diff options
Diffstat (limited to 'src')
-rw-r--r-- | src/H5AC2.c | 4790 | ||||
-rw-r--r-- | src/H5AC2pkg.h | 335 | ||||
-rw-r--r-- | src/H5AC2private.h | 315 | ||||
-rw-r--r-- | src/H5AC2public.h | 393 | ||||
-rw-r--r-- | src/H5C2.c | 11024 | ||||
-rw-r--r-- | src/H5C2pkg.h | 949 | ||||
-rw-r--r-- | src/H5C2private.h | 1350 | ||||
-rw-r--r-- | src/H5C2public.h | 55 |
8 files changed, 19211 insertions, 0 deletions
diff --git a/src/H5AC2.c b/src/H5AC2.c new file mode 100644 index 0000000..a053b36 --- /dev/null +++ b/src/H5AC2.c @@ -0,0 +1,4790 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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: H5AC2.c + * Jul 9 1997 + * Robb Matzke <matzke@llnl.gov> + * + * Purpose: Functions in this file implement a cache for + * things which exist on disk. All "things" associated + * with a particular HDF file share the same cache; each + * HDF file has it's own cache. + * + * Modifications: + * + * Robb Matzke, 4 Aug 1997 + * Added calls to H5E. + * + * Quincey Koziol, 22 Apr 2000 + * Turned on "H5AC2_SORT_BY_ADDR" + * + * John Mainzer, 5/19/04 + * Complete redesign and rewrite. See the header comments for + * H5AC2_t for an overview of what is going on. + * + * John Mainzer, 6/4/04 + * Factored the new cache code into a separate file (H5C.c) to + * facilitate re-use. Re-worked this file again to use H5C. + * + * John Mainzer, 10/18/07 + * Copied H5AC2.c to H5AC22.c and reworked to use H5C2 instead of H5C. + * All this is in support of cache API modifications needed for + * journaling. + * + *------------------------------------------------------------------------- + */ + +#define H5C2_PACKAGE /*suppress error about including H5C2pkg */ +#define H5AC2_PACKAGE /*suppress error about including H5AC2pkg */ +#define H5F_PACKAGE /*suppress error about including H5Fpkg */ + +/* Interface initialization */ +#define H5_INTERFACE_INIT_FUNC H5AC2_init_interface + +#ifdef H5_HAVE_PARALLEL +#include <mpi.h> +#endif /* H5_HAVE_PARALLEL */ + +#include "H5private.h" /* Generic Functions */ +#include "H5AC2pkg.h" /* Metadata cache */ +#include "H5C2pkg.h" /* Cache */ +#include "H5Dprivate.h" /* Dataset functions */ +#include "H5Eprivate.h" /* Error handling */ +#include "H5Fpkg.h" /* Files */ +#include "H5FDprivate.h" /* File drivers */ +#include "H5FLprivate.h" /* Free Lists */ +#include "H5Iprivate.h" /* IDs */ +#include "H5MMprivate.h" /* Memory management */ +#include "H5Pprivate.h" /* Property lists */ + + +#ifdef H5_HAVE_PARALLEL + +/* Declare a free list to manage the H5AC2_aux_t struct */ +H5FL_DEFINE_STATIC(H5AC2_aux_t); + +#endif /* H5_HAVE_PARALLEL */ + +/**************************************************************************** + * + * structure H5AC2_slist_entry_t + * + * The dirty entry list maintained via the d_slist_ptr field of H5AC2_aux_t + * and the cleaned entry list maintained via the c_slist_ptr field of + * H5AC2_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: + * + * magic: Unsigned 32 bit integer always set to + * H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC. This field is used to + * validate pointers to instances of H5AC2_slist_entry_t. + * + * 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 renamed. They are + * removed when they appear in a clean entries broadcast. + * + ****************************************************************************/ + +#ifdef H5_HAVE_PARALLEL + +#define H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC 0x00D0A02 + +typedef struct H5AC2_slist_entry_t +{ + uint32_t magic; + + haddr_t addr; +} H5AC2_slist_entry_t; + +/* Declare a free list to manage the H5AC2_slist_entry_t struct */ +H5FL_DEFINE_STATIC(H5AC2_slist_entry_t); + +#endif /* H5_HAVE_PARALLEL */ + + +/* + * Private file-scope variables. + */ + +/* Default dataset transfer property list for metadata I/O calls */ +/* (Collective set, "block before metadata write" set and "library internal" set) */ +/* (Global variable definition, declaration is in H5AC2private.h also) */ +hid_t H5AC2_dxpl_id=(-1); + +/* Private dataset transfer property list for metadata I/O calls */ +/* (Collective set and "library internal" set) */ +/* (Static variable definition) */ +static hid_t H5AC2_noblock_dxpl_id=(-1); + +/* Dataset transfer property list for independent metadata I/O calls */ +/* (just "library internal" set - i.e. independent transfer mode) */ +/* (Global variable definition, declaration is in H5AC2private.h also) */ +hid_t H5AC2_ind_dxpl_id=(-1); + + +/* + * Private file-scope function declarations: + */ + +static herr_t H5AC2_check_if_write_permitted(const H5F_t *f, + hid_t dxpl_id, + hbool_t * write_permitted_ptr); + +#ifdef H5_HAVE_PARALLEL +static herr_t H5AC2_broadcast_clean_list(H5AC2_t * cache_ptr); +#endif /* JRM */ + +static herr_t H5AC2_ext_config_2_int_config( + H5AC2_cache_config_t * ext_conf_ptr, + H5C2_auto_size_ctl_t * int_conf_ptr); + +#ifdef H5_HAVE_PARALLEL +static herr_t H5AC2_log_deleted_entry(H5AC2_t * cache_ptr, + H5AC2_info_t * entry_ptr, + haddr_t addr, + unsigned int flags); + +static herr_t H5AC2_log_dirtied_entry(H5AC2_t * cache_ptr, + H5C2_cache_entry_t * entry_ptr, + haddr_t addr, + hbool_t size_changed, + size_t new_size); + +static herr_t H5AC2_log_flushed_entry(H5C2_t * cache_ptr, + haddr_t addr, + hbool_t was_dirty, + unsigned flags, + int type_id); + +#if 0 /* this is useful debugging code -- JRM */ +static herr_t H5AC2_log_flushed_entry_dummy(H5C2_t * cache_ptr, + haddr_t addr, + hbool_t was_dirty, + unsigned flags, + int type_id); +#endif /* JRM */ + +static herr_t H5AC2_log_inserted_entry(H5F_t * f, + H5AC2_t * cache_ptr, + H5AC2_info_t * entry_ptr, + const H5AC2_class_t * type, + haddr_t addr, + size_t size); + +static herr_t H5AC2_propagate_flushed_and_still_clean_entries_list(H5F_t * f, + hid_t dxpl_id, + H5AC2_t * cache_ptr, + hbool_t do_barrier); + +static herr_t H5AC2_receive_and_apply_clean_list(H5F_t * f, + hid_t dxpl_id, + H5AC2_t * cache_ptr); + +static herr_t H5AC2_log_renamed_entry(H5AC2_t * cache_ptr, + haddr_t old_addr, + haddr_t new_addr); +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_init + * + * Purpose: Initialize the interface from some other layer. + * + * Return: Success: non-negative + * + * Failure: negative + * + * Programmer: Quincey Koziol + * Saturday, January 18, 2003 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_init(void) +{ + herr_t ret_value=SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5AC2_init, FAIL) + /* FUNC_ENTER() does all the work */ + +done: + FUNC_LEAVE_NOAPI(ret_value) +} + + +/*------------------------------------------------------------------------- + * Function: H5AC2_init_interface + * + * Purpose: Initialize interface-specific information + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Quincey Koziol + * Thursday, July 18, 2002 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ +static herr_t +H5AC2_init_interface(void) +{ +#ifdef H5_HAVE_PARALLEL + H5P_genclass_t *xfer_pclass; /* Dataset transfer property list class object */ + H5P_genplist_t *xfer_plist; /* Dataset transfer property list object */ + unsigned block_before_meta_write; /* "block before meta write" property value */ + unsigned library_internal=1; /* "library internal" property value */ + H5FD_mpio_xfer_t xfer_mode; /* I/O transfer mode property value */ + herr_t ret_value=SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI_NOINIT(H5AC2_init_interface) + + /* Sanity check */ + HDassert(H5P_CLS_DATASET_XFER_g!=(-1)); + + /* Get the dataset transfer property list class object */ + if (NULL == (xfer_pclass = H5I_object(H5P_CLS_DATASET_XFER_g))) + HGOTO_ERROR(H5E_CACHE, H5E_BADATOM, FAIL, "can't get property list class") + + /* Get an ID for the blocking, collective H5AC2 dxpl */ + if ((H5AC2_dxpl_id=H5P_create_id(xfer_pclass)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTCREATE, FAIL, "unable to register property list") + + /* Get the property list object */ + if (NULL == (xfer_plist = H5I_object(H5AC2_dxpl_id))) + HGOTO_ERROR(H5E_CACHE, H5E_BADATOM, FAIL, "can't get new property list object") + + /* Insert 'block before metadata write' property */ + block_before_meta_write=1; + if(H5P_insert(xfer_plist,H5AC2_BLOCK_BEFORE_META_WRITE_NAME,H5AC2_BLOCK_BEFORE_META_WRITE_SIZE,&block_before_meta_write,NULL,NULL,NULL,NULL,NULL,NULL)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't insert metadata cache dxpl property") + + /* Insert 'library internal' property */ + if(H5P_insert(xfer_plist,H5AC2_LIBRARY_INTERNAL_NAME,H5AC2_LIBRARY_INTERNAL_SIZE,&library_internal,NULL,NULL,NULL,NULL,NULL,NULL)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't insert metadata cache dxpl property") + + /* Set the transfer mode */ + xfer_mode=H5FD_MPIO_COLLECTIVE; + if (H5P_set(xfer_plist,H5D_XFER_IO_XFER_MODE_NAME,&xfer_mode)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "unable to set value") + + /* Get an ID for the non-blocking, collective H5AC2 dxpl */ + if ((H5AC2_noblock_dxpl_id=H5P_create_id(xfer_pclass)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTCREATE, FAIL, "unable to register property list") + + /* Get the property list object */ + if (NULL == (xfer_plist = H5I_object(H5AC2_noblock_dxpl_id))) + HGOTO_ERROR(H5E_CACHE, H5E_BADATOM, FAIL, "can't get new property list object") + + /* Insert 'block before metadata write' property */ + block_before_meta_write=0; + if(H5P_insert(xfer_plist,H5AC2_BLOCK_BEFORE_META_WRITE_NAME,H5AC2_BLOCK_BEFORE_META_WRITE_SIZE,&block_before_meta_write,NULL,NULL,NULL,NULL,NULL,NULL)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't insert metadata cache dxpl property") + + /* Insert 'library internal' property */ + if(H5P_insert(xfer_plist,H5AC2_LIBRARY_INTERNAL_NAME,H5AC2_LIBRARY_INTERNAL_SIZE,&library_internal,NULL,NULL,NULL,NULL,NULL,NULL)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't insert metadata cache dxpl property") + + /* Set the transfer mode */ + xfer_mode=H5FD_MPIO_COLLECTIVE; + if (H5P_set(xfer_plist,H5D_XFER_IO_XFER_MODE_NAME,&xfer_mode)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "unable to set value") + + /* Get an ID for the non-blocking, independent H5AC2 dxpl */ + if ((H5AC2_ind_dxpl_id=H5P_create_id(xfer_pclass)) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTCREATE, FAIL, "unable to register property list") + + /* Get the property list object */ + if (NULL == (xfer_plist = H5I_object(H5AC2_ind_dxpl_id))) + HGOTO_ERROR(H5E_CACHE, H5E_BADATOM, FAIL, "can't get new property list object") + + /* Insert 'block before metadata write' property */ + block_before_meta_write=0; + if(H5P_insert(xfer_plist,H5AC2_BLOCK_BEFORE_META_WRITE_NAME,H5AC2_BLOCK_BEFORE_META_WRITE_SIZE,&block_before_meta_write,NULL,NULL,NULL,NULL,NULL,NULL)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't insert metadata cache dxpl property") + + /* Insert 'library internal' property */ + if(H5P_insert(xfer_plist,H5AC2_LIBRARY_INTERNAL_NAME,H5AC2_LIBRARY_INTERNAL_SIZE,&library_internal,NULL,NULL,NULL,NULL,NULL,NULL)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't insert metadata cache dxpl property") + + /* Set the transfer mode */ + xfer_mode=H5FD_MPIO_INDEPENDENT; + if (H5P_set(xfer_plist,H5D_XFER_IO_XFER_MODE_NAME,&xfer_mode)<0) + HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "unable to set value") + +done: + FUNC_LEAVE_NOAPI(ret_value) + +#else /* H5_HAVE_PARALLEL */ + FUNC_ENTER_NOAPI_NOINIT_NOFUNC(H5AC2_init_interface) + + /* Sanity check */ + assert(H5P_LST_DATASET_XFER_g!=(-1)); + + H5AC2_dxpl_id=H5P_DATASET_XFER_DEFAULT; + H5AC2_noblock_dxpl_id=H5P_DATASET_XFER_DEFAULT; + H5AC2_ind_dxpl_id=H5P_DATASET_XFER_DEFAULT; + + FUNC_LEAVE_NOAPI(SUCCEED) +#endif /* H5_HAVE_PARALLEL */ +} /* end H5AC2_init_interface() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_term_interface + * + * Purpose: Terminate this interface. + * + * Return: Success: Positive if anything was done that might + * affect other interfaces; zero otherwise. + * + * Failure: Negative. + * + * Programmer: Quincey Koziol + * Thursday, July 18, 2002 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ +int +H5AC2_term_interface(void) +{ + int n=0; + + FUNC_ENTER_NOAPI_NOINIT_NOFUNC(H5AC2_term_interface) + + if (H5_interface_initialize_g) { +#ifdef H5_HAVE_PARALLEL + if(H5AC2_dxpl_id>0 || H5AC2_noblock_dxpl_id>0 || H5AC2_ind_dxpl_id>0) { + /* Indicate more work to do */ + n = 1; /* H5I */ + + /* Close H5AC2 dxpl */ + if (H5I_dec_ref(H5AC2_dxpl_id) < 0 || + H5I_dec_ref(H5AC2_noblock_dxpl_id) < 0 || + H5I_dec_ref(H5AC2_ind_dxpl_id) < 0) + H5E_clear_stack(NULL); /*ignore error*/ + else { + /* Reset static IDs */ + H5AC2_dxpl_id=(-1); + H5AC2_noblock_dxpl_id=(-1); + H5AC2_ind_dxpl_id=(-1); + + /* Reset interface initialization flag */ + H5_interface_initialize_g = 0; + } /* end else */ + } /* end if */ + else +#else /* H5_HAVE_PARALLEL */ + /* Reset static IDs */ + H5AC2_dxpl_id=(-1); + H5AC2_noblock_dxpl_id=(-1); + H5AC2_ind_dxpl_id=(-1); + +#endif /* H5_HAVE_PARALLEL */ + /* Reset interface initialization flag */ + H5_interface_initialize_g = 0; + } /* end if */ + + FUNC_LEAVE_NOAPI(n) +} /* end H5AC2_term_interface() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_create + * + * Purpose: Initialize the cache just after a file is opened. The + * SIZE_HINT is the number of cache slots desired. If you + * pass an invalid value then H5AC2_NSLOTS is used. You can + * turn off caching by using 1 for the SIZE_HINT value. + * + * Return: Success: Number of slots actually used. + * + * Failure: Negative + * + * Programmer: Robb Matzke + * matzke@llnl.gov + * Jul 9 1997 + * + * Modifications: + * + * Complete re-design and re-write to support the re-designed + * metadata cache. + * + * At present, the size_hint is ignored, and the + * max_cache_size and min_clean_size fields are hard + * coded. This should be fixed, but a parameter + * list change will be required, so I will leave it + * for now. + * + * Since no-one seems to care, the function now returns + * one on success. + * JRM - 4/28/04 + * + * Reworked the function again after abstracting its guts to + * the similar function in H5C2.c. The function is now a + * wrapper for H5C2_create(). + * JRM - 6/4/04 + * + * Deleted the old size_hint parameter and added the + * max_cache_size, and min_clean_size parameters. + * + * JRM - 3/10/05 + * + * Deleted the max_cache_size, and min_clean_size parameters, + * and added the config_ptr parameter. Added code to + * validate the resize configuration before we do anything. + * + * JRM - 3/24/05 + * + * Changed the type of config_ptr from H5AC2_auto_size_ctl_t * + * to H5AC2_cache_config_t *. Propagated associated changes + * through the function. + * JRM - 4/7/05 + * + * Added code allocating and initializing the auxilary + * structure (an instance of H5AC2_aux_t), and linking it + * to the instance of H5C2_t created by H5C2_create(). At + * present, the auxilary structure is only used in PHDF5. + * + * JRM - 6/28/05 + * + * Added code to set the prefix if required. + * + * JRM - 1/20/06 + * + * Added code to initialize the new write_done field. + * + * JRM - 5/11/06 + * + * Reworked code to conform with changes in the cache + * API. + * JRM - 10/18/07 + * + *------------------------------------------------------------------------- + */ + +static const char * H5AC2_entry_type_names[H5AC2_NTYPES] = +{ + "B-tree nodes", + "symbol table nodes", + "local heaps", + "global heaps", + "object headers", + "v2 B-tree headers", + "v2 B-tree internal nodes", + "v2 B-tree leaf nodes", + "fractal heap headers", + "fractal heap direct blocks", + "fractal heap indirect blocks", + "free space headers", + "free space sections", + "shared OH message master table", + "shared OH message index", + "test entry" /* for testing only -- not used for actual files */ +}; + +herr_t +H5AC2_create(const H5F_t *f, + H5AC2_cache_config_t *config_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; +#ifdef H5_HAVE_PARALLEL + char prefix[H5C2__PREFIX_LEN] = ""; + MPI_Comm mpi_comm = MPI_COMM_NULL; + int mpi_rank = -1; + int mpi_size = -1; + H5AC2_aux_t * aux_ptr = NULL; +#endif /* H5_HAVE_PARALLEL */ + + FUNC_ENTER_NOAPI(H5AC2_create, FAIL) + + HDassert ( f ); + HDassert ( NULL == f->shared->cache2 ); + HDassert ( config_ptr != NULL ) ; + HDassert ( NELMTS(H5AC2_entry_type_names) == H5AC2_NTYPES); + HDassert ( H5C2__MAX_NUM_TYPE_IDS == H5AC2_NTYPES); + + result = H5AC2_validate_config(config_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "Bad cache configuration"); + } + +#ifdef H5_HAVE_PARALLEL + if ( IS_H5FD_MPI(f) ) { + + if ( (mpi_comm = H5F_mpi_get_comm(f)) == MPI_COMM_NULL ) { + + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, \ + "can't get MPI communicator") + } + + if ( (mpi_rank = H5F_mpi_get_rank(f)) < 0 ) { + + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't get mpi rank") + } + + if ( (mpi_size = H5F_mpi_get_size(f)) < 0 ) { + + HGOTO_ERROR(H5E_VFL, H5E_CANTGET, FAIL, "can't get mpi size") + } + + /* There is no point in setting up the auxilary structure if size + * is less than or equal to 1, as there will never be any processes + * to broadcast the clean lists to. + */ + if ( mpi_size > 1 ) { + + if ( NULL == (aux_ptr = H5FL_CALLOC(H5AC2_aux_t)) ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "Can't allocate H5AC2 auxilary structure.") + + } else { + + aux_ptr->magic = H5AC2__H5AC2_AUX_T_MAGIC; + aux_ptr->mpi_comm = mpi_comm; + aux_ptr->mpi_rank = mpi_rank; + aux_ptr->mpi_size = mpi_size; + aux_ptr->write_permitted = FALSE; + aux_ptr->dirty_bytes_threshold = + H5AC2__DEFAULT_DIRTY_BYTES_THRESHOLD; + aux_ptr->dirty_bytes = 0; +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->dirty_bytes_propagations = 0; + 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 /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + aux_ptr->d_slist_ptr = NULL; + aux_ptr->d_slist_len = 0; + aux_ptr->c_slist_ptr = NULL; + aux_ptr->c_slist_len = 0; + aux_ptr->write_done = NULL; + + sprintf(prefix, "%d:", mpi_rank); + } + + if ( mpi_rank == 0 ) { + + aux_ptr->d_slist_ptr = + H5SL_create(H5SL_TYPE_HADDR,0.5,(size_t)16); + + if ( aux_ptr->d_slist_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTCREATE, FAIL, + "can't create dirtied entry list.") + } + + aux_ptr->c_slist_ptr = + H5SL_create(H5SL_TYPE_HADDR,0.5,(size_t)16); + + if ( aux_ptr->c_slist_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTCREATE, FAIL, + "can't create cleaned entry list.") + } + } + } + + if ( aux_ptr != NULL ) { + + if ( aux_ptr->mpi_rank == 0 ) { + + f->shared->cache2 = H5C2_create(f, + H5AC2__DEFAULT_MAX_CACHE_SIZE, + H5AC2__DEFAULT_MIN_CLEAN_SIZE, + (H5AC2_NTYPES - 1), + (const char **)H5AC2_entry_type_names, + H5AC2_check_if_write_permitted, + TRUE, + H5AC2_log_flushed_entry, + (void *)aux_ptr); + + } else { + + f->shared->cache2 = H5C2_create(f, + H5AC2__DEFAULT_MAX_CACHE_SIZE, + H5AC2__DEFAULT_MIN_CLEAN_SIZE, + (H5AC2_NTYPES - 1), + (const char **)H5AC2_entry_type_names, + NULL, + FALSE, +#if 0 /* this is useful debugging code -- keep it for a while */ /* JRM */ + H5AC2_log_flushed_entry_dummy, +#else /* JRM */ + NULL, +#endif /* JRM */ + (void *)aux_ptr); + } + + } else { + + f->shared->cache2 = H5C2_create(f, + H5AC2__DEFAULT_MAX_CACHE_SIZE, + H5AC2__DEFAULT_MIN_CLEAN_SIZE, + (H5AC2_NTYPES - 1), + (const char **)H5AC2_entry_type_names, + H5AC2_check_if_write_permitted, + TRUE, + NULL, + NULL); + } + } else { +#endif /* H5_HAVE_PARALLEL */ + /* The default max cache size and min clean size will frequently be + * overwritten shortly by the subsequent set resize config call. + * -- JRM + */ + + f->shared->cache2 = H5C2_create(f, + H5AC2__DEFAULT_MAX_CACHE_SIZE, + H5AC2__DEFAULT_MIN_CLEAN_SIZE, + (H5AC2_NTYPES - 1), + (const char **)H5AC2_entry_type_names, + H5AC2_check_if_write_permitted, + TRUE, + NULL, + NULL); +#ifdef H5_HAVE_PARALLEL + } +#endif /* H5_HAVE_PARALLEL */ + + if ( NULL == f->shared->cache2 ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "memory allocation failed") + + } +#ifdef H5_HAVE_PARALLEL + else if ( aux_ptr != NULL ) { + + result = H5C2_set_prefix(f->shared->cache2, prefix); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "H5C2_set_prefix() failed") + } + } +#endif /* H5_HAVE_PARALLEL */ + + result = H5AC2_set_cache_auto_resize_config(f->shared->cache2, config_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "auto resize configuration failed") + } + +done: + +#ifdef H5_HAVE_PARALLEL + + /* if there is a failure, try to tidy up the auxilary structure */ + + if ( ret_value != SUCCEED ) { + + if ( aux_ptr != NULL ) { + + if ( aux_ptr->d_slist_ptr != NULL ) { + + H5SL_close(aux_ptr->d_slist_ptr); + } + + if ( aux_ptr->c_slist_ptr != NULL ) { + + H5SL_close(aux_ptr->c_slist_ptr); + } + + aux_ptr->magic = 0; + H5FL_FREE(H5AC2_aux_t, aux_ptr); + aux_ptr = NULL; + } + } +#endif /* H5_HAVE_PARALLEL */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_create() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_dest + * + * Purpose: Flushes all data to disk and destroys the cache. + * This function fails if any object are protected since the + * resulting file might not be consistent. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Robb Matzke + * matzke@llnl.gov + * Jul 9 1997 + * + * Modifications: + * + * Complete re-design and re-write to support the re-designed + * metadata cache. + * JRM - 5/12/04 + * + * Abstracted the guts of the function to H5C_dest() in H5C.c, + * and then re-wrote the function as a wrapper for H5C_dest(). + * + * JRM - 6/7/04 + * + * Added code to free the auxiliary structure and its + * associated slist if present. + * JRM - 6/28/05 + * + * Added code to close the trace file if it is present. + * + * JRM - 6/8/06 + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_dest(H5F_t *f, hid_t dxpl_id) +{ + H5AC2_t *cache = NULL; + herr_t ret_value=SUCCEED; /* Return value */ +#ifdef H5_HAVE_PARALLEL + H5AC2_aux_t * aux_ptr = NULL; +#endif /* H5_HAVE_PARALLEL */ + + FUNC_ENTER_NOAPI(H5AC2_dest, FAIL) + + assert(f); + assert(f->shared->cache2); + cache = f->shared->cache2; +#ifdef H5_HAVE_PARALLEL + aux_ptr = cache->aux_ptr; + + if ( aux_ptr != NULL ) { + + HDassert ( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + } +#endif /* H5_HAVE_PARALLEL */ + +#if H5AC2__TRACE_FILE_ENABLED + if ( H5AC2_close_trace_file(cache) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5AC2_close_trace_file() failed.") + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + if ( H5C2_dest(cache, dxpl_id) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFREE, FAIL, "can't destroy cache") + } + + f->shared->cache2 = NULL; + +#ifdef H5_HAVE_PARALLEL + if ( aux_ptr != NULL ) { + + if ( aux_ptr->d_slist_ptr != NULL ) { + + H5SL_close(aux_ptr->d_slist_ptr); + } + + if ( aux_ptr->c_slist_ptr != NULL ) { + + H5SL_close(aux_ptr->c_slist_ptr); + } + + aux_ptr->magic = 0; + H5FL_FREE(H5AC2_aux_t, aux_ptr); + aux_ptr = NULL; + } +#endif /* H5_HAVE_PARALLEL */ + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_dest() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_expunge_entry + * + * Purpose: Expunge the target entry from the cache without writing it + * to disk even if it is dirty. The entry must not be either + * pinned or protected. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/30/06 + * + * Modifications: + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_expunge_entry(H5F_t *f, + hid_t dxpl_id, + const H5AC2_class_t *type, + haddr_t addr) +{ + herr_t result; + herr_t ret_value=SUCCEED; /* Return value */ + H5AC2_t * cache_ptr = NULL; +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_expunge_entry, FAIL) + + HDassert(f); + HDassert(f->shared); + HDassert(f->shared->cache2); + HDassert(type); + HDassert(H5F_addr_defined(addr)); + + cache_ptr = f->shared->cache2; + +#if H5AC2__TRACE_FILE_ENABLED + /* For the expunge entry call, only the addr, and type id are really + * necessary in the trace file. Write the return value to catch occult + * errors. + */ + if ( ( cache_ptr != NULL ) && + ( H5C2_get_trace_file_ptr(cache_ptr, &trace_file_ptr) >= 0 ) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_expunge_entry 0x%lx %d", + (unsigned long)addr, + (int)(type->id)); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + result = H5C2_expunge_entry(cache_ptr, + dxpl_id, + type, + addr); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTEXPUNGE, FAIL, \ + "H5C2_expunge_entry() failed.") + } + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_expunge_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_flush + * + * Purpose: Flush (and possibly destroy) the metadata cache associated + * with the specified file. + * + * This is a re-write of an earlier version of the function + * which was reputedly capable of flushing (and destroying + * if requested) individual entries, individual entries if + * they match the supplied type, all entries of a given type, + * as well as all entries in the cache. + * + * As only this last capability is actually used at present, + * I have not implemented the other capabilities in this + * version of the function. + * + * The type and addr parameters are retained to avoid source + * code changed, but values other than NULL and HADDR_UNDEF + * respectively are errors. If all goes well, they should + * be removed, and the function renamed to something more + * descriptive -- perhaps H5AC2_flush_cache. + * + * If the cache contains protected entries, the function will + * fail, as protected entries cannot be flushed. However + * all unprotected entries should be flushed before the + * function returns failure. + * + * For historical purposes, the original version of the + * purpose section is reproduced below: + * + * ============ Original Version of "Purpose:" ============ + * + * Flushes (and destroys if DESTROY is non-zero) the specified + * entry from the cache. If the entry TYPE is CACHE_FREE and + * ADDR is HADDR_UNDEF then all types of entries are + * flushed. If TYPE is CACHE_FREE and ADDR is defined then + * whatever is cached at ADDR is flushed. Otherwise the thing + * at ADDR is flushed if it is the correct type. + * + * If there are protected objects they will not be flushed. + * However, an attempt will be made to flush all non-protected + * items before this function returns failure. + * + * Return: Non-negative on success/Negative on failure if there was a + * request to flush all items and something was protected. + * + * Programmer: Robb Matzke + * matzke@llnl.gov + * Jul 9 1997 + * + * Modifications: + * Robb Matzke, 1999-07-27 + * The ADDR argument is passed by value. + * + * Complete re-write. See above for details. -- JRM 5/11/04 + * + * Abstracted the guts of the function to H5C_flush_cache() + * in H5C.c, and then re-wrote the function as a wrapper for + * H5C_flush_cache(). + * + * JRM - 6/7/04 + * + * JRM - 7/5/05 + * Modified function as part of a fix for a cache coherency + * bug in PHDF5. See the header comments on the H5AC2_aux_t + * structure for details. + * + * JRM -- 5/11/06 + * Added call to the write_done callback. + * + * JRM -- 6/6/06 + * Added trace file support. + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_flush(H5F_t *f, hid_t dxpl_id, unsigned flags) +{ + herr_t status; + herr_t ret_value = SUCCEED; /* Return value */ +#ifdef H5_HAVE_PARALLEL + H5AC2_aux_t * aux_ptr = NULL; + int mpi_code; +#endif /* H5_HAVE_PARALLEL */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + + FUNC_ENTER_NOAPI(H5AC2_flush, FAIL) + + HDassert(f); + HDassert(f->shared->cache2); + +#if H5AC2__TRACE_FILE_ENABLED + /* For the flush, only the flags are really necessary in the trace file. + * Write the result to catch occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_flush 0x%x", flags); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + +#ifdef H5_HAVE_PARALLEL + aux_ptr = f->shared->cache2->aux_ptr; + + if ( aux_ptr != NULL ) { + +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + HDfprintf(stdout, + "%d::H5AC2_flush: (u/uu/i/iu/r/ru) = %d/%d/%d/%d/%d/%d\n", + (int)(aux_ptr->mpi_rank), + (int)(aux_ptr->unprotect_dirty_bytes), + (int)(aux_ptr->unprotect_dirty_bytes_updates), + (int)(aux_ptr->insert_dirty_bytes), + (int)(aux_ptr->insert_dirty_bytes_updates), + (int)(aux_ptr->rename_dirty_bytes), + (int)(aux_ptr->rename_dirty_bytes_updates)); +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + + /* to prevent "messages from the future" we must synchronize all + * processes before we start the flush. Hence the following + * barrier. + */ + if ( MPI_SUCCESS != (mpi_code = MPI_Barrier(aux_ptr->mpi_comm)) ) { + + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_code) + } + + /* if the clear only flag is set, this flush will not involve any + * disk I/O. In such cases, it is not necessary to let process 0 + * flush first. + */ + if ( ( aux_ptr->mpi_rank == 0 ) && + ( (flags & H5AC2__FLUSH_CLEAR_ONLY_FLAG) == 0 ) ) { + + unsigned init_flush_flags = H5AC2__NO_FLAGS_SET; + + if ( ( (flags & H5AC2__FLUSH_MARKED_ENTRIES_FLAG) != 0 ) && + ( (flags & H5AC2__FLUSH_INVALIDATE_FLAG) == 0 ) ) { + + init_flush_flags |= H5AC2__FLUSH_MARKED_ENTRIES_FLAG; + } + + aux_ptr->write_permitted = TRUE; + + status = H5C2_flush_cache(f->shared->cache2, + dxpl_id, + init_flush_flags); + + aux_ptr->write_permitted = FALSE; + + if ( status < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't flush.") + } + + if ( aux_ptr->write_done != NULL ) { + + (aux_ptr->write_done)(); + } + + } /* end if ( aux_ptr->mpi_rank == 0 ) */ + + status = H5AC2_propagate_flushed_and_still_clean_entries_list(f, + H5AC2_noblock_dxpl_id, + f->shared->cache2, + FALSE); + } /* end if ( aux_ptr != NULL ) */ +#endif /* H5_HAVE_PARALLEL */ + + status = H5C2_flush_cache(f->shared->cache2, + dxpl_id, + flags); + + if ( status < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "Can't flush entry.") + } + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_flush() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_get_entry_status + * + * Purpose: Given a file address, determine whether the metadata + * cache contains an entry at that location. If it does, + * also determine whether the entry is dirty, protected, + * pinned, etc. and return that information to the caller + * in *status_ptr. + * + * If the specified entry doesn't exist, set *status_ptr + * to zero. + * + * On error, the value of *status_ptr is undefined. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 4/27/06 + * + * Modifications: + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_get_entry_status(H5F_t * f, + haddr_t addr, + unsigned * status_ptr) +{ + H5C2_t *cache_ptr = f->shared->cache2; + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; + hbool_t in_cache; + hbool_t is_dirty; + hbool_t is_protected; + hbool_t is_pinned; + size_t entry_size; + unsigned status = 0; + + FUNC_ENTER_NOAPI(H5AC2_get_entry_status, FAIL) + + if ( ( cache_ptr == NULL ) || + ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) || + ( ! H5F_addr_defined(addr) ) || + ( status_ptr == NULL ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad param(s) on entry.") + } + + result = H5C2_get_entry_status(cache_ptr, addr, &entry_size, &in_cache, + &is_dirty, &is_protected, &is_pinned); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_entry_status() failed.") + } + + if ( in_cache ) { + + status |= H5AC2_ES__IN_CACHE; + + if ( is_dirty ) + status |= H5AC2_ES__IS_DIRTY; + + if ( is_protected ) + status |= H5AC2_ES__IS_PROTECTED; + + if ( is_pinned ) + status |= H5AC2_ES__IS_PINNED; + } + + *status_ptr = status; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_get_entry_status() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_set + * + * Purpose: Adds the specified thing to the cache. The thing need not + * exist on disk yet, but it must have an address and disk + * space reserved. + * + * If H5AC2_DEBUG is defined then this function checks + * that the object being inserted isn't a protected object. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Robb Matzke + * matzke@llnl.gov + * Jul 9 1997 + * + * Modifications: + * Robb Matzke, 1999-07-27 + * The ADDR argument is passed by value. + * + * Bill Wendling, 2003-09-16 + * Added automatic "flush" if the FPHDF5 driver is being + * used. This'll write the metadata to the SAP where other, + * lesser processes can grab it. + * + * JRM - 5/13/04 + * Complete re-write for the new metadata cache. The new + * code is functionally almost identical to the old, although + * the sanity check for a protected entry is now an assert + * at the beginning of the function. + * + * JRM - 6/7/04 + * Abstracted the guts of the function to H5C_insert_entry() + * in H5C.c, and then re-wrote the function as a wrapper for + * H5C_insert_entry(). + * + * JRM - 1/6/05 + * Added the flags parameter. At present, this parameter is + * only used to set the new flush_marker field on the new + * entry. Since this doesn't apply to the SAP code, no change + * is needed there. Thus the only change to the body of the + * code is to pass the flags parameter through to + * H5C_insert_entry(). + * + * JRM - 6/6/05 + * Added code to force newly inserted entries to be dirty + * in the flexible parallel case. The normal case is handled + * in H5C.c. This is part of a series of changes directed at + * moving management of the dirty flag on cache entries into + * the cache code. + * + * JRM - 7/5/05 + * Added code to track dirty byte generation, and to trigger + * clean entry list propagation when it exceeds a user + * specified threshold. Note that this code only applies in + * the PHDF5 case. It should have no effect on either the + * serial or FPHSD5 cases. + * + * JRM - 6/6/06 + * Added trace file support. + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_set(H5F_t *f, hid_t dxpl_id, const H5AC2_class_t *type, haddr_t addr, size_t len, void *thing, unsigned int flags) +{ + herr_t result; + H5AC2_info_t *info; + H5AC2_t *cache; + herr_t ret_value=SUCCEED; /* Return value */ +#ifdef H5_HAVE_PARALLEL + H5AC2_aux_t * aux_ptr = NULL; +#endif /* H5_HAVE_PARALLEL */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + size_t trace_entry_size = 0; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_set, FAIL) + + HDassert(f); + HDassert(f->shared->cache2); + HDassert(type); + HDassert(type->serialize); + HDassert(H5F_addr_defined(addr)); + HDassert(thing); + +#if H5AC2__TRACE_FILE_ENABLED + /* For the insert, only the addr, size, type id and flags are really + * necessary in the trace file. Write the result to catch occult + * errors. + * + * Note that some data is not available right now -- put what we can + * in the trace buffer now, and fill in the rest at the end. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_set 0x%lx %ld %d 0x%x", + (unsigned long)addr, + (long)len, + type->id, + flags); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + /* Get local copy of this information */ + cache = f->shared->cache2; + info = (H5AC2_info_t *)thing; + + info->addr = addr; + info->type = type; + info->is_protected = FALSE; + +#ifdef H5_HAVE_PARALLEL + if ( NULL != (aux_ptr = f->shared->cache2->aux_ptr) ) { + + result = H5AC2_log_inserted_entry(f, + f->shared->cache2, + (H5AC2_info_t *)thing, + type, + addr, + len); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTINS, FAIL, \ + "H5AC2_log_inserted_entry() failed.") + } + } +#endif /* H5_HAVE_PARALLEL */ + + result = H5C2_insert_entry(cache, + dxpl_id, + type, + addr, + len, + thing, + flags); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTINS, FAIL, "H5C2_insert_entry() failed") + } + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + /* make note of the entry size */ + trace_entry_size = ((H5C2_cache_entry_t *)thing)->size; + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + +#ifdef H5_HAVE_PARALLEL + if ( ( aux_ptr != NULL ) && + ( aux_ptr->dirty_bytes >= aux_ptr->dirty_bytes_threshold ) ) { + + result = H5AC2_propagate_flushed_and_still_clean_entries_list(f, + H5AC2_noblock_dxpl_id, + f->shared->cache2, + TRUE); + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "Can't propagate clean entries list.") + } + } +#endif /* H5_HAVE_PARALLEL */ + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d %d\n", trace, + (int)trace_entry_size, + (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_set() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_mark_pinned_entry_dirty + * + * Purpose: Mark a pinned entry as dirty. The target entry MUST be + * be pinned, and MUST be unprotected. + * + * If the entry has changed size, the function updates + * data structures for the size change. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 4/11/06 + * + * Modifications: + * + * Added trace file support. JRM -- 6/6/06 + * + * Modified code in support of revised cache API needed + * to permit journaling. JRM -- 10/18/07 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_mark_pinned_entry_dirty(H5F_t * f, + void * thing, + hbool_t size_changed, + size_t new_size) +{ + H5C2_t *cache_ptr = f->shared->cache2; + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_mark_pinned_entry_dirty, FAIL) + +#if H5AC2__TRACE_FILE_ENABLED + /* For the mark pinned entry dirty call, only the addr, size_changed, + * and new_size are really necessary in the trace file. Write the result + * to catch occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_mark_pinned_entry_dirty 0x%lx %d %d", + (unsigned long)(((H5C2_cache_entry_t *)thing)->addr), + (int)size_changed, + (int)new_size); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + +#ifdef H5_HAVE_PARALLEL + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + + if ( ( ((H5AC2_info_t *)thing)->is_dirty == FALSE ) && + ( NULL != cache_ptr->aux_ptr) ) { + + H5AC2_info_t * entry_ptr; + + HDassert( ( size_changed == TRUE ) || ( size_changed == FALSE ) ); + + entry_ptr = (H5AC2_info_t *)thing; + + if ( ! ( entry_ptr->is_pinned ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "Entry isn't pinned??") + } + + if ( entry_ptr->is_protected ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "Entry is protected??") + } + + result = H5AC2_log_dirtied_entry(cache_ptr, + entry_ptr, + entry_ptr->addr, + size_changed, + new_size); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "H5AC2_log_dirtied_entry() failed.") + } + } +#endif /* H5_HAVE_PARALLEL */ + + result = H5C2_mark_pinned_entry_dirty(cache_ptr, + thing, + size_changed, + new_size); + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "H5C2_mark_pinned_entry_dirty() failed.") + + } + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_mark_pinned_entry_dirty() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_mark_pinned_or_protected_entry_dirty + * + * Purpose: Mark a pinned or protected entry as dirty. The target + * entry MUST be either pinned, protected, or both. + * + * Unlike H5AC2_mark_pinned_entry_dirty(), this function does + * not support size changes. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 5/16/06 + * + * Modifications: + * + * Added trace file support. JRM -- 6/6/06 + * + * Modified code in support of revised cache API needed + * to permit journaling. JRM -- 10/18/07 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_mark_pinned_or_protected_entry_dirty(H5F_t * f, + void * thing) +{ + H5C2_t * cache_ptr = f->shared->cache2; +#ifdef H5_HAVE_PARALLEL + H5AC2_info_t * info_ptr; +#endif /* H5_HAVE_PARALLEL */ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_mark_pinned_or_protected_entry_dirty, FAIL) + +#if H5AC2__TRACE_FILE_ENABLED + /* For the mark pinned or protected entry dirty call, only the addr + * is really necessary in the trace file. Write the result to catch + * occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_mark_pinned_or_protected_entry_dirty 0x%lx", + (unsigned long)(((H5C2_cache_entry_t *)thing)->addr)); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + +#ifdef H5_HAVE_PARALLEL + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + + info_ptr = (H5AC2_info_t *)thing; + + if ( ( info_ptr->is_dirty == FALSE ) && + ( ! ( info_ptr->is_protected ) ) && + ( info_ptr->is_pinned ) && + ( NULL != cache_ptr->aux_ptr) ) { + + result = H5AC2_log_dirtied_entry(cache_ptr, + info_ptr, + info_ptr->addr, + FALSE, + 0); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "H5AC2_log_dirtied_entry() failed.") + } + } +#endif /* H5_HAVE_PARALLEL */ + + result = H5C2_mark_pinned_or_protected_entry_dirty(cache_ptr, thing); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "H5C2_mark_pinned_entry_dirty() failed.") + + } + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_mark_pinned_entry_dirty() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_rename + * + * Purpose: Use this function to notify the cache that an object's + * file address changed. + * + * If H5AC2_DEBUG is defined then this function checks + * that the old and new addresses don't correspond to the + * address of a protected object. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Robb Matzke + * matzke@llnl.gov + * Jul 9 1997 + * + * Modifications: + * Robb Matzke, 1999-07-27 + * The OLD_ADDR and NEW_ADDR arguments are passed by value. + * + * JRM 5/17/04 + * Complete rewrite for the new meta-data cache. + * + * JRM - 6/7/04 + * Abstracted the guts of the function to H5C_rename_entry() + * in H5C.c, and then re-wrote the function as a wrapper for + * H5C_rename_entry(). + * + * JRM - 7/5/05 + * Added code to track dirty byte generation, and to trigger + * clean entry list propagation when it exceeds a user + * specified threshold. Note that this code only applies in + * the PHDF5 case. It should have no effect on either the + * serial or FPHSD5 cases. + * + * Note that this code presumes that the renamed entry will + * be present in all caches -- which it must be at present. + * To maintain this invarient, only rename entries immediately + * after you unprotect them. + * + * JRM - 6/6/06 + * Added trace file support. + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_rename(H5F_t *f, const H5AC2_class_t *type, haddr_t old_addr, haddr_t new_addr) +{ + herr_t result; + herr_t ret_value=SUCCEED; /* Return value */ +#ifdef H5_HAVE_PARALLEL + H5AC2_aux_t * aux_ptr = NULL; +#endif /* H5_HAVE_PARALLEL */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_rename, FAIL) + + HDassert(f); + HDassert(f->shared->cache2); + HDassert(type); + HDassert(H5F_addr_defined(old_addr)); + HDassert(H5F_addr_defined(new_addr)); + HDassert(H5F_addr_ne(old_addr, new_addr)); + +#if H5AC2__TRACE_FILE_ENABLED + /* For the rename call, only the old addr and new addr are really + * necessary in the trace file. Include the type id so we don't have to + * look it up. Also write the result to catch occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_rename 0x%lx 0x%lx %d", + (unsigned long)old_addr, + (unsigned long)new_addr, + (int)(type->id)); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + +#ifdef H5_HAVE_PARALLEL + if ( NULL != (aux_ptr = f->shared->cache2->aux_ptr) ) { + + result = H5AC2_log_renamed_entry(f->shared->cache2, + old_addr, + new_addr); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "H5AC2_log_renamed_entry() failed.") + } + } +#endif /* H5_HAVE_PARALLEL */ + + result = H5C2_rename_entry(f->shared->cache2, + type, + old_addr, + new_addr); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRENAME, FAIL, \ + "H5C2_rename_entry() failed.") + } + +#ifdef H5_HAVE_PARALLEL + if ( ( aux_ptr != NULL ) && + ( aux_ptr->dirty_bytes >= aux_ptr->dirty_bytes_threshold ) ) { + + result = H5AC2_propagate_flushed_and_still_clean_entries_list(f, + H5AC2_noblock_dxpl_id, + f->shared->cache2, + TRUE); + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "Can't propagate clean entries list.") + } + } +#endif /* H5_HAVE_PARALLEL */ + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_rename() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_pin_protected_entry() + * + * Purpose: Pin a protected cache entry. The entry must be protected + * at the time of call, and must be unpinned. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 4/27/06 + * + * Modifications: + * + * Added trace file support. 6/6/06 + * + * Modified code in support of revised cache API needed + * to permit journaling. JRM - 10/18/07 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_pin_protected_entry(H5F_t * f, + void * thing) +{ + H5C2_t *cache_ptr = f->shared->cache2; + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_pin_protected_entry, FAIL) + +#if H5AC2__TRACE_FILE_ENABLED + /* For the pin protected entry call, only the addr is really necessary + * in the trace file. Also write the result to catch occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_pin_protected_entry 0x%lx", + (unsigned long)(((H5C2_cache_entry_t *)thing)->addr)); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + result = H5C2_pin_protected_entry(cache_ptr, thing); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPIN, FAIL, \ + "H5C2_pin_protected_entry() failed.") + } + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_pin_protected_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_protect + * + * Purpose: If the target entry is not in the cache, load it. If + * necessary, attempt to evict one or more entries to keep + * the cache within its maximum size. + * + * Mark the target entry as protected, and return its address + * to the caller. The caller must call H5AC2_unprotect() when + * finished with the entry. + * + * While it is protected, the entry may not be either evicted + * or flushed -- nor may it be accessed by another call to + * H5AC2_protect. Any attempt to do so will result in a failure. + * + * This comment is a re-write of the original Purpose: section. + * For historical interest, the original version is reproduced + * below: + * + * Original Purpose section: + * + * Similar to H5AC2_find() except the object is removed from + * the cache and given to the caller, preventing other parts + * of the program from modifying the protected object or + * preempting it from the cache. + * + * The caller must call H5AC2_unprotect() when finished with + * the pointer. + * + * If H5AC2_DEBUG is defined then we check that the + * requested object isn't already protected. + * + * Return: Success: Ptr to the object. + * + * Failure: NULL + * + * Programmer: Robb Matzke + * matzke@llnl.gov + * Sep 2 1997 + * + * Modifications: + * Robb Matzke, 1999-07-27 + * The ADDR argument is passed by value. + * + * Bill Wendling, 2003-09-10 + * Added parameter to indicate whether this is a READ or + * WRITE type of protect. + * + * JRM -- 5/17/04 + * Complete re-write for the new client cache. See revised + * Purpose section above. + * + * JRM - 6/7/04 + * Abstracted the guts of the function to H5C2_protect() + * in H5C2.c, and then re-wrote the function as a wrapper for + * H5C2_protect(). + * + * JRM - 6/6/06 + * Added trace file support. + * + * JRM - 3/18/07 + * Modified code to support the new flags parameter for + * H5C2_protect(). For now, that means passing in the + * H5C2_READ_ONLY_FLAG if rw == H5AC2_READ. + * + * Also updated the trace file output to save the + * rw parameter, since we are now doing something with it. + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ +void * +H5AC2_protect(H5F_t *f, + hid_t dxpl_id, + const H5AC2_class_t *type, + haddr_t addr, + size_t len, + const void *udata, + H5AC2_protect_t rw) +{ + /* char * fcn_name = "H5AC2_protect"; */ + unsigned protect_flags = H5C2__NO_FLAGS_SET; + void * thing = (void *)NULL; + void * ret_value; /* Return value */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + size_t trace_entry_size = 0; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_protect, NULL) + + /* check args */ + HDassert(f); + HDassert(f->shared); + HDassert(f->shared->cache2); + HDassert(type); + HDassert(type->serialize); + HDassert(H5F_addr_defined(addr)); + + /* Check for invalid access request */ + if(0 == (f->intent & H5F_ACC_RDWR) && rw == H5AC2_WRITE) + HGOTO_ERROR(H5E_CACHE, H5E_BADVALUE, NULL, "no write intent on file") + +#if H5AC2__TRACE_FILE_ENABLED + /* For the protect call, only the addr and type id is really necessary + * in the trace file. Include the size of the entry protected as a + * sanity check. Also indicate whether the call was successful to + * catch occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + char * rw_string; + + if ( rw == H5AC2_WRITE ) { + + rw_string = "H5AC2_WRITE"; + + } else if ( rw == H5AC2_READ ) { + + rw_string = "H5AC2_READ"; + + } else { + + rw_string = "???"; + } + + sprintf(trace, "H5AC2_protect 0x%lx %ld %d %s", + (unsigned long)addr, + (long)len, + (int)(type->id), + rw_string); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + if ( rw == H5AC2_READ ) { + + protect_flags |= H5C2__READ_ONLY_FLAG; + } + + thing = H5C2_protect(f->shared->cache2, + dxpl_id, + type, + addr, + len, + udata, + protect_flags); + + if ( thing == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPROTECT, NULL, "H5C2_protect() failed.") + } + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + /* make note of the entry size */ + trace_entry_size = ((H5C2_cache_entry_t *)thing)->size; + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + /* Set return value */ + ret_value = thing; + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d %d\n", trace, + (int)trace_entry_size, + (int)(ret_value != NULL)); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_protect() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_resize_pinned_entry + * + * Purpose: Resize a pinned entry. The target entry MUST be + * be pinned, and MUST not be unprotected. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 7/5/06 + * + * Modifications: + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_resize_pinned_entry(H5F_t * f, + void * thing, + size_t new_size) +{ + H5C2_t *cache_ptr = f->shared->cache2; + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_resize_pinned_entry, FAIL) + +#if H5AC2__TRACE_FILE_ENABLED + /* For the resize pinned entry call, only the addr, and new_size are + * really necessary in the trace file. Write the result to catch + * occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_resize_pinned_entry 0x%lx %d", + (unsigned long)(((H5C2_cache_entry_t *)thing)->addr), + (int)new_size); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + +#ifdef H5_HAVE_PARALLEL + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + + if ( ( ((H5AC2_info_t *)thing)->is_dirty == FALSE ) && + ( NULL != cache_ptr->aux_ptr) ) { + + H5AC2_info_t * entry_ptr; + + entry_ptr = (H5AC2_info_t *)thing; + + if ( ! ( entry_ptr->is_pinned ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, \ + "Entry isn't pinned??") + } + + if ( entry_ptr->is_protected ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, \ + "Entry is protected??") + } + + result = H5AC2_log_dirtied_entry(cache_ptr, + entry_ptr, + entry_ptr->addr, + TRUE, + new_size); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "H5AC2_log_dirtied_entry() failed.") + } + } +#endif /* H5_HAVE_PARALLEL */ + + result = H5C2_resize_pinned_entry(cache_ptr, + thing, + new_size); + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, \ + "H5C2_resize_pinned_entry() failed.") + + } + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_resize_pinned_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_unpin_entry() + * + * Purpose: Unpin a cache entry. The entry must be unprotected at + * the time of call, and must be pinned. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 4/11/06 + * + * Modifications: + * + * Added code supporting the trace file. JRM -- 6/7/06 + * + * Modified code in support of revised cache API needed + * to permit journaling. JRM - 10/18/07 + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_unpin_entry(H5F_t * f, + void * thing) +{ + H5C2_t *cache_ptr = f->shared->cache2; + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_unpin_entry, FAIL) + +#if H5AC2__TRACE_FILE_ENABLED + /* For the unpin entry call, only the addr is really necessary + * in the trace file. Also write the result to catch occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_unpin_entry 0x%lx", + (unsigned long)(((H5C2_cache_entry_t *)thing)->addr)); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + result = H5C2_unpin_entry(cache_ptr, thing); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPIN, FAIL, "H5C2_unpin_entry() failed.") + } + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d\n", trace, (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_unpin_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_unprotect + * + * Purpose: Undo an H5AC2_protect() call -- specifically, mark the + * entry as unprotected, remove it from the protected list, + * and give it back to the replacement policy. + * + * The TYPE and ADDR arguments must be the same as those in + * the corresponding call to H5AC2_protect() and the THING + * argument must be the value returned by that call to + * H5AC2_protect(). + * + * If the deleted flag is TRUE, simply remove the target entry + * from the cache, clear it, and free it without writing it to + * disk. + * + * This verion of the function is a complete re-write to + * use the new metadata cache. While there isn't all that + * much difference between the old and new Purpose sections, + * the original version is given below. + * + * Original purpose section: + * + * This function should be called to undo the effect of + * H5AC2_protect(). The TYPE and ADDR arguments should be the + * same as the corresponding call to H5AC2_protect() and the + * THING argument should be the value returned by H5AC2_protect(). + * If the DELETED flag is set, then this object has been deleted + * from the file and should not be returned to the cache. + * + * If H5AC2_DEBUG is defined then this function fails + * if the TYPE and ADDR arguments are not what was used when the + * object was protected or if the object was never protected. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Robb Matzke + * matzke@llnl.gov + * Sep 2 1997 + * + * Modifications: + * Robb Matzke, 1999-07-27 + * The ADDR argument is passed by value. + * + * Quincey Koziol, 2003-03-19 + * Added "deleted" argument + * + * Bill Wendling, 2003-09-18 + * If this is an FPHDF5 driver and the data is dirty, + * perform a "flush" that writes the data to the SAP. + * + * John Mainzer 5/19/04 + * Complete re-write for the new metadata cache. + * + * JRM - 6/7/04 + * Abstracted the guts of the function to H5C_unprotect() + * in H5C.c, and then re-wrote the function as a wrapper for + * H5C_unprotect(). + * + * JRM - 1/6/05 + * Replaced the deleted parameter with the new flags parameter. + * Since the deleted parameter is not used by the FPHDF5 code, + * the only change in the body is to replace the deleted + * parameter with the flags parameter in the call to + * H5C_unprotect(). + * + * JRM - 6/6/05 + * Added the dirtied flag and supporting code. This is + * part of a collection of changes directed at moving + * management of cache entry dirty flags into the H5C code. + * + * JRM - 7/5/05 + * Added code to track dirty byte generation, and to trigger + * clean entry list propagation when it exceeds a user + * specified threshold. Note that this code only applies in + * the PHDF5 case. It should have no effect on either the + * serial or FPHSD5 cases. + * + * JRM - 9/8/05 + * Added code to track entry size changes. This is necessary + * as it can effect dirty byte creation counts, thereby + * throwing the caches out of sync in the PHDF5 case. + * + * JRM - 5/16/06 + * Added code to use the new dirtied field in + * H5C_cache_entry_t in the test to see if the entry has + * been dirtied. + * + * JRM - 6/7/06 + * Added support for the trace file. + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + * Note that the H5AC2__SIZE_CHANGED_FLAG must now be set if + * the size of the entry has changed. + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_unprotect(H5F_t *f, hid_t dxpl_id, const H5AC2_class_t *type, + haddr_t addr, size_t new_size, void *thing, unsigned flags) +{ + herr_t result; + herr_t ret_value=SUCCEED; /* Return value */ + hbool_t size_changed = FALSE; + hbool_t dirtied; +#ifdef H5_HAVE_PARALLEL + H5AC2_aux_t * aux_ptr = NULL; +#endif /* H5_HAVE_PARALLEL */ +#if H5AC2__TRACE_FILE_ENABLED + char trace[128] = ""; + unsigned trace_flags = 0; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_unprotect, FAIL) + + HDassert(f); + HDassert(f->shared->cache2); + HDassert(type); + HDassert(type->deserialize); + HDassert(H5F_addr_defined(addr)); + HDassert(thing); + HDassert( ((H5AC2_info_t *)thing)->addr == addr ); + HDassert( ((H5AC2_info_t *)thing)->type == type ); + +#if H5AC2__TRACE_FILE_ENABLED + /* For the unprotect call, only the addr, type id, flags, and possible + * new size are really necessary in the trace file. Write the return + * value to catch occult errors. + */ + if ( ( f != NULL ) && + ( f->shared != NULL ) && + ( f->shared->cache2 != NULL ) && + ( H5C2_get_trace_file_ptr(f->shared->cache2, &trace_file_ptr) >= 0) && + ( trace_file_ptr != NULL ) ) { + + sprintf(trace, "H5AC2_unprotect 0x%lx %d", + (unsigned long)addr, + (int)(type->id)); + + trace_flags = flags; + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + dirtied = ( ( (flags & H5AC2__DIRTIED_FLAG) == H5AC2__DIRTIED_FLAG ) || + ( ((H5AC2_info_t *)thing)->dirtied ) ); + + size_changed = ( (flags & H5AC2__SIZE_CHANGED_FLAG) == + H5AC2__SIZE_CHANGED_FLAG ); + +#ifdef H5_HAVE_PARALLEL + if ( ( dirtied ) && ( ((H5AC2_info_t *)thing)->is_dirty == FALSE ) && + ( NULL != (aux_ptr = f->shared->cache2->aux_ptr) ) ) { + + result = H5AC2_log_dirtied_entry(f->shared->cache2, + (H5AC2_info_t *)thing, + addr, + size_changed, + new_size); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "H5AC2_log_dirtied_entry() failed.") + } + } + + if ( ( (flags & H5C2__DELETED_FLAG) != 0 ) && + ( NULL != (aux_ptr = f->shared->cache2->aux_ptr) ) && + ( aux_ptr->mpi_rank == 0 ) ) { + + result = H5AC2_log_deleted_entry(f->shared->cache2, + (H5AC2_info_t *)thing, + addr, + flags); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "H5AC2_log_deleted_entry() failed.") + } + } +#endif /* H5_HAVE_PARALLEL */ + + result = H5C2_unprotect(f->shared->cache2, + dxpl_id, + type, + addr, + thing, + flags, + new_size); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "H5C2_unprotect() failed.") + } + +#ifdef H5_HAVE_PARALLEL + if ( ( aux_ptr != NULL ) && + ( aux_ptr->dirty_bytes >= aux_ptr->dirty_bytes_threshold ) ) { + + result = H5AC2_propagate_flushed_and_still_clean_entries_list(f, + H5AC2_noblock_dxpl_id, + f->shared->cache2, + TRUE); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "Can't propagate clean entries list.") + } + } +#endif /* H5_HAVE_PARALLEL */ + +done: + +#if H5AC2__TRACE_FILE_ENABLED + if ( trace_file_ptr != NULL ) { + + HDfprintf(trace_file_ptr, "%s %d %x %d\n", + trace, + (int)new_size, + (unsigned)flags, + (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_unprotect() */ + + +/*------------------------------------------------------------------------- + * Function: HA5C_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 + * + * Modifications: + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +herr_t +H5AC2_set_write_done_callback(H5C2_t * cache_ptr, + void (* write_done)(void)) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5AC2_aux_t * aux_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_set_write_done_callback, FAIL) + + /* This would normally be an assert, but we need to use an HGOTO_ERROR + * call to shut up the compiler. + */ + if ( ( ! cache_ptr ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr") + } + + aux_ptr = cache_ptr->aux_ptr; + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + + aux_ptr->write_done = write_done; + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_set_write_done_callback() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_stats + * + * Purpose: Prints statistics about the cache. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Robb Matzke + * Thursday, October 30, 1997 + * + * Modifications: + * John Mainzer 5/19/04 + * Re-write to support the new metadata cache. + * + * JRM - 6/7/04 + * Abstracted the guts of the function to H5C_stats() + * in H5C.c, and then re-wrote the function as a wrapper for + * H5C_stats(). + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ +herr_t +H5AC2_stats(const H5F_t *f) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5AC2_stats, FAIL) + + HDassert(f); + HDassert(f->shared->cache2); + + /* at present, this can't fail */ + (void)H5C2_stats(f->shared->cache2, f->name, FALSE); + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_stats() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_get_cache_auto_resize_config + * + * Purpose: Wrapper function for H5C2_get_cache_auto_resize_config(). + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 3/10/05 + * + * Modifications: + * + * JRM - 4/6/05 + * Reworked for the addition of struct H5AC2_cache_config_t. + * + * JRM - 10/25/05 + * Added support for the new dirty_bytes_threshold field of + * both H5AC2_cache_config_t and H5AC2_aux_t. + * + * JRM - 6/8/06 + * Added support for the new trace file related fields. + * + * JRM - 7/28/07 + * Added support for the new evictions enabled related fields. + * + * Observe that H5AC2_get_cache_auto_resize_config() and + * H5AC2_set_cache_auto_resize_config() are becoming generic + * metadata cache configuration routines as they gain + * switches for functions that are only tenuously related + * to auto resize configuration. + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_get_cache_auto_resize_config(H5AC2_t * cache_ptr, + H5AC2_cache_config_t *config_ptr) +{ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + hbool_t evictions_enabled; + H5C2_auto_size_ctl_t internal_config; + + FUNC_ENTER_NOAPI(H5AC2_get_cache_auto_resize_config, FAIL) + + if ( ( cache_ptr == NULL ) + || +#ifdef H5_HAVE_PARALLEL + ( ( cache_ptr->aux_ptr != NULL ) + && + ( ((H5AC2_aux_t *)(cache_ptr->aux_ptr))->magic + != + H5AC2__H5AC2_AUX_T_MAGIC + ) + ) + || +#endif /* H5_HAVE_PARALLEL */ + ( config_ptr == NULL ) + || + ( config_ptr->version != H5AC2__CURR_CACHE_CONFIG_VERSION ) + ) + { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Bad cache_ptr or config_ptr on entry.") + + } + + result = H5C2_get_cache_auto_resize_config((H5C2_t *)cache_ptr, + &internal_config); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_cache_auto_resize_config() failed.") + } + + result = H5C2_get_evictions_enabled((H5C2_t *)cache_ptr, + &evictions_enabled); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_resize_enabled() failed.") + } + + if ( internal_config.rpt_fcn == NULL ) { + + config_ptr->rpt_fcn_enabled = FALSE; + + } else { + + config_ptr->rpt_fcn_enabled = TRUE; + } + + config_ptr->open_trace_file = FALSE; + config_ptr->close_trace_file = FALSE; + config_ptr->trace_file_name[0] = '\0'; + config_ptr->evictions_enabled = evictions_enabled; + config_ptr->set_initial_size = internal_config.set_initial_size; + config_ptr->initial_size = internal_config.initial_size; + config_ptr->min_clean_fraction = internal_config.min_clean_fraction; + config_ptr->max_size = internal_config.max_size; + config_ptr->min_size = internal_config.min_size; + config_ptr->epoch_length = (long)(internal_config.epoch_length); + config_ptr->incr_mode = internal_config.incr_mode; + config_ptr->lower_hr_threshold = internal_config.lower_hr_threshold; + config_ptr->increment = internal_config.increment; + config_ptr->apply_max_increment = internal_config.apply_max_increment; + config_ptr->max_increment = internal_config.max_increment; + config_ptr->decr_mode = internal_config.decr_mode; + config_ptr->upper_hr_threshold = internal_config.upper_hr_threshold; + config_ptr->decrement = internal_config.decrement; + config_ptr->apply_max_decrement = internal_config.apply_max_decrement; + config_ptr->max_decrement = internal_config.max_decrement; + config_ptr->epochs_before_eviction = + (int)(internal_config.epochs_before_eviction); + config_ptr->apply_empty_reserve = internal_config.apply_empty_reserve; + config_ptr->empty_reserve = internal_config.empty_reserve; + +#ifdef H5_HAVE_PARALLEL + if ( cache_ptr->aux_ptr != NULL ) { + + config_ptr->dirty_bytes_threshold = + ((H5AC2_aux_t *)(cache_ptr->aux_ptr))->dirty_bytes_threshold; + + } else { +#endif /* H5_HAVE_PARALLEL */ + + config_ptr->dirty_bytes_threshold = H5AC2__DEFAULT_DIRTY_BYTES_THRESHOLD; + +#ifdef H5_HAVE_PARALLEL + } +#endif /* H5_HAVE_PARALLEL */ + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_get_cache_auto_resize_config() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_get_cache_size + * + * Purpose: Wrapper function for H5C2_get_cache_size(). + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 3/11/05 + * + * Modifications: + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_get_cache_size(H5AC2_t * cache_ptr, + size_t * max_size_ptr, + size_t * min_clean_size_ptr, + size_t * cur_size_ptr, + int32_t * cur_num_entries_ptr) +{ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5AC2_get_cache_size, FAIL) + + result = H5C2_get_cache_size((H5C2_t *)cache_ptr, + max_size_ptr, + min_clean_size_ptr, + cur_size_ptr, + cur_num_entries_ptr); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_cache_size() failed.") + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_get_cache_size() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_get_cache_hit_rate + * + * Purpose: Wrapper function for H5C2_get_cache_hit_rate(). + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 3/10/05 + * + * Modifications: + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_get_cache_hit_rate(H5AC2_t * cache_ptr, + double * hit_rate_ptr) + +{ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5AC2_get_cache_hit_rate, FAIL) + + result = H5C2_get_cache_hit_rate((H5C2_t *)cache_ptr, hit_rate_ptr); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_cache_hit_rate() failed.") + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_get_cache_hit_rate() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_reset_cache_hit_rate_stats() + * + * Purpose: Wrapper function for H5C2_reset_cache_hit_rate_stats(). + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer, 3/10/05 + * + * Modifications: + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_reset_cache_hit_rate_stats(H5AC2_t * cache_ptr) +{ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5AC2_reset_cache_hit_rate_stats, FAIL) + + result = H5C2_reset_cache_hit_rate_stats((H5C2_t *)cache_ptr); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_reset_cache_hit_rate_stats() failed.") + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_reset_cache_hit_rate_stats() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_set_cache_auto_resize_config + * + * Purpose: Wrapper function for H5C2_set_cache_auto_resize_config(). + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 3/10/05 + * + * Modifications: + * + * John Mainzer -- 4/6/05 + * Updated for the addition of H5AC2_cache_config_t. + * + * John Mainzer -- 10/25/05 + * Added support for the new dirty_bytes_threshold field of + * both H5AC2_cache_config_t and H5AC2_aux_t. + * + * John Mainzer -- 6/7/06 + * Added trace file support. + * + * John Mainzer -- 7/28/07 + * Added support for the new evictions enabled related fields. + * + * Observe that H5AC2_get_cache_auto_resize_config() and + * H5AC2_set_cache_auto_resize_config() are becoming generic + * metadata cache configuration routines as they gain + * switches for functions that are only tenuously related + * to auto resize configuration. + * + * JRM - 10/18/07 + * Modified code in support of revised cache API needed + * to permit journaling. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_set_cache_auto_resize_config(H5AC2_t * cache_ptr, + H5AC2_cache_config_t *config_ptr) +{ + /* const char * fcn_name = "H5AC2_set_cache_auto_resize_config"; */ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_auto_size_ctl_t internal_config; +#if H5AC2__TRACE_FILE_ENABLED + H5AC2_cache_config_t trace_config = H5AC2__DEFAULT_CACHE_CONFIG; + FILE * trace_file_ptr = NULL; +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_ENTER_NOAPI(H5AC2_set_cache_auto_resize_config, FAIL) + +#if H5AC2__TRACE_FILE_ENABLED + /* Make note of the new configuration. Don't look up the trace file + * pointer, as that may change before we use it. + */ + if ( config_ptr != NULL ) { + + trace_config = *config_ptr; + + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + if ( ( cache_ptr == NULL ) +#ifdef H5_HAVE_PARALLEL + || + ( ( cache_ptr->aux_ptr != NULL ) + && + ( + ((H5AC2_aux_t *)(cache_ptr->aux_ptr))->magic + != + H5AC2__H5AC2_AUX_T_MAGIC + ) + ) +#endif /* H5_HAVE_PARALLEL */ + ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "bad cache_ptr on entry.") + } + + result = H5AC2_validate_config(config_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "Bad cache configuration"); + } + + if ( config_ptr->open_trace_file ) { + + FILE * file_ptr = NULL; + + if ( H5C2_get_trace_file_ptr(cache_ptr, &file_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_trace_file_ptr() failed.") + } + + if ( ( ! ( config_ptr->close_trace_file ) ) && + ( file_ptr != NULL ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "Trace file already open.") + } + } + + if ( + ( + config_ptr->dirty_bytes_threshold + < + H5AC2__MIN_DIRTY_BYTES_THRESHOLD + ) + || + ( + config_ptr->dirty_bytes_threshold + > + H5AC2__MAX_DIRTY_BYTES_THRESHOLD + ) + ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, + "config_ptr->dirty_bytes_threshold out of range.") + } + + if ( config_ptr->close_trace_file ) { + + if ( H5AC2_close_trace_file(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5AC2_close_trace_file() failed.") + } + } + + if ( config_ptr->open_trace_file ) { + + if ( H5AC2_open_trace_file(cache_ptr, config_ptr->trace_file_name) < 0 ) + { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "H5AC2_open_trace_file() failed.") + } + } + + if ( H5AC2_ext_config_2_int_config(config_ptr, &internal_config) != + SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5AC2_ext_config_2_int_config() failed.") + } + + result = H5C2_set_cache_auto_resize_config((H5C2_t *)cache_ptr, + &internal_config); + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_set_cache_auto_resize_config() failed.") + } + + + result = H5C2_set_evictions_enabled((H5C2_t *)cache_ptr, + config_ptr->evictions_enabled); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_set_evictions_enabled() failed.") + } + +#ifdef H5_HAVE_PARALLEL + if ( cache_ptr->aux_ptr != NULL ) { + + ((H5AC2_aux_t *)(cache_ptr->aux_ptr))->dirty_bytes_threshold = + config_ptr->dirty_bytes_threshold; + } +#endif /* H5_HAVE_PARALLEL */ + +done: + +#if H5AC2__TRACE_FILE_ENABLED + /* For the set cache auto resize config call, only the contents + * of the config is necessary in the trace file. Write the return + * value to catch occult errors. + */ + if ( ( cache_ptr != NULL ) && + ( H5C2_get_trace_file_ptr(cache_ptr, &trace_file_ptr) >= 0 ) && + ( trace_file_ptr != NULL ) ) { + + HDfprintf(trace_file_ptr, + "%s %d %d %d %d \"%s\" %d %d %d %f %d %d %ld %d %f %f %d %d %d %f %f %d %d %d %d %f %d %d\n", + "H5AC2_set_cache_auto_resize_config", + trace_config.version, + (int)(trace_config.rpt_fcn_enabled), + (int)(trace_config.open_trace_file), + (int)(trace_config.close_trace_file), + trace_config.trace_file_name, + (int)(trace_config.evictions_enabled), + (int)(trace_config.set_initial_size), + (int)(trace_config.initial_size), + trace_config.min_clean_fraction, + (int)(trace_config.max_size), + (int)(trace_config.min_size), + trace_config.epoch_length, + (int)(trace_config.incr_mode), + trace_config.lower_hr_threshold, + trace_config.increment, + (int)(trace_config.apply_max_increment), + (int)(trace_config.max_increment), + (int)(trace_config.decr_mode), + trace_config.upper_hr_threshold, + trace_config.decrement, + (int)(trace_config.apply_max_decrement), + (int)(trace_config.max_decrement), + trace_config.epochs_before_eviction, + (int)(trace_config.apply_empty_reserve), + trace_config.empty_reserve, + trace_config.dirty_bytes_threshold, + (int)ret_value); + } +#endif /* H5AC2__TRACE_FILE_ENABLED */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_set_cache_auto_resize_config() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_validate_config() + * + * Purpose: Run a sanity check on the contents of the supplied + * instance of H5AC2_cache_config_t. + * + * Do nothing and return SUCCEED if no errors are detected, + * and flag an error and return FAIL otherwise. + * + * At present, this function operates by packing the data + * from the instance of H5AC2_cache_config_t into an instance + * of H5C2_auto_size_ctl_t, and then calling + * H5C2_validate_resize_config(). As H5AC2_cache_config_t and + * H5C2_auto_size_ctl_t diverge, we may have to change this. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 4/6/05 + * + * Modifications: + * + * - Added code testing the trace file configuration fields. + * These tests are not comprehensive, as many errors cannot + * be caught until the directives contained in these fields + * are applied. + * JRM - 5/15/06 + * + * - Added code testing the evictions enabled field. At + * present this consists of verifying that if + * evictions_enabled is FALSE, then automatic cache + * resizing in disabled. + * + * JRM - 7/28/07 + * + * Modified code in support of revised cache API needed + * to permit journaling. + * JRM - 10/18/07 + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_validate_config(H5AC2_cache_config_t * config_ptr) + +{ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + int name_len; + H5C2_auto_size_ctl_t internal_config; + + FUNC_ENTER_NOAPI(H5AC2_validate_config, FAIL) + + if ( config_ptr == NULL ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "NULL config_ptr on entry.") + } + + if ( config_ptr->version != H5AC2__CURR_CACHE_CONFIG_VERSION ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "Unknown config version.") + } + + if ( ( config_ptr->rpt_fcn_enabled != TRUE ) && + ( config_ptr->rpt_fcn_enabled != FALSE ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "config_ptr->rpt_fcn_enabled must be either TRUE or FALSE.") + } + + if ( ( config_ptr->open_trace_file != TRUE ) && + ( config_ptr->open_trace_file != FALSE ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "config_ptr->open_trace_file must be either TRUE or FALSE.") + } + + if ( ( config_ptr->close_trace_file != TRUE ) && + ( config_ptr->close_trace_file != FALSE ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "config_ptr->close_trace_file must be either TRUE or FALSE.") + } + + /* don't bother to test trace_file_name unless open_trace_file is TRUE */ + if ( config_ptr->open_trace_file ) { + + /* Can't really test the trace_file_name field without trying to + * open the file, so we will content ourselves with a couple of + * sanity checks on the length of the file name. + */ + name_len = HDstrlen(config_ptr->trace_file_name); + + if ( name_len <= 0 ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "config_ptr->trace_file_name is empty.") + + } else if ( name_len > H5AC2__MAX_TRACE_FILE_NAME_LEN ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "config_ptr->trace_file_name too long.") + } + } + + if ( ( config_ptr->evictions_enabled != TRUE ) && + ( config_ptr->evictions_enabled != FALSE ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "config_ptr->evictions_enabled must be either TRUE or FALSE.") + } + + if ( ( config_ptr->evictions_enabled == FALSE ) && + ( ( config_ptr->incr_mode != H5C2_incr__off ) || + ( config_ptr->incr_mode != H5C2_decr__off ) ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "Can't disable evictions while auto-resize is enabled.") + } + + if ( config_ptr->dirty_bytes_threshold < H5AC2__MIN_DIRTY_BYTES_THRESHOLD ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, + "dirty_bytes_threshold too small.") + } else + if ( config_ptr->dirty_bytes_threshold > H5AC2__MAX_DIRTY_BYTES_THRESHOLD ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, + "dirty_bytes_threshold too big.") + } + + if ( H5AC2_ext_config_2_int_config(config_ptr, &internal_config) != + SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5AC2_ext_config_2_int_config() failed.") + } + + result = H5C2_validate_resize_config(&internal_config, + H5C2_RESIZE_CFG__VALIDATE_ALL); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "error(s) in new config.") + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_validate_config() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_close_trace_file() + * + * Purpose: If a trace file is open, stop logging calls to the cache, + * and close the file. + * + * Note that the function does nothing if there is no trace + * file. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/2/06 + * + * Modifications: + * + * Modified code in support of revised cache API needed + * to permit journaling. + * JRM - 10/18/07 + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_close_trace_file(H5AC2_t * cache_ptr) + +{ + herr_t ret_value = SUCCEED; /* Return value */ + FILE * trace_file_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_close_trace_file, FAIL) + + if ( cache_ptr == NULL ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "NULL cache_ptr on entry.") + } + + if ( H5C2_get_trace_file_ptr(cache_ptr, &trace_file_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_trace_file_ptr() failed.") + } + + if ( trace_file_ptr != NULL ) { + + if ( H5C2_set_trace_file_ptr(cache_ptr, NULL) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_set_trace_file_ptr() failed.") + } + + if ( HDfclose(trace_file_ptr) != 0 ) { + + HGOTO_ERROR(H5E_FILE, H5E_CANTCLOSEFILE, FAIL, \ + "can't close metadata cache trace file") + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_close_trace_file() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_open_trace_file() + * + * Purpose: Open a trace file, and start logging calls to the cache. + * + * This logging is done at the H5C2 level, and will only take + * place if H5C2_TRACE_FILE_ENABLED (defined in H5C2private.h) + * is TRUE. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/1/06 + * + * Modifications: + * + * Modified code in support of revised cache API needed + * to permit journaling. + * JRM - 10/18/07 + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_open_trace_file(H5AC2_t * cache_ptr, + const char * trace_file_name) +{ + herr_t ret_value = SUCCEED; /* Return value */ + char file_name[H5AC2__MAX_TRACE_FILE_NAME_LEN + H5C2__PREFIX_LEN + 2]; + FILE * file_ptr = NULL; +#ifdef H5_HAVE_PARALLEL + H5AC2_aux_t * aux_ptr = NULL; +#endif /* H5_HAVE_PARALLEL */ + + FUNC_ENTER_NOAPI(H5AC2_open_trace_file, FAIL) + + HDassert(cache_ptr); + + if ( cache_ptr == NULL ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "cache_ptr NULL on entry.") + } + + if ( trace_file_name == NULL ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "NULL trace_file_name on entry.") + } + + if ( HDstrlen(trace_file_name) > H5AC2__MAX_TRACE_FILE_NAME_LEN ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "trace file name too long.") + } + + if ( H5C2_get_trace_file_ptr(cache_ptr, &file_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_get_trace_file_ptr() failed.") + } + + if ( file_ptr != NULL ) { + + HGOTO_ERROR(H5E_FILE, H5E_FILEOPEN, FAIL, "trace file already open.") + } + +#ifdef H5_HAVE_PARALLEL + + aux_ptr = (H5AC2_aux_t *)(cache_ptr->aux_ptr); + + if ( cache_ptr->aux_ptr == NULL ) { + + sprintf(file_name, "%s", trace_file_name); + + } else { + + if ( aux_ptr->magic != H5AC2__H5AC2_AUX_T_MAGIC ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad aux_ptr->magic.") + } + + sprintf(file_name, "%s.%d", trace_file_name, aux_ptr->mpi_rank); + + } + + if ( HDstrlen(file_name) > + H5AC2__MAX_TRACE_FILE_NAME_LEN + H5C2__PREFIX_LEN + 1 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "cooked trace file name too long.") + } + +#else /* H5_HAVE_PARALLEL */ + + sprintf(file_name, "%s", trace_file_name); + +#endif /* H5_HAVE_PARALLEL */ + + if ( (file_ptr = HDfopen(file_name, "w")) == NULL ) { + + /* trace file open failed */ + HGOTO_ERROR(H5E_FILE, H5E_CANTOPENFILE, FAIL, "trace file open failed.") + } + + HDfprintf(file_ptr, "### HDF5 metadata cache trace file version 1 ###\n"); + + if ( H5C2_set_trace_file_ptr(cache_ptr, file_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_set_trace_file_ptr() failed.") + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_open_trace_file() */ + + +/*************************************************************************/ +/**************************** Private Functions: *************************/ +/*************************************************************************/ + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_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 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +static herr_t +H5AC2_broadcast_clean_list(H5AC2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + haddr_t addr; + H5AC2_aux_t * aux_ptr = NULL; + H5SL_node_t * slist_node_ptr = NULL; + H5AC2_slist_entry_t * slist_entry_ptr = NULL; + MPI_Offset * buf_ptr = NULL; + size_t buf_size; + int i = 0; + int mpi_result; + int num_entries; + + FUNC_ENTER_NOAPI(H5AC2_broadcast_clean_list, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = (H5AC2_aux_t *)(cache_ptr->aux_ptr); + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + HDassert( aux_ptr->mpi_rank == 0 ); + HDassert( aux_ptr->c_slist_ptr != NULL ); + HDassert( H5SL_count(aux_ptr->c_slist_ptr) == + (size_t)(aux_ptr->c_slist_len) ); + + + /* 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 = aux_ptr->c_slist_len; + + mpi_result = MPI_Bcast(&num_entries, 1, MPI_INT, 0, aux_ptr->mpi_comm); + + if ( mpi_result != MPI_SUCCESS ) { + + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed 1", mpi_result) + + } + + if ( num_entries > 0 ) + { + /* allocate a buffer to store the list of entry base addresses in */ + + buf_size = sizeof(MPI_Offset) * (size_t)num_entries; + + buf_ptr = (MPI_Offset *)H5MM_malloc(buf_size); + + if ( buf_ptr == NULL ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "memory allocation failed for clean entry buffer") + } + + /* now load the entry base addresses into the buffer, emptying the + * cleaned entry list in passing + */ + + while ( NULL != (slist_node_ptr = H5SL_first(aux_ptr->c_slist_ptr) ) ) + { + slist_entry_ptr = H5SL_item(slist_node_ptr); + + HDassert(slist_entry_ptr->magic == H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC); + + HDassert( i < num_entries ); + + addr = slist_entry_ptr->addr; + + if ( H5FD_mpi_haddr_to_MPIOff(addr, &(buf_ptr[i])) < 0 ) { + + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, \ + "can't convert from haddr to MPI off") + } + + i++; + + /* now remove the entry from the cleaned entry list */ + if ( H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from cleaned entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->c_slist_len -= 1; + + HDassert( aux_ptr->c_slist_len >= 0 ); + + /* and also remove the matching entry from the dirtied list + * if it exists. + */ + if ( (slist_entry_ptr = H5SL_search(aux_ptr->d_slist_ptr, + (void *)(&addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == + H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC ); + HDassert( slist_entry_ptr->addr == addr ); + + if ( H5SL_remove(aux_ptr->d_slist_ptr, (void *)(&addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from dirty entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->d_slist_len -= 1; + + HDassert( aux_ptr->d_slist_len >= 0 ); + } + + } /* while */ + + + /* Now broadcast the list of cleaned entries -- if there is one. + * + * The peculiar structure of the following call to MPI_Bcast is + * due to MPI's (?) failure to believe in the MPI_Offset type. + * Thus the element type is MPI_BYTE, with size equal to the + * buf_size computed above. + */ + + mpi_result = MPI_Bcast((void *)buf_ptr, (int)buf_size, MPI_BYTE, 0, + aux_ptr->mpi_comm); + + if ( mpi_result != MPI_SUCCESS ) { + + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed 2", mpi_result) + } + } + +done: + + if ( buf_ptr != NULL ) { + + buf_ptr = (MPI_Offset *)H5MM_xfree((void *)buf_ptr); + } + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_broadcast_clean_list() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_check_if_write_permitted + * + * Purpose: Determine if a write is permitted under the current + * circumstances, and set *write_permitted_ptr accordingly. + * As a general rule it is, but when we are running in parallel + * mode with collective I/O, we must ensure that a read cannot + * cause a write. + * + * In the event of failure, the value of *write_permitted_ptr + * is undefined. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 5/15/04 + * + * Modifications: + * + * John Mainzer, 9/23/05 + * Rewrote function to return the value of the + * write_permitted field in aux structure if the structure + * exists and mpi_rank is 0. + * + * If the aux structure exists, but mpi_rank isn't 0, the + * function now returns FALSE. + * + * In all other cases, the function returns TRUE. + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +static herr_t +H5AC2_check_if_write_permitted(const H5F_t *f, + hid_t UNUSED dxpl_id, + hbool_t * write_permitted_ptr) +#else /* H5_HAVE_PARALLEL */ +static herr_t +H5AC2_check_if_write_permitted(const H5F_t UNUSED * f, + hid_t UNUSED dxpl_id, + hbool_t * write_permitted_ptr) +#endif /* H5_HAVE_PARALLEL */ +{ + hbool_t write_permitted = TRUE; + herr_t ret_value = SUCCEED; /* Return value */ +#ifdef H5_HAVE_PARALLEL + H5AC2_aux_t * aux_ptr = NULL; +#endif /* H5_HAVE_PARALLEL */ + + + FUNC_ENTER_NOAPI(H5AC2_check_if_write_permitted, FAIL) + +#ifdef H5_HAVE_PARALLEL + HDassert( f != NULL ); + HDassert( f->shared != NULL ); + HDassert( f->shared->cache2 != NULL ); + + aux_ptr = (H5AC2_aux_t *)(f->shared->cache2->aux_ptr); + + if ( aux_ptr != NULL ) { + + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + + if ( aux_ptr->mpi_rank == 0 ) { + + write_permitted = aux_ptr->write_permitted; + + } else { + + write_permitted = FALSE; + } + } +#endif /* H5_HAVE_PARALLEL */ + + *write_permitted_ptr = write_permitted; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_check_if_write_permitted() */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_ext_config_2_int_config() + * + * Purpose: Utility function to translate an instance of + * H5AC2_cache_config_t to an instance of H5C2_auto_size_ctl_t. + * + * Places translation in *int_conf_ptr and returns SUCCEED + * if successful. Returns FAIL on failure. + * + * Does only minimal sanity checking. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 1/26/06 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5AC2_ext_config_2_int_config(H5AC2_cache_config_t * ext_conf_ptr, + H5C2_auto_size_ctl_t * int_conf_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5AC2_ext_config_2_int_config, FAIL) + + if ( ( ext_conf_ptr == NULL ) || + ( ext_conf_ptr->version != H5AC2__CURR_CACHE_CONFIG_VERSION ) || + ( int_conf_ptr == NULL ) ) { + + } + + int_conf_ptr->version = H5C2__CURR_AUTO_SIZE_CTL_VER; + + if ( ext_conf_ptr->rpt_fcn_enabled ) { + + int_conf_ptr->rpt_fcn = H5C2_def_auto_resize_rpt_fcn; + + } else { + + int_conf_ptr->rpt_fcn = NULL; + } + + int_conf_ptr->set_initial_size = ext_conf_ptr->set_initial_size; + int_conf_ptr->initial_size = ext_conf_ptr->initial_size; + int_conf_ptr->min_clean_fraction = ext_conf_ptr->min_clean_fraction; + int_conf_ptr->max_size = ext_conf_ptr->max_size; + int_conf_ptr->min_size = ext_conf_ptr->min_size; + int_conf_ptr->epoch_length = + (int64_t)(ext_conf_ptr->epoch_length); + + int_conf_ptr->incr_mode = ext_conf_ptr->incr_mode; + int_conf_ptr->lower_hr_threshold = ext_conf_ptr->lower_hr_threshold; + int_conf_ptr->increment = ext_conf_ptr->increment; + int_conf_ptr->apply_max_increment = ext_conf_ptr->apply_max_increment; + int_conf_ptr->max_increment = ext_conf_ptr->max_increment; + + int_conf_ptr->decr_mode = ext_conf_ptr->decr_mode; + int_conf_ptr->upper_hr_threshold = ext_conf_ptr->upper_hr_threshold; + int_conf_ptr->decrement = ext_conf_ptr->decrement; + int_conf_ptr->apply_max_decrement = ext_conf_ptr->apply_max_decrement; + int_conf_ptr->max_decrement = ext_conf_ptr->max_decrement; + int_conf_ptr->epochs_before_eviction = + (int32_t)(ext_conf_ptr->epochs_before_eviction); + int_conf_ptr->apply_empty_reserve = ext_conf_ptr->apply_empty_reserve; + int_conf_ptr->empty_reserve = ext_conf_ptr->empty_reserve; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_ext_config_2_int_config() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_log_deleted_entry() + * + * Purpose: Log an entry for which H5C2__DELETED_FLAG has been set. + * + * If mpi_rank is 0, we must make sure that the entry doesn't + * appear in the cleaned or dirty entry lists. Otherwise, + * we have nothing to do. + * + * Return SUCCEED on success, and FAIL on failure. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 6/29/05 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +static herr_t +H5AC2_log_deleted_entry(H5AC2_t * cache_ptr, + H5AC2_info_t * entry_ptr, + haddr_t addr, + unsigned int flags) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5AC2_aux_t * aux_ptr = NULL; + H5AC2_slist_entry_t * slist_entry_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_log_deleted_entry, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = cache_ptr->aux_ptr; + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + + HDassert( entry_ptr != NULL ); + HDassert( entry_ptr->addr == addr ); + + HDassert( (flags & H5C2__DELETED_FLAG) != 0 ); + + if ( 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 ( (slist_entry_ptr = H5SL_search(aux_ptr->d_slist_ptr, + (void *)(&addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == + H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC ); + HDassert( slist_entry_ptr->addr == addr ); + + if ( H5SL_remove(aux_ptr->d_slist_ptr, (void *)(&addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from dirty entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->d_slist_len -= 1; + + HDassert( aux_ptr->d_slist_len >= 0 ); + } + + /* if the entry appears in the cleaned entry slist, remove it. */ + if ( (slist_entry_ptr = H5SL_search(aux_ptr->c_slist_ptr, + (void *)(&addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == + H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC ); + HDassert( slist_entry_ptr->addr == addr ); + + if ( H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from cleaned entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->c_slist_len -= 1; + + HDassert( aux_ptr->c_slist_len >= 0 ); + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_log_deleted_entry() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_log_dirtied_entry() + * + * Purpose: Update the dirty_bytes count for a newly dirtied entry. + * + * If mpi_rank isnt 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 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +static herr_t +H5AC2_log_dirtied_entry(H5AC2_t * cache_ptr, + H5AC2_info_t * entry_ptr, + haddr_t addr, + hbool_t size_changed, + size_t new_size) +{ + herr_t ret_value = SUCCEED; /* Return value */ + size_t entry_size; + H5AC2_aux_t * aux_ptr = NULL; + H5AC2_slist_entry_t * slist_entry_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_log_dirtied_entry, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = cache_ptr->aux_ptr; + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + + HDassert( entry_ptr != NULL ); + HDassert( entry_ptr->addr == addr ); + HDassert( entry_ptr->is_dirty == FALSE ); + + if ( size_changed ) { + + entry_size = new_size; + + } else { + + entry_size = entry_ptr->size; + } + + if ( aux_ptr->mpi_rank == 0 ) { + + HDassert( aux_ptr->d_slist_ptr != NULL ); + HDassert( aux_ptr->c_slist_ptr != NULL ); + + if ( H5SL_search(aux_ptr->d_slist_ptr, (void *)(&addr)) == NULL ) { + + /* 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_CALLOC(H5AC2_slist_entry_t)) ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "Can't allocate dirty slist entry .") + } + + slist_entry_ptr->magic = H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC; + 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->d_slist_len += 1; + aux_ptr->dirty_bytes += entry_size; +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->unprotect_dirty_bytes += entry_size; + aux_ptr->unprotect_dirty_bytes_updates += 1; +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + } + + if ( H5SL_search(aux_ptr->c_slist_ptr, (void *)(&addr)) != NULL ) { + + /* the entry is dirty. If it exists on the cleaned entries list, + * remove it. + */ + if ( (slist_entry_ptr = H5SL_search(aux_ptr->c_slist_ptr, + (void *)(&addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == + H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC ); + HDassert( slist_entry_ptr->addr == addr ); + + if ( H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from clean entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->c_slist_len -= 1; + + HDassert( aux_ptr->c_slist_len >= 0 ); + } + } + } else { + + aux_ptr->dirty_bytes += entry_size; +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->unprotect_dirty_bytes += entry_size; + aux_ptr->unprotect_dirty_bytes_updates += 1; +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_log_dirtied_entry() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_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 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +#if 0 /* This is useful debugging code. -- JRM */ +static herr_t +H5AC2_log_flushed_entry_dummy(H5C2_t * cache_ptr, + haddr_t addr, + hbool_t was_dirty, + unsigned flags, + int type_id) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5AC2_aux_t * aux_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_log_flushed_entry_dummy, FAIL) + + aux_ptr = cache_ptr->aux_ptr; + + if ( ( was_dirty ) && ( (flags & H5C2__FLUSH_CLEAR_ONLY_FLAG) == 0 ) ) { + + HDfprintf(stdout, + "%d:H5AC2_log_flushed_entry(): addr = %d, flags = %x, was_dirty = %d, type_id = %d\n", + (int)(aux_ptr->mpi_rank), (int)addr, flags, (int)was_dirty, type_id); + } +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_log_flushed_entry_dummy() */ +#endif /* JRM */ + +static herr_t +H5AC2_log_flushed_entry(H5C2_t * cache_ptr, + haddr_t addr, + hbool_t was_dirty, + unsigned flags, + int UNUSED type_id) +{ + herr_t ret_value = SUCCEED; /* Return value */ + hbool_t cleared; + H5AC2_aux_t * aux_ptr; + H5AC2_slist_entry_t * slist_entry_ptr = NULL; + + + FUNC_ENTER_NOAPI(H5AC2_log_flushed_entry, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = cache_ptr->aux_ptr; + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + HDassert( aux_ptr->mpi_rank == 0 ); + HDassert( aux_ptr->c_slist_ptr != NULL ); + + cleared = ( (flags & H5C2__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 ( (slist_entry_ptr = H5SL_search(aux_ptr->c_slist_ptr, + (void *)(&addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC); + HDassert( slist_entry_ptr->addr == addr ); + + if ( H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from clean entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->c_slist_len -= 1; + + HDassert( aux_ptr->c_slist_len >= 0 ); + } + + if ( (slist_entry_ptr = H5SL_search(aux_ptr->d_slist_ptr, + (void *)(&addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC); + HDassert( slist_entry_ptr->addr == addr ); + + if ( H5SL_remove(aux_ptr->d_slist_ptr, (void *)(&addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from dirty entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->d_slist_len -= 1; + + HDassert( aux_ptr->d_slist_len >= 0 ); + } + } else if ( was_dirty ) { + + if ( H5SL_search(aux_ptr->c_slist_ptr, (void *)(&addr)) == NULL ) { + + /* insert the address of the entry in the clean entry list. */ + + if ( NULL == (slist_entry_ptr = H5FL_CALLOC(H5AC2_slist_entry_t)) ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "Can't allocate clean slist entry .") + } + + slist_entry_ptr->magic = H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC; + 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.") + } + + aux_ptr->c_slist_len += 1; + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_log_flushed_entry() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_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 + * + * Modifications: + * + * JRM -- 10/24/07 + * Added the size parameter. + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +static herr_t +H5AC2_log_inserted_entry(H5F_t * f, + H5AC2_t * cache_ptr, + H5AC2_info_t * entry_ptr, + const H5AC2_class_t * type, + haddr_t addr, + size_t size) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5AC2_aux_t * aux_ptr = NULL; + H5AC2_slist_entry_t * slist_entry_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_log_inserted_entry, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = cache_ptr->aux_ptr; + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + + HDassert( entry_ptr != NULL ); + HDassert( entry_ptr->addr == addr ); + HDassert( entry_ptr->type == type ); + + if ( aux_ptr->mpi_rank == 0 ) { + + HDassert( aux_ptr->d_slist_ptr != NULL ); + HDassert( aux_ptr->c_slist_ptr != NULL ); + + if ( H5SL_search(aux_ptr->d_slist_ptr, (void *)(&addr)) == NULL ) { + + /* 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_CALLOC(H5AC2_slist_entry_t)) ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "Can't allocate dirty slist entry .") + } + + slist_entry_ptr->magic = H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC; + 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->d_slist_len += 1; + + } else { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Inserted entry already in dirty slist.") + } + + if ( H5SL_search(aux_ptr->c_slist_ptr, (void *)(&addr)) != NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Inserted entry in clean slist.") + } + } + + aux_ptr->dirty_bytes += size; + +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->insert_dirty_bytes += size; + aux_ptr->insert_dirty_bytes_updates += 1; +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_log_inserted_entry() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_log_renamed_entry() + * + * Purpose: Update the dirty_bytes count for a renamed entry. + * + * WARNING + * + * At present, the way that the rename call is used ensures + * that the renamed entry is present in all caches by + * renaming in a collective operation and immediately after + * unprotecting the target entry. + * + * This function uses this invarient, and will cause arcane + * failures if it is not met. If maintaining this invarient + * 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 + * rename 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 rename, 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 rename. 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 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +static herr_t +H5AC2_log_renamed_entry(H5AC2_t * cache_ptr, + haddr_t old_addr, + haddr_t new_addr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + hbool_t entry_in_cache; + hbool_t entry_dirty; + size_t entry_size; + H5AC2_aux_t * aux_ptr = NULL; + H5AC2_slist_entry_t * slist_entry_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_log_renamed_entry, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = cache_ptr->aux_ptr; + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + + /* get entry status, size, etc here */ + if ( H5C2_get_entry_status(cache_ptr, old_addr, &entry_size, &entry_in_cache, + &entry_dirty, NULL, NULL) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't get entry status.") + + } else if ( ! entry_in_cache ) { + + HDassert( entry_in_cache ); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "entry not in cache.") + } + + if ( 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 cleaned entry slist, under its old + * address, remove it. + */ + if ( (slist_entry_ptr = H5SL_search(aux_ptr->c_slist_ptr, + (void *)(&old_addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == + H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC ); + HDassert( slist_entry_ptr->addr == old_addr ); + + if ( H5SL_remove(aux_ptr->c_slist_ptr, (void *)(&old_addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from cleaned entry slist.") + } + + slist_entry_ptr->magic = 0; + H5FL_FREE(H5AC2_slist_entry_t, slist_entry_ptr); + slist_entry_ptr = NULL; + + aux_ptr->c_slist_len -= 1; + + HDassert( aux_ptr->c_slist_len >= 0 ); + } + + /* 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 ( (slist_entry_ptr = H5SL_search(aux_ptr->d_slist_ptr, + (void *)(&old_addr))) != NULL ) { + + HDassert( slist_entry_ptr->magic == + H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC ); + HDassert( slist_entry_ptr->addr == old_addr ); + + if ( H5SL_remove(aux_ptr->d_slist_ptr, (void *)(&old_addr)) + != slist_entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTDELETE, FAIL, \ + "Can't delete entry from dirty entry slist.") + } + + slist_entry_ptr->addr = new_addr; + + aux_ptr->d_slist_len -= 1; + + HDassert( aux_ptr->d_slist_len >= 0 ); + + } 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_CALLOC(H5AC2_slist_entry_t)) ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "Can't allocate dirty slist entry .") + } + + slist_entry_ptr->magic = H5AC2__H5AC2_SLIST_ENTRY_T_MAGIC; + slist_entry_ptr->addr = new_addr; + + aux_ptr->dirty_bytes += entry_size; + +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->rename_dirty_bytes += entry_size; + aux_ptr->rename_dirty_bytes_updates += 1; +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + } + + /* verify that there is no entry at new_addr in the dirty slist */ + if ( H5SL_search(aux_ptr->d_slist_ptr, (void *)(&new_addr)) != NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "dirty slist already contains entry at new_addr.") + } + + /* 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.") + } + + aux_ptr->d_slist_len += 1; + + } else if ( ! entry_dirty ) { + + aux_ptr->dirty_bytes += entry_size; + +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + aux_ptr->rename_dirty_bytes += entry_size; + aux_ptr->rename_dirty_bytes_updates += 1; +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_log_renamed_entry() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * Function: H5AC2_propagate_flushed_and_still_clean_entries_list + * + * Purpose: In PHDF5, 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 that proceedure. + * It must be called simultaniously on all processes that + * have the relevant file open. To this end, there must + * be a barrier immediately prior to this call. + * + * Typicaly, this will be done 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) Enable writes, flush the cache to its min clean size, + * and then disable writes again. + * + * 2) 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. + * + * 3) 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 + * + * Modifications: + * + * JRM -- 5/11/06 + * Added code to call the write_done callback. + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +herr_t +H5AC2_propagate_flushed_and_still_clean_entries_list(H5F_t * f, + hid_t dxpl_id, + H5AC2_t * cache_ptr, + hbool_t do_barrier) +{ + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; + int mpi_code; + H5AC2_aux_t * aux_ptr = NULL; + + FUNC_ENTER_NOAPI(H5AC2_propagate_flushed_and_still_clean_entries_list, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = (H5AC2_aux_t *)(cache_ptr->aux_ptr); + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + HDfprintf(stdout, + "%d:H5AC2_propagate...:%d: (u/uu/i/iu/r/ru) = %d/%d/%d/%d/%d/%d\n", + (int)(aux_ptr->mpi_rank), + (int)(aux_ptr->dirty_bytes_propagations), + (int)(aux_ptr->unprotect_dirty_bytes), + (int)(aux_ptr->unprotect_dirty_bytes_updates), + (int)(aux_ptr->insert_dirty_bytes), + (int)(aux_ptr->insert_dirty_bytes_updates), + (int)(aux_ptr->rename_dirty_bytes), + (int)(aux_ptr->rename_dirty_bytes_updates)); +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + + if ( do_barrier ) { + + /* 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_code = MPI_Barrier(aux_ptr->mpi_comm)) ) { + + HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_code) + } + } + + if ( aux_ptr->mpi_rank == 0 ) { + + aux_ptr->write_permitted = TRUE; + + result = H5C2_flush_to_min_clean(cache_ptr, dxpl_id); + + aux_ptr->write_permitted = FALSE; + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_flush_to_min_clean() failed.") + } + + if ( aux_ptr->write_done != NULL ) { + + (aux_ptr->write_done)(); + } + + if ( H5AC2_broadcast_clean_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Can't broadcast clean slist.") + } + + HDassert( aux_ptr->c_slist_len == 0 ); + + } else { + + if ( H5AC2_receive_and_apply_clean_list(f, dxpl_id, + cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Can't receive and/or process clean slist broadcast.") + } + } + + aux_ptr->dirty_bytes = 0; +#if H5AC2_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 /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_propagate_flushed_and_still_clean_entries_list() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * + * Function: H5AC2_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 + * + * Modifications: + * + * JRM --10/24/07 + * Reworked parameter list in support of the revised cache + * API. + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +static herr_t +H5AC2_receive_and_apply_clean_list(H5F_t * f, + hid_t dxpl_id, + H5AC2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5AC2_aux_t * aux_ptr = NULL; + haddr_t * haddr_buf_ptr = NULL; + MPI_Offset * MPI_Offset_buf_ptr = NULL; + size_t buf_size; + int i = 0; + int mpi_result; + int num_entries; + + FUNC_ENTER_NOAPI(H5AC2_receive_and_apply_clean_list, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + aux_ptr = (H5AC2_aux_t *)(cache_ptr->aux_ptr); + + HDassert( aux_ptr != NULL ); + HDassert( aux_ptr->magic == H5AC2__H5AC2_AUX_T_MAGIC ); + HDassert( aux_ptr->mpi_rank != 0 ); + + /* 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. + */ + mpi_result = MPI_Bcast(&num_entries, 1, MPI_INT, 0, aux_ptr->mpi_comm); + + if ( mpi_result != MPI_SUCCESS ) { + + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed 1", mpi_result) + } + + if ( num_entries > 0 ) + { + /* allocate a buffers to store the list of entry base addresses in */ + + buf_size = sizeof(MPI_Offset) * (size_t)num_entries; + + MPI_Offset_buf_ptr = (MPI_Offset *)H5MM_malloc(buf_size); + + if ( MPI_Offset_buf_ptr == NULL ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "memory allocation failed for receive buffer") + } + + haddr_buf_ptr = (haddr_t *)H5MM_malloc(sizeof(haddr_t) * + (size_t)num_entries); + + if ( haddr_buf_ptr == NULL ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "memory allocation failed for haddr buffer") + } + + + /* Now receive the list of cleaned entries + * + * The peculiar structure of the following call to MPI_Bcast is + * due to MPI's (?) failure to believe in the MPI_Offset type. + * Thus the element type is MPI_BYTE, with size equal to the + * buf_size computed above. + */ + + mpi_result = MPI_Bcast((void *)MPI_Offset_buf_ptr, (int)buf_size, + MPI_BYTE, 0, aux_ptr->mpi_comm); + + if ( mpi_result != MPI_SUCCESS ) { + + HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed 2", mpi_result) + } + + + /* translate the MPI_Offsets to haddr_t */ + i = 0; + while ( i < num_entries ) + { + haddr_buf_ptr[i] = H5FD_mpi_MPIOff_to_haddr(MPI_Offset_buf_ptr[i]); + + if ( haddr_buf_ptr[i] == HADDR_UNDEF ) { + + HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, \ + "can't convert MPI off to haddr") + } + + i++; + } + + + /* mark the indicated entries as clean */ + if ( H5C2_mark_entries_as_clean(cache_ptr, dxpl_id, + (int32_t)num_entries, + &(haddr_buf_ptr[0])) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Can't mark entries clean.") + + } + } + +done: + + if ( MPI_Offset_buf_ptr != NULL ) { + + MPI_Offset_buf_ptr = + (MPI_Offset *)H5MM_xfree((void *)MPI_Offset_buf_ptr); + } + + if ( haddr_buf_ptr != NULL ) { + + haddr_buf_ptr = (haddr_t *)H5MM_xfree((void *)haddr_buf_ptr); + } + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5AC2_receive_and_apply_clean_list() */ +#endif /* H5_HAVE_PARALLEL */ + diff --git a/src/H5AC2pkg.h b/src/H5AC2pkg.h new file mode 100644 index 0000000..202f4e3 --- /dev/null +++ b/src/H5AC2pkg.h @@ -0,0 +1,335 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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. * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +/* + * Programmer: John Mainzer -- 4/19/06 + * + * Purpose: This file contains declarations which are normally visible + * only within the H5AC2 package (just H5AC2.c at present). + * + * Source files outside the H5AC package should include + * H5ACprivate.h instead. + * + * The one exception to this rule is testpar/t_cache.c. The + * test code is easier to write if it can look at H5AC_aux_t. + * Indeed, this is the main reason why this file was created. + * + * Modifications: + * + * JRM - 10/18/07 + * Copied H5ACpkg.h to H5AC2pkg.h and reworked + * to use H5C2 instead of H5C. All this is in support + * of cache API modifications needed for journaling. + * + */ + +#ifndef H5AC2_PACKAGE +#error "Do not include this file outside the H5AC2 package!" +#endif + +#ifndef _H5AC2pkg_H +#define _H5AC2pkg_H + +#define H5C2_PACKAGE /*suppress error about including H5C2pkg */ + +/* Get package's private header */ +#include "H5AC2private.h" +#include "H5C2private.h" + + +/* Get needed headers */ +#include "H5C2pkg.h" /* Cache */ +#include "H5SLprivate.h" /* Skip lists */ + + +#define H5AC2_DEBUG_DIRTY_BYTES_CREATION 0 + +/*------------------------------------------------------------------------- + * It is a bit difficult to set ranges of allowable values on the + * dirty_bytes_threshold field of H5AC2_aux_t. The following are + * probably broader than they should be. + *------------------------------------------------------------------------- + */ + +#define H5AC2__MIN_DIRTY_BYTES_THRESHOLD (int32_t) \ + (H5C2__MIN_MAX_CACHE_SIZE / 2) +#define H5AC2__DEFAULT_DIRTY_BYTES_THRESHOLD (256 * 1024) +#define H5AC2__MAX_DIRTY_BYTES_THRESHOLD (int32_t) \ + (H5C2__MAX_MAX_CACHE_SIZE / 4) + +/**************************************************************************** + * + * structure H5AC2_aux_t + * + * While H5AC2 has become a wrapper for the cache implemented in H5C2.c, there + * are some features of the metadata cache that are specific to it, and which + * therefore do not belong in the more generic H5C2 cache code. + * + * In particular, there is the matter of synchronizing writes from the + * metadata cache to disk in the PHDF5 case. + * + * Prior to this update, the presumption was that all metadata caches would + * write the same data at the same time since all operations modifying + * metadata must be performed collectively. Given this assumption, it was + * safe to allow only the writes from process 0 to actually make it to disk, + * while metadata writes from all other processes were discarded. + * + * Unfortunately, this presumption is in error as operations that read + * metadata need not be collective, but can change the location of dirty + * entries in the metadata cache LRU lists. This can result in the same + * metadata write operation triggering writes from the metadata caches on + * some processes, but not all (causing a hang), or in different sets of + * entries being written from different caches (potentially resulting in + * metadata corruption in the file). + * + * To deal with this issue, I decided to apply a paradigm shift to the way + * metadata is written to disk. + * + * With this set of changes, only the metadata cache on process 0 is able + * to write metadata to disk, although metadata caches on all other + * processes can read metadata from disk as before. + * + * To keep all the other caches from getting plugged up with dirty metadata, + * process 0 periodically broadcasts a list of entries that it has flushed + * since that last notice, and which are currently clean. The other caches + * mark these entries as clean as well, which allows them to evict the + * entries as needed. + * + * One obvious problem in this approach is synchronizing the broadcasts + * and receptions, as different caches may see different amounts of + * activity. + * + * The current solution is for the caches to track the number of bytes + * of newly generated dirty metadata, and to broadcast and receive + * whenever this value exceeds some user specified threshold. + * + * Maintaining this count is easy for all processes not on process 0 -- + * all that is necessary is to add the size of the entry to the total + * whenever there is an insertion, a rename of a previously clean entry, + * or whever a previously clean entry is marked dirty in an unprotect. + * + * On process 0, we have to be careful not to count dirty bytes twice. + * If an entry is marked dirty, flushed, and marked dirty again, all + * within a single reporting period, it only th first marking should + * be added to the dirty bytes generated tally, as that is all that + * the other processes will see. + * + * At present, this structure exists to maintain the fields needed to + * implement the above scheme, and thus is only used in the parallel + * case. However, other uses may arise in the future. + * + * Instance of this structure are associated with metadata caches via + * the aux_ptr field of H5C2_t (see H5C2pkg.h). The H5AC2 code is + * responsible for allocating, maintaining, and discarding instances + * of H5AC2_aux_t. + * + * The remainder of this header comments documents the individual fields + * of the structure. + * + * JRM - 6/27/05 + * + * magic: Unsigned 32 bit integer always set to + * H5AC2__H5AC2_AUX_T_MAGIC. This field is used to validate + * pointers to instances of H5AC2_aux_t. + * + * mpi_comm: MPI communicator associated with the file for which the + * cache has been created. + * + * mpi_rank: MPI rank of this process within mpi_comm. + * + * mpi_size: Number of processes in mpi_comm. + * + * write_permitted: Boolean flag used to control whether the cache + * is permitted to write to file. + * + * dirty_bytes_threshold: Integer field containing the dirty bytes + * generation threashold. Whenever dirty byte creation + * exceeds this value, the metadata cache on process 0 + * broadcasts a list of the entries it has flushed since + * the last broadcast (or since the beginning of execution) + * and which are currently clean (if they are still in the + * cache) + * + * Similarly, metadata caches on processes other than process + * 0 will attempt to receive a list of clean entries whenever + * the threshold is exceeded. + * + * dirty_bytes: Integer field containing the number of bytes of dirty + * metadata generated since the beginning of the computation, + * or (more typically) since the last clean entries list + * broadcast. This field is reset to zero after each such + * broadcast. + * + * dirty_bytes_propagations: This field only exists when the + * H5AC2_DEBUG_DIRTY_BYTES_CREATION #define is TRUE. + * + * It is used to track the number of times the cleaned list + * has been propagated from process 0 to the other + * processes. + * + * unprotect_dirty_bytes: This field only exists when the + * H5AC2_DEBUG_DIRTY_BYTES_CREATION #define is TRUE. + * + * It is used to track the number of dirty bytes created + * via unprotect operations since the last time the cleaned + * list was propagated. + * + * unprotect_dirty_bytes_updates: This field only exists when the + * H5AC2_DEBUG_DIRTY_BYTES_CREATION #define is TRUE. + * + * It is used to track the number of times dirty bytes have + * been created via unprotect operations since the last time + * the cleaned list was propagated. + * + * insert_dirty_bytes: This field only exists when the + * H5AC2_DEBUG_DIRTY_BYTES_CREATION #define is TRUE. + * + * It is used to track the number of dirty bytes created + * via insert operations since the last time the cleaned + * list was propagated. + * + * insert_dirty_bytes_updates: This field only exists when the + * H5AC2_DEBUG_DIRTY_BYTES_CREATION #define is TRUE. + * + * It is used to track the number of times dirty bytes have + * been created via insert operations since the last time + * the cleaned list was propagated. + * + * rename_dirty_bytes: This field only exists when the + * H5AC2_DEBUG_DIRTY_BYTES_CREATION #define is TRUE. + * + * It is used to track the number of dirty bytes created + * via rename operations since the last time the cleaned + * list was propagated. + * + * rename_dirty_bytes_updates: This field only exists when the + * H5AC2_DEBUG_DIRTY_BYTES_CREATION #define is TRUE. + * + * It is used to track the number of times dirty bytes have + * been created via rename operations since the last time + * the cleaned list was propagated. + * + * d_slist_ptr: Pointer to an instance of H5SL_t used to maintain a list + * of entries that have been dirtied since the last time they + * were listed in a clean entries broadcast. This list is + * only maintained by the metadata cache on process 0 -- it + * it used to maintain a view of the dirty entries as seen + * by the other caches, so as to keep the dirty bytes count + * in synchronization with them. + * + * Thus on process 0, the dirty_bytes count is incremented + * only if either + * + * 1) an entry is inserted in the metadata cache, or + * + * 2) a previously clean entry is renamed, and it does not + * already appear in the dirty entry list, or + * + * 3) a previously clean entry is unprotected with the + * dirtied flag set and the entry does not already appear + * in the dirty entry list. + * + * Entries are added to the dirty entry list whever they cause + * the dirty bytes count to be increased. They are removed + * when they appear in a clean entries broadcast. Note that + * renames must be reflected in the dirty entry list. + * + * To reitterate, this field is only used on process 0 -- it + * should be NULL on all other processes. + * + * d_slist_len: Integer field containing the number of entries in the + * dirty entry list. This field should always contain the + * value 0 on all processes other than process 0. It exists + * primarily for sanity checking. + * + * c_slist_ptr: Pointer to an instance of H5SL_t used to maintain a list + * of entries that were dirty, have been flushed + * to disk since the last clean entries broadcast, and are + * still clean. Since only process 0 can write to disk, this + * list only exists on process 0. + * + * In essence, this slist is used to assemble the contents of + * the next clean entries broadcast. The list emptied after + * each broadcast. + * + * c_slist_len: Integer field containing the number of entries in the clean + * entries list (*c_slist_ptr). This field should always + * contain the value 0 on all processes other than process 0. + * It exists primarily for sanity checking. + * + * write_done: In the parallel test bed, it is necessary to ensure that + * all writes to the server process from cache 0 complete + * before it enters the barrier call with the other caches. + * + * The write_done callback allows t_cache to do this without + * requiring an ACK on each write. Since these ACKs greatly + * increase the run time on some platforms, this is a + * significant optimization. + * + * This field must be set to NULL when the callback is not + * needed. + * + ****************************************************************************/ + +#ifdef H5_HAVE_PARALLEL + +#define H5AC2__H5AC2_AUX_T_MAGIC (unsigned)0x00D0A02 + +typedef struct H5AC2_aux_t +{ + uint32_t magic; + + MPI_Comm mpi_comm; + + int mpi_rank; + + int mpi_size; + + hbool_t write_permitted; + + int32_t dirty_bytes_threshold; + + int32_t dirty_bytes; + +#if H5AC2_DEBUG_DIRTY_BYTES_CREATION + + int32_t dirty_bytes_propagations; + + int32_t unprotect_dirty_bytes; + int32_t unprotect_dirty_bytes_updates; + + int32_t insert_dirty_bytes; + int32_t insert_dirty_bytes_updates; + + int32_t rename_dirty_bytes; + int32_t rename_dirty_bytes_updates; + +#endif /* H5AC2_DEBUG_DIRTY_BYTES_CREATION */ + + H5SL_t * d_slist_ptr; + + int32_t d_slist_len; + + H5SL_t * c_slist_ptr; + + int32_t c_slist_len; + + void (* write_done)(void); + +} H5AC2_aux_t; /* struct H5AC2_aux_t */ + +#endif /* H5_HAVE_PARALLEL */ + +#endif /* _H5C2pkg_H */ diff --git a/src/H5AC2private.h b/src/H5AC2private.h new file mode 100644 index 0000000..de76b40 --- /dev/null +++ b/src/H5AC2private.h @@ -0,0 +1,315 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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: H5AC2private.h + * Jul 9 1997 + * Robb Matzke <matzke@llnl.gov> + * + * Purpose: Constants and typedefs available to the rest of the + * library. + * + * Modifications: JRM - 6/4/04 + * Complete re-write for a new caching algorithm + * located in H5C.c + * + * JRM - 10/18/07 + * Copied H5ACprivate.h to H5AC2private.h and reworked + * to use H5C2 instead of H5C. All this is in support + * of cache API modifications needed for journaling. + * + *------------------------------------------------------------------------- + */ + +#ifndef _H5AC2private_H +#define _H5AC2private_H + +#include "H5AC2public.h" /*public prototypes */ + +/* Pivate headers needed by this header */ +#include "H5private.h" /* Generic Functions */ +#include "H5Fprivate.h" /* File access */ +#include "H5C2private.h" /* cache */ + +#ifdef H5_METADATA_TRACE_FILE +#define H5AC2__TRACE_FILE_ENABLED 1 +#else /* H5_METADATA_TRACE_FILE */ +#define H5AC2__TRACE_FILE_ENABLED 0 +#endif /* H5_METADATA_TRACE_FILE */ + +/* Types of metadata objects cached */ +typedef enum { + H5AC2_BT_ID = 0, /*B-tree nodes */ + H5AC2_SNODE_ID, /*symbol table nodes */ + H5AC2_LHEAP_ID, /*local heap */ + H5AC2_GHEAP_ID, /*global heap */ + H5AC2_OHDR_ID, /*object header */ + H5AC2_BT2_HDR_ID, /*v2 B-tree header */ + H5AC2_BT2_INT_ID, /*v2 B-tree internal node */ + H5AC2_BT2_LEAF_ID, /*v2 B-tree leaf node */ + H5AC2_FHEAP_HDR_ID, /*fractal heap header */ + H5AC2_FHEAP_DBLOCK_ID,/*fractal heap direct block */ + H5AC2_FHEAP_IBLOCK_ID,/*fractal heap indirect block */ + H5AC2_FSPACE_HDR_ID, /*free space header */ + H5AC2_FSPACE_SINFO_ID,/*free space sections */ + H5AC2_SOHM_TABLE_ID, /*shared object header message master table */ + H5AC2_SOHM_LIST_ID, /*shared message index stored as a list */ + H5AC2_TEST_ID, /*test entry -- not used for actual files */ + H5AC2_NTYPES /* Number of types, must be last */ +} H5AC2_type_t; + +/* H5AC2_DUMP_STATS_ON_CLOSE should always be FALSE when + * H5C2_COLLECT_CACHE_STATS is FALSE. + * + * When H5C2_COLLECT_CACHE_STATS is TRUE, H5AC2_DUMP_STATS_ON_CLOSE must + * be FALSE for "make check" to succeed, but may be set to TRUE at other + * times for debugging purposes. + * + * Hence the following, somewhat odd set of #defines. + */ +#if H5C2_COLLECT_CACHE_STATS + +#define H5AC2_DUMP_STATS_ON_CLOSE 0 + +#else /* H5C2_COLLECT_CACHE_STATS */ + +#define H5AC2_DUMP_STATS_ON_CLOSE 0 + +#endif /* H5C2_COLLECT_CACHE_STATS */ + +/* Default max metadata cache size and min clean size are give here. + * At present, these are the same as those given in H5C2private.h. + */ + +#define H5AC2__DEFAULT_MAX_CACHE_SIZE H5C2__DEFAULT_MAX_CACHE_SIZE +#define H5AC2__DEFAULT_MIN_CLEAN_SIZE H5C2__DEFAULT_MIN_CLEAN_SIZE + + +/* + * Class methods pertaining to caching. Each type of cached object will + * have a constant variable with permanent life-span that describes how + * to cache the object. That variable will be of type H5AC2_class_t and + * have the following required fields... + * + * LOAD: Loads an object from disk to memory. The function + * should allocate some data structure and return it. + * + * FLUSH: Writes some data structure back to disk. It would be + * wise for the data structure to include dirty flags to + * indicate whether it really needs to be written. This + * function is also responsible for freeing memory allocated + * by the LOAD method if the DEST argument is non-zero (by + * calling the DEST method). + * + * DEST: Just frees memory allocated by the LOAD method. + * + * CLEAR: Just marks object as non-dirty. + * + * SIZE: Report the size (on disk) of the specified cache object. + * Note that the space allocated on disk may not be contiguous. + */ + +#define H5AC2__SERIALIZE_RESIZED_FLAG H5C2__SERIALIZE_RESIZED_FLAG +#define H5AC2__SERIALIZE_RENAMED_FLAG H5C2__SERIALIZE_RENAMED_FLAG + +typedef H5C2_deserialize_func_t H5AC2_deserialize_func_t; +typedef H5C2_image_len_func_t H5AC2_image_len_func_t; +typedef H5C2_serialize_func_t H5AC2_serialize_func_t; +typedef H5C2_free_icr_func_t H5CA2_free_icr_func_t; +typedef H5C2_clear_dirty_bits_func_t H5AC2_clear_dirty_bits_func_t; + +typedef H5C2_class_t H5AC2_class_t; + + + +typedef H5C2_cache_entry_t H5AC2_info_t; + + +/*===----------------------------------------------------------------------=== + * Protect Types + *===----------------------------------------------------------------------=== + * + * These are for the wrapper functions to H5AC2_protect. They specify what + * type of operation you're planning on doing to the metadata. The + * Flexible Parallel HDF5 locking can then act accordingly. + */ + +typedef enum H5AC2_protect_t { + H5AC2_WRITE, /* Protect object for writing */ + H5AC2_READ /* Protect object for reading */ +} H5AC2_protect_t; + + +/* Typedef for metadata cache (defined in H5C2pkg.h) */ +typedef H5C2_t H5AC2_t; + +/* Metadata specific properties for FAPL */ +/* (Only used for parallel I/O) */ +#ifdef H5_HAVE_PARALLEL +/* Definitions for "block before metadata write" property */ +#define H5AC2_BLOCK_BEFORE_META_WRITE_NAME "H5AC2_block_before_meta_write" +#define H5AC2_BLOCK_BEFORE_META_WRITE_SIZE sizeof(unsigned) +#define H5AC2_BLOCK_BEFORE_META_WRITE_DEF 0 + +/* Definitions for "library internal" property */ +#define H5AC2_LIBRARY_INTERNAL_NAME "H5AC2_library_internal" +#define H5AC2_LIBRARY_INTERNAL_SIZE sizeof(unsigned) +#define H5AC2_LIBRARY_INTERNAL_DEF 0 +#endif /* H5_HAVE_PARALLEL */ + +/* Dataset transfer property list for flush calls */ +/* (Collective set, "block before metadata write" set and "library internal" set) */ +/* (Global variable declaration, definition is in H5AC2.c) */ +extern hid_t H5AC2_dxpl_id; + +/* Dataset transfer property list for independent metadata I/O calls */ +/* (just "library internal" set - i.e. independent transfer mode) */ +/* (Global variable declaration, definition is in H5AC2.c) */ +extern hid_t H5AC2_ind_dxpl_id; + + +/* Default cache configuration. */ + +#define H5AC2__DEFAULT_CACHE_CONFIG \ +{ \ + /* int version = */ H5C2__CURR_AUTO_SIZE_CTL_VER, \ + /* hbool_t rpt_fcn_enabled = */ FALSE, \ + /* hbool_t open_trace_file = */ FALSE, \ + /* hbool_t close_trace_file = */ FALSE, \ + /* char trace_file_name[] = */ "", \ + /* hbool_t evictions_enabled = */ TRUE, \ + /* hbool_t set_initial_size = */ TRUE, \ + /* size_t initial_size = */ ( 1 * 1024 * 1024 ), \ + /* double min_clean_fraction = */ 0.5, \ + /* size_t max_size = */ (16 * 1024 * 1024 ), \ + /* size_t min_size = */ ( 1 * 1024 * 1024 ), \ + /* long int epoch_length = */ 50000, \ + /* enum H5C2_cache_incr_mode incr_mode = */ H5C2_incr__threshold, \ + /* double lower_hr_threshold = */ 0.9, \ + /* double increment = */ 2.0, \ + /* hbool_t apply_max_increment = */ TRUE, \ + /* size_t max_increment = */ (4 * 1024 * 1024), \ + /* enum H5C2_cache_decr_mode decr_mode = */ \ + H5C2_decr__age_out_with_threshold, \ + /* double upper_hr_threshold = */ 0.999, \ + /* double decrement = */ 0.9, \ + /* hbool_t apply_max_decrement = */ TRUE, \ + /* size_t max_decrement = */ (1 * 1024 * 1024), \ + /* int epochs_before_eviction = */ 3, \ + /* hbool_t apply_empty_reserve = */ TRUE, \ + /* double empty_reserve = */ 0.1, \ + /* int dirty_bytes_threshold = */ (256 * 1024) \ +} + + +/* + * Library prototypes. + */ + +/* #defines of flags used in the flags parameters in some of the + * following function calls. Note that they are just copies of + * the equivalent flags from H5C2private.h. + */ + +#define H5AC2__NO_FLAGS_SET H5C2__NO_FLAGS_SET +#define H5AC2__SET_FLUSH_MARKER_FLAG H5C2__SET_FLUSH_MARKER_FLAG +#define H5AC2__DELETED_FLAG H5C2__DELETED_FLAG +#define H5AC2__DIRTIED_FLAG H5C2__DIRTIED_FLAG +#define H5AC2__SIZE_CHANGED_FLAG H5C2__SIZE_CHANGED_FLAG +#define H5AC2__PIN_ENTRY_FLAG H5C2__PIN_ENTRY_FLAG +#define H5AC2__UNPIN_ENTRY_FLAG H5C2__UNPIN_ENTRY_FLAG +#define H5AC2__FLUSH_INVALIDATE_FLAG H5C2__FLUSH_INVALIDATE_FLAG +#define H5AC2__FLUSH_CLEAR_ONLY_FLAG H5C2__FLUSH_CLEAR_ONLY_FLAG +#define H5AC2__FLUSH_MARKED_ENTRIES_FLAG H5C2__FLUSH_MARKED_ENTRIES_FLAG +#define H5AC2__FLUSH_IGNORE_PROTECTED_FLAG H5C2__FLUSH_IGNORE_PROTECTED_FLAG + + +/* #defines of flags used to report entry status in the + * H5AC2_get_entry_status() call. + */ + +#define H5AC2_ES__IN_CACHE 0x0001 +#define H5AC2_ES__IS_DIRTY 0x0002 +#define H5AC2_ES__IS_PROTECTED 0x0004 +#define H5AC2_ES__IS_PINNED 0x0008 + + +/* external function declarations: */ + +H5_DLL herr_t H5AC2_init(void); +H5_DLL herr_t H5AC2_create(const H5F_t *f, H5AC2_cache_config_t *config_ptr); +H5_DLL herr_t H5AC2_get_entry_status(H5F_t * f, haddr_t addr, + unsigned * status_ptr); +H5_DLL herr_t H5AC2_set(H5F_t *f, hid_t dxpl_id, const H5AC2_class_t *type, + haddr_t addr, size_t len, void *thing, + unsigned int flags); +H5_DLL herr_t H5AC2_pin_protected_entry(H5F_t * f, void * thing); +H5_DLL void * H5AC2_protect(H5F_t *f, hid_t dxpl_id, const H5AC2_class_t *type, + haddr_t addr, size_t len, const void *udata, + H5AC2_protect_t rw); +H5_DLL herr_t H5AC2_resize_pinned_entry(H5F_t * f, + void * thing, + size_t new_size); +H5_DLL herr_t H5AC2_unpin_entry(H5F_t * f, + void * thing); +H5_DLL herr_t H5AC2_unprotect(H5F_t *f, hid_t dxpl_id, + const H5AC2_class_t *type, haddr_t addr, + size_t new_size, void *thing, unsigned flags); +H5_DLL herr_t H5AC2_flush(H5F_t *f, hid_t dxpl_id, unsigned flags); +H5_DLL herr_t H5AC2_mark_pinned_entry_dirty(H5F_t * f, + void * thing, + hbool_t size_changed, + size_t new_size); +H5_DLL herr_t H5AC2_mark_pinned_or_protected_entry_dirty(H5F_t * f, + void * thing); +H5_DLL herr_t H5AC2_rename(H5F_t *f, const H5AC2_class_t *type, + haddr_t old_addr, haddr_t new_addr); + +H5_DLL herr_t H5AC2_dest(H5F_t *f, hid_t dxpl_id); + +H5_DLL herr_t H5AC2_expunge_entry(H5F_t *f, hid_t dxpl_id, + const H5AC2_class_t *type, haddr_t addr); + +H5_DLL herr_t H5AC2_set_write_done_callback(H5C2_t * cache_ptr, + void (* write_done)(void)); +H5_DLL herr_t H5AC2_stats(const H5F_t *f); + +H5_DLL herr_t H5AC2_get_cache_auto_resize_config(H5AC2_t * cache_ptr, + H5AC2_cache_config_t *config_ptr); + +H5_DLL herr_t H5AC2_get_cache_size(H5AC2_t * cache_ptr, + size_t * max_size_ptr, + size_t * min_clean_size_ptr, + size_t * cur_size_ptr, + int32_t * cur_num_entries_ptr); + +H5_DLL herr_t H5AC2_get_cache_hit_rate(H5AC2_t * cache_ptr, + double * hit_rate_ptr); + +H5_DLL herr_t H5AC2_reset_cache_hit_rate_stats(H5AC2_t * cache_ptr); + +H5_DLL herr_t H5AC2_set_cache_auto_resize_config(H5AC2_t * cache_ptr, + H5AC2_cache_config_t *config_ptr); + +H5_DLL herr_t H5AC2_validate_config(H5AC2_cache_config_t * config_ptr); + +H5_DLL herr_t H5AC2_close_trace_file(H5AC2_t * cache_ptr); + +H5_DLL herr_t H5AC2_open_trace_file(H5AC2_t * cache_ptr, + const char * trace_file_name); + +#endif /* !_H5AC2private_H */ + diff --git a/src/H5AC2public.h b/src/H5AC2public.h new file mode 100644 index 0000000..3c88ef1 --- /dev/null +++ b/src/H5AC2public.h @@ -0,0 +1,393 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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: H5ACpublic.h + * Jul 10 1997 + * Robb Matzke <matzke@llnl.gov> + * + * Purpose: Public include file for cache functions. + * + * Modifications: + * + *------------------------------------------------------------------------- + */ +#ifndef _H5AC2public_H +#define _H5AC2public_H + +/* Public headers needed by this file */ +#include "H5public.h" +#include "H5C2public.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define H5AC2__MAX_TRACE_FILE_NAME_LEN 1024 + +/**************************************************************************** + * + * structure H5AC2_cache_config_t + * + * H5AC2_cache_config_t is a public structure intended for use in public APIs. + * At least in its initial incarnation, it is basicaly a copy of struct + * H5C2_auto_size_ctl_t, minus the report_fcn field, and plus the + * dirty_bytes_threshold field. + * + * The report_fcn field is omitted, as including it would require us to + * make H5C2_t structure public. + * + * The dirty_bytes_threshold field does not appear in H5C2_auto_size_ctl_t, + * as synchronization between caches on different processes is handled at + * the H5AC2 level, not at the level of H5C2. Note however that there is + * considerable interaction between this value and the other fields in this + * structure. + * + * Similarly, the open_trace_file, close_trace_file, and trace_file_name + * fields do not appear in H5C2_auto_size_ctl_t, as most trace file + * issues are handled at the H5AC2 level. The one exception is storage of + * the pointer to the trace file, which is handled by H5C2. + * + * The structure is in H5AC2public.h as we may wish to allow different + * configuration options for metadata and raw data caches. + * + * The fields of the structure are discussed individually below: + * + * version: Integer field containing the version number of this version + * of the H5AC2_cache_config_t structure. Any instance of + * H5AC2_cache_config_t passed to the cache must have a known + * version number, or an error will be flagged. + * + * rpt_fcn_enabled: Boolean field used to enable and disable the default + * reporting function. This function is invoked every time the + * automatic cache resize code is run, and reports on its activities. + * + * This is a debugging function, and should normally be turned off. + * + * open_trace_file: Boolean field indicating whether the trace_file_name + * field should be used to open a trace file for the cache. + * + * The trace file is a debuging feature that allow the capture of + * top level metadata cache requests for purposes of debugging and/or + * optimization. This field should normally be set to FALSE, as + * trace file collection imposes considerable overhead. + * + * This field should only be set to TRUE when the trace_file_name + * contains the full path of the desired trace file, and either + * there is no open trace file on the cache, or the close_trace_file + * field is also TRUE. + * + * close_trace_file: Boolean field indicating whether the current trace + * file (if any) should be closed. + * + * See the above comments on the open_trace_file field. This field + * should be set to FALSE unless there is an open trace file on the + * cache that you wish to close. + * + * trace_file_name: Full path of the trace file to be opened if the + * open_trace_file field is TRUE. + * + * In the parallel case, an ascii representation of the mpi rank of + * the process will be appended to the file name to yield a unique + * trace file name for each process. + * + * The length of the path must not exceed H5AC2__MAX_TRACE_FILE_NAME_LEN + * characters. + * + * evictions_enabled: Boolean field used to either report the current + * evictions enabled status of the cache, or to set the cache's + * evictions enabled status. + * + * In general, the metadata cache should always be allowed to + * evict entries. However, in some cases it is advantageous to + * disable evictions briefly, and thereby postpone metadata + * writes. However, this must be done with care, as the cache + * can grow quickly. If you do this, re-enable evictions as + * soon as possible and monitor cache size. + * + * At present, evictions can only be disabled if automatic + * cache resizing is also disabled (that is, ( incr_mode == + * H5C2_incr__off ) && ( decr_mode == H5C2_decr__off )). There + * is no logical reason why this should be so, but it simplifies + * implementation and testing, and I can't think of any reason + * why it would be desireable. If you can think of one, I'll + * revisit the issue. + * + * set_initial_size: Boolean flag indicating whether the size of the + * initial size of the cache is to be set to the value given in + * the initial_size field. If set_initial_size is FALSE, the + * initial_size field is ignored. + * + * initial_size: If enabled, this field contain the size the cache is + * to be set to upon receipt of this structure. Needless to say, + * initial_size must lie in the closed interval [min_size, max_size]. + * + * min_clean_fraction: double in the range 0 to 1 indicating the fraction + * of the cache that is to be kept clean. This field is only used + * in parallel mode. Typical values are 0.1 to 0.5. + * + * max_size: Maximum size to which the cache can be adjusted. The + * supplied value must fall in the closed interval + * [MIN_MAX_CACHE_SIZE, MAX_MAX_CACHE_SIZE]. Also, max_size must + * be greater than or equal to min_size. + * + * min_size: Minimum size to which the cache can be adjusted. The + * supplied value must fall in the closed interval + * [H5C2__MIN_MAX_CACHE_SIZE, H5C2__MAX_MAX_CACHE_SIZE]. Also, min_size + * must be less than or equal to max_size. + * + * epoch_length: Number of accesses on the cache over which to collect + * hit rate stats before running the automatic cache resize code, + * if it is enabled. + * + * At the end of an epoch, we discard prior hit rate data and start + * collecting afresh. The epoch_length must lie in the closed + * interval [H5C2__MIN_AR_EPOCH_LENGTH, H5C2__MAX_AR_EPOCH_LENGTH]. + * + * + * Cache size increase control fields: + * + * incr_mode: Instance of the H5C2_cache_incr_mode enumerated type whose + * value indicates how we determine whether the cache size should be + * increased. At present there are two possible values: + * + * H5C2_incr__off: Don't attempt to increase the size of the cache + * automatically. + * + * When this increment mode is selected, the remaining fields + * in the cache size increase section ar ignored. + * + * H5C2_incr__threshold: Attempt to increase the size of the cache + * whenever the average hit rate over the last epoch drops + * below the value supplied in the lower_hr_threshold + * field. + * + * Note that this attempt will fail if the cache is already + * at its maximum size, or if the cache is not already using + * all available space. + * + * Note that you must set decr_mode to H5C2_incr__off if you + * disable metadata cache entry evictions. + * + * lower_hr_threshold: Lower hit rate threshold. If the increment mode + * (incr_mode) is H5C2_incr__threshold and the hit rate drops below the + * value supplied in this field in an epoch, increment the cache size by + * size_increment. Note that cache size may not be incremented above + * max_size, and that the increment may be further restricted by the + * max_increment field if it is enabled. + * + * When enabled, this field must contain a value in the range [0.0, 1.0]. + * Depending on the incr_mode selected, it may also have to be less than + * upper_hr_threshold. + * + * increment: Double containing the multiplier used to derive the new + * cache size from the old if a cache size increment is triggered. + * The increment must be greater than 1.0, and should not exceed 2.0. + * + * The new cache size is obtained my multiplying the current max cache + * size by the increment, and then clamping to max_size and to stay + * within the max_increment as necessary. + * + * apply_max_increment: Boolean flag indicating whether the max_increment + * field should be used to limit the maximum cache size increment. + * + * max_increment: If enabled by the apply_max_increment field described + * above, this field contains the maximum number of bytes by which the + * cache size can be increased in a single re-size. + * + * + * Cache size decrease control fields: + * + * decr_mode: Instance of the H5C2_cache_decr_mode enumerated type whose + * value indicates how we determine whether the cache size should be + * decreased. At present there are four possibilities. + * + * H5C2_decr__off: Don't attempt to decrease the size of the cache + * automatically. + * + * When this increment mode is selected, the remaining fields + * in the cache size decrease section are ignored. + * + * H5C2_decr__threshold: Attempt to decrease the size of the cache + * whenever the average hit rate over the last epoch rises + * above the value supplied in the upper_hr_threshold + * field. + * + * H5C2_decr__age_out: At the end of each epoch, search the cache for + * entries that have not been accessed for at least the number + * of epochs specified in the epochs_before_eviction field, and + * evict these entries. Conceptually, the maximum cache size + * is then decreased to match the new actual cache size. However, + * this reduction may be modified by the min_size, the + * max_decrement, and/or the empty_reserve. + * + * H5C2_decr__age_out_with_threshold: Same as age_out, but we only + * attempt to reduce the cache size when the hit rate observed + * over the last epoch exceeds the value provided in the + * upper_hr_threshold field. + * + * Note that you must set decr_mode to H5C2_decr__off if you + * disable metadata cache entry evictions. + * + * upper_hr_threshold: Upper hit rate threshold. The use of this field + * varies according to the current decr_mode: + * + * H5C2_decr__off or H5C2_decr__age_out: The value of this field is + * ignored. + * + * H5C2_decr__threshold: If the hit rate exceeds this threshold in any + * epoch, attempt to decrement the cache size by size_decrement. + * + * Note that cache size may not be decremented below min_size. + * + * Note also that if the upper_threshold is 1.0, the cache size + * will never be reduced. + * + * H5C2_decr__age_out_with_threshold: If the hit rate exceeds this + * threshold in any epoch, attempt to reduce the cache size + * by evicting entries that have not been accessed for more + * than the specified number of epochs. + * + * decrement: This field is only used when the decr_mode is + * H5C2_decr__threshold. + * + * The field is a double containing the multiplier used to derive the + * new cache size from the old if a cache size decrement is triggered. + * The decrement must be in the range 0.0 (in which case the cache will + * try to contract to its minimum size) to 1.0 (in which case the + * cache will never shrink). + * + * apply_max_decrement: Boolean flag used to determine whether decrements + * in cache size are to be limited by the max_decrement field. + * + * max_decrement: Maximum number of bytes by which the cache size can be + * decreased in a single re-size. Note that decrements may also be + * restricted by the min_size of the cache, and (in age out modes) by + * the empty_reserve field. + * + * epochs_before_eviction: Integer field used in H5C2_decr__age_out and + * H5C2_decr__age_out_with_threshold decrement modes. + * + * This field contains the number of epochs an entry must remain + * unaccessed before it is evicted in an attempt to reduce the + * cache size. If applicable, this field must lie in the range + * [1, H5C2__MAX_EPOCH_MARKERS]. + * + * apply_empty_reserve: Boolean field controlling whether the empty_reserve + * field is to be used in computing the new cache size when the + * decr_mode is H5C2_decr__age_out or H5C2_decr__age_out_with_threshold. + * + * empty_reserve: To avoid a constant racheting down of cache size by small + * amounts in the H5C2_decr__age_out and H5C2_decr__age_out_with_threshold + * modes, this field allows one to require that any cache size + * reductions leave the specified fraction of unused space in the cache. + * + * The value of this field must be in the range [0.0, 1.0]. I would + * expect typical values to be in the range of 0.01 to 0.1. + * + * + * Parallel Configuration Fields: + * + * In PHDF5, all operations that modify metadata must be executed collectively. + * We used to think that this was enough to ensure consistency across the + * metadata caches, but since we allow processes to read metadata individually, + * the order of dirty entries in the LRU list can vary across processes, + * which can result in inconsistencies between the caches. + * + * To prevent this, only the metadata cache on process 0 is allowed to write + * to file, and then only after synchronizing with the other caches. After + * it writes entries to file, it sends the base addresses of the now clean + * entries to the other caches, so they can mark these entries clean as well. + * + * The different caches know when to synchronize caches by counting the + * number of bytes of dirty metadata created by the collective operations + * modifying metadata. Whenever this count exceeds a user specified + * threshold (see below), process 0 flushes down to its minimum clean size, + * and then sends the list of newly cleaned entries to the other caches. + * + * dirty_bytes_threshold: Threshold of dirty byte creation used to + * synchronize updates between caches. (See above for outline and + * motivation.) + * + * This value MUST be consistant across all processes accessing the + * file. This field is ignored unless HDF5 has been compiled for + * parallel. + * + ****************************************************************************/ + +#define H5AC2__CURR_CACHE_CONFIG_VERSION 1 + +typedef struct H5AC2_cache_config_t +{ + /* general configuration fields: */ + int version; + + hbool_t rpt_fcn_enabled; + + hbool_t open_trace_file; + hbool_t close_trace_file; + char trace_file_name[H5AC2__MAX_TRACE_FILE_NAME_LEN+1]; + + hbool_t evictions_enabled; + + hbool_t set_initial_size; + size_t initial_size; + + double min_clean_fraction; + + size_t max_size; + size_t min_size; + + long int epoch_length; + + + /* size increase control fields: */ + enum H5C2_cache_incr_mode incr_mode; + + double lower_hr_threshold; + + double increment; + + hbool_t apply_max_increment; + size_t max_increment; + + + /* size decrease control fields: */ + enum H5C2_cache_decr_mode decr_mode; + + double upper_hr_threshold; + + double decrement; + + hbool_t apply_max_decrement; + size_t max_decrement; + + int epochs_before_eviction; + + hbool_t apply_empty_reserve; + double empty_reserve; + + + /* parallel configuration fields: */ + int dirty_bytes_threshold; + +} H5AC2_cache_config_t; + + +#ifdef __cplusplus +} +#endif +#endif diff --git a/src/H5C2.c b/src/H5C2.c new file mode 100644 index 0000000..373191f --- /dev/null +++ b/src/H5C2.c @@ -0,0 +1,11024 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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: H5C2.c + * June 1 2004 + * John Mainzer + * + * Purpose: Functions in this file implement a generic cache for + * things which exist on disk, and which may be + * unambiguously referenced by their disk addresses. + * + * The code in this module was initially written in + * support of a complete re-write of the metadata cache + * in H5AC.c However, other uses for the cache code + * suggested themselves, and thus this file was created + * in an attempt to support re-use. + * + * For a detailed overview of the cache, please see the + * header comment for H5C2_t in H5C2pkg.h. + * + * Modifications: + * + * QAK - 11/27/2004 + * Switched over to using skip list routines instead of TBBT + * routines. + * + * JRM - 12/15/04 + * Added code supporting manual and automatic cache resizing. + * See the header for H5C2_auto_size_ctl_t in H5C2private.h for + * an overview. + * + * Some elements of the automatic cache resize code depend on + * the LRU list. Thus if we ever choose to support a new + * replacement policy, we will either have to disable those + * elements of the auto resize code when running the new + * policy, or modify them to make use of similar information + * maintained by the new policy code. + * + *------------------------------------------------------------------------- + */ + +/************************************************************************** + * + * To Do: + * + * Code Changes: + * + * - Remove extra functionality in H5C2_flush_single_entry()? + * + * - Change protect/unprotect to lock/unlock. + * + * - Change the way the dirty flag is set. Probably pass it in + * as a parameter in unprotect & insert. + * + * - Size should also be passed in as a parameter in insert and + * unprotect -- or some other way should be found to advise the + * cache of changes in entry size. + * + * - Flush entries in increasing address order in + * H5C2_make_space_in_cache(). + * + * - Also in H5C2_make_space_in_cache(), use high and low water marks + * to reduce the number of I/O calls. + * + * - When flushing, attempt to combine contiguous entries to reduce + * I/O overhead. Can't do this just yet as some entries are not + * contiguous. Do this in parallel only or in serial as well? + * + * - Create MPI type for dirty objects when flushing in parallel. + * + * - Now that TBBT routines aren't used, fix nodes in memory to + * point directly to the skip list node from the LRU list, eliminating + * skip list lookups when evicting objects from the cache. + * + * Tests: + * + * - Trim execution time. (This is no longer a major issue with the + * shift from the TBBT to a hash table for indexing.) + * + * - Add random tests. + * + **************************************************************************/ + +#define H5C2_PACKAGE /*suppress error about including H5C2pkg */ +#define H5F_PACKAGE /*suppress error about including H5Fpkg */ + + +#include "H5private.h" /* Generic Functions */ +#include "H5C2pkg.h" /* Cache */ +#include "H5Dprivate.h" /* Dataset functions */ +#include "H5Eprivate.h" /* Error handling */ +#include "H5Fpkg.h" /* Files */ +#include "H5FDprivate.h" /* File drivers */ +#include "H5FLprivate.h" /* Free Lists */ +#include "H5Iprivate.h" /* IDs */ +#include "H5MMprivate.h" /* Memory management */ +#include "H5Pprivate.h" /* Property lists */ +#include "H5SLprivate.h" /* Skip lists */ + + +/**************************************************************************** + * + * We maintain doubly linked lists of instances of H5C2_cache_entry_t for a + * variety of reasons -- protected list, LRU list, and the clean and dirty + * LRU lists at present. The following macros support linking and unlinking + * of instances of H5C2_cache_entry_t by both their regular and auxilary next + * and previous pointers. + * + * The size and length fields are also maintained. + * + * Note that the relevant pair of prev and next pointers are presumed to be + * NULL on entry in the insertion macros. + * + * Finally, observe that the sanity checking macros evaluate to the empty + * string when H5C2_DO_SANITY_CHECKS is FALSE. They also contain calls + * to the HGOTO_ERROR macro, which may not be appropriate in all cases. + * If so, we will need versions of the insertion and deletion macros which + * do not reference the sanity checking macros. + * JRM - 5/5/04 + * + * Changes: + * + * - Removed the line: + * + * ( ( (Size) == (entry_ptr)->size ) && ( (len) != 1 ) ) || + * + * from the H5C2__DLL_PRE_REMOVE_SC macro. With the addition of the + * epoch markers used in the age out based cache size reduction algorithm, + * this invarient need not hold, as the epoch markers are of size 0. + * + * One could argue that I should have given the epoch markers a positive + * size, but this would break the index_size = LRU_list_size + pl_size + * + pel_size invarient. + * + * Alternatively, I could pass the current decr_mode in to the macro, + * and just skip the check whenever epoch markers may be in use. + * + * However, any size errors should be caught when the cache is flushed + * and destroyed. Until we are tracking such an error, this should be + * good enough. + * JRM - 12/9/04 + * + * + * - In the H5C2__DLL_PRE_INSERT_SC macro, replaced the lines: + * + * ( ( (len) == 1 ) && + * ( ( (head_ptr) != (tail_ptr) ) || ( (Size) <= 0 ) || + * ( (head_ptr) == NULL ) || ( (head_ptr)->size != (Size) ) + * ) + * ) || + * + * with: + * + * ( ( (len) == 1 ) && + * ( ( (head_ptr) != (tail_ptr) ) || + * ( (head_ptr) == NULL ) || ( (head_ptr)->size != (Size) ) + * ) + * ) || + * + * Epoch markers have size 0, so we can now have a non-empty list with + * zero size. Hence the "( (Size) <= 0 )" clause cause false failures + * in the sanity check. Since "Size" is typically a size_t, it can't + * take on negative values, and thus the revised clause "( (Size) < 0 )" + * caused compiler warnings. + * JRM - 12/22/04 + * + * - In the H5C2__DLL_SC macro, replaced the lines: + * + * ( ( (len) == 1 ) && + * ( ( (head_ptr) != (tail_ptr) ) || ( (cache_ptr)->size <= 0 ) || + * ( (head_ptr) == NULL ) || ( (head_ptr)->size != (Size) ) + * ) + * ) || + * + * with + * + * ( ( (len) == 1 ) && + * ( ( (head_ptr) != (tail_ptr) ) || + * ( (head_ptr) == NULL ) || ( (head_ptr)->size != (Size) ) + * ) + * ) || + * + * Epoch markers have size 0, so we can now have a non-empty list with + * zero size. Hence the "( (Size) <= 0 )" clause cause false failures + * in the sanity check. Since "Size" is typically a size_t, it can't + * take on negative values, and thus the revised clause "( (Size) < 0 )" + * caused compiler warnings. + * JRM - 1/10/05 + * + * - Added the H5C2__DLL_UPDATE_FOR_SIZE_CHANGE macro and the associated + * sanity checking macros. These macro are used to update the size of + * a DLL when one of its entries changes size. + * + * JRM - 9/8/05 + * + ****************************************************************************/ + +#if H5C2_DO_SANITY_CHECKS + +#define H5C2__DLL_PRE_REMOVE_SC(entry_ptr, head_ptr, tail_ptr, len, Size, fv) \ +if ( ( (head_ptr) == NULL ) || \ + ( (tail_ptr) == NULL ) || \ + ( (entry_ptr) == NULL ) || \ + ( (len) <= 0 ) || \ + ( (Size) < (entry_ptr)->size ) || \ + ( ( (entry_ptr)->prev == NULL ) && ( (head_ptr) != (entry_ptr) ) ) || \ + ( ( (entry_ptr)->next == NULL ) && ( (tail_ptr) != (entry_ptr) ) ) || \ + ( ( (len) == 1 ) && \ + ( ! ( ( (head_ptr) == (entry_ptr) ) && \ + ( (tail_ptr) == (entry_ptr) ) && \ + ( (entry_ptr)->next == NULL ) && \ + ( (entry_ptr)->prev == NULL ) && \ + ( (Size) == (entry_ptr)->size ) \ + ) \ + ) \ + ) \ + ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "DLL pre remove SC failed") \ +} + +#define H5C2__DLL_SC(head_ptr, tail_ptr, len, Size, fv) \ +if ( ( ( ( (head_ptr) == NULL ) || ( (tail_ptr) == NULL ) ) && \ + ( (head_ptr) != (tail_ptr) ) \ + ) || \ + ( (len) < 0 ) || \ + ( (Size) < 0 ) || \ + ( ( (len) == 1 ) && \ + ( ( (head_ptr) != (tail_ptr) ) || \ + ( (head_ptr) == NULL ) || ( (head_ptr)->size != (Size) ) \ + ) \ + ) || \ + ( ( (len) >= 1 ) && \ + ( ( (head_ptr) == NULL ) || ( (head_ptr)->prev != NULL ) || \ + ( (tail_ptr) == NULL ) || ( (tail_ptr)->next != NULL ) \ + ) \ + ) \ + ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "DLL sanity check failed") \ +} + +#define H5C2__DLL_PRE_INSERT_SC(entry_ptr, head_ptr, tail_ptr, len, Size, fv) \ +if ( ( (entry_ptr) == NULL ) || \ + ( (entry_ptr)->next != NULL ) || \ + ( (entry_ptr)->prev != NULL ) || \ + ( ( ( (head_ptr) == NULL ) || ( (tail_ptr) == NULL ) ) && \ + ( (head_ptr) != (tail_ptr) ) \ + ) || \ + ( (len) < 0 ) || \ + ( ( (len) == 1 ) && \ + ( ( (head_ptr) != (tail_ptr) ) || \ + ( (head_ptr) == NULL ) || ( (head_ptr)->size != (Size) ) \ + ) \ + ) || \ + ( ( (len) >= 1 ) && \ + ( ( (head_ptr) == NULL ) || ( (head_ptr)->prev != NULL ) || \ + ( (tail_ptr) == NULL ) || ( (tail_ptr)->next != NULL ) \ + ) \ + ) \ + ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "DLL pre insert SC failed") \ +} + +#define H5C2__DLL_PRE_SIZE_UPDATE_SC(dll_len, dll_size, old_size, new_size) \ +if ( ( (dll_len) <= 0 ) || \ + ( (dll_size) <= 0 ) || \ + ( (old_size) <= 0 ) || \ + ( (old_size) > (dll_size) ) || \ + ( (new_size) <= 0 ) || \ + ( ( (dll_len) == 1 ) && ( (old_size) != (dll_size) ) ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "DLL pre size update SC failed") \ +} + +#define H5C2__DLL_POST_SIZE_UPDATE_SC(dll_len, dll_size, old_size, new_size) \ +if ( ( (new_size) > (dll_size) ) || \ + ( ( (dll_len) == 1 ) && ( (new_size) != (dll_size) ) ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "DLL post size update SC failed") \ +} + +#else /* H5C2_DO_SANITY_CHECKS */ + +#define H5C2__DLL_PRE_REMOVE_SC(entry_ptr, head_ptr, tail_ptr, len, Size, fv) +#define H5C2__DLL_SC(head_ptr, tail_ptr, len, Size, fv) +#define H5C2__DLL_PRE_INSERT_SC(entry_ptr, head_ptr, tail_ptr, len, Size, fv) +#define H5C2__DLL_PRE_SIZE_UPDATE_SC(dll_len, dll_size, old_size, new_size) +#define H5C2__DLL_POST_SIZE_UPDATE_SC(dll_len, dll_size, old_size, new_size) + +#endif /* H5C2_DO_SANITY_CHECKS */ + + +#define H5C2__DLL_APPEND(entry_ptr, head_ptr, tail_ptr, len, Size, fail_val) \ + H5C2__DLL_PRE_INSERT_SC(entry_ptr, head_ptr, tail_ptr, len, Size, \ + fail_val) \ + if ( (head_ptr) == NULL ) \ + { \ + (head_ptr) = (entry_ptr); \ + (tail_ptr) = (entry_ptr); \ + } \ + else \ + { \ + (tail_ptr)->next = (entry_ptr); \ + (entry_ptr)->prev = (tail_ptr); \ + (tail_ptr) = (entry_ptr); \ + } \ + (len)++; \ + (Size) += (entry_ptr)->size; + +#define H5C2__DLL_PREPEND(entry_ptr, head_ptr, tail_ptr, len, Size, fail_val) \ + H5C2__DLL_PRE_INSERT_SC(entry_ptr, head_ptr, tail_ptr, len, Size, \ + fail_val) \ + if ( (head_ptr) == NULL ) \ + { \ + (head_ptr) = (entry_ptr); \ + (tail_ptr) = (entry_ptr); \ + } \ + else \ + { \ + (head_ptr)->prev = (entry_ptr); \ + (entry_ptr)->next = (head_ptr); \ + (head_ptr) = (entry_ptr); \ + } \ + (len)++; \ + (Size) += entry_ptr->size; + +#define H5C2__DLL_REMOVE(entry_ptr, head_ptr, tail_ptr, len, Size, fail_val) \ + H5C2__DLL_PRE_REMOVE_SC(entry_ptr, head_ptr, tail_ptr, len, Size, \ + fail_val) \ + { \ + if ( (head_ptr) == (entry_ptr) ) \ + { \ + (head_ptr) = (entry_ptr)->next; \ + if ( (head_ptr) != NULL ) \ + { \ + (head_ptr)->prev = NULL; \ + } \ + } \ + else \ + { \ + (entry_ptr)->prev->next = (entry_ptr)->next; \ + } \ + if ( (tail_ptr) == (entry_ptr) ) \ + { \ + (tail_ptr) = (entry_ptr)->prev; \ + if ( (tail_ptr) != NULL ) \ + { \ + (tail_ptr)->next = NULL; \ + } \ + } \ + else \ + { \ + (entry_ptr)->next->prev = (entry_ptr)->prev; \ + } \ + entry_ptr->next = NULL; \ + entry_ptr->prev = NULL; \ + (len)--; \ + (Size) -= entry_ptr->size; \ + } + +#define H5C2__DLL_UPDATE_FOR_SIZE_CHANGE(dll_len, dll_size, old_size, new_size) \ + H5C2__DLL_PRE_SIZE_UPDATE_SC(dll_len, dll_size, old_size, new_size) \ + (dll_size) -= (old_size); \ + (dll_size) += (new_size); \ + H5C2__DLL_POST_SIZE_UPDATE_SC(dll_len, dll_size, old_size, new_size) + +#if H5C2_DO_SANITY_CHECKS + +#define H5C2__AUX_DLL_PRE_REMOVE_SC(entry_ptr, hd_ptr, tail_ptr, len, Size, fv) \ +if ( ( (hd_ptr) == NULL ) || \ + ( (tail_ptr) == NULL ) || \ + ( (entry_ptr) == NULL ) || \ + ( (len) <= 0 ) || \ + ( (Size) < (entry_ptr)->size ) || \ + ( ( (Size) == (entry_ptr)->size ) && ( ! ( (len) == 1 ) ) ) || \ + ( ( (entry_ptr)->aux_prev == NULL ) && ( (hd_ptr) != (entry_ptr) ) ) || \ + ( ( (entry_ptr)->aux_next == NULL ) && ( (tail_ptr) != (entry_ptr) ) ) || \ + ( ( (len) == 1 ) && \ + ( ! ( ( (hd_ptr) == (entry_ptr) ) && ( (tail_ptr) == (entry_ptr) ) && \ + ( (entry_ptr)->aux_next == NULL ) && \ + ( (entry_ptr)->aux_prev == NULL ) && \ + ( (Size) == (entry_ptr)->size ) \ + ) \ + ) \ + ) \ + ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "aux DLL pre remove SC failed") \ +} + +#define H5C2__AUX_DLL_SC(head_ptr, tail_ptr, len, Size, fv) \ +if ( ( ( ( (head_ptr) == NULL ) || ( (tail_ptr) == NULL ) ) && \ + ( (head_ptr) != (tail_ptr) ) \ + ) || \ + ( (len) < 0 ) || \ + ( (Size) < 0 ) || \ + ( ( (len) == 1 ) && \ + ( ( (head_ptr) != (tail_ptr) ) || ( (Size) <= 0 ) || \ + ( (head_ptr) == NULL ) || ( (head_ptr)->size != (Size) ) \ + ) \ + ) || \ + ( ( (len) >= 1 ) && \ + ( ( (head_ptr) == NULL ) || ( (head_ptr)->aux_prev != NULL ) || \ + ( (tail_ptr) == NULL ) || ( (tail_ptr)->aux_next != NULL ) \ + ) \ + ) \ + ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "AUX DLL sanity check failed") \ +} + +#define H5C2__AUX_DLL_PRE_INSERT_SC(entry_ptr, hd_ptr, tail_ptr, len, Size, fv) \ +if ( ( (entry_ptr) == NULL ) || \ + ( (entry_ptr)->aux_next != NULL ) || \ + ( (entry_ptr)->aux_prev != NULL ) || \ + ( ( ( (hd_ptr) == NULL ) || ( (tail_ptr) == NULL ) ) && \ + ( (hd_ptr) != (tail_ptr) ) \ + ) || \ + ( (len) < 0 ) || \ + ( ( (len) == 1 ) && \ + ( ( (hd_ptr) != (tail_ptr) ) || ( (Size) <= 0 ) || \ + ( (hd_ptr) == NULL ) || ( (hd_ptr)->size != (Size) ) \ + ) \ + ) || \ + ( ( (len) >= 1 ) && \ + ( ( (hd_ptr) == NULL ) || ( (hd_ptr)->aux_prev != NULL ) || \ + ( (tail_ptr) == NULL ) || ( (tail_ptr)->aux_next != NULL ) \ + ) \ + ) \ + ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, (fv), "AUX DLL pre insert SC failed") \ +} + +#else /* H5C2_DO_SANITY_CHECKS */ + +#define H5C2__AUX_DLL_PRE_REMOVE_SC(entry_ptr, hd_ptr, tail_ptr, len, Size, fv) +#define H5C2__AUX_DLL_SC(head_ptr, tail_ptr, len, Size, fv) +#define H5C2__AUX_DLL_PRE_INSERT_SC(entry_ptr, hd_ptr, tail_ptr, len, Size, fv) + +#endif /* H5C2_DO_SANITY_CHECKS */ + + +#define H5C2__AUX_DLL_APPEND(entry_ptr, head_ptr, tail_ptr, len, Size, fail_val)\ + H5C2__AUX_DLL_PRE_INSERT_SC(entry_ptr, head_ptr, tail_ptr, len, Size, \ + fail_val) \ + if ( (head_ptr) == NULL ) \ + { \ + (head_ptr) = (entry_ptr); \ + (tail_ptr) = (entry_ptr); \ + } \ + else \ + { \ + (tail_ptr)->aux_next = (entry_ptr); \ + (entry_ptr)->aux_prev = (tail_ptr); \ + (tail_ptr) = (entry_ptr); \ + } \ + (len)++; \ + (Size) += entry_ptr->size; + +#define H5C2__AUX_DLL_PREPEND(entry_ptr, head_ptr, tail_ptr, len, Size, fv) \ + H5C2__AUX_DLL_PRE_INSERT_SC(entry_ptr, head_ptr, tail_ptr, len, Size, \ + fv) \ + if ( (head_ptr) == NULL ) \ + { \ + (head_ptr) = (entry_ptr); \ + (tail_ptr) = (entry_ptr); \ + } \ + else \ + { \ + (head_ptr)->aux_prev = (entry_ptr); \ + (entry_ptr)->aux_next = (head_ptr); \ + (head_ptr) = (entry_ptr); \ + } \ + (len)++; \ + (Size) += entry_ptr->size; + +#define H5C2__AUX_DLL_REMOVE(entry_ptr, head_ptr, tail_ptr, len, Size, fv) \ + H5C2__AUX_DLL_PRE_REMOVE_SC(entry_ptr, head_ptr, tail_ptr, len, Size, \ + fv) \ + { \ + if ( (head_ptr) == (entry_ptr) ) \ + { \ + (head_ptr) = (entry_ptr)->aux_next; \ + if ( (head_ptr) != NULL ) \ + { \ + (head_ptr)->aux_prev = NULL; \ + } \ + } \ + else \ + { \ + (entry_ptr)->aux_prev->aux_next = (entry_ptr)->aux_next; \ + } \ + if ( (tail_ptr) == (entry_ptr) ) \ + { \ + (tail_ptr) = (entry_ptr)->aux_prev; \ + if ( (tail_ptr) != NULL ) \ + { \ + (tail_ptr)->aux_next = NULL; \ + } \ + } \ + else \ + { \ + (entry_ptr)->aux_next->aux_prev = (entry_ptr)->aux_prev; \ + } \ + entry_ptr->aux_next = NULL; \ + entry_ptr->aux_prev = NULL; \ + (len)--; \ + (Size) -= entry_ptr->size; \ + } + + +/*********************************************************************** + * + * Stats collection macros + * + * The following macros must handle stats collection when this collection + * is enabled, and evaluate to the empty string when it is not. + * + * The sole exception to this rule is + * H5C2__UPDATE_CACHE_HIT_RATE_STATS(), which is always active as + * the cache hit rate stats are always collected and available. + * + * Changes: + * + * JRM -- 3/21/06 + * Added / updated macros for pinned entry related stats. + * + * JRM -- 8/9/06 + * More pinned entry stats related updates. + * + * JRM -- 3/31/07 + * Updated H5C2__UPDATE_STATS_FOR_PROTECT() to keep stats on + * read and write protects. + * + ***********************************************************************/ + +#define H5C2__UPDATE_CACHE_HIT_RATE_STATS(cache_ptr, hit) \ + (cache_ptr->cache_accesses)++; \ + if ( hit ) { \ + (cache_ptr->cache_hits)++; \ + } \ + +#if H5C2_COLLECT_CACHE_STATS + +#define H5C2__UPDATE_STATS_FOR_DIRTY_PIN(cache_ptr, entry_ptr) \ + (((cache_ptr)->dirty_pins)[(entry_ptr)->type->id])++; + +#define H5C2__UPDATE_STATS_FOR_UNPROTECT(cache_ptr) \ + if ( (cache_ptr)->slist_len > (cache_ptr)->max_slist_len ) \ + (cache_ptr)->max_slist_len = (cache_ptr)->slist_len; \ + if ( (cache_ptr)->slist_size > (cache_ptr)->max_slist_size ) \ + (cache_ptr)->max_slist_size = (cache_ptr)->slist_size; \ + if ( (cache_ptr)->pel_len > (cache_ptr)->max_pel_len ) \ + (cache_ptr)->max_pel_len = (cache_ptr)->pel_len; \ + if ( (cache_ptr)->pel_size > (cache_ptr)->max_pel_size ) \ + (cache_ptr)->max_pel_size = (cache_ptr)->pel_size; + +#define H5C2__UPDATE_STATS_FOR_RENAME(cache_ptr, entry_ptr) \ + if ( cache_ptr->flush_in_progress ) { \ + ((cache_ptr)->cache_flush_renames[(entry_ptr)->type->id])++; \ + } \ + if ( entry_ptr->flush_in_progress ) { \ + ((cache_ptr)->entry_flush_renames[(entry_ptr)->type->id])++; \ + } \ + (((cache_ptr)->renames)[(entry_ptr)->type->id])++; + +#define H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE(cache_ptr, entry_ptr, new_size)\ + if ( cache_ptr->flush_in_progress ) { \ + ((cache_ptr)->cache_flush_size_changes[(entry_ptr)->type->id])++; \ + } \ + if ( entry_ptr->flush_in_progress ) { \ + ((cache_ptr)->entry_flush_size_changes[(entry_ptr)->type->id])++; \ + } \ + if ( (entry_ptr)->size < (new_size) ) { \ + ((cache_ptr)->size_increases[(entry_ptr)->type->id])++; \ + if ( (cache_ptr)->index_size > (cache_ptr)->max_index_size ) \ + (cache_ptr)->max_index_size = (cache_ptr)->index_size; \ + if ( (cache_ptr)->slist_size > (cache_ptr)->max_slist_size ) \ + (cache_ptr)->max_slist_size = (cache_ptr)->slist_size; \ + if ( (cache_ptr)->pl_size > (cache_ptr)->max_pl_size ) \ + (cache_ptr)->max_pl_size = (cache_ptr)->pl_size; \ + } else if ( (entry_ptr)->size > (new_size) ) { \ + ((cache_ptr)->size_decreases[(entry_ptr)->type->id])++; \ + } + +#define H5C2__UPDATE_STATS_FOR_HT_INSERTION(cache_ptr) \ + (cache_ptr)->total_ht_insertions++; + +#define H5C2__UPDATE_STATS_FOR_HT_DELETION(cache_ptr) \ + (cache_ptr)->total_ht_deletions++; + +#define H5C2__UPDATE_STATS_FOR_HT_SEARCH(cache_ptr, success, depth) \ + if ( success ) { \ + (cache_ptr)->successful_ht_searches++; \ + (cache_ptr)->total_successful_ht_search_depth += depth; \ + } else { \ + (cache_ptr)->failed_ht_searches++; \ + (cache_ptr)->total_failed_ht_search_depth += depth; \ + } + +#define H5C2__UPDATE_STATS_FOR_UNPIN(cache_ptr, entry_ptr) \ + ((cache_ptr)->unpins)[(entry_ptr)->type->id]++; + +#if H5C2_COLLECT_CACHE_ENTRY_STATS + +#define H5C2__RESET_CACHE_ENTRY_STATS(entry_ptr) \ + (entry_ptr)->accesses = 0; \ + (entry_ptr)->clears = 0; \ + (entry_ptr)->flushes = 0; \ + (entry_ptr)->pins = 0; + +#define H5C2__UPDATE_STATS_FOR_CLEAR(cache_ptr, entry_ptr) \ + (((cache_ptr)->clears)[(entry_ptr)->type->id])++; \ + if ( (entry_ptr)->is_pinned ) { \ + (((cache_ptr)->pinned_clears)[(entry_ptr)->type->id])++; \ + } \ + ((entry_ptr)->clears)++; + +#define H5C2__UPDATE_STATS_FOR_FLUSH(cache_ptr, entry_ptr) \ + (((cache_ptr)->flushes)[(entry_ptr)->type->id])++; \ + if ( (entry_ptr)->is_pinned ) { \ + (((cache_ptr)->pinned_flushes)[(entry_ptr)->type->id])++; \ + } \ + ((entry_ptr)->flushes)++; + +#define H5C2__UPDATE_STATS_FOR_EVICTION(cache_ptr, entry_ptr) \ + (((cache_ptr)->evictions)[(entry_ptr)->type->id])++; \ + if ( (entry_ptr)->accesses > \ + ((cache_ptr)->max_accesses)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_accesses)[(entry_ptr)->type->id] \ + = (entry_ptr)->accesses; \ + } \ + if ( (entry_ptr)->accesses < \ + ((cache_ptr)->min_accesses)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->min_accesses)[(entry_ptr)->type->id] \ + = (entry_ptr)->accesses; \ + } \ + if ( (entry_ptr)->clears > \ + ((cache_ptr)->max_clears)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_clears)[(entry_ptr)->type->id] \ + = (entry_ptr)->clears; \ + } \ + if ( (entry_ptr)->flushes > \ + ((cache_ptr)->max_flushes)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_flushes)[(entry_ptr)->type->id] \ + = (entry_ptr)->flushes; \ + } \ + if ( (entry_ptr)->size > \ + ((cache_ptr)->max_size)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_size)[(entry_ptr)->type->id] \ + = (entry_ptr)->size; \ + } \ + if ( (entry_ptr)->pins > \ + ((cache_ptr)->max_pins)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_pins)[(entry_ptr)->type->id] \ + = (entry_ptr)->pins; \ + } + +#define H5C2__UPDATE_STATS_FOR_INSERTION(cache_ptr, entry_ptr) \ + (((cache_ptr)->insertions)[(entry_ptr)->type->id])++; \ + if ( (entry_ptr)->is_pinned ) { \ + (((cache_ptr)->pinned_insertions)[(entry_ptr)->type->id])++; \ + ((cache_ptr)->pins)[(entry_ptr)->type->id]++; \ + (entry_ptr)->pins++; \ + if ( (cache_ptr)->pel_len > (cache_ptr)->max_pel_len ) \ + (cache_ptr)->max_pel_len = (cache_ptr)->pel_len; \ + if ( (cache_ptr)->pel_size > (cache_ptr)->max_pel_size ) \ + (cache_ptr)->max_pel_size = (cache_ptr)->pel_size; \ + } \ + if ( (cache_ptr)->index_len > (cache_ptr)->max_index_len ) \ + (cache_ptr)->max_index_len = (cache_ptr)->index_len; \ + if ( (cache_ptr)->index_size > (cache_ptr)->max_index_size ) \ + (cache_ptr)->max_index_size = (cache_ptr)->index_size; \ + if ( (cache_ptr)->slist_len > (cache_ptr)->max_slist_len ) \ + (cache_ptr)->max_slist_len = (cache_ptr)->slist_len; \ + if ( (cache_ptr)->slist_size > (cache_ptr)->max_slist_size ) \ + (cache_ptr)->max_slist_size = (cache_ptr)->slist_size; \ + if ( (entry_ptr)->size > \ + ((cache_ptr)->max_size)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_size)[(entry_ptr)->type->id] \ + = (entry_ptr)->size; \ + } + +#define H5C2__UPDATE_STATS_FOR_PROTECT(cache_ptr, entry_ptr, hit) \ + if ( hit ) \ + ((cache_ptr)->hits)[(entry_ptr)->type->id]++; \ + else \ + ((cache_ptr)->misses)[(entry_ptr)->type->id]++; \ + if ( ! ((entry_ptr)->is_read_only) ) { \ + ((cache_ptr)->write_protects)[(entry_ptr)->type->id]++; \ + } else { \ + ((cache_ptr)->read_protects)[(entry_ptr)->type->id]++; \ + if ( ((entry_ptr)->ro_ref_count) > \ + ((cache_ptr)->max_read_protects)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_read_protects)[(entry_ptr)->type->id] = \ + ((entry_ptr)->ro_ref_count); \ + } \ + } \ + if ( (cache_ptr)->index_len > (cache_ptr)->max_index_len ) \ + (cache_ptr)->max_index_len = (cache_ptr)->index_len; \ + if ( (cache_ptr)->index_size > (cache_ptr)->max_index_size ) \ + (cache_ptr)->max_index_size = (cache_ptr)->index_size; \ + if ( (cache_ptr)->pl_len > (cache_ptr)->max_pl_len ) \ + (cache_ptr)->max_pl_len = (cache_ptr)->pl_len; \ + if ( (cache_ptr)->pl_size > (cache_ptr)->max_pl_size ) \ + (cache_ptr)->max_pl_size = (cache_ptr)->pl_size; \ + if ( (entry_ptr)->size > \ + ((cache_ptr)->max_size)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_size)[(entry_ptr)->type->id] \ + = (entry_ptr)->size; \ + } \ + ((entry_ptr)->accesses)++; + +#define H5C2__UPDATE_STATS_FOR_PIN(cache_ptr, entry_ptr) \ + ((cache_ptr)->pins)[(entry_ptr)->type->id]++; \ + (entry_ptr)->pins++; \ + if ( (cache_ptr)->pel_len > (cache_ptr)->max_pel_len ) \ + (cache_ptr)->max_pel_len = (cache_ptr)->pel_len; \ + if ( (cache_ptr)->pel_size > (cache_ptr)->max_pel_size ) \ + (cache_ptr)->max_pel_size = (cache_ptr)->pel_size; + +#else /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + +#define H5C2__RESET_CACHE_ENTRY_STATS(entry_ptr) + +#define H5C2__UPDATE_STATS_FOR_CLEAR(cache_ptr, entry_ptr) \ + if ( (entry_ptr)->is_pinned ) { \ + (((cache_ptr)->pinned_clears)[(entry_ptr)->type->id])++; \ + } \ + (((cache_ptr)->clears)[(entry_ptr)->type->id])++; + +#define H5C2__UPDATE_STATS_FOR_FLUSH(cache_ptr, entry_ptr) \ + (((cache_ptr)->flushes)[(entry_ptr)->type->id])++; \ + if ( (entry_ptr)->is_pinned ) { \ + (((cache_ptr)->pinned_flushes)[(entry_ptr)->type->id])++; \ + } + +#define H5C2__UPDATE_STATS_FOR_EVICTION(cache_ptr, entry_ptr) \ + (((cache_ptr)->evictions)[(entry_ptr)->type->id])++; + +#define H5C2__UPDATE_STATS_FOR_INSERTION(cache_ptr, entry_ptr) \ + (((cache_ptr)->insertions)[(entry_ptr)->type->id])++; \ + if ( (entry_ptr)->is_pinned ) { \ + (((cache_ptr)->pinned_insertions)[(entry_ptr)->type->id])++; \ + ((cache_ptr)->pins)[(entry_ptr)->type->id]++; \ + if ( (cache_ptr)->pel_len > (cache_ptr)->max_pel_len ) \ + (cache_ptr)->max_pel_len = (cache_ptr)->pel_len; \ + if ( (cache_ptr)->pel_size > (cache_ptr)->max_pel_size ) \ + (cache_ptr)->max_pel_size = (cache_ptr)->pel_size; \ + } \ + if ( (cache_ptr)->index_len > (cache_ptr)->max_index_len ) \ + (cache_ptr)->max_index_len = (cache_ptr)->index_len; \ + if ( (cache_ptr)->index_size > (cache_ptr)->max_index_size ) \ + (cache_ptr)->max_index_size = (cache_ptr)->index_size; \ + if ( (cache_ptr)->slist_len > (cache_ptr)->max_slist_len ) \ + (cache_ptr)->max_slist_len = (cache_ptr)->slist_len; \ + if ( (cache_ptr)->slist_size > (cache_ptr)->max_slist_size ) \ + (cache_ptr)->max_slist_size = (cache_ptr)->slist_size; + +#define H5C2__UPDATE_STATS_FOR_PROTECT(cache_ptr, entry_ptr, hit) \ + if ( hit ) \ + ((cache_ptr)->hits)[(entry_ptr)->type->id]++; \ + else \ + ((cache_ptr)->misses)[(entry_ptr)->type->id]++; \ + if ( ! ((entry_ptr)->is_read_only) ) { \ + ((cache_ptr)->write_protects)[(entry_ptr)->type->id]++; \ + } else { \ + ((cache_ptr)->read_protects)[(entry_ptr)->type->id]++; \ + if ( ((entry_ptr)->ro_ref_count) > \ + ((cache_ptr)->max_read_protects)[(entry_ptr)->type->id] ) { \ + ((cache_ptr)->max_read_protects)[(entry_ptr)->type->id] = \ + ((entry_ptr)->ro_ref_count); \ + } \ + } \ + if ( (cache_ptr)->index_len > (cache_ptr)->max_index_len ) \ + (cache_ptr)->max_index_len = (cache_ptr)->index_len; \ + if ( (cache_ptr)->index_size > (cache_ptr)->max_index_size ) \ + (cache_ptr)->max_index_size = (cache_ptr)->index_size; \ + if ( (cache_ptr)->pl_len > (cache_ptr)->max_pl_len ) \ + (cache_ptr)->max_pl_len = (cache_ptr)->pl_len; \ + if ( (cache_ptr)->pl_size > (cache_ptr)->max_pl_size ) \ + (cache_ptr)->max_pl_size = (cache_ptr)->pl_size; + +#define H5C2__UPDATE_STATS_FOR_PIN(cache_ptr, entry_ptr) \ + ((cache_ptr)->pins)[(entry_ptr)->type->id]++; \ + if ( (cache_ptr)->pel_len > (cache_ptr)->max_pel_len ) \ + (cache_ptr)->max_pel_len = (cache_ptr)->pel_len; \ + if ( (cache_ptr)->pel_size > (cache_ptr)->max_pel_size ) \ + (cache_ptr)->max_pel_size = (cache_ptr)->pel_size; + +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + +#else /* H5C2_COLLECT_CACHE_STATS */ + +#define H5C2__RESET_CACHE_ENTRY_STATS(entry_ptr) +#define H5C2__UPDATE_STATS_FOR_DIRTY_PIN(cache_ptr, entry_ptr) +#define H5C2__UPDATE_STATS_FOR_UNPROTECT(cache_ptr) +#define H5C2__UPDATE_STATS_FOR_RENAME(cache_ptr, entry_ptr) +#define H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE(cache_ptr, entry_ptr, new_size) +#define H5C2__UPDATE_STATS_FOR_HT_INSERTION(cache_ptr) +#define H5C2__UPDATE_STATS_FOR_HT_DELETION(cache_ptr) +#define H5C2__UPDATE_STATS_FOR_HT_SEARCH(cache_ptr, success, depth) +#define H5C2__UPDATE_STATS_FOR_INSERTION(cache_ptr, entry_ptr) +#define H5C2__UPDATE_STATS_FOR_CLEAR(cache_ptr, entry_ptr) +#define H5C2__UPDATE_STATS_FOR_FLUSH(cache_ptr, entry_ptr) +#define H5C2__UPDATE_STATS_FOR_EVICTION(cache_ptr, entry_ptr) +#define H5C2__UPDATE_STATS_FOR_PROTECT(cache_ptr, entry_ptr, hit) +#define H5C2__UPDATE_STATS_FOR_PIN(cache_ptr, entry_ptr) +#define H5C2__UPDATE_STATS_FOR_UNPIN(cache_ptr, entry_ptr) + +#endif /* H5C2_COLLECT_CACHE_STATS */ + + +/*********************************************************************** + * + * Hash table access and manipulation macros: + * + * The following macros handle searches, insertions, and deletion in + * the hash table. + * + * When modifying these macros, remember to modify the similar macros + * in tst/cache.c + * + ***********************************************************************/ + +/* H5C2__HASH_TABLE_LEN is defined in H5C2pkg.h. It mut be a power of two. */ + +#define H5C2__HASH_MASK ((size_t)(H5C2__HASH_TABLE_LEN - 1) << 3) + +#define H5C2__HASH_FCN(x) (int)(((x) & H5C2__HASH_MASK) >> 3) + +#if H5C2_DO_SANITY_CHECKS + +#define H5C2__PRE_HT_INSERT_SC(cache_ptr, entry_ptr, fail_val) \ +if ( ( (cache_ptr) == NULL ) || \ + ( (cache_ptr)->magic != H5C2__H5C2_T_MAGIC ) || \ + ( (entry_ptr) == NULL ) || \ + ( ! H5F_addr_defined((entry_ptr)->addr) ) || \ + ( (entry_ptr)->ht_next != NULL ) || \ + ( (entry_ptr)->ht_prev != NULL ) || \ + ( (entry_ptr)->size <= 0 ) || \ + ( (k = H5C2__HASH_FCN((entry_ptr)->addr)) < 0 ) || \ + ( k >= H5C2__HASH_TABLE_LEN ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, fail_val, \ + "Pre HT insert SC failed") \ +} + +#define H5C2__PRE_HT_REMOVE_SC(cache_ptr, entry_ptr) \ +if ( ( (cache_ptr) == NULL ) || \ + ( (cache_ptr)->magic != H5C2__H5C2_T_MAGIC ) || \ + ( (cache_ptr)->index_len < 1 ) || \ + ( (entry_ptr) == NULL ) || \ + ( (cache_ptr)->index_size < (entry_ptr)->size ) || \ + ( ! H5F_addr_defined((entry_ptr)->addr) ) || \ + ( (entry_ptr)->size <= 0 ) || \ + ( H5C2__HASH_FCN((entry_ptr)->addr) < 0 ) || \ + ( H5C2__HASH_FCN((entry_ptr)->addr) >= H5C2__HASH_TABLE_LEN ) || \ + ( ((cache_ptr)->index)[(H5C2__HASH_FCN((entry_ptr)->addr))] \ + == NULL ) || \ + ( ( ((cache_ptr)->index)[(H5C2__HASH_FCN((entry_ptr)->addr))] \ + != (entry_ptr) ) && \ + ( (entry_ptr)->ht_prev == NULL ) ) || \ + ( ( ((cache_ptr)->index)[(H5C2__HASH_FCN((entry_ptr)->addr))] == \ + (entry_ptr) ) && \ + ( (entry_ptr)->ht_prev != NULL ) ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Pre HT remove SC failed") \ +} + +#define H5C2__PRE_HT_SEARCH_SC(cache_ptr, Addr, fail_val) \ +if ( ( (cache_ptr) == NULL ) || \ + ( (cache_ptr)->magic != H5C2__H5C2_T_MAGIC ) || \ + ( ! H5F_addr_defined(Addr) ) || \ + ( H5C2__HASH_FCN(Addr) < 0 ) || \ + ( H5C2__HASH_FCN(Addr) >= H5C2__HASH_TABLE_LEN ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, fail_val, "Pre HT search SC failed") \ +} + +#define H5C2__POST_SUC_HT_SEARCH_SC(cache_ptr, entry_ptr, Addr, k, fail_val) \ +if ( ( (cache_ptr) == NULL ) || \ + ( (cache_ptr)->magic != H5C2__H5C2_T_MAGIC ) || \ + ( (cache_ptr)->index_len < 1 ) || \ + ( (entry_ptr) == NULL ) || \ + ( (cache_ptr)->index_size < (entry_ptr)->size ) || \ + ( H5F_addr_ne((entry_ptr)->addr, (Addr)) ) || \ + ( (entry_ptr)->size <= 0 ) || \ + ( ((cache_ptr)->index)[k] == NULL ) || \ + ( ( ((cache_ptr)->index)[k] != (entry_ptr) ) && \ + ( (entry_ptr)->ht_prev == NULL ) ) || \ + ( ( ((cache_ptr)->index)[k] == (entry_ptr) ) && \ + ( (entry_ptr)->ht_prev != NULL ) ) || \ + ( ( (entry_ptr)->ht_prev != NULL ) && \ + ( (entry_ptr)->ht_prev->ht_next != (entry_ptr) ) ) || \ + ( ( (entry_ptr)->ht_next != NULL ) && \ + ( (entry_ptr)->ht_next->ht_prev != (entry_ptr) ) ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, fail_val, \ + "Post successful HT search SC failed") \ +} + +#define H5C2__POST_HT_SHIFT_TO_FRONT(cache_ptr, entry_ptr, k, fail_val) \ +if ( ( (cache_ptr) == NULL ) || \ + ( ((cache_ptr)->index)[k] != (entry_ptr) ) || \ + ( (entry_ptr)->ht_prev != NULL ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, fail_val, \ + "Post HT shift to front SC failed") \ +} + +#define H5C2__PRE_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size) \ +if ( ( (cache_ptr) == NULL ) || \ + ( (cache_ptr)->index_len <= 0 ) || \ + ( (cache_ptr)->index_size <= 0 ) || \ + ( (new_size) <= 0 ) || \ + ( (old_size) > (cache_ptr)->index_size ) || \ + ( (new_size) <= 0 ) || \ + ( ( (cache_ptr)->index_len == 1 ) && \ + ( (cache_ptr)->index_size != (old_size) ) ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Pre HT entry size change SC failed") \ +} + +#define H5C2__POST_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size) \ +if ( ( (cache_ptr) == NULL ) || \ + ( (cache_ptr)->index_len <= 0 ) || \ + ( (cache_ptr)->index_size <= 0 ) || \ + ( (new_size) > (cache_ptr)->index_size ) || \ + ( ( (cache_ptr)->index_len == 1 ) && \ + ( (cache_ptr)->index_size != (new_size) ) ) ) { \ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Post HT entry size change SC failed") \ +} + +#else /* H5C2_DO_SANITY_CHECKS */ + +#define H5C2__PRE_HT_INSERT_SC(cache_ptr, entry_ptr, fail_val) +#define H5C2__PRE_HT_REMOVE_SC(cache_ptr, entry_ptr) +#define H5C2__PRE_HT_SEARCH_SC(cache_ptr, Addr, fail_val) +#define H5C2__POST_SUC_HT_SEARCH_SC(cache_ptr, entry_ptr, Addr, k, fail_val) +#define H5C2__POST_HT_SHIFT_TO_FRONT(cache_ptr, entry_ptr, k, fail_val) +#define H5C2__PRE_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size) +#define H5C2__POST_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size) + +#endif /* H5C2_DO_SANITY_CHECKS */ + + +#define H5C2__INSERT_IN_INDEX(cache_ptr, entry_ptr, fail_val) \ +{ \ + int k; \ + H5C2__PRE_HT_INSERT_SC(cache_ptr, entry_ptr, fail_val) \ + k = H5C2__HASH_FCN((entry_ptr)->addr); \ + if ( ((cache_ptr)->index)[k] == NULL ) \ + { \ + ((cache_ptr)->index)[k] = (entry_ptr); \ + } \ + else \ + { \ + (entry_ptr)->ht_next = ((cache_ptr)->index)[k]; \ + (entry_ptr)->ht_next->ht_prev = (entry_ptr); \ + ((cache_ptr)->index)[k] = (entry_ptr); \ + } \ + (cache_ptr)->index_len++; \ + (cache_ptr)->index_size += (entry_ptr)->size; \ + H5C2__UPDATE_STATS_FOR_HT_INSERTION(cache_ptr) \ +} + +#define H5C2__DELETE_FROM_INDEX(cache_ptr, entry_ptr) \ +{ \ + int k; \ + H5C2__PRE_HT_REMOVE_SC(cache_ptr, entry_ptr) \ + k = H5C2__HASH_FCN((entry_ptr)->addr); \ + if ( (entry_ptr)->ht_next ) \ + { \ + (entry_ptr)->ht_next->ht_prev = (entry_ptr)->ht_prev; \ + } \ + if ( (entry_ptr)->ht_prev ) \ + { \ + (entry_ptr)->ht_prev->ht_next = (entry_ptr)->ht_next; \ + } \ + if ( ((cache_ptr)->index)[k] == (entry_ptr) ) \ + { \ + ((cache_ptr)->index)[k] = (entry_ptr)->ht_next; \ + } \ + (entry_ptr)->ht_next = NULL; \ + (entry_ptr)->ht_prev = NULL; \ + (cache_ptr)->index_len--; \ + (cache_ptr)->index_size -= (entry_ptr)->size; \ + H5C2__UPDATE_STATS_FOR_HT_DELETION(cache_ptr) \ +} + +#define H5C2__SEARCH_INDEX(cache_ptr, Addr, entry_ptr, fail_val) \ +{ \ + int k; \ + int depth = 0; \ + H5C2__PRE_HT_SEARCH_SC(cache_ptr, Addr, fail_val) \ + k = H5C2__HASH_FCN(Addr); \ + entry_ptr = ((cache_ptr)->index)[k]; \ + while ( ( entry_ptr ) && ( H5F_addr_ne(Addr, (entry_ptr)->addr) ) ) \ + { \ + (entry_ptr) = (entry_ptr)->ht_next; \ + (depth)++; \ + } \ + if ( entry_ptr ) \ + { \ + H5C2__POST_SUC_HT_SEARCH_SC(cache_ptr, entry_ptr, Addr, k, fail_val) \ + if ( entry_ptr != ((cache_ptr)->index)[k] ) \ + { \ + if ( (entry_ptr)->ht_next ) \ + { \ + (entry_ptr)->ht_next->ht_prev = (entry_ptr)->ht_prev; \ + } \ + HDassert( (entry_ptr)->ht_prev != NULL ); \ + (entry_ptr)->ht_prev->ht_next = (entry_ptr)->ht_next; \ + ((cache_ptr)->index)[k]->ht_prev = (entry_ptr); \ + (entry_ptr)->ht_next = ((cache_ptr)->index)[k]; \ + (entry_ptr)->ht_prev = NULL; \ + ((cache_ptr)->index)[k] = (entry_ptr); \ + H5C2__POST_HT_SHIFT_TO_FRONT(cache_ptr, entry_ptr, k, fail_val) \ + } \ + } \ + H5C2__UPDATE_STATS_FOR_HT_SEARCH(cache_ptr, (entry_ptr != NULL), depth) \ +} + +#define H5C2__SEARCH_INDEX_NO_STATS(cache_ptr, Addr, entry_ptr, fail_val) \ +{ \ + int k; \ + int depth = 0; \ + H5C2__PRE_HT_SEARCH_SC(cache_ptr, Addr, fail_val) \ + k = H5C2__HASH_FCN(Addr); \ + entry_ptr = ((cache_ptr)->index)[k]; \ + while ( ( entry_ptr ) && ( H5F_addr_ne(Addr, (entry_ptr)->addr) ) ) \ + { \ + (entry_ptr) = (entry_ptr)->ht_next; \ + (depth)++; \ + } \ + if ( entry_ptr ) \ + { \ + H5C2__POST_SUC_HT_SEARCH_SC(cache_ptr, entry_ptr, Addr, k, fail_val) \ + if ( entry_ptr != ((cache_ptr)->index)[k] ) \ + { \ + if ( (entry_ptr)->ht_next ) \ + { \ + (entry_ptr)->ht_next->ht_prev = (entry_ptr)->ht_prev; \ + } \ + HDassert( (entry_ptr)->ht_prev != NULL ); \ + (entry_ptr)->ht_prev->ht_next = (entry_ptr)->ht_next; \ + ((cache_ptr)->index)[k]->ht_prev = (entry_ptr); \ + (entry_ptr)->ht_next = ((cache_ptr)->index)[k]; \ + (entry_ptr)->ht_prev = NULL; \ + ((cache_ptr)->index)[k] = (entry_ptr); \ + H5C2__POST_HT_SHIFT_TO_FRONT(cache_ptr, entry_ptr, k, fail_val) \ + } \ + } \ +} + +#define H5C2__UPDATE_INDEX_FOR_SIZE_CHANGE(cache_ptr, old_size, new_size) \ +{ \ + H5C2__PRE_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size) \ + (cache_ptr)->index_size -= old_size; \ + (cache_ptr)->index_size += new_size; \ + H5C2__POST_HT_ENTRY_SIZE_CHANGE_SC(cache_ptr, old_size, new_size) \ +} + + +/************************************************************************** + * + * Skip list insertion and deletion macros: + * + * These used to be functions, but I converted them to macros to avoid some + * function call overhead. + * + **************************************************************************/ + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__INSERT_ENTRY_IN_SLIST + * + * Purpose: Insert the specified instance of H5C2_cache_entry_t into + * the skip list in the specified instance of H5C2_t. Update + * the associated length and size fields. + * + * Return: N/A + * + * Programmer: John Mainzer, 5/10/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated function to set the in_tree flag when inserting + * an entry into the tree. Also modified the function to + * update the tree size and len fields instead of the similar + * index fields. + * + * All of this is part of the modifications to support the + * hash table. + * + * JRM -- 7/27/04 + * Converted the function H5C2_insert_entry_in_tree() into + * the macro H5C2__INSERT_ENTRY_IN_TREE in the hopes of + * wringing a little more speed out of the cache. + * + * Note that we don't bother to check if the entry is already + * in the tree -- if it is, H5SL_insert() will fail. + * + * QAK -- 11/27/04 + * Switched over to using skip list routines. + * + * JRM -- 6/27/06 + * Added fail_val parameter. + * + * JRM -- 8/25/06 + * Added the H5C2_DO_SANITY_CHECKS version of the macro. + * + * This version maintains the slist_len_increase and + * slist_size_increase fields that are used in sanity + * checks in the flush routines. + * + * All this is needed as the fractal heap needs to be + * able to dirty, resize and/or rename entries during the + * flush. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_DO_SANITY_CHECKS + +#define H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( (entry_ptr)->size > 0 ); \ + HDassert( H5F_addr_defined((entry_ptr)->addr) ); \ + HDassert( !((entry_ptr)->in_slist) ); \ + \ + if ( H5SL_insert((cache_ptr)->slist_ptr, entry_ptr, &(entry_ptr)->addr) \ + < 0 ) \ + HGOTO_ERROR(H5E_CACHE, H5E_BADVALUE, (fail_val), \ + "Can't insert entry in skip list") \ + \ + (entry_ptr)->in_slist = TRUE; \ + (cache_ptr)->slist_len++; \ + (cache_ptr)->slist_size += (entry_ptr)->size; \ + (cache_ptr)->slist_len_increase++; \ + (cache_ptr)->slist_size_increase += (entry_ptr)->size; \ + \ + HDassert( (cache_ptr)->slist_len > 0 ); \ + HDassert( (cache_ptr)->slist_size > 0 ); \ + \ +} /* H5C2__INSERT_ENTRY_IN_SLIST */ + +#else /* H5C2_DO_SANITY_CHECKS */ + +#define H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( (entry_ptr)->size > 0 ); \ + HDassert( H5F_addr_defined((entry_ptr)->addr) ); \ + HDassert( !((entry_ptr)->in_slist) ); \ + \ + if ( H5SL_insert((cache_ptr)->slist_ptr, entry_ptr, &(entry_ptr)->addr) \ + < 0 ) \ + HGOTO_ERROR(H5E_CACHE, H5E_BADVALUE, (fail_val), \ + "Can't insert entry in skip list") \ + \ + (entry_ptr)->in_slist = TRUE; \ + (cache_ptr)->slist_len++; \ + (cache_ptr)->slist_size += (entry_ptr)->size; \ + \ + HDassert( (cache_ptr)->slist_len > 0 ); \ + HDassert( (cache_ptr)->slist_size > 0 ); \ + \ +} /* H5C2__INSERT_ENTRY_IN_SLIST */ + +#endif /* H5C2_DO_SANITY_CHECKS */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__REMOVE_ENTRY_FROM_SLIST + * + * Purpose: Remove the specified instance of H5C2_cache_entry_t from the + * index skip list in the specified instance of H5C2_t. Update + * the associated length and size fields. + * + * Return: N/A + * + * Programmer: John Mainzer, 5/10/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated function for the addition of the hash table. + * + * JRM - 7/27/04 + * Converted from the function H5C2_remove_entry_from_tree() + * to the macro H5C2__REMOVE_ENTRY_FROM_TREE in the hopes of + * wringing a little more performance out of the cache. + * + * QAK -- 11/27/04 + * Switched over to using skip list routines. + * + * JRM -- 3/28/07 + * Updated sanity checks for the new is_read_only and + * ro_ref_count fields in H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#define H5C2__REMOVE_ENTRY_FROM_SLIST(cache_ptr, entry_ptr) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + HDassert( (entry_ptr)->in_slist ); \ + HDassert( (cache_ptr)->slist_ptr ); \ + \ + if ( H5SL_remove((cache_ptr)->slist_ptr, &(entry_ptr)->addr) \ + != (entry_ptr) ) \ + \ + HGOTO_ERROR(H5E_CACHE, H5E_BADVALUE, FAIL, \ + "Can't delete entry from skip list.") \ + \ + HDassert( (cache_ptr)->slist_len > 0 ); \ + (cache_ptr)->slist_len--; \ + HDassert( (cache_ptr)->slist_size >= (entry_ptr)->size ); \ + (cache_ptr)->slist_size -= (entry_ptr)->size; \ + (entry_ptr)->in_slist = FALSE; \ +} /* H5C2__REMOVE_ENTRY_FROM_SLIST */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__UPDATE_SLIST_FOR_SIZE_CHANGE + * + * Purpose: Update cache_ptr->slist_size for a change in the size of + * and entry in the slist. + * + * Return: N/A + * + * Programmer: John Mainzer, 9/07/05 + * + * Modifications: + * + * JRM -- 8/27/06 + * Added the H5C2_DO_SANITY_CHECKS version of the macro. + * + * This version maintains the slist_size_increase field + * that are used in sanity checks in the flush routines. + * + * All this is needed as the fractal heap needs to be + * able to dirty, resize and/or rename entries during the + * flush. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_DO_SANITY_CHECKS + +#define H5C2__UPDATE_SLIST_FOR_SIZE_CHANGE(cache_ptr, old_size, new_size) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (old_size) > 0 ); \ + HDassert( (new_size) > 0 ); \ + HDassert( (old_size) <= (cache_ptr)->slist_size ); \ + HDassert( (cache_ptr)->slist_len > 0 ); \ + HDassert( ((cache_ptr)->slist_len > 1) || \ + ( (cache_ptr)->slist_size == (old_size) ) ); \ + \ + (cache_ptr)->slist_size -= (old_size); \ + (cache_ptr)->slist_size += (new_size); \ + \ + (cache_ptr)->slist_size_increase -= (int64_t)(old_size); \ + (cache_ptr)->slist_size_increase += (int64_t)(new_size); \ + \ + HDassert( (new_size) <= (cache_ptr)->slist_size ); \ + HDassert( ( (cache_ptr)->slist_len > 1 ) || \ + ( (cache_ptr)->slist_size == (new_size) ) ); \ +} /* H5C2__REMOVE_ENTRY_FROM_SLIST */ + +#else /* H5C2_DO_SANITY_CHECKS */ + +#define H5C2__UPDATE_SLIST_FOR_SIZE_CHANGE(cache_ptr, old_size, new_size) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (old_size) > 0 ); \ + HDassert( (new_size) > 0 ); \ + HDassert( (old_size) <= (cache_ptr)->slist_size ); \ + HDassert( (cache_ptr)->slist_len > 0 ); \ + HDassert( ((cache_ptr)->slist_len > 1) || \ + ( (cache_ptr)->slist_size == (old_size) ) ); \ + \ + (cache_ptr)->slist_size -= (old_size); \ + (cache_ptr)->slist_size += (new_size); \ + \ + HDassert( (new_size) <= (cache_ptr)->slist_size ); \ + HDassert( ( (cache_ptr)->slist_len > 1 ) || \ + ( (cache_ptr)->slist_size == (new_size) ) ); \ +} /* H5C2__REMOVE_ENTRY_FROM_SLIST */ + +#endif /* H5C2_DO_SANITY_CHECKS */ + + +/************************************************************************** + * + * Replacement policy update macros: + * + * These used to be functions, but I converted them to macros to avoid some + * function call overhead. + * + **************************************************************************/ + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__FAKE_RP_FOR_MOST_RECENT_ACCESS + * + * Purpose: For efficiency, we sometimes change the order of flushes -- + * but doing so can confuse the replacement policy. This + * macro exists to allow us to specify an entry as the + * most recently touched so we can repair any such + * confusion. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the macro + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 10/13/05 + * + * Modifications: + * + * JRM -- 3/20/06 + * Modified macro to ignore pinned entries. Pinned entries + * do not appear in the data structures maintained by the + * replacement policy code, and thus this macro has nothing + * to do if called for such an entry. + * + * JRM -- 3/28/07 + * Added sanity checks using the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__FAKE_RP_FOR_MOST_RECENT_ACCESS(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( ! ((entry_ptr)->is_pinned) ) { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list, and re-insert it at the head.\ + */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* Use the dirty flag to infer whether the entry is on the clean or \ + * dirty LRU list, and remove it. Then insert it at the head of \ + * the same LRU list. \ + * \ + * At least initially, all entries should be clean. That may \ + * change, so we may as well deal with both cases now. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + } else { \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + /* End modified LRU specific code. */ \ + } \ +} /* H5C2__FAKE_RP_FOR_MOST_RECENT_ACCESS */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__FAKE_RP_FOR_MOST_RECENT_ACCESS(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( ! ((entry_ptr)->is_pinned) ) { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list, and re-insert it at the head \ + */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + } \ +} /* H5C2__FAKE_RP_FOR_MOST_RECENT_ACCESS */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_EVICTION + * + * Purpose: Update the replacement policy data structures for an + * eviction of the specified cache entry. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the function + * should switch on the current policy and act accordingly. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 5/10/04 + * + * Modifications: + * + * JRM - 7/27/04 + * Converted the function H5C2_update_rp_for_eviction() to the + * macro H5C2__UPDATE_RP_FOR_EVICTION in an effort to squeeze + * a bit more performance out of the cache. + * + * At least for the first cut, I am leaving the comments and + * white space in the macro. If they cause dificulties with + * the pre-processor, I'll have to remove them. + * + * JRM - 7/28/04 + * Split macro into two version, one supporting the clean and + * dirty LRU lists, and the other not. Yet another attempt + * at optimization. + * + * JRM - 3/20/06 + * Pinned entries can't be evicted, so this entry should never + * be called on a pinned entry. Added assert to verify this. + * + * JRM -- 3/28/07 + * Added sanity checks for the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_EVICTION(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( !((entry_ptr)->is_pinned) ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list. */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* If the entry is clean when it is evicted, it should be on the \ + * clean LRU list, if it was dirty, it should be on the dirty LRU list. \ + * Remove it from the appropriate list according to the value of the \ + * dirty flag. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + } else { \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ +} /* H5C2__UPDATE_RP_FOR_EVICTION */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_EVICTION(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( !((entry_ptr)->is_pinned) ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list. */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ +} /* H5C2__UPDATE_RP_FOR_EVICTION */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_FLUSH + * + * Purpose: Update the replacement policy data structures for a flush + * of the specified cache entry. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the function + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 5/6/04 + * + * Modifications: + * + * JRM - 7/27/04 + * Converted the function H5C2_update_rp_for_flush() to the + * macro H5C2__UPDATE_RP_FOR_FLUSH in an effort to squeeze + * a bit more performance out of the cache. + * + * At least for the first cut, I am leaving the comments and + * white space in the macro. If they cause dificulties with + * pre-processor, I'll have to remove them. + * + * JRM - 7/28/04 + * Split macro into two versions, one supporting the clean and + * dirty LRU lists, and the other not. Yet another attempt + * at optimization. + * + * JRM - 3/20/06 + * While pinned entries can be flushed, they don't reside in + * the replacement policy data structures when unprotected. + * Thus I modified this macro to do nothing if the entry is + * pinned. + * + * JRM - 3/28/07 + * Added sanity checks based on the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_FLUSH(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( ! ((entry_ptr)->is_pinned) ) { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list, and re-insert it at the \ + * head. \ + */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* since the entry is being flushed or cleared, one would think \ + * that it must be dirty -- but that need not be the case. Use the \ + * dirty flag to infer whether the entry is on the clean or dirty \ + * LRU list, and remove it. Then insert it at the head of the \ + * clean LRU list. \ + * \ + * The function presumes that a dirty entry will be either cleared \ + * or flushed shortly, so it is OK if we put a dirty entry on the \ + * clean LRU list. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + } else { \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + } \ +} /* H5C2__UPDATE_RP_FOR_FLUSH */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_FLUSH(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( ! ((entry_ptr)->is_pinned) ) { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list, and re-insert it at the \ + * head. \ + */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + } \ +} /* H5C2__UPDATE_RP_FOR_FLUSH */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_INSERTION + * + * Purpose: Update the replacement policy data structures for an + * insertion of the specified cache entry. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the function + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 5/17/04 + * + * Modifications: + * + * JRM - 7/27/04 + * Converted the function H5C2_update_rp_for_insertion() to the + * macro H5C2__UPDATE_RP_FOR_INSERTION in an effort to squeeze + * a bit more performance out of the cache. + * + * At least for the first cut, I am leaving the comments and + * white space in the macro. If they cause dificulties with + * pre-processor, I'll have to remove them. + * + * JRM - 7/28/04 + * Split macro into two version, one supporting the clean and + * dirty LRU lists, and the other not. Yet another attempt + * at optimization. + * + * JRM - 3/10/06 + * This macro should never be called on a pinned entry. + * Inserted an assert to verify this. + * + * JRM - 8/9/06 + * Not any more. We must now allow insertion of pinned + * entries. Updated macro to support this. + * + * JRM - 3/28/07 + * Added sanity checks using the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_INSERTION(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* insert the entry at the head of the LRU list. */ \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* insert the entry at the head of the clean or dirty LRU list as \ + * appropriate. \ + */ \ + \ + if ( entry_ptr->is_dirty ) { \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + } else { \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + /* End modified LRU specific code. */ \ + } \ +} + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_INSERTION(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* insert the entry at the head of the LRU list. */ \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + } \ +} + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_PROTECT + * + * Purpose: Update the replacement policy data structures for a + * protect of the specified cache entry. + * + * To do this, unlink the specified entry from any data + * structures used by the replacement policy, and add the + * entry to the protected list. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the function + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 5/17/04 + * + * Modifications: + * + * JRM - 7/27/04 + * Converted the function H5C2_update_rp_for_protect() to the + * macro H5C2__UPDATE_RP_FOR_PROTECT in an effort to squeeze + * a bit more performance out of the cache. + * + * At least for the first cut, I am leaving the comments and + * white space in the macro. If they cause dificulties with + * pre-processor, I'll have to remove them. + * + * JRM - 7/28/04 + * Split macro into two version, one supporting the clean and + * dirty LRU lists, and the other not. Yet another attempt + * at optimization. + * + * JRM - 3/17/06 + * Modified macro to attempt to remove pinned entriese from + * the pinned entry list instead of from the data structures + * maintained by the replacement policy. + * + * JRM - 3/28/07 + * Added sanity checks based on the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_PROTECT(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list. */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* Similarly, remove the entry from the clean or dirty LRU list \ + * as appropriate. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + \ + } else { \ + \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + /* End modified LRU specific code. */ \ + } \ + \ + /* Regardless of the replacement policy, or whether the entry is \ + * pinned, now add the entry to the protected list. \ + */ \ + \ + H5C2__DLL_APPEND((entry_ptr), (cache_ptr)->pl_head_ptr, \ + (cache_ptr)->pl_tail_ptr, \ + (cache_ptr)->pl_len, \ + (cache_ptr)->pl_size, (fail_val)) \ +} /* H5C2__UPDATE_RP_FOR_PROTECT */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_PROTECT(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list. */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + } \ + \ + /* Regardless of the replacement policy, or whether the entry is \ + * pinned, now add the entry to the protected list. \ + */ \ + \ + H5C2__DLL_APPEND((entry_ptr), (cache_ptr)->pl_head_ptr, \ + (cache_ptr)->pl_tail_ptr, \ + (cache_ptr)->pl_len, \ + (cache_ptr)->pl_size, (fail_val)) \ +} /* H5C2__UPDATE_RP_FOR_PROTECT */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_RENAME + * + * Purpose: Update the replacement policy data structures for a + * rename of the specified cache entry. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the function + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 5/17/04 + * + * Modifications: + * + * JRM - 7/27/04 + * Converted the function H5C2_update_rp_for_rename() to the + * macro H5C2__UPDATE_RP_FOR_RENAME in an effort to squeeze + * a bit more performance out of the cache. + * + * At least for the first cut, I am leaving the comments and + * white space in the macro. If they cause dificulties with + * pre-processor, I'll have to remove them. + * + * JRM - 7/28/04 + * Split macro into two version, one supporting the clean and + * dirty LRU lists, and the other not. Yet another attempt + * at optimization. + * + * JRM - 6/23/05 + * Added the was_dirty parameter. It is possible that + * the entry was clean when it was renamed -- if so it + * it is in the clean LRU regardless of the current + * value of the is_dirty field. + * + * At present, all renamed entries are forced to be + * dirty. This macro is a bit more general that that, + * to allow it to function correctly should that policy + * be relaxed in the future. + * + * JRM - 3/17/06 + * Modified macro to do nothing if the entry is pinned. + * In this case, the entry is on the pinned entry list, not + * in the replacement policy data structures, so there is + * nothing to be done. + * + * JRM - 3/28/07 + * Added sanity checks using the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_RENAME(cache_ptr, entry_ptr, was_dirty, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( ! ((entry_ptr)->is_pinned) ) { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list, and re-insert it at the head. \ + */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* remove the entry from either the clean or dirty LUR list as \ + * indicated by the was_dirty parameter \ + */ \ + if ( was_dirty ) { \ + \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + \ + } else { \ + \ + H5C2__AUX_DLL_REMOVE((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + /* insert the entry at the head of either the clean or dirty LRU \ + * list as appropriate. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + \ + } else { \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + /* End modified LRU specific code. */ \ + } \ +} /* H5C2__UPDATE_RP_FOR_RENAME */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_RENAME(cache_ptr, entry_ptr, was_dirty, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + if ( ! ((entry_ptr)->is_pinned) ) { \ + \ + /* modified LRU specific code */ \ + \ + /* remove the entry from the LRU list, and re-insert it at the head. \ + */ \ + \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + } \ +} /* H5C2__UPDATE_RP_FOR_RENAME */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_SIZE_CHANGE + * + * Purpose: Update the replacement policy data structures for a + * size change of the specified cache entry. + * + * To do this, determine if the entry is pinned. If it is, + * update the size of the pinned entry list. + * + * If it isn't pinned, the entry must handled by the + * replacement policy. Update the appropriate replacement + * policy data structures. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the function + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 8/23/06 + * + * Modifications: + * + * JRM -- 3/28/07 + * Added sanity checks based on the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_SIZE_CHANGE(cache_ptr, entry_ptr, new_size) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + HDassert( new_size > 0 ); \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, \ + (entry_ptr)->size, \ + (new_size)); \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* Update the size of the LRU list */ \ + \ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ + \ + /* Similarly, update the size of the clean or dirty LRU list as \ + * appropriate. At present, the entry must be clean, but that \ + * could change. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + \ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ + \ + } else { \ + \ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ + } \ + \ + /* End modified LRU specific code. */ \ + } \ + \ +} /* H5C2__UPDATE_RP_FOR_SIZE_CHANGE */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_SIZE_CHANGE(cache_ptr, entry_ptr, new_size) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->size > 0 ); \ + HDassert( new_size > 0 ); \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, \ + (entry_ptr)->size, \ + (new_size)); \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* Update the size of the LRU list */ \ + \ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (entry_ptr)->size, \ + (new_size)); \ + \ + /* End modified LRU specific code. */ \ + } \ + \ +} /* H5C2__UPDATE_RP_FOR_SIZE_CHANGE */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_UNPIN + * + * Purpose: Update the replacement policy data structures for an + * unpin of the specified cache entry. + * + * To do this, unlink the specified entry from the protected + * entry list, and re-insert it in the data structures used + * by the current replacement policy. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the macro + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 3/22/06 + * + * Modifications: + * + * JRM -- 3/28/07 + * Added sanity checks based on the new is_read_only and + * ro_ref_count fields of struct H5C2_cache_entry_t. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_UNPIN(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->is_pinned); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + /* Regardless of the replacement policy, remove the entry from the \ + * pinned entry list. \ + */ \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + /* modified LRU specific code */ \ + \ + /* insert the entry at the head of the LRU list. */ \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* Similarly, insert the entry at the head of either the clean or \ + * dirty LRU list as appropriate. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + \ + } else { \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + /* End modified LRU specific code. */ \ + \ +} /* H5C2__UPDATE_RP_FOR_UNPIN */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_UNPIN(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( !((entry_ptr)->is_protected) ); \ + HDassert( !((entry_ptr)->is_read_only) ); \ + HDassert( ((entry_ptr)->ro_ref_count) == 0 ); \ + HDassert( (entry_ptr)->is_pinned); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + /* Regardless of the replacement policy, remove the entry from the \ + * pinned entry list. \ + */ \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + /* modified LRU specific code */ \ + \ + /* insert the entry at the head of the LRU list. */ \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + \ +} /* H5C2__UPDATE_RP_FOR_UNPIN */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/*------------------------------------------------------------------------- + * + * Macro: H5C2__UPDATE_RP_FOR_UNPROTECT + * + * Purpose: Update the replacement policy data structures for an + * unprotect of the specified cache entry. + * + * To do this, unlink the specified entry from the protected + * list, and re-insert it in the data structures used by the + * current replacement policy. + * + * At present, we only support the modified LRU policy, so + * this function deals with that case unconditionally. If + * we ever support other replacement policies, the function + * should switch on the current policy and act accordingly. + * + * Return: N/A + * + * Programmer: John Mainzer, 5/19/04 + * + * Modifications: + * + * JRM - 7/27/04 + * Converted the function H5C2_update_rp_for_unprotect() to + * the macro H5C2__UPDATE_RP_FOR_UNPROTECT in an effort to + * squeeze a bit more performance out of the cache. + * + * At least for the first cut, I am leaving the comments and + * white space in the macro. If they cause dificulties with + * pre-processor, I'll have to remove them. + * + * JRM - 7/28/04 + * Split macro into two version, one supporting the clean and + * dirty LRU lists, and the other not. Yet another attempt + * at optimization. + * + * JRM - 3/17/06 + * Modified macro to put pinned entries on the pinned entry + * list instead of inserting them in the data structures + * maintained by the replacement policy. + * + *------------------------------------------------------------------------- + */ + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + +#define H5C2__UPDATE_RP_FOR_UNPROTECT(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( (entry_ptr)->is_protected); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + /* Regardless of the replacement policy, remove the entry from the \ + * protected list. \ + */ \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->pl_head_ptr, \ + (cache_ptr)->pl_tail_ptr, (cache_ptr)->pl_len, \ + (cache_ptr)->pl_size, (fail_val)) \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* insert the entry at the head of the LRU list. */ \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* Similarly, insert the entry at the head of either the clean or \ + * dirty LRU list as appropriate. \ + */ \ + \ + if ( (entry_ptr)->is_dirty ) { \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->dLRU_head_ptr, \ + (cache_ptr)->dLRU_tail_ptr, \ + (cache_ptr)->dLRU_list_len, \ + (cache_ptr)->dLRU_list_size, (fail_val)) \ + \ + } else { \ + \ + H5C2__AUX_DLL_PREPEND((entry_ptr), (cache_ptr)->cLRU_head_ptr, \ + (cache_ptr)->cLRU_tail_ptr, \ + (cache_ptr)->cLRU_list_len, \ + (cache_ptr)->cLRU_list_size, (fail_val)) \ + } \ + \ + /* End modified LRU specific code. */ \ + } \ + \ +} /* H5C2__UPDATE_RP_FOR_UNPROTECT */ + +#else /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + +#define H5C2__UPDATE_RP_FOR_UNPROTECT(cache_ptr, entry_ptr, fail_val) \ +{ \ + HDassert( (cache_ptr) ); \ + HDassert( (cache_ptr)->magic == H5C2__H5C2_T_MAGIC ); \ + HDassert( (entry_ptr) ); \ + HDassert( (entry_ptr)->is_protected); \ + HDassert( (entry_ptr)->size > 0 ); \ + \ + /* Regardless of the replacement policy, remove the entry from the \ + * protected list. \ + */ \ + H5C2__DLL_REMOVE((entry_ptr), (cache_ptr)->pl_head_ptr, \ + (cache_ptr)->pl_tail_ptr, (cache_ptr)->pl_len, \ + (cache_ptr)->pl_size, (fail_val)) \ + \ + if ( (entry_ptr)->is_pinned ) { \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->pel_head_ptr, \ + (cache_ptr)->pel_tail_ptr, \ + (cache_ptr)->pel_len, \ + (cache_ptr)->pel_size, (fail_val)) \ + \ + } else { \ + \ + /* modified LRU specific code */ \ + \ + /* insert the entry at the head of the LRU list. */ \ + \ + H5C2__DLL_PREPEND((entry_ptr), (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, (fail_val)) \ + \ + /* End modified LRU specific code. */ \ + } \ +} /* H5C2__UPDATE_RP_FOR_UNPROTECT */ + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + +/* + * Private file-scope variables. + */ + +/* Declare a free list to manage the H5C2_t struct */ +H5FL_DEFINE_STATIC(H5C2_t); + + +/* + * Private file-scope function declarations: + */ + +static herr_t H5C2__auto_adjust_cache_size(H5C2_t * cache_ptr, + hid_t dxpl_id, + hbool_t write_permitted); + +static herr_t H5C2__autoadjust__ageout(H5C2_t * cache_ptr, + double hit_rate, + enum H5C2_resize_status * status_ptr, + size_t * new_max_cache_size_ptr, + hid_t dxpl_id, + hbool_t write_permitted); + +static herr_t H5C2__autoadjust__ageout__cycle_epoch_marker(H5C2_t * cache_ptr); + +static herr_t H5C2__autoadjust__ageout__evict_aged_out_entries(hid_t dxpl_id, + H5C2_t * cache_ptr, + hbool_t write_permitted); + +static herr_t H5C2__autoadjust__ageout__insert_new_marker(H5C2_t * cache_ptr); + +static herr_t H5C2__autoadjust__ageout__remove_all_markers(H5C2_t * cache_ptr); + +static herr_t H5C2__autoadjust__ageout__remove_excess_markers(H5C2_t * cache_ptr); + +static herr_t H5C2_flush_single_entry(H5F_t * f, + hid_t dxpl_id, + H5C2_t * cache_ptr, + const H5C2_class_t * type_ptr, + haddr_t addr, + unsigned flags, + hbool_t del_entry_from_slist_on_destroy); + +static herr_t H5C2_flush_invalidate_cache(hid_t dxpl_id, + H5C2_t * cache_ptr, + unsigned flags); + +static void * H5C2_load_entry(H5F_t * f, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + size_t len, + hbool_t chk_len, + const void * udata_ptr); + +static herr_t H5C2_make_space_in_cache(hid_t dxpl_id, + H5C2_t * cache_ptr, + size_t space_needed, + hbool_t write_permitted); + +#if H5C2_DO_EXTREME_SANITY_CHECKS +static herr_t H5C2_validate_lru_list(H5C2_t * cache_ptr); +static herr_t H5C2_verify_not_in_index(H5C2_t * cache_ptr, + H5C2_cache_entry_t * entry_ptr); +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + +/**************************************************************************** + * + * #defines and declarations for epoch marker cache entries. + * + * As a strategy for automatic cache size reduction, the cache may insert + * marker entries in the LRU list at the end of each epoch. These markers + * are then used to identify entries that have not been accessed for n + * epochs so that they can be evicted from the cache. + * + ****************************************************************************/ + +/* Note that H5C2__MAX_EPOCH_MARKERS is defined in H5C2pkg.h, not here because + * it is needed to dimension arrays in H5C2_t. + */ + +#define H5C2__EPOCH_MARKER_TYPE H5C2__MAX_NUM_TYPE_IDS + +static void * H5C2_epoch_marker_deserialize(haddr_t addr, + size_t len, + const void * image_ptr, + const void * udata_ptr, + hbool_t * dirty_ptr); +static herr_t H5C2_epoch_marker_image_len(void * thing, + size_t *image_len_ptr); +static herr_t H5C2_epoch_marker_serialize(haddr_t addr, + size_t len, + void * image_ptr, + void * thing, + unsigned * flags_ptr, + haddr_t * new_addr_ptr, + size_t * new_len_ptr, + void ** new_image_ptr_ptr); +static herr_t H5C2_epoch_marker_free_icr(haddr_t addr, + size_t len, + void * thing); +static herr_t H5C2_epoch_marker_clear_dirty_bits(haddr_t addr, + size_t len, + void * thing); + +const H5C2_class_t epoch_marker_class_2 = +{ + /* id = */ H5C2__EPOCH_MARKER_TYPE, + /* name = */ "epoch marker", + /* mem_type = */ H5FD_MEM_DEFAULT, /* value doesn't matter */ + /* deserialize = */ &H5C2_epoch_marker_deserialize, + /* image_len = */ &H5C2_epoch_marker_image_len, + /* serialize = */ &H5C2_epoch_marker_serialize, + /* free_icr = */ &H5C2_epoch_marker_free_icr, + /* clear_dirty_bits = */ &H5C2_epoch_marker_clear_dirty_bits, +}; + + +/*************************************************************************** + * Class functions for H5C2__EPOCH_MAKER_TYPE: + * + * None of these functions should ever be called, so there is no point in + * documenting them separately. + * JRM - 11/16/04 + * + ***************************************************************************/ + +static void * +H5C2_epoch_marker_deserialize(haddr_t UNUSED addr, + size_t UNUSED len, + const void UNUSED * image_ptr, + const void UNUSED * udata_ptr, + hbool_t UNUSED * dirty_ptr) +{ + void * ret_value = NULL; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_epoch_marker_serialize, NULL) + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, NULL, "called unreachable fcn.") + +done: + + FUNC_LEAVE_NOAPI(ret_value) +} + +static herr_t +H5C2_epoch_marker_image_len(void UNUSED *thing, + size_t UNUSED *image_len_ptr) +{ + herr_t ret_value = FAIL; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_epoch_marker_image_len, FAIL) + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "called unreachable fcn.") + +done: + + FUNC_LEAVE_NOAPI(ret_value) +} + +static herr_t +H5C2_epoch_marker_serialize(haddr_t UNUSED addr, + size_t UNUSED len, + void UNUSED * image_ptr, + void UNUSED * thing, + unsigned UNUSED * flags_ptr, + haddr_t UNUSED * new_addr_ptr, + size_t UNUSED * new_len_ptr, + void UNUSED ** new_image_ptr_ptr) +{ + herr_t ret_value = FAIL; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_epoch_marker_serialize, FAIL) + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "called unreachable fcn.") + +done: + + FUNC_LEAVE_NOAPI(ret_value) +} + +static herr_t +H5C2_epoch_marker_free_icr(haddr_t UNUSED addr, + size_t UNUSED len, + void UNUSED * thing) +{ + herr_t ret_value = FAIL; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_epoch_marker_free_icr, FAIL) + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "called unreachable fcn.") + +done: + + FUNC_LEAVE_NOAPI(ret_value) +} + +static herr_t +H5C2_epoch_marker_clear_dirty_bits(haddr_t UNUSED addr, + size_t UNUSED len, + void UNUSED * thing) +{ + herr_t ret_value = FAIL; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_epoch_marker_clear_dirty_bits, FAIL) + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "called unreachable fcn.") + +done: + + FUNC_LEAVE_NOAPI(ret_value) +} + + + +/*------------------------------------------------------------------------- + * Function: H5C2_create + * + * Purpose: Allocate, initialize, and return the address of a new + * instance of H5C2_t. + * + * In general, the max_cache_size parameter must be positive, + * and the min_clean_size parameter must lie in the closed + * interval [0, max_cache_size]. + * + * The check_write_permitted parameter must either be NULL, + * or point to a function of type H5C2_write_permitted_func_t. + * If it is NULL, the cache will use the write_permitted + * flag to determine whether writes are permitted. + * + * Return: Success: Pointer to the new instance. + * + * Failure: NULL + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * JRM -- 7/20/04 + * Updated for the addition of the hash table. + * + * JRM -- 10/5/04 + * Added call to H5C2_reset_cache_hit_rate_stats(). Also + * added initialization for cache_is_full flag and for + * resize_ctl. + * + * JRM -- 11/12/04 + * Added initialization for the new size_decreased field. + * + * JRM -- 11/17/04 + * Added/updated initialization for the automatic cache + * size control data structures. + * + * JRM -- 6/24/05 + * Added support for the new write_permitted field of + * the H5C2_t structure. + * + * JRM -- 7/5/05 + * Added the new log_flush parameter and supporting code. + * + * JRM -- 9/21/05 + * Added the new aux_ptr parameter and supporting code. + * + * JRM -- 1/20/06 + * Added initialization of the new prefix field in H5C2_t. + * + * JRM -- 3/16/06 + * Added initialization for the pinned entry related fields. + * + * JRM -- 5/31/06 + * Added initialization for the trace_file_ptr field. + * + * JRM -- 8/19/06 + * Added initialization for the flush_in_progress field. + * + * JRM -- 8/25/06 + * Added initialization for the slist_len_increase and + * slist_size_increase fields. These fields are used + * for sanity checking in the flush process, and are not + * compiled in unless H5C2_DO_SANITY_CHECKS is TRUE. + * + * JRM -- 3/28/07 + * Added initialization for the new is_read_only and + * ro_ref_count fields. + * + * JRM -- 7/10/07 + * Added the f parameter, along with initialization of + * the field of the same in in H5C2_t. Also removed the + * type name table, as type names are now included in + * instances of H5C2_class_t. + * + * JRM -- 3/28/07 + * Added initialization for the new is_read_only and + * ro_ref_count fields. + * + *------------------------------------------------------------------------- + */ + +H5C2_t * +H5C2_create(const H5F_t * f, + size_t max_cache_size, + size_t min_clean_size, + int max_type_id, + const char * (* type_name_table_ptr), + H5C2_write_permitted_func_t check_write_permitted, + hbool_t write_permitted, + H5C2_log_flush_func_t log_flush, + void * aux_ptr) +{ + int i; + H5C2_t * cache_ptr = NULL; + H5C2_t * ret_value = NULL; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_create, NULL) + + HDassert( f ); + + HDassert( max_cache_size >= H5C2__MIN_MAX_CACHE_SIZE ); + HDassert( max_cache_size <= H5C2__MAX_MAX_CACHE_SIZE ); + HDassert( min_clean_size <= max_cache_size ); + + HDassert( max_type_id >= 0 ); + HDassert( max_type_id < H5C2__MAX_NUM_TYPE_IDS ); + + HDassert( ( write_permitted == TRUE ) || ( write_permitted == FALSE ) ); + + for ( i = 0; i <= max_type_id; i++ ) { + + HDassert( (type_name_table_ptr)[i] ); + HDassert( HDstrlen(( type_name_table_ptr)[i]) > 0 ); + } + + if ( NULL == (cache_ptr = H5FL_CALLOC(H5C2_t)) ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, \ + "memory allocation failed") + } + + if ( (cache_ptr->slist_ptr = H5SL_create(H5SL_TYPE_HADDR,0.5,(size_t)16)) + == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTCREATE, NULL, "can't create skip list.") + } + + /* If we get this far, we should succeed. Go ahead and initialize all + * the fields. + */ + + cache_ptr->magic = H5C2__H5C2_T_MAGIC; + + cache_ptr->f = (H5F_t *)f; + + cache_ptr->flush_in_progress = FALSE; + + cache_ptr->trace_file_ptr = NULL; + + cache_ptr->aux_ptr = aux_ptr; + + cache_ptr->max_type_id = max_type_id; + + cache_ptr->type_name_table_ptr = type_name_table_ptr; + + cache_ptr->max_cache_size = max_cache_size; + cache_ptr->min_clean_size = min_clean_size; + + cache_ptr->check_write_permitted = check_write_permitted; + cache_ptr->write_permitted = write_permitted; + + cache_ptr->log_flush = log_flush; + + cache_ptr->evictions_enabled = TRUE; + + cache_ptr->index_len = 0; + cache_ptr->index_size = (size_t)0; + + cache_ptr->slist_len = 0; + cache_ptr->slist_size = (size_t)0; + +#if H5C2_DO_SANITY_CHECKS + cache_ptr->slist_len_increase = 0; + cache_ptr->slist_size_increase = 0; +#endif /* H5C2_DO_SANITY_CHECKS */ + + for ( i = 0; i < H5C2__HASH_TABLE_LEN; i++ ) + { + (cache_ptr->index)[i] = NULL; + } + + cache_ptr->pl_len = 0; + cache_ptr->pl_size = (size_t)0; + cache_ptr->pl_head_ptr = NULL; + cache_ptr->pl_tail_ptr = NULL; + + cache_ptr->pel_len = 0; + cache_ptr->pel_size = (size_t)0; + cache_ptr->pel_head_ptr = NULL; + cache_ptr->pel_tail_ptr = NULL; + + cache_ptr->LRU_list_len = 0; + cache_ptr->LRU_list_size = (size_t)0; + cache_ptr->LRU_head_ptr = NULL; + cache_ptr->LRU_tail_ptr = NULL; + + cache_ptr->cLRU_list_len = 0; + cache_ptr->cLRU_list_size = (size_t)0; + cache_ptr->cLRU_head_ptr = NULL; + cache_ptr->cLRU_tail_ptr = NULL; + + cache_ptr->dLRU_list_len = 0; + cache_ptr->dLRU_list_size = (size_t)0; + cache_ptr->dLRU_head_ptr = NULL; + cache_ptr->dLRU_tail_ptr = NULL; + + cache_ptr->size_increase_possible = FALSE; + cache_ptr->size_decrease_possible = FALSE; + cache_ptr->resize_enabled = FALSE; + cache_ptr->cache_full = FALSE; + cache_ptr->size_decreased = FALSE; + + (cache_ptr->resize_ctl).version = H5C2__CURR_AUTO_SIZE_CTL_VER; + (cache_ptr->resize_ctl).rpt_fcn = NULL; + (cache_ptr->resize_ctl).set_initial_size = FALSE; + (cache_ptr->resize_ctl).initial_size = H5C2__DEF_AR_INIT_SIZE; + (cache_ptr->resize_ctl).min_clean_fraction = H5C2__DEF_AR_MIN_CLEAN_FRAC; + (cache_ptr->resize_ctl).max_size = H5C2__DEF_AR_MAX_SIZE; + (cache_ptr->resize_ctl).min_size = H5C2__DEF_AR_MIN_SIZE; + (cache_ptr->resize_ctl).epoch_length = H5C2__DEF_AR_EPOCH_LENGTH; + + (cache_ptr->resize_ctl).incr_mode = H5C2_incr__off; + (cache_ptr->resize_ctl).lower_hr_threshold = H5C2__DEF_AR_LOWER_THRESHHOLD; + (cache_ptr->resize_ctl).increment = H5C2__DEF_AR_INCREMENT; + (cache_ptr->resize_ctl).apply_max_increment = TRUE; + (cache_ptr->resize_ctl).max_increment = H5C2__DEF_AR_MAX_INCREMENT; + + (cache_ptr->resize_ctl).decr_mode = H5C2_decr__off; + (cache_ptr->resize_ctl).upper_hr_threshold = H5C2__DEF_AR_UPPER_THRESHHOLD; + (cache_ptr->resize_ctl).decrement = H5C2__DEF_AR_DECREMENT; + (cache_ptr->resize_ctl).apply_max_decrement = TRUE; + (cache_ptr->resize_ctl).max_decrement = H5C2__DEF_AR_MAX_DECREMENT; + (cache_ptr->resize_ctl).epochs_before_eviction = H5C2__DEF_AR_EPCHS_B4_EVICT; + (cache_ptr->resize_ctl).apply_empty_reserve = TRUE; + (cache_ptr->resize_ctl).empty_reserve = H5C2__DEF_AR_EMPTY_RESERVE; + + cache_ptr->epoch_markers_active = 0; + + /* no need to initialize the ring buffer itself */ + cache_ptr->epoch_marker_ringbuf_first = 1; + cache_ptr->epoch_marker_ringbuf_last = 0; + cache_ptr->epoch_marker_ringbuf_size = 0; + + for ( i = 0; i < H5C2__MAX_EPOCH_MARKERS; i++ ) + { + (cache_ptr->epoch_marker_active)[i] = FALSE; + + ((cache_ptr->epoch_markers)[i]).magic = + H5C2__H5C2_CACHE_ENTRY_T_MAGIC; + ((cache_ptr->epoch_markers)[i]).addr = (haddr_t)i; + ((cache_ptr->epoch_markers)[i]).size = (size_t)0; + ((cache_ptr->epoch_markers)[i]).type = &epoch_marker_class_2; + ((cache_ptr->epoch_markers)[i]).is_dirty = FALSE; + ((cache_ptr->epoch_markers)[i]).dirtied = FALSE; + ((cache_ptr->epoch_markers)[i]).is_protected = FALSE; + ((cache_ptr->epoch_markers)[i]).is_read_only = FALSE; + ((cache_ptr->epoch_markers)[i]).ro_ref_count = 0; + ((cache_ptr->epoch_markers)[i]).is_pinned = FALSE; + ((cache_ptr->epoch_markers)[i]).in_slist = FALSE; + ((cache_ptr->epoch_markers)[i]).ht_next = NULL; + ((cache_ptr->epoch_markers)[i]).ht_prev = NULL; + ((cache_ptr->epoch_markers)[i]).next = NULL; + ((cache_ptr->epoch_markers)[i]).prev = NULL; + ((cache_ptr->epoch_markers)[i]).aux_next = NULL; + ((cache_ptr->epoch_markers)[i]).aux_prev = NULL; +#if H5C2_COLLECT_CACHE_ENTRY_STATS + ((cache_ptr->epoch_markers)[i]).accesses = 0; + ((cache_ptr->epoch_markers)[i]).clears = 0; + ((cache_ptr->epoch_markers)[i]).flushes = 0; + ((cache_ptr->epoch_markers)[i]).pins = 0; +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + } + + if ( H5C2_reset_cache_hit_rate_stats(cache_ptr) != SUCCEED ) { + + /* this should be impossible... */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, NULL, \ + "H5C2_reset_cache_hit_rate_stats failed.") + } + + H5C2_stats__reset(cache_ptr); + + cache_ptr->skip_file_checks = FALSE; + cache_ptr->skip_dxpl_id_checks = FALSE; + cache_ptr->prefix[0] = '\0'; /* empty string */ + + /* Set return value */ + ret_value = cache_ptr; + +done: + + if ( ret_value == 0 ) { + + if ( cache_ptr != NULL ) { + + if ( cache_ptr->slist_ptr != NULL ) + H5SL_close(cache_ptr->slist_ptr); + + cache_ptr->magic = 0; + H5FL_FREE(H5C2_t, cache_ptr); + cache_ptr = NULL; + + } /* end if */ + + } /* end if */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_create() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_def_auto_resize_rpt_fcn + * + * Purpose: Print results of a automatic cache resize. + * + * This function should only be used where HDprintf() behaves + * well -- i.e. not on Windows. + * + * Return: void + * + * Programmer: John Mainzer + * 10/27/04 + * + * Modifications: + * + * JRM -- 11/22/04 + * Reworked function to adapt it to the addition of the + * ageout method of cache size reduction. + * + * JRM -- 1/19/06 + * Updated function for display the new prefix field of + * H5C2_t in output. + * + *------------------------------------------------------------------------- + */ +void +H5C2_def_auto_resize_rpt_fcn(H5C2_t * cache_ptr, +#ifndef NDEBUG + int32_t version, +#else /* NDEBUG */ + int32_t UNUSED version, +#endif /* NDEBUG */ + double hit_rate, + enum H5C2_resize_status status, + size_t old_max_cache_size, + size_t new_max_cache_size, + size_t old_min_clean_size, + size_t new_min_clean_size) +{ + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( version == H5C2__CURR_AUTO_RESIZE_RPT_FCN_VER ); + + switch ( status ) + { + case in_spec2: + HDfprintf(stdout, + "%sAuto cache resize -- no change. (hit rate = %lf)\n", + cache_ptr->prefix, hit_rate); + break; + + case increase2: + HDassert( hit_rate < (cache_ptr->resize_ctl).lower_hr_threshold ); + HDassert( old_max_cache_size < new_max_cache_size ); + + HDfprintf(stdout, + "%sAuto cache resize -- hit rate (%lf) out of bounds low (%6.5lf).\n", + cache_ptr->prefix, hit_rate, + (cache_ptr->resize_ctl).lower_hr_threshold); + + HDfprintf(stdout, + "%s cache size increased from (%Zu/%Zu) to (%Zu/%Zu).\n", + cache_ptr->prefix, + old_max_cache_size, + old_min_clean_size, + new_max_cache_size, + new_min_clean_size); + break; + + case decrease2: + HDassert( old_max_cache_size > new_max_cache_size ); + + switch ( (cache_ptr->resize_ctl).decr_mode ) + { + case H5C2_decr__threshold: + HDassert( hit_rate > + (cache_ptr->resize_ctl).upper_hr_threshold ); + + HDfprintf(stdout, + "%sAuto cache resize -- decrease by threshold. HR = %lf > %6.5lf\n", + cache_ptr->prefix, hit_rate, + (cache_ptr->resize_ctl).upper_hr_threshold); + + HDfprintf(stdout, "%sout of bounds high (%6.5lf).\n", + cache_ptr->prefix, + (cache_ptr->resize_ctl).upper_hr_threshold); + break; + + case H5C2_decr__age_out: + HDfprintf(stdout, + "%sAuto cache resize -- decrease by ageout. HR = %lf\n", + cache_ptr->prefix, hit_rate); + break; + + case H5C2_decr__age_out_with_threshold: + HDassert( hit_rate > + (cache_ptr->resize_ctl).upper_hr_threshold ); + + HDfprintf(stdout, + "%sAuto cache resize -- decrease by ageout with threshold. HR = %lf > %6.5lf\n", + cache_ptr->prefix, hit_rate, + (cache_ptr->resize_ctl).upper_hr_threshold); + break; + + default: + HDfprintf(stdout, + "%sAuto cache resize -- decrease by unknown mode. HR = %lf\n", + cache_ptr->prefix, hit_rate); + } + + HDfprintf(stdout, + "%s cache size decreased from (%Zu/%Zu) to (%Zu/%Zu).\n", + cache_ptr->prefix, + old_max_cache_size, + old_min_clean_size, + new_max_cache_size, + new_min_clean_size); + break; + + case at_max_size2: + HDfprintf(stdout, + "%sAuto cache resize -- hit rate (%lf) out of bounds low (%6.5lf).\n", + cache_ptr->prefix, hit_rate, + (cache_ptr->resize_ctl).lower_hr_threshold); + HDfprintf(stdout, + "%s cache already at maximum size so no change.\n", + cache_ptr->prefix); + break; + + case at_min_size2: + HDfprintf(stdout, + "%sAuto cache resize -- hit rate (%lf) -- can't decrease.\n", + cache_ptr->prefix, hit_rate); + HDfprintf(stdout, "%s cache already at minimum size.\n", + cache_ptr->prefix); + break; + + case increase_disabled2: + HDfprintf(stdout, + "%sAuto cache resize -- increase disabled -- HR = %lf.", + cache_ptr->prefix, hit_rate); + break; + + case decrease_disabled2: + HDfprintf(stdout, + "%sAuto cache resize -- decrease disabled -- HR = %lf.\n", + cache_ptr->prefix, hit_rate); + break; + + case not_full2: + HDassert( hit_rate < (cache_ptr->resize_ctl).lower_hr_threshold ); + + HDfprintf(stdout, + "%sAuto cache resize -- hit rate (%lf) out of bounds low (%6.5lf).\n", + cache_ptr->prefix, hit_rate, + (cache_ptr->resize_ctl).lower_hr_threshold); + HDfprintf(stdout, + "%s cache not full so no increase in size.\n", + cache_ptr->prefix); + break; + + default: + HDfprintf(stdout, "%sAuto cache resize -- unknown status code.\n", + cache_ptr->prefix); + break; + } + + return; + +} /* H5C2_def_auto_resize_rpt_fcn() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_dest + * + * Purpose: Flush all data to disk and destroy the cache. + * + * This function fails if any object are protected since the + * resulting file might not be consistent. + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used on the first write occasioned + * by the destroy (primary_dxpl_id), and on all subsequent + * writes (secondary_dxpl_id). This is useful in the metadata + * cache, but may not be needed elsewhere. If so, just use the + * same dxpl_id for both parameters. + * + * Note that *cache_ptr has been freed upon successful return. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * JRM -- 7/11/07 + * Reworked parameter list for the revised cache API. + * The function lost its pointer to an instance of + * H5F_t (now supplied via cache_ptr), and one of its + * dxlp ids. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_dest(H5C2_t * cache_ptr, + hid_t dxpl_id) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_dest, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + if ( H5C2_flush_cache(cache_ptr, dxpl_id, + H5C2__FLUSH_INVALIDATE_FLAG) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "unable to flush cache") + } + + if ( cache_ptr->slist_ptr != NULL ) { + + H5SL_close(cache_ptr->slist_ptr); + cache_ptr->slist_ptr = NULL; + } + + cache_ptr->magic = 0; + + H5FL_FREE(H5C2_t, cache_ptr); + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_dest() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_dest_empty + * + * Purpose: Destroy an empty cache. + * + * This function fails if the cache is not empty on entry. + * + * Note that *cache_ptr has been freed upon successful return. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_dest_empty(H5C2_t * cache_ptr) +{ + herr_t ret_value=SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_dest_empty, FAIL) + + /* This would normally be an assert, but we need to use an HGOTO_ERROR + * call to shut up the compiler. + */ + if ( ( ! cache_ptr ) || + ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) || + ( cache_ptr->index_len != 0 ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Bad cache_ptr or non-empty cache on entry.") + } + + + if ( cache_ptr->slist_ptr != NULL ) { + + H5SL_close(cache_ptr->slist_ptr); + cache_ptr->slist_ptr = NULL; + } + + cache_ptr->magic = 0; + + H5FL_FREE(H5C2_t, cache_ptr); + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_dest_empty() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_expunge_entry + * + * Purpose: Use this function to tell the cache to expunge an entry + * from the cache without writing it to disk even if it is + * dirty. The entry may not be either pinned or protected. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/29/06 + * + * Modifications: + * + * JRM -- 7/11/07 + * Reworked the parameter list for the revised cache API. + * The function lost its file pointer (now passed in the + * *cache_ptr), and one of the dxpl ids. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_expunge_entry(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr) +{ + /* const char * fcn_name = "H5C2_expunge_entry()"; */ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr = NULL; + + FUNC_ENTER_NOAPI(H5C2_expunge_entry, FAIL) + + HDassert( H5F_addr_defined(addr) ); + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( type ); + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + H5C2__SEARCH_INDEX(cache_ptr, addr, entry_ptr, FAIL) + + if ( ( entry_ptr == NULL ) || ( entry_ptr->type != type ) ) { + + /* the target doesn't exist in the cache, so we are done. */ + HGOTO_DONE(SUCCEED) + } + + HDassert( entry_ptr->addr == addr ); + HDassert( entry_ptr->type == type ); + + if ( entry_ptr->is_protected ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTEXPUNGE, FAIL, \ + "Target entry is protected.") + } + + if ( entry_ptr->is_pinned ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTEXPUNGE, FAIL, \ + "Target entry is pinned.") + } + + /* If we get this far, call H5C2_flush_single_entry() with the + * H5C2__FLUSH_INVALIDATE_FLAG and the H5C2__FLUSH_CLEAR_ONLY_FLAG. + * This will clear the entry, and then delete it from the cache. + */ + + result = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__FLUSH_INVALIDATE_FLAG | + H5C2__FLUSH_CLEAR_ONLY_FLAG, + TRUE); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTEXPUNGE, FAIL, \ + "H5C2_flush_single_entry() failed.") + } + +done: + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_expunge_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_flush_cache + * + * Purpose: Flush (and possibly destroy) the entries contained in the + * specified cache. + * + * If the cache contains protected entries, the function will + * fail, as protected entries cannot be flushed. However + * all unprotected entries should be flushed before the + * function returns failure. + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used on the first write occasioned + * by the flush (primary_dxpl_id), and on all subsequent + * writes (secondary_dxpl_id). This is useful in the metadata + * cache, but may not be needed elsewhere. If so, just use the + * same dxpl_id for both parameters. + * + * Return: Non-negative on success/Negative on failure or if there was + * a request to flush all items and something was protected. + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * JRM -- 7/20/04 + * Modified the function for the addition of the hash table. + * + * JRM -- 11/22/04 + * Added code to remove all epoch markers (if any) from the + * LRU list before a destroy. Strictly speaking, this isn't + * necessary, as the marker entries reside only in the LRU + * list, never in the index or in the tree. However, it + * never hurts to tidy up. + * + * JRM -- 1/6/05 + * Reworked code to support the new + * H5C2__FLUSH_MARKED_ENTRIES_FLAG, and for the replacement of + * H5F_FLUSH_INVALIDATE flag with H5C2__FLUSH_INVALIDATE_FLAG. + * + * Note that the H5C2__FLUSH_INVALIDATE_FLAG takes precidence + * over the H5C2__FLUSH_MARKED_ENTRIES_FLAG. Thus if both are + * set, the functions behaves as if just the + * H5C2__FLUSH_INVALIDATE_FLAG was set. + * + * The H5C2__FLUSH_CLEAR_ONLY_FLAG flag can co-exist with + * either the H5C2__FLUSH_MARKED_ENTRIES_FLAG, or the + * H5C2__FLUSH_INVALIDATE_FLAG. In all cases, it is simply + * passed along to H5C2_flush_single_entry(). In the case of + * H5C2__FLUSH_MARKED_ENTRIES_FLAG, it will only apply to + * the marked entries. + * + * JRM -- 10/15/05 + * Added code supporting the new + * H5C2__FLUSH_IGNORE_PROTECTED_FLAG. We need this flag, as + * we now use this function to flush large number of entries + * in increasing address order. We do this by marking the + * entries to be flushed, calling this function to flush them, + * and then restoring LRU order. + * + * However, it is possible that the cache will contain other, + * unmarked protected entries, when we make this call. This + * new flag allows us to ignore them. + * + * Note that even with this flag set, it is still an error + * to try to flush a protected entry. + * + * JRM -- 3/25/06 + * Updated function to handle pinned entries. + * + * JRM -- 8/19/06 + * Added code managing the new flush_in_progress field of + * H5C2_t. + * + * Also reworked function to allow for the possibility that + * entries will be dirtied, resized, or renamed during flush + * callbacks. As a result, we may have to make multiple + * passes through the skip list before the cache is flushed. + * + * JRM -- 7/11/07 + * Reworked function to support the new metadata cache API. + * The function lost the H5F_t parameter (now passed via + * *cache_ptr), and one of the dxpl id parameters. + * + * JRM -- 10/13/07 + * Added code to detect and manage the case in which a + * serialize callback changes the s-list out from under + * the function. The only way I can think of in which this + * can happen is if a serialize function loads an entry + * into the cache that isn't there already. Quincey tells + * me that this will never happen, but I'm not sure I + * believe him. + * + * Note that this is a pretty bad scenario if it ever + * happens. The code I have added should allow us to + * handle the situation, but one can argue that I should + * just scream and die if I ever detect the condidtion. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_flush_cache(H5C2_t * cache_ptr, + hid_t dxpl_id, + unsigned flags) +{ + /* const char * fcn_name = "H5C2_flush_cache()"; */ + herr_t status; + herr_t ret_value = SUCCEED; + hbool_t destroy; + hbool_t flushed_entries_last_pass; + hbool_t flush_marked_entries; + hbool_t ignore_protected; + hbool_t tried_to_flush_protected_entry = FALSE; + int32_t passes = 0; + int32_t protected_entries = 0; + H5SL_node_t * node_ptr = NULL; + H5C2_cache_entry_t * entry_ptr = NULL; + H5C2_cache_entry_t * next_entry_ptr = NULL; +#if H5C2_DO_SANITY_CHECKS + int64_t flushed_entries_count; + size_t flushed_entries_size; + int64_t initial_slist_len; + size_t initial_slist_size; +#endif /* H5C2_DO_SANITY_CHECKS */ + + FUNC_ENTER_NOAPI(H5C2_flush_cache, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( cache_ptr->skip_file_checks || cache_ptr->f ); + HDassert( cache_ptr->slist_ptr ); + + ignore_protected = ( (flags & H5C2__FLUSH_IGNORE_PROTECTED_FLAG) != 0 ); + + destroy = ( (flags & H5C2__FLUSH_INVALIDATE_FLAG) != 0 ); + + /* note that flush_marked_entries is set to FALSE if destroy is TRUE */ + flush_marked_entries = ( ( (flags & H5C2__FLUSH_MARKED_ENTRIES_FLAG) != 0 ) + && + ( ! destroy ) + ); + + HDassert( ! ( destroy && ignore_protected ) ); + + HDassert( ! ( cache_ptr->flush_in_progress ) ); + + cache_ptr->flush_in_progress = TRUE; + + if ( destroy ) { + + status = H5C2_flush_invalidate_cache(dxpl_id, + cache_ptr, + flags); + + if ( status < 0 ) { + + /* This shouldn't happen -- if it does, we are toast so + * just scream and die. + */ + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "flush invalidate failed.") + } + } else { + /* When we are only flushing marked entries, the slist will usually + * still contain entries when we have flushed everything we should. + * Thus we track whether we have flushed any entries in the last + * pass, and terminate if we haven't. + */ + + flushed_entries_last_pass = TRUE; + + while ( ( passes < H5C2__MAX_PASSES_ON_FLUSH ) && + ( cache_ptr->slist_len != 0 ) && + ( protected_entries == 0 ) && + ( flushed_entries_last_pass ) ) + { + flushed_entries_last_pass = FALSE; + node_ptr = H5SL_first(cache_ptr->slist_ptr); + if ( node_ptr != NULL ) { + next_entry_ptr = (H5C2_cache_entry_t *)H5SL_item(node_ptr); + + if ( next_entry_ptr == NULL ) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "next_entry_ptr == NULL 1 ?!?!"); + } + HDassert( next_entry_ptr->magic == + H5C2__H5C2_CACHE_ENTRY_T_MAGIC ); + HDassert( next_entry_ptr->is_dirty ); + HDassert( next_entry_ptr->in_slist ); + } else { + next_entry_ptr = NULL; + } + + HDassert( node_ptr != NULL ); + +#if H5C2_DO_SANITY_CHECKS + /* For sanity checking, try to verify that the skip list has + * the expected size and number of entries at the end of each + * internal while loop (see below). + * + * Doing this get a bit tricky, as depending on flags, we may + * or may not flush all the entries in the slist. + * + * To make things more entertaining, with the advent of the + * fractal heap, the entry serialize callback can cause entries + * to be dirtied, resized, and/or renamed. + * + * To deal with this, we first make note of the initial + * skip list length and size: + */ + initial_slist_len = cache_ptr->slist_len; + initial_slist_size = cache_ptr->slist_size; + + /* We then zero counters that we use to track the number + * and total size of entries flushed: + */ + flushed_entries_count = 0; + flushed_entries_size = 0; + + /* As mentioned above, there is the possibility that + * entries will be dirtied, resized, and/or flushed during + * our pass through the skip list. To capture the number + * of entries added, and the skip list size delta, + * zero the slist_len_increase and slist_size_increase of + * the cache's instance of H5C2_t. These fields will be + * updated elsewhere to account for slist insertions and/or + * dirty entry size changes. + */ + cache_ptr->slist_len_increase = 0; + cache_ptr->slist_size_increase = 0; + + /* at the end of the loop, use these values to compute the + * expected slist length and size and compare this with the + * value recorded in the cache's instance of H5C2_t. + */ +#endif /* H5C2_DO_SANITY_CHECKS */ + + while ( node_ptr != NULL ) + { + entry_ptr = next_entry_ptr; + + /* With the advent of the fractal heap, it is possible + * that the serialize callback will dirty and/or resize + * other entries in the cache. In particular, while + * Quincey has promised me that this will never happen, + * it is possible that the serialize callback for an + * entry may protect an entry that is not in the cache, + * perhaps causing the cache to flush and possibly + * evict the entry associated with node_ptr to make + * space for the new entry. + * + * Thus we do a bit of extra sanity checking on entry_ptr, + * and break out of this scan of the skip list if we + * detect minor problems. We have a bit of leaway on the + * number of passes though the skip list, so this shouldn't + * be an issue in the flush in and of itself, as it should + * be all but impossible for this to happen more than once + * in any flush. + * + * Observe that that breaking out of the scan early + * shouldn't break the sanity checks just after the end + * of this while loop. + * + * If an entry has merely been marked clean and removed from + * the s-list, we simply break out of the scan. + * + * If the entry has been evicted, we flag an error and + * exit. + */ + + if ( entry_ptr->magic != H5C2__H5C2_CACHE_ENTRY_T_MAGIC ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "entry_ptr->magic invalid ?!?!"); + + } else if ( ( ! entry_ptr->is_dirty ) || + ( ! entry_ptr->in_slist ) ) { + + /* the s-list has been modified out from under us. + * set node_ptr to NULL and break out of the loop. + */ + node_ptr = NULL; + break; + } + + /* increment node pointer now, before we delete its target + * from the slist. + */ + node_ptr = H5SL_next(node_ptr); + if ( node_ptr != NULL ) { + next_entry_ptr = (H5C2_cache_entry_t *)H5SL_item(node_ptr); + + if ( next_entry_ptr == NULL ) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "next_entry_ptr == NULL 2 ?!?!"); + } + HDassert( next_entry_ptr->magic == + H5C2__H5C2_CACHE_ENTRY_T_MAGIC ); + HDassert( next_entry_ptr->is_dirty ); + HDassert( next_entry_ptr->in_slist ); + } else { + next_entry_ptr = NULL; + } + + HDassert( entry_ptr != NULL ); + HDassert( entry_ptr->in_slist ); + + if ( ( ! flush_marked_entries ) || + ( entry_ptr->flush_marker ) ) { + + if ( entry_ptr->is_protected ) { + + /* we probably have major problems -- but lets flush + * everything we can before we decide whether to flag + * an error. + */ + tried_to_flush_protected_entry = TRUE; + protected_entries++; + + } else if ( entry_ptr->is_pinned ) { + + /* Test to see if we are can flush the entry now. + * If we can, go ahead and flush. Note that we + * aren't trying to do a destroy here, so that + * is not an issue. + */ + if ( TRUE ) { /* When we get to multithreaded cache, + * we will need either locking code, + * and/or a test to see if the entry + * is in flushable condition here. + */ +#if H5C2_DO_SANITY_CHECKS + flushed_entries_count++; + flushed_entries_size += entry_ptr->size; +#endif /* H5C2_DO_SANITY_CHECKS */ + status = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + flags, + FALSE); + if ( status < 0 ) { + + /* This shouldn't happen -- if it does, we + * are toast so just scream and die. + */ + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "dirty pinned entry flush failed.") + } + } + } else { +#if H5C2_DO_SANITY_CHECKS + flushed_entries_count++; + flushed_entries_size += entry_ptr->size; +#endif /* H5C2_DO_SANITY_CHECKS */ + status = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + flags, + FALSE); + if ( status < 0 ) { + + /* This shouldn't happen -- if it does, we are + * toast so just scream and die. + */ + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "Can't flush entry.") + } + flushed_entries_last_pass = TRUE; + } + } + } /* while ( node_ptr != NULL ) */ + +#if H5C2_DO_SANITY_CHECKS + /* Verify that the slist size and length are as expected. */ + + HDassert( (initial_slist_len + cache_ptr->slist_len_increase - + flushed_entries_count) == cache_ptr->slist_len ); + HDassert( (initial_slist_size + cache_ptr->slist_size_increase - + flushed_entries_size) == cache_ptr->slist_size ); +#endif /* H5C2_DO_SANITY_CHECKS */ + + passes++; + + } /* while */ + + HDassert( protected_entries <= cache_ptr->pl_len ); + + if ( ( ( cache_ptr->pl_len > 0 ) && ( !ignore_protected ) ) + || + ( tried_to_flush_protected_entry ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "cache has protected items") + } + + if ( ( cache_ptr->slist_len != 0 ) && + ( passes >= H5C2__MAX_PASSES_ON_FLUSH ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "flush pass limit exceeded.") + } + +#if H5C2_DO_SANITY_CHECKS + if ( ! flush_marked_entries ) { + + HDassert( cache_ptr->slist_len == 0 ); + HDassert( cache_ptr->slist_size == 0 ); + } +#endif /* H5C2_DO_SANITY_CHECKS */ + + } + +done: + + cache_ptr->flush_in_progress = FALSE; + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_flush_cache() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_flush_to_min_clean + * + * Purpose: Flush dirty entries until the caches min clean size is + * attained. + * + * This function is used in the implementation of the + * metadata cache in PHDF5. To avoid "messages from the + * future", the cache on process 0 can't be allowed to + * flush entries until the other processes have reached + * the same point in the calculation. If this constraint + * is not met, it is possible that the other processes will + * read metadata generated at a future point in the + * computation. + * + * + * Return: Non-negative on success/Negative on failure or if + * write is not permitted. + * + * Programmer: John Mainzer + * 9/16/05 + * + * Modifications: + * + * Re-wrote function to flush dirty entries in increasing + * address order, while maintaining LRU order in the LRU list + * upon return. + * + * Do this by scanning up the dirty LRU list for entries to + * flush to reach min clean size, setting their flush_marker + * flags, and recording their addresses in the order + * encountered. + * + * Then call H5C2_flush_cache() to flush the marked entries. + * + * Finally, use the list of marked entries to force the + * correct LRU list order after the flush. + * + * JRM - 10/13/05 + * + * This change had the oposite of the desired effect. Lets + * leave it in (albeit commented out for now). If we can't + * find a case where it helps, lets get rid of it. + * + * + * Added some sanity checks to the change which verify the + * expected values of the new is_read_only and ro_ref_count + * fields. + * JRM - 3/29/07 + * + * Modified parameter list for the new metadata cache API. + * THe function lost its H5F_t parameter (now passed via + * *cache_ptr), and one of its dxpl ids. + * + * JRM - 7/11/07 + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_flush_to_min_clean(H5C2_t * cache_ptr, + hid_t dxpl_id) +{ + herr_t result; + herr_t ret_value = SUCCEED; + hbool_t write_permitted; +#if 0 /* modified code -- commented out for now */ + int i; + int flushed_entries_count = 0; + size_t flushed_entries_size = 0; + size_t space_needed = 0; + haddr_t * flushed_entries_list = NULL; + H5C2_cache_entry_t * entry_ptr = NULL; +#endif /* JRM */ + + FUNC_ENTER_NOAPI(H5C2_flush_to_min_clean, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( cache_ptr->skip_file_checks || cache_ptr->f ); + + if ( cache_ptr->check_write_permitted != NULL ) { + + result = (cache_ptr->check_write_permitted)(cache_ptr->f, + dxpl_id, + &write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Can't get write_permitted") + } + } else { + + write_permitted = cache_ptr->write_permitted; + } + + if ( ! write_permitted ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "cache write is not permitted!?!\n"); + } +#if 1 /* original code */ + result = H5C2_make_space_in_cache(dxpl_id, + cache_ptr, + (size_t)0, + write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_make_space_in_cache failed.") + } +#else /* modified code -- commented out for now */ + 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); + } + } 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; + } + } + + if ( space_needed > 0 ) { /* we have work to do */ + + HDassert( cache_ptr->slist_len > 0 ); + + /* allocate an array to keep a list of the entries that we + * mark for flush. We need this list to touch up the LRU + * list after the flush. + */ + flushed_entries_list = (haddr_t *)H5MM_malloc(sizeof(haddr_t) * + (size_t)(cache_ptr->slist_len)); + + if ( flushed_entries_list == NULL ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "memory allocation failed for flushed entries list") + } + + /* Scan the dirty LRU list from tail forward and mark sufficient + * entries to free up the necessary space. Keep a list of the + * entries marked in the order in which they are encountered. + */ + entry_ptr = cache_ptr->dLRU_tail_ptr; + + while ( ( flushed_entries_size < space_needed ) && + ( flushed_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 ); + + entry_ptr->flush_marker = TRUE; + flushed_entries_size += entry_ptr->size; + flushed_entries_list[flushed_entries_count] = entry_ptr->addr; + flushed_entries_count++; + entry_ptr = entry_ptr->aux_prev; + } + + if ( ( flushed_entries_count > cache_ptr->slist_len) || + ( flushed_entries_size < space_needed ) ) { + HDfprintf(stdout, "flushed_entries_count = %d <= %d = slist_size\n", + (int)flushed_entries_count, (int)(cache_ptr->slist_size)); + HDfprintf(stdout, + "flushed_entries_size = %d < %d = space_needed.\n", + (int)flushed_entries_size, (int)space_needed); + } + + HDassert( flushed_entries_count <= cache_ptr->slist_len ); + HDassert( flushed_entries_size >= space_needed ); + + + /* Flush the marked entries */ + result = H5C2_flush_cache(f, primary_dxpl_id, secondary_dxpl_id, + cache_ptr, H5C2__FLUSH_MARKED_ENTRIES_FLAG | + H5C2__FLUSH_IGNORE_PROTECTED_FLAG); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "H5C2_flush_cache failed.") + } + + /* Now touch up the LRU list so as to place the flushed entries in + * the order they they would be in if we had flushed them in the + * order we encountered them in. + */ + + i = 0; + while ( i < flushed_entries_count ) + { + H5C2__SEARCH_INDEX_NO_STATS(cache_ptr, flushed_entries_list[i], \ + entry_ptr, FAIL) + + /* At present, the above search must always succeed. However, + * that may change. Write the code so we need only remove the + * following assert in that event. + */ + HDassert( entry_ptr != NULL ); + H5C2__FAKE_RP_FOR_MOST_RECENT_ACCESS(cache_ptr, entry_ptr, FAIL) + i++; + } + } /* if ( space_needed > 0 ) */ +#endif /* end modified code -- commented out for now */ + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_flush_to_min_clean() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_get_cache_auto_resize_config + * + * Purpose: Copy the current configuration of the cache automatic + * re-sizing function into the instance of H5C2_auto_size_ctl_t + * pointed to by config_ptr. + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 10/8/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_get_cache_auto_resize_config(H5C2_t * cache_ptr, + H5C2_auto_size_ctl_t *config_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_get_cache_auto_resize_config, FAIL) + + if ( ( cache_ptr == NULL ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + if ( config_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad config_ptr on entry.") + } + + *config_ptr = cache_ptr->resize_ctl; + + config_ptr->set_initial_size = FALSE; + config_ptr->initial_size = cache_ptr->max_cache_size; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_get_cache_auto_resize_config() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_get_cache_size + * + * Purpose: Return the cache maximum size, the minimum clean size, the + * current size, and the current number of entries in + * *max_size_ptr, *min_clean_size_ptr, *cur_size_ptr, and + * *cur_num_entries_ptr respectively. If any of these + * parameters are NULL, skip that value. + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 10/8/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_get_cache_size(H5C2_t * cache_ptr, + size_t * max_size_ptr, + size_t * min_clean_size_ptr, + size_t * cur_size_ptr, + int32_t * cur_num_entries_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_get_cache_size, FAIL) + + if ( ( cache_ptr == NULL ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + if ( max_size_ptr != NULL ) { + + *max_size_ptr = cache_ptr->max_cache_size; + } + + if ( min_clean_size_ptr != NULL ) { + + *min_clean_size_ptr = cache_ptr->min_clean_size; + } + + if ( cur_size_ptr != NULL ) { + + *cur_size_ptr = cache_ptr->index_size; + } + + if ( cur_num_entries_ptr != NULL ) { + + *cur_num_entries_ptr = cache_ptr->index_len; + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_get_cache_size() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_get_cache_hit_rate + * + * Purpose: Compute and return the current cache hit rate in + * *hit_rate_ptr. If there have been no accesses since the + * last time the cache hit rate stats were reset, set + * *hit_rate_ptr to 0.0. On error, *hit_rate_ptr is + * undefined. + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 10/7/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_get_cache_hit_rate(H5C2_t * cache_ptr, + double * hit_rate_ptr) + +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_get_cache_hit_rate, FAIL) + + if ( ( cache_ptr == NULL ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + if ( hit_rate_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad hit_rate_ptr on entry.") + } + + HDassert( cache_ptr->cache_hits >= 0 ); + HDassert( cache_ptr->cache_accesses >= cache_ptr->cache_hits ); + + if ( cache_ptr->cache_accesses > 0 ) { + + *hit_rate_ptr = ((double)(cache_ptr->cache_hits)) / + ((double)(cache_ptr->cache_accesses)); + + } else { + + *hit_rate_ptr = 0.0; + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_get_cache_hit_rate() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_get_entry_status + * + * Purpose: This function is used to determine whether the cache + * contains an entry with the specified base address. If + * the entry exists, it also reports some status information + * on the entry. + * + * Status information is reported in the locations pointed + * to by the size_ptr, in_cache_ptr, is_dirty_ptr, and + * is_protected_ptr. While in_cache_ptr must be defined, + * the remaining pointers may be NULL, in which case the + * associated data is not reported. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 7/1/05 + * + * Modifications: + * + * JRM -- 4/26/06 + * Added the is_pinned_ptr parameter and supporting code. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_get_entry_status(H5C2_t * cache_ptr, + haddr_t addr, + size_t * size_ptr, + hbool_t * in_cache_ptr, + hbool_t * is_dirty_ptr, + hbool_t * is_protected_ptr, + hbool_t * is_pinned_ptr) +{ + /* const char * fcn_name = "H5C2_get_entry_status()"; */ + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr = NULL; + + FUNC_ENTER_NOAPI(H5C2_get_entry_status, FAIL) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( H5F_addr_defined(addr) ); + HDassert( in_cache_ptr != NULL ); + + /* this test duplicates tow of the above asserts, but we need an + * invocation of HGOTO_ERROR to keep the compiler happy. + */ + if ( ( cache_ptr == NULL ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + H5C2__SEARCH_INDEX(cache_ptr, addr, entry_ptr, FAIL) + + if ( entry_ptr == NULL ) { + + /* the entry doesn't exist in the cache -- report this + * and quit. + */ + *in_cache_ptr = FALSE; + + } else { + + *in_cache_ptr = TRUE; + + if ( size_ptr != NULL ) { + + *size_ptr = entry_ptr->size; + } + + if ( is_dirty_ptr != NULL ) { + + *is_dirty_ptr = entry_ptr->is_dirty; + } + + if ( is_protected_ptr != NULL ) { + + *is_protected_ptr = entry_ptr->is_protected; + } + + if ( is_pinned_ptr != NULL ) { + + *is_pinned_ptr = entry_ptr->is_pinned; + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_get_entry_status() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_get_evictions_enabled() + * + * Purpose: Copy the current value of cache_ptr->evictions_enabled into + * *evictions_enabled_ptr. + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 7/27/07 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_get_evictions_enabled(H5C2_t * cache_ptr, + hbool_t * evictions_enabled_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_get_evictions_enabled, FAIL) + + if ( ( cache_ptr == NULL ) || + ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + if ( evictions_enabled_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Bad evictions_enabled_ptr on entry.") + } + + *evictions_enabled_ptr = cache_ptr->evictions_enabled; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_get_evictions_enabled() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_get_trace_file_ptr + * + * Purpose: Get the trace_file_ptr field from the cache. + * + * This field will either be NULL (which indicates that trace + * file logging is turned off), or contain a pointer to the + * open file to which trace file data is to be written. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 1/20/06 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_get_trace_file_ptr(H5C2_t * cache_ptr, + FILE ** trace_file_ptr_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_get_trace_file_ptr, FAIL) + + /* This would normally be an assert, but we need to use an HGOTO_ERROR + * call to shut up the compiler. + */ + if ( ( ! cache_ptr ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr") + } + + if ( trace_file_ptr_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "NULL trace_file_ptr_ptr") + } + + *trace_file_ptr_ptr = cache_ptr->trace_file_ptr; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_get_trace_file_ptr() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_insert_entry + * + * Purpose: Adds the specified thing to the cache. The thing need not + * exist on disk yet, but it must have an address and disk + * space reserved. + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used on the first write occasioned + * by the insertion (primary_dxpl_id), and on all subsequent + * writes (secondary_dxpl_id). This is useful in the + * metadata cache, but may not be needed elsewhere. If so, + * just use the same dxpl_id for both parameters. + * + * The primary_dxpl_id is the dxpl_id passed to the + * check_write_permitted function if such a function has been + * provided. + * + * Observe that this function cannot occasion a read. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated function for the addition of the hash table. + * + * JRM -- 10/28/04 + * Added code to set the cache_full flag to TRUE when ever + * we need to make space in the cache. + * + * JRM -- 11/22/04 + * Updated function for the addition of the first_flush_ptr + * parameter to H5C2_make_space_in_cache(). + * + * JRM -- 1/6/05 + * Added the flags parameter, and code supporting + * H5C2__SET_FLUSH_MARKER_FLAG. Note that this flag is + * ignored unless the new entry is dirty. + * + * JRM -- 6/6/05 + * Added code to force all inserted entries to be dirty. + * This is part of a set of changes moving management of the + * is_dirty field of H5C2_cache_entry_t into the H5C2 code. + * + * JRM -- 6/24/05 + * Added support for the new write_permitted field of + * the H5C2_t structure. + * + * JRM -- 3/16/06 + * Added initialization for the new is_pinned field of the + * H5C2_cache_entry_t structure. + * + * JRM -- 5/3/06 + * Added initialization for the new dirtied field of the + * H5C2_cache_entry_t structure. + * + * JRM -- 8/9/06 + * Added code supporting insertion of pinned entries. + * + * JRM -- 8/21/06 + * Added initialization for the new flush_in_progress and + * destroy_in_progress fields. + * + * JRM -- 3/29/07 + * Added initialization for the new is_read_only and + * ro_ref_count fields. + * + * JRM -- 7/11/07 + * Reworked the parameter list for the revised metadata + * cache API. The function lost its pointer to H5F_t + * (now supplied via *cache_ptr), and one of dxpl id + * parameters. It gained a entry length parameter. + * Numerous internal changes to support the API change. + * + * JRM -- 8/1/07 + * Added code to disable evictions when the new + * evictions_enabled field is FALSE. + * + * JRM -- 10/12/07 + * Added initialization for the new magic field. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_insert_entry(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + size_t len, + void * thing, + unsigned int flags) +{ + herr_t result; + herr_t ret_value = SUCCEED; /* Return value */ + hbool_t insert_pinned; + hbool_t set_flush_marker; + hbool_t write_permitted = TRUE; + H5C2_cache_entry_t * entry_ptr; + H5C2_cache_entry_t * test_entry_ptr; + + FUNC_ENTER_NOAPI(H5C2_insert_entry, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( cache_ptr->skip_file_checks || cache_ptr->f ); + HDassert( type ); + HDassert( H5F_addr_defined(addr) ); + HDassert( thing ); + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_verify_not_in_index(cache_ptr, (H5C2_cache_entry_t *)thing) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "thing already in index.\n"); + } +#endif /* H5C2_DO_SANITY_CHECKS */ + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + set_flush_marker = ( (flags & H5C2__SET_FLUSH_MARKER_FLAG) != 0 ); + insert_pinned = ( (flags & H5C2__PIN_ENTRY_FLAG) != 0 ); + + entry_ptr = (H5C2_cache_entry_t *)thing; + + entry_ptr->magic = H5C2__H5C2_CACHE_ENTRY_T_MAGIC; + + entry_ptr->addr = addr; + entry_ptr->type = type; + + /* newly inserted entries are assumed to be dirty */ + entry_ptr->is_dirty = TRUE; + + /* not protected, so can't be dirtied */ + entry_ptr->dirtied = FALSE; + + entry_ptr->size = len; + + HDassert( entry_ptr->size < H5C2_MAX_ENTRY_SIZE ); + + entry_ptr->in_slist = FALSE; + +#ifdef H5_HAVE_PARALLEL + entry_ptr->clear_on_unprotect = FALSE; +#endif /* H5_HAVE_PARALLEL */ + + entry_ptr->flush_in_progress = FALSE; + entry_ptr->destroy_in_progress = FALSE; + + entry_ptr->ht_next = NULL; + entry_ptr->ht_prev = NULL; + + entry_ptr->next = NULL; + entry_ptr->prev = NULL; + + entry_ptr->aux_next = NULL; + entry_ptr->aux_prev = NULL; + + H5C2__RESET_CACHE_ENTRY_STATS(entry_ptr) + + if ( ( cache_ptr->evictions_enabled ) && + ( (cache_ptr->index_size + entry_ptr->size) > + cache_ptr->max_cache_size ) ) { + + size_t space_needed; + + cache_ptr->cache_full = TRUE; + + if ( cache_ptr->check_write_permitted != NULL ) { + + result = (cache_ptr->check_write_permitted)(cache_ptr->f, + dxpl_id, + &write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTINS, FAIL, \ + "Can't get write_permitted") + } + } else { + + write_permitted = cache_ptr->write_permitted; + } + + HDassert( entry_ptr->size <= H5C2_MAX_ENTRY_SIZE ); + + space_needed = entry_ptr->size; + + if ( space_needed > cache_ptr->max_cache_size ) { + + space_needed = cache_ptr->max_cache_size; + } + + /* Note that space_needed is just the amount of space that + * needed to insert the new entry without exceeding the cache + * size limit. The subsequent call to H5C2_make_space_in_cache() + * may evict the entries required to free more or less space + * depending on conditions. It MAY be less if the cache is + * currently undersized, or more if the cache is oversized. + * + * The cache can exceed its maximum size limit via the following + * mechanisms: + * + * First, it is possible for the cache to grow without + * bound as long as entries are protected and not unprotected. + * + * Second, when writes are not permitted it is also possible + * for the cache to grow without bound. + * + * Finally, we usually don't check to see if the cache is + * oversized at the end of an unprotect. As a result, it is + * possible to have a vastly oversized cache with no protected + * entries as long as all the protects preceed the unprotects. + * + * Since items 1 and 2 are not changing any time soon, I see + * no point in worrying about the third. + */ + + result = H5C2_make_space_in_cache(dxpl_id, + cache_ptr, + space_needed, + write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTINS, FAIL, \ + "H5C2_make_space_in_cache failed.") + } + } + + /* verify that the new entry isn't already in the hash table -- scream + * and die if it is. + */ + + H5C2__SEARCH_INDEX(cache_ptr, addr, test_entry_ptr, FAIL) + + if ( test_entry_ptr != NULL ) { + + if ( test_entry_ptr == entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTINS, FAIL, \ + "entry already in cache.") + + } else { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTINS, FAIL, \ + "duplicate entry in cache.") + } + } + + /* we don't initialize the protected field until here as it is + * possible that the entry is already in the cache, and already + * protected. If it is, we don't want to make things worse by + * marking it unprotected. + */ + + entry_ptr->is_protected = FALSE; + entry_ptr->is_read_only = FALSE; + entry_ptr->ro_ref_count = 0; + + entry_ptr->is_pinned = insert_pinned; + + H5C2__INSERT_IN_INDEX(cache_ptr, entry_ptr, FAIL) + + /* New entries are presumed to be dirty, so this if statement is + * unnecessary. Rework it once the rest of the code changes are + * in and tested. -- JRM + */ + if ( entry_ptr->is_dirty ) { + + entry_ptr->flush_marker = set_flush_marker; + H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, FAIL) + + } else { + + entry_ptr->flush_marker = FALSE; + } + + H5C2__UPDATE_RP_FOR_INSERTION(cache_ptr, entry_ptr, FAIL) + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + H5C2__UPDATE_STATS_FOR_INSERTION(cache_ptr, entry_ptr) + +done: + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_insert_entry() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_mark_entries_as_clean + * + * Purpose: When the H5C2 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 H5C2 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 + * H5C2__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 + * + * Modifications: + * + * Reworked function to flush entries in LRU order instead + * of increasing address order. The hope is that this will + * improve the hit rate on the slave caches. + * + * JRM - 10/13/05 + * + * Leave the old code in place for now (commented out) for + * benchmarking. + * + * JRM -- 4/13/06 + * Updated function to deal with pinned entries. + * + * JRM -- 7/11/07 + * Revised function for the new metadata cache API. The + * function lost its point to H5F_t (now supplied via + * *cache_ptr), and one of its dxpl ids. Also internal + * changes supporting the revised API. + * + * JRM -- 10/13/07 + * Didn't modify this function to detect the case in which + * the LRU is modified out from under it. It shouldn't be + * an issue here, as this function is only called in the + * parallel case, and serialize callbacks must not modify + * other entries in parallel case. If they do, they will + * cause inconsistancies in metadata across processes. + * + * Further, since this function only clears entries, and + * thus the serialize functions are never called, the + * situation will be even worse, as the changes will only + * exist on process 0. + * + * Bottom line -- the calls to H5C2_flush_single_entry() + * in this function will not call serialize, thus no change + * in the LRU is possible. Even if they did, the serialize() + * callbacks are banned from making such changes in the + * parallel case. + * + *------------------------------------------------------------------------- + */ + +#ifdef H5_HAVE_PARALLEL +herr_t +H5C2_mark_entries_as_clean(H5C2_t * cache_ptr, + hid_t dxpl_id, + int32_t ce_array_len, + haddr_t * ce_array_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + int entries_cleared; + int entries_examined; + int i; + int initial_list_len; + haddr_t addr; +#if H5C2_DO_SANITY_CHECKS + int pinned_entries_marked = 0; + int protected_entries_marked = 0; + int other_entries_marked = 0; + haddr_t last_addr; +#endif /* H5C2_DO_SANITY_CHECKS */ + H5C2_cache_entry_t * clear_ptr = NULL; + H5C2_cache_entry_t * entry_ptr = NULL; + + FUNC_ENTER_NOAPI(H5C2_mark_entries_as_clean, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + HDassert( ce_array_len > 0 ); + HDassert( ce_array_ptr != NULL ); + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HDassert(0); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + for ( i = 0; i < ce_array_len; i++ ) + { + addr = ce_array_ptr[i]; + +#if H5C2_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 H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HDassert(0); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ +#endif /* H5C2_DO_SANITY_CHECKS */ + + HDassert( H5F_addr_defined(addr) ); + + H5C2__SEARCH_INDEX(cache_ptr, addr, entry_ptr, FAIL) + + if ( entry_ptr == NULL ) { +#if H5C2_DO_SANITY_CHECKS + HDfprintf(stdout, + "H5C2_mark_entries_as_clean: entry[%d] = %ld not in cache.\n", + (int)i, + (long)addr); +#endif /* H5C2_DO_SANITY_CHECKS */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Listed entry not in cache?!?!?.") + + } else if ( ! entry_ptr->is_dirty ) { + +#if H5C2_DO_SANITY_CHECKS + HDfprintf(stdout, + "H5C2_mark_entries_as_clean: entry %ld is not dirty!?!\n", + (long)addr); +#endif /* H5C2_DO_SANITY_CHECKS */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Listed entry not dirty?!?!?.") +#if 0 /* original code */ + } else if ( entry_ptr->is_protected ) { + + entry_ptr->clear_on_unprotect = TRUE; + + } else { + + if ( H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + addr, + H5C2__FLUSH_CLEAR_ONLY_FLAG, + TRUE) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } + } +#else /* modified code */ + } 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 H5C2_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 /* H5C2_DO_SANITY_CHECKS */ + } +#endif /* end modified code */ + } +#if 1 /* modified code */ + /* 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. + */ + + 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 ( H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + clear_ptr->type, + clear_ptr->addr, + H5C2__FLUSH_CLEAR_ONLY_FLAG, + TRUE) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } + } else { + + entry_ptr = entry_ptr->prev; + } + entries_examined++; + } + +#if H5C2_DO_SANITY_CHECKS + HDassert( entries_cleared == other_entries_marked ); +#endif /* H5C2_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 ( H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + clear_ptr->type, + clear_ptr->addr, + H5C2__FLUSH_CLEAR_ONLY_FLAG, + TRUE) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't clear entry.") + } + } else { + + entry_ptr = entry_ptr->next; + } + } + +#if H5C2_DO_SANITY_CHECKS + HDassert( entries_cleared == pinned_entries_marked + other_entries_marked ); + HDassert( entries_cleared + protected_entries_marked == ce_array_len ); +#endif /* H5C2_DO_SANITY_CHECKS */ + + HDassert( ( entries_cleared == ce_array_len ) || + ( (ce_array_len - entries_cleared) <= cache_ptr->pl_len ) ); + +#if H5C2_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 /* H5C2_DO_SANITY_CHECKS */ +#endif /* modified code */ + +done: + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HDassert(0); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_mark_entries_as_clean() */ +#endif /* H5_HAVE_PARALLEL */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_mark_pinned_entry_dirty + * + * Purpose: Mark a pinned entry as dirty. The target entry MUST be + * be pinned, and MUST be unprotected. + * + * If the entry has changed size, the function updates + * data structures for the size change. + * + * If the entry is not already dirty, the function places + * the entry on the skip list. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 3/22/06 + * + * Modifications: + * + * None + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_mark_pinned_entry_dirty(H5C2_t * cache_ptr, + void * thing, + hbool_t size_changed, + size_t new_size) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr; + + FUNC_ENTER_NOAPI(H5C2_mark_pinned_entry_dirty, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + HDassert( ( size_changed == TRUE ) || ( size_changed == FALSE ) ); + + entry_ptr = (H5C2_cache_entry_t *)thing; + + if ( ! ( entry_ptr->is_pinned ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "Entry isn't pinned??") + } + + if ( entry_ptr->is_protected ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "Entry is protected??") + } + + /* mark the entry as dirty if it isn't already */ + entry_ptr->is_dirty = TRUE; + + /* update for change in entry size if necessary */ + if ( ( size_changed ) && ( entry_ptr->size != new_size ) ) { + + /* update the protected entry list */ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr->pel_len), \ + (cache_ptr->pel_size), \ + (entry_ptr->size), (new_size)); + + /* update the hash table */ + H5C2__UPDATE_INDEX_FOR_SIZE_CHANGE((cache_ptr), (entry_ptr->size),\ + (new_size)); + + /* if the entry is in the skip list, update that too */ + if ( entry_ptr->in_slist ) { + + H5C2__UPDATE_SLIST_FOR_SIZE_CHANGE((cache_ptr), (entry_ptr->size),\ + (new_size)); + } + + /* update statistics just before changing the entry size */ + H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE((cache_ptr), (entry_ptr), \ + (new_size)); + + /* finally, update the entry size proper */ + entry_ptr->size = new_size; + } + + if ( ! (entry_ptr->in_slist) ) { + + H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, FAIL) + } + + H5C2__UPDATE_STATS_FOR_DIRTY_PIN(cache_ptr, entry_ptr) + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_mark_pinned_entry_dirty() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_mark_pinned_or_protected_entry_dirty + * + * Purpose: Mark a pinned or protected entry as dirty. The target entry + * MUST be either pinned or protected, and MAY be both. + * + * At present, this funtion does not support size change. + * + * In the protected case, this call is the functional + * equivalent of setting the H5C2__DIRTIED_FLAG on an unprotect + * call. + * + * In the pinned but not protected case, if the entry is not + * already dirty, the function places function marks the entry + * dirty and places it on the skip list. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 5/15/06 + * + * Modifications: + * + * JRM -- 3/29/07 + * Added sanity check to verify that the pinned entry + * is not protected read only. + * + * This sanity check is commented out for now -- uncomment + * it once we deal with the problem of entries being protected + * read only, and then dirtied. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_mark_pinned_or_protected_entry_dirty(H5C2_t * cache_ptr, + void * thing) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr; + + FUNC_ENTER_NOAPI(H5C2_mark_pinned_or_protected_entry_dirty, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + + entry_ptr = (H5C2_cache_entry_t *)thing; + + if ( entry_ptr->is_protected ) { +#if 1 /* JRM - uncomment this when possible */ + HDassert( ! ((entry_ptr)->is_read_only) ); +#endif + /* set the dirtied flag */ + entry_ptr->dirtied = TRUE; + + } else if ( entry_ptr->is_pinned ) { + + /* mark the entry as dirty if it isn't already */ + entry_ptr->is_dirty = TRUE; + + + if ( ! (entry_ptr->in_slist) ) { + + H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, FAIL) + } + + H5C2__UPDATE_STATS_FOR_DIRTY_PIN(cache_ptr, entry_ptr) + + } else { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTMARKDIRTY, FAIL, \ + "Entry is neither pinned nor protected??") + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_mark_pinned_or_protected_entry_dirty() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_rename_entry + * + * Purpose: Use this function to notify the cache that an entry's + * file address changed. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated function for the addition of the hash table. + * + * JRM -- 6/6/05 + * Updated function to force all renamed entries to be + * dirty. This is part of a series of code modifications + * moving management of the is_dirty field of + * H5C2_cache_entry_t into the H5C2 code. + * + * JRM -- 4/3/06 + * Updated function to disallow renaming of pinned entries. + * + * JRM -- 4/27/06 + * Updated function to support renaming of pinned entries. + * + * JRM -- 8/24/06 + * Updated function to refrain from alterning the index, the + * replacement policy data structures, and skip list when + * the function is called within the flush callback for the + * target entry and the target entry is being destroyed. + * + * Note that in this case H5C2_flush_single_entry() will handle + * all these details for us. + * + * JRM -- 10/13/07 + * Renames of the target entry in a serialize callback are + * now handled by H5C2_flush_single_entry() -- hence the above + * modification is now obsolete. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_rename_entry(H5C2_t * cache_ptr, + const H5C2_class_t * type, + haddr_t old_addr, + haddr_t new_addr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + hbool_t was_dirty; + H5C2_cache_entry_t * entry_ptr = NULL; + H5C2_cache_entry_t * test_entry_ptr = NULL; +#if H5C2_DO_SANITY_CHECKS + hbool_t removed_entry_from_slist = FALSE; +#endif /* H5C2_DO_SANITY_CHECKS */ + + FUNC_ENTER_NOAPI(H5C2_rename_entry, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( type ); + HDassert( H5F_addr_defined(old_addr) ); + HDassert( H5F_addr_defined(new_addr) ); + HDassert( H5F_addr_ne(old_addr, new_addr) ); + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + H5C2__SEARCH_INDEX(cache_ptr, old_addr, entry_ptr, FAIL) + + if ( ( entry_ptr == NULL ) || ( entry_ptr->type != type ) ) { + + /* the old item doesn't exist in the cache, so we are done. */ + HGOTO_DONE(SUCCEED) + } + + HDassert( entry_ptr->addr == old_addr ); + HDassert( entry_ptr->type == type ); + + if ( entry_ptr->is_protected ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRENAME, FAIL, \ + "Target entry is protected.") + } + + H5C2__SEARCH_INDEX(cache_ptr, new_addr, test_entry_ptr, FAIL) + + if ( test_entry_ptr != NULL ) { /* we are hosed */ + + if ( test_entry_ptr->type == type ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRENAME, FAIL, \ + "Target already renamed & reinserted???.") + + } else { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRENAME, FAIL, \ + "New address already in use?.") + + } + } + + /* If we get this far we have work to do. Remove *entry_ptr from + * the hash table (and skip list if necessary), change its address to the + * new address, mark it as dirty (if it isn't already) and then re-insert. + * + * Update the replacement policy for a hit to avoid an eviction before + * the renamed entry is touched. Update stats for a rename. + * + * Note that we do not check the size of the cache, or evict anything. + * Since this is a simple re-name, cache size should be unaffected. + * + * Check to see if the target entry is in the process of being destroyed + * before we delete from the index, etc. If it is, all we do is + * change the addr. If the entry is only in the process of being flushed, + * don't mark it as dirty either, lest we confuse the flush call back. + */ + + if ( ! ( entry_ptr->destroy_in_progress ) ) { + + H5C2__DELETE_FROM_INDEX(cache_ptr, entry_ptr) + + if ( entry_ptr->in_slist ) { + + HDassert( cache_ptr->slist_ptr ); + + H5C2__REMOVE_ENTRY_FROM_SLIST(cache_ptr, entry_ptr) + +#if H5C2_DO_SANITY_CHECKS + + removed_entry_from_slist = TRUE; + +#endif /* H5C2_DO_SANITY_CHECKS */ + } + } + + entry_ptr->addr = new_addr; + + if ( ! ( entry_ptr->destroy_in_progress ) ) { + + was_dirty = entry_ptr->is_dirty; + + if ( ! ( entry_ptr->flush_in_progress ) ) { + + entry_ptr->is_dirty = TRUE; + } + + H5C2__INSERT_IN_INDEX(cache_ptr, entry_ptr, FAIL) + + if ( ! ( entry_ptr->flush_in_progress ) ) { + + H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, FAIL) + +#if H5C2_DO_SANITY_CHECKS + + if ( removed_entry_from_slist ) { + + /* we just removed the entry from the slist. Thus we + * must touch up cache_ptr->slist_len_increase and + * cache_ptr->slist_size_increase to keep from skewing + * the sanity checks. + */ + HDassert( cache_ptr->slist_len_increase > 1 ); + HDassert( cache_ptr->slist_size_increase > entry_ptr->size ); + + cache_ptr->slist_len_increase -= 1; + cache_ptr->slist_size_increase -= entry_ptr->size; + } + +#endif /* H5C2_DO_SANITY_CHECKS */ + + H5C2__UPDATE_RP_FOR_RENAME(cache_ptr, entry_ptr, was_dirty, FAIL) + } + } + + H5C2__UPDATE_STATS_FOR_RENAME(cache_ptr, entry_ptr) + +done: + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_rename_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_resize_pinned_entry + * + * Purpose: Resize a pinned entry. The target entry MUST be + * be pinned, and MUST not be unprotected. + * + * Resizing an entry dirties it, so if the entry is not + * already dirty, the function places the entry on the + * skip list. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 7/5/06 + * + * Modifications: + * + * None + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_resize_pinned_entry(H5C2_t * cache_ptr, + void * thing, + size_t new_size) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr; + + FUNC_ENTER_NOAPI(H5C2_resize_pinned_entry, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + + entry_ptr = (H5C2_cache_entry_t *)thing; + + if ( new_size <= 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, \ + "New size is non-positive.") + } + + if ( ! ( entry_ptr->is_pinned ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, \ + "Entry isn't pinned??") + } + + if ( entry_ptr->is_protected ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTRESIZE, FAIL, \ + "Entry is protected??") + } + + /* resizing dirties entries -- mark the entry as dirty if it + * isn't already + */ + entry_ptr->is_dirty = TRUE; + + /* update for change in entry size if necessary */ + if ( entry_ptr->size != new_size ) { + + /* update the protected entry list */ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr->pel_len), \ + (cache_ptr->pel_size), \ + (entry_ptr->size), (new_size)); + + /* update the hash table */ + H5C2__UPDATE_INDEX_FOR_SIZE_CHANGE((cache_ptr), (entry_ptr->size),\ + (new_size)); + + /* if the entry is in the skip list, update that too */ + if ( entry_ptr->in_slist ) { + + H5C2__UPDATE_SLIST_FOR_SIZE_CHANGE((cache_ptr), (entry_ptr->size),\ + (new_size)); + } + + /* update statistics just before changing the entry size */ + H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE((cache_ptr), (entry_ptr), \ + (new_size)); + + /* finally, update the entry size proper */ + entry_ptr->size = new_size; + } + + if ( ! (entry_ptr->in_slist) ) { + + H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, FAIL) + } + + H5C2__UPDATE_STATS_FOR_DIRTY_PIN(cache_ptr, entry_ptr) + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_resize_pinned_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_pin_protected_entry() + * + * Purpose: Pin a protected cache entry. The entry must be protected + * at the time of call, and must be unpinned. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 4/26/06 + * + * Modifications: + * + * JRM -- 4/26/06 + * Modified routine to allow it to operate on protected + * entries. + * + * JRM -- 2/16/07 + * Added conditional compile to avoid unused parameter + * warning in production compile. + * + * JRM -- 4/4/07 + * Fixed typo -- canged macro call to + * H5C2__UPDATE_STATS_FOR_UNPIN to call to + * H5C2__UPDATE_STATS_FOR_PIN. + * + *------------------------------------------------------------------------- + */ +#ifndef NDEBUG +herr_t +H5C2_pin_protected_entry(H5C2_t * cache_ptr, + void * thing) +#else +herr_t +H5C2_pin_protected_entry(H5C2_t UNUSED * cache_ptr, + void * thing) +#endif +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr; + + FUNC_ENTER_NOAPI(H5C2_pin_protected_entry, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + + entry_ptr = (H5C2_cache_entry_t *)thing; + + HDassert( H5F_addr_defined(entry_ptr->addr) ); + + if ( ! ( entry_ptr->is_protected ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPIN, FAIL, "Entry isn't protected") + } + + if ( entry_ptr->is_pinned ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPIN, FAIL, "Entry is already pinned") + } + + entry_ptr->is_pinned = TRUE; + + H5C2__UPDATE_STATS_FOR_PIN(cache_ptr, entry_ptr) + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_pin_protected_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_protect + * + * Purpose: If the target entry is not in the cache, load it. If + * necessary, attempt to evict one or more entries to keep + * the cache within its maximum size. + * + * Mark the target entry as protected, and return its address + * to the caller. The caller must call H5C2_unprotect() when + * finished with the entry. + * + * While it is protected, the entry may not be either evicted + * or flushed -- nor may it be accessed by another call to + * H5C2_protect. Any attempt to do so will result in a failure. + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used on the first write occasioned + * by the insertion (primary_dxpl_id), and on all subsequent + * writes (secondary_dxpl_id). This is useful in the + * metadata cache, but may not be needed elsewhere. If so, + * just use the same dxpl_id for both parameters. + * + * All reads are performed with the primary_dxpl_id. + * + * Similarly, the primary_dxpl_id is passed to the + * check_write_permitted function if it is called. + * + * Return: Success: Ptr to the desired entry + * + * Failure: NULL + * + * Programmer: John Mainzer - 6/2/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated for the addition of the hash table. + * + * JRM -- 10/28/04 + * Added code to set cache_full to TRUE whenever we try to + * make space in the cache. + * + * JRM -- 11/12/04 + * Added code to call to H5C2_make_space_in_cache() + * after the call to H5C2__auto_adjust_cache_size() if that + * function sets the size_decreased flag is TRUE. + * + * JRM -- 4/25/05 + * The size_decreased flag can also be set to TRUE in + * H5C2_set_cache_auto_resize_config() if a new configuration + * forces an immediate reduction in cache size. Modified + * the code to deal with this eventuallity. + * + * JRM -- 6/24/05 + * Added support for the new write_permitted field of H5C2_t. + * + * JRM -- 10/22/05 + * Hand optimizations. + * + * JRM -- 5/3/06 + * Added code to set the new dirtied field in + * H5C2_cache_entry_t to FALSE prior to return. + * + * JRM -- 6/23/06 + * Modified code to allow dirty entries to be loaded from + * disk. This is necessary as a bug fix in the object + * header code requires us to modify a header as it is read. + * + * JRM -- 3/28/07 + * Added the flags parameter and supporting code. At least + * for now, this parameter is used to allow the entry to + * be protected read only, thus allowing multiple protects. + * + * Also added code to allow multiple read only protects + * of cache entries. + * + * JRM -- 7/27/07 + * Added code supporting the new evictions_enabled fieled + * in H5C2_t. + * + * JRM -- 7/11/07 + * Revised function for the new metadata cache API. The + * function lost its point to H5F_t (now supplied via + * *cache_ptr), one of its dxpl ids. and one of the udata + * fields. Gained the len parameter. Also internal + * changes supporting the revised API. + * + *------------------------------------------------------------------------- + */ + +void * +H5C2_protect(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + size_t len, + const void * udata, + unsigned flags) +{ + /* const char * fcn_name = "H5C2_protect()"; */ + hbool_t hit; + hbool_t first_flush; + hbool_t have_write_permitted = FALSE; + hbool_t read_only = FALSE; + hbool_t write_permitted; + hbool_t chk_len = FALSE; + herr_t result; + void * thing; + H5C2_cache_entry_t * entry_ptr; + void * ret_value; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_protect, NULL) + + /* check args */ + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( cache_ptr->skip_file_checks || cache_ptr->f ); + HDassert( type ); + HDassert( H5F_addr_defined(addr) ); + HDassert( len > 0 ); + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HDassert(0); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, NULL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + if ( (flags & H5C2__READ_ONLY_FLAG) != 0 ) + { + read_only = TRUE; + } + + if ( (flags & H5C2__CHECK_SIZE_FLAG) != 0 ) + { + chk_len = TRUE; + } + + /* first check to see if the target is in cache */ + H5C2__SEARCH_INDEX(cache_ptr, addr, entry_ptr, NULL) + + if ( entry_ptr != NULL ) { + + hit = TRUE; + thing = (void *)entry_ptr; + + } else { + + /* must try to load the entry from disk. */ + + hit = FALSE; + + thing = H5C2_load_entry(cache_ptr->f, dxpl_id, type, + addr, len, chk_len, udata); + + if ( thing == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "can't load entry") + } + + entry_ptr = (H5C2_cache_entry_t *)thing; + + /* try to free up some space if necessary and if + * evictions are permitted + */ + if ( ( cache_ptr->evictions_enabled ) && + ( (cache_ptr->index_size + entry_ptr->size) > + cache_ptr->max_cache_size ) ) { + + size_t space_needed; + + cache_ptr->cache_full = TRUE; + + if ( cache_ptr->check_write_permitted != NULL ) { + + result = (cache_ptr->check_write_permitted)(cache_ptr->f, + dxpl_id, + &write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPROTECT, NULL, \ + "Can't get write_permitted 1") + + } else { + + have_write_permitted = TRUE; + + first_flush = TRUE; + } + } else { + + write_permitted = cache_ptr->write_permitted; + + have_write_permitted = TRUE; + + first_flush = TRUE; + } + + HDassert( entry_ptr->size <= H5C2_MAX_ENTRY_SIZE ); + + space_needed = entry_ptr->size; + + if ( space_needed > cache_ptr->max_cache_size ) { + + space_needed = cache_ptr->max_cache_size; + } + + /* Note that space_needed is just the amount of space that + * needed to insert the new entry without exceeding the cache + * size limit. The subsequent call to H5C2_make_space_in_cache() + * may evict the entries required to free more or less space + * depending on conditions. It MAY be less if the cache is + * currently undersized, or more if the cache is oversized. + * + * The cache can exceed its maximum size limit via the following + * mechanisms: + * + * First, it is possible for the cache to grow without + * bound as long as entries are protected and not unprotected. + * + * Second, when writes are not permitted it is also possible + * for the cache to grow without bound. + * + * Third, the user may choose to disable evictions -- causing + * the cache to grow without bound until evictions are + * re-enabled. + * + * Finally, we usually don't check to see if the cache is + * oversized at the end of an unprotect. As a result, it is + * possible to have a vastly oversized cache with no protected + * entries as long as all the protects preceed the unprotects. + * + * Since items 1, 2, and 3 are not changing any time soon, I + * see no point in worrying about the fourth. + */ + + result = H5C2_make_space_in_cache(dxpl_id, cache_ptr, + space_needed, write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPROTECT, NULL, \ + "H5C2_make_space_in_cache failed 1.") + } + } + + /* Insert the entry in the hash table. It can't be dirty yet, so + * we don't even check to see if it should go in the skip list. + * + * This is no longer true -- due to a bug fix, we may modify + * data on load to repair a file. + */ + H5C2__INSERT_IN_INDEX(cache_ptr, entry_ptr, NULL) + + if ( ( entry_ptr->is_dirty ) && ( ! (entry_ptr->in_slist) ) ) { + + H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, NULL) + } + + /* insert the entry in the data structures used by the replacement + * policy. We are just going to take it out again when we update + * the replacement policy for a protect, but this simplifies the + * code. If we do this often enough, we may want to optimize this. + */ + H5C2__UPDATE_RP_FOR_INSERTION(cache_ptr, entry_ptr, NULL) + } + + HDassert( entry_ptr->addr == addr ); + HDassert( entry_ptr->type == type ); + + if ( entry_ptr->is_protected ) { + + if ( ( read_only ) && ( entry_ptr->is_read_only ) ) { + + HDassert( entry_ptr->ro_ref_count > 0 ); + + (entry_ptr->ro_ref_count)++; + + } else { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPROTECT, NULL, \ + "Target already protected & not read only?!?.") + } + } else { + + H5C2__UPDATE_RP_FOR_PROTECT(cache_ptr, entry_ptr, NULL) + + entry_ptr->is_protected = TRUE; + + if ( read_only ) { + + entry_ptr->is_read_only = TRUE; + entry_ptr->ro_ref_count = 1; + } + + entry_ptr->dirtied = FALSE; + } + + H5C2__UPDATE_CACHE_HIT_RATE_STATS(cache_ptr, hit) + + H5C2__UPDATE_STATS_FOR_PROTECT(cache_ptr, entry_ptr, hit) + + ret_value = thing; + + if ( ( cache_ptr->evictions_enabled ) && + ( ( cache_ptr->size_decreased ) || + ( ( cache_ptr->resize_enabled ) && + ( cache_ptr->cache_accesses >= + (cache_ptr->resize_ctl).epoch_length ) ) ) ) { + + if ( ! have_write_permitted ) { + + if ( cache_ptr->check_write_permitted != NULL ) { + + result = (cache_ptr->check_write_permitted)(cache_ptr->f, + dxpl_id, + &write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPROTECT, NULL, \ + "Can't get write_permitted 2") + + } else { + + have_write_permitted = TRUE; + + first_flush = TRUE; + } + } else { + + write_permitted = cache_ptr->write_permitted; + + have_write_permitted = TRUE; + + first_flush = TRUE; + } + } + + if ( ( cache_ptr->resize_enabled ) && + ( cache_ptr->cache_accesses >= + (cache_ptr->resize_ctl).epoch_length ) ) { + + result = H5C2__auto_adjust_cache_size(cache_ptr, + dxpl_id, + write_permitted); + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPROTECT, NULL, \ + "Cache auto-resize failed.") + } + } + + if ( cache_ptr->size_decreased ) { + + cache_ptr->size_decreased = FALSE; + + /* check to see if the cache is now oversized due to the cache + * size reduction. If it is, try to evict enough entries to + * bring the cache size down to the current maximum cache size. + */ + if ( cache_ptr->index_size > cache_ptr->max_cache_size ) { + + cache_ptr->cache_full = TRUE; + + result = H5C2_make_space_in_cache(dxpl_id, + cache_ptr, + (size_t)0, + write_permitted); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPROTECT, NULL, \ + "H5C2_make_space_in_cache failed 2.") + } + } + } + } + +done: + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HDassert(0); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, NULL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_protect() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_reset_cache_hit_rate_stats() + * + * Purpose: Reset the cache hit rate computation fields. + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer, 10/5/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_reset_cache_hit_rate_stats(H5C2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_reset_cache_hit_rate_stats, FAIL) + + if ( ( cache_ptr == NULL ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + cache_ptr->cache_hits = 0; + cache_ptr->cache_accesses = 0; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_reset_cache_hit_rate_stats() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_set_cache_auto_resize_config + * + * Purpose: Set the cache automatic resize configuration to the + * provided values if they are in range, and fail if they + * are not. + * + * If the new configuration enables automatic cache resizing, + * coerce the cache max size and min clean size into agreement + * with the new policy and re-set the full cache hit rate + * stats. + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 10/8/04 + * + * Modifications: + * + * JRM -- 11/18/04 + * Reworked function to match major changes in + * H5C2_auto_size_ctl_t. + * + * JRM -- 4/25/05 + * Added code to set cache_ptr->size_decreased to TRUE + * if the new configuration forces an immediate reduction + * in cache size. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_set_cache_auto_resize_config(H5C2_t * cache_ptr, + H5C2_auto_size_ctl_t *config_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; + size_t new_max_cache_size; + size_t new_min_clean_size; + + FUNC_ENTER_NOAPI(H5C2_set_cache_auto_resize_config, FAIL) + + if ( ( cache_ptr == NULL ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + if ( config_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "NULL config_ptr on entry.") + } + + if ( config_ptr->version != H5C2__CURR_AUTO_SIZE_CTL_VER ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Unknown config version.") + } + + /* check general configuration section of the config: */ + if ( SUCCEED != H5C2_validate_resize_config(config_ptr, + H5C2_RESIZE_CFG__VALIDATE_GENERAL) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADRANGE, FAIL, \ + "error in general configuration fields of new config.") + } + + /* check size increase control fields of the config: */ + if ( SUCCEED != H5C2_validate_resize_config(config_ptr, + H5C2_RESIZE_CFG__VALIDATE_INCREMENT) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADRANGE, FAIL, \ + "error in the size increase control fields of new config.") + } + + /* check size decrease control fields of the config: */ + if ( SUCCEED != H5C2_validate_resize_config(config_ptr, + H5C2_RESIZE_CFG__VALIDATE_DECREMENT) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADRANGE, FAIL, \ + "error in the size decrease control fields of new config.") + } + + /* check for conflicts between size increase and size decrease controls: */ + if ( SUCCEED != H5C2_validate_resize_config(config_ptr, + H5C2_RESIZE_CFG__VALIDATE_INTERACTIONS) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADRANGE, FAIL, \ + "conflicting threshold fields in new config.") + } + + cache_ptr->size_increase_possible = TRUE; /* will set to FALSE if needed */ + cache_ptr->size_decrease_possible = TRUE; /* will set to FALSE if needed */ + + switch ( config_ptr->incr_mode ) + { + case H5C2_incr__off: + cache_ptr->size_increase_possible = FALSE; + break; + + case H5C2_incr__threshold: + if ( ( config_ptr->lower_hr_threshold <= 0.0 ) || + ( config_ptr->increment <= 1.0 ) || + ( ( config_ptr->apply_max_increment ) && + ( config_ptr->max_increment <= 0 ) ) ) { + + cache_ptr->size_increase_possible = FALSE; + } + break; + + default: /* should be unreachable */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Unknown incr_mode?!?!?.") + } + + switch ( config_ptr->decr_mode ) + { + case H5C2_decr__off: + cache_ptr->size_decrease_possible = FALSE; + break; + + case H5C2_decr__threshold: + if ( ( config_ptr->upper_hr_threshold >= 1.0 ) || + ( config_ptr->decrement >= 1.0 ) || + ( ( config_ptr->apply_max_decrement ) && + ( config_ptr->max_decrement <= 0 ) ) ) { + + cache_ptr->size_decrease_possible = FALSE; + } + break; + + case H5C2_decr__age_out: + if ( ( ( config_ptr->apply_empty_reserve ) && + ( config_ptr->empty_reserve >= 1.0 ) ) || + ( ( config_ptr->apply_max_decrement ) && + ( config_ptr->max_decrement <= 0 ) ) ) { + + cache_ptr->size_decrease_possible = FALSE; + } + break; + + case H5C2_decr__age_out_with_threshold: + if ( ( ( config_ptr->apply_empty_reserve ) && + ( config_ptr->empty_reserve >= 1.0 ) ) || + ( ( config_ptr->apply_max_decrement ) && + ( config_ptr->max_decrement <= 0 ) ) || + ( config_ptr->upper_hr_threshold >= 1.0 ) ) { + + cache_ptr->size_decrease_possible = FALSE; + } + break; + + default: /* should be unreachable */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Unknown decr_mode?!?!?.") + } + + if ( config_ptr->max_size == config_ptr->min_size ) { + + cache_ptr->size_increase_possible = FALSE; + cache_ptr->size_decrease_possible = FALSE; + } + + cache_ptr->resize_enabled = cache_ptr->size_increase_possible || + cache_ptr->size_decrease_possible; + + cache_ptr->resize_ctl = *config_ptr; + + /* Resize the cache to the supplied initial value if requested, or as + * necessary to force it within the bounds of the current automatic + * cache resizing configuration. + * + * Note that the min_clean_fraction may have changed, so we + * go through the exercise even if the current size is within + * range and an initial size has not been provided. + */ + if ( (cache_ptr->resize_ctl).set_initial_size ) { + + new_max_cache_size = (cache_ptr->resize_ctl).initial_size; + } + else if ( cache_ptr->max_cache_size > (cache_ptr->resize_ctl).max_size ) { + + new_max_cache_size = (cache_ptr->resize_ctl).max_size; + } + else if ( cache_ptr->max_cache_size < (cache_ptr->resize_ctl).min_size ) { + + new_max_cache_size = (cache_ptr->resize_ctl).min_size; + + } else { + + new_max_cache_size = cache_ptr->max_cache_size; + } + + new_min_clean_size = (size_t) + ((double)new_max_cache_size * + ((cache_ptr->resize_ctl).min_clean_fraction)); + + + /* since new_min_clean_size is of type size_t, we have + * + * ( 0 <= new_min_clean_size ) + * + * by definition. + */ + HDassert( new_min_clean_size <= new_max_cache_size ); + HDassert( (cache_ptr->resize_ctl).min_size <= new_max_cache_size ); + HDassert( new_max_cache_size <= (cache_ptr->resize_ctl).max_size ); + + if ( new_max_cache_size < cache_ptr->max_cache_size ) { + + cache_ptr->size_decreased = TRUE; + } + + cache_ptr->max_cache_size = new_max_cache_size; + cache_ptr->min_clean_size = new_min_clean_size; + + if ( H5C2_reset_cache_hit_rate_stats(cache_ptr) != SUCCEED ) { + + /* this should be impossible... */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_reset_cache_hit_rate_stats failed.") + } + + /* remove excess epoch markers if any */ + if ( ( config_ptr->decr_mode == H5C2_decr__age_out_with_threshold ) || + ( config_ptr->decr_mode == H5C2_decr__age_out ) ) { + + if ( cache_ptr->epoch_markers_active > + (cache_ptr->resize_ctl).epochs_before_eviction ) { + + result = + H5C2__autoadjust__ageout__remove_excess_markers(cache_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "can't remove excess epoch markers.") + } + } + } else if ( cache_ptr->epoch_markers_active > 0 ) { + + result = H5C2__autoadjust__ageout__remove_all_markers(cache_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "error removing all epoch markers.") + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_set_cache_auto_resize_config() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_set_evictions_enabled() + * + * Purpose: Set cache_ptr->evictions_enabled to the value of the + * evictions enabled parameter. + * + * Return: SUCCEED on success, and FAIL on failure. + * + * Programmer: John Mainzer + * 7/27/07 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_set_evictions_enabled(H5C2_t * cache_ptr, + hbool_t evictions_enabled) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_set_evictions_enabled, FAIL) + + if ( ( cache_ptr == NULL ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr on entry.") + } + + if ( ( evictions_enabled != TRUE ) && ( evictions_enabled != FALSE ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Bad evictions_enabled on entry.") + } + + /* There is no fundamental reason why we should not permit + * evictions to be disabled while automatic resize is enabled. + * However, I can't think of any good reason why one would + * want to, and allowing it would greatly complicate testing + * the feature. Hence the following: + */ + if ( ( evictions_enabled != TRUE ) && + ( ( cache_ptr->resize_ctl.incr_mode != H5C2_incr__off ) || + ( cache_ptr->resize_ctl.decr_mode != H5C2_decr__off ) ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Can't disable evictions when auto resize enabled.") + } + + cache_ptr->evictions_enabled = evictions_enabled; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_set_evictions_enabled() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_set_prefix + * + * Purpose: Set the values of the prefix field of H5C2_t. This + * filed is used to label some debugging output. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 1/20/06 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_set_prefix(H5C2_t * cache_ptr, + char * prefix) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_set_prefix, FAIL) + + /* This would normally be an assert, but we need to use an HGOTO_ERROR + * call to shut up the compiler. + */ + if ( ( ! cache_ptr ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr") + } + + HDassert( prefix ); + HDassert( HDstrlen(prefix) < H5C2__PREFIX_LEN ) ; + + HDstrcpy(&(cache_ptr->prefix[0]), prefix); + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_set_prefix() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_set_skip_flags + * + * Purpose: Set the values of the skip sanity check flags. + * + * This function and the skip sanity check flags were created + * for the convenience of the test bed. However it is + * possible that there may be other uses for the flags. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/11/04 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_set_skip_flags(H5C2_t * cache_ptr, + hbool_t skip_file_checks, + hbool_t skip_dxpl_id_checks) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_set_skip_flags, FAIL) + + /* This would normally be an assert, but we need to use an HGOTO_ERROR + * call to shut up the compiler. + */ + if ( ( ! cache_ptr ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr") + } + + cache_ptr->skip_file_checks = skip_file_checks; + cache_ptr->skip_dxpl_id_checks = skip_dxpl_id_checks; + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_set_skip_flags() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_set_trace_file_ptr + * + * Purpose: Set the trace_file_ptr field for the cache. + * + * This field must either be NULL (which turns of trace + * file logging), or be a pointer to an open file to which + * trace file data is to be written. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 1/20/06 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_set_trace_file_ptr(H5C2_t * cache_ptr, + FILE * trace_file_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_set_trace_file_ptr, FAIL) + + /* This would normally be an assert, but we need to use an HGOTO_ERROR + * call to shut up the compiler. + */ + if ( ( ! cache_ptr ) || ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr") + } + + cache_ptr->trace_file_ptr = trace_file_ptr; + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_set_trace_file_ptr() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_stats + * + * Purpose: Prints statistics about the cache. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated function for the addition of the hash table. + * + * JRM -- 9/8/05 + * Updated function for the addition of cache entry size + * change statistics. + * + * JRM -- 1/13/06 + * Added code to use the prefix field of H5C2_t to allow + * tagging of statistics output. + * + * JRM -- 3/21/06 + * Added code supporting the pinned entry related stats. + * + * JRM -- 8/9/06 + * More code supporting pinned entry related stats. + * + * JRM -- 8/23/06 + * Added code supporting new flush related statistics. + * + * JRM -- 3/31/07 + * Added code supporting the new write_protects, + * read_protects, and max_read_protects fields. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_stats(H5C2_t * cache_ptr, + const char * cache_name, + hbool_t +#if !H5C2_COLLECT_CACHE_STATS + UNUSED +#endif /* H5C2_COLLECT_CACHE_STATS */ + display_detailed_stats) +{ + herr_t ret_value = SUCCEED; /* Return value */ + +#if H5C2_COLLECT_CACHE_STATS + int i; + int64_t total_hits = 0; + int64_t total_misses = 0; + int64_t total_write_protects = 0; + int64_t total_read_protects = 0; + int64_t max_read_protects = 0; + int64_t total_insertions = 0; + int64_t total_pinned_insertions = 0; + int64_t total_clears = 0; + int64_t total_flushes = 0; + int64_t total_evictions = 0; + int64_t total_renames = 0; + int64_t total_entry_flush_renames = 0; + int64_t total_cache_flush_renames = 0; + int64_t total_size_increases = 0; + int64_t total_size_decreases = 0; + int64_t total_entry_flush_size_changes = 0; + int64_t total_cache_flush_size_changes = 0; + int64_t total_pins = 0; + int64_t total_unpins = 0; + int64_t total_dirty_pins = 0; + int64_t total_pinned_flushes = 0; + int64_t total_pinned_clears = 0; + int32_t aggregate_max_accesses = 0; + int32_t aggregate_min_accesses = 1000000; + int32_t aggregate_max_clears = 0; + int32_t aggregate_max_flushes = 0; + size_t aggregate_max_size = 0; + int32_t aggregate_max_pins = 0; + double hit_rate; + double average_successful_search_depth = 0.0; + double average_failed_search_depth = 0.0; +#endif /* H5C2_COLLECT_CACHE_STATS */ + + FUNC_ENTER_NOAPI(H5C2_stats, FAIL) + + /* This would normally be an assert, but we need to use an HGOTO_ERROR + * call to shut up the compiler. + */ + if ( ( ! cache_ptr ) || + ( cache_ptr->magic != H5C2__H5C2_T_MAGIC ) || + ( !cache_name ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Bad cache_ptr or cache_name") + } + +#if H5C2_COLLECT_CACHE_STATS + + for ( i = 0; i <= cache_ptr->max_type_id; i++ ) { + + total_hits += cache_ptr->hits[i]; + total_misses += cache_ptr->misses[i]; + total_write_protects += cache_ptr->write_protects[i]; + total_read_protects += cache_ptr->read_protects[i]; + if ( max_read_protects < cache_ptr->max_read_protects[i] ) { + max_read_protects = cache_ptr->max_read_protects[i]; + } + total_insertions += cache_ptr->insertions[i]; + total_pinned_insertions += cache_ptr->pinned_insertions[i]; + total_clears += cache_ptr->clears[i]; + total_flushes += cache_ptr->flushes[i]; + total_evictions += cache_ptr->evictions[i]; + total_renames += cache_ptr->renames[i]; + total_entry_flush_renames + += cache_ptr->entry_flush_renames[i]; + total_cache_flush_renames + += cache_ptr->cache_flush_renames[i]; + total_size_increases += cache_ptr->size_increases[i]; + total_size_decreases += cache_ptr->size_decreases[i]; + total_entry_flush_size_changes + += cache_ptr->entry_flush_size_changes[i]; + total_cache_flush_size_changes + += cache_ptr->cache_flush_size_changes[i]; + total_pins += cache_ptr->pins[i]; + total_unpins += cache_ptr->unpins[i]; + total_dirty_pins += cache_ptr->dirty_pins[i]; + total_pinned_flushes += cache_ptr->pinned_flushes[i]; + total_pinned_clears += cache_ptr->pinned_clears[i]; +#if H5C2_COLLECT_CACHE_ENTRY_STATS + if ( aggregate_max_accesses < cache_ptr->max_accesses[i] ) + aggregate_max_accesses = cache_ptr->max_accesses[i]; + if ( aggregate_min_accesses > aggregate_max_accesses ) + aggregate_min_accesses = aggregate_max_accesses; + if ( aggregate_min_accesses > cache_ptr->min_accesses[i] ) + aggregate_min_accesses = cache_ptr->min_accesses[i]; + if ( aggregate_max_clears < cache_ptr->max_clears[i] ) + aggregate_max_clears = cache_ptr->max_clears[i]; + if ( aggregate_max_flushes < cache_ptr->max_flushes[i] ) + aggregate_max_flushes = cache_ptr->max_flushes[i]; + if ( aggregate_max_size < cache_ptr->max_size[i] ) + aggregate_max_size = cache_ptr->max_size[i]; + if ( aggregate_max_pins < cache_ptr->max_pins[i] ) + aggregate_max_pins = cache_ptr->max_pins[i]; +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + } + + if ( ( total_hits > 0 ) || ( total_misses > 0 ) ) { + + hit_rate = 100.0 * ((double)(total_hits)) / + ((double)(total_hits + total_misses)); + } else { + hit_rate = 0.0; + } + + if ( cache_ptr->successful_ht_searches > 0 ) { + + average_successful_search_depth = + ((double)(cache_ptr->total_successful_ht_search_depth)) / + ((double)(cache_ptr->successful_ht_searches)); + } + + if ( cache_ptr->failed_ht_searches > 0 ) { + + average_failed_search_depth = + ((double)(cache_ptr->total_failed_ht_search_depth)) / + ((double)(cache_ptr->failed_ht_searches)); + } + + + HDfprintf(stdout, "\n%sH5C2: cache statistics for %s\n", + cache_ptr->prefix, cache_name); + + HDfprintf(stdout, "\n"); + + HDfprintf(stdout, + "%s hash table insertion / deletions = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->total_ht_insertions), + (long)(cache_ptr->total_ht_deletions)); + + HDfprintf(stdout, + "%s HT successful / failed searches = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->successful_ht_searches), + (long)(cache_ptr->failed_ht_searches)); + + HDfprintf(stdout, + "%s Av. HT suc / failed search depth = %f / %f\n", + cache_ptr->prefix, + average_successful_search_depth, + average_failed_search_depth); + + HDfprintf(stdout, + "%s current (max) index size / length = %ld (%ld) / %ld (%ld)\n", + cache_ptr->prefix, + (long)(cache_ptr->index_size), + (long)(cache_ptr->max_index_size), + (long)(cache_ptr->index_len), + (long)(cache_ptr->max_index_len)); + + HDfprintf(stdout, + "%s current (max) slist size / length = %ld (%ld) / %ld (%ld)\n", + cache_ptr->prefix, + (long)(cache_ptr->slist_size), + (long)(cache_ptr->max_slist_size), + (long)(cache_ptr->slist_len), + (long)(cache_ptr->max_slist_len)); + + HDfprintf(stdout, + "%s current (max) PL size / length = %ld (%ld) / %ld (%ld)\n", + cache_ptr->prefix, + (long)(cache_ptr->pl_size), + (long)(cache_ptr->max_pl_size), + (long)(cache_ptr->pl_len), + (long)(cache_ptr->max_pl_len)); + + HDfprintf(stdout, + "%s current (max) PEL size / length = %ld (%ld) / %ld (%ld)\n", + cache_ptr->prefix, + (long)(cache_ptr->pel_size), + (long)(cache_ptr->max_pel_size), + (long)(cache_ptr->pel_len), + (long)(cache_ptr->max_pel_len)); + + HDfprintf(stdout, + "%s current LRU list size / length = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->LRU_list_size), + (long)(cache_ptr->LRU_list_len)); + + HDfprintf(stdout, + "%s current clean LRU size / length = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->cLRU_list_size), + (long)(cache_ptr->cLRU_list_len)); + + HDfprintf(stdout, + "%s current dirty LRU size / length = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->dLRU_list_size), + (long)(cache_ptr->dLRU_list_len)); + + HDfprintf(stdout, + "%s Total hits / misses / hit_rate = %ld / %ld / %f\n", + cache_ptr->prefix, + (long)total_hits, + (long)total_misses, + hit_rate); + + HDfprintf(stdout, + "%s Total write / read (max) protects = %ld / %ld (%d)\n", + cache_ptr->prefix, + (long)total_write_protects, + (long)total_read_protects, + max_read_protects); + + HDfprintf(stdout, + "%s Total clears / flushes / evictions = %ld / %ld / %ld\n", + cache_ptr->prefix, + (long)total_clears, + (long)total_flushes, + (long)total_evictions); + + HDfprintf(stdout, + "%s Total insertions(pinned) / renames = %ld(%ld) / %ld\n", + cache_ptr->prefix, + (long)total_insertions, + (long)total_pinned_insertions, + (long)total_renames); + + HDfprintf(stdout, + "%s Total entry / cache flush renames = %ld / %ld\n", + cache_ptr->prefix, + (long)total_entry_flush_renames, + (long)total_cache_flush_renames); + + HDfprintf(stdout, "%s Total entry size incrs / decrs = %ld / %ld\n", + cache_ptr->prefix, + (long)total_size_increases, + (long)total_size_decreases); + + HDfprintf(stdout, "%s Ttl entry/cache flush size changes = %ld / %ld\n", + cache_ptr->prefix, + (long)total_entry_flush_size_changes, + (long)total_cache_flush_size_changes); + + HDfprintf(stdout, + "%s Total entry pins (dirty) / unpins = %ld (%ld) / %ld\n", + cache_ptr->prefix, + (long)total_pins, + (long)total_dirty_pins, + (long)total_unpins); + + HDfprintf(stdout, "%s Total pinned flushes / clears = %ld / %ld\n", + cache_ptr->prefix, + (long)total_pinned_flushes, + (long)total_pinned_clears); + +#if H5C2_COLLECT_CACHE_ENTRY_STATS + + HDfprintf(stdout, "%s aggregate max / min accesses = %d / %d\n", + cache_ptr->prefix, + (int)aggregate_max_accesses, + (int)aggregate_min_accesses); + + HDfprintf(stdout, "%s aggregate max_clears / max_flushes = %d / %d\n", + cache_ptr->prefix, + (int)aggregate_max_clears, + (int)aggregate_max_flushes); + + HDfprintf(stdout, "%s aggregate max_size / max_pins = %d / %d\n", + cache_ptr->prefix, + (int)aggregate_max_size, + (int)aggregate_max_pins); + +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + + if ( display_detailed_stats ) + { + + for ( i = 0; i <= cache_ptr->max_type_id; i++ ) { + + HDfprintf(stdout, "\n"); + + HDfprintf(stdout, "%s Stats on %s:\n", + cache_ptr->prefix, + ((cache_ptr->type_name_table_ptr))[i]); + + if ( ( cache_ptr->hits[i] > 0 ) || ( cache_ptr->misses[i] > 0 ) ) { + + hit_rate = 100.0 * ((double)(cache_ptr->hits[i])) / + ((double)(cache_ptr->hits[i] + cache_ptr->misses[i])); + } else { + hit_rate = 0.0; + } + + HDfprintf(stdout, + "%s hits / misses / hit_rate = %ld / %ld / %f\n", + cache_ptr->prefix, + (long)(cache_ptr->hits[i]), + (long)(cache_ptr->misses[i]), + hit_rate); + + HDfprintf(stdout, + "%s write / read (max) protects = %ld / %ld (%d)\n", + cache_ptr->prefix, + (long)(cache_ptr->write_protects[i]), + (long)(cache_ptr->read_protects[i]), + (int)(cache_ptr->max_read_protects[i])); + + HDfprintf(stdout, + "%s clears / flushes / evictions = %ld / %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->clears[i]), + (long)(cache_ptr->flushes[i]), + (long)(cache_ptr->evictions[i])); + + HDfprintf(stdout, + "%s insertions(pinned) / renames = %ld(%ld) / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->insertions[i]), + (long)(cache_ptr->pinned_insertions[i]), + (long)(cache_ptr->renames[i])); + + HDfprintf(stdout, + "%s entry / cache flush renames = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->entry_flush_renames[i]), + (long)(cache_ptr->cache_flush_renames[i])); + + HDfprintf(stdout, + "%s size increases / decreases = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->size_increases[i]), + (long)(cache_ptr->size_decreases[i])); + + HDfprintf(stdout, + "%s entry/cache flush size changes = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->entry_flush_size_changes[i]), + (long)(cache_ptr->cache_flush_size_changes[i])); + + + HDfprintf(stdout, + "%s entry pins / unpins = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->pins[i]), + (long)(cache_ptr->unpins[i])); + + HDfprintf(stdout, + "%s entry dirty pins/pin'd flushes = %ld / %ld\n", + cache_ptr->prefix, + (long)(cache_ptr->dirty_pins[i]), + (long)(cache_ptr->pinned_flushes[i])); + +#if H5C2_COLLECT_CACHE_ENTRY_STATS + + HDfprintf(stdout, + "%s entry max / min accesses = %d / %d\n", + cache_ptr->prefix, + cache_ptr->max_accesses[i], + cache_ptr->min_accesses[i]); + + HDfprintf(stdout, + "%s entry max_clears / max_flushes = %d / %d\n", + cache_ptr->prefix, + cache_ptr->max_clears[i], + cache_ptr->max_flushes[i]); + + HDfprintf(stdout, + "%s entry max_size / max_pins = %d / %d\n", + cache_ptr->prefix, + (int)(cache_ptr->max_size[i]), + (int)(cache_ptr->max_pins[i])); + + +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + + } + } + + HDfprintf(stdout, "\n"); + +#endif /* H5C2_COLLECT_CACHE_STATS */ + +done: + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_stats() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_stats__reset + * + * Purpose: Reset the stats fields to their initial values. + * + * Return: void + * + * Programmer: John Mainzer, 4/28/04 + * + * Modifications: + * + * JRM - 7/21/04 + * Updated for hash table related statistics. + * + * JRM - 9/8/05 + * Updated for size increase / decrease statistics. + * + * JRM - 3/20/06 + * Updated for pin / unpin related statistics. + * + * JRM - 8/9/06 + * Further updates for pin related statistics. + * + * JRM 8/23/06 + * Added initialization code for new flush related statistics. + * + * JRM 2/16/07 + * Added conditional compile code to avoid unused parameter + * warning in the production build. + * + * JRM 3/31/07 + * Added initialization for the new write_protects, + * read_protects, and max_read_protects fields. + * + *------------------------------------------------------------------------- + */ + +void +#ifndef NDEBUG +H5C2_stats__reset(H5C2_t * cache_ptr) +#else /* NDEBUG */ +#if H5C2_COLLECT_CACHE_STATS +H5C2_stats__reset(H5C2_t * cache_ptr) +#else /* H5C2_COLLECT_CACHE_STATS */ +H5C2_stats__reset(H5C2_t UNUSED * cache_ptr) +#endif /* H5C2_COLLECT_CACHE_STATS */ +#endif /* NDEBUG */ +{ +#if H5C2_COLLECT_CACHE_STATS + int i; +#endif /* H5C2_COLLECT_CACHE_STATS */ + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + +#if H5C2_COLLECT_CACHE_STATS + for ( i = 0; i <= cache_ptr->max_type_id; i++ ) + { + cache_ptr->hits[i] = 0; + cache_ptr->misses[i] = 0; + cache_ptr->write_protects[i] = 0; + cache_ptr->read_protects[i] = 0; + cache_ptr->max_read_protects[i] = 0; + cache_ptr->insertions[i] = 0; + cache_ptr->pinned_insertions[i] = 0; + cache_ptr->clears[i] = 0; + cache_ptr->flushes[i] = 0; + cache_ptr->evictions[i] = 0; + cache_ptr->renames[i] = 0; + cache_ptr->entry_flush_renames[i] = 0; + cache_ptr->cache_flush_renames[i] = 0; + cache_ptr->pins[i] = 0; + cache_ptr->unpins[i] = 0; + cache_ptr->dirty_pins[i] = 0; + cache_ptr->pinned_flushes[i] = 0; + cache_ptr->pinned_clears[i] = 0; + cache_ptr->size_increases[i] = 0; + cache_ptr->size_decreases[i] = 0; + cache_ptr->entry_flush_size_changes[i] = 0; + cache_ptr->cache_flush_size_changes[i] = 0; + } + + cache_ptr->total_ht_insertions = 0; + cache_ptr->total_ht_deletions = 0; + cache_ptr->successful_ht_searches = 0; + cache_ptr->total_successful_ht_search_depth = 0; + cache_ptr->failed_ht_searches = 0; + cache_ptr->total_failed_ht_search_depth = 0; + + cache_ptr->max_index_len = 0; + cache_ptr->max_index_size = (size_t)0; + + cache_ptr->max_slist_len = 0; + cache_ptr->max_slist_size = (size_t)0; + + cache_ptr->max_pl_len = 0; + cache_ptr->max_pl_size = (size_t)0; + + cache_ptr->max_pel_len = 0; + cache_ptr->max_pel_size = (size_t)0; + +#if H5C2_COLLECT_CACHE_ENTRY_STATS + + for ( i = 0; i <= cache_ptr->max_type_id; i++ ) + { + cache_ptr->max_accesses[i] = 0; + cache_ptr->min_accesses[i] = 1000000; + cache_ptr->max_clears[i] = 0; + cache_ptr->max_flushes[i] = 0; + cache_ptr->max_size[i] = (size_t)0; + cache_ptr->max_pins[i] = 0; + } + +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ +#endif /* H5C2_COLLECT_CACHE_STATS */ + + return; + +} /* H5C2_stats__reset() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_unpin_entry() + * + * Purpose: Unpin a cache entry. The entry must be unprotected at + * the time of call, and must be pinned. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 3/22/06 + * + * Modifications: + * + * JRM -- 4/26/06 + * Modified routine to allow it to operate on protected + * entries. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_unpin_entry(H5C2_t * cache_ptr, + void * thing) +{ + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr; + + FUNC_ENTER_NOAPI(H5C2_unpin_entry, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( thing ); + + entry_ptr = (H5C2_cache_entry_t *)thing; + + if ( ! ( entry_ptr->is_pinned ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPIN, FAIL, "Entry isn't pinned") + } + + if ( ! ( entry_ptr->is_protected ) ) { + + H5C2__UPDATE_RP_FOR_UNPIN(cache_ptr, entry_ptr, FAIL) + } + + entry_ptr->is_pinned = FALSE; + + H5C2__UPDATE_STATS_FOR_UNPIN(cache_ptr, entry_ptr) + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_unpin_entry() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_unprotect + * + * Purpose: Undo an H5C2_protect() call -- specifically, mark the + * entry as unprotected, remove it from the protected list, + * and give it back to the replacement policy. + * + * The TYPE and ADDR arguments must be the same as those in + * the corresponding call to H5C2_protect() and the THING + * argument must be the value returned by that call to + * H5C2_protect(). + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used on the first write occasioned + * by the unprotect (primary_dxpl_id), and on all subsequent + * writes (secondary_dxpl_id). Since an uprotect cannot + * occasion a write at present, all this is moot for now. + * However, things change, and in any case, + * H5C2_flush_single_entry() needs primary_dxpl_id and + * secondary_dxpl_id in its parameter list. + * + * The function can't cause a read either, so the dxpl_id + * parameters are moot in this case as well. + * + * Return: Non-negative on success/Negative on failure + * + * If the deleted flag is TRUE, simply remove the target entry + * from the cache, clear it, and free it without writing it to + * disk. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 6/2/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated the function for the addition of the hash table. + * In particular, we now add dirty entries to the tree if + * they aren't in the tree already. + * + * JRM -- 1/6/05 + * Added the flags parameter, and code supporting + * H5C2__SET_FLUSH_MARKER_FLAG. Note that this flag is + * ignored unless the new entry is dirty. Also note that + * once the flush_marker field of an entry is set, the + * only way it can be reset is by being flushed. + * + * JRM -- 6/3/05 + * Added the dirtied parameter and supporting code. This + * is part of an effort to move management of the is_dirty + * field into the cache code. This has become necessary + * to repair a cache coherency bug in PHDF5. + * + * JRM -- 7/5/05 + * Added code supporting the new clear_on_unprotect field + * of H5C2_cache_entry_t. This change is also part of the + * above mentioned cache coherency bug fix in PHDF5. + * + * JRM -- 9/8/05 + * Added the size_changed and new_size parameters and the + * supporting code. Since the metadata cache synchronizes + * on dirty bytes creation in the PHDF5 case, we must now + * track changes in entry size. + * + * Note that the new_size parameter is ignored unless the + * size_changed parameter is TRUE. In this case, the new_size + * must be positive. + * + * Also observe that if size_changed is TRUE, dirtied must be + * TRUE. + * + * JRM -- 9/23/05 + * Moved the size_changed parameter into flags. + * + * JRM -- 3/21/06 + * Unpdated function to pin and unpin entries as directed via + * the new H5C2__PIN_ENTRY_FLAG and H5C2__UNPIN_ENTRY_FLAG flags. + * + * JRM -- 5/3/06 + * Added code to make use of the new dirtied field in + * H5C2_cache_entry_t. If this field is TRUE, it is the + * equivalent of setting the H5C2__DIRTIED_FLAG. + * + * JRM -- 3/29/07 + * Modified function to allow a entry to be protected + * more than once if the entry is protected read only. + * + * Also added sanity checks using the new is_read_only and + * ro_ref_count parameters. + * + * JRM -- 9/8/07 + * Revised function for the new metadata cache API. The + * function lost its pointer to H5F_t (now supplied via + * *cache_ptr), and one of its dxpl ids. Also internal + * changes supporting the revised API. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_unprotect(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + void * thing, + unsigned int flags, + size_t new_size) +{ + /* const char * fcn_name = "H5C2_unprotect()"; */ + hbool_t deleted; + hbool_t dirtied; + hbool_t set_flush_marker; + hbool_t size_changed; + hbool_t pin_entry; + hbool_t unpin_entry; +#ifdef H5_HAVE_PARALLEL + hbool_t clear_entry = FALSE; +#endif /* H5_HAVE_PARALLEL */ + herr_t ret_value = SUCCEED; /* Return value */ + H5C2_cache_entry_t * entry_ptr; + H5C2_cache_entry_t * test_entry_ptr; + + FUNC_ENTER_NOAPI(H5C2_unprotect, FAIL) + + deleted = ( (flags & H5C2__DELETED_FLAG) != 0 ); + dirtied = ( (flags & H5C2__DIRTIED_FLAG) != 0 ); + set_flush_marker = ( (flags & H5C2__SET_FLUSH_MARKER_FLAG) != 0 ); + size_changed = ( (flags & H5C2__SIZE_CHANGED_FLAG) != 0 ); + pin_entry = ( (flags & H5C2__PIN_ENTRY_FLAG) != 0 ); + unpin_entry = ( (flags & H5C2__UNPIN_ENTRY_FLAG) != 0 ); + + /* Changing the size of an entry dirties it. Thus, set the + * dirtied flag if the size_changed flag is set. + */ + + dirtied |= size_changed; + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( cache_ptr->skip_file_checks || cache_ptr->f ); + HDassert( type ); + HDassert( H5F_addr_defined(addr) ); + HDassert( thing ); + HDassert( ( size_changed == TRUE ) || ( size_changed == FALSE ) ); + HDassert( ( ! size_changed ) || ( dirtied ) ); + HDassert( ( ! size_changed ) || ( new_size > 0 ) ); + HDassert( ! ( pin_entry && unpin_entry ) ); + + entry_ptr = (H5C2_cache_entry_t *)thing; + + HDassert( entry_ptr->addr == addr ); + HDassert( entry_ptr->type == type ); + + /* also set the dirtied variable if the dirtied field is set in + * the entry. + */ + dirtied |= entry_ptr->dirtied; + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + + /* if the entry has multiple read only protects, just decrement + * the ro_ref_counter. Don't actually unprotect until the ref count + * drops to zero. + */ + if ( entry_ptr->ro_ref_count > 1 ) { + + HDassert( entry_ptr->is_protected ); + HDassert( entry_ptr->is_read_only ); + + if ( dirtied ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "Read only entry modified(1)??") + } + + (entry_ptr->ro_ref_count)--; + + /* Pin or unpin the entry as requested. */ + if ( pin_entry ) { + + if ( entry_ptr->is_pinned ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPIN, FAIL, \ + "Entry already pinned???") + } + entry_ptr->is_pinned = TRUE; + H5C2__UPDATE_STATS_FOR_PIN(cache_ptr, entry_ptr) + + } else if ( unpin_entry ) { + + if ( ! ( entry_ptr->is_pinned ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPIN, FAIL, \ + "Entry already unpinned???") + } + entry_ptr->is_pinned = FALSE; + H5C2__UPDATE_STATS_FOR_UNPIN(cache_ptr, entry_ptr) + + } + + } else { + + if ( entry_ptr->is_read_only ) { + + HDassert( entry_ptr->ro_ref_count == 1 ); + + if ( dirtied ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "Read only entry modified(2)??") + } + + entry_ptr->is_read_only = FALSE; + entry_ptr->ro_ref_count = 0; + } + +#ifdef H5_HAVE_PARALLEL + /* When the H5C2 code is used to implement the metadata cache in the + * PHDF5 case, only the cache on process 0 is allowed to write to file. + * All the other metadata caches must hold dirty entries until they + * are told that the entries are clean. + * + * The clear_on_unprotect flag in the H5C2_cache_entry_t structure + * exists to deal with the case in which an entry is protected when + * its cache receives word that the entry is now clean. In this case, + * the clear_on_unprotect flag is set, and the entry is flushed with + * the H5C2__FLUSH_CLEAR_ONLY_FLAG. + * + * All this is a bit awkward, but until the metadata cache entries + * are contiguous, with only one dirty flag, we have to let the supplied + * functions deal with the reseting the is_dirty flag. + */ + if ( entry_ptr->clear_on_unprotect ) { + + HDassert( entry_ptr->is_dirty ); + + entry_ptr->clear_on_unprotect = FALSE; + + if ( ! dirtied ) { + + clear_entry = TRUE; + } + } +#endif /* H5_HAVE_PARALLEL */ + + if ( ! (entry_ptr->is_protected) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "Entry already unprotected??") + } + + /* mark the entry as dirty if appropriate */ + entry_ptr->is_dirty = ( (entry_ptr->is_dirty) || dirtied ); + + /* update for change in entry size if necessary */ + if ( ( size_changed ) && ( entry_ptr->size != new_size ) ) { + + /* update the protected list */ + H5C2__DLL_UPDATE_FOR_SIZE_CHANGE((cache_ptr->pl_len), \ + (cache_ptr->pl_size), \ + (entry_ptr->size), (new_size)); + + /* update the hash table */ + H5C2__UPDATE_INDEX_FOR_SIZE_CHANGE((cache_ptr), (entry_ptr->size),\ + (new_size)); + + /* if the entry is in the skip list, update that too */ + if ( entry_ptr->in_slist ) { + + H5C2__UPDATE_SLIST_FOR_SIZE_CHANGE((cache_ptr), \ + (entry_ptr->size),\ + (new_size)); + } + + /* update statistics just before changing the entry size */ + H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE((cache_ptr), (entry_ptr), \ + (new_size)); + + /* finally, update the entry size proper */ + entry_ptr->size = new_size; + } + + /* Pin or unpin the entry as requested. */ + if ( pin_entry ) { + + if ( entry_ptr->is_pinned ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTPIN, FAIL, \ + "Entry already pinned???") + } + entry_ptr->is_pinned = TRUE; + H5C2__UPDATE_STATS_FOR_PIN(cache_ptr, entry_ptr) + + } else if ( unpin_entry ) { + + if ( ! ( entry_ptr->is_pinned ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPIN, FAIL, \ + "Entry already unpinned???") + } + entry_ptr->is_pinned = FALSE; + H5C2__UPDATE_STATS_FOR_UNPIN(cache_ptr, entry_ptr) + + } + + /* H5C2__UPDATE_RP_FOR_UNPROTECT will places the unprotected entry on + * the pinned entry list if entry_ptr->is_pined is TRUE. + */ + H5C2__UPDATE_RP_FOR_UNPROTECT(cache_ptr, entry_ptr, FAIL) + + entry_ptr->is_protected = FALSE; + + /* if the entry is dirty, 'or' its flush_marker with the set flush flag, + * and then add it to the skip list if it isn't there already. + */ + + if ( entry_ptr->is_dirty ) { + + entry_ptr->flush_marker |= set_flush_marker; + + if ( ! (entry_ptr->in_slist) ) { + + H5C2__INSERT_ENTRY_IN_SLIST(cache_ptr, entry_ptr, FAIL) + } + } + + /* this implementation of the "deleted" option is a bit inefficient, as + * we re-insert the entry to be deleted into the replacement policy + * data structures, only to remove them again. Depending on how often + * we do this, we may want to optimize a bit. + * + * On the other hand, this implementation is reasonably clean, and + * makes good use of existing code. + * JRM - 5/19/04 + */ + if ( deleted ) { + + /* we can't delete a pinned entry */ + HDassert ( ! (entry_ptr->is_pinned ) ); + + /* verify that the target entry is in the cache. */ + + H5C2__SEARCH_INDEX(cache_ptr, addr, test_entry_ptr, FAIL) + + if ( test_entry_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "entry not in hash table?!?.") + } + else if ( test_entry_ptr != entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "hash table contains multiple entries for addr?!?.") + } + + if ( H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + type, + addr, + (H5C2__FLUSH_CLEAR_ONLY_FLAG | + H5C2__FLUSH_INVALIDATE_FLAG), + TRUE) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, "Can't flush.") + } + } +#ifdef H5_HAVE_PARALLEL + else if ( clear_entry ) { + + /* verify that the target entry is in the cache. */ + + H5C2__SEARCH_INDEX(cache_ptr, addr, test_entry_ptr, FAIL) + + if ( test_entry_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "entry not in hash table?!?.") + } + else if ( test_entry_ptr != entry_ptr ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, \ + "hash table contains multiple entries for addr?!?.") + } + + if ( H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + type, + addr, + H5C2__FLUSH_CLEAR_ONLY_FLAG, + TRUE) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTUNPROTECT, FAIL, "Can't clear.") + } + } +#endif /* H5_HAVE_PARALLEL */ + } + + H5C2__UPDATE_STATS_FOR_UNPROTECT(cache_ptr) + +done: + +#if H5C2_DO_EXTREME_SANITY_CHECKS + if ( H5C2_validate_lru_list(cache_ptr) < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "LRU sanity check failed.\n"); + } +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_unprotect() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_validate_resize_config() + * + * Purpose: Run a sanity check on the specified sections of the + * provided instance of struct H5C2_auto_size_ctl_t. + * + * Do nothing and return SUCCEED if no errors are detected, + * and flag an error and return FAIL otherwise. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: John Mainzer + * 3/23/05 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +herr_t +H5C2_validate_resize_config(H5C2_auto_size_ctl_t * config_ptr, + unsigned int tests) +{ + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_NOAPI(H5C2_validate_resize_config, FAIL) + + if ( config_ptr == NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "NULL config_ptr on entry.") + } + + if ( config_ptr->version != H5C2__CURR_AUTO_SIZE_CTL_VER ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Unknown config version.") + } + + + if ( (tests & H5C2_RESIZE_CFG__VALIDATE_GENERAL) != 0 ) { + + if ( ( config_ptr->set_initial_size != TRUE ) && + ( config_ptr->set_initial_size != FALSE ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "set_initial_size must be either TRUE or FALSE"); + } + + if ( config_ptr->max_size > H5C2__MAX_MAX_CACHE_SIZE ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "max_size too big"); + } + + if ( config_ptr->min_size < H5C2__MIN_MAX_CACHE_SIZE ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "min_size too small"); + } + + if ( config_ptr->min_size > config_ptr->max_size ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "min_size > max_size"); + } + + if ( ( config_ptr->set_initial_size ) && + ( ( config_ptr->initial_size < config_ptr->min_size ) || + ( config_ptr->initial_size > config_ptr->max_size ) ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "initial_size must be in the interval [min_size, max_size]"); + } + + if ( ( config_ptr->min_clean_fraction < 0.0 ) || + ( config_ptr->min_clean_fraction > 1.0 ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "min_clean_fraction must be in the interval [0.0, 1.0]"); + } + + if ( config_ptr->epoch_length < H5C2__MIN_AR_EPOCH_LENGTH ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "epoch_length too small"); + } + + if ( config_ptr->epoch_length > H5C2__MAX_AR_EPOCH_LENGTH ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "epoch_length too big"); + } + } /* H5C2_RESIZE_CFG__VALIDATE_GENERAL */ + + + if ( (tests & H5C2_RESIZE_CFG__VALIDATE_INCREMENT) != 0 ) { + + if ( ( config_ptr->incr_mode != H5C2_incr__off ) && + ( config_ptr->incr_mode != H5C2_incr__threshold ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "Invalid incr_mode"); + } + + if ( config_ptr->incr_mode == H5C2_incr__threshold ) { + + if ( ( config_ptr->lower_hr_threshold < 0.0 ) || + ( config_ptr->lower_hr_threshold > 1.0 ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "lower_hr_threshold must be in the range [0.0, 1.0]"); + } + + if ( config_ptr->increment < 1.0 ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "increment must be greater than or equal to 1.0"); + } + + if ( ( config_ptr->apply_max_increment != TRUE ) && + ( config_ptr->apply_max_increment != FALSE ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "apply_max_increment must be either TRUE or FALSE"); + } + + /* no need to check max_increment, as it is a size_t, + * and thus must be non-negative. + */ + } /* H5C2_incr__threshold */ + + } /* H5C2_RESIZE_CFG__VALIDATE_INCREMENT */ + + + if ( (tests & H5C2_RESIZE_CFG__VALIDATE_DECREMENT) != 0 ) { + + if ( ( config_ptr->decr_mode != H5C2_decr__off ) && + ( config_ptr->decr_mode != H5C2_decr__threshold ) && + ( config_ptr->decr_mode != H5C2_decr__age_out ) && + ( config_ptr->decr_mode != H5C2_decr__age_out_with_threshold ) + ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "Invalid decr_mode"); + } + + if ( config_ptr->decr_mode == H5C2_decr__threshold ) { + + if ( config_ptr->upper_hr_threshold > 1.0 ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "upper_hr_threshold must be <= 1.0"); + } + + if ( ( config_ptr->decrement > 1.0 ) || + ( config_ptr->decrement < 0.0 ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "decrement must be in the interval [0.0, 1.0]"); + } + + /* no need to check max_decrement as it is a size_t + * and thus must be non-negative. + */ + } /* H5C2_decr__threshold */ + + if ( ( config_ptr->decr_mode == H5C2_decr__age_out ) || + ( config_ptr->decr_mode == H5C2_decr__age_out_with_threshold ) + ) { + + if ( config_ptr->epochs_before_eviction < 1 ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "epochs_before_eviction must be positive"); + } + + if ( config_ptr->epochs_before_eviction > H5C2__MAX_EPOCH_MARKERS ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "epochs_before_eviction too big"); + } + + if ( ( config_ptr->apply_empty_reserve != TRUE ) && + ( config_ptr->apply_empty_reserve != FALSE ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "apply_empty_reserve must be either TRUE or FALSE"); + } + + if ( ( config_ptr->apply_empty_reserve ) && + ( ( config_ptr->empty_reserve > 1.0 ) || + ( config_ptr->empty_reserve < 0.0 ) ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "empty_reserve must be in the interval [0.0, 1.0]"); + } + + /* no need to check max_decrement as it is a size_t + * and thus must be non-negative. + */ + } /* H5C2_decr__age_out || H5C2_decr__age_out_with_threshold */ + + if ( config_ptr->decr_mode == H5C2_decr__age_out_with_threshold ) { + + if ( ( config_ptr->upper_hr_threshold > 1.0 ) || + ( config_ptr->upper_hr_threshold < 0.0 ) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "upper_hr_threshold must be in the interval [0.0, 1.0]"); + } + } /* H5C2_decr__age_out_with_threshold */ + + } /* H5C2_RESIZE_CFG__VALIDATE_DECREMENT */ + + + if ( (tests & H5C2_RESIZE_CFG__VALIDATE_INTERACTIONS) != 0 ) { + + if ( ( config_ptr->incr_mode == H5C2_incr__threshold ) + && + ( ( config_ptr->decr_mode == H5C2_decr__threshold ) + || + ( config_ptr->decr_mode == H5C2_decr__age_out_with_threshold ) + ) + && + ( config_ptr->lower_hr_threshold + >= + config_ptr->upper_hr_threshold + ) + ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, \ + "conflicting threshold fields in config.") + } + } /* H5C2_RESIZE_CFG__VALIDATE_INTERACTIONS */ + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_validate_resize_config() */ + + +/*************************************************************************/ +/**************************** Private Functions: *************************/ +/*************************************************************************/ + +/*------------------------------------------------------------------------- + * + * Function: H5C2__auto_adjust_cache_size + * + * Purpose: Obtain the current full cache hit rate, and compare it + * with the hit rate thresholds for modifying cache size. + * If one of the thresholds has been crossed, adjusts the + * size of the cache accordingly. + * + * The function then resets the full cache hit rate + * statistics, and exits. + * + * Return: Non-negative on success/Negative on failure or if there was + * an attempt to flush a protected item. + * + * + * Programmer: John Mainzer, 10/7/04 + * + * Modifications: + * + * JRM -- 11/18/04 + * Major re-write to support ageout method of cache size + * reduction, and to adjust to changes in the + * H5C2_auto_size_ctl_t structure. + * + * JRM -- 9/8/07 + * Reworked to accomodate cache API changes needed to + * support metadata journaling. Mostly, this involved + * removing a bunch of parameters that used to be + * passed through to other calls, and are no longer + * needed. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2__auto_adjust_cache_size(H5C2_t * cache_ptr, + hid_t dxpl_id, + hbool_t write_permitted) +{ + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; + hbool_t inserted_epoch_marker = FALSE; + size_t new_max_cache_size = 0; + size_t old_max_cache_size = 0; + size_t new_min_clean_size = 0; + size_t old_min_clean_size = 0; + double hit_rate; + enum H5C2_resize_status status = in_spec2; /* will change if needed */ + + FUNC_ENTER_NOAPI_NOINIT(H5C2__auto_adjust_cache_size) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( cache_ptr->cache_accesses >= + (cache_ptr->resize_ctl).epoch_length ); + HDassert( 0.0 <= (cache_ptr->resize_ctl).min_clean_fraction ); + HDassert( (cache_ptr->resize_ctl).min_clean_fraction <= 100.0 ); + + if ( !cache_ptr->resize_enabled ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Auto cache resize disabled.") + } + + HDassert( ( (cache_ptr->resize_ctl).incr_mode != H5C2_incr__off ) || \ + ( (cache_ptr->resize_ctl).decr_mode != H5C2_decr__off ) ); + + if ( H5C2_get_cache_hit_rate(cache_ptr, &hit_rate) != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't get hit rate.") + } + + HDassert( ( 0.0 <= hit_rate ) && ( hit_rate <= 1.0 ) ); + + switch ( (cache_ptr->resize_ctl).incr_mode ) + { + case H5C2_incr__off: + if ( cache_ptr->size_increase_possible ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "size_increase_possible but H5C2_incr__off?!?!?") + } + break; + + case H5C2_incr__threshold: + if ( hit_rate < (cache_ptr->resize_ctl).lower_hr_threshold ) { + + if ( ! cache_ptr->size_increase_possible ) { + + status = increase_disabled2; + + } else if ( cache_ptr->max_cache_size >= + (cache_ptr->resize_ctl).max_size ) { + + HDassert( cache_ptr->max_cache_size == \ + (cache_ptr->resize_ctl).max_size ); + status = at_max_size2; + + } else if ( ! cache_ptr->cache_full ) { + + status = not_full2; + + } else { + + new_max_cache_size = (size_t) + (((double)(cache_ptr->max_cache_size)) * + (cache_ptr->resize_ctl).increment); + + /* clip to max size if necessary */ + if ( new_max_cache_size > + (cache_ptr->resize_ctl).max_size ) { + + new_max_cache_size = (cache_ptr->resize_ctl).max_size; + } + + /* clip to max increment if necessary */ + if ( ( (cache_ptr->resize_ctl).apply_max_increment ) && + ( (cache_ptr->max_cache_size + + (cache_ptr->resize_ctl).max_increment) < + new_max_cache_size ) ) { + + new_max_cache_size = cache_ptr->max_cache_size + + (cache_ptr->resize_ctl).max_increment; + } + + status = increase2; + } + } + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unknown incr_mode.") + } + + /* If the decr_mode is either age out or age out with threshold, we + * must run the marker maintenance code, whether we run the size + * reduction code or not. We do this in two places -- here we + * insert a new marker if the number of active epoch markers is + * is less than the the current epochs before eviction, and after + * the ageout call, we cycle the markers. + * + * However, we can't call the ageout code or cycle the markers + * unless there was a full complement of markers in place on + * entry. The inserted_epoch_marker flag is used to track this. + */ + + if ( ( ( (cache_ptr->resize_ctl).decr_mode == H5C2_decr__age_out ) + || + ( (cache_ptr->resize_ctl).decr_mode == + H5C2_decr__age_out_with_threshold + ) + ) + && + ( cache_ptr->epoch_markers_active < + (cache_ptr->resize_ctl).epochs_before_eviction + ) + ) { + + result = H5C2__autoadjust__ageout__insert_new_marker(cache_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "can't insert new epoch marker.") + + } else { + + inserted_epoch_marker = TRUE; + } + } + + /* don't run the cache size decrease code unless the cache size + * increase code is disabled, or the size increase code sees no need + * for action. In either case, status == in_spec2 at this point. + */ + + if ( status == in_spec2 ) { + + switch ( (cache_ptr->resize_ctl).decr_mode ) + { + case H5C2_decr__off: + break; + + case H5C2_decr__threshold: + if ( hit_rate > (cache_ptr->resize_ctl).upper_hr_threshold ) { + + if ( ! cache_ptr->size_decrease_possible ) { + + status = decrease_disabled2; + + } else if ( cache_ptr->max_cache_size <= + (cache_ptr->resize_ctl).min_size ) { + + HDassert( cache_ptr->max_cache_size == + (cache_ptr->resize_ctl).min_size ); + status = at_min_size2; + + } else { + + new_max_cache_size = (size_t) + (((double)(cache_ptr->max_cache_size)) * + (cache_ptr->resize_ctl).decrement); + + /* clip to min size if necessary */ + if ( new_max_cache_size < + (cache_ptr->resize_ctl).min_size ) { + + new_max_cache_size = + (cache_ptr->resize_ctl).min_size; + } + + /* clip to max decrement if necessary */ + if ( ( (cache_ptr->resize_ctl).apply_max_decrement ) && + ( ((cache_ptr->resize_ctl).max_decrement + + new_max_cache_size) < + cache_ptr->max_cache_size ) ) { + + new_max_cache_size = cache_ptr->max_cache_size - + (cache_ptr->resize_ctl).max_decrement; + } + + status = decrease2; + } + } + break; + + case H5C2_decr__age_out_with_threshold: + case H5C2_decr__age_out: + if ( ! inserted_epoch_marker ) { + + if ( ! cache_ptr->size_decrease_possible ) { + + status = decrease_disabled2; + + } else { + + result = H5C2__autoadjust__ageout(cache_ptr, + hit_rate, + &status, + &new_max_cache_size, + dxpl_id, + write_permitted); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "ageout code failed.") + } + } + } + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unknown incr_mode.") + } + } + + /* cycle the epoch markers here if appropriate */ + if ( ( ( (cache_ptr->resize_ctl).decr_mode == H5C2_decr__age_out ) + || + ( (cache_ptr->resize_ctl).decr_mode == + H5C2_decr__age_out_with_threshold + ) + ) + && + ( ! inserted_epoch_marker ) + ) { + + /* move last epoch marker to the head of the LRU list */ + result = H5C2__autoadjust__ageout__cycle_epoch_marker(cache_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "error cycling epoch marker.") + } + } + + if ( ( status == increase2 ) || ( status == decrease2 ) ) { + + old_max_cache_size = cache_ptr->max_cache_size; + old_min_clean_size = cache_ptr->min_clean_size; + + new_min_clean_size = (size_t) + ((double)new_max_cache_size * + ((cache_ptr->resize_ctl).min_clean_fraction)); + + /* new_min_clean_size is of size_t, and thus must be non-negative. + * Hence we have + * + * ( 0 <= new_min_clean_size ). + * + * by definition. + */ + HDassert( new_min_clean_size <= new_max_cache_size ); + HDassert( (cache_ptr->resize_ctl).min_size <= new_max_cache_size ); + HDassert( new_max_cache_size <= (cache_ptr->resize_ctl).max_size ); + + cache_ptr->max_cache_size = new_max_cache_size; + cache_ptr->min_clean_size = new_min_clean_size; + + if ( status == increase2 ) { + + cache_ptr->cache_full = FALSE; + + } else if ( status == decrease2 ) { + + cache_ptr->size_decreased = TRUE; + } + } + + if ( (cache_ptr->resize_ctl).rpt_fcn != NULL ) { + + (*((cache_ptr->resize_ctl).rpt_fcn)) + (cache_ptr, + H5C2__CURR_AUTO_RESIZE_RPT_FCN_VER, + hit_rate, + status, + old_max_cache_size, + new_max_cache_size, + old_min_clean_size, + new_min_clean_size); + } + + if ( H5C2_reset_cache_hit_rate_stats(cache_ptr) != SUCCEED ) { + + /* this should be impossible... */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "H5C2_reset_cache_hit_rate_stats failed.") + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2__auto_adjust_cache_size() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__autoadjust__ageout + * + * Purpose: Implement the ageout automatic cache size decrement + * algorithm. Note that while this code evicts aged out + * entries, the code does not change the maximum cache size. + * Instead, the function simply computes the new value (if + * any change is indicated) and reports this value in + * *new_max_cache_size_ptr. + * + * Return: Non-negative on success/Negative on failure or if there was + * an attempt to flush a protected item. + * + * + * Programmer: John Mainzer, 11/18/04 + * + * Modifications: + * + * JRM -- 9/9/07 + * Reworked function to support API changes in support of + * metadata cacheing. In essence, the change involved + * removal of arguments that are no longer needed by the + * callbacks, and thus no-longer need be passed through. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2__autoadjust__ageout(H5C2_t * cache_ptr, + double hit_rate, + enum H5C2_resize_status * status_ptr, + size_t * new_max_cache_size_ptr, + hid_t dxpl_id, + hbool_t write_permitted) +{ + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; + size_t test_size; + + FUNC_ENTER_NOAPI_NOINIT(H5C2__autoadjust__ageout) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( ( status_ptr ) && ( *status_ptr == in_spec2 ) ); + HDassert( ( new_max_cache_size_ptr ) && ( *new_max_cache_size_ptr == 0 ) ); + + /* remove excess epoch markers if any */ + if ( cache_ptr->epoch_markers_active > + (cache_ptr->resize_ctl).epochs_before_eviction ) { + + result = H5C2__autoadjust__ageout__remove_excess_markers(cache_ptr); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "can't remove excess epoch markers.") + } + } + + if ( ( (cache_ptr->resize_ctl).decr_mode == H5C2_decr__age_out ) + || + ( ( (cache_ptr->resize_ctl).decr_mode == + H5C2_decr__age_out_with_threshold + ) + && + ( hit_rate >= (cache_ptr->resize_ctl).upper_hr_threshold ) + ) + ) { + + if ( cache_ptr->max_cache_size > (cache_ptr->resize_ctl).min_size ){ + + /* evict aged out cache entries if appropriate... */ + result = H5C2__autoadjust__ageout__evict_aged_out_entries + ( + dxpl_id, + cache_ptr, + write_permitted + ); + + if ( result != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "error flushing aged out entries.") + } + + /* ... and then reduce cache size if appropriate */ + if ( cache_ptr->index_size < cache_ptr->max_cache_size ) { + + if ( (cache_ptr->resize_ctl).apply_empty_reserve ) { + + test_size = (size_t)(((double)cache_ptr->index_size) / + (1 - (cache_ptr->resize_ctl).empty_reserve)); + + if ( test_size < cache_ptr->max_cache_size ) { + + *status_ptr = decrease2; + *new_max_cache_size_ptr = test_size; + } + } else { + + *status_ptr = decrease2; + *new_max_cache_size_ptr = cache_ptr->index_size; + } + + if ( *status_ptr == decrease2 ) { + + /* clip to min size if necessary */ + if ( *new_max_cache_size_ptr < + (cache_ptr->resize_ctl).min_size ) { + + *new_max_cache_size_ptr = + (cache_ptr->resize_ctl).min_size; + } + + /* clip to max decrement if necessary */ + if ( ( (cache_ptr->resize_ctl).apply_max_decrement ) && + ( ((cache_ptr->resize_ctl).max_decrement + + *new_max_cache_size_ptr) < + cache_ptr->max_cache_size ) ) { + + *new_max_cache_size_ptr = cache_ptr->max_cache_size - + (cache_ptr->resize_ctl).max_decrement; + } + } + } + } else { + + *status_ptr = at_min_size2; + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2__autoadjust__ageout() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__autoadjust__ageout__cycle_epoch_marker + * + * Purpose: Remove the oldest epoch marker from the LRU list, + * and reinsert it at the head of the LRU list. Also + * remove the epoch marker's index from the head of the + * ring buffer, and re-insert it at the tail of the ring + * buffer. + * + * Return: SUCCEED on success/FAIL on failure. + * + * Programmer: John Mainzer, 11/22/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2__autoadjust__ageout__cycle_epoch_marker(H5C2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + int i; + + FUNC_ENTER_NOAPI_NOINIT(H5C2__autoadjust__ageout__cycle_epoch_marker) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + if ( cache_ptr->epoch_markers_active <= 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "No active epoch markers on entry?!?!?.") + } + + /* remove the last marker from both the ring buffer and the LRU list */ + + i = cache_ptr->epoch_marker_ringbuf[cache_ptr->epoch_marker_ringbuf_first]; + + cache_ptr->epoch_marker_ringbuf_first = + (cache_ptr->epoch_marker_ringbuf_first + 1) % + (H5C2__MAX_EPOCH_MARKERS + 1); + + cache_ptr->epoch_marker_ringbuf_size -= 1; + + if ( cache_ptr->epoch_marker_ringbuf_size < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "ring buffer underflow.") + } + + if ( (cache_ptr->epoch_marker_active)[i] != TRUE ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unused marker in LRU?!?") + } + + H5C2__DLL_REMOVE((&((cache_ptr->epoch_markers)[i])), \ + (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (FAIL)) + + /* now, re-insert it at the head of the LRU list, and at the tail of + * the ring buffer. + */ + + HDassert( ((cache_ptr->epoch_markers)[i]).addr == (haddr_t)i ); + HDassert( ((cache_ptr->epoch_markers)[i]).next == NULL ); + HDassert( ((cache_ptr->epoch_markers)[i]).prev == NULL ); + + cache_ptr->epoch_marker_ringbuf_last = + (cache_ptr->epoch_marker_ringbuf_last + 1) % + (H5C2__MAX_EPOCH_MARKERS + 1); + + (cache_ptr->epoch_marker_ringbuf)[cache_ptr->epoch_marker_ringbuf_last] = i; + + cache_ptr->epoch_marker_ringbuf_size += 1; + + if ( cache_ptr->epoch_marker_ringbuf_size > H5C2__MAX_EPOCH_MARKERS ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "ring buffer overflow.") + } + + H5C2__DLL_PREPEND((&((cache_ptr->epoch_markers)[i])), \ + (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (FAIL)) +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2__autoadjust__ageout__cycle_epoch_marker() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__autoadjust__ageout__evict_aged_out_entries + * + * Purpose: Evict clean entries in the cache that haven't + * been accessed for at least + * (cache_ptr->resize_ctl).epochs_before_eviction epochs, + * and flush dirty entries that haven't been accessed for + * that amount of time. + * + * Depending on configuration, the function will either + * flush or evict all such entries, or all such entries it + * encounters until it has freed the maximum amount of space + * allowed under the maximum decrement. + * + * If we are running in parallel mode, writes may not be + * permitted. If so, the function simply skips any dirty + * entries it may encounter. + * + * The function makes no attempt to maintain the minimum + * clean size, as there is no guarantee that the cache size + * will be changed. + * + * If there is no cache size change, the minimum clean size + * constraint will be met through a combination of clean + * entries and free space in the cache. + * + * If there is a cache size reduction, the minimum clean size + * will be re-calculated, and will be enforced the next time + * we have to make space in the cache. + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used depending on the value of + * *first_flush_ptr. The idea is to use the primary_dxpl_id + * on the first write in a sequence of writes, and to use + * the secondary_dxpl_id on all subsequent writes. + * + * This is useful in the metadata cache, but may not be + * needed elsewhere. If so, just use the same dxpl_id for + * both parameters. + * + * Observe that this function cannot occasion a read. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 11/22/04 + * + * Modifications: + * + * JRM -- 9/9/07 + * Reworked function to support API changes in support of + * metadata cacheing. In essence, the change involved + * removal of arguments that are no longer needed by the + * callbacks, and thus no-longer need be passed through. + * + * JRM -- 10/13/07 + * Reworked code to allow it to the function to handle the + * case in which the LRU list is modified out from under the + * function by a serialize function. This can happen if + * the serialize function associated with the entry being + * flushed either accesses the next item in the LRU list, + * or (as Quincey assures me is impossible), it accesses + * an entry not currently in cache, causing the eviction + * of the next entry in the LRU. + * + * We handle this situation by detecting it, and restarting + * the scan of the LRU when it occurs. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2__autoadjust__ageout__evict_aged_out_entries(hid_t dxpl_id, + H5C2_t * cache_ptr, + hbool_t write_permitted) +{ + /* const char * fcn_name = + "H5C2__autoadjust__ageout__evict_aged_out_entries()"; */ + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; + size_t eviction_size_limit; + size_t bytes_evicted = 0; + hbool_t prev_is_dirty = FALSE; + H5C2_cache_entry_t * entry_ptr; + H5C2_cache_entry_t * next_ptr; + H5C2_cache_entry_t * prev_ptr; + + FUNC_ENTER_NOAPI_NOINIT(H5C2__autoadjust__ageout__evict_aged_out_entries) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + /* if there is a limit on the amount that the cache size can be decrease + * in any one round of the cache size reduction algorithm, load that + * limit into eviction_size_limit. Otherwise, set eviction_size_limit + * to the equivalent of infinity. The current size of the index will + * do nicely. + */ + if ( (cache_ptr->resize_ctl).apply_max_decrement ) { + + eviction_size_limit = (cache_ptr->resize_ctl).max_decrement; + + } else { + + eviction_size_limit = cache_ptr->index_size; /* i.e. infinity */ + } + + if ( write_permitted ) { + + entry_ptr = cache_ptr->LRU_tail_ptr; + + while ( ( entry_ptr != NULL ) && + ( (entry_ptr->type)->id != H5C2__EPOCH_MARKER_TYPE ) && + ( bytes_evicted < eviction_size_limit ) ) + { + HDassert( ! (entry_ptr->is_protected) ); + + next_ptr = entry_ptr->next; + prev_ptr = entry_ptr->prev; + + if ( prev_ptr != NULL ) { + + prev_is_dirty = prev_ptr->is_dirty; + } + + if ( entry_ptr->is_dirty ) { + + result = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__NO_FLAGS_SET, + FALSE); + } else { + + bytes_evicted += entry_ptr->size; + + result = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__FLUSH_INVALIDATE_FLAG, + TRUE); + } + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unable to flush entry") + } + + if ( prev_ptr != NULL ) { + + if ( prev_ptr->magic != H5C2__H5C2_CACHE_ENTRY_T_MAGIC ) { + + /* something horrible has happened to *prev_ptr -- + * scream and die. + */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "*prev_ptr corrupt") + + } else if ( ( prev_ptr->is_dirty != prev_is_dirty ) + || + ( prev_ptr->next != next_ptr ) + || + ( prev_ptr->is_protected ) + || + ( prev_ptr->is_pinned ) ) { + + /* something has happened to the LRU -- start over + * from the tail. + */ + entry_ptr = cache_ptr->LRU_tail_ptr; + + } else { + + entry_ptr = prev_ptr; + + } + } else { + + entry_ptr = NULL; + + } + } /* end while */ + + /* for now at least, don't bother to maintain the minimum clean size, + * as the cache should now be less than its maximum size. Due to + * the vaguries of the cache size reduction algorthim, we may not + * reduce the size of the cache. + * + * If we do, we will calculate a new minimum clean size, which will + * be enforced the next time we try to make space in the cache. + * + * If we don't, no action is necessary, as we have just evicted and/or + * or flushed a bunch of entries and therefore the sum of the clean + * and free space in the cache must be greater than or equal to the + * min clean space requirement (assuming that requirement was met on + * entry). + */ + + } else /* ! write_permitted */ { + + /* since we are not allowed to write, all we can do is evict + * any clean entries that we may encounter before we either + * hit the eviction size limit, or encounter the epoch marker. + * + * If we are operating read only, this isn't an issue, as there + * will not be any dirty entries. + * + * If we are operating in R/W mode, all the dirty entries we + * skip will be flushed the next time we attempt to make space + * when writes are permitted. This may have some local + * performance implications, but it shouldn't cause any net + * slowdown. + */ + + HDassert( H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS ); + + entry_ptr = cache_ptr->LRU_tail_ptr; + + while ( ( entry_ptr != NULL ) && + ( (entry_ptr->type)->id != H5C2__EPOCH_MARKER_TYPE ) && + ( bytes_evicted < eviction_size_limit ) ) + { + HDassert( ! (entry_ptr->is_protected) ); + + prev_ptr = entry_ptr->prev; + + if ( ! (entry_ptr->is_dirty) ) { + + result = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__FLUSH_INVALIDATE_FLAG, + TRUE); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unable to flush clean entry") + } + } + /* just skip the entry if it is dirty, as we can't do + * anything with it now since we can't write. + * + * Since all entries are clean, serialize() will no be called, + * and thus we needn't test to see if the LRU has been changed + * out from under us. + */ + + entry_ptr = prev_ptr; + + } /* end while */ + } + + if ( cache_ptr->index_size < cache_ptr->max_cache_size ) { + + cache_ptr->cache_full = FALSE; + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2__autoadjust__ageout__evict_aged_out_entries() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__autoadjust__ageout__insert_new_marker + * + * Purpose: Find an unused marker cache entry, mark it as used, and + * insert it at the head of the LRU list. Also add the + * marker's index in the epoch_markers array. + * + * Return: SUCCEED on success/FAIL on failure. + * + * Programmer: John Mainzer, 11/19/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2__autoadjust__ageout__insert_new_marker(H5C2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + int i; + + FUNC_ENTER_NOAPI_NOINIT(H5C2__autoadjust__ageout__insert_new_marker) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + if ( cache_ptr->epoch_markers_active >= + (cache_ptr->resize_ctl).epochs_before_eviction ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Already have a full complement of markers.") + } + + /* find an unused marker */ + i = 0; + while ( ( (cache_ptr->epoch_marker_active)[i] ) && + ( i < H5C2__MAX_EPOCH_MARKERS ) ) + { + i++; + } + + HDassert( i < H5C2__MAX_EPOCH_MARKERS ); + + if ( (cache_ptr->epoch_marker_active)[i] != FALSE ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Can't find unused marker.") + } + + HDassert( ((cache_ptr->epoch_markers)[i]).addr == (haddr_t)i ); + HDassert( ((cache_ptr->epoch_markers)[i]).next == NULL ); + HDassert( ((cache_ptr->epoch_markers)[i]).prev == NULL ); + + (cache_ptr->epoch_marker_active)[i] = TRUE; + + cache_ptr->epoch_marker_ringbuf_last = + (cache_ptr->epoch_marker_ringbuf_last + 1) % + (H5C2__MAX_EPOCH_MARKERS + 1); + + (cache_ptr->epoch_marker_ringbuf)[cache_ptr->epoch_marker_ringbuf_last] = i; + + cache_ptr->epoch_marker_ringbuf_size += 1; + + if ( cache_ptr->epoch_marker_ringbuf_size > H5C2__MAX_EPOCH_MARKERS ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "ring buffer overflow.") + } + + H5C2__DLL_PREPEND((&((cache_ptr->epoch_markers)[i])), \ + (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (FAIL)) + + cache_ptr->epoch_markers_active += 1; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2__autoadjust__ageout__insert_new_marker() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__autoadjust__ageout__remove_all_markers + * + * Purpose: Remove all epoch markers from the LRU list and mark them + * as inactive. + * + * Return: SUCCEED on success/FAIL on failure. + * + * Programmer: John Mainzer, 11/22/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2__autoadjust__ageout__remove_all_markers(H5C2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + int i; + int ring_buf_index; + + FUNC_ENTER_NOAPI_NOINIT(H5C2__autoadjust__ageout__remove_all_markers) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + while ( cache_ptr->epoch_markers_active > 0 ) + { + /* get the index of the last epoch marker in the LRU list + * and remove it from the ring buffer. + */ + + ring_buf_index = cache_ptr->epoch_marker_ringbuf_first; + i = (cache_ptr->epoch_marker_ringbuf)[ring_buf_index]; + + cache_ptr->epoch_marker_ringbuf_first = + (cache_ptr->epoch_marker_ringbuf_first + 1) % + (H5C2__MAX_EPOCH_MARKERS + 1); + + cache_ptr->epoch_marker_ringbuf_size -= 1; + + if ( cache_ptr->epoch_marker_ringbuf_size < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "ring buffer underflow.") + } + + if ( (cache_ptr->epoch_marker_active)[i] != TRUE ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unused marker in LRU?!?") + } + + /* remove the epoch marker from the LRU list */ + H5C2__DLL_REMOVE((&((cache_ptr->epoch_markers)[i])), \ + (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (FAIL)) + + /* mark the epoch marker as unused. */ + (cache_ptr->epoch_marker_active)[i] = FALSE; + + HDassert( ((cache_ptr->epoch_markers)[i]).addr == (haddr_t)i ); + HDassert( ((cache_ptr->epoch_markers)[i]).next == NULL ); + HDassert( ((cache_ptr->epoch_markers)[i]).prev == NULL ); + + /* decrement the number of active epoch markers */ + cache_ptr->epoch_markers_active -= 1; + + HDassert( cache_ptr->epoch_markers_active == \ + cache_ptr->epoch_marker_ringbuf_size ); + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2__autoadjust__ageout__remove_all_markers() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2__autoadjust__ageout__remove_excess_markers + * + * Purpose: Remove epoch markers from the end of the LRU list and + * mark them as inactive until the number of active markers + * equals the the current value of + * (cache_ptr->resize_ctl).epochs_before_eviction. + * + * Return: SUCCEED on success/FAIL on failure. + * + * Programmer: John Mainzer, 11/19/04 + * + * Modifications: + * + * None. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2__autoadjust__ageout__remove_excess_markers(H5C2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + int i; + int ring_buf_index; + + FUNC_ENTER_NOAPI_NOINIT(H5C2__autoadjust__ageout__remove_excess_markers) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + if ( cache_ptr->epoch_markers_active <= + (cache_ptr->resize_ctl).epochs_before_eviction ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "no excess markers on entry.") + } + + while ( cache_ptr->epoch_markers_active > + (cache_ptr->resize_ctl).epochs_before_eviction ) + { + /* get the index of the last epoch marker in the LRU list + * and remove it from the ring buffer. + */ + + ring_buf_index = cache_ptr->epoch_marker_ringbuf_first; + i = (cache_ptr->epoch_marker_ringbuf)[ring_buf_index]; + + cache_ptr->epoch_marker_ringbuf_first = + (cache_ptr->epoch_marker_ringbuf_first + 1) % + (H5C2__MAX_EPOCH_MARKERS + 1); + + cache_ptr->epoch_marker_ringbuf_size -= 1; + + if ( cache_ptr->epoch_marker_ringbuf_size < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "ring buffer underflow.") + } + + if ( (cache_ptr->epoch_marker_active)[i] != TRUE ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "unused marker in LRU?!?") + } + + /* remove the epoch marker from the LRU list */ + H5C2__DLL_REMOVE((&((cache_ptr->epoch_markers)[i])), \ + (cache_ptr)->LRU_head_ptr, \ + (cache_ptr)->LRU_tail_ptr, \ + (cache_ptr)->LRU_list_len, \ + (cache_ptr)->LRU_list_size, \ + (FAIL)) + + /* mark the epoch marker as unused. */ + (cache_ptr->epoch_marker_active)[i] = FALSE; + + HDassert( ((cache_ptr->epoch_markers)[i]).addr == (haddr_t)i ); + HDassert( ((cache_ptr->epoch_markers)[i]).next == NULL ); + HDassert( ((cache_ptr->epoch_markers)[i]).prev == NULL ); + + /* decrement the number of active epoch markers */ + cache_ptr->epoch_markers_active -= 1; + + HDassert( cache_ptr->epoch_markers_active == \ + cache_ptr->epoch_marker_ringbuf_size ); + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2__autoadjust__ageout__remove_excess_markers() */ + + +/*------------------------------------------------------------------------- + * Function: H5C2_flush_invalidate_cache + * + * Purpose: Flush and destroy the entries contained in the target + * cache. + * + * If the cache contains protected entries, the function will + * fail, as protected entries cannot be either flushed or + * destroyed. However all unprotected entries should be + * flushed and destroyed before the function returns failure. + * + * While pinned entries can usually be flushed, they cannot + * be destroyed. However, they should be unpinned when all + * the entries that reference them have been destroyed (thus + * reduding the pinned entry's reference count to 0, allowing + * it to be unpinned). + * + * If pinned entries are present, the function makes repeated + * passes through the cache, flushing all dirty entries + * (including the pinned dirty entries where permitted) and + * destroying all unpinned entries. This process is repeated + * until either the cache is empty, or the number of pinned + * entries stops decreasing on each pass. + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used on the first write occasioned + * by the flush (primary_dxpl_id), and on all subsequent + * writes (secondary_dxpl_id). + * + * Return: Non-negative on success/Negative on failure or if there was + * a request to flush all items and something was protected. + * + * Programmer: John Mainzer + * 3/24/065 + * + * Modifications: + * + * To support the fractal heap, the cache must now deal with + * entries being dirtied, resized, and/or renamed inside + * flush callbacks. Updated function to support this. + * + * -- JRM 8/27/06 + * + * Reworked argument list and code to reflect the + * removal of the secondary dxpl id, and the decision + * to store f in H5C2_t, removing the need to pass it + * in all the time. + * + *------------------------------------------------------------------------- + */ +herr_t +H5C2_flush_invalidate_cache(hid_t dxpl_id, + H5C2_t * cache_ptr, + unsigned flags) +{ + /* const char * fcn_name = "H5C2_flush_invalidate_cache()"; */ + herr_t status; + herr_t ret_value = SUCCEED; + hbool_t done = FALSE; + int32_t protected_entries = 0; + int32_t i; + int32_t cur_pel_len; + int32_t old_pel_len; + int32_t passes = 0; + unsigned cooked_flags; + H5SL_node_t * node_ptr = NULL; + H5C2_cache_entry_t * entry_ptr = NULL; + H5C2_cache_entry_t * next_entry_ptr = NULL; +#if H5C2_DO_SANITY_CHECKS + int64_t actual_slist_len = 0; + int64_t initial_slist_len = 0; + size_t actual_slist_size = 0; + size_t initial_slist_size = 0; +#endif /* H5C2_DO_SANITY_CHECKS */ + + FUNC_ENTER_NOAPI(H5C2_flush_invalidate_cache, FAIL) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( cache_ptr->skip_file_checks || cache_ptr->f ); + HDassert( cache_ptr->slist_ptr ); + + /* Filter out the flags that are not relevant to the flush/invalidate. + * At present, only the H5C2__FLUSH_CLEAR_ONLY_FLAG is kept. + */ + cooked_flags = flags & H5C2__FLUSH_CLEAR_ONLY_FLAG; + + /* remove ageout markers if present */ + if ( cache_ptr->epoch_markers_active > 0 ) { + + status = H5C2__autoadjust__ageout__remove_all_markers(cache_ptr); + + if ( status != SUCCEED ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "error removing all epoch markers.") + } + } + + /* The flush proceedure here is a bit strange. + * + * In the outer while loop we make at least one pass through the + * cache, and then repeat until either all the pinned entries + * unpin themselves, or until the number of pinned entries stops + * declining. In this later case, we scream and die. + * + * Since the fractal heap can dirty, resize, and/or rename entries + * in is flush callback, it is possible that the cache will still + * contain dirty entries at this point. If so, we must make up to + * H5C2__MAX_PASSES_ON_FLUSH more passes through the skip list + * to allow it to empty. If is is not empty at this point, we again + * scream and die. + * + * Further, since clean entries can be dirtied, resized, and/or renamed + * as the result of a flush call back (either the entries own, or that + * for some other cache entry), we can no longer promise to flush + * the cache entries in increasing address order. + * + * Instead, we just do the best we can -- making a pass through + * the skip list, and then a pass through the "clean" entries, and + * then repeating as needed. Thus it is quite possible that an + * entry will be evicted from the cache only to be re-loaded later + * in the flush process (From what Quincey tells me, the pin + * mechanism makes this impossible, but even it it is true now, + * we shouldn't count on it in the future.) + * + * The bottom line is that entries will probably be flushed in close + * to increasing address order, but there are no guarantees. + */ + + cur_pel_len = cache_ptr->pel_len; + old_pel_len = cache_ptr->pel_len; + + while ( ! done ) + { + /* first, try to flush-destroy any dirty entries. Do this by + * making a scan through the slist. Note that new dirty entries + * may be created by the flush call backs. Thus it is possible + * that the slist will not be empty after we finish the scan. + */ + + if ( cache_ptr->slist_len == 0 ) { + + node_ptr = NULL; + next_entry_ptr = NULL; + HDassert( cache_ptr->slist_size == 0 ); + + } else { + + node_ptr = H5SL_first(cache_ptr->slist_ptr); + + if ( node_ptr == NULL ) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "slist_len != 0 && node_ptr == NULL"); + } + + next_entry_ptr = (H5C2_cache_entry_t *)H5SL_item(node_ptr); + + if ( next_entry_ptr == NULL ) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "next_entry_ptr == NULL 1 ?!?!"); + } + + HDassert( next_entry_ptr->magic == H5C2__H5C2_CACHE_ENTRY_T_MAGIC ); + HDassert( next_entry_ptr->is_dirty ); + HDassert( next_entry_ptr->in_slist ); + + } + +#if H5C2_DO_SANITY_CHECKS + /* Depending on circumstances, H5C2_flush_single_entry() will + * remove dirty entries from the slist as it flushes them. + * Thus for sanity checks we must make note of the initial + * slist length and size before we do any flushes. + */ + initial_slist_len = cache_ptr->slist_len; + initial_slist_size = cache_ptr->slist_size; + + /* There is also the possibility that entries will be + * dirtied, resized, and/or renamed as the result of + * calls to the flush callbacks. We use the slist_len_increase + * and slist_size_increase increase fields in struct H5C2_t + * to track these changes for purpose of sanity checking. + * To this end, we must zero these fields before we start + * the pass through the slist. + */ + cache_ptr->slist_len_increase = 0; + cache_ptr->slist_size_increase = 0; + + /* Finally, reset the actual_slist_len and actual_slist_size + * fields to zero, as these fields are used to accumulate + * the slist lenght and size that we see as we scan through + * the slist. + */ + actual_slist_len = 0; + actual_slist_size = 0; +#endif /* H5C2_DO_SANITY_CHECKS */ + + while ( node_ptr != NULL ) + { + entry_ptr = next_entry_ptr; + + /* With the advent of the fractal heap, it is possible + * that the serialize callback will dirty and/or resize + * other entries in the cache. In particular, while + * Quincey has promised me that this will never happen, + * it is possible that the serialize callback for an + * entry may protect an entry that is not in the cache, + * perhaps causing the cache to flush and possibly + * evict the entry associated with node_ptr to make + * space for the new entry. + * + * Thus we do a bit of extra sanity checking on entry_ptr, + * and break out of this scan of the skip list if we + * detect major problems. We have a bit of leaway on the + * number of passes though the skip list, so this shouldn't + * be an issue in the flush in and of itself, as it should + * be all but impossible for this to happen more than once + * in any flush. + * + * Observe that that breaking out of the scan early + * shouldn't break the sanity checks just after the end + * of this while loop. + * + * If an entry has merely been marked clean and removed from + * the s-list, we simply break out of the scan. + * + * If the entry has been evicted, we flag an error and + * exit. + */ + + if ( entry_ptr->magic != H5C2__H5C2_CACHE_ENTRY_T_MAGIC ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "entry_ptr->magic is invalid ?!?!"); + + } else if ( ( ! entry_ptr->is_dirty ) || + ( ! entry_ptr->in_slist ) ) { + + /* the s-list has been modified out from under us. + * break out of the loop. + */ + break; + } + + /* increment node pointer now, before we delete its target + * from the slist. + */ + node_ptr = H5SL_next(node_ptr); + + if ( node_ptr != NULL ) { + + next_entry_ptr = (H5C2_cache_entry_t *)H5SL_item(node_ptr); + + if ( next_entry_ptr == NULL ) { + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "next_entry_ptr == NULL 2 ?!?!"); + } + + HDassert( next_entry_ptr->magic == + H5C2__H5C2_CACHE_ENTRY_T_MAGIC ); + HDassert( next_entry_ptr->is_dirty ); + HDassert( next_entry_ptr->in_slist ); + + } else { + + next_entry_ptr = NULL; + } + + /* Note that we now remove nodes from the slist as we flush + * the associated entries, instead of leaving them there + * until we are done, and then destroying all nodes in + * the slist. + * + * While this optimization used to be easy, with the possibility + * of new entries being added to the slist in the midst of the + * flush, we must keep the slist in cannonical form at all + * times. + */ + + HDassert( entry_ptr != NULL ); + HDassert( entry_ptr->in_slist ); + +#if H5C2_DO_SANITY_CHECKS + /* update actual_slist_len & actual_slist_size before + * the flush. Note that the entry will be removed + * from the slist after the flush, and thus may be + * resized by the flush callback. This is OK, as + * we will catch the size delta in + * cache_ptr->slist_size_increase. + * + * Note that we include pinned entries in this count, even + * though we will not actually flush them. + */ + actual_slist_len++; + actual_slist_size += entry_ptr->size; +#endif /* H5C2_DO_SANITY_CHECKS */ + + if ( entry_ptr->is_protected ) { + + /* we have major problems -- but lets flush + * everything we can before we flag an error. + */ + protected_entries++; + + } else if ( entry_ptr->is_pinned ) { + + /* Test to see if we are can flush the entry now. + * If we can, go ahead and flush, but don't tell + * H5C2_flush_single_entry() to destroy the entry + * as pinned entries can't be evicted. + */ + if ( TRUE ) { /* When we get to multithreaded cache, + * we will need either locking code, and/or + * a test to see if the entry is in flushable + * condition here. + */ + + status = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__NO_FLAGS_SET, + FALSE); + if ( status < 0 ) { + + /* This shouldn't happen -- if it does, we are toast + * so just scream and die. + */ + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "dirty pinned entry flush failed.") + } + } + } else { + + status = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + (cooked_flags | + H5C2__FLUSH_INVALIDATE_FLAG), + TRUE); + if ( status < 0 ) { + + /* This shouldn't happen -- if it does, we are toast so + * just scream and die. + */ + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "dirty entry flush destroy failed.") + } + } + + } /* end while loop scanning skip list */ + +#if H5C2_DO_SANITY_CHECKS + /* It is possible that entries were added to the slist during + * the scan, either before or after scan pointer. The following + * asserts take this into account. + * + * Don't bother with the sanity checks if node_ptr != NULL, as + * in this case we broke out of the loop because it got changed + * out from under us. + */ + + if ( node_ptr == NULL ) { + + HDassert( (actual_slist_len + cache_ptr->slist_len) == + (initial_slist_len + cache_ptr->slist_len_increase) ); + HDassert( (actual_slist_size + cache_ptr->slist_size) == + (initial_slist_size + cache_ptr->slist_size_increase) ); + } +#endif /* H5C2_DO_SANITY_CHECKS */ + + /* Since we are doing a destroy, we must make a pass through + * the hash table and try to flush - destroy all entries that + * remain. + * + * It used to be that all entries remaining in the cache at + * this point had to be clean, but with the fractal heap mods + * this may not be the case. If so, we will flush entries out + * of increasing address order. + * + * Writes to disk are possible here. + */ + for ( i = 0; i < H5C2__HASH_TABLE_LEN; i++ ) + { + next_entry_ptr = cache_ptr->index[i]; + + while ( next_entry_ptr != NULL ) + { + entry_ptr = next_entry_ptr; + + HDassert( entry_ptr->magic == H5C2__H5C2_CACHE_ENTRY_T_MAGIC ); + + next_entry_ptr = entry_ptr->ht_next; + + HDassert ( ( next_entry_ptr == NULL ) || + ( next_entry_ptr->magic == + H5C2__H5C2_CACHE_ENTRY_T_MAGIC ) ); + + if ( entry_ptr->is_protected ) { + + /* we have major problems -- but lets flush and destroy + * everything we can before we flag an error. + */ + protected_entries++; + + if ( ! entry_ptr->in_slist ) { + + HDassert( !(entry_ptr->is_dirty) ); + } + } else if ( ! ( entry_ptr->is_pinned ) ) { + + status = + H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + (cooked_flags | + H5C2__FLUSH_INVALIDATE_FLAG), + TRUE); + if ( status < 0 ) { + + /* This shouldn't happen -- if it does, we are toast so + * just scream and die. + */ + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "Entry flush destroy failed.") + } + } + /* We can't do anything if the entry is pinned. The + * hope is that the entry will be unpinned as the + * result of destroys of entries that reference it. + * + * We detect this by noting the change in the number + * of pinned entries from pass to pass. If it stops + * shrinking before it hits zero, we scream and die. + */ + + /* if the serialize function on the entry we last evicted + * loaded an entry into cache (as Quincey has promised me + * it never will), and if the cache was full, it is + * possible that *nexte_entry_ptr was flushed or evicted. + * + * Test to see if this happened here, and set next_entry_ptr + * to NULL if it did. Note that if this test is triggred, + * we are accessing a deallocated piece of dynamically + * allocated memory, so we just scream and die. + */ + if ( ( next_entry_ptr != NULL ) && + ( next_entry_ptr->magic != + H5C2__H5C2_CACHE_ENTRY_T_MAGIC ) ) { + + /* Something horrible has happened to + * *next_entry_ptr -- scream and die. + */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "next_entry_ptr->magic is invalid?!?!?.") + } + } /* end while loop scanning hash table bin */ + } /* end for loop scanning hash table */ + + old_pel_len = cur_pel_len; + cur_pel_len = cache_ptr->pel_len; + + if ( ( cur_pel_len > 0 ) && ( cur_pel_len >= old_pel_len ) ) { + + /* The number of pinned entries is positive, and it is not + * declining. Scream and die. + */ + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "Can't unpin all pinned entries 1.") + + } else if ( ( cur_pel_len == 0 ) && ( old_pel_len == 0 ) ) { + + /* increment the pass count */ + passes++; + } + + if ( passes >= H5C2__MAX_PASSES_ON_FLUSH ) { + + /* we have exceeded the maximum number of passes through the + * cache to flush and destroy all entries. Scream and die. + */ + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "Maximum passes on flush exceeded.") + } + + if ( cache_ptr->index_len <= 0 ) { + + done = TRUE; + HDassert( cache_ptr->index_size == 0 ); + HDassert( cache_ptr->slist_len == 0 ); + HDassert( cache_ptr->slist_size == 0 ); + HDassert( cache_ptr->pel_len == 0 ); + HDassert( cache_ptr->pel_size == 0 ); + HDassert( cache_ptr->pl_len == 0 ); + HDassert( cache_ptr->pl_size == 0 ); + HDassert( cache_ptr->LRU_list_len == 0 ); + HDassert( cache_ptr->LRU_list_size == 0 ); + } + + } /* main while loop */ + + HDassert( protected_entries <= cache_ptr->pl_len ); + + if ( protected_entries > 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "Cache has protected entries.") + + } else if ( cur_pel_len > 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "Can't unpin all pinned entries 2.") + + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_flush_invalidate_cache() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_flush_single_entry + * + * Purpose: Flush or clear (and evict if requested) the cache entry + * with the specified address and type. If the type is NULL, + * any unprotected entry at the specified address will be + * flushed (and possibly evicted). + * + * Attempts to flush a protected entry will result in an + * error. + * + * *first_flush_ptr should be true if only one + * flush is contemplated before the next load, or if this + * is the first of a sequence of flushes that will be + * completed before the next load. *first_flush_ptr is set + * to false if a flush actually takes place, and should be + * left false until the end of the sequence. + * + * The primary_dxpl_id is used if *first_flush_ptr is TRUE + * on entry, and a flush actually takes place. The + * secondary_dxpl_id is used in any subsequent flush where + * *first_flush_ptr is FALSE on entry. + * + * If the H5C2__FLUSH_INVALIDATE_FLAG flag is set, the entry will + * be cleared and not flushed -- in the case *first_flush_ptr, + * primary_dxpl_id, and secondary_dxpl_id are all irrelevent, + * and the call can't be part of a sequence of flushes. + * + * If the caller knows the address of the TBBT node at + * which the target entry resides, it can avoid a lookup + * by supplying that address in the tgt_node_ptr parameter. + * If this parameter is NULL, the function will do a TBBT + * search for the entry instead. + * + * The function does nothing silently if there is no entry + * at the supplied address, or if the entry found has the + * wrong type. + * + * Return: Non-negative on success/Negative on failure or if there was + * an attempt to flush a protected item. + * + * Programmer: John Mainzer, 5/5/04 + * + * Modifications: + * + * JRM -- 7/21/04 + * Updated function for the addition of the hash table. + * + * QAK -- 11/26/04 + * Updated function for the switch from TBBTs to skip lists. + * + * JRM -- 1/6/05 + * Updated function to reset the flush_marker field. + * Also replace references to H5F_FLUSH_INVALIDATE and + * H5F_FLUSH_CLEAR_ONLY with references to + * H5C2__FLUSH_INVALIDATE_FLAG and H5C2__FLUSH_CLEAR_ONLY_FLAG + * respectively. + * + * JRM -- 6/24/05 + * Added code to remove dirty entries from the slist after + * they have been flushed. Also added a sanity check that + * will scream if we attempt a write when writes are + * completely disabled. + * + * JRM -- 7/5/05 + * Added code to call the new log_flush callback whenever + * a dirty entry is written to disk. Note that the callback + * is not called if the H5C2__FLUSH_CLEAR_ONLY_FLAG is set, + * as there is no write to file in this case. + * + * JRM -- 8/21/06 + * Added code maintaining the flush_in_progress and + * destroy_in_progress fields in H5C2_cache_entry_t. + * + * Also added flush_flags parameter to the call to + * type_ptr->flush() so that the flush routine can report + * whether the entry has been resized or renamed. Added + * code using the flush_flags variable to detect the case + * in which the target entry is resized during flush, and + * update the caches data structures accordingly. + * + * + * JRM -- 3/29/07 + * Added sanity checks on the new is_read_only and + * ro_ref_count fields. + * + * JRM -- 6/25/07 + * Rewrite of function to use the new metadata cache callback + * functions. These functions move all metadata file I/O into + * the cache proper, which is necessary for metadata journaling. + * + * To date, the functions of the H5C2__FLUSH_INVALIDATE_FLAG + * and H5C2__FLUSH_CLEAR_ONLY_FLAG have not been documented + * in H5C2, as these flags were just passed through to the + * client callbacks. As much of the callback functionality + * is now in the cache, the function of these flags should + * be documented explicitly here in H5C2. + * + * If the H5C2__FLUSH_INVALIDATE_FLAG is set, the entry is to + * be written to disk if dirty, and then evicted from the + * cache and discarded. As an optimization, the destroyed + * is deleted from the slist only on request. + * + * If the H5C2__FLUSH_CLEAR_ONLY_FLAG is set, the entry is + * to be marked clean if it is dirty. Under no circumstances + * will it be written to disk. + * + * If both the H5C2__FLUSH_INVALIDATE_FLAG and the + * H5C2__FLUSH_CLEAR_ONLY_FLAG are set, the entry is marked + * clean and then evicted from the cache without writing + * to disk. If dirty, the entry is removed from the slist + * or not as requested. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2_flush_single_entry(H5F_t * f, + hid_t dxpl_id, + H5C2_t * cache_ptr, + const H5C2_class_t * type_ptr, + haddr_t addr, + unsigned flags, + hbool_t del_entry_from_slist_on_destroy) +{ + /* const char * fcn_name = "H5C2_flush_single_entry()"; */ + hbool_t destroy; + hbool_t clear_only; + hbool_t was_dirty; + herr_t ret_value = SUCCEED; /* Return value */ + herr_t status; + int type_id; + unsigned serialize_flags = 0; + haddr_t new_addr; + size_t new_len; + void * new_image_ptr; + H5C2_cache_entry_t * entry_ptr = NULL; + + FUNC_ENTER_NOAPI_NOINIT(H5C2_flush_single_entry) + + HDassert( f ); + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( type_ptr ); + HDassert( H5F_addr_defined(addr) ); + + destroy = ( (flags & H5C2__FLUSH_INVALIDATE_FLAG) != 0 ); + clear_only = ( (flags & H5C2__FLUSH_CLEAR_ONLY_FLAG) != 0); + + /* attempt to find the target entry in the hash table */ + H5C2__SEARCH_INDEX(cache_ptr, addr, entry_ptr, FAIL) + +#if H5C2_DO_SANITY_CHECKS + if ( entry_ptr != NULL ) { + + HDassert( ! ( ( destroy ) && ( entry_ptr->is_pinned ) ) ); + + if ( entry_ptr->in_slist ) { + + if ( ( ( entry_ptr->flush_marker ) && ( ! entry_ptr->is_dirty ) ) + || + ( entry_ptr->addr != addr ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "entry in slist failed sanity checks.") + } + } else { + + if ( ( entry_ptr->is_dirty ) || + ( entry_ptr->flush_marker ) || + ( entry_ptr->addr != addr ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "entry failed sanity checks.") + } + } + } +#if 1 + /* this should be useful for debugging from time to time. + * lets leave it in for now. -- JRM 12/15/04 + */ + else { + HDfprintf(stdout, + "H5C2_flush_single_entry(): non-existant entry. addr = 0x%lx\n", + (long)addr); + HDfflush(stdout); + } +#endif +#endif /* H5C2_DO_SANITY_CHECKS */ + + if ( ( entry_ptr != NULL ) && ( entry_ptr->is_protected ) ) + { + /* Attempt to flush a protected entry -- scream and die. */ + HGOTO_ERROR(H5E_CACHE, H5E_PROTECT, FAIL, \ + "Attempt to flush a protected entry.") + } + + if ( ( entry_ptr != NULL ) && + ( ( type_ptr == NULL ) || ( type_ptr->id == entry_ptr->type->id ) ) ) + { + /* we have work to do */ + + /* We will set flush_in_progress back to FALSE at the end if the + * entry still exists at that point. + */ + entry_ptr->flush_in_progress = TRUE; + +#ifdef H5_HAVE_PARALLEL +#ifndef NDEBUG + + /* If MPI based VFD is used, do special parallel I/O sanity checks. + * Note that we only do these sanity checks when the clear_only flag + * is not set, and the entry to be flushed is dirty. Don't bother + * otherwise as no file I/O can result. + */ + if ( ( ! clear_only ) && + ( entry_ptr->is_dirty ) && + ( IS_H5FD_MPI(f) ) ) { + + H5P_genplist_t *dxpl; /* Dataset transfer property list */ + H5FD_mpio_xfer_t xfer_mode; /* I/O xfer mode property value */ + + /* Get the dataset transfer property list */ + if ( NULL == (dxpl = H5I_object(dxpl_id)) ) { + + HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, \ + "not a dataset creation property list") + } + + /* Get the transfer mode property */ + if( H5P_get(dxpl, H5D_XFER_IO_XFER_MODE_NAME, &xfer_mode) < 0 ) { + + HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, \ + "can't retrieve xfer mode") + } + + /* Sanity check transfer mode */ + /* I'm surprised that this sanity check is working at + * present -- will need to look into it at some point. + * + * JRM -- 7/7/07 + */ + + HDassert( xfer_mode == H5FD_MPIO_COLLECTIVE ); + } + +#endif /* NDEBUG */ +#endif /* H5_HAVE_PARALLEL */ + + was_dirty = entry_ptr->is_dirty; + type_id = entry_ptr->type->id; + + entry_ptr->flush_marker = FALSE; + + if ( clear_only ) { + H5C2__UPDATE_STATS_FOR_CLEAR(cache_ptr, entry_ptr) + } else { + H5C2__UPDATE_STATS_FOR_FLUSH(cache_ptr, entry_ptr) + } + + if ( destroy ) { + H5C2__UPDATE_STATS_FOR_EVICTION(cache_ptr, entry_ptr) + } + + /* Always remove the entry from the hash table on a destroy. On a + * flush with destroy, it is cheaper to discard the skip list all at + * once rather than remove the entries one by one, so we only delete + * from the slist only if requested. + * + * Note that it is possible that the entry will be renamed during + * its call to flush. This will upset H5C2_rename_entry() if we + * don't tell it that it doesn't have to worry about updating the + * index and SLIST. Use the destroy_in_progress field for this + * purpose. + */ + if ( destroy ) { + + entry_ptr->destroy_in_progress = TRUE; + + H5C2__DELETE_FROM_INDEX(cache_ptr, entry_ptr) + + if ( ( entry_ptr->in_slist ) && + ( del_entry_from_slist_on_destroy ) ) { + + H5C2__REMOVE_ENTRY_FROM_SLIST(cache_ptr, entry_ptr) + } + } + + /* Update the replacement policy for the flush or eviction. */ + if ( destroy ) { /* AKA eviction */ + +#if 0 /* JRM */ + /* This test code may come in handy -- lets keep it for a while. + * + * Note that it will cause spurious errors in the serial case + * unless we are maintaining the clean and dirty LRU lists. + */ + { + if ( entry_ptr->is_dirty ) + { + if ( cache_ptr->dLRU_head_ptr == NULL ) + HDfprintf(stdout, + "%s: cache_ptr->dLRU_head_ptr == NULL.\n", + fcn_name); + + if ( cache_ptr->dLRU_tail_ptr == NULL ) + HDfprintf(stdout, + "%s: cache_ptr->dLRU_tail_ptr == NULL.\n", + fcn_name); + + if ( cache_ptr->dLRU_list_len <= 0 ) + HDfprintf(stdout, + "%s: cache_ptr->dLRU_list_len <= 0.\n", + fcn_name); + + if ( cache_ptr->dLRU_list_size <= 0 ) + HDfprintf(stdout, + "%s: cache_ptr->dLRU_list_size <= 0.\n", + fcn_name); + + if ( cache_ptr->dLRU_list_size < entry_ptr->size ) + HDfprintf(stdout, + "%s: cache_ptr->dLRU_list_size < entry_ptr->size.\n", + fcn_name); + + if ( ( (cache_ptr->dLRU_list_size) == entry_ptr->size ) && + ( ! ( (cache_ptr->dLRU_list_len) == 1 ) ) ) + HDfprintf(stdout, + "%s: dLRU_list_size == size && dLRU_list_len != 1\n", + fcn_name); + + if ( ( entry_ptr->aux_prev == NULL ) && + ( cache_ptr->dLRU_head_ptr != entry_ptr ) ) + HDfprintf(stdout, + "%s: entry_ptr->aux_prev == NULL && dLRU_head_ptr != entry_ptr\n", + fcn_name); + + if ( ( entry_ptr->aux_next == NULL ) && + ( cache_ptr->dLRU_tail_ptr != entry_ptr ) ) + HDfprintf(stdout, + "%s: entry_ptr->aux_next == NULL && dLRU_tail_ptr != entry_ptr\n", + fcn_name); + + if ( ( cache_ptr->dLRU_list_len == 1 ) && + ( ! ( ( cache_ptr->dLRU_head_ptr == entry_ptr ) && + ( cache_ptr->dLRU_tail_ptr == entry_ptr ) && + ( entry_ptr->aux_next == NULL ) && + ( entry_ptr->aux_prev == NULL ) && + ( cache_ptr->dLRU_list_size == entry_ptr->size ) + ) + ) + ) + { + HDfprintf(stdout, + "%s: single entry dlru sanity check fails\n", + fcn_name); + } + + } + else + { + if ( cache_ptr->cLRU_head_ptr == NULL ) + HDfprintf(stdout, + "%s: cache_ptr->cLRU_head_ptr == NULL.\n", + fcn_name); + + if ( cache_ptr->cLRU_tail_ptr == NULL ) + HDfprintf(stdout, + "%s: cache_ptr->cLRU_tail_ptr == NULL.\n", + fcn_name); + + if ( cache_ptr->cLRU_list_len <= 0 ) + HDfprintf(stdout, + "%s: cache_ptr->cLRU_list_len <= 0.\n", + fcn_name); + + if ( cache_ptr->cLRU_list_size <= 0 ) + HDfprintf(stdout, + "%s: cache_ptr->cLRU_list_size <= 0.\n", + fcn_name); + + if ( cache_ptr->cLRU_list_size < entry_ptr->size ) + HDfprintf(stdout, + "%s: cache_ptr->cLRU_list_size < entry_ptr->size.\n", + fcn_name); + + if ( ( (cache_ptr->cLRU_list_size) == entry_ptr->size ) && + ( ! ( (cache_ptr->cLRU_list_len) == 1 ) ) ) + HDfprintf(stdout, + "%s: cLRU_list_size == size && cLRU_list_len != 1\n", + fcn_name); + + if ( ( entry_ptr->aux_prev == NULL ) && + ( cache_ptr->cLRU_head_ptr != entry_ptr ) ) + HDfprintf(stdout, "%s: entry_ptr->aux_prev == NULL && cLRU_head_ptr != entry_ptr\n", fcn_name); + + if ( ( entry_ptr->aux_next == NULL ) && + ( cache_ptr->cLRU_tail_ptr != entry_ptr ) ) + HDfprintf(stdout, "%s: entry_ptr->aux_next == NULL && cLRU_tail_ptr != entry_ptr\n", fcn_name); + + if ( ( cache_ptr->cLRU_list_len == 1 ) && + ( ! ( ( cache_ptr->cLRU_head_ptr == entry_ptr ) && + ( cache_ptr->cLRU_tail_ptr == entry_ptr ) && + ( entry_ptr->aux_next == NULL ) && + ( entry_ptr->aux_prev == NULL ) && + ( cache_ptr->cLRU_list_size == entry_ptr->size ) + ) + ) + ) + { + HDfprintf(stdout, + "%s: single entry clru sanity check fails\n", + fcn_name); + } + } + } +#endif /* JRM */ + + H5C2__UPDATE_RP_FOR_EVICTION(cache_ptr, entry_ptr, FAIL) + + } else { + /* We are either doing a flush or a clear. + * + * A clear and a flush are the same from the point of view of + * the replacement policy and the slist. Hence no + * differentiation between them. + * JRM -- 7/7/07 + */ + + H5C2__UPDATE_RP_FOR_FLUSH(cache_ptr, entry_ptr, FAIL) + + if ( entry_ptr->in_slist ) + { + H5C2__REMOVE_ENTRY_FROM_SLIST(cache_ptr, entry_ptr) + } + } + + /* Clear the dirty flag only, if requested */ + if ( clear_only ) + { + if ( entry_ptr->is_dirty ) + { +#ifndef NDEBUG + /* only call the clear_dirty_bits callback if debugging + * is enabled. + */ + if ( entry_ptr->type->clear_dirty_bits(addr, + entry_ptr->size, + (void *)entry_ptr) + != SUCCEED ) + { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "clear_dirty_bits() failed.") + } +#endif /* NDEBUG */ + } + entry_ptr->is_dirty = FALSE; + + } + else if ( entry_ptr->is_dirty ) + { + /* The entry is dirty, and we are doing either a flush, + * or a flush destroy. In either case, serialize the + * entry and write it to disk. + * + * If the entry is clean, we do nothing at this point. + * + * If the serialize function changes the size or location + * of the entry, and we are not doing a flush destroy, we + * will have to touch up the cache to account for the + * change(s). + */ + +#if H5C2_DO_SANITY_CHECKS + if ( ( cache_ptr->check_write_permitted == NULL ) && + ( ! (cache_ptr->write_permitted) ) ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Write when writes are always forbidden!?!?!") + } +#endif /* H5C2_DO_SANITY_CHECKS */ + + if ( entry_ptr->image_ptr == NULL ) + { + entry_ptr->image_ptr = H5MM_malloc(entry_ptr->size); + + if ( entry_ptr->image_ptr == NULL ) + { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, \ + "memory allocation failed for on disk image buffer.") + } + } + + if ( entry_ptr->type->serialize(entry_ptr->addr, + entry_ptr->size, + entry_ptr->image_ptr, + (void *)entry_ptr, + &serialize_flags, + &new_addr, + &new_len, + &new_image_ptr) != SUCCEED ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unable to serialize entry") + } + + if ( serialize_flags != 0 ) + { + if ( destroy ) + { + /* We have already removed the entry from the + * cache's data structures, so no need to update + * them for the re-size and/or rename. All we need + * to do is update the cache entry so we will have + * the correct values when we actually write the + * image of the entry to disk. + * + * Note that if the serialize function changes the + * size of the disk image of the entry, it must + * deallocate the old image, and allocate a new. + */ + + switch ( serialize_flags ) + { + case H5C2__SERIALIZE_RESIZED_FLAG: + H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE(cache_ptr, + entry_ptr, + new_len) + entry_ptr->size = new_len; + entry_ptr->image_ptr = new_image_ptr; + break; + + case (H5C2__SERIALIZE_RESIZED_FLAG | + H5C2__SERIALIZE_RENAMED_FLAG): + H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE(cache_ptr, + entry_ptr, + new_len) + entry_ptr->addr = new_addr; + entry_ptr->size = new_len; + entry_ptr->image_ptr = new_image_ptr; + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unexpected serialize flag(s)") + break; + } + } + else + { + /* The entry is not being destroyed, and thus has not + * been removed from the cache's data structures. + * + * Thus, in addition to updating the entry for the + * re-size and/or rename, we must also update the + * cache data structures. + */ + + switch ( serialize_flags ) + { + case H5C2__SERIALIZE_RESIZED_FLAG: + H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE(cache_ptr, + entry_ptr, + new_len) + /* The replacement policy code thinks the + * entry is already clean, so modify is_dirty + * to meet this expectation. + */ + entry_ptr->is_dirty = FALSE; + + /* update the hash table for the size change*/ + H5C2__UPDATE_INDEX_FOR_SIZE_CHANGE((cache_ptr), \ + (entry_ptr->size),\ + (new_len)); + + /* The entry can't be protected since we are in + * the process of flushing it. Thus we must + * update the replacement policy data structures + * for the size change. The macro deals with + * the pinned case. + */ + H5C2__UPDATE_RP_FOR_SIZE_CHANGE(cache_ptr, \ + entry_ptr, \ + new_len); + + /* The entry can't be in the slist, so no need + * to update the slist for the size change. + */ + + /* finally, set is_dirty to TRUE again, and + * update the size and image_ptr. + */ + entry_ptr->is_dirty = TRUE; + entry_ptr->size = new_len; + entry_ptr->image_ptr = new_image_ptr; + break; + + case (H5C2__SERIALIZE_RESIZED_FLAG | + H5C2__SERIALIZE_RENAMED_FLAG): + H5C2__UPDATE_STATS_FOR_ENTRY_SIZE_CHANGE(cache_ptr, + entry_ptr, + new_len) + /* The replacement policy code thinks the + * entry is already clean, so modify is_dirty + * to meet this expectation. + */ + entry_ptr->is_dirty = FALSE; + + /* first update the hash table for the rename */ + H5C2__DELETE_FROM_INDEX(cache_ptr, entry_ptr) + entry_ptr->addr = new_addr; + H5C2__INSERT_IN_INDEX(cache_ptr, entry_ptr, FAIL) + + /* update the hash table for the size change */ + H5C2__UPDATE_INDEX_FOR_SIZE_CHANGE((cache_ptr), \ + (entry_ptr->size),\ + (new_len)); + + /* The entry can't be protected since we are in + * the process of flushing it. Thus we must + * update the replacement policy data structures + * for the size change. The macro deals with + * the pinned case. + */ + H5C2__UPDATE_RP_FOR_SIZE_CHANGE(cache_ptr, \ + entry_ptr, \ + new_len); + + /* The entry can't be in the slist, so no need + * to update the slist for the size change. + */ + + /* finally, set is_dirty to TRUE again, and + * update the size and image_ptr. + */ + entry_ptr->is_dirty = TRUE; + + entry_ptr->size = new_len; + entry_ptr->image_ptr = new_image_ptr; + break; + + default: + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unexpected serialize flag(s)") + break; + } + } + } + + /* now write the image to disk */ + if ( H5F_block_write(f, type_ptr->mem_type, entry_ptr->addr, + entry_ptr->size, dxpl_id, + entry_ptr->image_ptr) < 0 ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "Can't write image to file.") + } + +#ifdef H5_HAVE_PARALLEL + if ( serialize_flags != 0 ) { + + /* In the parallel case, resizes and renames in + * the serialize operation can cause problems. + * If they occur, scream and die. + * + * At present, in the parallel case, the aux_ptr + * will only be set if there is more than one + * process. Thus we can use this to detect + * the parallel case. + * + * This works for now, but if we start using the + * aux_ptr for other purposes, we will have to + * change this test accordingly. + * + * NB: While this test detects entryies that attempt + * to resize or rename themselves during a flush + * in the parallel case, it will not detect an + * entry that dirties, resizes, and/or renames + * other entries during its flush. + * + * From what Quincey tells me, this test is + * sufficient for now, as any flush routine that + * does the latter will also do the former. + * + * If that ceases to be the case, further + * tests will be necessary. + */ + if ( cache_ptr->aux_ptr != NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "resize/rename in serialize occured in parallel case.") + + } + } +#endif /* H5_HAVE_PARALLEL */ + + entry_ptr->is_dirty = FALSE; + } + + if ( destroy ) /* time to discard the entry */ + { + /* start by freeing the buffer for the on disk image */ + if ( entry_ptr->image_ptr != NULL ) { + + entry_ptr->image_ptr = H5MM_xfree(entry_ptr->image_ptr); + + if ( entry_ptr->image_ptr != NULL ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "*image_ptr free failed.") + } + } + + /* we are about to discard the in core representation -- + * set the magic field to bad magic so we can detect a + * freed entry if we see one. + */ + entry_ptr->magic = H5C2__H5C2_CACHE_ENTRY_T_BAD_MAGIC; + + if ( type_ptr->free_icr(entry_ptr->addr, entry_ptr->size, + (void *)entry_ptr) != SUCCEED ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "free_icr callback failed.") + } + } + else /* just flushing or clearing */ + { + entry_ptr->flush_in_progress = FALSE; + } + + if ( cache_ptr->log_flush ) { + + /* JRM */ /* may want to update this */ + status = (cache_ptr->log_flush)(cache_ptr, addr, was_dirty, + flags, type_id); + + if ( status < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "log_flush callback failed.") + } + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_flush_single_entry() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_load_entry + * + * Purpose: Attempt to load the entry at the specified disk address + * and with the specified type into memory. If successful. + * return the in memory address of the entry. Return NULL + * on failure. + * + * Note that this function simply loads the entry into + * core. It does not insert it into the cache. + * + * Return: Non-NULL on success / NULL on failure. + * + * Programmer: John Mainzer, 5/18/04 + * + * Modifications: + * + * JRM - 7/21/04 + * Updated function for the addition of the hash table. + * + * JRM - 6/23/06 + * Deleted assertion that verified that a newly loaded + * entry is clean. Due to a bug fix, this need not be + * the case, as our code will attempt to repair errors + * on load. + * + * JRM - 8/21/06 + * Added initialization for the new flush_in_progress and + * destroy in progress fields. + * + * JRM - 3/29/07 + * Added initialization for the new is_read_only and + * ro_ref_count fields. + * + * JRM - 6/25/07 + * Reworked function to use the new client callback + * functions that are needed to implement metadata + * journaling. Removed skip_file_checks parameter. + * + * JRM -- 10/12/07 + * Added initialization for the new magic field. + * + *------------------------------------------------------------------------- + */ + +static void * +H5C2_load_entry(H5F_t * f, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + size_t len, + hbool_t chk_len, + const void * udata_ptr) +{ + /* const char * fcn_name = "H5C2_load_entry()"; */ + hbool_t dirty = FALSE; + haddr_t abs_eoa; /* Absolute end of file address */ + haddr_t rel_eoa; /* Relative end of file address */ + void * image_ptr = NULL; + void * thing = NULL; + void * ret_value = NULL; + H5C2_cache_entry_t * entry_ptr = NULL; + size_t new_len; + + FUNC_ENTER_NOAPI_NOINIT(H5C2_load_entry) + + HDassert( f ); + HDassert( type ); + HDassert( H5F_addr_defined(addr) ); + HDassert( len > 0 ); + HDassert( type->deserialize ); + HDassert( ( ! chk_len ) || ( type->image_len ) ); + + /* Make certain we don't speculatively read off the end of the file */ + if ( HADDR_UNDEF == (abs_eoa = H5F_get_eoa(f)) ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, \ + "unable to determine file size") + } + + /* Adjust absolute EOA address to relative EOA address */ + rel_eoa = abs_eoa - H5F_get_base_addr(f); + + HDassert( rel_eoa > addr ); + + if ( (rel_eoa - addr) < (haddr_t)len ) + { + /* either the client is speculatively reading beyond the end of + * file, or there is a major screw up. Adjust the len in the + * former case, and scream and die in the latter + */ + if ( chk_len ) /* i.e. possible speculative read beyond eof */ + { + /* Quincey: Did I use this correctly? In the example I + * stole it from, the from type was hsize_t even though + * the source was a haddr_t. I changed the from to match + * the source. Is this as it should be? + */ + /* JRM */ + H5_ASSIGN_OVERFLOW(len, (rel_eoa - addr), \ + /* from: */ haddr_t, /* to: */ size_t); + HDassert( len > 0 ); + } + else + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, \ + "tried to read beyond eof") + + } + } + + image_ptr = H5MM_malloc(len); + + if ( image_ptr == NULL ) + { + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, \ + "memory allocation failed for on disk image buffer.") + } + + if ( H5F_block_read(f, type->mem_type, addr, len, dxpl_id, image_ptr) < 0 ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "Can't read image") + } + + thing = type->deserialize(addr, len, image_ptr, udata_ptr, &dirty); + + if ( thing == NULL ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "Can't deserialize image") + } + + if ( chk_len ) + { + if ( type->image_len(thing, &new_len) != SUCCEED ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "image_len() failed") + } + else if ( new_len > len ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "new_len > len") + } + else if ( new_len <= 0 ) + { + HGOTO_ERROR(H5E_CACHE, H5E_CANTLOAD, NULL, "new_len <= 0") + } + else if ( new_len < len ) + { + thing = H5MM_realloc(thing, new_len); + + if ( thing == NULL ) + { + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, \ + "thing null after H5MM_realloc().") + } + else + { + len = new_len; + } + } + } + + entry_ptr = (H5C2_cache_entry_t *)thing; + + /* In general, an entry should be clean just after it is loaded. + * + * However, when this code is used in the metadata cache, it is + * possible that object headers will be dirty at this point, as + * the deserialize function will alter object headers if necessary to + * fix an old bug. + * + * In the following assert: + * + * HDassert( ( dirty == FALSE ) || ( type->id == 4 ) ); + * + * note that type id 4 is associated with object headers in the metadata + * cache. + * + * When we get to using H5C2 for other purposes, we may wish to + * tighten up the assert so that the loophole only applies to the + * metadata cache. + */ + + HDassert( ( dirty == FALSE ) || ( type->id == 4 ) ); + HDassert( entry_ptr->size < H5C2_MAX_ENTRY_SIZE ); + + entry_ptr->magic = H5C2__H5C2_CACHE_ENTRY_T_MAGIC; + entry_ptr->addr = addr; + entry_ptr->size = len; + entry_ptr->image_ptr = image_ptr; + entry_ptr->type = type; + entry_ptr->is_dirty = dirty; + entry_ptr->dirtied = FALSE; + entry_ptr->is_protected = FALSE; + entry_ptr->is_read_only = FALSE; + entry_ptr->ro_ref_count = 0; + entry_ptr->is_pinned = FALSE; + entry_ptr->in_slist = FALSE; + entry_ptr->flush_marker = FALSE; +#ifdef H5_HAVE_PARALLEL + entry_ptr->clear_on_unprotect = FALSE; +#endif /* H5_HAVE_PARALLEL */ + entry_ptr->flush_in_progress = FALSE; + entry_ptr->destroy_in_progress = FALSE; + + entry_ptr->ht_next = NULL; + entry_ptr->ht_prev = NULL; + + entry_ptr->next = NULL; + entry_ptr->prev = NULL; + + entry_ptr->aux_next = NULL; + entry_ptr->aux_prev = NULL; + + H5C2__RESET_CACHE_ENTRY_STATS(entry_ptr); + + ret_value = thing; + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_load_entry() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_make_space_in_cache + * + * Purpose: Attempt to evict cache entries until the index_size + * is at least needed_space below max_cache_size. + * + * In passing, also attempt to bring cLRU_list_size to a + * value greater than min_clean_size. + * + * Depending on circumstances, both of these goals may + * be impossible, as in parallel mode, we must avoid generating + * a write as part of a read (to avoid deadlock in collective + * I/O), and in all cases, it is possible (though hopefully + * highly unlikely) that the protected list may exceed the + * maximum size of the cache. + * + * Thus the function simply does its best, returning success + * unless an error is encountered. + * + * The primary_dxpl_id and secondary_dxpl_id parameters + * specify the dxpl_ids used on the first write occasioned + * by the call (primary_dxpl_id), and on all subsequent + * writes (secondary_dxpl_id). This is useful in the metadata + * cache, but may not be needed elsewhere. If so, just use the + * same dxpl_id for both parameters. + * + * Observe that this function cannot occasion a read. + * + * Return: Non-negative on success/Negative on failure. + * + * Programmer: John Mainzer, 5/14/04 + * + * Modifications: + * + * JRM --7/21/04 + * Minor modifications in support of the addition of a hash + * table to facilitate lookups. + * + * JRM -- 11/22/04 + * Added the first_flush_ptr parameter, which replaces the + * old first_flush local variable. This allows the function + * to coordinate on the first flush issue with other functions. + * + * JRM -- 12/13/04 + * Added code to skip over epoch markers if present. + * + * JRM -- 1/3/06 + * Modified function to work correctly when the the cache + * is not full. This case occurs when we need to flush to + * min clean size before the cache has filled. + * + * JRM -- 3/29/07 + * Added sanity checks using the new is_read_only and + * ro_ref_count fields. + * + * JRM -- 8/24/07 + * Reworked parameter list and code for the removal of the + * secondary dxpl id, and the decision to store the file + * pointer f in *cache_ptr. + * + * JRM -- 10/12/07 + * Added code to detect the case in which the LRU list has + * been modified by a serialize callback, and cause the + * function to re-start its scan at the tail of the LRU. + * + *------------------------------------------------------------------------- + */ + +static herr_t +H5C2_make_space_in_cache(hid_t dxpl_id, + H5C2_t * cache_ptr, + size_t space_needed, + hbool_t write_permitted) +{ + /* const char * fcn_name = "H5C2_make_space_in_cache()"; */ + herr_t ret_value = SUCCEED; /* Return value */ + herr_t result; + int32_t entries_examined = 0; + int32_t initial_list_len; +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + size_t empty_space; +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + hbool_t prev_is_dirty = FALSE; + H5C2_cache_entry_t * entry_ptr; + H5C2_cache_entry_t * prev_ptr; + H5C2_cache_entry_t * next_ptr; + + FUNC_ENTER_NOAPI_NOINIT(H5C2_make_space_in_cache) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + if ( write_permitted ) { + + initial_list_len = cache_ptr->LRU_list_len; + entry_ptr = cache_ptr->LRU_tail_ptr; + + while ( ( (cache_ptr->index_size + space_needed) + > + cache_ptr->max_cache_size + ) + && + ( entries_examined <= (2 * initial_list_len) ) + && + ( entry_ptr != NULL ) + ) + { + HDassert( ! (entry_ptr->is_protected) ); + HDassert( ! (entry_ptr->is_read_only) ); + HDassert( (entry_ptr->ro_ref_count) == 0 ); + + next_ptr = entry_ptr->next; + prev_ptr = entry_ptr->prev; + + if ( prev_ptr != NULL ) { + + prev_is_dirty = prev_ptr->is_dirty; + } + + if ( (entry_ptr->type)->id != H5C2__EPOCH_MARKER_TYPE ) { + + if ( entry_ptr->is_dirty ) { + + result = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__NO_FLAGS_SET, + FALSE); + } else { + + result = + H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__FLUSH_INVALIDATE_FLAG, + TRUE); + } + } else { + + /* Skip epoch markers. Set result to SUCCEED to avoid + * triggering the error code below. + */ + result = SUCCEED; + } + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unable to flush entry 1") + } + + if ( prev_ptr != NULL ) { + + if ( prev_ptr->magic != H5C2__H5C2_CACHE_ENTRY_T_MAGIC ) { + + /* something horrible has happened to *prev_ptr -- + * scream and die. + */ + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "*prev_ptr corrupt 1") + + } else if ( ( prev_ptr->is_dirty != prev_is_dirty ) + || + ( prev_ptr->next != next_ptr ) + || + ( prev_ptr->is_protected ) + || + ( prev_ptr->is_pinned ) ) { + + /* something has happened to the LRU -- start over + * from the tail. + */ + entry_ptr = cache_ptr->LRU_tail_ptr; + + } else { + + entry_ptr = prev_ptr; + + } + } else { + + entry_ptr = NULL; + + } + + entries_examined++; + + } + +#if H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS + + initial_list_len = cache_ptr->dLRU_list_len; + entry_ptr = cache_ptr->dLRU_tail_ptr; + + if ( cache_ptr->index_size < cache_ptr->max_cache_size ) { + + empty_space = cache_ptr->max_cache_size - cache_ptr->index_size; + + } else { + + empty_space = 0; + } + + while ( ( (cache_ptr->cLRU_list_size + empty_space) + < cache_ptr->min_clean_size ) && + ( entries_examined <= initial_list_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 ); + + prev_ptr = entry_ptr->aux_prev; + + next_ptr = entry_ptr->aux_next; + + if ( prev_ptr != NULL ) { + + HDassert( prev_ptr->is_dirty ); + } + + result = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__NO_FLAGS_SET, + FALSE); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unable to flush entry 2") + } + + if ( prev_ptr != NULL ) { + + if (prev_ptr->magic != H5C2__H5C2_CACHE_ENTRY_T_MAGIC) { + + /* something horrible has happened to *prev_ptr -- + * scream and die. + */ + + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "*prev_ptr corrupt 2") + + } else if ( ( ! ( prev_ptr->is_dirty ) ) + || + ( prev_ptr->aux_next != next_ptr ) + || + ( prev_ptr->is_protected ) + || + ( prev_ptr->is_pinned ) ) { + + /* something has happened to the dirty LRU -- start over + * from the tail. + */ + +#if 0 /* This debuging code may be useful in the future -- keep it for now. */ + if ( ! ( prev_ptr->is_dirty ) ) { + HDfprintf(stdout, "%s: ! prev_ptr->is_dirty\n", + fcn_name); + } + if ( prev_ptr->aux_next != next_ptr ) { + HDfprintf(stdout, "%s: prev_ptr->next != next_ptr\n", + fcn_name); + } + if ( prev_ptr->is_protected ) { + HDfprintf(stdout, "%s: prev_ptr->is_protected\n", + fcn_name); + } + if ( prev_ptr->is_pinned ) { + HDfprintf(stdout, "%s:prev_ptr->is_pinned\n", + fcn_name); + } + + HDfprintf(stdout, "%s: re-starting scan of dirty list\n", + fcn_name); +#endif /* JRM */ + entry_ptr = cache_ptr->dLRU_tail_ptr; + + } else { + + entry_ptr = prev_ptr; + + } + } else { + + entry_ptr = NULL; + + } + + entries_examined++; + } + +#endif /* H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS */ + + } else { + + HDassert( H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS ); + + initial_list_len = cache_ptr->cLRU_list_len; + entry_ptr = cache_ptr->cLRU_tail_ptr; + + while ( ( (cache_ptr->index_size + space_needed) + > + cache_ptr->max_cache_size + ) + && + ( entries_examined <= initial_list_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) ); + + prev_ptr = entry_ptr->aux_prev; + + result = H5C2_flush_single_entry(cache_ptr->f, + dxpl_id, + cache_ptr, + entry_ptr->type, + entry_ptr->addr, + H5C2__FLUSH_INVALIDATE_FLAG, + TRUE); + + if ( result < 0 ) { + + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, \ + "unable to flush entry") + } + + /* we are scanning the clean LRU, so the serialize function + * will not be called on any entry -- thus there is no + * concern about the list being modified out from under + * this function. + */ + + entry_ptr = prev_ptr; + entries_examined++; + } + } + +done: + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_make_space_in_cache() */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_validate_lru_list + * + * Purpose: Debugging function that scans the LRU list for errors. + * + * If an error is detected, the function generates a + * diagnostic and returns FAIL. If no error is detected, + * the function returns SUCCEED. + * + * Return: FAIL if error is detected, SUCCEED otherwise. + * + * Programmer: John Mainzer, 7/14/05 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#if H5C2_DO_EXTREME_SANITY_CHECKS + +static herr_t +H5C2_validate_lru_list(H5C2_t * cache_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + int32_t len = 0; + size_t size = 0; + H5C2_cache_entry_t * entry_ptr = NULL; + + FUNC_ENTER_NOAPI_NOINIT(H5C2_validate_lru_list) + + HDassert( cache_ptr ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + + if ( ( ( cache_ptr->LRU_head_ptr == NULL ) + || + ( cache_ptr->LRU_tail_ptr == NULL ) + ) + && + ( cache_ptr->LRU_head_ptr != cache_ptr->LRU_tail_ptr ) + ) { + + HDfprintf(stdout,"H5C2_validate_lru_list: Check 1 failed.\n"); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 1 failed") + } + + if ( ( cache_ptr->LRU_list_len < 0 ) || ( cache_ptr->LRU_list_size < 0 ) ) { + + HDfprintf(stdout,"H5C2_validate_lru_list: Check 2 failed.\n"); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 2 failed") + } + + if ( ( cache_ptr->LRU_list_len == 1 ) + && + ( ( cache_ptr->LRU_head_ptr != cache_ptr->LRU_tail_ptr ) + || + ( cache_ptr->LRU_head_ptr == NULL ) + || + ( cache_ptr->LRU_head_ptr->size != cache_ptr->LRU_list_size ) + ) + ) { + + HDfprintf(stdout,"H5C2_validate_lru_list: Check 3 failed.\n"); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 3 failed") + } + + if ( ( cache_ptr->LRU_list_len >= 1 ) + && + ( ( cache_ptr->LRU_head_ptr == NULL ) + || + ( cache_ptr->LRU_head_ptr->prev != NULL ) + || + ( cache_ptr->LRU_tail_ptr == NULL ) + || + ( cache_ptr->LRU_tail_ptr->next != NULL ) + ) + ) { + + HDfprintf(stdout,"H5C2_validate_lru_list: Check 4 failed.\n"); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 4 failed") + } + + entry_ptr = cache_ptr->LRU_head_ptr; + while ( entry_ptr != NULL ) + { + + if ( ( entry_ptr != cache_ptr->LRU_head_ptr ) && + ( ( entry_ptr->prev == NULL ) || + ( entry_ptr->prev->next != entry_ptr ) ) ) { + + HDfprintf(stdout,"H5C2_validate_lru_list: Check 5 failed.\n"); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 5 failed") + } + + if ( ( entry_ptr != cache_ptr->LRU_tail_ptr ) && + ( ( entry_ptr->next == NULL ) || + ( entry_ptr->next->prev != entry_ptr ) ) ) { + + HDfprintf(stdout,"H5C2_validate_lru_list: Check 6 failed.\n"); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 6 failed") + } + + len++; + size += entry_ptr->size; + entry_ptr = entry_ptr->next; + } + + if ( ( cache_ptr->LRU_list_len != len ) || + ( cache_ptr->LRU_list_size != size ) ) { + + HDfprintf(stdout,"H5C2_validate_lru_list: Check 7 failed.\n"); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, "Check 7 failed") + } + +done: + + if ( ret_value != SUCCEED ) { + + HDassert(0); + } + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_validate_lru_list() */ + +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ + + +/*------------------------------------------------------------------------- + * + * Function: H5C2_verify_not_in_index + * + * Purpose: Debugging function that scans the hash table to verify + * that the specified instance of H5C2_cache_entry_t is not + * present. + * + * If an error is detected, the function generates a + * diagnostic and returns FAIL. If no error is detected, + * the function returns SUCCEED. + * + * Return: FAIL if error is detected, SUCCEED otherwise. + * + * Programmer: John Mainzer, 7/14/05 + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#if H5C2_DO_EXTREME_SANITY_CHECKS + +static herr_t +H5C2_verify_not_in_index(H5C2_t * cache_ptr, + H5C2_cache_entry_t * entry_ptr) +{ + herr_t ret_value = SUCCEED; /* Return value */ + int32_t i; + int32_t depth; + H5C2_cache_entry_t * scan_ptr = NULL; + + FUNC_ENTER_NOAPI_NOINIT(H5C2_verify_not_in_index) + + HDassert( cache_ptr != NULL ); + HDassert( cache_ptr->magic == H5C2__H5C2_T_MAGIC ); + HDassert( entry_ptr != NULL ); + + for ( i = 0; i < H5C2__HASH_TABLE_LEN; i++ ) + { + depth = 0; + scan_ptr = cache_ptr->index[i]; + + while ( scan_ptr != NULL ) + { + if ( scan_ptr == entry_ptr ) { + + HDfprintf(stdout, + "H5C2_verify_not_in_index: entry in index (%d/%d)\n", + i, depth); + HGOTO_ERROR(H5E_CACHE, H5E_SYSTEM, FAIL, \ + "Entry already in index.") + } + depth++; + scan_ptr = scan_ptr->ht_next; + } + } + +done: + + if ( ret_value != SUCCEED ) { + + HDassert(0); + } + + FUNC_LEAVE_NOAPI(ret_value) + +} /* H5C2_verify_not_in_index() */ + +#endif /* H5C2_DO_EXTREME_SANITY_CHECKS */ diff --git a/src/H5C2pkg.h b/src/H5C2pkg.h new file mode 100644 index 0000000..a1753b4 --- /dev/null +++ b/src/H5C2pkg.h @@ -0,0 +1,949 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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. * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +/* + * Programmer: John Mainzer -- 10/12/04 + * + * Purpose: This file contains declarations which are normally visible + * only within the H5C2 package (just H5C2.c at present). + * + * Source files outside the H5C2 package should include + * H5C2private.h instead. + * + * The one exception to this rule is test/cache2.c. The test + * code is easier to write if it can look at the cache's + * internal data structures. Indeed, this is the main + * reason why this file was created. + */ + +#ifndef H5C2_PACKAGE +#error "Do not include this file outside the H5C2 package!" +#endif + +#ifndef _H5C2pkg_H +#define _H5C2pkg_H + + +/* Get package's private header */ +#include "H5C2private.h" + + +/* Get needed headers */ +#include "H5SLprivate.h" /* Skip lists */ + +/* With the introduction of the fractal heap, it is now possible for + * entries to be dirtied, resized, and/or renamed in the flush callbacks. + * As a result, on flushes, it may be necessary to make multiple passes + * through the slist before it is empty. The H5C2__MAX_PASSES_ON_FLUSH + * #define is used to set an upper limit on the number of passes. + * The current value was obtained via personal communication with + * Quincey. I have applied a fudge factor of 2. + */ + +#define H5C2__MAX_PASSES_ON_FLUSH 4 + + +#define H5C2__HASH_TABLE_LEN (64 * 1024) /* must be a power of 2 */ + + +/**************************************************************************** + * + * structure H5C2_t + * + * Catchall structure for all variables specific to an instance of the cache. + * + * While the individual fields of the structure are discussed below, the + * following overview may be helpful. + * + * Entries in the cache are stored in an instance of H5TB_TREE, indexed on + * the entry's disk address. While the H5TB_TREE is less efficient than + * hash table, it keeps the entries in address sorted order. As flushes + * in parallel mode are more efficient if they are issued in increasing + * address order, this is a significant benefit. Also the H5TB_TREE code + * was readily available, which reduced development time. + * + * While the cache was designed with multiple replacement policies in mind, + * at present only a modified form of LRU is supported. + * + * JRM - 4/26/04 + * + * Profiling has indicated that searches in the instance of H5TB_TREE are + * too expensive. To deal with this issue, I have augmented the cache + * with a hash table in which all entries will be stored. Given the + * advantages of flushing entries in increasing address order, the TBBT + * is retained, but only dirty entries are stored in it. At least for + * now, we will leave entries in the TBBT after they are flushed. + * + * Note that index_size and index_len now refer to the total size of + * and number of entries in the hash table. + * + * JRM - 7/19/04 + * + * The TBBT has since been replaced with a skip list. This change + * greatly predates this note. + * + * JRM - 9/26/05 + * + * magic: Unsigned 32 bit integer always set to H5C2__H5C2_T_MAGIC. + * This field is used to validate pointers to instances of + * H5C2_t. + * + * f: Pointer to the instance of H5F_t associated with this + * instance of the metadata cache. This field is set at + * create, and then used until the file is closed (at which + * point, the cache will be shut down as well). + * + * flush_in_progress: Boolean flag indicating whether a flush is in + * progress. + * + * trace_file_ptr: File pointer pointing to the trace file, which is used + * to record cache operations for use in simulations and design + * studies. This field will usually be NULL, indicating that + * no trace file should be recorded. + * + * Since much of the code supporting the parallel metadata + * cache is in H5AC, we don't write the trace file from + * H5C2. Instead, H5AC reads the trace_file_ptr as needed. + * + * When we get to using H5C2 in other places, we may add + * code to write trace file data at the H5C2 level as well. + * + * aux_ptr: Pointer to void used to allow wrapper code to associate + * its data with an instance of H5C2_t. The H5C2 cache code + * sets this field to NULL, and otherwise leaves it alone. + * + * max_type_id: Integer field containing the maximum type id number assigned + * to a type of entry in the cache. All type ids from 0 to + * max_type_id inclusive must be defined. The names of the + * types are stored in the type_name_table discussed below, and + * indexed by the ids. + * + * type_name_table_ptr: Pointer to an array of pointer to char of length + * max_type_id + 1. The strings pointed to by the entries + * in the array are the names of the entry types associated + * with the indexing type IDs. + * + * max_cache_size: Nominal maximum number of bytes that may be stored in the + * cache. This value should be viewed as a soft limit, as the + * cache can exceed this value under the following circumstances: + * + * a) All entries in the cache are protected, and the cache is + * asked to insert a new entry. In this case the new entry + * will be created. If this causes the cache to exceed + * max_cache_size, it will do so. The cache will attempt + * to reduce its size as entries are unprotected. + * + * b) When running in parallel mode, the cache may not be + * permitted to flush a dirty entry in response to a read. + * If there are no clean entries available to evict, the + * cache will exceed its maximum size. Again the cache + * will attempt to reduce its size to the max_cache_size + * limit on the next cache write. + * + * c) When an entry increases in size, the cache may exceed + * the max_cache_size limit until the next time the cache + * attempts to load or insert an entry. + * + * min_clean_size: Nominal minimum number of clean bytes in the cache. + * The cache attempts to maintain this number of bytes of + * clean data so as to avoid case b) above. Again, this is + * a soft limit. + * + * + * In addition to the call back functions required for each entry, the + * cache requires the following call back functions for this instance of + * the cache as a whole: + * + * check_write_permitted: In certain applications, the cache may not + * be allowed to write to disk at certain time. If specified, + * the check_write_permitted function is used to determine if + * a write is permissible at any given point in time. + * + * If no such function is specified (i.e. this field is NULL), + * the cache uses the following write_permitted field to + * determine whether writes are permitted. + * + * write_permitted: If check_write_permitted is NULL, this boolean flag + * indicates whether writes are permitted. + * + * log_flush: If provided, this function is called whenever a dirty + * entry is flushed to disk. + * + * + * In cases where memory is plentiful, and performance is an issue, it + * is useful to disable all cache evictions, and thereby postpone metadata + * writes. The following field is used to implement this. + * + * evictions_enabled: Boolean flag that is initialized to TRUE. When + * this flag is set to FALSE, the metadata cache will not + * attempt to evict entries to make space for newly protected + * entries, and instead the will grow without limit. + * + * Needless to say, this feature must be used with care. + * + * + * The cache requires an index to facilitate searching for entries. The + * following fields support that index. + * + * index_len: Number of entries currently in the hash table used to index + * the cache. + * + * index_size: Number of bytes of cache entries currently stored in the + * hash table used to index the cache. + * + * This value should not be mistaken for footprint of the + * cache in memory. The average cache entry is small, and + * the cache has a considerable overhead. Multiplying the + * index_size by two should yield a conservative estimate + * of the cache's memory footprint. + * + * index: Array of pointer to H5C2_cache_entry_t of size + * H5C2__HASH_TABLE_LEN. At present, this value is a power + * of two, not the usual prime number. + * + * I hope that the variable size of cache elements, the large + * hash table size, and the way in which HDF5 allocates space + * will combine to avoid problems with periodicity. If so, we + * can use a trivial hash function (a bit-and and a 3 bit left + * shift) with some small savings. + * + * If not, it will become evident in the statistics. Changing + * to the usual prime number length hash table will require + * changing the H5C2__HASH_FCN macro and the deletion of the + * H5C2__HASH_MASK #define. No other changes should be required. + * + * + * When we flush the cache, we need to write entries out in increasing + * address order. An instance of a skip list is used to store dirty entries in + * sorted order. Whether it is cheaper to sort the dirty entries as needed, + * or to maintain the list is an open question. At a guess, it depends + * on how frequently the cache is flushed. We will see how it goes. + * + * For now at least, I will not remove dirty entries from the list as they + * are flushed. (this has been changed -- dirty entries are now removed from + * the skip list as they are flushed. JRM - 10/25/05) + * + * slist_len: Number of entries currently in the skip list + * used to maintain a sorted list of dirty entries in the + * cache. + * + * slist_size: Number of bytes of cache entries currently stored in the + * skip list used to maintain a sorted list of + * dirty entries in the cache. + * + * slist_ptr: pointer to the instance of H5SL_t used maintain a sorted + * list of dirty entries in the cache. This sorted list has + * two uses: + * + * a) It allows us to flush dirty entries in increasing address + * order, which results in significant savings. + * + * b) It facilitates checking for adjacent dirty entries when + * attempting to evict entries from the cache. While we + * don't use this at present, I hope that this will allow + * some optimizations when I get to it. + * + * With the addition of the fractal heap, the cache must now deal with + * the case in which entries may be dirtied, renamed, or have their sizes + * changed during a flush. To allow sanity checks in this situation, the + * following two fields have been added. They are only compiled in when + * H5C2_DO_SANITY_CHECKS is TRUE. + * + * slist_len_increase: Number of entries that have been added to the + * slist since the last time this field was set to zero. + * + * slist_size_increase: Total size of all entries that have been added + * to the slist since the last time this field was set to + * zero. + * + * + * When a cache entry is protected, it must be removed from the LRU + * list(s) as it cannot be either flushed or evicted until it is unprotected. + * The following fields are used to implement the protected list (pl). + * + * pl_len: Number of entries currently residing on the protected list. + * + * pl_size: Number of bytes of cache entries currently residing on the + * protected list. + * + * pl_head_ptr: Pointer to the head of the doubly linked list of protected + * entries. Note that cache entries on this list are linked + * by their next and prev fields. + * + * This field is NULL if the list is empty. + * + * pl_tail_ptr: Pointer to the tail of the doubly linked list of protected + * entries. Note that cache entries on this list are linked + * by their next and prev fields. + * + * This field is NULL if the list is empty. + * + * + * For very frequently used entries, the protect/unprotect overhead can + * become burdensome. To avoid this overhead, I have modified the cache + * to allow entries to be "pinned". A pinned entry is similar to a + * protected entry, in the sense that it cannot be evicted, and that + * the entry can be modified at any time. + * + * Pinning an entry has the following implications: + * + * 1) A pinned entry cannot be evicted. Thus unprotected + * pinned entries reside in the pinned entry list, instead + * of the LRU list(s) (or other lists maintained by the current + * replacement policy code). + * + * 2) A pinned entry can be accessed or modified at any time. + * Therefore, the cache must check with the entry owner + * before flushing it. If permission is denied, the + * cache just skips the entry in the flush. + * + * 3) A pinned entry can be marked as dirty (and possibly + * change size) while it is unprotected. + * + * 4) The flush-destroy code must allow pinned entries to + * be unpinned (and possibly unprotected) during the + * flush. + * + * Since pinned entries cannot be evicted, they must be kept on a pinned + * entry list, instead of being entrusted to the replacement policy code. + * + * Maintaining the pinned entry list requires the following fields: + * + * pel_len: Number of entries currently residing on the pinned + * entry list. + * + * pel_size: Number of bytes of cache entries currently residing on + * the pinned entry list. + * + * pel_head_ptr: Pointer to the head of the doubly linked list of pinned + * but not protected entries. Note that cache entries on + * this list are linked by their next and prev fields. + * + * This field is NULL if the list is empty. + * + * pel_tail_ptr: Pointer to the tail of the doubly linked list of pinned + * but not protected entries. Note that cache entries on + * this list are linked by their next and prev fields. + * + * This field is NULL if the list is empty. + * + * + * The cache must have a replacement policy, and the fields supporting this + * policy must be accessible from this structure. + * + * While there has been interest in several replacement policies for + * this cache, the initial development schedule is tight. Thus I have + * elected to support only a modified LRU policy for the first cut. + * + * To further simplify matters, I have simply included the fields needed + * by the modified LRU in this structure. When and if we add support for + * other policies, it will probably be easiest to just add the necessary + * fields to this structure as well -- we only create one instance of this + * structure per file, so the overhead is not excessive. + * + * + * Fields supporting the modified LRU policy: + * + * See most any OS text for a discussion of the LRU replacement policy. + * + * When operating in parallel mode, we must ensure that a read does not + * cause a write. If it does, the process will hang, as the write will + * be collective and the other processes will not know to participate. + * + * To deal with this issue, I have modified the usual LRU policy by adding + * clean and dirty LRU lists to the usual LRU list. + * + * The clean LRU list is simply the regular LRU list with all dirty cache + * entries removed. + * + * Similarly, the dirty LRU list is the regular LRU list with all the clean + * cache entries removed. + * + * When reading in parallel mode, we evict from the clean LRU list only. + * This implies that we must try to ensure that the clean LRU list is + * reasonably well stocked at all times. + * + * We attempt to do this by trying to flush enough entries on each write + * to keep the cLRU_list_size >= min_clean_size. + * + * Even if we start with a completely clean cache, a sequence of protects + * without unprotects can empty the clean LRU list. In this case, the + * cache must grow temporarily. At the next write, we will attempt to + * evict enough entries to reduce index_size to less than max_cache_size. + * While this will usually be possible, all bets are off if enough entries + * are protected. + * + * Discussions of the individual fields used by the modified LRU replacement + * policy follow: + * + * LRU_list_len: Number of cache entries currently on the LRU list. + * + * Observe that LRU_list_len + pl_len must always equal + * index_len. + * + * LRU_list_size: Number of bytes of cache entries currently residing on the + * LRU list. + * + * Observe that LRU_list_size + pl_size must always equal + * index_size. + * + * LRU_head_ptr: Pointer to the head of the doubly linked LRU list. Cache + * entries on this list are linked by their next and prev fields. + * + * This field is NULL if the list is empty. + * + * LRU_tail_ptr: Pointer to the tail of the doubly linked LRU list. Cache + * entries on this list are linked by their next and prev fields. + * + * This field is NULL if the list is empty. + * + * cLRU_list_len: Number of cache entries currently on the clean LRU list. + * + * Observe that cLRU_list_len + dLRU_list_len must always + * equal LRU_list_len. + * + * cLRU_list_size: Number of bytes of cache entries currently residing on + * the clean LRU list. + * + * Observe that cLRU_list_size + dLRU_list_size must always + * equal LRU_list_size. + * + * cLRU_head_ptr: Pointer to the head of the doubly linked clean LRU list. + * Cache entries on this list are linked by their aux_next and + * aux_prev fields. + * + * This field is NULL if the list is empty. + * + * cLRU_tail_ptr: Pointer to the tail of the doubly linked clean LRU list. + * Cache entries on this list are linked by their aux_next and + * aux_prev fields. + * + * This field is NULL if the list is empty. + * + * dLRU_list_len: Number of cache entries currently on the dirty LRU list. + * + * Observe that cLRU_list_len + dLRU_list_len must always + * equal LRU_list_len. + * + * dLRU_list_size: Number of cache entries currently on the dirty LRU list. + * + * Observe that cLRU_list_len + dLRU_list_len must always + * equal LRU_list_len. + * + * dLRU_head_ptr: Pointer to the head of the doubly linked dirty LRU list. + * Cache entries on this list are linked by their aux_next and + * aux_prev fields. + * + * This field is NULL if the list is empty. + * + * dLRU_tail_ptr: Pointer to the tail of the doubly linked dirty LRU list. + * Cache entries on this list are linked by their aux_next and + * aux_prev fields. + * + * This field is NULL if the list is empty. + * + * + * Automatic cache size adjustment: + * + * While the default cache size is adequate for most cases, we can run into + * cases where the default is too small. Ideally, we will let the user + * adjust the cache size as required. However, this is not possible in all + * cases. Thus I have added automatic cache size adjustment code. + * + * The configuration for the automatic cache size adjustment is stored in + * the structure described below: + * + * size_increase_possible: Depending on the configuration data given + * in the resize_ctl field, it may or may not be possible + * to increase the size of the cache. Rather than test for + * all the ways this can happen, we simply set this flag when + * we receive a new configuration. + * + * size_decrease_possible: Depending on the configuration data given + * in the resize_ctl field, it may or may not be possible + * to decrease the size of the cache. Rather than test for + * all the ways this can happen, we simply set this flag when + * we receive a new configuration. + * + * cache_full: Boolean flag used to keep track of whether the cache is + * full, so we can refrain from increasing the size of a + * cache which hasn't used up the space alotted to it. + * + * The field is initialized to FALSE, and then set to TRUE + * whenever we attempt to make space in the cache. + * + * resize_enabled: This is another convenience flag which is set whenever + * a new set of values for resize_ctl are provided. Very + * simply, + * + * resize_enabled = size_increase_possible || + * size_decrease_possible; + * + * size_decreased: Boolean flag set to TRUE whenever the maximun cache + * size is decreased. The flag triggers a call to + * H5C2_make_space_in_cache() on the next call to H5C2_protect(). + * + * resize_ctl: Instance of H5C2_auto_size_ctl_t containing configuration + * data for automatic cache resizing. + * + * epoch_markers_active: Integer field containing the number of epoch + * markers currently in use in the LRU list. This value + * must be in the range [0, H5C2__MAX_EPOCH_MARKERS - 1]. + * + * epoch_marker_active: Array of boolean of length H5C2__MAX_EPOCH_MARKERS. + * This array is used to track which epoch markers are currently + * in use. + * + * epoch_marker_ringbuf: Array of int of length H5C2__MAX_EPOCH_MARKERS + 1. + * + * To manage the epoch marker cache entries, it is necessary + * to track their order in the LRU list. This is done with + * epoch_marker_ringbuf. When markers are inserted at the + * head of the LRU list, the index of the marker in the + * epoch_markers array is inserted at the tail of the ring + * buffer. When it becomes the epoch_marker_active'th marker + * in the LRU list, it will have worked its way to the head + * of the ring buffer as well. This allows us to remove it + * without scanning the LRU list if such is required. + * + * epoch_marker_ringbuf_first: Integer field containing the index of the + * first entry in the ring buffer. + * + * epoch_marker_ringbuf_last: Integer field containing the index of the + * last entry in the ring buffer. + * + * epoch_marker_ringbuf_size: Integer field containing the number of entries + * in the ring buffer. + * + * epoch_markers: Array of instances of H5C2_cache_entry_t of length + * H5C2__MAX_EPOCH_MARKERS. The entries are used as markers + * in the LRU list to identify cache entries that haven't + * been accessed for some (small) specified number of + * epochs. These entries (if any) can then be evicted and + * the cache size reduced -- ideally without evicting any + * of the current working set. Needless to say, the epoch + * length and the number of epochs before an unused entry + * must be chosen so that all, or almost all, the working + * set will be accessed before the limit. + * + * Epoch markers only appear in the LRU list, never in + * the index or slist. While they are of type + * H5C2__EPOCH_MARKER_TYPE, and have associated class + * functions, these functions should never be called. + * + * The addr fields of these instances of H5C2_cache_entry_t + * are set to the index of the instance in the epoch_markers + * array, the size is set to 0, and the type field points + * to the constant structure epoch_marker_class defined + * in H5C2.c. The next and prev fields are used as usual + * to link the entry into the LRU list. + * + * All other fields are unused. + * + * + * Cache hit rate collection fields: + * + * We supply the current cache hit rate on request, so we must keep a + * simple cache hit rate computation regardless of whether statistics + * collection is enabled. The following fields support this capability. + * + * cache_hits: Number of cache hits since the last time the cache hit + * rate statistics were reset. Note that when automatic cache + * re-sizing is enabled, this field will be reset every automatic + * resize epoch. + * + * cache_accesses: Number of times the cache has been accessed while + * since the last since the last time the cache hit rate statistics + * were reset. Note that when automatic cache re-sizing is enabled, + * this field will be reset every automatic resize epoch. + * + * + * Statistics collection fields: + * + * When enabled, these fields are used to collect statistics as described + * below. The first set are collected only when H5C2_COLLECT_CACHE_STATS + * is true. + * + * hits: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type id + * equal to the array index has been in cache when requested in + * the current epoch. + * + * misses: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type id + * equal to the array index has not been in cache when + * requested in the current epoch. + * + * write_protects: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The + * cells are used to record the number of times an entry with + * type id equal to the array index has been write protected + * in the current epoch. + * + * Observe that (hits + misses) = (write_protects + read_protects). + * + * read_protects: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The + * cells are used to record the number of times an entry with + * type id equal to the array index has been read protected in + * the current epoch. + * + * Observe that (hits + misses) = (write_protects + read_protects). + * + * max_read_protects: Array of int32 of length H5C2__MAX_NUM_TYPE_IDS + 1. + * The cells are used to maximum number of simultaneous read + * protects on any entry with type id equal to the array index + * in the current epoch. + * + * insertions: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type + * id equal to the array index has been inserted into the + * cache in the current epoch. + * + * pinned_insertions: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. + * The cells are used to record the number of times an entry + * with type id equal to the array index has been inserted + * pinned into the cache in the current epoch. + * + * clears: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type + * id equal to the array index has been cleared in the current + * epoch. + * + * flushes: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type id + * equal to the array index has been written to disk in the + * current epoch. + * + * evictions: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type id + * equal to the array index has been evicted from the cache in + * the current epoch. + * + * renames: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type + * id equal to the array index has been renamed in the current + * epoch. + * + * entry_flush_renames: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. + * The cells are used to record the number of times an entry + * with type id equal to the array index has been renamed + * during its flush callback in the current epoch. + * + * cache_flush_renames: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. + * The cells are used to record the number of times an entry + * with type id equal to the array index has been renamed + * during a cache flush in the current epoch. + * + * pins: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type + * id equal to the array index has been pinned in the current + * epoch. + * + * unpins: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type + * id equal to the array index has been unpinned in the current + * epoch. + * + * dirty_pins: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the number of times an entry with type + * id equal to the array index has been marked dirty while pinned + * in the current epoch. + * + * pinned_flushes: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The + * cells are used to record the number of times an entry + * with type id equal to the array index has been flushed while + * pinned in the current epoch. + * + * pinned_cleared: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. The + * cells are used to record the number of times an entry + * with type id equal to the array index has been cleared while + * pinned in the current epoch. + * + * size_increases: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. + * The cells are used to record the number of times an entry + * with type id equal to the array index has increased in + * size in the current epoch. + * + * size_decreases: Array of int64 of length H5C2__MAX_NUM_TYPE_IDS + 1. + * The cells are used to record the number of times an entry + * with type id equal to the array index has decreased in + * size in the current epoch. + * + * entry_flush_size_changes: Array of int64 of length + * H5C2__MAX_NUM_TYPE_IDS + 1. The cells are used to record + * the number of times an entry with type id equal to the + * array index has changed size while in its flush callback. + * + * cache_flush_size_changes: Array of int64 of length + * H5C2__MAX_NUM_TYPE_IDS + 1. The cells are used to record + * the number of times an entry with type id equal to the + * array index has changed size during a cache flush + * + * total_ht_insertions: Number of times entries have been inserted into the + * hash table in the current epoch. + * + * total_ht_deletions: Number of times entries have been deleted from the + * hash table in the current epoch. + * + * successful_ht_searches: int64 containing the total number of successful + * searches of the hash table in the current epoch. + * + * total_successful_ht_search_depth: int64 containing the total number of + * entries other than the targets examined in successful + * searches of the hash table in the current epoch. + * + * failed_ht_searches: int64 containing the total number of unsuccessful + * searches of the hash table in the current epoch. + * + * total_failed_ht_search_depth: int64 containing the total number of + * entries examined in unsuccessful searches of the hash + * table in the current epoch. + * + * max_index_len: Largest value attained by the index_len field in the + * current epoch. + * + * max_index_size: Largest value attained by the index_size field in the + * current epoch. + * + * max_slist_len: Largest value attained by the slist_len field in the + * current epoch. + * + * max_slist_size: Largest value attained by the slist_size field in the + * current epoch. + * + * max_pl_len: Largest value attained by the pl_len field in the + * current epoch. + * + * max_pl_size: Largest value attained by the pl_size field in the + * current epoch. + * + * max_pel_len: Largest value attained by the pel_len field in the + * current epoch. + * + * max_pel_size: Largest value attained by the pel_size field in the + * current epoch. + * + * The remaining stats are collected only when both H5C2_COLLECT_CACHE_STATS + * and H5C2_COLLECT_CACHE_ENTRY_STATS are true. + * + * max_accesses: Array of int32 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the maximum number of times any single + * entry with type id equal to the array index has been + * accessed in the current epoch. + * + * min_accesses: Array of int32 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the minimum number of times any single + * entry with type id equal to the array index has been + * accessed in the current epoch. + * + * max_clears: Array of int32 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the maximum number of times any single + * entry with type id equal to the array index has been cleared + * in the current epoch. + * + * max_flushes: Array of int32 of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the maximum number of times any single + * entry with type id equal to the array index has been + * flushed in the current epoch. + * + * max_size: Array of size_t of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the maximum size of any single entry + * with type id equal to the array index that has resided in + * the cache in the current epoch. + * + * max_pins: Array of size_t of length H5C2__MAX_NUM_TYPE_IDS + 1. The cells + * are used to record the maximum number of times that any single + * entry with type id equal to the array index that has been + * marked as pinned in the cache in the current epoch. + * + * + * Fields supporting testing: + * + * For test purposes, it is useful to turn off some asserts and sanity + * checks. The following flags support this. + * + * skip_file_checks: Boolean flag used to skip sanity checks on file + * parameters passed to the cache. In the test bed, there + * is no reason to have a file open, as the cache proper + * just passes these parameters through without using them. + * + * When this flag is set, all sanity checks on the file + * parameters are skipped. The field defaults to FALSE. + * + * skip_dxpl_id_checks: Boolean flag used to skip sanity checks on the + * dxpl_id parameters passed to the cache. These are not + * used directly by the cache, so skipping the checks + * simplifies the test bed. + * + * When this flag is set, all sanity checks on the dxpl_id + * parameters are skipped. The field defaults to FALSE. + * + * prefix Array of char used to prefix debugging output. The + * field is intended to allow marking of output of with + * the processes mpi rank. + * + ****************************************************************************/ + +#define H5C2__H5C2_T_MAGIC 0x005CAC0F +#define H5C2__MAX_NUM_TYPE_IDS 16 +#define H5C2__PREFIX_LEN 32 + +struct H5C2_t +{ + uint32_t magic; + + H5F_t * f; + + hbool_t flush_in_progress; + + FILE * trace_file_ptr; + + void * aux_ptr; + + int32_t max_type_id; + const char * (* type_name_table_ptr); + + size_t max_cache_size; + size_t min_clean_size; + + H5C2_write_permitted_func_t check_write_permitted; + hbool_t write_permitted; + + H5C2_log_flush_func_t log_flush; + + hbool_t evictions_enabled; + + int32_t index_len; + size_t index_size; + H5C2_cache_entry_t * (index[H5C2__HASH_TABLE_LEN]); + + + int32_t slist_len; + size_t slist_size; + H5SL_t * slist_ptr; +#if H5C2_DO_SANITY_CHECKS + int64_t slist_len_increase; + int64_t slist_size_increase; +#endif /* H5C2_DO_SANITY_CHECKS */ + + int32_t pl_len; + size_t pl_size; + H5C2_cache_entry_t * pl_head_ptr; + H5C2_cache_entry_t * pl_tail_ptr; + + int32_t pel_len; + size_t pel_size; + H5C2_cache_entry_t * pel_head_ptr; + H5C2_cache_entry_t * pel_tail_ptr; + + int32_t LRU_list_len; + size_t LRU_list_size; + H5C2_cache_entry_t * LRU_head_ptr; + H5C2_cache_entry_t * LRU_tail_ptr; + + int32_t cLRU_list_len; + size_t cLRU_list_size; + H5C2_cache_entry_t * cLRU_head_ptr; + H5C2_cache_entry_t * cLRU_tail_ptr; + + int32_t dLRU_list_len; + size_t dLRU_list_size; + H5C2_cache_entry_t * dLRU_head_ptr; + H5C2_cache_entry_t * dLRU_tail_ptr; + + hbool_t size_increase_possible; + hbool_t size_decrease_possible; + hbool_t resize_enabled; + hbool_t cache_full; + hbool_t size_decreased; + H5C2_auto_size_ctl_t resize_ctl; + + int32_t epoch_markers_active; + hbool_t epoch_marker_active[H5C2__MAX_EPOCH_MARKERS]; + int32_t epoch_marker_ringbuf[H5C2__MAX_EPOCH_MARKERS+1]; + int32_t epoch_marker_ringbuf_first; + int32_t epoch_marker_ringbuf_last; + int32_t epoch_marker_ringbuf_size; + H5C2_cache_entry_t epoch_markers[H5C2__MAX_EPOCH_MARKERS]; + + int64_t cache_hits; + int64_t cache_accesses; + +#if H5C2_COLLECT_CACHE_STATS + + /* stats fields */ + int64_t hits[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t misses[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t write_protects[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t read_protects[H5C2__MAX_NUM_TYPE_IDS + 1]; + int32_t max_read_protects[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t insertions[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t pinned_insertions[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t clears[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t flushes[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t evictions[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t renames[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t entry_flush_renames[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t cache_flush_renames[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t pins[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t unpins[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t dirty_pins[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t pinned_flushes[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t pinned_clears[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t size_increases[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t size_decreases[H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t entry_flush_size_changes + [H5C2__MAX_NUM_TYPE_IDS + 1]; + int64_t cache_flush_size_changes + [H5C2__MAX_NUM_TYPE_IDS + 1]; + + int64_t total_ht_insertions; + int64_t total_ht_deletions; + int64_t successful_ht_searches; + int64_t total_successful_ht_search_depth; + int64_t failed_ht_searches; + int64_t total_failed_ht_search_depth; + + int32_t max_index_len; + size_t max_index_size; + + int32_t max_slist_len; + size_t max_slist_size; + + int32_t max_pl_len; + size_t max_pl_size; + + int32_t max_pel_len; + size_t max_pel_size; + +#if H5C2_COLLECT_CACHE_ENTRY_STATS + + int32_t max_accesses[H5C2__MAX_NUM_TYPE_IDS + 1]; + int32_t min_accesses[H5C2__MAX_NUM_TYPE_IDS + 1]; + int32_t max_clears[H5C2__MAX_NUM_TYPE_IDS + 1]; + int32_t max_flushes[H5C2__MAX_NUM_TYPE_IDS + 1]; + size_t max_size[H5C2__MAX_NUM_TYPE_IDS + 1]; + int32_t max_pins[H5C2__MAX_NUM_TYPE_IDS + 1]; + +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + +#endif /* H5C2_COLLECT_CACHE_STATS */ + + hbool_t skip_file_checks; + hbool_t skip_dxpl_id_checks; + char prefix[H5C2__PREFIX_LEN]; +}; + +#endif /* _H5C2pkg_H */ + diff --git a/src/H5C2private.h b/src/H5C2private.h new file mode 100644 index 0000000..a949f6d --- /dev/null +++ b/src/H5C2private.h @@ -0,0 +1,1350 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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: H5C2private.h + * 6/3/04 + * John Mainzer + * + * Purpose: Constants and typedefs available to the rest of the + * library. + * + * Modifications: + * + *------------------------------------------------------------------------- + */ + +#ifndef _H5C2private_H +#define _H5C2private_H + +#include "H5C2public.h" /* public prototypes */ + +/* Pivate headers needed by this header */ +#include "H5private.h" /* Generic Functions */ +#include "H5Fprivate.h" /* File access */ + + +#define H5C2_DO_SANITY_CHECKS 1 +#define H5C2_DO_EXTREME_SANITY_CHECKS 0 + +/* This sanity checking constant was picked out of the air. Increase + * or decrease it if appropriate. Its purposes is to detect corrupt + * object sizes, so it probably doesn't matter if it is a bit big. + * + * JRM - 5/17/04 + */ +#define H5C2_MAX_ENTRY_SIZE ((size_t)(10 * 1024 * 1024)) + +/* H5C2_COLLECT_CACHE_STATS controls overall collection of statistics + * on cache activity. In general, this #define should be set to 0. + */ +#define H5C2_COLLECT_CACHE_STATS 1 + +/* H5C2_COLLECT_CACHE_ENTRY_STATS controls collection of statistics + * in individual cache entries. + * + * H5C2_COLLECT_CACHE_ENTRY_STATS should only be defined to true if + * H5C2_COLLECT_CACHE_STATS is also defined to true. + */ +#if H5C2_COLLECT_CACHE_STATS + +#define H5C2_COLLECT_CACHE_ENTRY_STATS 1 + +#else + +#define H5C2_COLLECT_CACHE_ENTRY_STATS 0 + +#endif /* H5C2_COLLECT_CACHE_STATS */ + + +#ifdef H5_HAVE_PARALLEL + +/* we must maintain the clean and dirty LRU lists when we are compiled + * with parallel support. + */ +#define H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS 1 + +#else /* H5_HAVE_PARALLEL */ + +/* The clean and dirty LRU lists don't buy us anything here -- we may + * want them on for testing on occasion, but in general they should be + * off. + */ +#define H5C2_MAINTAIN_CLEAN_AND_DIRTY_LRU_LISTS 0 + +#endif /* H5_HAVE_PARALLEL */ + + +/* Typedef for the main structure for the cache (defined in H5C2pkg.h) */ + +typedef struct H5C2_t H5C2_t; + + +/*************************************************************************** + * + * Struct H5C2_class_t + * + * Instances of H5C2_class_t are used to specify the callback functions + * used by the metadata cache for each class of metadata cache entry. + * The fields of the structure are discussed below: + * + * id: Integer field containing the unique ID of the class of metadata + * cache entries. + * + * name: Pointer to a string containing the name of the class of metadata + * cache entries. + * + * mem_type: Instance of H5FD_mem_t, that is used to supply the + * mem type passed into H5F_block_read(). + * + * deserialize: Pointer to the deserialize function. + * + * This function must be able to read an on disk image of a metadata + * cache entry, allocate and load the equivalent in core representation, + * and return a pointer to that representation. + * + * The typedef for the deserialize callback is as follows: + * + * typedef void *(*H5C_deserialize_func_t)(haddr_t addr, + * size_t len, + * const void * image_ptr, + * void * udata_ptr, + * boolean * dirty_ptr); + * + * The parameters of the deserialize callback are as follows: + * + * addr: Base address in file of the image to be deserialized. + * + * This parameter is supplied mainly for sanity checking. + * Sanity checks should be performed when compiled in debug + * mode, but the parameter may be unused when compiled in + * production mode. + * + * len: Length in bytes of the in file image to be deserialized. + * + * This parameter is supplied mainly for sanity checking. + * Sanity checks should be performed when compiled in debug + * mode, but the parameter may be unused when compiled in + * production mode. + * + * image_ptr: Pointer to a buffer of length len containing the + * contents of the file starting at addr and continuing + * for len bytes. + * + * udata_ptr: Pointer to user data provided in the protect call, which + * must be passed through to the deserialize callback. + * + * dirty_ptr: Pointer to boolean which the deserialize function + * must use to mark the entry dirty if it has to modify + * the entry to clean up file corruption left over from + * an old bug in the HDF5 library. + * + * Processing in the deserialize function should proceed as follows: + * + * If the image contains valid data, and is of the correct length, + * the deserialize function must allocate space for an in core + * represntation of that data, load the contents of the image into + * the space allocated for the in core representation, and return + * a pointer to the in core representation. Observe that an + * instance of H5C2_cache_entry_t must be the first item in this + * representation. It will have to be initialized appropriately + * after the callback returns. + * + * Note that the structure of the in core representation is otherwise + * up to the cache client. All that is required is that the pointer + * returned be sufficient for the clients purposes when it is returned + * on a protect call. + * + * If the deserialize function has to clean up file corruption + * left over from an old bug in the HDF5 library, it must set + * *dirty_ptr to TRUE. If it doesn't, no action is needed as + * *dirty_ptr will be set to FALSE before the deserialize call. + * + * If the operation fails for any reason (i.e. bad data in buffer, bad + * buffer length, malloc failure, etc.) the function must return NULL and + * push error information on the error stack with the error API routines. + * + * If the protect call which occasioned the call to the deserialize + * callback had the check length flag set, after the deserialize call + * returns, the cache must call the image_len callback (see below) and + * update its on disk image length accordingly. + * + * + * image_len: Pointer to the image length callback. + * + * In the best of all possible worlds, we would not have this callback. + * It exists to allow clients to reduce the size of the on disk image of + * an entry in the deserialize callback. + * + * The typedef for the image_len callback is as follows: + * + * typedef herr_t (*H5C_image_len_func_t)(void *thing, + * size_t *image_len_ptr); + * + * The parameters of the image_len callback are as follows: + * + * thing: Pointer to the in core representation of the entry. + * + * image_len_ptr: Pointer to size_t in which the callback will return + * the length of the on disk image of the cache entry. + * + * Processing in the image_len function should proceed as follows: + * + * If successful, the function will place the length of the on disk + * image associated with the in core representation provided in the + * thing parameter in *image_len_ptr, and then return SUCCEED. + * + * On failure, the function must return FAIL and push error information + * onto the error stack with the error API routines. + * + * + * serialize: Pointer to the serialize callback. + * + * The serialize callback is invoked by the metadata cache whenever + * it needs a current on disk image of the metadata entry for purposes + * either constructing a journal or flushing the entry to disk. + * + * At this point, one would think that the base address and length of + * the length of the entry's image on disk would be well known. + * However, that need not be the case as fractal heap blocks can + * change size (and therefor possible location as well) on + * serialization if compression is enabled. In the old H5C code, + * this happened on a flush, and occasioned a rename in the midst + * of the flush. To avoid this in H5C2, the serialize callback + * will return the new base address, length, and image pointer to + * the caller when necessary. The caller must then update the + * metadata cache's internal structures accordingly. + * + * The typedef for the serialize callback is as follows: + * + * typedef herr_t (*H5C_serialize_func_t)(haddr_t addr, + * size_t len, + * void * image_ptr, + * void * thing, + * unsigned * flags_ptr, + * haddr_t * new_addr_ptr, + * size_t * new_len_ptr, + * void ** new_image_ptr_ptr); + * + * The parameters of the serialize callback are as follows: + * + * addr: Base address in file of the entry to be serialized. + * + * This parameter is supplied mainly for sanity checking. + * Sanity checks should be performed when compiled in debug + * mode, but the parameter may be unused when compiled in + * production mode. + * + * len: Length in bytes of the in file image of the entry to be + * serialized. Also the size of *image_ptr (below). + * + * This parameter is supplied mainly for sanity checking. + * Sanity checks should be performed when compiled in debug + * mode, but the parameter may be unused when compiled in + * production mode. + * + * image_ptr: Pointer to a buffer of length len bytes into which a + * serialized image of the target metadata cache entry is + * to be written. + * + * Note that this buffer will not in general be initialized + * to any particular value. Thus the serialize function may + * not assume any initial value and must set each byte in + * the buffer. + * + * thing: Pointer to void containing the address of the in core + * representation of the target metadata cache entry. + * This is the same pointer returned by a protect of the + * addr and len given above. + * + * flags_ptr: Pointer to an unsigned integer used to return flags + * indicating whether the resize function resized or renamed + * the entry. If the entry was neither resized or renamed, + * the serialize function must set *flags_ptr to zero. + * H5C2__SERIALIZE_RESIZED_FLAG and H5C2__SERIALIZE_RENAMED_FLAG + * must be set to indicate a resize and a rename respectively. + * + * If the H5C2__SERIALIZE_RESIZED_FLAG is set, the new length + * and image pointer must be stored in *new_len_ptr and + * *new_image_ptr_ptr respectively. + * + * If the H5C2__SERIALIZE_RENAMED_FLAG flag is also set, the + * new image base address must be stored in *new_addr_ptr. + * Observe that the H5C2__SERIALIZE_RENAMED_FLAG must not + * appear without the H5C2__SERIALIZE_RESIZED_FLAG. + * + * Except as noted above, the locations pointed to by the + * remaining parameters are undefined, and should be ignored + * by the caller. + * + * new_addr_ptr: Pointer to haddr_t. If the entry is renamed by + * the serialize function, the new on disk base address must + * be stored in *new_addr_ptr. If the entry is not renamed + * by the serialize function, *new_addr_ptr is undefined. + * + * new_len_ptr: Pointer to size_t. If the entry is resized by the + * serialize function, the new length of the on disk image + * must be stored in *new_len_ptr. If the entry is not + * resized by the serialize function, *new_len_ptr is + * undefined. + * + * new_image_ptr_ptr: Pointer to pointer to void. If the entry is + * resized by the serialize function, the pointer to the + * new buffer containing the on disk image must be stored + * in *new_image_ptr_ptr. If the entry is not resized by + * the serialize function, *new_image_ptr_ptr is undefined. + * + * Processing in the serialize function should proceed as follows: + * + * The serialize function must examine the in core representation + * indicated by the thing parameter, and write a serialized image + * of its contents into the provided buffer. + * + * If the serialize function does not change the size or location + * of the on disk image, it must set *flags_ptr to zero. + * + * If the size of the on disk image must be changed, the serialize + * function must free the old image buffer (base address in image_ptr), + * allocate a new one, load the image into the new buffer, load the + * base address of the new buffer into *new_image_ptr_ptr, load the + * length of the new image into *new_len_ptr, and set the + * H5C2__SERIALIZE_RESIZED_FLAG in *flags_ptr. + * + * If in addition, the base address of the on disk image must + * be changed, the serialize function must also set *new_addr_ptr + * to the new base address, and set the H5C2__SERIALIZE_RENAMED_FLAG + * in *flags_ptr. + * + * If it is successful, the function must return SUCCEED. + * + * If it fails for any reason, the function must return FAIL and + * push error information on the error stack with the error API + * routines. + * + * + * free_icr: Pointer to the free ICR Callback. + * + * The free ICR callback is invoked by the metadata cache when it + * wishes to evict an entry, and needs the client to free the memory + * allocated for the in core representation. + * + * The typedef for the free ICR callback is as follows: + * + * typedef herr_t (*N5C_free_icr_func_t)(haddr_t addr, + * size_t len, + * void * thing); + * + * The parameters of the free ICR callback are as follows: + * + * addr: Base address in file of the entry being evicted. + * + * This parameter is supplied mainly for sanity checking. + * Sanity checks should be performed when compiled in debug + * mode, but the parameter may be unused when compiled in + * production mode. + * + * len: Length of the in file image of the entry being evicted + * in bytes. + * + * This parameter is supplied mainly for sanity checking. + * Sanity checks should be performed when compiled in debug + * mode, but the parameter may be unused when compiled in + * production mode. + * + * thing: Pointer to void containing the address of the in core + * representation of the target metadata cache entry. This + * is the same pointer that would be returned by a protect + * of the addr and len above. + * + * Processing in the free ICR function should proceed as follows: + * + * The free ICR function must free all memory allocated to the + * in core representation. + * + * If the function is successful, it must return SUCCEED. + * + * If it fails for any reason, the function must return FAIL and + * push error information on the error stack with the error API + * routines. + * + * At least when compiled with debug, it would be useful if the + * free ICR call would fail if the in core representation has been + * modified since the last serialize of clear callback. + * + * + * clear_dirty_bits: Pointer to the clear dirty bits callback. + * + * For sanity checking purposes, it will be useful if cache clients + * track whether an in core representation has been modified since + * the last time it was serialized. This data is used to flag an + * error if the cache attempts to free an in core representation + * that has not been serialized since the last time it was modified. + * + * If this happens, either the client forgot to tell the cache that + * an entry is dirty, or the cache forgot to flush a dirty entry + * before evicting it. In either case we want to know before we + * get file corruption complaints. + * + * However, in some cases, we want to mark an entry as clean even + * though it has not been flushed to disk -- most particularly in + * the parallel case. Thus we need some way to tell the client + * that a free of the associated ICR is OK even though it has + * been modified since the last serialization. Hence the clear + * dirty bits callback. + * + * Since the clear dirty bits callback is purely for sanity checking, + * it is called only when we compile with debug. + * + * The typedef for the clear callback is as follows: + * + * typedef herr_t (*N5C_clear_dirty_bits_func_t)(haddr_t addr, + * size_t len, + * void * thing); + * + * The parameters of the clear callback are as follows: + * + * addr: Base address in file of the entry whose dirty bits + * are being cleared + * + * len: Length in bytes of the in file image of the entry + * whose dirty bits are being cleared. + * + * thing: Pointer to void containing the address of the in + * core representation of the target metadata cache entry. + * This is the same pointer that would be returned by a + * protect of the addr and len above. + * + * Processing in the clear callback function should proceed as follows: + * + * The function must clear any dirty bits associated with the ICR. + * + * If successful, the function must return SUCCEED. + * + * If it fails for any reason, the function must return FAIL and + * push error information on the error stack with the error API + * routines. + * + ***************************************************************************/ +typedef void *(*H5C2_deserialize_func_t)(haddr_t addr, + size_t len, + const void * image_ptr, + const void * udata_ptr, + hbool_t * dirty_ptr); + +typedef herr_t (*H5C2_image_len_func_t)(void *thing, + size_t *image_len_ptr); + +#define H5C2__SERIALIZE_RESIZED_FLAG 0x1 +#define H5C2__SERIALIZE_RENAMED_FLAG 0x2 + +typedef herr_t (*H5C2_serialize_func_t)(haddr_t addr, + size_t len, + void * image_ptr, + void * thing, + unsigned * flags_ptr, + haddr_t * new_addr_ptr, + size_t * new_len_ptr, + void ** new_image_ptr_ptr); + +typedef herr_t (*H5C2_free_icr_func_t)(haddr_t addr, + size_t len, + void * thing); + +typedef herr_t (*H5C2_clear_dirty_bits_func_t)(haddr_t addr, + size_t len, + void * thing); + +typedef struct H5C2_class_t { + int id; + const char * name; + H5FD_mem_t mem_type; + H5C2_deserialize_func_t deserialize; + H5C2_image_len_func_t image_len; + H5C2_serialize_func_t serialize; + H5C2_free_icr_func_t free_icr; + H5C2_clear_dirty_bits_func_t clear_dirty_bits; +} H5C2_class_t; + + +/* Type defintions of call back functions used by the cache as a whole */ + +typedef herr_t (*H5C2_write_permitted_func_t)(const H5F_t *f, + hid_t dxpl_id, + hbool_t * write_permitted_ptr); + +typedef herr_t (*H5C2_log_flush_func_t)(H5C2_t * cache_ptr, + haddr_t addr, + hbool_t was_dirty, + unsigned flags, + int type_id); + +/* Upper and lower limits on cache size. These limits are picked + * out of a hat -- you should be able to change them as necessary. + * + * However, if you need a very big cache, you should also increase the + * size of the hash table (H5C2__HASH_TABLE_LEN in H5C2pkg.h). The current + * upper bound on cache size is rather large for the current hash table + * size. + */ + +#define H5C2__MAX_MAX_CACHE_SIZE ((size_t)(128 * 1024 * 1024)) +#define H5C2__MIN_MAX_CACHE_SIZE ((size_t)(1024)) + + +/* Default max cache size and min clean size are give here to make + * them generally accessable. + */ + +#define H5C2__DEFAULT_MAX_CACHE_SIZE ((size_t)(4 * 1024 * 1024)) +#define H5C2__DEFAULT_MIN_CLEAN_SIZE ((size_t)(2 * 1024 * 1024)) + + +/**************************************************************************** + * + * structure H5C2_cache_entry_t + * + * Instances of the H5C2_cache_entry_t structure are used to store cache + * entries in a hash table and sometimes in a skip list. + * See H5SL.c for the particulars of the skip list. + * + * In typical application, this structure is the first field in a + * structure to be cached. For historical reasons, the external module + * is responsible for managing the is_dirty field (this is no longer + * completely true. See the comment on the is_dirty field for details). + * All other fields are managed by the cache. + * + * The fields of this structure are discussed individually below: + * + * JRM - 4/26/04 + * + * magic: Unsigned 32 bit integer that must always be set to + * H5C2__H5C2_CACHE_ENTRY_T_MAGIC when the entry is valid. + * The field must be set to H5C2__H5C2_CACHE_ENTRY_T_BAD_MAGIC + * just before the entry is freed. + * + * This is necessary, as the LRU list can be changed out + * from under H5C2_make_space_in_cache() by the serialize + * callback which may change the size of an existing entry, + * and/or load a new entry while serializing the target entry. + * + * This in turn can cause a recursive call to + * H5C2_make_space_in_cache() which may either flush or evict + * the next entry that the first invocation of that function + * was about to examine. + * + * The magic field allows H5C2_make_space_in_cache() to + * detect this case, and re-start its scan from the bottom + * of the LRU when this situation occurs. + * + * addr: Base address of the cache entry on disk. + * + * size: Length of the cache entry on disk. Note that unlike normal + * caches, the entries in this cache are of variable length. + * The entries should never overlap, and when we do writebacks, + * we will want to writeback adjacent entries where possible. + * + * NB: At present, entries need not be contiguous on disk. Until + * we fix this, we can't do much with writing back adjacent + * entries. + * + * Update: This has now been changed -- all metadata cache + * entries must now be associated with a single contiguous + * block of memory on disk. The image of this block (i.e. + * the on disk image) is stored in *image_ptr (discussed below). + * + * image_ptr: Pointer to void. When not NULL, this field points to a + * dynamically allocated block of size bytes in which the + * on disk image of the metadata cache entry is stored. + * + * If the entry is dirty, the serialize callback must be used + * to update this image before it is written to disk + * + * type: Pointer to the instance of H5C2_class_t containing pointers + * to the methods for cache entries of the current type. This + * field should be NULL when the instance of H5C2_cache_entry_t + * is not in use. + * + * The name is not particularly descriptive, but is retained + * to avoid changes in existing code. + * + * is_dirty: Boolean flag indicating whether the contents of the cache + * entry has been modified since the last time it was written + * to disk. + * + * NOTE: For historical reasons, this field is not maintained + * by the cache. Instead, the module using the cache + * sets this flag when it modifies the entry, and the + * flush and clear functions supplied by that module + * reset the dirty when appropriate. + * + * This is a bit quirky, so we may want to change this + * someday. However it will require a change in the + * cache interface. + * + * Update: Management of the is_dirty field has been largely + * moved into the cache. The only remaining exceptions + * are the flush and clear functions supplied by the + * modules using the cache. These still clear the + * is_dirty field as before. -- JRM 7/5/05 + * + * Update: Management of the is_dirty field is now entirely + * in the cache. -- JRM 7/5/07 + * + * dirtied: Boolean flag used to indicate that the entry has been + * dirtied while protected. + * + * This field is set to FALSE in the protect call, and may + * be set to TRUE by the + * H5C2_mark_pinned_or_protected_entry_dirty() + * call at an time prior to the unprotect call. + * + * The H5C2_mark_pinned_or_protected_entry_dirty() call exists + * as a convenience function for the fractal heap code which + * may not know if an entry is protected or pinned, but knows + * that is either protected or pinned. The dirtied field was + * added as in the parallel case, it is necessary to know + * whether a protected entry was dirty prior to the protect call. + * + * is_protected: Boolean flag indicating whether this entry is protected + * (or locked, to use more conventional terms). When it is + * protected, the entry cannot be flushed or accessed until + * it is unprotected (or unlocked -- again to use more + * conventional terms). + * + * Note that protected entries are removed from the LRU lists + * and inserted on the protected list. + * + * is_read_only: Boolean flag that is only meaningful if is_protected is + * TRUE. In this circumstance, it indicates whether the + * entry has been protected read only, or read/write. + * + * If the entry has been protected read only (i.e. is_protected + * and is_read_only are both TRUE), we allow the entry to be + * protected more than once. + * + * In this case, the number of readers is maintained in the + * ro_ref_count field (see below), and unprotect calls simply + * decrement that field until it drops to zero, at which point + * the entry is actually unprotected. + * + * ro_ref_count: Integer field used to maintain a count of the number of + * outstanding read only protects on this entry. This field + * must be zero whenever either is_protected or is_read_only + * are TRUE. + * + * is_pinned: Boolean flag indicating whether the entry has been pinned + * in the cache. + * + * For very hot entries, the protect / unprotect overhead + * can become excessive. Thus the cache has been extended + * to allow an entry to be "pinned" in the cache. + * + * Pinning an entry in the cache has several implications: + * + * 1) A pinned entry cannot be evicted. Thus unprotected + * pinned entries must be stored in the pinned entry + * list, instead of being managed by the replacement + * policy code (LRU at present). + * + * 2) A pinned entry can be accessed or modified at any time. + * Therefore, the cache must check with the entry owner + * before flushing it. If permission is denied, the + * cache does not flush the entry. + * + * 3) A pinned entry can be marked as dirty (and possibly + * change size) while it is unprotected. + * + * 4) The flush-destroy code must allow pinned entries to + * be unpinned (and possibly unprotected) during the + * flush. + * + * JRM -- 3/16/06 + * + * in_slist: Boolean flag indicating whether the entry is in the skip list + * As a general rule, entries are placed in the list when they + * are marked dirty. However they may remain in the list after + * being flushed. + * + * Update: Dirty entries are now removed from the skip list + * when they are flushed. + * + * flush_marker: Boolean flag indicating that the entry is to be flushed + * the next time H5C2_flush_cache() is called with the + * H5AC__FLUSH_MARKED_ENTRIES_FLAG. The flag is reset when + * the entry is flushed for whatever reason. + * + * clear_on_unprotect: Boolean flag used only in PHDF5. When H5C2 is used + * to implement the metadata cache In the parallel case, only + * the cache with mpi rank 0 is allowed to actually write to + * file -- all other caches must retain dirty entries until they + * are advised that the entry is clean. + * + * This flag is used in the case that such an advisory is + * received when the entry is protected. If it is set when an + * entry is unprotected, and the dirtied flag is not set in + * the unprotect, the entry's is_dirty flag is reset by flushing + * it with the H5C2__FLUSH_CLEAR_ONLY_FLAG. + * + * flush_in_progress: Boolean flag that is set to true iff the entry + * is in the process of being flushed. This allows the cache + * to detect when a call is the result of a flush callback. + * + * destroy_in_progress: Boolean flag that is set to true iff the entry + * is in the process of being flushed and destroyed. + * + * + * Fields supporting the hash table: + * + * Fields in the cache are indexed by a more or less conventional hash table. + * If there are multiple entries in any hash bin, they are stored in a doubly + * linked list. + * + * ht_next: Next pointer used by the hash table to store multiple + * entries in a single hash bin. This field points to the + * next entry in the doubly linked list of entries in the + * hash bin, or NULL if there is no next entry. + * + * ht_prev: Prev pointer used by the hash table to store multiple + * entries in a single hash bin. This field points to the + * previous entry in the doubly linked list of entries in + * the hash bin, or NULL if there is no previuos entry. + * + * + * Fields supporting replacement policies: + * + * The cache must have a replacement policy, and it will usually be + * necessary for this structure to contain fields supporting that policy. + * + * While there has been interest in several replacement policies for + * this cache, the initial development schedule is tight. Thus I have + * elected to support only a modified LRU policy for the first cut. + * + * When additional replacement policies are added, the fields in this + * section will be used in different ways or not at all. Thus the + * documentation of these fields is repeated for each replacement policy. + * + * Modified LRU: + * + * When operating in parallel mode, we must ensure that a read does not + * cause a write. If it does, the process will hang, as the write will + * be collective and the other processes will not know to participate. + * + * To deal with this issue, I have modified the usual LRU policy by adding + * clean and dirty LRU lists to the usual LRU list. When reading in + * parallel mode, we evict from the clean LRU list only. This implies + * that we must try to ensure that the clean LRU list is reasonably well + * stocked. See the comments on H5C2_t in H5C2pkg.h for more details. + * + * Note that even if we start with a completely clean cache, a sequence + * of protects without unprotects can empty the clean LRU list. In this + * case, the cache must grow temporarily. At the next write, we will + * attempt to evict enough entries to get the cache down to its nominal + * maximum size. + * + * The use of the replacement policy fields under the Modified LRU policy + * is discussed below: + * + * next: Next pointer in either the LRU or the protected list, + * depending on the current value of protected. If there + * is no next entry on the list, this field should be set + * to NULL. + * + * prev: Prev pointer in either the LRU or the protected list, + * depending on the current value of protected. If there + * is no previous entry on the list, this field should be + * set to NULL. + * + * aux_next: Next pointer on either the clean or dirty LRU lists. + * This entry should be NULL when protected is true. When + * protected is false, and dirty is true, it should point + * to the next item on the dirty LRU list. When protected + * is false, and dirty is false, it should point to the + * next item on the clean LRU list. In either case, when + * there is no next item, it should be NULL. + * + * aux_prev: Previous pointer on either the clean or dirty LRU lists. + * This entry should be NULL when protected is true. When + * protected is false, and dirty is true, it should point + * to the previous item on the dirty LRU list. When protected + * is false, and dirty is false, it should point to the + * previous item on the clean LRU list. In either case, when + * there is no previous item, it should be NULL. + * + * + * Cache entry stats collection fields: + * + * These fields should only be compiled in when both H5C2_COLLECT_CACHE_STATS + * and H5C2_COLLECT_CACHE_ENTRY_STATS are true. When present, they allow + * collection of statistics on individual cache entries. + * + * accesses: int32_t containing the number of times this cache entry has + * been referenced in its lifetime. + * + * clears: int32_t containing the number of times this cache entry has + * been cleared in its life time. + * + * flushes: int32_t containing the number of times this cache entry has + * been flushed to file in its life time. + * + * pins: int32_t containing the number of times this cache entry has + * been pinned in cache in its life time. + * + ****************************************************************************/ + +#define H5C2__H5C2_CACHE_ENTRY_T_MAGIC 0x005CAC0A +#define H5C2__H5C2_CACHE_ENTRY_T_BAD_MAGIC 0xDeadBeef + +typedef struct H5C2_cache_entry_t +{ + uint32_t magic; + haddr_t addr; + size_t size; + void * image_ptr; + const H5C2_class_t * type; + hbool_t is_dirty; + hbool_t dirtied; + hbool_t is_protected; + hbool_t is_read_only; + int ro_ref_count; + hbool_t is_pinned; + hbool_t in_slist; + hbool_t flush_marker; +#ifdef H5_HAVE_PARALLEL + hbool_t clear_on_unprotect; +#endif /* H5_HAVE_PARALLEL */ + hbool_t flush_in_progress; + hbool_t destroy_in_progress; + + /* fields supporting the hash table: */ + + struct H5C2_cache_entry_t * ht_next; + struct H5C2_cache_entry_t * ht_prev; + + /* fields supporting replacement policies: */ + + struct H5C2_cache_entry_t * next; + struct H5C2_cache_entry_t * prev; + struct H5C2_cache_entry_t * aux_next; + struct H5C2_cache_entry_t * aux_prev; + +#if H5C2_COLLECT_CACHE_ENTRY_STATS + + /* cache entry stats fields */ + + int32_t accesses; + int32_t clears; + int32_t flushes; + int32_t pins; + +#endif /* H5C2_COLLECT_CACHE_ENTRY_STATS */ + +} H5C2_cache_entry_t; + + +/**************************************************************************** + * + * structure H5C2_auto_size_ctl_t + * + * Instances of H5C2_auto_size_ctl_t are used to get and set the control + * fields for automatic cache re-sizing. + * + * The fields of the structure are discussed individually below: + * + * version: Integer field containing the version number of this version + * of the H5C2_auto_size_ctl_t structure. Any instance of + * H5C2_auto_size_ctl_t passed to the cache must have a known + * version number, or an error will be flagged. + * + * report_fcn: Pointer to the function that is to be called to report + * activities each time the auto cache resize code is executed. If the + * field is NULL, no call is made. + * + * If the field is not NULL, it must contain the address of a function + * of type H5C2_auto_resize_report_fcn. + * + * set_initial_size: Boolean flag indicating whether the size of the + * initial size of the cache is to be set to the value given in + * the initial_size field. If set_initial_size is FALSE, the + * initial_size field is ignored. + * + * initial_size: If enabled, this field contain the size the cache is + * to be set to upon receipt of this structure. Needless to say, + * initial_size must lie in the closed interval [min_size, max_size]. + * + * min_clean_fraction: double in the range 0 to 1 indicating the fraction + * of the cache that is to be kept clean. This field is only used + * in parallel mode. Typical values are 0.1 to 0.5. + * + * max_size: Maximum size to which the cache can be adjusted. The + * supplied value must fall in the closed interval + * [MIN_MAX_CACHE_SIZE, MAX_MAX_CACHE_SIZE]. Also, max_size must + * be greater than or equal to min_size. + * + * min_size: Minimum size to which the cache can be adjusted. The + * supplied value must fall in the closed interval + * [MIN_MAX_CACHE_SIZE, MAX_MAX_CACHE_SIZE]. Also, min_size must + * be less than or equal to max_size. + * + * epoch_length: Number of accesses on the cache over which to collect + * hit rate stats before running the automatic cache resize code, + * if it is enabled. + * + * At the end of an epoch, we discard prior hit rate data and start + * collecting afresh. The epoch_length must lie in the closed + * interval [H5C2__MIN_AR_EPOCH_LENGTH, H5C2__MAX_AR_EPOCH_LENGTH]. + * + * + * Cache size increase control fields: + * + * incr_mode: Instance of the H5C2_cache_incr_mode enumerated type whose + * value indicates how we determine whether the cache size should be + * increased. At present there are two possible values: + * + * H5C2_incr__off: Don't attempt to increase the size of the cache + * automatically. + * + * When this increment mode is selected, the remaining fields + * in the cache size increase section ar ignored. + * + * H5C2_incr__threshold: Attempt to increase the size of the cache + * whenever the average hit rate over the last epoch drops + * below the value supplied in the lower_hr_threshold + * field. + * + * Note that this attempt will fail if the cache is already + * at its maximum size, or if the cache is not already using + * all available space. + * + * lower_hr_threshold: Lower hit rate threshold. If the increment mode + * (incr_mode) is H5C2_incr__threshold and the hit rate drops below the + * value supplied in this field in an epoch, increment the cache size by + * size_increment. Note that cache size may not be incremented above + * max_size, and that the increment may be further restricted by the + * max_increment field if it is enabled. + * + * When enabled, this field must contain a value in the range [0.0, 1.0]. + * Depending on the incr_mode selected, it may also have to be less than + * upper_hr_threshold. + * + * increment: Double containing the multiplier used to derive the new + * cache size from the old if a cache size increment is triggered. + * The increment must be greater than 1.0, and should not exceed 2.0. + * + * The new cache size is obtained by multiplying the current max cache + * size by the increment, and then clamping to max_size and to stay + * within the max_increment as necessary. + * + * apply_max_increment: Boolean flag indicating whether the max_increment + * field should be used to limit the maximum cache size increment. + * + * max_increment: If enabled by the apply_max_increment field described + * above, this field contains the maximum number of bytes by which the + * cache size can be increased in a single re-size. + * + * + * Cache size decrease control fields: + * + * decr_mode: Instance of the H5C2_cache_decr_mode enumerated type whose + * value indicates how we determine whether the cache size should be + * decreased. At present there are four possibilities. + * + * H5C2_decr__off: Don't attempt to decrease the size of the cache + * automatically. + * + * When this increment mode is selected, the remaining fields + * in the cache size decrease section are ignored. + * + * H5C2_decr__threshold: Attempt to decrease the size of the cache + * whenever the average hit rate over the last epoch rises + * above the value supplied in the upper_hr_threshold + * field. + * + * H5C2_decr__age_out: At the end of each epoch, search the cache for + * entries that have not been accessed for at least the number + * of epochs specified in the epochs_before_eviction field, and + * evict these entries. Conceptually, the maximum cache size + * is then decreased to match the new actual cache size. However, + * this reduction may be modified by the min_size, the + * max_decrement, and/or the empty_reserve. + * + * H5C2_decr__age_out_with_threshold: Same as age_out, but we only + * attempt to reduce the cache size when the hit rate observed + * over the last epoch exceeds the value provided in the + * upper_hr_threshold field. + * + * upper_hr_threshold: Upper hit rate threshold. The use of this field + * varies according to the current decr_mode: + * + * H5C2_decr__off or H5C2_decr__age_out: The value of this field is + * ignored. + * + * H5C2_decr__threshold: If the hit rate exceeds this threshold in any + * epoch, attempt to decrement the cache size by size_decrement. + * + * Note that cache size may not be decremented below min_size. + * + * Note also that if the upper_threshold is 1.0, the cache size + * will never be reduced. + * + * H5C2_decr__age_out_with_threshold: If the hit rate exceeds this + * threshold in any epoch, attempt to reduce the cache size + * by evicting entries that have not been accessed for more + * than the specified number of epochs. + * + * decrement: This field is only used when the decr_mode is + * H5C2_decr__threshold. + * + * The field is a double containing the multiplier used to derive the + * new cache size from the old if a cache size decrement is triggered. + * The decrement must be in the range 0.0 (in which case the cache will + * try to contract to its minimum size) to 1.0 (in which case the + * cache will never shrink). + * + * apply_max_decrement: Boolean flag used to determine whether decrements + * in cache size are to be limited by the max_decrement field. + * + * max_decrement: Maximum number of bytes by which the cache size can be + * decreased in a single re-size. Note that decrements may also be + * restricted by the min_size of the cache, and (in age out modes) by + * the empty_reserve field. + * + * epochs_before_eviction: Integer field used in H5C2_decr__age_out and + * H5C2_decr__age_out_with_threshold decrement modes. + * + * This field contains the number of epochs an entry must remain + * unaccessed before it is evicted in an attempt to reduce the + * cache size. If applicable, this field must lie in the range + * [1, H5C2__MAX_EPOCH_MARKERS]. + * + * apply_empty_reserve: Boolean field controlling whether the empty_reserve + * field is to be used in computing the new cache size when the + * decr_mode is H5C2_decr__age_out or H5C2_decr__age_out_with_threshold. + * + * empty_reserve: To avoid a constant racheting down of cache size by small + * amounts in the H5C2_decr__age_out and H5C2_decr__age_out_with_threshold + * modes, this field allows one to require that any cache size + * reductions leave the specified fraction of unused space in the cache. + * + * The value of this field must be in the range [0.0, 1.0]. I would + * expect typical values to be in the range of 0.01 to 0.1. + * + ****************************************************************************/ + +#define H5C2_RESIZE_CFG__VALIDATE_GENERAL 0x1 +#define H5C2_RESIZE_CFG__VALIDATE_INCREMENT 0x2 +#define H5C2_RESIZE_CFG__VALIDATE_DECREMENT 0x4 +#define H5C2_RESIZE_CFG__VALIDATE_INTERACTIONS 0x8 +#define H5C2_RESIZE_CFG__VALIDATE_ALL \ +( \ + H5C2_RESIZE_CFG__VALIDATE_GENERAL | \ + H5C2_RESIZE_CFG__VALIDATE_INCREMENT | \ + H5C2_RESIZE_CFG__VALIDATE_DECREMENT | \ + H5C2_RESIZE_CFG__VALIDATE_INTERACTIONS \ +) + +#define H5C2__CURR_AUTO_SIZE_CTL_VER 1 +#define H5C2__CURR_AUTO_RESIZE_RPT_FCN_VER 1 + +#define H5C2__MAX_EPOCH_MARKERS 10 + +#define H5C2__DEF_AR_UPPER_THRESHHOLD 0.9999 +#define H5C2__DEF_AR_LOWER_THRESHHOLD 0.9 +#define H5C2__DEF_AR_MAX_SIZE ((size_t)(16 * 1024 * 1024)) +#define H5C2__DEF_AR_INIT_SIZE ((size_t)( 1 * 1024 * 1024)) +#define H5C2__DEF_AR_MIN_SIZE ((size_t)( 1 * 1024 * 1024)) +#define H5C2__DEF_AR_MIN_CLEAN_FRAC 0.5 +#define H5C2__DEF_AR_INCREMENT 2.0 +#define H5C2__DEF_AR_MAX_INCREMENT ((size_t)( 2 * 1024 * 1024)) +#define H5C2__DEF_AR_DECREMENT 0.9 +#define H5C2__DEF_AR_MAX_DECREMENT ((size_t)( 1 * 1024 * 1024)) +#define H5C2__DEF_AR_EPCHS_B4_EVICT 3 +#define H5C2__DEF_AR_EMPTY_RESERVE 0.05 +#define H5C2__MIN_AR_EPOCH_LENGTH 100 +#define H5C2__DEF_AR_EPOCH_LENGTH 50000 +#define H5C2__MAX_AR_EPOCH_LENGTH 1000000 + +enum H5C2_resize_status +{ + in_spec2, + increase2, + decrease2, + at_max_size2, + at_min_size2, + increase_disabled2, + decrease_disabled2, + not_full2 +}; /* enum H5C2_resize_conditions */ + +typedef void (*H5C2_auto_resize_rpt_fcn)(H5C2_t * cache_ptr, + int32_t version, + double hit_rate, + enum H5C2_resize_status status, + size_t old_max_cache_size, + size_t new_max_cache_size, + size_t old_min_clean_size, + size_t new_min_clean_size); + +typedef struct H5C2_auto_size_ctl_t +{ + /* general configuration fields: */ + int32_t version; + H5C2_auto_resize_rpt_fcn rpt_fcn; + + hbool_t set_initial_size; + size_t initial_size; + + double min_clean_fraction; + + size_t max_size; + size_t min_size; + + int64_t epoch_length; + + + /* size increase control fields: */ + enum H5C2_cache_incr_mode incr_mode; + + double lower_hr_threshold; + + double increment; + + hbool_t apply_max_increment; + size_t max_increment; + + + /* size decrease control fields: */ + enum H5C2_cache_decr_mode decr_mode; + + double upper_hr_threshold; + + double decrement; + + hbool_t apply_max_decrement; + size_t max_decrement; + + int32_t epochs_before_eviction; + + hbool_t apply_empty_reserve; + double empty_reserve; + +} H5C2_auto_size_ctl_t; + + +/* + * Library prototypes. + */ + +/* #defines of flags used in the flags parameters in some of the + * following function calls. Note that not all flags are applicable + * to all function calls. Flags that don't apply to a particular + * function are ignored in that function. + * + * These flags apply to all function calls: + * + * H5C2__NO_FLAGS_SET (generic "no flags set" for all fcn calls) + * + * + * These flags apply to H5C2_insert_entry(): + * + * H5C2__SET_FLUSH_MARKER_FLAG + * H5C2__PIN_ENTRY_FLAG + * + * These flags apply to H5C2_protect() + * + * H5C2__READ_ONLY_FLAG + * H5C2__CHECK_SIZE_FLAG + * + * These flags apply to H5C2_unprotect(): + * + * H5C2__SET_FLUSH_MARKER_FLAG + * H5C2__DELETED_FLAG + * H5C2__DIRTIED_FLAG + * H5C2__SIZE_CHANGED_FLAG + * H5C2__PIN_ENTRY_FLAG + * H5C2__UNPIN_ENTRY_FLAG + * + * + * These flags apply to H5C2_flush_cache(): + * + * H5C2__FLUSH_INVALIDATE_FLAG + * H5C2__FLUSH_CLEAR_ONLY_FLAG + * H5C2__FLUSH_MARKED_ENTRIES_FLAG + * H5C2__FLUSH_IGNORE_PROTECTED_FLAG (can't use this flag in combination + * with H5C2__FLUSH_INVALIDATE_FLAG) + * + * These flags apply to H5C2_flush_single_entry(): + * + * H5C2__FLUSH_INVALIDATE_FLAG + * H5C2__FLUSH_CLEAR_ONLY_FLAG + * H5C2__FLUSH_MARKED_ENTRIES_FLAG + */ + +#define H5C2__NO_FLAGS_SET 0x0000 +#define H5C2__SET_FLUSH_MARKER_FLAG 0x0001 +#define H5C2__DELETED_FLAG 0x0002 +#define H5C2__DIRTIED_FLAG 0x0004 +#define H5C2__SIZE_CHANGED_FLAG 0x0008 +#define H5C2__PIN_ENTRY_FLAG 0x0010 +#define H5C2__UNPIN_ENTRY_FLAG 0x0020 +#define H5C2__FLUSH_INVALIDATE_FLAG 0x0040 +#define H5C2__FLUSH_CLEAR_ONLY_FLAG 0x0080 +#define H5C2__FLUSH_MARKED_ENTRIES_FLAG 0x0100 +#define H5C2__FLUSH_IGNORE_PROTECTED_FLAG 0x0200 +#define H5C2__READ_ONLY_FLAG 0x0400 +#define H5C2__CHECK_SIZE_FLAG 0x0800 + +H5_DLL H5C2_t * H5C2_create(const H5F_t * f, + size_t max_cache_size, + size_t min_clean_size, + int max_type_id, + const char * (* type_name_table_ptr), + H5C2_write_permitted_func_t check_write_permitted, + hbool_t write_permitted, + H5C2_log_flush_func_t log_flush, + void * aux_ptr); + +H5_DLL void H5C2_def_auto_resize_rpt_fcn(H5C2_t * cache_ptr, + int32_t version, + double hit_rate, + enum H5C2_resize_status status, + size_t old_max_cache_size, + size_t new_max_cache_size, + size_t old_min_clean_size, + size_t new_min_clean_size); + +H5_DLL herr_t H5C2_dest(H5C2_t * cache_ptr, + hid_t dxpl_id); + +H5_DLL herr_t H5C2_dest_empty(H5C2_t * cache_ptr); + +H5_DLL herr_t H5C2_expunge_entry(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr); + +H5_DLL herr_t H5C2_flush_cache(H5C2_t * cache_ptr, + hid_t dxpl_id, + unsigned flags); + + +H5_DLL herr_t H5C2_flush_to_min_clean(H5C2_t * cache_ptr, + hid_t dxpl_id); + +H5_DLL herr_t H5C2_get_cache_auto_resize_config(H5C2_t * cache_ptr, + H5C2_auto_size_ctl_t *config_ptr); + +H5_DLL herr_t H5C2_get_cache_size(H5C2_t * cache_ptr, + size_t * max_size_ptr, + size_t * min_clean_size_ptr, + size_t * cur_size_ptr, + int32_t * cur_num_entries_ptr); + +H5_DLL herr_t H5C2_get_cache_hit_rate(H5C2_t * cache_ptr, + double * hit_rate_ptr); + +H5_DLL herr_t H5C2_get_entry_status(H5C2_t * cache_ptr, + haddr_t addr, + size_t * size_ptr, + hbool_t * in_cache_ptr, + hbool_t * is_dirty_ptr, + hbool_t * is_protected_ptr, + hbool_t * is_pinned_ptr); + +H5_DLL herr_t H5C2_get_evictions_enabled(H5C2_t * cache_ptr, + hbool_t * evictions_enabled_ptr); + +H5_DLL herr_t H5C2_get_trace_file_ptr(H5C2_t * cache_ptr, + FILE ** trace_file_ptr_ptr); + +H5_DLL herr_t H5C2_insert_entry(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + size_t len, + void * thing, + unsigned int flags); + +H5_DLL herr_t H5C2_mark_entries_as_clean(H5C2_t * cache_ptr, + hid_t dxpl_id, + int32_t ce_array_len, + haddr_t * ce_array_ptr); + +H5_DLL herr_t H5C2_mark_pinned_entry_dirty(H5C2_t * cache_ptr, + void * thing, + hbool_t size_changed, + size_t new_size); + +H5_DLL herr_t H5C2_mark_pinned_or_protected_entry_dirty(H5C2_t * cache_ptr, + void * thing); + +H5_DLL herr_t H5C2_rename_entry(H5C2_t * cache_ptr, + const H5C2_class_t * type, + haddr_t old_addr, + haddr_t new_addr); + +H5_DLL herr_t H5C2_pin_protected_entry(H5C2_t * cache_ptr, + void * thing); + +H5_DLL void * H5C2_protect(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + size_t len, + const void * udata, + unsigned flags); + +H5_DLL herr_t H5C2_reset_cache_hit_rate_stats(H5C2_t * cache_ptr); + +H5_DLL herr_t H5C2_resize_pinned_entry(H5C2_t * cache_ptr, + void * thing, + size_t new_size); + +H5_DLL herr_t H5C2_set_cache_auto_resize_config(H5C2_t * cache_ptr, + H5C2_auto_size_ctl_t *config_ptr); + +H5_DLL herr_t H5C2_set_evictions_enabled(H5C2_t * cache_ptr, + hbool_t evictions_enabled); + +H5_DLL herr_t H5C2_set_prefix(H5C2_t * cache_ptr, char * prefix); + +H5_DLL herr_t H5C2_set_skip_flags(H5C2_t * cache_ptr, + hbool_t skip_file_checks, + hbool_t skip_dxpl_id_checks); + +H5_DLL herr_t H5C2_set_trace_file_ptr(H5C2_t * cache_ptr, + FILE * trace_file_ptr); + +H5_DLL herr_t H5C2_stats(H5C2_t * cache_ptr, + const char * cache_name, + hbool_t display_detailed_stats); + +H5_DLL void H5C2_stats__reset(H5C2_t * cache_ptr); + +H5_DLL herr_t H5C2_unpin_entry(H5C2_t * cache_ptr, void * thing); + +H5_DLL herr_t H5C2_unprotect(H5C2_t * cache_ptr, + hid_t dxpl_id, + const H5C2_class_t * type, + haddr_t addr, + void * thing, + unsigned int flags, + size_t new_size); + +H5_DLL herr_t H5C2_validate_resize_config(H5C2_auto_size_ctl_t * config_ptr, + unsigned int tests); + +#endif /* !_H5C2private_H */ + diff --git a/src/H5C2public.h b/src/H5C2public.h new file mode 100644 index 0000000..e1adff3 --- /dev/null +++ b/src/H5C2public.h @@ -0,0 +1,55 @@ +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * + * 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: H5C2public.h + * June 4, 2005 + * John Mainzer + * + * Purpose: Public include file for cache functions. + * + * Modifications: + * + *------------------------------------------------------------------------- + */ +#ifndef _H5C2public_H +#define _H5C2public_H + +/* Public headers needed by this file */ +#include "H5public.h" + +#ifdef __cplusplus +extern "C" { +#endif + +enum H5C2_cache_incr_mode +{ + H5C2_incr__off, + H5C2_incr__threshold +}; + +enum H5C2_cache_decr_mode +{ + H5C2_decr__off, + H5C2_decr__threshold, + H5C2_decr__age_out, + H5C2_decr__age_out_with_threshold +}; + +#ifdef __cplusplus +} +#endif +#endif |