */ if (flags & H5Z_FLAG_REVERSE) { @@ -1240,7 +1240,7 @@ H5Z__filter_scaleoffset(unsigned flags, size_t cd_nelmts, const unsigned cd_valu p.minbits = minbits; /* calculate size of output buffer after decompression */ - size_out = d_nelmts * p.size; + size_out = d_nelmts * (size_t)p.size; /* allocate memory space for decompressed buffer */ if (NULL == (outbuf = (unsigned char *)H5MM_malloc(size_out))) @@ -1403,7 +1403,7 @@ H5Z__scaleoffset_convert(void *buf, unsigned d_nelmts, unsigned dtype_size) unsigned char *buffer, temp; buffer = (unsigned char *)buf; - for (i = 0; i < d_nelmts * dtype_size; i += dtype_size) + for (i = 0; i < d_nelmts * (size_t)dtype_size; i += dtype_size) for (j = 0; j < dtype_size / 2; j++) { /* swap pair of bytes */ temp = buffer[i + j]; @@ -1681,7 +1681,7 @@ H5Z__scaleoffset_decompress(unsigned char *data, unsigned d_nelmts, unsigned cha unsigned buf_len; /* must initialize to zeros */ - for (i = 0; i < d_nelmts * p.size; i++) + for (i = 0; i < d_nelmts * (size_t)p.size; i++) data[i] = 0; /* initialization before the loop */ diff --git a/src/H5module.h b/src/H5module.h index 6d3cba8..f7d3cd6 100644 --- a/src/H5module.h +++ b/src/H5module.h @@ -11,9 +11,9 @@ * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ /* - * Purpose: This file contains declarations which define macros for the - * H5 package. Including this header means that the source file - * is part of the H5 package. + * Purpose: This file contains declarations which define macros for the + * H5 package. Including this header means that the source file + * is part of the H5 package. */ #ifndef H5module_H #define H5module_H diff --git a/src/H5mpi.c b/src/H5mpi.c index aea0104..f5d709a 100644 --- a/src/H5mpi.c +++ b/src/H5mpi.c @@ -549,4 +549,237 @@ done: FUNC_LEAVE_NOAPI(ret_value) } /* end H5_mpio_create_large_type() */ +/*------------------------------------------------------------------------- + * Function: H5_mpio_gatherv_alloc + * + * Purpose: A wrapper around MPI_(All)gatherv that performs allocation + * of the receive buffer on the caller's behalf. This + * routine's parameters are as follows: + * + * `send_buf` - The buffer that data will be sent from for + * the calling MPI rank. Analogous to + * MPI_(All)gatherv's `sendbuf` parameter. + * + * `send_count` - The number of `send_type` elements in the + * send buffer. Analogous to MPI_(All)gatherv's + * `sendcount` parameter. + * + * `send_type` - The MPI Datatype of the elements in the send + * buffer. Analogous to MPI_(All)gatherv's + * `sendtype` parameter. + * + * `recv_counts` - An array containing the number of elements + * to be received from each MPI rank. + * Analogous to MPI_(All)gatherv's `recvcount` + * parameter. + * + * `displacements` - An array containing the displacements + * in the receive buffer where data from + * each MPI rank should be placed. Analogous + * to MPI_(All)gatherv's `displs` parameter. + * + * `recv_type` - The MPI Datatype of the elements in the + * receive buffer. Analogous to + * MPI_(All)gatherv's `recvtype` parameter. + * + * `allgather` - Specifies whether the gather operation to be + * performed should be MPI_Allgatherv (TRUE) or + * MPI_Gatherv (FALSE). + * + * `root` - For MPI_Gatherv operations, specifies the rank + * that will receive the data sent by other ranks. + * Analogous to MPI_Gatherv's `root` parameter. For + * MPI_Allgatherv operations, this parameter is + * ignored. + * + * `comm` - Specifies the MPI Communicator for the operation. + * Analogous to MPI_(All)gatherv's `comm` parameter. + * + * `mpi_rank` - Specifies the calling rank's rank value, as + * obtained by calling MPI_Comm_rank on the + * MPI Communicator `comm`. + * + * `mpi_size` - Specifies the MPI Communicator size, as + * obtained by calling MPI_Comm_size on the + * MPI Communicator `comm`. + * + * `out_buf` - Resulting buffer that is allocated and + * returned to the caller after data has been + * gathered into it. Returned only to the rank + * specified by `root` for MPI_Gatherv + * operations, or to all ranks for + * MPI_Allgatherv operations. + * + * `out_buf_num_entries` - The number of elements in the + * resulting buffer, in terms of + * the MPI Datatype provided for + * `recv_type`. + * + * Notes: This routine is collective across `comm`. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +herr_t +H5_mpio_gatherv_alloc(void *send_buf, int send_count, MPI_Datatype send_type, const int recv_counts[], + const int displacements[], MPI_Datatype recv_type, hbool_t allgather, int root, + MPI_Comm comm, int mpi_rank, int mpi_size, void **out_buf, size_t *out_buf_num_entries) +{ + size_t recv_buf_num_entries = 0; + void * recv_buf = NULL; +#if H5_CHECK_MPI_VERSION(3, 0) + MPI_Count type_lb; + MPI_Count type_extent; +#else + MPI_Aint type_lb; + MPI_Aint type_extent; +#endif + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_NOAPI(FAIL) + + HDassert(send_buf || send_count == 0); + if (allgather || (mpi_rank == root)) + HDassert(out_buf && out_buf_num_entries); + + /* Retrieve the extent of the MPI Datatype being used */ +#if H5_CHECK_MPI_VERSION(3, 0) + if (MPI_SUCCESS != (mpi_code = MPI_Type_get_extent_x(recv_type, &type_lb, &type_extent))) +#else + if (MPI_SUCCESS != (mpi_code = MPI_Type_get_extent(recv_type, &type_lb, &type_extent))) +#endif + HMPI_GOTO_ERROR(FAIL, "MPI_Type_get_extent(_x) failed", mpi_code) + + if (type_extent < 0) + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "MPI recv_type had a negative extent") + + /* + * Calculate the total size of the buffer being + * returned and allocate it + */ + if (allgather || (mpi_rank == root)) { + size_t i; + size_t buf_size; + + for (i = 0, recv_buf_num_entries = 0; i < (size_t)mpi_size; i++) + recv_buf_num_entries += (size_t)recv_counts[i]; + buf_size = recv_buf_num_entries * (size_t)type_extent; + + /* If our buffer size is 0, there's nothing to do */ + if (buf_size == 0) + HGOTO_DONE(SUCCEED) + + if (NULL == (recv_buf = H5MM_malloc(buf_size))) + /* Push an error, but still participate in collective gather operation */ + HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "couldn't allocate receive buffer") + } + + /* Perform gather operation */ + if (allgather) { + if (MPI_SUCCESS != (mpi_code = MPI_Allgatherv(send_buf, send_count, send_type, recv_buf, recv_counts, + displacements, recv_type, comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Allgatherv failed", mpi_code) + } + else { + if (MPI_SUCCESS != (mpi_code = MPI_Gatherv(send_buf, send_count, send_type, recv_buf, recv_counts, + displacements, recv_type, root, comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Gatherv failed", mpi_code) + } + + if (allgather || (mpi_rank == root)) { + *out_buf = recv_buf; + *out_buf_num_entries = recv_buf_num_entries; + } + +done: + if (ret_value < 0) { + if (recv_buf) + H5MM_free(recv_buf); + } + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5_mpio_gatherv_alloc() */ + +/*------------------------------------------------------------------------- + * Function: H5_mpio_gatherv_alloc_simple + * + * Purpose: A slightly simplified interface to H5_mpio_gatherv_alloc + * which calculates the receive counts and receive buffer + * displacements for the caller. + * + * Notes: This routine is collective across `comm`. + * + * Return: Non-negative on success/Negative on failure + * + *------------------------------------------------------------------------- + */ +herr_t +H5_mpio_gatherv_alloc_simple(void *send_buf, int send_count, MPI_Datatype send_type, MPI_Datatype recv_type, + hbool_t allgather, int root, MPI_Comm comm, int mpi_rank, int mpi_size, + void **out_buf, size_t *out_buf_num_entries) +{ + int * recv_counts_disps_array = NULL; + int mpi_code; + herr_t ret_value = SUCCEED; + + FUNC_ENTER_NOAPI(FAIL) + + HDassert(send_buf || send_count == 0); + if (allgather || (mpi_rank == root)) + HDassert(out_buf && out_buf_num_entries); + + /* + * Allocate array to store the receive counts of each rank, as well as + * the displacements into the final array where each rank will place + * their data. The first half of the array contains the receive counts + * (in rank order), while the latter half contains the displacements + * (also in rank order). + */ + if (allgather || (mpi_rank == root)) { + if (NULL == + (recv_counts_disps_array = H5MM_malloc(2 * (size_t)mpi_size * sizeof(*recv_counts_disps_array)))) + /* Push an error, but still participate in collective gather operation */ + HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, + "couldn't allocate receive counts and displacements array") + } + + /* Collect each rank's send count to interested ranks */ + if (allgather) { + if (MPI_SUCCESS != + (mpi_code = MPI_Allgather(&send_count, 1, MPI_INT, recv_counts_disps_array, 1, MPI_INT, comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code) + } + else { + if (MPI_SUCCESS != + (mpi_code = MPI_Gather(&send_count, 1, MPI_INT, recv_counts_disps_array, 1, MPI_INT, root, comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Gather failed", mpi_code) + } + + /* Set the displacements into the receive buffer for the gather operation */ + if (allgather || (mpi_rank == root)) { + size_t i; + int * displacements_ptr; + + displacements_ptr = &recv_counts_disps_array[mpi_size]; + + *displacements_ptr = 0; + for (i = 1; i < (size_t)mpi_size; i++) + displacements_ptr[i] = displacements_ptr[i - 1] + recv_counts_disps_array[i - 1]; + } + + /* Perform gather operation */ + if (H5_mpio_gatherv_alloc(send_buf, send_count, send_type, recv_counts_disps_array, + &recv_counts_disps_array[mpi_size], recv_type, allgather, root, comm, mpi_rank, + mpi_size, out_buf, out_buf_num_entries) < 0) + HGOTO_ERROR(H5E_LIB, H5E_CANTGATHER, FAIL, "can't gather data") + +done: + if (recv_counts_disps_array) + H5MM_free(recv_counts_disps_array); + + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5_mpio_gatherv_alloc_simple() */ + #endif /* H5_HAVE_PARALLEL */ diff --git a/src/H5private.h b/src/H5private.h index 765e7b6..d98d76d 100644 --- a/src/H5private.h +++ b/src/H5private.h @@ -44,10 +44,12 @@ #include <sys/time.h> #endif #ifdef H5_HAVE_UNISTD_H +#include <unistd.h> +#endif #ifdef H5_HAVE_PWD_H #include <pwd.h> #endif -#include <unistd.h> +#ifdef H5_HAVE_WAITPID #include <sys/wait.h> #endif @@ -170,7 +172,7 @@ */ #define BEGIN_MPE_LOG \ if (H5_MPEinit_g) { \ - sprintf(p_event_start, "start %s", __func__); \ + snprintf(p_event_start, sizeof(p_event_start), "start %s", __func__); \ if (eventa(__func__) == -1 && eventb(__func__) == -1) { \ const char *p_color = "red"; \ eventa(__func__) = MPE_Log_get_event_number(); \ @@ -385,6 +387,25 @@ #define HSSIZET_MAX ((hssize_t)LLONG_MAX) #define HSSIZET_MIN (~(HSSIZET_MAX)) +#ifdef H5_HAVE_PARALLEL + +/* Define a type for safely sending size_t values with MPI */ +#if SIZE_MAX == UCHAR_MAX +#define H5_SIZE_T_AS_MPI_TYPE MPI_UNSIGNED_CHAR +#elif SIZE_MAX == USHRT_MAX +#define H5_SIZE_T_AS_MPI_TYPE MPI_UNSIGNED_SHORT +#elif SIZE_MAX == UINT_MAX +#define H5_SIZE_T_AS_MPI_TYPE MPI_UNSIGNED +#elif SIZE_MAX == ULONG_MAX +#define H5_SIZE_T_AS_MPI_TYPE MPI_UNSIGNED_LONG +#elif SIZE_MAX == ULLONG_MAX +#define H5_SIZE_T_AS_MPI_TYPE MPI_UNSIGNED_LONG_LONG +#else +#error "no suitable MPI type for size_t" +#endif + +#endif /* H5_HAVE_PARALLEL */ + /* * Types and max sizes for POSIX I/O. * OS X (Darwin) is odd since the max I/O size does not match the types. @@ -506,6 +527,9 @@ #define H5_GCC_CLANG_DIAG_ON(x) #endif +/* Function pointer typedef for qsort */ +typedef int (*H5_sort_func_cb_t)(const void *, const void *); + /* Typedefs and functions for timing certain parts of the library. */ /* A set of elapsed/user/system times emitted as a time point by the @@ -574,7 +598,7 @@ typedef off_t h5_stat_size_t; #define HDoff_t off_t #endif -#/* Redefine all the POSIX and C functions. We should never see an +/* Redefine all the POSIX and C functions. We should never see an * undecorated POSIX or C function (or any other non-HDF5 function) * in the source. */ @@ -1130,57 +1154,6 @@ H5_DLL H5_ATTR_CONST int Nflock(int fd, int operation); #ifndef HDprintf #define HDprintf printf /*varargs*/ #endif -#ifndef HDpthread_attr_destroy -#define HDpthread_attr_destroy(A) pthread_attr_destroy(A) -#endif -#ifndef HDpthread_attr_init -#define HDpthread_attr_init(A) pthread_attr_init(A) -#endif -#ifndef HDpthread_attr_setscope -#define HDpthread_attr_setscope(A, S) pthread_attr_setscope(A, S) -#endif -#ifndef HDpthread_cond_init -#define HDpthread_cond_init(C, A) pthread_cond_init(C, A) -#endif -#ifndef HDpthread_cond_signal -#define HDpthread_cond_signal(C) pthread_cond_signal(C) -#endif -#ifndef HDpthread_cond_wait -#define HDpthread_cond_wait(C, M) pthread_cond_wait(C, M) -#endif -#ifndef HDpthread_create -#define HDpthread_create(R, A, F, U) pthread_create(R, A, F, U) -#endif -#ifndef HDpthread_equal -#define HDpthread_equal(T1, T2) pthread_equal(T1, T2) -#endif -#ifndef HDpthread_getspecific -#define HDpthread_getspecific(K) pthread_getspecific(K) -#endif -#ifndef HDpthread_join -#define HDpthread_join(T, V) pthread_join(T, V) -#endif -#ifndef HDpthread_key_create -#define HDpthread_key_create(K, D) pthread_key_create(K, D) -#endif -#ifndef HDpthread_mutex_init -#define HDpthread_mutex_init(M, A) pthread_mutex_init(M, A) -#endif -#ifndef HDpthread_mutex_lock -#define HDpthread_mutex_lock(M) pthread_mutex_lock(M) -#endif -#ifndef HDpthread_mutex_unlock -#define HDpthread_mutex_unlock(M) pthread_mutex_unlock(M) -#endif -#ifndef HDpthread_self -#define HDpthread_self() pthread_self() -#endif -#ifndef HDpthread_setcancelstate -#define HDpthread_setcancelstate(N, O) pthread_setcancelstate(N, O) -#endif -#ifndef HDpthread_setspecific -#define HDpthread_setspecific(K, V) pthread_setspecific(K, V) -#endif #ifndef HDputc #define HDputc(C, F) putc(C, F) #endif @@ -1766,6 +1739,15 @@ typedef struct H5_debug_t { } H5_debug_t; #ifdef H5_HAVE_PARALLEL + +/* + * Check that the MPI library version is at least version + * `mpi_version` and subversion `mpi_subversion` + */ +#define H5_CHECK_MPI_VERSION(mpi_version, mpi_subversion) \ + ((MPI_VERSION > (mpi_version)) || \ + ((MPI_VERSION == (mpi_version)) && (MPI_SUBVERSION >= (mpi_subversion)))) + extern hbool_t H5_coll_api_sanity_check_g; #endif /* H5_HAVE_PARALLEL */ @@ -2006,6 +1988,14 @@ extern hbool_t H5_libterm_g; /* Is the library being shutdown? */ #endif /* H5_HAVE_THREADSAFE */ +/* Extern global to determine if we should use selection I/O if available (this + * variable should be removed once selection I/O performs as well as the + * previous scalar I/O implementation + * + * NOTE: Must be exposed via H5_DLLVAR so parallel tests pass on Windows. + */ +H5_DLLVAR hbool_t H5_use_selection_io_g; + #ifdef H5_HAVE_CODESTACK /* Include required function stack header */ @@ -2500,6 +2490,16 @@ H5_DLL herr_t H5CX_pop(hbool_t update_dxpl_props); #define HDcompile_assert(e) do { typedef struct { unsigned int b: (e); } x; } while(0) */ +/* Private typedefs */ + +/* Union for const/non-const pointer for use by functions that manipulate + * pointers but do not write to their targets or return pointers to const + * specified locations. This helps us avoid compiler warnings. */ +typedef union { + void * vp; + const void *cvp; +} H5_flexible_const_ptr_t; + /* Private functions, not part of the publicly documented API */ H5_DLL herr_t H5_init_library(void); H5_DLL void H5_term_library(void); @@ -2600,7 +2600,8 @@ struct h5_long_options { */ }; -H5_DLL int H5_get_option(int argc, const char **argv, const char *opt, const struct h5_long_options *l_opt); +H5_DLL int H5_get_option(int argc, const char *const *argv, const char *opt, + const struct h5_long_options *l_opt); #ifdef H5_HAVE_PARALLEL /* Generic MPI functions */ @@ -2614,6 +2615,14 @@ H5_DLL herr_t H5_mpi_comm_cmp(MPI_Comm comm1, MPI_Comm comm2, int *result); H5_DLL herr_t H5_mpi_info_cmp(MPI_Info info1, MPI_Info info2, int *result); H5_DLL herr_t H5_mpio_create_large_type(hsize_t num_elements, MPI_Aint stride_bytes, MPI_Datatype old_type, MPI_Datatype *new_type); +H5_DLL herr_t H5_mpio_gatherv_alloc(void *send_buf, int send_count, MPI_Datatype send_type, + const int recv_counts[], const int displacements[], + MPI_Datatype recv_type, hbool_t allgather, int root, MPI_Comm comm, + int mpi_rank, int mpi_size, void **out_buf, size_t *out_buf_num_entries); +H5_DLL herr_t H5_mpio_gatherv_alloc_simple(void *send_buf, int send_count, MPI_Datatype send_type, + MPI_Datatype recv_type, hbool_t allgather, int root, MPI_Comm comm, + int mpi_rank, int mpi_size, void **out_buf, + size_t *out_buf_num_entries); #endif /* H5_HAVE_PARALLEL */ /* Functions for debugging */ diff --git a/src/H5public.h b/src/H5public.h index 6a3911c..3f9848a 100644 --- a/src/H5public.h +++ b/src/H5public.h @@ -83,7 +83,7 @@ /** * For tweaks, bug-fixes, or development */ -#define H5_VERS_RELEASE 1 +#define H5_VERS_RELEASE 2 /** * For pre-releases like \c snap0. Empty string for official releases. */ @@ -91,7 +91,7 @@ /** * Full version string */ -#define H5_VERS_INFO "HDF5 library version: 1.13.1-1" +#define H5_VERS_INFO "HDF5 library version: 1.13.2-1" #define H5check() H5check_version(H5_VERS_MAJOR, H5_VERS_MINOR, H5_VERS_RELEASE) @@ -289,6 +289,11 @@ typedef long long ssize_t; * \internal Defined as a (minimum) 64-bit integer type. */ typedef uint64_t hsize_t; + +#ifdef H5_HAVE_PARALLEL +#define HSIZE_AS_MPI_TYPE MPI_UINT64_T +#endif + /** * The size of file objects. Used when negative values are needed to indicate errors. * @@ -323,7 +328,7 @@ typedef uint64_t haddr_t; #define HADDR_MAX (HADDR_UNDEF - 1) #ifdef H5_HAVE_PARALLEL -#define HADDR_AS_MPI_TYPE MPI_LONG_LONG_INT +#define HADDR_AS_MPI_TYPE MPI_UINT64_T #endif //! <!-- [H5_iter_order_t_snip] --> diff --git a/src/H5system.c b/src/H5system.c index 9a966b0..a369e3d 100644 --- a/src/H5system.c +++ b/src/H5system.c @@ -862,7 +862,7 @@ H5_nanosleep(uint64_t nanosec) #else - const uint64_t nanosec_per_sec = 1000 * 1000 * 1000; + const uint64_t nanosec_per_sec = 1000 * 1000L * 1000; struct timespec sleeptime; /* Struct to hold time to sleep */ /* Set up time to sleep @@ -956,7 +956,7 @@ const char *H5_optarg; /* Flag argument (or value) */ *------------------------------------------------------------------------- */ int -H5_get_option(int argc, const char **argv, const char *opts, const struct h5_long_options *l_opts) +H5_get_option(int argc, const char *const *argv, const char *opts, const struct h5_long_options *l_opts) { static int sp = 1; /* character index in current token */ int optchar = '?'; /* option character passed back to user */ @@ -1033,7 +1033,7 @@ H5_get_option(int argc, const char **argv, const char *opts, const struct h5_lon HDfree(arg); } else { - register char *cp; /* pointer into current token */ + char *cp; /* pointer into current token */ /* short command line option */ optchar = argv[H5_optind][sp]; diff --git a/src/H5timer.c b/src/H5timer.c index b2cc5f0..b5dba97 100644 --- a/src/H5timer.c +++ b/src/H5timer.c @@ -193,17 +193,26 @@ H5_now_usec(void) struct timespec ts; HDclock_gettime(CLOCK_MONOTONIC, &ts); - now = (uint64_t)(ts.tv_sec * (1000 * 1000)) + (uint64_t)(ts.tv_nsec / 1000); + + /* Cast all values in this expression to uint64_t to ensure that all intermediate + * calculations are done in 64 bit, to prevent overflow */ + now = ((uint64_t)ts.tv_sec * ((uint64_t)1000 * (uint64_t)1000)) + + ((uint64_t)ts.tv_nsec / (uint64_t)1000); } #elif defined(H5_HAVE_GETTIMEOFDAY) { struct timeval now_tv; HDgettimeofday(&now_tv, NULL); - now = (uint64_t)(now_tv.tv_sec * (1000 * 1000)) + (uint64_t)now_tv.tv_usec; + + /* Cast all values in this expression to uint64_t to ensure that all intermediate + * calculations are done in 64 bit, to prevent overflow */ + now = ((uint64_t)now_tv.tv_sec * ((uint64_t)1000 * (uint64_t)1000)) + (uint64_t)now_tv.tv_usec; } #else /* H5_HAVE_GETTIMEOFDAY */ - now = (uint64_t)(HDtime(NULL) * (1000 * 1000)); + /* Cast all values in this expression to uint64_t to ensure that all intermediate calculations + * are done in 64 bit, to prevent overflow */ + now = ((uint64_t)HDtime(NULL) * ((uint64_t)1000 * (uint64_t)1000)); #endif /* H5_HAVE_GETTIMEOFDAY */ return (now); diff --git a/src/H5trace.c b/src/H5trace.c index baf6a10..8790a88 100644 --- a/src/H5trace.c +++ b/src/H5trace.c @@ -1248,7 +1248,7 @@ H5_trace_args(H5RS_str_t *rs, const char *type, va_list ap) { H5FD_class_t cls = HDva_arg(ap, H5FD_class_t); - H5RS_asprintf_cat(rs, "{'%s', " H5_PRINTF_HADDR_FMT ", ", cls.name, cls.maxaddr); + H5RS_asprintf_cat(rs, "{'%s', %" PRIuHADDR ", ", cls.name, cls.maxaddr); H5_trace_args_close_degree(rs, cls.fc_degree); H5RS_acat(rs, ", ...}"); } /* end block */ @@ -4023,7 +4023,7 @@ H5_trace(const double *returning, const char *func, const char *type, ...) H5_timer_get_times(function_timer, &function_times); H5_timer_get_times(running_timer, &running_times); - HDsprintf(tmp, "%.6f", (function_times.elapsed - running_times.elapsed)); + HDsnprintf(tmp, sizeof(tmp), "%.6f", (function_times.elapsed - running_times.elapsed)); H5RS_asprintf_cat(rs, " %*s ", (int)HDstrlen(tmp), ""); } for (i = 0; i < current_depth; i++) diff --git a/src/Makefile.am b/src/Makefile.am index c4023ae..edfd9b0 100644 --- a/src/Makefile.am +++ b/src/Makefile.am @@ -74,12 +74,11 @@ libhdf5_la_SOURCES= H5.c H5checksum.c H5dbg.c H5lib_settings.c H5system.c \ H5HFspace.c H5HFstat.c H5HFtest.c H5HFtiny.c \ H5HG.c H5HGcache.c H5HGdbg.c H5HGquery.c \ H5HL.c H5HLcache.c H5HLdbg.c H5HLint.c H5HLprfx.c H5HLdblk.c \ - H5HP.c \ H5I.c H5Idbg.c H5Iint.c H5Itest.c \ H5L.c H5Ldeprec.c H5Lexternal.c H5Lint.c \ H5M.c \ H5MF.c H5MFaggr.c H5MFdbg.c H5MFsection.c \ - H5MM.c H5MP.c H5MPtest.c \ + H5MM.c \ H5O.c H5Odeprec.c H5Oainfo.c H5Oalloc.c H5Oattr.c H5Oattribute.c \ H5Obogus.c H5Obtreek.c H5Ocache.c H5Ocache_image.c H5Ochunk.c \ H5Ocont.c H5Ocopy.c H5Ocopy_ref.c H5Odbg.c H5Odrvinfo.c H5Odtype.c \ |