diff options
author | John Mainzer <mainzer@hdfgroup.org> | 2005-09-27 05:20:11 (GMT) |
---|---|---|
committer | John Mainzer <mainzer@hdfgroup.org> | 2005-09-27 05:20:11 (GMT) |
commit | c100b0bf2639c03579ce1b2c4013b36c6f40350b (patch) | |
tree | 5c3b2834af1206110243886f357857900a8b108e /src/H5HL.c | |
parent | f9fc749ca218a878dbea4022ba1c2fb527f7822c (diff) | |
download | hdf5-c100b0bf2639c03579ce1b2c4013b36c6f40350b.zip hdf5-c100b0bf2639c03579ce1b2c4013b36c6f40350b.tar.gz hdf5-c100b0bf2639c03579ce1b2c4013b36c6f40350b.tar.bz2 |
[svn-r11470] Purpose:
Repair synchronization bug in the metadata cache in PHDF5
Also repair numerous other bugs that surfaced in testing the
bug fix.
Description:
While operations modifying metadata must be collective, we allow
independant reads. This allows metadata caches on different processes
to adjust to different sizes, and to place the entries on their dirty
lists in different orders. Since only process 0 actually writes
metadata to disk (all other processes thought they did, but the writes
were discarded on the theory that they had to be collective), this made
it possible for another process to modify metadata, flush it, and then
read it back in in its original form (pre-modification) form. The
possibilities for file corruption should be obvious.
Solution:
Make the policy that only process 0 can write to file explicit, and
visible to the metadata caches. Thus only process 0 may flush dirty
entries -- all other caches must retain dirty entries until they are
informed by process 0 that the entries are clean.
Synchronization is handled by counting the bytes of dirty cache entries
created, and then synching up between the caches whenever the sum
exceeds an (eventually user specified) limit. Dirty metadata creation
is consistent across all processes because all operations modifying
metadata must be collective.
This change uncovered may bugs which are repaired in this checkin.
It also required modification of H5HL and H5O to allocate file space
on insertion rather than on flush from cache.
Platforms tested:
H5committest, heping(parallel & serial)
Misc. update:
Diffstat (limited to 'src/H5HL.c')
-rw-r--r-- | src/H5HL.c | 189 |
1 files changed, 150 insertions, 39 deletions
@@ -155,7 +155,6 @@ H5HL_create(H5F_t *f, hid_t dxpl_id, size_t size_hint, haddr_t *addr_p/*out*/) heap->addr = *addr_p + (hsize_t)sizeof_hdr; heap->disk_alloc = size_hint; heap->mem_alloc = size_hint; - heap->disk_resrv = 0; if (NULL==(heap->chunk = H5FL_BLK_CALLOC(heap_chunk,(sizeof_hdr + size_hint)))) HGOTO_ERROR (H5E_RESOURCE, H5E_NOSPACE, FAIL, "memory allocation failed"); @@ -320,6 +319,20 @@ done: * * Modifications: * + * John Mainzer, 8/10/05 + * Reworked this function for a different role. + * + * It used to be called during cache eviction, where it + * attempted to size the disk space allocation for the + * actuall size of the heap. However, this causes problems + * in the parallel case, as the reuslting disk allocations + * may not be synchronized. + * + * It is now called from H5HL_remove(), where it is used to + * reduce heap size in response to an entry deletion. This + * means that the function should either do nothing, or + * reduce the size of the disk allocation. + * *------------------------------------------------------------------------- */ static herr_t @@ -331,19 +344,13 @@ H5HL_minimize_heap_space(H5F_t *f, hid_t dxpl_id, H5HL_t *heap) FUNC_ENTER_NOAPI(H5HL_minimize_heap_space, FAIL) /* check args */ - assert(f); - assert(heap); + HDassert( f ); + HDassert( heap ); + HDassert( heap->disk_alloc == heap->mem_alloc ); sizeof_hdr = H5HL_SIZEOF_HDR(f); /* cache H5HL header size for file */ /* - * When the heap is being flushed to disk, release the file space reserved - * for it. - */ - H5MF_free_reserved(f, (hsize_t)heap->disk_resrv); - heap->disk_resrv = 0; - - /* * Check to see if we can reduce the size of the heap in memory by * eliminating free blocks at the tail of the buffer before flushing the * buffer out. @@ -427,13 +434,15 @@ H5HL_minimize_heap_space(H5F_t *f, hid_t dxpl_id, H5HL_t *heap) } /* - * If the heap grew larger or smaller than disk storage then move the + * If the heap grew smaller than disk storage then move the * data segment of the heap to another contiguous block of disk * storage. */ if (heap->mem_alloc != heap->disk_alloc) { haddr_t old_addr = heap->addr, new_addr; + HDassert( heap->mem_alloc < heap->disk_alloc ); + /* Release old space on disk */ H5_CHECK_OVERFLOW(heap->disk_alloc, size_t, hsize_t); H5MF_xfree(f, H5FD_MEM_LHEAP, dxpl_id, old_addr, (hsize_t)heap->disk_alloc); @@ -453,7 +462,8 @@ H5HL_minimize_heap_space(H5F_t *f, hid_t dxpl_id, H5HL_t *heap) done: FUNC_LEAVE_NOAPI(ret_value) -} + +} /* H5HL_minimize_heap_space() */ /*------------------------------------------------------------------------- @@ -543,6 +553,13 @@ H5HL_serialize(H5F_t *f, H5HL_t *heap, uint8_t *buf) * * Bill Wendling, 2003-09-16 * Separated out the bit that serializes the heap. + * + * John Mainzer, 2005-08-10 + * Removed call to H5HL_minimize_heap_space(). It does disk space + * allocation, which can cause problems if done at flush time. + * Instead, disk space allocation/deallocation is now done at + * insert/remove time. + * *------------------------------------------------------------------------- */ static herr_t @@ -553,21 +570,18 @@ H5HL_flush(H5F_t *f, hid_t dxpl_id, hbool_t destroy, haddr_t addr, H5HL_t *heap) FUNC_ENTER_NOAPI(H5HL_flush, FAIL); /* check arguments */ - assert(f); - assert(H5F_addr_defined(addr)); - assert(heap); + HDassert( f ); + HDassert( H5F_addr_defined(addr) ); + HDassert( heap ); + HDassert( heap->disk_alloc == heap->mem_alloc ); if (heap->cache_info.is_dirty) { haddr_t hdr_end_addr; size_t sizeof_hdr = H5HL_SIZEOF_HDR(f); /* cache H5HL header size for file */ - /* Minimize the heap space size if possible */ - if (H5HL_minimize_heap_space(f, dxpl_id, heap) < 0) - HGOTO_ERROR(H5E_HEAP, H5E_CANTFREE, FAIL, "unable to minimize local heap space") - /* Write the header */ if (H5HL_serialize(f, heap, heap->chunk) < 0) - HGOTO_ERROR(H5E_BTREE, H5E_CANTSERIALIZE, FAIL, "unable to serialize local heap") + HGOTO_ERROR(H5E_HEAP, H5E_CANTSERIALIZE, FAIL, "unable to serialize local heap") /* Copy buffer to disk */ hdr_end_addr = addr + (hsize_t)sizeof_hdr; @@ -951,6 +965,10 @@ H5HL_remove_free(H5HL_t *heap, H5HL_free_t *fl) * H5AC_unprotect() instead of manipulating the is_dirty * field of the cache info directly. * + * John Mainzer, 8/10/05 + * Modified code to allocate file space as needed, instead + * of allocating it on eviction. + * *------------------------------------------------------------------------- */ size_t @@ -959,10 +977,12 @@ H5HL_insert(H5F_t *f, hid_t dxpl_id, haddr_t addr, size_t buf_size, const void * H5HL_t *heap = NULL; unsigned heap_flags = H5AC__NO_FLAGS_SET; H5HL_free_t *fl = NULL, *max_fl = NULL; + htri_t tri_result; + herr_t result; size_t offset = 0; size_t need_size, old_size, need_more; + size_t new_disk_alloc; hbool_t found; - size_t disk_resrv; /* Amount of additional space to reserve in file */ size_t sizeof_hdr; /* Cache H5HL header size for file */ size_t ret_value; /* Return value */ @@ -1028,18 +1048,54 @@ H5HL_insert(H5F_t *f, hid_t dxpl_id, haddr_t addr, size_t buf_size, const void * if (found==FALSE) { need_more = MAX3(need_size, heap->mem_alloc, H5HL_SIZEOF_FREE(f)); - /* Reserve space in the file to hold the increased heap size - */ - if( heap->disk_resrv == heap->mem_alloc) - disk_resrv = need_more; - else - disk_resrv = heap->mem_alloc + need_more - heap->disk_resrv; + new_disk_alloc = heap->disk_alloc + need_more; + HDassert( heap->disk_alloc < new_disk_alloc ); + H5_CHECK_OVERFLOW(heap->disk_alloc, size_t, hsize_t); + H5_CHECK_OVERFLOW(new_disk_alloc, size_t, hsize_t); + + /* extend the current heap if we can... */ + tri_result = H5MF_can_extend(f, H5FD_MEM_LHEAP, heap->addr, + (hsize_t)(heap->disk_alloc), + (hsize_t)need_more); + if ( tri_result == TRUE ) { + + result = H5MF_extend(f, H5FD_MEM_LHEAP, heap->addr, + (hsize_t)(heap->disk_alloc), + (hsize_t)need_more); + if ( result < 0 ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, (size_t)(-1), \ + "can't extend heap on disk"); + } + + heap->disk_alloc = new_disk_alloc; + + } else { /* ...if we can't, allocate a new chunk & release the old */ + + haddr_t old_addr = heap->addr; + haddr_t new_addr; + + /* The new allocation may fail -- to avoid the possiblity of + * file corruption, allocate the new heap first, and then + * deallocate the old. + */ + + /* allocate new disk space for the heap */ + if ( (new_addr = H5MF_alloc(f, H5FD_MEM_LHEAP, dxpl_id, + (hsize_t)new_disk_alloc)) == HADDR_UNDEF ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, (size_t)(-1), \ + "unable to allocate file space for heap") + } - if( H5MF_reserve(f, (hsize_t)disk_resrv) < 0 ) - HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, (size_t)(-1), "unable to reserve space in file"); + /* Release old space on disk */ + H5MF_xfree(f, H5FD_MEM_LHEAP, dxpl_id, old_addr, + (hsize_t)heap->disk_alloc); + H5E_clear_stack(NULL); /* don't really care if the free failed */ - /* Update heap's record of how much space it has reserved */ - heap->disk_resrv += disk_resrv; + heap->addr = new_addr; + heap->disk_alloc = new_disk_alloc; + } if (max_fl && max_fl->offset + max_fl->size == heap->mem_alloc) { /* @@ -1117,7 +1173,8 @@ done: HDONE_ERROR(H5E_HEAP, H5E_PROTECT, (size_t)(-1), "unable to release object header"); FUNC_LEAVE_NOAPI(ret_value); -} + +} /* H5HL_insert() */ #ifdef NOT_YET @@ -1217,6 +1274,11 @@ done: * H5AC_unprotect() instead of manipulating the is_dirty * field of the cache info directly. * + * John Mainzer, 8/10/05 + * Modified code to attempt to decrease heap size if the + * entry removal results in a free list entry at the end + * of the heap that is at least half the size of the heap. + * *------------------------------------------------------------------------- */ herr_t @@ -1225,15 +1287,15 @@ H5HL_remove(H5F_t *f, hid_t dxpl_id, haddr_t addr, size_t offset, size_t size) H5HL_t *heap = NULL; unsigned heap_flags = H5AC__NO_FLAGS_SET; H5HL_free_t *fl = NULL, *fl2 = NULL; - herr_t ret_value=SUCCEED; /* Return value */ + herr_t ret_value = SUCCEED; /* Return value */ FUNC_ENTER_NOAPI(H5HL_remove, FAIL); /* check arguments */ - assert(f); - assert(H5F_addr_defined(addr)); - assert(size > 0); - assert (offset==H5HL_ALIGN (offset)); + HDassert( f ); + HDassert( H5F_addr_defined(addr) ); + HDassert( size > 0 ); + HDassert( offset == H5HL_ALIGN(offset) ); if (0==(f->intent & H5F_ACC_RDWR)) HGOTO_ERROR (H5E_HEAP, H5E_WRITEERROR, FAIL, "no write intent on file"); @@ -1243,8 +1305,9 @@ H5HL_remove(H5F_t *f, hid_t dxpl_id, haddr_t addr, size_t offset, size_t size) if (NULL == (heap = H5AC_protect(f, dxpl_id, H5AC_LHEAP, addr, NULL, NULL, H5AC_WRITE))) HGOTO_ERROR(H5E_HEAP, H5E_PROTECT, FAIL, "unable to load heap"); - assert(offset < heap->mem_alloc); - assert(offset + size <= heap->mem_alloc); + HDassert( offset < heap->mem_alloc ); + HDassert( offset + size <= heap->mem_alloc ); + HDassert( heap->disk_alloc == heap->mem_alloc ); fl = heap->freelist; heap_flags |= H5AC__DIRTIED_FLAG; @@ -1268,10 +1331,29 @@ H5HL_remove(H5F_t *f, hid_t dxpl_id, haddr_t addr, size_t offset, size_t size) assert (fl->offset==H5HL_ALIGN (fl->offset)); assert (fl->size==H5HL_ALIGN (fl->size)); fl2 = H5HL_remove_free(heap, fl2); + if ( ( (fl->offset + fl->size) == heap->mem_alloc ) && + ( (2 * fl->size) > heap->mem_alloc ) ) { + + if ( H5HL_minimize_heap_space(f, dxpl_id, heap) != + SUCCEED ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTFREE, FAIL, \ + "heap size minimization failed"); + } + } HGOTO_DONE(SUCCEED); } fl2 = fl2->next; } + if ( ( (fl->offset + fl->size) == heap->mem_alloc ) && + ( (2 * fl->size) > heap->mem_alloc ) ) { + + if ( H5HL_minimize_heap_space(f, dxpl_id, heap) != SUCCEED ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTFREE, FAIL, \ + "heap size minimization failed"); + } + } HGOTO_DONE(SUCCEED); } else if (fl->offset + fl->size == offset) { @@ -1283,10 +1365,29 @@ H5HL_remove(H5F_t *f, hid_t dxpl_id, haddr_t addr, size_t offset, size_t size) fl->size += fl2->size; assert (fl->size==H5HL_ALIGN (fl->size)); fl2 = H5HL_remove_free(heap, fl2); + if ( ( (fl->offset + fl->size) == heap->mem_alloc ) && + ( (2 * fl->size) > heap->mem_alloc ) ) { + + if ( H5HL_minimize_heap_space(f, dxpl_id, heap) != + SUCCEED ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTFREE, FAIL, \ + "heap size minimization failed"); + } + } HGOTO_DONE(SUCCEED); } fl2 = fl2->next; } + if ( ( (fl->offset + fl->size) == heap->mem_alloc ) && + ( (2 * fl->size) > heap->mem_alloc ) ) { + + if ( H5HL_minimize_heap_space(f, dxpl_id, heap) != SUCCEED ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTFREE, FAIL, \ + "heap size minimization failed"); + } + } HGOTO_DONE(SUCCEED); } fl = fl->next; @@ -1321,6 +1422,16 @@ H5HL_remove(H5F_t *f, hid_t dxpl_id, haddr_t addr, size_t offset, size_t size) heap->freelist->prev = fl; heap->freelist = fl; + if ( ( (fl->offset + fl->size) == heap->mem_alloc ) && + ( (2 * fl->size) > heap->mem_alloc ) ) { + + if ( H5HL_minimize_heap_space(f, dxpl_id, heap) != SUCCEED ) { + + HGOTO_ERROR(H5E_RESOURCE, H5E_CANTFREE, FAIL, \ + "heap size minimization failed"); + } + } + done: if (heap && H5AC_unprotect(f, dxpl_id, H5AC_LHEAP, addr, heap, heap_flags) != SUCCEED) HDONE_ERROR(H5E_HEAP, H5E_PROTECT, FAIL, "unable to release object header"); |