From c77dc57fda2f912911998eab2f67d18ba93485ab Mon Sep 17 00:00:00 2001 From: Vailin Choi Date: Thu, 17 Sep 2015 22:34:04 -0500 Subject: [svn-r27822] Bring revisions #27262 - 27294 from trunk to revise_chunks. Tested on jam, ostrich, platypus, osx1010test, emu. --- src/H5ACmpio.c | 2295 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++ src/H5Cmpio.c | 1220 ++++++++++++++++++++++++++++++ 2 files changed, 3515 insertions(+) create mode 100644 src/H5ACmpio.c create mode 100644 src/H5Cmpio.c diff --git a/src/H5ACmpio.c b/src/H5ACmpio.c new file mode 100644 index 0000000..367d6ef --- /dev/null +++ b/src/H5ACmpio.c @@ -0,0 +1,2295 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * Copyright by The HDF Group. * + * Copyright by the Board of Trustees of the University of Illinois. * + * All rights reserved. * + * * + * This file is part of HDF5. The full HDF5 copyright notice, including * + * terms governing use, modification, and redistribution, is contained in * + * the files COPYING and Copyright.html. COPYING can be found at the root * + * of the source code distribution tree; Copyright.html can be found at the * + * root level of an installed copy of the electronic HDF5 document set and * + * is linked from the top-level documents page. It can also be found at * + * http://hdfgroup.org/HDF5/doc/Copyright.html. If you do not have * + * access to either file, you may request a copy from help@hdfgroup.org. * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +/*------------------------------------------------------------------------- + * + * Created: H5ACmpio.c + * Jun 20 2015 + * Quincey Koziol + * + * Purpose: Functions in this file implement support for parallel + * I/O cache functionality + * + *------------------------------------------------------------------------- + */ + +/****************/ +/* Module Setup */ +/****************/ + +#define H5AC_PACKAGE /*suppress error about including H5ACpkg */ +#define H5C_PACKAGE /*suppress error about including H5Cpkg */ +#define H5F_PACKAGE /*suppress error about including H5Fpkg */ + +/* Interface initialization */ +#define H5_INTERFACE_INIT_FUNC H5AC__init_mpio_interface + +/***********/ +/* Headers */ +/***********/ +#include "H5private.h" /* Generic Functions */ +#include "H5ACpkg.h" /* Metadata cache */ +#include "H5Cpkg.h" /* Cache */ +#include "H5Eprivate.h" /* Error handling */ +#include "H5Fpkg.h" /* Files */ +#include "H5MMprivate.h" /* Memory management */ + +#ifdef H5_HAVE_PARALLEL + +/****************/ +/* Local Macros */ +/****************/ + + +/******************/ +/* Local Typedefs */ +/******************/ + +/**************************************************************************** + * + * structure H5AC_slist_entry_t + * + * The dirty entry list maintained via the d_slist_ptr field of H5AC_aux_t + * and the cleaned entry list maintained via the c_slist_ptr field of + * H5AC_aux_t are just lists of the file offsets of the dirty/cleaned + * entries. Unfortunately, the slist code makes us define a dynamically + * allocated structure to store these offsets in. This structure serves + * that purpose. Its fields are as follows: + * + * addr: file offset of a metadata entry. Entries are added to this + * list (if they aren't there already) when they are marked + * dirty in an unprotect, inserted, or moved. They are + * removed when they appear in a clean entries broadcast. + * + ****************************************************************************/ +typedef struct H5AC_slist_entry_t +{ + haddr_t addr; +} H5AC_slist_entry_t; + +/* User data for address list building callbacks */ +typedef struct H5AC_addr_list_ud_t +{ + H5AC_aux_t * aux_ptr; /* 'Auxiliary' parallel cache info */ + haddr_t * addr_buf_ptr; /* Array to store addresses */ + int i; /* Counter for position in array */ +} H5AC_addr_list_ud_t; + + +/********************/ +/* Local Prototypes */ +/********************/ + +static herr_t H5AC__broadcast_candidate_list(H5AC_t *cache_ptr, + int *num_entries_ptr, haddr_t **haddr_buf_ptr_ptr); +static herr_t H5AC__broadcast_clean_list(H5AC_t *cache_ptr); +static herr_t H5AC__construct_candidate_list(H5AC_t *cache_ptr, + H5AC_aux_t *aux_ptr, int sync_point_op); +static herr_t H5AC__copy_candidate_list_to_buffer(const H5AC_t *cache_ptr, + int *num_entries_ptr, haddr_t **haddr_buf_ptr_ptr); +static herr_t H5AC__propagate_and_apply_candidate_list(H5F_t *f, hid_t dxpl_id); +static herr_t H5AC__propagate_flushed_and_still_clean_entries_list(H5F_t *f, + hid_t dxpl_id); +static herr_t H5AC__receive_haddr_list(MPI_Comm mpi_comm, int *num_entries_ptr, + haddr_t **haddr_buf_ptr_ptr); +static herr_t H5AC__receive_candidate_list(const H5AC_t *cache_ptr, + int *num_entries_ptr, haddr_t **haddr_buf_ptr_ptr); +static herr_t H5AC__receive_and_apply_clean_list(H5F_t *f, hid_t dxpl_id); +static herr_t H5AC__tidy_cache_0_lists(H5AC_t *cache_ptr, int num_candidates, + haddr_t *candidates_list_ptr); +static herr_t H5AC__rsp__dist_md_write__flush(H5F_t *f, hid_t dxpl_id); +static herr_t H5AC__rsp__dist_md_write__flush_to_min_clean(H5F_t *f, hid_t dxpl_id); +static herr_t H5AC__rsp__p0_only__flush(H5F_t *f, hid_t dxpl_id); +static herr_t H5AC__rsp__p0_only__flush_to_min_clean(H5F_t *f, hid_t dxpl_id); + + +/*********************/ +/* Package Variables */ +/*********************/ + +/* Declare a free list to manage the H5AC_aux_t struct */ +H5FL_DEFINE(H5AC_aux_t); + + +/*****************************/ +/* Library Private Variables */ +/*****************************/ + + +/*******************/ +/* Local Variables */ +/*******************/ + +/* Declare a free list to manage the H5AC_slist_entry_t struct */ +H5FL_DEFINE_STATIC(H5AC_slist_entry_t); + + +/*------------------------------------------------------------------------- + * Function: H5AC__init_mpio_interface + * + * Purpose: Initialize interface-specific information + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Quincey Koziol + * 6/20/15 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__init_mpio_interface(void) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Funnel all work to H5AC_init() */ + if(H5AC_init() < 0) + HGOTO_ERROR(H5E_FUNC, H5E_CANTINIT, FAIL, "interface initialization failed") + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__init_mpio_interface() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__set_sync_point_done_callback + * + * Purpose: Set the value of the sync_point_done callback. This + * callback is used by the parallel test code to verify + * that the expected writes and only the expected writes + * take place during a sync point. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 5/9/10 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__set_sync_point_done_callback(H5C_t * cache_ptr, + void (* sync_point_done)(int num_writes, haddr_t * written_entries_tbl)) +{ + H5AC_aux_t * aux_ptr; + + FUNC_ENTER_PACKAGE_NOERR + + /* Sanity checks */ + HDassert(cache_ptr && (cache_ptr->magic == H5C__H5C_T_MAGIC)); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + + aux_ptr->sync_point_done = sync_point_done; + + FUNC_LEAVE_NOAPI(SUCCEED) +} /* H5AC__set_sync_point_done_callback() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__set_write_done_callback + * + * Purpose: Set the value of the write_done callback. This callback + * is used to improve performance of the parallel test bed + * for the cache. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 5/11/06 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__set_write_done_callback(H5C_t * cache_ptr, void (* write_done)(void)) +{ + H5AC_aux_t * aux_ptr; + + FUNC_ENTER_PACKAGE_NOERR + + /* Sanity checks */ + HDassert(cache_ptr && (cache_ptr->magic == H5C__H5C_T_MAGIC)); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC ); + + aux_ptr->write_done = write_done; + + FUNC_LEAVE_NOAPI(SUCCEED) +} /* H5AC__set_write_done_callback() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC_add_candidate() + * + * Purpose: Add the supplied metadata entry address to the candidate + * list. Verify that each entry added does not appear in + * the list prior to its insertion. + * + * This function is intended for used in constructing list + * of entried to be flushed during sync points. It shouldn't + * be called anywhere else. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 3/17/10 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC_add_candidate(H5AC_t * cache_ptr, haddr_t addr) +{ + H5AC_aux_t * aux_ptr; + H5AC_slist_entry_t * slist_entry_ptr = NULL; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + /* Sanity checks */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + HDassert(aux_ptr->candidate_slist_ptr != NULL); + + /* Construct an entry for the supplied address, and insert + * it into the candidate slist. + */ + if(NULL == (slist_entry_ptr = H5FL_MALLOC(H5AC_slist_entry_t))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "Can't allocate candidate slist entry") + slist_entry_ptr->addr = addr; + + if(H5SL_insert(aux_ptr->candidate_slist_ptr, slist_entry_ptr, &(slist_entry_ptr->addr)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTINSERT, FAIL, "can't insert entry into dirty entry slist") + +done: + /* Clean up on error */ + if(ret_value < 0) + if(slist_entry_ptr) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC_add_candidate() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__broadcast_candidate_list() + * + * Purpose: Broadcast the contents of the process 0 candidate entry + * slist. In passing, also remove all entries from said + * list. As the application of this will be handled by + * the same functions on all processes, construct and + * return a copy of the list in the same format as that + * received by the other processes. Note that if this + * copy is returned in *haddr_buf_ptr_ptr, the caller + * must free it. + * + * This function must only be called by the process with + * MPI_rank 0. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 7/1/05 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__broadcast_candidate_list(H5AC_t *cache_ptr, int *num_entries_ptr, + haddr_t **haddr_buf_ptr_ptr) +{ + H5AC_aux_t * aux_ptr = NULL; + haddr_t * haddr_buf_ptr = NULL; + int mpi_result; + int num_entries; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->mpi_rank == 0); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + HDassert(aux_ptr->candidate_slist_ptr != NULL); + HDassert(num_entries_ptr != NULL); + HDassert(*num_entries_ptr == 0); + HDassert(haddr_buf_ptr_ptr != NULL); + HDassert(*haddr_buf_ptr_ptr == NULL); + + /* First broadcast the number of entries in the list so that the + * receivers can set up buffers to receive them. If there aren't + * any, we are done. + */ + num_entries = (int)H5SL_count(aux_ptr->candidate_slist_ptr); + if(MPI_SUCCESS != (mpi_result = MPI_Bcast(&num_entries, 1, MPI_INT, 0, aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_result) + + if(num_entries > 0) { + size_t buf_size = 0; + int chk_num_entries = 0; + + /* convert the candidate list into the format we + * are used to receiving from process 0, and also load it + * into a buffer for transmission. + */ + if(H5AC__copy_candidate_list_to_buffer(cache_ptr, &chk_num_entries, &haddr_buf_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate buffer.") + HDassert(chk_num_entries == num_entries); + HDassert(haddr_buf_ptr != NULL); + + /* Now broadcast the list of candidate entries */ + buf_size = sizeof(haddr_t) * (size_t)num_entries; + if(MPI_SUCCESS != (mpi_result = MPI_Bcast((void *)haddr_buf_ptr, (int)buf_size, MPI_BYTE, 0, aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_result) + } /* end if */ + + /* Pass the number of entries and the buffer pointer + * back to the caller. Do this so that we can use the same code + * to apply the candidate list to all the processes. + */ + *num_entries_ptr = num_entries; + *haddr_buf_ptr_ptr = haddr_buf_ptr; + +done: + if(ret_value < 0) + if(haddr_buf_ptr) + haddr_buf_ptr = (haddr_t *)H5MM_xfree((void *)haddr_buf_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__broadcast_candidate_list() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__broadcast_clean_list_cb() + * + * Purpose: Skip list callback for building array of addresses for + * broadcasting the clean list. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: Quincey Koziol, 6/12/15 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__broadcast_clean_list_cb(void *_item, void H5_ATTR_UNUSED *_key, + void *_udata) +{ + H5AC_slist_entry_t * slist_entry_ptr = (H5AC_slist_entry_t *)_item; /* Address of item */ + H5AC_addr_list_ud_t * udata = (H5AC_addr_list_ud_t *)_udata; /* Context for callback */ + haddr_t addr; + + FUNC_ENTER_STATIC_NOERR + + /* Sanity checks */ + HDassert(slist_entry_ptr); + HDassert(udata); + + /* Store the entry's address in the buffer */ + addr = slist_entry_ptr->addr; + udata->addr_buf_ptr[udata->i] = addr; + udata->i++; + + /* now release the entry */ + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + + /* and also remove the matching entry from the dirtied list + * if it exists. + */ + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(udata->aux_ptr->d_slist_ptr, (void *)(&addr)))) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + + FUNC_LEAVE_NOAPI(SUCCEED) +} /* H5AC__broadcast_clean_list_cb() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__broadcast_clean_list() + * + * Purpose: Broadcast the contents of the process 0 cleaned entry + * slist. In passing, also remove all entries from said + * list, and also remove any matching entries from the dirtied + * slist. + * + * This function must only be called by the process with + * MPI_rank 0. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 7/1/05 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__broadcast_clean_list(H5AC_t * cache_ptr) +{ + haddr_t * addr_buf_ptr = NULL; + H5AC_aux_t * aux_ptr; + int mpi_result; + int num_entries = 0; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)cache_ptr->aux_ptr; + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->mpi_rank == 0); + HDassert(aux_ptr->c_slist_ptr != NULL); + + /* First broadcast the number of entries in the list so that the + * receives can set up a buffer to receive them. If there aren't + * any, we are done. + */ + num_entries = (int)H5SL_count(aux_ptr->c_slist_ptr); + if(MPI_SUCCESS != (mpi_result = MPI_Bcast(&num_entries, 1, MPI_INT, 0, aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_result) + + if(num_entries > 0) { + H5AC_addr_list_ud_t udata; + size_t buf_size; + + /* allocate a buffer to store the list of entry base addresses in */ + buf_size = sizeof(haddr_t) * (size_t)num_entries; + if(NULL == (addr_buf_ptr = (haddr_t *)H5MM_malloc(buf_size))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "memory allocation failed for addr buffer") + + /* Set up user data for callback */ + udata.aux_ptr = aux_ptr; + udata.addr_buf_ptr = addr_buf_ptr; + udata.i = 0; + + /* Free all the clean list entries, building the address list in the callback */ + /* (Callback also removes the matching entries from the dirtied list) */ + if(H5SL_free(aux_ptr->c_slist_ptr, H5AC__broadcast_clean_list_cb, &udata) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFREE, FAIL, "Can't build address list for clean entries") + + /* Now broadcast the list of cleaned entries */ + if(MPI_SUCCESS != (mpi_result = MPI_Bcast((void *)addr_buf_ptr, (int)buf_size, MPI_BYTE, 0, aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_result) + } /* end if */ + + /* if it is defined, call the sync point done callback. Note + * that this callback is defined purely for testing purposes, + * and should be undefined under normal operating circumstances. + */ + if(aux_ptr->sync_point_done) + (aux_ptr->sync_point_done)(num_entries, addr_buf_ptr); + +done: + if(addr_buf_ptr) + addr_buf_ptr = (haddr_t *)H5MM_xfree((void *)addr_buf_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__broadcast_clean_list() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__construct_candidate_list() + * + * Purpose: In the parallel case when the metadata_write_strategy is + * H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED, process 0 uses + * this function to construct the list of cache entries to + * be flushed. This list is then propagated to the other + * caches, and then flushed in a distributed fashion. + * + * The sync_point_op parameter is used to determine the extent + * of the flush. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 3/17/10 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__construct_candidate_list(H5AC_t *cache_ptr, H5AC_aux_t *aux_ptr, + int sync_point_op) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + HDassert((sync_point_op == H5AC_SYNC_POINT_OP__FLUSH_CACHE) || (aux_ptr->mpi_rank == 0)); + HDassert(aux_ptr->d_slist_ptr != NULL); + HDassert(aux_ptr->c_slist_ptr != NULL); + HDassert(H5SL_count(aux_ptr->c_slist_ptr) == 0); + HDassert(aux_ptr->candidate_slist_ptr != NULL); + HDassert(H5SL_count(aux_ptr->candidate_slist_ptr) == 0); + HDassert((sync_point_op == H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN) || (sync_point_op == H5AC_SYNC_POINT_OP__FLUSH_CACHE)); + + switch(sync_point_op) { + case H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN: + if(H5C_construct_candidate_list__min_clean((H5C_t *)cache_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5C_construct_candidate_list__min_clean() failed.") + break; + + case H5AC_SYNC_POINT_OP__FLUSH_CACHE: + if(H5C_construct_candidate_list__clean_cache((H5C_t *)cache_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5C_construct_candidate_list__clean_cache() failed.") + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unknown sync point operation.") + break; + } /* end switch */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__construct_candidate_list() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__copy_candidate_list_to_buffer_cb + * + * Purpose: Skip list callback for building array of addresses for + * broadcasting the candidate list. + * + * Return: Return SUCCEED on success, and FAIL on failure. + * + * Programmer: Quincey Koziol, 6/12/15 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__copy_candidate_list_to_buffer_cb(void *_item, void H5_ATTR_UNUSED *_key, + void *_udata) +{ + H5AC_slist_entry_t * slist_entry_ptr = (H5AC_slist_entry_t *)_item; /* Address of item */ + H5AC_addr_list_ud_t * udata = (H5AC_addr_list_ud_t *)_udata; /* Context for callback */ + + FUNC_ENTER_STATIC_NOERR + + /* Sanity checks */ + HDassert(slist_entry_ptr); + HDassert(udata); + + /* Store the entry's address in the buffer */ + udata->addr_buf_ptr[udata->i] = slist_entry_ptr->addr; + udata->i++; + + /* now release the entry */ + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + + FUNC_LEAVE_NOAPI(SUCCEED) +} /* H5AC__copy_candidate_list_to_buffer_cb() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__copy_candidate_list_to_buffer + * + * Purpose: Allocate buffer(s) and copy the contents of the candidate + * entry slist into it (them). In passing, remove all + * entries from the candidate slist. Note that the + * candidate slist must not be empty. + * + * If MPI_Offset_buf_ptr_ptr is not NULL, allocate a buffer + * of MPI_Offset, copy the contents of the candidate + * entry list into it with the appropriate conversions, + * and return the base address of the buffer in + * *MPI_Offset_buf_ptr. Note that this is the buffer + * used by process 0 to transmit the list of entries to + * be flushed to all other processes (in this file group). + * + * Similarly, allocate a buffer of haddr_t, load the contents + * of the candidate list into this buffer, and return its + * base address in *haddr_buf_ptr_ptr. Note that this + * latter buffer is constructed unconditionally. + * + * In passing, also remove all entries from the candidate + * entry slist. + * + * Return: Return SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer, 4/19/10 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__copy_candidate_list_to_buffer(const H5AC_t *cache_ptr, int *num_entries_ptr, + haddr_t **haddr_buf_ptr_ptr) +{ + H5AC_aux_t * aux_ptr = NULL; + H5AC_addr_list_ud_t udata; + haddr_t * haddr_buf_ptr = NULL; + size_t buf_size; + int num_entries = 0; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + HDassert(aux_ptr->candidate_slist_ptr != NULL); + HDassert(H5SL_count(aux_ptr->candidate_slist_ptr) > 0); + HDassert(num_entries_ptr != NULL); + HDassert(*num_entries_ptr == 0); + HDassert(haddr_buf_ptr_ptr != NULL); + HDassert(*haddr_buf_ptr_ptr == NULL); + + num_entries = (int)H5SL_count(aux_ptr->candidate_slist_ptr); + + /* allocate a buffer(s) to store the list of candidate entry + * base addresses in + */ + buf_size = sizeof(haddr_t) * (size_t)num_entries; + if(NULL == (haddr_buf_ptr = (haddr_t *)H5MM_malloc(buf_size))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "memory allocation failed for haddr buffer") + + /* Set up user data for callback */ + udata.aux_ptr = aux_ptr; + udata.addr_buf_ptr = haddr_buf_ptr; + udata.i = 0; + + /* Free all the candidate list entries, building the address list in the callback */ + if(H5SL_free(aux_ptr->candidate_slist_ptr, H5AC__copy_candidate_list_to_buffer_cb, &udata) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFREE, FAIL, "Can't build address list for candidate entries") + + /* Pass the number of entries and the buffer pointer + * back to the caller. + */ + *num_entries_ptr = num_entries; + *haddr_buf_ptr_ptr = haddr_buf_ptr; + +done: + if(ret_value < 0) + if(haddr_buf_ptr) + haddr_buf_ptr = (haddr_t *)H5MM_xfree((void *)haddr_buf_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__copy_candidate_list_to_buffer() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__log_deleted_entry() + * + * Purpose: Log an entry which has been deleted. + * + * Only called for mpi_rank 0. We must make sure that the entry + * doesn't appear in the cleaned or dirty entry lists. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 6/29/05 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__log_deleted_entry(const H5AC_info_t *entry_ptr) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + H5AC_slist_entry_t * slist_entry_ptr = NULL; + haddr_t addr; + + FUNC_ENTER_PACKAGE_NOERR + + /* Sanity checks */ + HDassert(entry_ptr); + addr = entry_ptr->addr; + cache_ptr = entry_ptr->cache_ptr; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->mpi_rank == 0); + HDassert(aux_ptr->d_slist_ptr != NULL); + HDassert(aux_ptr->c_slist_ptr != NULL); + + /* if the entry appears in the dirtied entry slist, remove it. */ + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->d_slist_ptr, (void *)(&addr)))) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + + /* if the entry appears in the cleaned entry slist, remove it. */ + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&addr)))) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + + FUNC_LEAVE_NOAPI(SUCCEED) +} /* H5AC__log_deleted_entry() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__log_dirtied_entry() + * + * Purpose: Update the dirty_bytes count for a newly dirtied entry. + * + * If mpi_rank isn't 0, this simply means adding the size + * of the entries to the dirty_bytes count. + * + * If mpi_rank is 0, we must first check to see if the entry + * appears in the dirty entries slist. If it is, do nothing. + * If it isn't, add the size to th dirty_bytes count, add the + * entry to the dirty entries slist, and remove it from the + * cleaned list (if it is present there). + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 6/29/05 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__log_dirtied_entry(const H5AC_info_t *entry_ptr) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Sanity checks */ + HDassert(entry_ptr); + HDassert(entry_ptr->is_dirty == FALSE); + cache_ptr = entry_ptr->cache_ptr; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + + if(aux_ptr->mpi_rank == 0) { + H5AC_slist_entry_t *slist_entry_ptr; + haddr_t addr = entry_ptr->addr; + + /* Sanity checks */ + HDassert(aux_ptr->d_slist_ptr != NULL); + HDassert(aux_ptr->c_slist_ptr != NULL); + + if(NULL == H5SL_search(aux_ptr->d_slist_ptr, (void *)(&addr))) { + /* insert the address of the entry in the dirty entry list, and + * add its size to the dirty_bytes count. + */ + if(NULL == (slist_entry_ptr = H5FL_MALLOC(H5AC_slist_entry_t))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "Can't allocate dirty slist entry .") + slist_entry_ptr->addr = addr; + + if(H5SL_insert(aux_ptr->d_slist_ptr, slist_entry_ptr, &(slist_entry_ptr->addr)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTINSERT, FAIL, "can't insert entry into dirty entry slist.") + + aux_ptr->dirty_bytes += entry_ptr->size; +#if H5AC_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->unprotect_dirty_bytes += entry_ptr->size; + aux_ptr->unprotect_dirty_bytes_updates += 1; +#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ + } /* end if */ + + /* the entry is dirty. If it exists on the cleaned entries list, + * remove it. + */ + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&addr)))) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + } /* end if */ + else { + aux_ptr->dirty_bytes += entry_ptr->size; +#if H5AC_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->unprotect_dirty_bytes += entry_size; + aux_ptr->unprotect_dirty_bytes_updates += 1; +#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ + } /* end else */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__log_dirtied_entry() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__log_flushed_entry() + * + * Purpose: Update the clean entry slist for the flush of an entry -- + * specifically, if the entry has been cleared, remove it + * from both the cleaned and dirtied lists if it is present. + * Otherwise, if the entry was dirty, insert the indicated + * entry address in the clean slist if it isn't there already. + * + * This function is only used in PHDF5, and should only + * be called for the process with mpi rank 0. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 6/29/05 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__log_flushed_entry(H5C_t *cache_ptr, haddr_t addr, hbool_t was_dirty, + unsigned flags) +{ + hbool_t cleared; + H5AC_aux_t * aux_ptr; + H5AC_slist_entry_t * slist_entry_ptr = NULL; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Sanity check */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->mpi_rank == 0); + HDassert(aux_ptr->c_slist_ptr != NULL); + + /* Set local flags */ + cleared = ((flags & H5C__FLUSH_CLEAR_ONLY_FLAG) != 0); + + if(cleared) { + /* If the entry has been cleared, must remove it from both the + * cleaned list and the dirtied list. + */ + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&addr)))) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->d_slist_ptr, (void *)(&addr)))) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + } /* end if */ + else if(was_dirty) { + if(NULL == H5SL_search(aux_ptr->c_slist_ptr, (void *)(&addr))) { + /* insert the address of the entry in the clean entry list. */ + if(NULL == (slist_entry_ptr = H5FL_MALLOC(H5AC_slist_entry_t))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "Can't allocate clean slist entry .") + slist_entry_ptr->addr = addr; + + if(H5SL_insert(aux_ptr->c_slist_ptr, slist_entry_ptr, &(slist_entry_ptr->addr)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTINSERT, FAIL, "can't insert entry into clean entry slist.") + } /* end if */ + } /* end else-if */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__log_flushed_entry() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__log_inserted_entry() + * + * Purpose: Update the dirty_bytes count for a newly inserted entry. + * + * If mpi_rank isnt 0, this simply means adding the size + * of the entry to the dirty_bytes count. + * + * If mpi_rank is 0, we must also add the entry to the + * dirty entries slist. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 6/30/05 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__log_inserted_entry(const H5AC_info_t *entry_ptr) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Sanity checks */ + HDassert(entry_ptr); + cache_ptr = entry_ptr->cache_ptr; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + + if(aux_ptr->mpi_rank == 0) { + H5AC_slist_entry_t *slist_entry_ptr; + + HDassert(aux_ptr->d_slist_ptr != NULL); + HDassert(aux_ptr->c_slist_ptr != NULL); + + /* Entry to insert should not be in dirty list currently */ + if(NULL != H5SL_search(aux_ptr->d_slist_ptr, (const void *)(&entry_ptr->addr))) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Inserted entry already in dirty slist.") + + /* insert the address of the entry in the dirty entry list, and + * add its size to the dirty_bytes count. + */ + if(NULL == (slist_entry_ptr = H5FL_MALLOC(H5AC_slist_entry_t))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "Can't allocate dirty slist entry .") + slist_entry_ptr->addr = entry_ptr->addr; + if(H5SL_insert(aux_ptr->d_slist_ptr, slist_entry_ptr, &(slist_entry_ptr->addr)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTINSERT, FAIL, "can't insert entry into dirty entry slist.") + + /* Entry to insert should not be in clean list either */ + if(NULL != H5SL_search(aux_ptr->c_slist_ptr, (const void *)(&entry_ptr->addr))) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Inserted entry in clean slist.") + } /* end if */ + + aux_ptr->dirty_bytes += entry_ptr->size; + +#if H5AC_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->insert_dirty_bytes += size; + aux_ptr->insert_dirty_bytes_updates += 1; +#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__log_inserted_entry() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__log_moved_entry() + * + * Purpose: Update the dirty_bytes count for a moved entry. + * + * WARNING + * + * At present, the way that the move call is used ensures + * that the moved entry is present in all caches by + * moving in a collective operation and immediately after + * unprotecting the target entry. + * + * This function uses this invariant, and will cause arcane + * failures if it is not met. If maintaining this invariant + * becomes impossible, we will have to rework this function + * extensively, and likely include a bit of IPC for + * synchronization. A better option might be to subsume + * move in the unprotect operation. + * + * Given that the target entry is in all caches, the function + * proceeds as follows: + * + * For processes with mpi rank other 0, it simply checks to + * see if the entry was dirty prior to the move, and adds + * the entries size to the dirty bytes count. + * + * In the process with mpi rank 0, the function first checks + * to see if the entry was dirty prior to the move. If it + * was, and if the entry doesn't appear in the dirtied list + * under its old address, it adds the entry's size to the + * dirty bytes count. + * + * The rank 0 process then removes any references to the + * entry under its old address from the cleands and dirtied + * lists, and inserts an entry in the dirtied list under the + * new address. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 6/30/05 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__log_moved_entry(const H5F_t *f, haddr_t old_addr, haddr_t new_addr) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + hbool_t entry_in_cache; + hbool_t entry_dirty; + size_t entry_size; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Sanity checks */ + HDassert(f); + HDassert(f->shared); + cache_ptr = (H5AC_t *)f->shared->cache; + HDassert(cache_ptr); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + + /* get entry status, size, etc here */ + if(H5C_get_entry_status(f, old_addr, &entry_size, &entry_in_cache, + &entry_dirty, NULL, NULL, NULL, NULL, NULL) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't get entry status.") + if(!entry_in_cache) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "entry not in cache.") + + if(aux_ptr->mpi_rank == 0) { + H5AC_slist_entry_t * slist_entry_ptr; + + HDassert(aux_ptr->d_slist_ptr != NULL); + HDassert(aux_ptr->c_slist_ptr != NULL); + + /* if the entry appears in the cleaned entry slist, under its old + * address, remove it. + */ + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&old_addr)))) + slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, slist_entry_ptr); + + /* if the entry appears in the dirtied entry slist under its old + * address, remove it, but don't free it. Set addr to new_addr. + */ + if(NULL != (slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->d_slist_ptr, (void *)(&old_addr)))) + slist_entry_ptr->addr = new_addr; + else { + /* otherwise, allocate a new entry that is ready + * for insertion, and increment dirty_bytes. + * + * Note that the fact that the entry wasn't in the dirtied + * list under its old address implies that it must have + * been clean to start with. + */ + HDassert(!entry_dirty); + if(NULL == (slist_entry_ptr = H5FL_MALLOC(H5AC_slist_entry_t))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "Can't allocate dirty slist entry .") + slist_entry_ptr->addr = new_addr; + + aux_ptr->dirty_bytes += entry_size; + +#if H5AC_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->move_dirty_bytes += entry_size; + aux_ptr->move_dirty_bytes_updates += 1; +#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ + } /* end else */ + + /* insert / reinsert the entry in the dirty slist */ + if(H5SL_insert(aux_ptr->d_slist_ptr, slist_entry_ptr, &(slist_entry_ptr->addr)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTINSERT, FAIL, "can't insert entry into dirty entry slist.") + } /* end if */ + else if(!entry_dirty) { + aux_ptr->dirty_bytes += entry_size; + +#if H5AC_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->move_dirty_bytes += entry_size; + aux_ptr->move_dirty_bytes_updates += 1; +#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ + } /* end else-if */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__log_moved_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__propagate_and_apply_candidate_list + * + * Purpose: Prior to the addition of support for multiple metadata + * write strategies, in PHDF5, only the metadata cache with + * mpi rank 0 was allowed to write to file. All other + * metadata caches on processes with rank greater than 0 + * were required to retain dirty entries until they were + * notified that the entry was clean. + * + * This constraint is relaxed with the distributed + * metadata write strategy, in which a list of candidate + * metadata cache entries is constructed by the process 0 + * cache and then distributed to the caches of all the other + * processes. Once the listed is distributed, many (if not + * all) processes writing writing a unique subset of the + * entries, and marking the remainder clean. The subsets + * are chosen so that each entry in the list of candidates + * is written by exactly one cache, and all entries are + * marked as being clean in all caches. + * + * While the list of candidate cache entries is prepared + * elsewhere, this function is the main routine for distributing + * and applying the list. It must be run simultaniously on + * all processes that have the relevant file open. To ensure + * proper synchronization, there is a barrier at the beginning + * of this function. + * + * At present, this function is called under one of two + * circumstances: + * + * 1) Dirty byte creation exceeds some user specified value. + * + * While metadata reads may occur independently, all + * operations writing metadata must be collective. Thus + * all metadata caches see the same sequence of operations, + * and therefore the same dirty data creation. + * + * This fact is used to synchronize the caches for purposes + * of propagating the list of candidate entries, by simply + * calling this function from all caches whenever some user + * specified threshold on dirty data is exceeded. (the + * process 0 cache creates the candidate list just before + * calling this function). + * + * 2) Under direct user control -- this operation must be + * collective. + * + * The operations to be managed by this function are as + * follows: + * + * All processes: + * + * 1) Participate in an opening barrier. + * + * For the process with mpi rank 0: + * + * 1) Load the contents of the candidate list + * (candidate_slist_ptr) into a buffer, and broadcast that + * buffer to all the other caches. Clear the candidate + * list in passing. + * + * If there is a positive number of candidates, proceed with + * the following: + * + * 2) Apply the candidate entry list. + * + * 3) Particpate in a closing barrier. + * + * 4) Remove from the dirty list (d_slist_ptr) and from the + * flushed and still clean entries list (c_slist_ptr), + * all addresses that appeared in the candidate list, as + * these entries are now clean. + * + * + * For all processes with mpi rank greater than 0: + * + * 1) Receive the candidate entry list broadcast + * + * If there is a positive number of candidates, proceed with + * the following: + * + * 2) Apply the candidate entry list. + * + * 3) Particpate in a closing barrier. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * 3/17/10 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__propagate_and_apply_candidate_list(H5F_t *f, hid_t dxpl_id) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + haddr_t * candidates_list_ptr = NULL; + int mpi_result; + int num_candidates = 0; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + + /* to prevent "messages from the future" we must synchronize all + * processes before we write any entries. + */ + if(MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_result) + + if(aux_ptr->mpi_rank == 0) { + if(H5AC__broadcast_candidate_list(cache_ptr, &num_candidates, &candidates_list_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't broadcast candidate slist.") + + HDassert(H5SL_count(aux_ptr->candidate_slist_ptr) == 0); + } /* end if */ + else { + if(H5AC__receive_candidate_list(cache_ptr, &num_candidates, &candidates_list_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't receive candidate broadcast.") + } /* end else */ + + if(num_candidates > 0) { + herr_t result; + + /* all processes apply the candidate list. + * H5C_apply_candidate_list() handles the details of + * distributing the writes across the processes. + */ + + /* Enable writes during this operation */ + aux_ptr->write_permitted = TRUE; + + /* Apply the candidate list */ + result = H5C_apply_candidate_list(f, dxpl_id, cache_ptr, num_candidates, + candidates_list_ptr, aux_ptr->mpi_rank, aux_ptr->mpi_size); + + /* Disable writes again */ + aux_ptr->write_permitted = FALSE; + + /* Check for error on the write operation */ + if(result < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't apply candidate list.") + + /* this code exists primarily for the test bed -- it allows us to + * enforce posix semantics on the server that pretends to be a + * file system in our parallel tests. + */ + if(aux_ptr->write_done) + (aux_ptr->write_done)(); + + /* to prevent "messages from the past" we must synchronize all + * processes again before we go on. + */ + if(MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_result) + + /* if this is process zero, tidy up the dirtied, + * and flushed and still clean lists. + */ + if(aux_ptr->mpi_rank == 0) + if(H5AC__tidy_cache_0_lists(cache_ptr, num_candidates, candidates_list_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't tidy up process 0 lists.") + } /* end if */ + + /* if it is defined, call the sync point done callback. Note + * that this callback is defined purely for testing purposes, + * and should be undefined under normal operating circumstances. + */ + if(aux_ptr->sync_point_done) + (aux_ptr->sync_point_done)(num_candidates, candidates_list_ptr); + +done: + if(candidates_list_ptr) + candidates_list_ptr = (haddr_t *)H5MM_xfree((void *)candidates_list_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__propagate_and_apply_candidate_list() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__propagate_flushed_and_still_clean_entries_list + * + * Purpose: In PHDF5, if the process 0 only metadata write strategy + * is selected, only the metadata cache with mpi rank 0 is + * allowed to write to file. All other metadata caches on + * processes with rank greater than 0 must retain dirty + * entries until they are notified that the entry is now + * clean. + * + * This function is the main routine for handling this + * notification proceedure. It must be called + * simultaniously on all processes that have the relevant + * file open. To this end, it is called only during a + * sync point, with a barrier prior to the call. + * + * Note that any metadata entry writes by process 0 will + * occur after the barrier and just before this call. + * + * Typicaly, calls to this function will be triggered in + * one of two ways: + * + * 1) Dirty byte creation exceeds some user specified value. + * + * While metadata reads may occur independently, all + * operations writing metadata must be collective. Thus + * all metadata caches see the same sequence of operations, + * and therefore the same dirty data creation. + * + * This fact is used to synchronize the caches for purposes + * of propagating the list of flushed and still clean + * entries, by simply calling this function from all + * caches whenever some user specified threshold on dirty + * data is exceeded. + * + * 2) Under direct user control -- this operation must be + * collective. + * + * The operations to be managed by this function are as + * follows: + * + * For the process with mpi rank 0: + * + * 1) Load the contents of the flushed and still clean entries + * list (c_slist_ptr) into a buffer, and broadcast that + * buffer to all the other caches. + * + * 2) Clear the flushed and still clean entries list + * (c_slist_ptr). + * + * + * For all processes with mpi rank greater than 0: + * + * 1) Receive the flushed and still clean entries list broadcast + * + * 2) Mark the specified entries as clean. + * + * + * For all processes: + * + * 1) Reset the dirtied bytes count to 0. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * July 5, 2005 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__propagate_flushed_and_still_clean_entries_list(H5F_t *f, hid_t dxpl_id) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__PROCESS_0_ONLY); + + if(aux_ptr->mpi_rank == 0) { + if(H5AC__broadcast_clean_list(cache_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't broadcast clean slist.") + HDassert(H5SL_count(aux_ptr->c_slist_ptr) == 0); + } /* end if */ + else { + if(H5AC__receive_and_apply_clean_list(f, dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't receive and/or process clean slist broadcast.") + } /* end else */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__propagate_flushed_and_still_clean_entries_list() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC_receive_haddr_list() + * + * Purpose: Receive the list of entry addresses from process 0, + * and return it in a buffer pointed to by *haddr_buf_ptr_ptr. + * Note that the caller must free this buffer if it is + * returned. + * + * This function must only be called by the process with + * MPI_rank greater than 0. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: Quincey Koziol, 6/11/2015 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__receive_haddr_list(MPI_Comm mpi_comm, int *num_entries_ptr, + haddr_t **haddr_buf_ptr_ptr) +{ + haddr_t * haddr_buf_ptr = NULL; + int mpi_result; + int num_entries; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(num_entries_ptr != NULL); + HDassert(*num_entries_ptr == 0); + HDassert(haddr_buf_ptr_ptr != NULL); + HDassert(*haddr_buf_ptr_ptr == NULL); + + /* First receive the number of entries in the list so that we + * can set up a buffer to receive them. If there aren't + * any, we are done. + */ + if(MPI_SUCCESS != (mpi_result = MPI_Bcast(&num_entries, 1, MPI_INT, 0, mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_result) + + if(num_entries > 0) { + size_t buf_size; + + /* allocate buffers to store the list of entry base addresses in */ + buf_size = sizeof(haddr_t) * (size_t)num_entries; + if(NULL == (haddr_buf_ptr = (haddr_t *)H5MM_malloc(buf_size))) + HGOTO_ERROR(H5E_CACHE, H5E_CANTALLOC, FAIL, "memory allocation failed for haddr buffer") + + /* Now receive the list of candidate entries */ + if(MPI_SUCCESS != (mpi_result = MPI_Bcast((void *)haddr_buf_ptr, (int)buf_size, MPI_BYTE, 0, mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_result) + } /* end if */ + + /* finally, pass the number of entries and the buffer pointer + * back to the caller. + */ + *num_entries_ptr = num_entries; + *haddr_buf_ptr_ptr = haddr_buf_ptr; + +done: + if(ret_value < 0) + if(haddr_buf_ptr) + haddr_buf_ptr = (haddr_t *)H5MM_xfree((void *)haddr_buf_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC_receive_haddr_list() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__receive_and_apply_clean_list() + * + * Purpose: Receive the list of cleaned entries from process 0, + * and mark the specified entries as clean. + * + * This function must only be called by the process with + * MPI_rank greater than 0. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 7/4/05 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__receive_and_apply_clean_list(H5F_t *f, hid_t dxpl_id) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + haddr_t * haddr_buf_ptr = NULL; + int num_entries = 0; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity check */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->mpi_rank != 0); + + /* Retrieve the clean list from process 0 */ + if(H5AC__receive_haddr_list(aux_ptr->mpi_comm, &num_entries, &haddr_buf_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "can't receive clean list") + + if(num_entries > 0) + /* mark the indicated entries as clean */ + if(H5C_mark_entries_as_clean(f, dxpl_id, (int32_t)num_entries, haddr_buf_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't mark entries clean.") + + /* if it is defined, call the sync point done callback. Note + * that this callback is defined purely for testing purposes, + * and should be undefined under normal operating circumstances. + */ + if(aux_ptr->sync_point_done) + (aux_ptr->sync_point_done)(num_entries, haddr_buf_ptr); + +done: + if(haddr_buf_ptr) + haddr_buf_ptr = (haddr_t *)H5MM_xfree((void *)haddr_buf_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__receive_and_apply_clean_list() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC__receive_candidate_list() + * + * Purpose: Receive the list of candidate entries from process 0, + * and return it in a buffer pointed to by *haddr_buf_ptr_ptr. + * Note that the caller must free this buffer if it is + * returned. + * + * This function must only be called by the process with + * MPI_rank greater than 0. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 3/17/10 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__receive_candidate_list(const H5AC_t *cache_ptr, int *num_entries_ptr, + haddr_t **haddr_buf_ptr_ptr) +{ + H5AC_aux_t * aux_ptr; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->mpi_rank != 0); + HDassert(aux_ptr-> metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + HDassert(num_entries_ptr != NULL); + HDassert(*num_entries_ptr == 0); + HDassert(haddr_buf_ptr_ptr != NULL); + HDassert(*haddr_buf_ptr_ptr == NULL); + + /* Retrieve the candidate list from process 0 */ + if(H5AC__receive_haddr_list(aux_ptr->mpi_comm, num_entries_ptr, haddr_buf_ptr_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "can't receive clean list") + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__receive_candidate_list() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__rsp__dist_md_write__flush + * + * Purpose: Routine for handling the details of running a sync point + * that is triggered by a flush -- which in turn must have been + * triggered by either a flush API call or a file close -- + * when the distributed metadata write strategy is selected. + * + * Upon entry, each process generates it own candidate list, + * being a sorted list of all dirty metadata entries currently + * in the metadata cache. Note that this list must be idendical + * across all processes, as all processes see the same stream + * of dirty metadata coming in, and use the same lists of + * candidate entries at each sync point. (At first glance, this + * argument sounds circular, but think of it in the sense of + * a recursive proof). + * + * If this this list is empty, we are done, and the function + * returns + * + * Otherwise, after the sorted list dirty metadata entries is + * constructed, each process uses the same algorithm to assign + * each entry on the candidate list to exactly one process for + * flushing. + * + * At this point, all processes participate in a barrier to + * avoid messages from the past/future bugs. + * + * Each process then flushes the entries assigned to it, and + * marks all other entries on the candidate list as clean. + * + * Finally, all processes participate in a second barrier to + * avoid messages from the past/future bugs. + * + * At the end of this process, process 0 and only process 0 + * must tidy up its lists of dirtied and cleaned entries. + * These lists are not used in the distributed metadata write + * strategy, but they must be maintained should we shift + * to a strategy that uses them. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * April 28, 2010 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__rsp__dist_md_write__flush(H5F_t *f, hid_t dxpl_id) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + haddr_t * haddr_buf_ptr = NULL; + int mpi_result; + int num_entries = 0; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + + /* first construct the candidate list -- initially, this will be in the + * form of a skip list. We will convert it later. + */ + if(H5C_construct_candidate_list__clean_cache(cache_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate list.") + + if(H5SL_count(aux_ptr->candidate_slist_ptr) > 0) { + herr_t result; + + /* convert the candidate list into the format we + * are used to receiving from process 0. + */ + if(H5AC__copy_candidate_list_to_buffer(cache_ptr, &num_entries, &haddr_buf_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate buffer.") + + /* initial sync point barrier */ + if(MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_result) + + /* Enable writes during this operation */ + aux_ptr->write_permitted = TRUE; + + /* Apply the candidate list */ + result = H5C_apply_candidate_list(f, dxpl_id, cache_ptr, num_entries, + haddr_buf_ptr, aux_ptr->mpi_rank, aux_ptr->mpi_size); + + /* Disable writes again */ + aux_ptr->write_permitted = FALSE; + + /* Check for error on the write operation */ + if(result < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't apply candidate list.") + + /* this code exists primarily for the test bed -- it allows us to + * enforce posix semantics on the server that pretends to be a + * file system in our parallel tests. + */ + if(aux_ptr->write_done) + (aux_ptr->write_done)(); + + /* final sync point barrier */ + if(MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_result) + + /* if this is process zero, tidy up the dirtied, + * and flushed and still clean lists. + */ + if(aux_ptr->mpi_rank == 0) + if(H5AC__tidy_cache_0_lists(cache_ptr, num_entries, haddr_buf_ptr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't tidy up process 0 lists.") + } /* end if */ + + /* if it is defined, call the sync point done callback. Note + * that this callback is defined purely for testing purposes, + * and should be undefined under normal operating circumstances. + */ + if(aux_ptr->sync_point_done) + (aux_ptr->sync_point_done)(num_entries, haddr_buf_ptr); + +done: + if(haddr_buf_ptr) + haddr_buf_ptr = (haddr_t *)H5MM_xfree((void *)haddr_buf_ptr); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__rsp__dist_md_write__flush() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__rsp__dist_md_write__flush_to_min_clean + * + * Purpose: Routine for handling the details of running a sync point + * triggered by the accumulation of dirty metadata (as + * opposed to a flush call to the API) when the distributed + * metadata write strategy is selected. + * + * After invocation and initial sanity checking this function + * first checks to see if evictions are enabled -- if they + * are not, the function does nothing and returns. + * + * Otherwise, process zero constructs a list of entries to + * be flushed in order to bring the process zero cache back + * within its min clean requirement. Note that this list + * (the candidate list) may be empty. + * + * Then, all processes participate in a barrier. + * + * After the barrier, process 0 broadcasts the number of + * entries in the candidate list prepared above, and all + * other processes receive this number. + * + * If this number is zero, we are done, and the function + * returns without further action. + * + * Otherwise, process 0 broadcasts the sorted list of + * candidate entries, and all other processes receive it. + * + * Then, each process uses the same algorithm to assign + * each entry on the candidate list to exactly one process + * for flushing. + * + * Each process then flushes the entries assigned to it, and + * marks all other entries on the candidate list as clean. + * + * Finally, all processes participate in a second barrier to + * avoid messages from the past/future bugs. + * + * At the end of this process, process 0 and only process 0 + * must tidy up its lists of dirtied and cleaned entries. + * These lists are not used in the distributed metadata write + * strategy, but they must be maintained should we shift + * to a strategy that uses them. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * April 28, 2010 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__rsp__dist_md_write__flush_to_min_clean(H5F_t *f, hid_t dxpl_id) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + hbool_t evictions_enabled; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + + /* Query if evictions are allowed */ + if(H5C_get_evictions_enabled((const H5C_t *)cache_ptr, &evictions_enabled) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "H5C_get_evictions_enabled() failed.") + + if(evictions_enabled) { + /* construct candidate list -- process 0 only */ + if(aux_ptr->mpi_rank == 0) + if(H5AC__construct_candidate_list(cache_ptr, aux_ptr, H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't construct candidate list.") + + /* propagate and apply candidate list -- all processes */ + if(H5AC__propagate_and_apply_candidate_list(f, dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't propagate and apply candidate list.") + } /* evictions enabled */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__rsp__dist_md_write__flush_to_min_clean() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__rsp__p0_only__flush + * + * Purpose: Routine for handling the details of running a sync point + * that is triggered a flush -- which in turn must have been + * triggered by either a flush API call or a file close -- + * when the process 0 only metadata write strategy is selected. + * + * First, all processes participate in a barrier. + * + * Then process zero flushes all dirty entries, and broadcasts + * they number of clean entries (if any) to all the other + * caches. + * + * If this number is zero, we are done. + * + * Otherwise, process 0 broadcasts the list of cleaned + * entries, and all other processes which are part of this + * file group receive it, and mark the listed entries as + * clean in their caches. + * + * Since all processes have the same set of dirty + * entries at the beginning of the sync point, and all + * entries that will be written are written before + * process zero broadcasts the number of cleaned entries, + * there is no need for a closing barrier. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * April 28, 2010 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__rsp__p0_only__flush(H5F_t *f, hid_t dxpl_id) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + int mpi_result; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__PROCESS_0_ONLY); + + /* to prevent "messages from the future" we must + * synchronize all processes before we start the flush. + * Hence the following barrier. + */ + if(MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_result) + + /* Flush data to disk, from rank 0 process */ + if(aux_ptr->mpi_rank == 0) { + herr_t result; + + /* Enable writes during this operation */ + aux_ptr->write_permitted = TRUE; + + /* Flush the cache */ + result = H5C_flush_cache(f, dxpl_id, H5AC__NO_FLAGS_SET); + + /* Disable writes again */ + aux_ptr->write_permitted = FALSE; + + /* Check for error on the write operation */ + if(result < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't flush.") + + /* this code exists primarily for the test bed -- it allows us to + * enforce posix semantics on the server that pretends to be a + * file system in our parallel tests. + */ + if(aux_ptr->write_done) + (aux_ptr->write_done)(); + } /* end if */ + + /* Propagate cleaned entries to other ranks. */ + if(H5AC__propagate_flushed_and_still_clean_entries_list(f, H5AC_dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't propagate clean entries list.") + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__rsp__p0_only__flush() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__rsp__p0_only__flush_to_min_clean + * + * Purpose: Routine for handling the details of running a sync point + * triggered by the accumulation of dirty metadata (as + * opposed to a flush call to the API) when the process 0 + * only metadata write strategy is selected. + * + * After invocation and initial sanity checking this function + * first checks to see if evictions are enabled -- if they + * are not, the function does nothing and returns. + * + * Otherwise, all processes participate in a barrier. + * + * After the barrier, if this is process 0, the function + * causes the cache to flush sufficient entries to get the + * cache back within its minimum clean fraction, and broadcast + * the number of entries which have been flushed since + * the last sync point, and are still clean. + * + * If this number is zero, we are done. + * + * Otherwise, process 0 broadcasts the list of cleaned + * entries, and all other processes which are part of this + * file group receive it, and mark the listed entries as + * clean in their caches. + * + * Since all processes have the same set of dirty + * entries at the beginning of the sync point, and all + * entries that will be written are written before + * process zero broadcasts the number of cleaned entries, + * there is no need for a closing barrier. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * April 28, 2010 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__rsp__p0_only__flush_to_min_clean(H5F_t *f, hid_t dxpl_id) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + hbool_t evictions_enabled; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_STATIC + + /* Sanity checks */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__PROCESS_0_ONLY); + + /* Query if evictions are allowed */ + if(H5C_get_evictions_enabled((const H5C_t *)cache_ptr, &evictions_enabled) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "H5C_get_evictions_enabled() failed.") + + /* Flush if evictions are allowed -- following call + * will cause process 0 to flush to min clean size, + * and then propagate the newly clean entries to the + * other processes. + * + * Otherwise, do nothing. + */ + if(evictions_enabled) { + int mpi_result; + + /* to prevent "messages from the future" we must synchronize all + * processes before we start the flush. This synchronization may + * already be done -- hence the do_barrier parameter. + */ + if(MPI_SUCCESS != (mpi_result = MPI_Barrier(aux_ptr->mpi_comm))) + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_result) + + if(0 == aux_ptr->mpi_rank) { + herr_t result; + + /* here, process 0 flushes as many entries as necessary to + * comply with the currently specified min clean size. + * Note that it is quite possible that no entries will be + * flushed. + */ + + /* Enable writes during this operation */ + aux_ptr->write_permitted = TRUE; + + /* Flush the cache */ + result = H5C_flush_to_min_clean(f, dxpl_id); + + /* Disable writes again */ + aux_ptr->write_permitted = FALSE; + + /* Check for error on the write operation */ + if(result < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5C_flush_to_min_clean() failed.") + + /* this call exists primarily for the test code -- it is used + * to enforce POSIX semantics on the process used to simulate + * reads and writes in t_cache.c. + */ + if(aux_ptr->write_done) + (aux_ptr->write_done)(); + } /* end if */ + + if(H5AC__propagate_flushed_and_still_clean_entries_list(f, dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't propagate clean entries list.") + } /* end if */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__rsp__p0_only__flush_to_min_clean() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__run_sync_point + * + * Purpose: Top level routine for managing a sync point between all + * meta data caches in the parallel case. Since all caches + * see the same sequence of dirty metadata, we simply count + * bytes of dirty metadata, and run a sync point whenever the + * number of dirty bytes of metadata seen since the last + * sync point exceeds a threshold that is common across all + * processes. We also run sync points in response to + * HDF5 API calls triggering either a flush or a file close. + * + * In earlier versions of PHDF5, only the metadata cache with + * mpi rank 0 was allowed to write to file. All other + * metadata caches on processes with rank greater than 0 were + * required to retain dirty entries until they were notified + * that the entry is was clean. + * + * This function was created to make it easier for us to + * experiment with other options, as it is a single point + * for the execution of sync points. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * March 11, 2010 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__run_sync_point(H5F_t *f, hid_t dxpl_id, int sync_point_op) +{ + H5AC_t * cache_ptr; + H5AC_aux_t * aux_ptr; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Sanity checks */ + HDassert(f != NULL); + cache_ptr = f->shared->cache; + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert((sync_point_op == H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN) || + (sync_point_op == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED)); + +#if H5AC_DEBUG_DIRTY_BYTES_CREATION +HDfprintf(stdout, "%d:H5AC_propagate...:%u: (u/uu/i/iu/r/ru) = %zu/%u/%zu/%u/%zu/%u\n", + aux_ptr->mpi_rank, + aux_ptr->dirty_bytes_propagations, + aux_ptr->unprotect_dirty_bytes, + aux_ptr->unprotect_dirty_bytes_updates, + aux_ptr->insert_dirty_bytes, + aux_ptr->insert_dirty_bytes_updates, + aux_ptr->rename_dirty_bytes, + aux_ptr->rename_dirty_bytes_updates); +#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ + + switch(aux_ptr->metadata_write_strategy) { + case H5AC_METADATA_WRITE_STRATEGY__PROCESS_0_ONLY: + switch(sync_point_op) { + case H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN: + if(H5AC__rsp__p0_only__flush_to_min_clean(f, dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "H5AC__rsp__p0_only__flush_to_min_clean() failed.") + break; + + case H5AC_SYNC_POINT_OP__FLUSH_CACHE: + if(H5AC__rsp__p0_only__flush(f, dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "H5AC__rsp__p0_only__flush() failed.") + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unknown flush op"); + break; + } /* end switch */ + break; + + case H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED: + switch(sync_point_op) { + case H5AC_SYNC_POINT_OP__FLUSH_TO_MIN_CLEAN: + if(H5AC__rsp__dist_md_write__flush_to_min_clean(f, dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "H5AC__rsp__dist_md_write__flush_to_min_clean() failed.") + break; + + case H5AC_SYNC_POINT_OP__FLUSH_CACHE: + if(H5AC__rsp__dist_md_write__flush(f, dxpl_id) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTGET, FAIL, "H5AC__rsp__dist_md_write__flush() failed.") + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unknown flush op"); + break; + } /* end switch */ + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Unknown metadata write strategy.") + break; + } /* end switch */ + + /* reset the dirty bytes count */ + aux_ptr->dirty_bytes = 0; + +#if H5AC_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->dirty_bytes_propagations += 1; + aux_ptr->unprotect_dirty_bytes = 0; + aux_ptr->unprotect_dirty_bytes_updates = 0; + aux_ptr->insert_dirty_bytes = 0; + aux_ptr->insert_dirty_bytes_updates = 0; + aux_ptr->rename_dirty_bytes = 0; + aux_ptr->rename_dirty_bytes_updates = 0; +#endif /* H5AC_DEBUG_DIRTY_BYTES_CREATION */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__run_sync_point() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__tidy_cache_0_lists() + * + * Purpose: In the distributed metadata write strategy, not all dirty + * entries are written by process 0 -- thus we must tidy + * up the dirtied, and flushed and still clean lists + * maintained by process zero after each sync point. + * + * This procedure exists to tend to this issue. + * + * At this point, all entries that process 0 cleared should + * have been removed from both the dirty and flushed and + * still clean lists, and entries that process 0 has flushed + * should have been removed from the dirtied list and added + * to the flushed and still clean list. + * + * However, since the distributed metadata write strategy + * doesn't make use of these lists, the objective is simply + * to maintain these lists in consistent state that allows + * them to be used should the metadata write strategy change + * to one that uses these lists. + * + * Thus for our purposes, all we need to do is remove from + * the dirtied and flushed and still clean lists all + * references to entries that appear in the candidate list. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: John Mainzer + * 4/20/10 + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC__tidy_cache_0_lists(H5AC_t *cache_ptr, int num_candidates, + haddr_t *candidates_list_ptr) +{ + H5AC_aux_t * aux_ptr; + int i; + + FUNC_ENTER_STATIC_NOERR + + /* Sanity checks */ + HDassert(cache_ptr != NULL); + HDassert(cache_ptr->magic == H5C__H5C_T_MAGIC); + aux_ptr = (H5AC_aux_t *)(cache_ptr->aux_ptr); + HDassert(aux_ptr != NULL); + HDassert(aux_ptr->magic == H5AC__H5AC_AUX_T_MAGIC); + HDassert(aux_ptr->metadata_write_strategy == H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED); + HDassert(aux_ptr->mpi_rank == 0); + HDassert(num_candidates > 0); + HDassert(candidates_list_ptr != NULL); + + /* clean up dirtied and flushed and still clean lists by removing + * all entries on the candidate list. Cleared entries should + * have been removed from both the dirty and cleaned lists at + * this point, flushed entries should have been added to the + * cleaned list. However, for this metadata write strategy, + * we just want to remove all references to the candidate entries. + */ + for(i = 0; i < num_candidates; i++) { + H5AC_slist_entry_t * d_slist_entry_ptr; + H5AC_slist_entry_t * c_slist_entry_ptr; + haddr_t addr; + + addr = candidates_list_ptr[i]; + + /* addr may be either on the dirtied list, or on the flushed + * and still clean list. Remove it. + */ + if(NULL != (d_slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->d_slist_ptr, (void *)&addr))) + d_slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, d_slist_entry_ptr); + if(NULL != (c_slist_entry_ptr = (H5AC_slist_entry_t *)H5SL_remove(aux_ptr->c_slist_ptr, (void *)&addr))) + c_slist_entry_ptr = H5FL_FREE(H5AC_slist_entry_t, c_slist_entry_ptr); + } /* end for */ + + FUNC_LEAVE_NOAPI(SUCCEED) +} /* H5AC__tidy_cache_0_lists() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC__flush_entries + * + * Purpose: Flush the metadata cache associated with the specified file, + * only writing from rank 0, but propagating the cleaned entries + * to all ranks. + * + * Return: Non-negative on success/Negative on failure if there was a + * request to flush all items and something was protected. + * + * Programmer: Quincey Koziol + * koziol@hdfgroup.org + * Aug 22 2009 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC__flush_entries(H5F_t *f, hid_t dxpl_id) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Sanity checks */ + HDassert(f); + HDassert(f->shared->cache); + + /* Check if we have >1 ranks */ + if(f->shared->cache->aux_ptr) + if(H5AC__run_sync_point(f, dxpl_id, H5AC_SYNC_POINT_OP__FLUSH_CACHE) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't run sync point.") + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5AC__flush_entries() */ +#endif /* H5_HAVE_PARALLEL */ + diff --git a/src/H5Cmpio.c b/src/H5Cmpio.c new file mode 100644 index 0000000..ec34285 --- /dev/null +++ b/src/H5Cmpio.c @@ -0,0 +1,1220 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * Copyright by The HDF Group. * + * Copyright by the Board of Trustees of the University of Illinois. * + * All rights reserved. * + * * + * This file is part of HDF5. The full HDF5 copyright notice, including * + * terms governing use, modification, and redistribution, is contained in * + * the files COPYING and Copyright.html. COPYING can be found at the root * + * of the source code distribution tree; Copyright.html can be found at the * + * root level of an installed copy of the electronic HDF5 document set and * + * is linked from the top-level documents page. It can also be found at * + * http://hdfgroup.org/HDF5/doc/Copyright.html. If you do not have * + * access to either file, you may request a copy from help@hdfgroup.org. * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +/*------------------------------------------------------------------------- + * + * Created: H5Cmpio.c + * June 20 2015 + * Quincey Koziol + * + * Purpose: Functions in this file implement support for parallel I/O for + * generic cache code. + * + *------------------------------------------------------------------------- + */ + +/****************/ +/* Module Setup */ +/****************/ + +#define H5C_PACKAGE /*suppress error about including H5Cpkg */ +#define H5F_PACKAGE /*suppress error about including H5Fpkg */ + + +/***********/ +/* Headers */ +/***********/ +#include "H5private.h" /* Generic Functions */ +#include "H5ACprivate.h" /* Metadata cache */ +#include "H5Cpkg.h" /* Cache */ +#include "H5Eprivate.h" /* Error handling */ +#include "H5Fpkg.h" /* Files */ +#include "H5Iprivate.h" /* IDs */ +#include "H5MMprivate.h" /* Memory management */ + + +#ifdef H5_HAVE_PARALLEL + +/****************/ +/* Local Macros */ +/****************/ +#define H5C_APPLY_CANDIDATE_LIST__DEBUG 0 + + +/******************/ +/* Local Typedefs */ +/******************/ + + +/********************/ +/* Local Prototypes */ +/********************/ + + +/*********************/ +/* Package Variables */ +/*********************/ + + +/*****************************/ +/* Library Private Variables */ +/*****************************/ + + +/*******************/ +/* Local Variables */ +/*******************/ + + + +/*------------------------------------------------------------------------- + * Function: H5C_apply_candidate_list + * + * Purpose: Apply the supplied candidate list. + * + * We used to do this by simply having each process write + * every mpi_size-th entry in the candidate list, starting + * at index mpi_rank, and mark all the others clean. + * + * However, this can cause unnecessary contention in a file + * system by increasing the number of processes writing to + * adjacent locations in the HDF5 file. + * + * To attempt to minimize this, we now arange matters such + * that each process writes n adjacent entries in the + * candidate list, and marks all others clean. We must do + * this in such a fashion as to guarantee that each entry + * on the candidate list is written by exactly one process, + * and marked clean by all others. + * + * To do this, first construct a table mapping mpi_rank + * to the index of the first entry in the candidate list to + * be written by the process of that mpi_rank, and then use + * the table to control which entries are written and which + * are marked as clean as a function of the mpi_rank. + * + * Note that the table must be identical on all processes, as + * all see the same candidate list, mpi_size, and mpi_rank -- + * the inputs used to construct the table. + * + * We construct the table as follows. Let: + * + * n = num_candidates / mpi_size; + * + * m = num_candidates % mpi_size; + * + * Now allocate an array of integers of length mpi_size + 1, + * and call this array candidate_assignment_table. + * + * Conceptually, if the number of candidates is a multiple + * of the mpi_size, we simply pass through the candidate list + * and assign n entries to each process to flush, with the + * index of the first entry to flush in the location in + * the candidate_assignment_table indicated by the mpi_rank + * of the process. + * + * In the more common case in which the candidate list isn't + * isn't a multiple of the mpi_size, we pretend it is, and + * give num_candidates % mpi_size processes one extra entry + * each to make things work out. + * + * Once the table is constructed, we determine the first and + * last entry this process is to flush as follows: + * + * first_entry_to_flush = candidate_assignment_table[mpi_rank] + * + * last_entry_to_flush = + * candidate_assignment_table[mpi_rank + 1] - 1; + * + * With these values determined, we simply scan through the + * candidate list, marking all entries in the range + * [first_entry_to_flush, last_entry_to_flush] for flush, + * and all others to be cleaned. + * + * Finally, we scan the LRU from tail to head, flushing + * or marking clean the candidate entries as indicated. + * If necessary, we scan the pinned list as well. + * + * Note that this function will fail if any protected or + * clean entries appear on the candidate list. + * + * This function is used in managing sync points, and + * shouldn't be used elsewhere. + * + * Return: Success: SUCCEED + * + * Failure: FAIL + * + * Programmer: John Mainzer + * 3/17/10 + * + * Changes: Ported code to detect next entry status changes as the + * the result of a flush from the serial code in the scan of + * the LRU. Also added code to detect and adapt to the + * removal from the cache of the next entry in the scan of + * the LRU. + * + * Note that at present, all of these changes should not + * be required as the operations on entries as they are + * flushed that can cause these condiditions are not premitted + * in the parallel case. However, Quincey indicates that + * this may change, and thus has requested the modification. + * + * Note the assert(FALSE) in the if statement whose body + * restarts the scan of the LRU. As the body of the if + * statement should be unreachable, it should never be + * triggered until the constraints on the parallel case + * are relaxed. Please remove the assertion at that time. + * + * Also added warning on the Pinned Entry List scan, as it + * is potentially subject to the same issue. As there is + * no cognate of this scan in the serial code, I don't have + * a fix to port to it. + * + * JRM -- 4/10/19 + * + *------------------------------------------------------------------------- + */ +herr_t +H5C_apply_candidate_list(H5F_t * f, + hid_t dxpl_id, + H5C_t * cache_ptr, + int num_candidates, + haddr_t * candidates_list_ptr, + int mpi_rank, + int mpi_size) +{ + hbool_t restart_scan; + hbool_t prev_is_dirty; + int i; + int m; + int n; + int first_entry_to_flush; + int last_entry_to_flush; + int entries_to_clear = 0; + int entries_to_flush = 0; + int entries_to_flush_or_clear_last = 0; + int entries_to_flush_collectively = 0; + int entries_cleared = 0; + int entries_flushed = 0; + int entries_delayed = 0; + int entries_flushed_or_cleared_last = 0; + int entries_flushed_collectively = 0; + int entries_examined = 0; + int initial_list_len; + int * candidate_assignment_table = NULL; + haddr_t addr; + H5C_cache_entry_t * clear_ptr = NULL; + H5C_cache_entry_t * next_ptr = NULL; + H5C_cache_entry_t * entry_ptr = NULL; + H5C_cache_entry_t * flush_ptr = NULL; + H5C_cache_entry_t * delayed_ptr = NULL; +#if H5C_DO_SANITY_CHECKS + haddr_t last_addr; +#endif /* H5C_DO_SANITY_CHECKS */ +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + char tbl_buf[1024]; +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C__H5C_T_MAGIC ); + HDassert( num_candidates > 0 ); + HDassert( num_candidates <= cache_ptr->slist_len ); + HDassert( candidates_list_ptr != NULL ); + HDassert( 0 <= mpi_rank ); + HDassert( mpi_rank < mpi_size ); + +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + HDfprintf(stdout, "%s:%d: setting up candidate assignment table.\n", + FUNC, mpi_rank); + for ( i = 0; i < 1024; i++ ) tbl_buf[i] = '\0'; + sprintf(&(tbl_buf[0]), "candidate list = "); + for ( i = 0; i < num_candidates; i++ ) + { + sprintf(&(tbl_buf[HDstrlen(tbl_buf)]), " 0x%llx", + (long long)(*(candidates_list_ptr + i))); + } + sprintf(&(tbl_buf[HDstrlen(tbl_buf)]), "\n"); + HDfprintf(stdout, "%s", tbl_buf); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + n = num_candidates / mpi_size; + m = num_candidates % mpi_size; + HDassert(n >= 0); + + if(NULL == (candidate_assignment_table = (int *)H5MM_malloc(sizeof(int) * (size_t)(mpi_size + 1)))) + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "memory allocation failed for candidate assignment table") + + candidate_assignment_table[0] = 0; + candidate_assignment_table[mpi_size] = num_candidates; + + if(m == 0) { /* mpi_size is an even divisor of num_candidates */ + HDassert(n > 0); + for(i = 1; i < mpi_size; i++) + candidate_assignment_table[i] = candidate_assignment_table[i - 1] + n; + } /* end if */ + else { + for(i = 1; i <= m; i++) + candidate_assignment_table[i] = candidate_assignment_table[i - 1] + n + 1; + + if(num_candidates < mpi_size) { + for(i = m + 1; i < mpi_size; i++) + candidate_assignment_table[i] = num_candidates; + } /* end if */ + else { + for(i = m + 1; i < mpi_size; i++) + candidate_assignment_table[i] = candidate_assignment_table[i - 1] + n; + } /* end else */ + } /* end else */ + HDassert((candidate_assignment_table[mpi_size - 1] + n) == num_candidates); + +#if H5C_DO_SANITY_CHECKS + /* verify that the candidate assignment table has the expected form */ + for ( i = 1; i < mpi_size - 1; i++ ) + { + int a, b; + + a = candidate_assignment_table[i] - candidate_assignment_table[i - 1]; + b = candidate_assignment_table[i + 1] - candidate_assignment_table[i]; + + HDassert( n + 1 >= a ); + HDassert( a >= b ); + HDassert( b >= n ); + } +#endif /* H5C_DO_SANITY_CHECKS */ + + first_entry_to_flush = candidate_assignment_table[mpi_rank]; + last_entry_to_flush = candidate_assignment_table[mpi_rank + 1] - 1; + +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + for ( i = 0; i < 1024; i++ ) + tbl_buf[i] = '\0'; + sprintf(&(tbl_buf[0]), "candidate assignment table = "); + for(i = 0; i <= mpi_size; i++) + sprintf(&(tbl_buf[HDstrlen(tbl_buf)]), " %d", candidate_assignment_table[i]); + sprintf(&(tbl_buf[HDstrlen(tbl_buf)]), "\n"); + HDfprintf(stdout, "%s", tbl_buf); + + HDfprintf(stdout, "%s:%d: flush entries [%d, %d].\n", + FUNC, mpi_rank, first_entry_to_flush, last_entry_to_flush); + + HDfprintf(stdout, "%s:%d: marking entries.\n", FUNC, mpi_rank); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + for(i = 0; i < num_candidates; i++) { + addr = candidates_list_ptr[i]; + HDassert( H5F_addr_defined(addr) ); + +#if H5C_DO_SANITY_CHECKS + if ( i > 0 ) { + if ( last_addr == addr ) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Duplicate entry in cleaned list.\n") + } else if ( last_addr > addr ) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "candidate list not sorted.\n") + } + } + + last_addr = addr; +#endif /* H5C_DO_SANITY_CHECKS */ + + H5C__SEARCH_INDEX(cache_ptr, addr, entry_ptr, FAIL) + if(entry_ptr == NULL) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Listed candidate entry not in cache?!?!?.") + } else if(!entry_ptr->is_dirty) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Listed entry not dirty?!?!?.") + } else if ( entry_ptr->is_protected ) { + /* For now at least, we can't deal with protected entries. + * If we encounter one, scream and die. If it becomes an + * issue, we should be able to work around this. + */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Listed entry is protected?!?!?.") + } else { + /* determine whether the entry is to be cleared or flushed, + * and mark it accordingly. We will scan the protected and + * pinned list shortly, and clear or flush according to these + * markings. + */ + if((i >= first_entry_to_flush) && (i <= last_entry_to_flush)) { + entries_to_flush++; + entry_ptr->flush_immediately = TRUE; + } /* end if */ + else { + entries_to_clear++; + entry_ptr->clear_on_unprotect = TRUE; + } /* end else */ + } /* end else */ + } /* end for */ + +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + HDfprintf(stdout, "%s:%d: num candidates/to clear/to flush = %d/%d/%d.\n", + FUNC, mpi_rank, (int)num_candidates, (int)entries_to_clear, + (int)entries_to_flush); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + + /* We have now marked all the entries on the candidate list for + * either flush or clear -- now scan the LRU and the pinned list + * for these entries and do the deed. + * + * Note that we are doing things in this round about manner so as + * to preserve the order of the LRU list to the best of our ability. + * If we don't do this, my experiments indicate that we will have a + * noticably poorer hit ratio as a result. + */ + +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + HDfprintf(stdout, "%s:%d: scanning LRU list. len = %d.\n", FUNC, mpi_rank, + (int)(cache_ptr->LRU_list_len)); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + /* ===================================================================== * + * Now scan the LRU and PEL lists, flushing or clearing entries as + * needed. + * + * The flush_me_last and flush_me_collectively flags may dictate how or + * when some entries can be flushed, and should be addressed here. + * However, in their initial implementation, these flags only apply to the + * superblock, so there's only a relatively small change to this function + * to account for this one case where they come into play. If these flags + * are ever expanded upon, this function and the following flushing steps + * should be reworked to account for additional cases. + * ===================================================================== */ + + HDassert(entries_to_flush >= 0); + + restart_scan = FALSE; + entries_examined = 0; + initial_list_len = cache_ptr->LRU_list_len; + entry_ptr = cache_ptr->LRU_tail_ptr; + + /* Examine each entry in the LRU list */ + while ( ( entry_ptr != NULL ) + && + ( entries_examined <= (entries_to_flush + 1) * initial_list_len ) + && + ( (entries_cleared + entries_flushed) < num_candidates ) ) { + + if ( entry_ptr->prev != NULL ) + prev_is_dirty = entry_ptr->prev->is_dirty; + + /* If this process needs to clear this entry. */ + if(entry_ptr->clear_on_unprotect) { + + HDassert(entry_ptr->is_dirty); + + next_ptr = entry_ptr->next; + entry_ptr->clear_on_unprotect = FALSE; + clear_ptr = entry_ptr; + entry_ptr = entry_ptr->prev; + entries_cleared++; + +#if ( H5C_APPLY_CANDIDATE_LIST__DEBUG > 1 ) + HDfprintf(stdout, "%s:%d: clearing 0x%llx.\n", FUNC, mpi_rank, + (long long)clear_ptr->addr); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + /* No need to check for the next entry in the scan being + * removed from the cache, as this call to H5C__flush_single_entry() + * will not call either the pre_serialize or serialize callbacks. + */ + + if(H5C__flush_single_entry(f, dxpl_id, clear_ptr, H5C__FLUSH_CLEAR_ONLY_FLAG | H5C__DEL_FROM_SLIST_ON_DESTROY_FLAG, NULL) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } /* end if */ + + /* Else, if this process needs to flush this entry. */ + else if (entry_ptr->flush_immediately) { + + HDassert(entry_ptr->is_dirty); + + next_ptr = entry_ptr->next; + entry_ptr->flush_immediately = FALSE; + flush_ptr = entry_ptr; + entry_ptr = entry_ptr->prev; + entries_flushed++; + +#if ( H5C_APPLY_CANDIDATE_LIST__DEBUG > 1 ) + HDfprintf(stdout, "%s:%d: flushing 0x%llx.\n", FUNC, mpi_rank, + (long long)flush_ptr->addr); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + /* reset entries_removed_counter and + * last_entry_removed_ptr prior to the call to + * H5C__flush_single_entry() so that we can spot + * unexpected removals of entries from the cache, + * and set the restart_scan flag if proceeding + * would be likely to cause us to scan an entry + * that is no longer in the cache. + * + * Note that as of this writing (April 2015) this + * case cannot occur in the parallel case. However + * Quincey is making noises about changing this, hence + * the insertion of this test. + * + * Note also that there is no test code to verify + * that this code actually works (although similar code + * in the serial version exists and is tested). + * + * Implementing a test will likely require implementing + * flush op like facilities in the parallel tests. At + * a guess this will not be terribly painful, but it + * will take a bit of time. + */ + cache_ptr->entries_removed_counter = 0; + cache_ptr->last_entry_removed_ptr = NULL; + + if(H5C__flush_single_entry(f, dxpl_id, flush_ptr, H5C__DEL_FROM_SLIST_ON_DESTROY_FLAG, NULL) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't flush entry.") + + if ( ( cache_ptr->entries_removed_counter > 1 ) || + ( cache_ptr->last_entry_removed_ptr == entry_ptr ) ) + + restart_scan = TRUE; + + } /* end else-if */ + + /* Otherwise, no action to be taken on this entry. Grab the next. */ + else { + entry_ptr = entry_ptr->prev; + + if ( entry_ptr != NULL ) + next_ptr = entry_ptr->next; + + } /* end else */ + + if ( ( entry_ptr != NULL ) + && + ( ( restart_scan ) + || + ( entry_ptr->is_dirty != prev_is_dirty ) + || + ( entry_ptr->next != next_ptr ) + || + ( entry_ptr->is_protected ) + || + ( entry_ptr->is_pinned ) + ) + ) { + + /* something has happened to the LRU -- start over + * from the tail. + * + * Recall that this code should be un-reachable at present, + * as all the operations by entries on flush that could cause + * it to be reachable are disallowed in the parallel case at + * present. Hence the following assertion which should be + * removed if the above changes. + */ + + HDassert( ! restart_scan ); + HDassert( entry_ptr->is_dirty == prev_is_dirty ); + HDassert( entry_ptr->next == next_ptr ); + HDassert( ! entry_ptr->is_protected ); + HDassert( ! entry_ptr->is_pinned ); + + HDassert(FALSE); /* see comment above */ + + restart_scan = FALSE; + entry_ptr = cache_ptr->LRU_tail_ptr; +/* + H5C__UPDATE_STATS_FOR_LRU_SCAN_RESTART(cache_ptr) +*/ + } + + entries_examined++; + } /* end while */ + +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + HDfprintf(stdout, "%s:%d: entries examined/cleared/flushed = %d/%d/%d.\n", + FUNC, mpi_rank, entries_examined, + entries_cleared, entries_flushed); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + /* It is also possible that some of the cleared entries are on the + * pinned list. Must scan that also. + * + * WARNING: + * + * As we now allow unpinning, and removal of other entries as a side + * effect of flushing an entry, it is possible that the next entry + * in a PEL scan could either be no longer pinned, or no longer in + * the cache by the time we get to it. + * + * At present, this is not possible in this case, as we disallow such + * operations in the parallel version of the library. However, Quincey + * has been making noises about relaxing this. If and when he does, + * we have a potential problem here. + * + * The same issue exists in the serial cache, and there are tests + * to detect this problem when it occurs, and adjust to it. As seen + * above in the LRU scan, I have ported such tests to the parallel + * code where a close cognate exists in the serial code. + * + * I haven't done so here, as there are no PEL scans where the problem + * can occur in the serial code. Needless to say, this will have to + * be repaired if the constraints on pre_serialize and serialize + * callbacks are relaxed in the parallel version of the metadata cache. + * + * JRM -- 4/1/15 + */ + +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + HDfprintf(stdout, "%s:%d: scanning pinned entry list. len = %d\n", + FUNC, mpi_rank, (int)(cache_ptr->pel_len)); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + entry_ptr = cache_ptr->pel_head_ptr; + while((entry_ptr != NULL) && + ((entries_cleared + entries_flushed + entries_delayed) + < num_candidates)) { + + /* If entry is marked for flush or for clear */ + if((entry_ptr->clear_on_unprotect||entry_ptr->flush_immediately)) { + + /* If this entry needs to be flushed last */ + if (entry_ptr->flush_me_last) { + + /* At this time, only the superblock supports being + flushed last. Conveniently, it also happens to be the only + entry that supports being flushed collectively, as well. Also + conveniently, it's always pinned, so we only need to check + for it while scanning the PEL here. Finally, it's never + included in a candidate list that excludes other dirty + entries in a cache, so we can handle this relatively simple + case here. + + For now, this function asserts this and saves the entry + to flush it after scanning the rest of the PEL list. + + If there are ever more entries that either need to be + flushed last and/or flushed collectively, this whole routine + will need to be reworked to handle all additional cases. As + it is the simple case of a single pinned entry needing + flushed last and collectively is just a minor addition to + this routine, but signficantly buffing up the usage of + flush_me_last or flush_me_collectively will require a more + intense rework of this function and potentially the function + of candidate lists as a whole. */ + + HDassert(entry_ptr->flush_me_collectively); + entries_to_flush_or_clear_last++; + entries_to_flush_collectively++; + HDassert(entries_to_flush_or_clear_last == 1); + HDassert(entries_to_flush_collectively == 1); + + /* Delay the entry. It will be flushed later. */ + delayed_ptr = entry_ptr; + entries_delayed++; + HDassert(entries_delayed == 1); + + } /* end if */ + + /* Else, this process needs to clear this entry. */ + else if (entry_ptr->clear_on_unprotect) { + HDassert(!entry_ptr->flush_immediately); + entry_ptr->clear_on_unprotect = FALSE; + clear_ptr = entry_ptr; + entry_ptr = entry_ptr->next; + entries_cleared++; + +#if ( H5C_APPLY_CANDIDATE_LIST__DEBUG > 1 ) + HDfprintf(stdout, "%s:%d: clearing 0x%llx.\n", FUNC, mpi_rank, + (long long)clear_ptr->addr); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + if(H5C__flush_single_entry(f, dxpl_id, clear_ptr, H5C__FLUSH_CLEAR_ONLY_FLAG | H5C__DEL_FROM_SLIST_ON_DESTROY_FLAG, NULL) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } /* end else-if */ + + /* Else, if this process needs to independently flush this entry. */ + else if (entry_ptr->flush_immediately) { + entry_ptr->flush_immediately = FALSE; + flush_ptr = entry_ptr; + entry_ptr = entry_ptr->next; + entries_flushed++; + +#if ( H5C_APPLY_CANDIDATE_LIST__DEBUG > 1 ) + HDfprintf(stdout, "%s:%d: flushing 0x%llx.\n", FUNC, mpi_rank, + (long long)flush_ptr->addr); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + if(H5C__flush_single_entry(f, dxpl_id, flush_ptr, H5C__DEL_FROM_SLIST_ON_DESTROY_FLAG, NULL) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } /* end else-if */ + } /* end if */ + + /* Otherwise, this entry is not marked for flush or clear. Grab the next. */ + else { + entry_ptr = entry_ptr->next; + } /* end else */ + + } /* end while */ + +#if H5C_APPLY_CANDIDATE_LIST__DEBUG + HDfprintf(stdout, + "%s:%d: pel entries examined/cleared/flushed = %d/%d/%d.\n", + FUNC, mpi_rank, entries_examined, + entries_cleared, entries_flushed); + HDfprintf(stdout, "%s:%d: done.\n", FUNC, mpi_rank); + + HDfsync(stdout); +#endif /* H5C_APPLY_CANDIDATE_LIST__DEBUG */ + + /* ====================================================================== * + * Now, handle all delayed entries. * + * * + * This can *only* be the superblock at this time, so it's relatively * + * easy to deal with. We're collectively flushing the entry saved from * + * above. This will need to be handled differently if there are ever more * + * than one entry needing this special treatment.) * + * ====================================================================== */ + + if (delayed_ptr) { + + if (delayed_ptr->clear_on_unprotect) { + entry_ptr->clear_on_unprotect = FALSE; + entries_cleared++; + } else if (delayed_ptr->flush_immediately) { + entry_ptr->flush_immediately = FALSE; + entries_flushed++; + } /* end if */ + + if(H5C__flush_single_entry(f, dxpl_id, delayed_ptr, H5C__DEL_FROM_SLIST_ON_DESTROY_FLAG, NULL) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't flush entry collectively.") + + entries_flushed_collectively++; + entries_flushed_or_cleared_last++; + } /* end if */ + + /* ====================================================================== * + * Finished flushing everything. * + * ====================================================================== */ + + HDassert((entries_flushed == entries_to_flush)); + HDassert((entries_cleared == entries_to_clear)); + HDassert((entries_flushed_or_cleared_last == entries_to_flush_or_clear_last)); + HDassert((entries_flushed_collectively == entries_to_flush_collectively)); + + if((entries_flushed != entries_to_flush) || + (entries_cleared != entries_to_clear) || + (entries_flushed_or_cleared_last != entries_to_flush_or_clear_last) || + (entries_flushed_collectively != entries_to_flush_collectively)) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "entry count mismatch.") + +done: + if(candidate_assignment_table != NULL) + candidate_assignment_table = (int *)H5MM_xfree((void *)candidate_assignment_table); + + FUNC_LEAVE_NOAPI(ret_value) +} /* H5C_apply_candidate_list() */ + + +/*------------------------------------------------------------------------- + * Function: H5C_construct_candidate_list__clean_cache + * + * Purpose: Construct the list of entries that should be flushed to + * clean all entries in the cache. + * + * This function is used in managing sync points, and + * shouldn't be used elsewhere. + * + * Return: Success: SUCCEED + * + * Failure: FAIL + * + * Programmer: John Mainzer + * 3/17/10 + * + *------------------------------------------------------------------------- + */ +herr_t +H5C_construct_candidate_list__clean_cache(H5C_t * cache_ptr) +{ + size_t space_needed; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C__H5C_T_MAGIC ); + + /* As a sanity check, set space needed to the size of the skip list. + * This should be the sum total of the sizes of all the dirty entries + * in the metadata cache. + */ + space_needed = cache_ptr->slist_size; + + /* Recall that while we shouldn't have any protected entries at this + * point, it is possible that some dirty entries may reside on the + * pinned list at this point. + */ + HDassert( cache_ptr->slist_size <= + (cache_ptr->dLRU_list_size + cache_ptr->pel_size) ); + HDassert( cache_ptr->slist_len <= + (cache_ptr->dLRU_list_len + cache_ptr->pel_len) ); + + if(space_needed > 0) { /* we have work to do */ + H5C_cache_entry_t *entry_ptr; + int nominated_entries_count = 0; + size_t nominated_entries_size = 0; + haddr_t nominated_addr; + + HDassert( cache_ptr->slist_len > 0 ); + + /* Scan the dirty LRU list from tail forward and nominate sufficient + * entries to free up the necessary space. + */ + entry_ptr = cache_ptr->dLRU_tail_ptr; + while((nominated_entries_size < space_needed) && + (nominated_entries_count < cache_ptr->slist_len) && + (entry_ptr != NULL)) { + HDassert( ! (entry_ptr->is_protected) ); + HDassert( ! (entry_ptr->is_read_only) ); + HDassert( entry_ptr->ro_ref_count == 0 ); + HDassert( entry_ptr->is_dirty ); + HDassert( entry_ptr->in_slist ); + + nominated_addr = entry_ptr->addr; + if(H5AC_add_candidate((H5AC_t *)cache_ptr, nominated_addr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5AC_add_candidate() failed(1).") + + nominated_entries_size += entry_ptr->size; + nominated_entries_count++; + entry_ptr = entry_ptr->aux_prev; + } /* end while */ + HDassert( entry_ptr == NULL ); + + /* it is possible that there are some dirty entries on the + * protected entry list as well -- scan it too if necessary + */ + entry_ptr = cache_ptr->pel_head_ptr; + while((nominated_entries_size < space_needed) && + (nominated_entries_count < cache_ptr->slist_len) && + (entry_ptr != NULL)) { + if(entry_ptr->is_dirty) { + HDassert( ! (entry_ptr->is_protected) ); + HDassert( ! (entry_ptr->is_read_only) ); + HDassert( entry_ptr->ro_ref_count == 0 ); + HDassert( entry_ptr->is_dirty ); + HDassert( entry_ptr->in_slist ); + + nominated_addr = entry_ptr->addr; + if(H5AC_add_candidate((H5AC_t *)cache_ptr, nominated_addr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5AC_add_candidate() failed(2).") + + nominated_entries_size += entry_ptr->size; + nominated_entries_count++; + } /* end if */ + + entry_ptr = entry_ptr->next; + } /* end while */ + + HDassert( nominated_entries_count == cache_ptr->slist_len ); + HDassert( nominated_entries_size == space_needed ); + } /* end if */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5C_construct_candidate_list__clean_cache() */ + + +/*------------------------------------------------------------------------- + * Function: H5C_construct_candidate_list__min_clean + * + * Purpose: Construct the list of entries that should be flushed to + * get the cache back within its min clean constraints. + * + * This function is used in managing sync points, and + * shouldn't be used elsewhere. + * + * Return: Success: SUCCEED + * + * Failure: FAIL + * + * Programmer: John Mainzer + * 3/17/10 + * + *------------------------------------------------------------------------- + */ +herr_t +H5C_construct_candidate_list__min_clean(H5C_t * cache_ptr) +{ + size_t space_needed = 0; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C__H5C_T_MAGIC ); + + /* compute the number of bytes (if any) that must be flushed to get the + * cache back within its min clean constraints. + */ + if(cache_ptr->max_cache_size > cache_ptr->index_size) { + if(((cache_ptr->max_cache_size - cache_ptr->index_size) + + cache_ptr->cLRU_list_size) >= cache_ptr->min_clean_size) + space_needed = 0; + else + space_needed = cache_ptr->min_clean_size - + ((cache_ptr->max_cache_size - cache_ptr->index_size) + + cache_ptr->cLRU_list_size); + } /* end if */ + else { + if(cache_ptr->min_clean_size <= cache_ptr->cLRU_list_size) + space_needed = 0; + else + space_needed = cache_ptr->min_clean_size - + cache_ptr->cLRU_list_size; + } /* end else */ + + if(space_needed > 0) { /* we have work to do */ + H5C_cache_entry_t *entry_ptr; + int nominated_entries_count = 0; + size_t nominated_entries_size = 0; + + HDassert( cache_ptr->slist_len > 0 ); + + /* Scan the dirty LRU list from tail forward and nominate sufficient + * entries to free up the necessary space. + */ + entry_ptr = cache_ptr->dLRU_tail_ptr; + while((nominated_entries_size < space_needed) && + (nominated_entries_count < cache_ptr->slist_len) && + (entry_ptr != NULL) && + (!entry_ptr->flush_me_last)) { + haddr_t nominated_addr; + + HDassert( ! (entry_ptr->is_protected) ); + HDassert( ! (entry_ptr->is_read_only) ); + HDassert( entry_ptr->ro_ref_count == 0 ); + HDassert( entry_ptr->is_dirty ); + HDassert( entry_ptr->in_slist ); + + nominated_addr = entry_ptr->addr; + if(H5AC_add_candidate((H5AC_t *)cache_ptr, nominated_addr) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5AC_add_candidate() failed.") + + nominated_entries_size += entry_ptr->size; + nominated_entries_count++; + entry_ptr = entry_ptr->aux_prev; + } /* end while */ + HDassert( nominated_entries_count <= cache_ptr->slist_len ); + HDassert( nominated_entries_size >= space_needed ); + } /* end if */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5C_construct_candidate_list__min_clean() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C_mark_entries_as_clean + * + * Purpose: When the H5C code is used to implement the metadata caches + * in PHDF5, only the cache with MPI_rank 0 is allowed to + * actually write entries to disk -- all other caches must + * retain dirty entries until they are advised that the + * entries are clean. + * + * This function exists to allow the H5C code to receive these + * notifications. + * + * The function receives a list of entry base addresses + * which must refer to dirty entries in the cache. If any + * of the entries are either clean or don't exist, the + * function flags an error. + * + * The function scans the list of entries and flushes all + * those that are currently unprotected with the + * H5C__FLUSH_CLEAR_ONLY_FLAG. Those that are currently + * protected are flagged for clearing when they are + * unprotected. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 7/5/05 + * + * Changes: Tidied up code, removeing some old commented out + * code that had been left in pending success of the + * new version. + * + * Note that unlike H5C_apply_candidate_list(), + * H5C_mark_entries_as_clean() makes all its calls to + * H6C_flush_single_entry() with the + * H5C__FLUSH_CLEAR_ONLY_FLAG set. As a result, + * the pre_serialize() and serialize calls are not made. + * + * This then implies that (assuming such actions were + * permitted in the parallel case) no loads, dirties, + * resizes, or removals of other entries can occur as + * a side effect of the flush. Hence, there is no need + * for the checks for entry removal / status change + * that I ported to H5C_apply_candidate_list(). + * + * However, if (in addition to allowing such operations + * in the parallel case), we allow such operations outside + * of the pre_serialize / serialize routines, this may + * cease to be the case -- requiring a review of this + * function. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C_mark_entries_as_clean(H5F_t * f, + hid_t dxpl_id, + int32_t ce_array_len, + haddr_t * ce_array_ptr) +{ + H5C_t * cache_ptr; + int entries_cleared; + int entries_examined; + int i; + int initial_list_len; + haddr_t addr; +#if H5C_DO_SANITY_CHECKS + int pinned_entries_marked = 0; + int protected_entries_marked = 0; + int other_entries_marked = 0; + haddr_t last_addr; +#endif /* H5C_DO_SANITY_CHECKS */ + H5C_cache_entry_t * clear_ptr = NULL; + H5C_cache_entry_t * entry_ptr = NULL; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(FAIL) + + HDassert( f ); + HDassert( f->shared ); + cache_ptr = f->shared->cache; + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C__H5C_T_MAGIC ); + + HDassert( ce_array_len > 0 ); + HDassert( ce_array_ptr != NULL ); + +#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 ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "an extreme sanity check failed on entry.\n"); + } +#endif /* H5C_DO_EXTREME_SANITY_CHECKS */ + + for ( i = 0; i < ce_array_len; i++ ) + { + addr = ce_array_ptr[i]; + +#if H5C_DO_SANITY_CHECKS + if ( i == 0 ) { + + last_addr = addr; + + } else { + + if ( last_addr == addr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Duplicate entry in cleaned list.\n"); + + } else if ( last_addr > addr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "cleaned list not sorted.\n"); + } + } + +#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 ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "an extreme sanity check failed in for loop.\n"); + } +#endif /* H5C_DO_EXTREME_SANITY_CHECKS */ +#endif /* H5C_DO_SANITY_CHECKS */ + + HDassert( H5F_addr_defined(addr) ); + + H5C__SEARCH_INDEX(cache_ptr, addr, entry_ptr, FAIL) + + if ( entry_ptr == NULL ) { +#if H5C_DO_SANITY_CHECKS + HDfprintf(stdout, + "H5C_mark_entries_as_clean: entry[%d] = %ld not in cache.\n", + (int)i, + (long)addr); +#endif /* H5C_DO_SANITY_CHECKS */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Listed entry not in cache?!?!?.") + + } else if ( ! entry_ptr->is_dirty ) { + +#if H5C_DO_SANITY_CHECKS + HDfprintf(stdout, + "H5C_mark_entries_as_clean: entry %ld is not dirty!?!\n", + (long)addr); +#endif /* H5C_DO_SANITY_CHECKS */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Listed entry not dirty?!?!?.") + + } else { + + /* Mark the entry to be cleared on unprotect. We will + * scan the LRU list shortly, and clear all those entries + * not currently protected. + */ + entry_ptr->clear_on_unprotect = TRUE; +#if H5C_DO_SANITY_CHECKS + if ( entry_ptr->is_protected ) { + + protected_entries_marked++; + + } else if ( entry_ptr->is_pinned ) { + + pinned_entries_marked++; + + } else { + + other_entries_marked++; + } +#endif /* H5C_DO_SANITY_CHECKS */ + } + } + + /* Scan through the LRU list from back to front, and flush the + * entries whose clear_on_unprotect flags are set. Observe that + * any protected entries will not be on the LRU, and therefore + * will not be flushed at this time. + * + * Note that unlike H5C_apply_candidate_list(), + * H5C_mark_entries_as_clean() makes all its calls to + * H6C_flush_single_entry() with the H5C__FLUSH_CLEAR_ONLY_FLAG + * set. As a result, the pre_serialize() and serialize calls are + * not made. + * + * This then implies that (assuming such actions were + * permitted in the parallel case) no loads, dirties, + * resizes, or removals of other entries can occur as + * a side effect of the flush. Hence, there is no need + * for the checks for entry removal / status change + * that I ported to H5C_apply_candidate_list(). + * + * However, if (in addition to allowing such operations + * in the parallel case), we allow such operations outside + * of the pre_serialize / serialize routines, this may + * cease to be the case -- requiring a review of this + * point. + * JRM -- 4/7/15 + */ + + entries_cleared = 0; + entries_examined = 0; + initial_list_len = cache_ptr->LRU_list_len; + entry_ptr = cache_ptr->LRU_tail_ptr; + + while ( ( entry_ptr != NULL ) && + ( entries_examined <= initial_list_len ) && + ( entries_cleared < ce_array_len ) ) + { + if ( entry_ptr->clear_on_unprotect ) { + + entry_ptr->clear_on_unprotect = FALSE; + clear_ptr = entry_ptr; + entry_ptr = entry_ptr->prev; + entries_cleared++; + + if(H5C__flush_single_entry(f, dxpl_id, clear_ptr, H5C__FLUSH_CLEAR_ONLY_FLAG | H5C__DEL_FROM_SLIST_ON_DESTROY_FLAG, NULL) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } else { + + entry_ptr = entry_ptr->prev; + } + entries_examined++; + } + +#if H5C_DO_SANITY_CHECKS + HDassert( entries_cleared == other_entries_marked ); +#endif /* H5C_DO_SANITY_CHECKS */ + + /* It is also possible that some of the cleared entries are on the + * pinned list. Must scan that also. + */ + + entry_ptr = cache_ptr->pel_head_ptr; + + while ( entry_ptr != NULL ) + { + if ( entry_ptr->clear_on_unprotect ) { + + entry_ptr->clear_on_unprotect = FALSE; + clear_ptr = entry_ptr; + entry_ptr = entry_ptr->next; + entries_cleared++; + + if(H5C__flush_single_entry(f, dxpl_id, clear_ptr, H5C__FLUSH_CLEAR_ONLY_FLAG | H5C__DEL_FROM_SLIST_ON_DESTROY_FLAG, NULL) < 0 ) + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } else { + + entry_ptr = entry_ptr->next; + } + } + +#if H5C_DO_SANITY_CHECKS + HDassert( entries_cleared == pinned_entries_marked + other_entries_marked ); + HDassert( entries_cleared + protected_entries_marked == ce_array_len ); +#endif /* H5C_DO_SANITY_CHECKS */ + + HDassert( ( entries_cleared == ce_array_len ) || + ( (ce_array_len - entries_cleared) <= cache_ptr->pl_len ) ); + +#if H5C_DO_SANITY_CHECKS + i = 0; + entry_ptr = cache_ptr->pl_head_ptr; + while ( entry_ptr != NULL ) + { + if ( entry_ptr->clear_on_unprotect ) { + + i++; + } + entry_ptr = entry_ptr->next; + } + HDassert( (entries_cleared + i) == ce_array_len ); +#endif /* H5C_DO_SANITY_CHECKS */ + +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 ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "an extreme sanity check failed on exit.\n"); + } +#endif /* H5C_DO_EXTREME_SANITY_CHECKS */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C_mark_entries_as_clean() */ +#endif /* H5_HAVE_PARALLEL */ + -- cgit v0.12