diff options
Diffstat (limited to 'src')
41 files changed, 5511 insertions, 2046 deletions
@@ -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 */ @@ -954,6 +954,7 @@ 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 +975,7 @@ 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 +1010,10 @@ 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 +1021,7 @@ 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 +1040,7 @@ 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, @@ -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) @@ -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 */ @@ -2600,16 +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 */ @@ -3082,8 +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 */ @@ -3093,8 +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 */ @@ -3261,8 +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 */ @@ -3528,8 +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 */ @@ -6063,8 +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 */ @@ -7411,6 +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; @@ -7739,7 +7746,7 @@ done: /*------------------------------------------------------------------------- * - * Function: H5C__validate_lru_list + * Function: H5C_validate_lru_list * * Purpose: Debugging function that scans the LRU list for errors. * @@ -7754,15 +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); @@ -7771,51 +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. @@ -7831,15 +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); @@ -7848,54 +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. @@ -7911,15 +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); @@ -7928,54 +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 @@ -7989,8 +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; @@ -8018,7 +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 */ /*------------------------------------------------------------------------- @@ -8508,8 +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 */ @@ -1423,7 +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 * 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/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/H5Dchunk.c b/src/H5Dchunk.c index 5d7c1b2..e4d8706 100644 --- a/src/H5Dchunk.c +++ b/src/H5Dchunk.c @@ -239,10 +239,14 @@ 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 { @@ -287,9 +291,6 @@ 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); @@ -306,8 +307,6 @@ static herr_t H5D__chunk_mem_cb(void *elem, const H5T_t *type, unsigned ndims, 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 +314,9 @@ 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 */ @@ -1362,7 +1361,7 @@ done: * *------------------------------------------------------------------------- */ -static void * +void * H5D__chunk_mem_alloc(size_t size, const H5O_pline_t *pline) { void *ret_value = NULL; /* Return value */ @@ -1393,7 +1392,7 @@ 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 +1416,7 @@ 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 +1436,7 @@ 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 */ @@ -3178,7 +3177,9 @@ 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 +3253,10 @@ 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 +3302,8 @@ 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 +4319,8 @@ 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 +4367,8 @@ 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 +4640,26 @@ 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") - /* Store the chunk's address for later */ - chunk_info.addr[chunk_info.num_io] = udata.chunk_block.offset; - chunk_info.num_io++; + chunk_fill_info.chunk_info = tmp_realloc; + } + + /* 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 +4732,7 @@ 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 +4748,8 @@ 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 +4943,35 @@ 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 +4987,89 @@ 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; - /* MSC - should not need this if MPI_type_create_hindexed_block is working */ - block_lens[i] = block_len; + /* 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) + +#if MPI_VERSION >= 3 && MPI_SUBVERSION >= 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 +5077,71 @@ 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 +5184,25 @@ 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); + entry1 = (const struct chunk_coll_fill_info *)_entry1; + entry2 = (const struct chunk_coll_fill_info *)_entry2; -#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 - - 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 */ /*------------------------------------------------------------------------- @@ -6826,7 +6916,7 @@ 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 +7211,89 @@ 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/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 1ea3f07..e226a0a 100644 --- a/src/H5Dio.c +++ b/src/H5Dio.c @@ -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 */ @@ -808,98 +821,35 @@ H5D__ioinfo_adjust(H5D_io_info_t *io_info, const H5D_t *dset, const H5S_t *file_ io_info->io_ops.single_write = H5D__mpio_select_write; } /* 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 cf8a6ef..527fc7b 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(). + * filtered I/O in the function H5D__mpio_collective_filtered_chunk_io_setup(). This + * struct's fields are as follows: * - * This struct's fields are as follows: + * 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: * - * 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. + * chunk_idx - The index of the chunk in the dataset's chunk index. * - * 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. + * 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. * - * 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. + * need_insert - A flag which determines whether or not a chunk needs to be re-inserted into + * the chunk index after the write operation. * - * 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. + * 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. * - * 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. + * 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. * - * 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_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_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. + * 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. * - * Its fields are as follows: + * 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. * - * 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. + * 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. * - * 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. + * chunk_buf_size - The size in bytes of the data buffer allocated for the chunk * - * 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. + * 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. * - * Its fields are as follows: + * new_owner - The MPI rank which has been selected to perform the modifications to this 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. + * 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. * - * new_owner - The processor which has been selected to perform the write to this chunk. + * 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. * - * 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. + * hh - A handle for hash tables provided by the uthash.h header * - * Its fields are as follows: - * - * 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. - * - * 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); \ + fflush(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__); \ + fflush(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)strlen(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)strlen(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 = time(NULL); + HDfprintf(debug_log_file, "##### %s", asctime(localtime(&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. - */ - 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. + 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. */ - 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__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 (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 (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 MPI_VERSION >= 3 + /* + * 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; + + 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: The difference between the two - * H5D_filtered_collective_io_info_t's original owner fields + * Return: -1, 0, 1 * - * Programmer: Jordan Henderson - * Monday, Apr. 10th, 2017 + *------------------------------------------------------------------------- + */ +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 + * + * Purpose: Routine to compare the original owning MPI rank for two + * H5D_chunk_redistribute_info_t structures + * + * 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. @@ -2337,30 +2771,47 @@ 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) { + 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 */ @@ -2413,10 +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); @@ -2424,6 +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); @@ -2469,22 +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 @@ -2492,12 +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) @@ -2544,6 +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)))) @@ -2610,6 +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) @@ -2623,34 +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 @@ -2660,19 +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); @@ -2683,275 +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); + + /* + * Phase 1 - Participate in collective gathering of every rank's + * list of chunks to the ranks which are performing the redistribution + * operation. + */ - 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") + 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; - 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") + 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]; + } + } + + /* + * Construct MPI derived types for extracting information + * necessary for MPI communication + */ + 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") + + /* + * 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); + } - /* 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. + /* + * Phase 2 - Involved ranks now redistribute any shared chunks to new + * owners as necessary. */ - 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") - /* 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_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; - 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") + /* 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)); - 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") + /* 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); - 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; + /* + * 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; - /* Process each set of duplicate entries caused by another process writing to the same chunk */ - do { - chunk_entry = &shared_chunks_info_array[i]; + /* Reset the initial number of writers to this chunk */ + num_writers = 0; - send_counts[chunk_entry->owners.original_owner] += (int)sizeof(*chunk_entry); + /* 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]; + + 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]++; + } - /* 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 + /* + * 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; + + 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 MPI_VERSION >= 3 + 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") - /* 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. + /* + * 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; @@ -2960,466 +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 (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0) + 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 (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") - /* 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]; + H5_CHECK_OVERFLOW(num_send_requests, size_t, int) + H5_CHECK_OVERFLOW(num_msgs_incoming, size_t, int) - 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; + /* + * 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") - 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") + 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 (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") + /* Process any incoming messages until everyone is done */ + do { + MPI_Status status; + int msg_flag; - for (j = 0; j < chunk_entry->num_writers - 1; j++) { - int count = 0; + /* 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) - /* 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 */ + /* + * If a message was found, allocate a buffer for the message and + * post a non-blocking receive to receive it + */ + if (msg_flag) { +#if MPI_VERSION >= 3 + MPI_Count msg_size = 0; - *local_chunk_array_num_entries = last_assigned_idx; + 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) - /* 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") + H5_CHECK_OVERFLOW(msg_size, MPI_Count, int) +#else + int msg_size = 0; - 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 */ + if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&status, MPI_BYTE, &msg_size))) + HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code) +#endif -done: - /* Now that all async send requests have completed, free up the send - * buffers used in the async operations + if (msg_size <= 0) + HGOTO_ERROR(H5E_DATASET, H5E_BADVALUE, FAIL, "invalid chunk modification message size") + + 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") + + 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; - /* 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. + /* 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); + } + + /* + * 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 - iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->fspace); + 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 (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") + /* Set size of full chunks in dataset */ + file_chunk_size = io_info->dset->shared->layout.u.chunk.size; - 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") + /* 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); + } - iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace); + /* + * 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); + + 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 (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_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 MPI_VERSION >= 3 + 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 MPI_VERSION >= 3 + 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 MPI_VERSION >= 3 && MPI_SUBVERSION >= 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 (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 (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__filtered_collective_chunk_entry_io() */ +} /* 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..a424929 100644 --- a/src/H5Dpkg.h +++ b/src/H5Dpkg.h @@ -559,6 +559,7 @@ 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 +592,10 @@ 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 +640,13 @@ 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 +705,11 @@ 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 +738,8 @@ 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/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. @@ -269,7 +269,7 @@ H5ESget_requests(hid_t es_id, H5_iter_order_t order, hid_t *connector_ids, void herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_API(FAIL) - H5TRACE5("e", "iIo*i**xx", es_id, order, connector_ids, requests, count); + 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))) @@ -928,9 +928,10 @@ 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) 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 1feff43..4aa8a96 100644 --- a/src/H5FDmpio.c +++ b/src/H5FDmpio.c @@ -188,6 +188,41 @@ 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 */ /*------------------------------------------------------------------------- @@ -994,7 +1029,6 @@ 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 */ @@ -1172,6 +1206,7 @@ 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)); @@ -1199,8 +1234,6 @@ 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) { @@ -1304,6 +1337,21 @@ 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); @@ -1366,8 +1414,8 @@ 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, bytes_read, H5FD__mem_t_to_str(type)); #endif /* @@ -1377,6 +1425,9 @@ 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); @@ -1489,20 +1540,6 @@ 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) { @@ -1548,6 +1585,21 @@ 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); @@ -1583,8 +1635,8 @@ 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, bytes_written, H5FD__mem_t_to_str(type)); #endif /* Each process will keep track of its perceived EOF value locally, and 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..67e153e 100644 --- a/src/H5Fprivate.h +++ b/src/H5Fprivate.h @@ -962,6 +962,9 @@ 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/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/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..2c3caa8 100644 --- a/src/H5Pfapl.c +++ b/src/H5Pfapl.c @@ -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/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/H5Tnative.c b/src/H5Tnative.c index 44c6d45..d56ac99 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/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. */ @@ -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 5bdc51c..46c1a10 100644 --- a/src/H5Zscaleoffset.c +++ b/src/H5Zscaleoffset.c @@ -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/H5mpi.c b/src/H5mpi.c index aea0104..15fb785 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 MPI_VERSION >= 3 + 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 MPI_VERSION >= 3 + 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 fa52e7f..d67163f 100644 --- a/src/H5private.h +++ b/src/H5private.h @@ -387,6 +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. @@ -508,6 +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 @@ -2602,7 +2624,8 @@ 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 */ @@ -2616,6 +2639,14 @@ 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..ee9077e 100644 --- a/src/H5system.c +++ b/src/H5system.c @@ -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 */ 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 5d15fee..8790a88 100644 --- a/src/H5trace.c +++ b/src/H5trace.c @@ -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++) |