diff options
Diffstat (limited to 'src/H5PB.c')
| -rw-r--r-- | src/H5PB.c | 67 |
1 files changed, 56 insertions, 11 deletions
@@ -1133,6 +1133,37 @@ done: } /* end H5PB_read() */ +static int +vfd_swmr_mdf_idx_entry_remove(H5F_shared_t *shared, uint64_t page) +{ + ptrdiff_t idx_idx, last_idx; + H5FD_vfd_swmr_idx_entry_t *idx_entry; + + idx_entry = vfd_swmr_pageno_to_mdf_idx_entry(shared->mdf_idx, + shared->mdf_idx_entries_used, page); + + if (idx_entry == NULL) + return 0; + + if (shared->vfd_swmr_writer && idx_entry->md_file_page_offset != 0 && + vfd_swmr_idx_entry_defer_free(shared, idx_entry) != 0) + return -1; + + idx_idx = idx_entry - shared->mdf_idx; + last_idx = shared->mdf_idx_entries_used - 1; + + shared->mdf_idx[idx_idx] = shared->mdf_idx[last_idx]; + if (shared->mdf_idx_entries_used > idx_idx + 1) { + const size_t ntocopy = + (size_t)(shared->mdf_idx_entries_used - (idx_idx + 1)); + memmove(&shared->mdf_idx[idx_idx], + &shared->mdf_idx[idx_idx + 1], + ntocopy * sizeof(shared->mdf_idx[idx_idx + 1])); + } + shared->mdf_idx_entries_used--; + return 0; +} + /*------------------------------------------------------------------------- * @@ -1239,22 +1270,34 @@ H5PB_remove_entry(H5F_shared_t *shared, haddr_t addr) HGOTO_ERROR(H5E_PAGEBUF, H5E_SYSTEM, FAIL, "forced eviction failed") - /* Do we need to remove the entry from the metadata file index in - * the VFD SWMR case? + /* We need to remove the entry from the shadow file index in + * the VFD SWMR case, so do that next. * - * Probably yes -- suppose a page is deallocated, and a multipage - * metadata entry is allocated at the same base address. This would - * change the metadata file entry size. + * If a multipage metadata entry is deallocated, and a new, single-page + * metadata entry is allocated at the same base address, then + * the old shadow index entry will still tell the size of the previous + * image, which is greater than a page, and a shadow-file flush will + * access bytes past the end of the entry's image. * - * However, this is sufficiently improbably that it doesn't cause - * problems (that I know of) at present. + * When we add code to allow entries + * to age out of the metadata file index, that may provide + * code that we can reuse to perform this invalidation. * - * Unless it does, hold off on this until we add code to allow entries - * to age out of the metadata file index, as that will give us the - * necessary infrastructure. + * It's also possible (I think) for the index-entry size to be set + * to one page, and then for a multipage entry to appear later at that + * same index entry. The recorded size will still say the same, but + * the image will be bigger. So the shadow file will never see the + * entire image written, just the first page of the image. * - * JRM -- 12/6/18 + * XXX The H5PB__evict_entry() call immediately prior should have called + * XXX vfd_swmr_mdf_idx_entry_remove() for this page. Need to + * XXX move this detailed comment and delete the redundant call to + * XXX vfd_swmr_mdf_idx_entry_remove(), no? */ + if (vfd_swmr_mdf_idx_entry_remove(shared, page) == -1) { + HGOTO_ERROR(H5E_PAGEBUF, H5E_SYSTEM, FAIL, + "failed to remove shadow index entry") + } } done: @@ -2369,6 +2412,8 @@ H5PB__evict_entry(H5F_shared_t *shared, H5PB_entry_t *entry_ptr, hbool_t force) /* remove the entry from the hash table */ H5PB__DELETE_FROM_INDEX(pb_ptr, entry_ptr, FAIL) + vfd_swmr_mdf_idx_entry_remove(shared, entry_ptr->page); + /* update stats for eviction */ H5PB__UPDATE_STATS_FOR_EVICTION(pb_ptr, entry_ptr) |