Print this page
1862 incremental zfs receive fails for sparse file > 8PB
dmu_tx_count_free is doing a horrible over-estimation of used memory. It
assumes that the file is fully non-sparse and calculates a worst-case estimate
of how much memory is needed to hold all metadata for the file. If a large
hole needs to be freed, the estimation goes into the TB-range, which obviously
fails later on.
This patch tries to calculate a more realistic estimate by counting the l1
blocks (the loop for this is already present) and assumes a worst-case
distribution of those blocks over the full length given.
Reviewed by: Matt Ahrens <matthew.ahrens@delphix.com>
Reviewed by: Simon Klinkert <klinkert@webgods.de>

*** 427,436 **** --- 427,437 ---- uint64_t space = 0, unref = 0, skipped = 0; dnode_t *dn = txh->txh_dnode; dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; spa_t *spa = txh->txh_tx->tx_pool->dp_spa; int epbs; + uint64_t l0span = 0, nl1blks = 0; if (dn->dn_nlevels == 0) return; /*
*** 459,468 **** --- 460,470 ---- } if (blkid + nblks > dn->dn_maxblkid) nblks = dn->dn_maxblkid - blkid; } + l0span = nblks; /* save for later use to calc level > 1 overhead */ if (dn->dn_nlevels == 1) { int i; for (i = 0; i < nblks; i++) { blkptr_t *bp = dn->dn_phys->dn_blkptr; ASSERT3U(blkid + i, <, dn->dn_nblkptr);
*** 471,498 **** dprintf_bp(bp, "can free old%s", ""); space += bp_get_dsize(spa, bp); } unref += BP_GET_ASIZE(bp); } nblks = 0; } - /* - * Add in memory requirements of higher-level indirects. - * This assumes a worst-possible scenario for dn_nlevels. - */ - { - uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs); - int level = (dn->dn_nlevels > 1) ? 2 : 1; - - while (level++ < DN_MAX_LEVELS) { - txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift; - blkcnt = 1 + (blkcnt >> epbs); - } - ASSERT(blkcnt <= dn->dn_nblkptr); - } - lastblk = blkid + nblks - 1; while (nblks) { dmu_buf_impl_t *dbuf; uint64_t ibyte, new_blkid; int epb = 1 << epbs; --- 473,486 ---- dprintf_bp(bp, "can free old%s", ""); space += bp_get_dsize(spa, bp); } unref += BP_GET_ASIZE(bp); } + nl1blks = 1; nblks = 0; } lastblk = blkid + nblks - 1; while (nblks) { dmu_buf_impl_t *dbuf; uint64_t ibyte, new_blkid; int epb = 1 << epbs;
*** 559,573 **** --- 547,585 ---- } unref += BP_GET_ASIZE(bp); } dbuf_rele(dbuf, FTAG); + ++nl1blks; blkid += tochk; nblks -= tochk; } rw_exit(&dn->dn_struct_rwlock); + /* + * Add in memory requirements of higher-level indirects. + * This assumes a worst-possible scenario for dn_nlevels and a + * worst-possible distribution of l1-blocks over the region to free. + */ + { + uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs); + int level = 2; + /* + * Here we don't use DN_MAX_LEVEL, but calculate it with the + * given datablkshift and indblkshift. This makes the + * difference between 19 and 8 on large files. + */ + int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) / + (dn->dn_indblkshift - SPA_BLKPTRSHIFT); + + while (level++ < maxlevel) { + txh->txh_memory_tohold += MIN(blkcnt, (nl1blks >> epbs)) + << dn->dn_indblkshift; + blkcnt = 1 + (blkcnt >> epbs); + } + } + /* account for new level 1 indirect blocks that might show up */ if (skipped > 0) { txh->txh_fudge += skipped << dn->dn_indblkshift; skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs); txh->txh_memory_tohold += skipped << dn->dn_indblkshift;