4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
27 #include <sys/zfs_context.h>
29 #include <sys/spa_impl.h>
33 #include <sys/dmu_tx.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/zio_compress.h>
38 #include <sys/dsl_scan.h>
41 * Enable/disable prefetching of dedup-ed blocks which are going to be freed.
43 int zfs_dedup_prefetch = 1;
45 SYSCTL_DECL(_vfs_zfs);
46 SYSCTL_NODE(_vfs_zfs, OID_AUTO, dedup, CTLFLAG_RW, 0, "ZFS DEDUP");
47 TUNABLE_INT("vfs.zfs.dedup.prefetch", &zfs_dedup_prefetch);
48 SYSCTL_INT(_vfs_zfs_dedup, OID_AUTO, prefetch, CTLFLAG_RW, &zfs_dedup_prefetch,
49 0, "Enable/disable prefetching of dedup-ed blocks which are going to be freed");
51 static const ddt_ops_t *ddt_ops[DDT_TYPES] = {
55 static const char *ddt_class_name[DDT_CLASSES] = {
62 ddt_object_create(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
65 spa_t *spa = ddt->ddt_spa;
66 objset_t *os = ddt->ddt_os;
67 uint64_t *objectp = &ddt->ddt_object[type][class];
68 boolean_t prehash = zio_checksum_table[ddt->ddt_checksum].ci_dedup;
69 char name[DDT_NAMELEN];
71 ddt_object_name(ddt, type, class, name);
73 ASSERT(*objectp == 0);
74 VERIFY(ddt_ops[type]->ddt_op_create(os, objectp, tx, prehash) == 0);
75 ASSERT(*objectp != 0);
77 VERIFY(zap_add(os, DMU_POOL_DIRECTORY_OBJECT, name,
78 sizeof (uint64_t), 1, objectp, tx) == 0);
80 VERIFY(zap_add(os, spa->spa_ddt_stat_object, name,
81 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
82 &ddt->ddt_histogram[type][class], tx) == 0);
86 ddt_object_destroy(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
89 spa_t *spa = ddt->ddt_spa;
90 objset_t *os = ddt->ddt_os;
91 uint64_t *objectp = &ddt->ddt_object[type][class];
93 char name[DDT_NAMELEN];
95 ddt_object_name(ddt, type, class, name);
97 ASSERT(*objectp != 0);
98 VERIFY(ddt_object_count(ddt, type, class, &count) == 0 && count == 0);
99 ASSERT(ddt_histogram_empty(&ddt->ddt_histogram[type][class]));
100 VERIFY(zap_remove(os, DMU_POOL_DIRECTORY_OBJECT, name, tx) == 0);
101 VERIFY(zap_remove(os, spa->spa_ddt_stat_object, name, tx) == 0);
102 VERIFY(ddt_ops[type]->ddt_op_destroy(os, *objectp, tx) == 0);
103 bzero(&ddt->ddt_object_stats[type][class], sizeof (ddt_object_t));
109 ddt_object_load(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
111 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
112 dmu_object_info_t doi;
114 char name[DDT_NAMELEN];
117 ddt_object_name(ddt, type, class, name);
119 error = zap_lookup(ddt->ddt_os, DMU_POOL_DIRECTORY_OBJECT, name,
120 sizeof (uint64_t), 1, &ddt->ddt_object[type][class]);
125 error = zap_lookup(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
126 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
127 &ddt->ddt_histogram[type][class]);
130 * Seed the cached statistics.
132 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0);
134 error = ddt_object_count(ddt, type, class, &count);
138 ddo->ddo_count = count;
139 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
140 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
147 ddt_object_sync(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
150 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
151 dmu_object_info_t doi;
153 char name[DDT_NAMELEN];
155 ddt_object_name(ddt, type, class, name);
157 VERIFY(zap_update(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
158 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
159 &ddt->ddt_histogram[type][class], tx) == 0);
162 * Cache DDT statistics; this is the only time they'll change.
164 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0);
165 VERIFY(ddt_object_count(ddt, type, class, &count) == 0);
167 ddo->ddo_count = count;
168 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
169 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
173 ddt_object_lookup(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
176 if (!ddt_object_exists(ddt, type, class))
177 return (SET_ERROR(ENOENT));
179 return (ddt_ops[type]->ddt_op_lookup(ddt->ddt_os,
180 ddt->ddt_object[type][class], dde));
184 ddt_object_prefetch(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
187 if (!ddt_object_exists(ddt, type, class))
190 ddt_ops[type]->ddt_op_prefetch(ddt->ddt_os,
191 ddt->ddt_object[type][class], dde);
195 ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
196 ddt_entry_t *dde, dmu_tx_t *tx)
198 ASSERT(ddt_object_exists(ddt, type, class));
200 return (ddt_ops[type]->ddt_op_update(ddt->ddt_os,
201 ddt->ddt_object[type][class], dde, tx));
205 ddt_object_remove(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
206 ddt_entry_t *dde, dmu_tx_t *tx)
208 ASSERT(ddt_object_exists(ddt, type, class));
210 return (ddt_ops[type]->ddt_op_remove(ddt->ddt_os,
211 ddt->ddt_object[type][class], dde, tx));
215 ddt_object_walk(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
216 uint64_t *walk, ddt_entry_t *dde)
218 ASSERT(ddt_object_exists(ddt, type, class));
220 return (ddt_ops[type]->ddt_op_walk(ddt->ddt_os,
221 ddt->ddt_object[type][class], dde, walk));
225 ddt_object_count(ddt_t *ddt, enum ddt_type type, enum ddt_class class, uint64_t *count)
227 ASSERT(ddt_object_exists(ddt, type, class));
229 return (ddt_ops[type]->ddt_op_count(ddt->ddt_os,
230 ddt->ddt_object[type][class], count));
234 ddt_object_info(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
235 dmu_object_info_t *doi)
237 if (!ddt_object_exists(ddt, type, class))
238 return (SET_ERROR(ENOENT));
240 return (dmu_object_info(ddt->ddt_os, ddt->ddt_object[type][class],
245 ddt_object_exists(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
247 return (!!ddt->ddt_object[type][class]);
251 ddt_object_name(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
254 (void) sprintf(name, DMU_POOL_DDT,
255 zio_checksum_table[ddt->ddt_checksum].ci_name,
256 ddt_ops[type]->ddt_op_name, ddt_class_name[class]);
260 ddt_bp_fill(const ddt_phys_t *ddp, blkptr_t *bp, uint64_t txg)
264 for (int d = 0; d < SPA_DVAS_PER_BP; d++)
265 bp->blk_dva[d] = ddp->ddp_dva[d];
266 BP_SET_BIRTH(bp, txg, ddp->ddp_phys_birth);
270 ddt_bp_create(enum zio_checksum checksum,
271 const ddt_key_t *ddk, const ddt_phys_t *ddp, blkptr_t *bp)
276 ddt_bp_fill(ddp, bp, ddp->ddp_phys_birth);
278 bp->blk_cksum = ddk->ddk_cksum;
281 BP_SET_LSIZE(bp, DDK_GET_LSIZE(ddk));
282 BP_SET_PSIZE(bp, DDK_GET_PSIZE(ddk));
283 BP_SET_COMPRESS(bp, DDK_GET_COMPRESS(ddk));
284 BP_SET_CHECKSUM(bp, checksum);
285 BP_SET_TYPE(bp, DMU_OT_DEDUP);
288 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
292 ddt_key_fill(ddt_key_t *ddk, const blkptr_t *bp)
294 ddk->ddk_cksum = bp->blk_cksum;
297 DDK_SET_LSIZE(ddk, BP_GET_LSIZE(bp));
298 DDK_SET_PSIZE(ddk, BP_GET_PSIZE(bp));
299 DDK_SET_COMPRESS(ddk, BP_GET_COMPRESS(bp));
303 ddt_phys_fill(ddt_phys_t *ddp, const blkptr_t *bp)
305 ASSERT(ddp->ddp_phys_birth == 0);
307 for (int d = 0; d < SPA_DVAS_PER_BP; d++)
308 ddp->ddp_dva[d] = bp->blk_dva[d];
309 ddp->ddp_phys_birth = BP_PHYSICAL_BIRTH(bp);
313 ddt_phys_clear(ddt_phys_t *ddp)
315 bzero(ddp, sizeof (*ddp));
319 ddt_phys_addref(ddt_phys_t *ddp)
325 ddt_phys_decref(ddt_phys_t *ddp)
327 ASSERT((int64_t)ddp->ddp_refcnt > 0);
332 ddt_phys_free(ddt_t *ddt, ddt_key_t *ddk, ddt_phys_t *ddp, uint64_t txg)
336 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
338 zio_free(ddt->ddt_spa, txg, &blk);
342 ddt_phys_select(const ddt_entry_t *dde, const blkptr_t *bp)
344 ddt_phys_t *ddp = (ddt_phys_t *)dde->dde_phys;
346 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
347 if (DVA_EQUAL(BP_IDENTITY(bp), &ddp->ddp_dva[0]) &&
348 BP_PHYSICAL_BIRTH(bp) == ddp->ddp_phys_birth)
355 ddt_phys_total_refcnt(const ddt_entry_t *dde)
359 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++)
360 refcnt += dde->dde_phys[p].ddp_refcnt;
366 ddt_stat_generate(ddt_t *ddt, ddt_entry_t *dde, ddt_stat_t *dds)
368 spa_t *spa = ddt->ddt_spa;
369 ddt_phys_t *ddp = dde->dde_phys;
370 ddt_key_t *ddk = &dde->dde_key;
371 uint64_t lsize = DDK_GET_LSIZE(ddk);
372 uint64_t psize = DDK_GET_PSIZE(ddk);
374 bzero(dds, sizeof (*dds));
376 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
378 uint64_t refcnt = ddp->ddp_refcnt;
380 if (ddp->ddp_phys_birth == 0)
383 for (int d = 0; d < SPA_DVAS_PER_BP; d++)
384 dsize += dva_get_dsize_sync(spa, &ddp->ddp_dva[d]);
386 dds->dds_blocks += 1;
387 dds->dds_lsize += lsize;
388 dds->dds_psize += psize;
389 dds->dds_dsize += dsize;
391 dds->dds_ref_blocks += refcnt;
392 dds->dds_ref_lsize += lsize * refcnt;
393 dds->dds_ref_psize += psize * refcnt;
394 dds->dds_ref_dsize += dsize * refcnt;
399 ddt_stat_add(ddt_stat_t *dst, const ddt_stat_t *src, uint64_t neg)
401 const uint64_t *s = (const uint64_t *)src;
402 uint64_t *d = (uint64_t *)dst;
403 uint64_t *d_end = (uint64_t *)(dst + 1);
405 ASSERT(neg == 0 || neg == -1ULL); /* add or subtract */
408 *d++ += (*s++ ^ neg) - neg;
412 ddt_stat_update(ddt_t *ddt, ddt_entry_t *dde, uint64_t neg)
415 ddt_histogram_t *ddh;
418 ddt_stat_generate(ddt, dde, &dds);
420 bucket = highbit(dds.dds_ref_blocks) - 1;
423 ddh = &ddt->ddt_histogram[dde->dde_type][dde->dde_class];
425 ddt_stat_add(&ddh->ddh_stat[bucket], &dds, neg);
429 ddt_histogram_add(ddt_histogram_t *dst, const ddt_histogram_t *src)
431 for (int h = 0; h < 64; h++)
432 ddt_stat_add(&dst->ddh_stat[h], &src->ddh_stat[h], 0);
436 ddt_histogram_stat(ddt_stat_t *dds, const ddt_histogram_t *ddh)
438 bzero(dds, sizeof (*dds));
440 for (int h = 0; h < 64; h++)
441 ddt_stat_add(dds, &ddh->ddh_stat[h], 0);
445 ddt_histogram_empty(const ddt_histogram_t *ddh)
447 const uint64_t *s = (const uint64_t *)ddh;
448 const uint64_t *s_end = (const uint64_t *)(ddh + 1);
458 ddt_get_dedup_object_stats(spa_t *spa, ddt_object_t *ddo_total)
460 /* Sum the statistics we cached in ddt_object_sync(). */
461 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
462 ddt_t *ddt = spa->spa_ddt[c];
463 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
464 for (enum ddt_class class = 0; class < DDT_CLASSES;
467 &ddt->ddt_object_stats[type][class];
468 ddo_total->ddo_count += ddo->ddo_count;
469 ddo_total->ddo_dspace += ddo->ddo_dspace;
470 ddo_total->ddo_mspace += ddo->ddo_mspace;
475 /* ... and compute the averages. */
476 if (ddo_total->ddo_count != 0) {
477 ddo_total->ddo_dspace /= ddo_total->ddo_count;
478 ddo_total->ddo_mspace /= ddo_total->ddo_count;
483 ddt_get_dedup_histogram(spa_t *spa, ddt_histogram_t *ddh)
485 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
486 ddt_t *ddt = spa->spa_ddt[c];
487 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
488 for (enum ddt_class class = 0; class < DDT_CLASSES;
490 ddt_histogram_add(ddh,
491 &ddt->ddt_histogram_cache[type][class]);
498 ddt_get_dedup_stats(spa_t *spa, ddt_stat_t *dds_total)
500 ddt_histogram_t *ddh_total;
502 ddh_total = kmem_zalloc(sizeof (ddt_histogram_t), KM_SLEEP);
503 ddt_get_dedup_histogram(spa, ddh_total);
504 ddt_histogram_stat(dds_total, ddh_total);
505 kmem_free(ddh_total, sizeof (ddt_histogram_t));
509 ddt_get_dedup_dspace(spa_t *spa)
511 ddt_stat_t dds_total = { 0 };
513 ddt_get_dedup_stats(spa, &dds_total);
514 return (dds_total.dds_ref_dsize - dds_total.dds_dsize);
518 ddt_get_pool_dedup_ratio(spa_t *spa)
520 ddt_stat_t dds_total = { 0 };
522 ddt_get_dedup_stats(spa, &dds_total);
523 if (dds_total.dds_dsize == 0)
526 return (dds_total.dds_ref_dsize * 100 / dds_total.dds_dsize);
530 ddt_ditto_copies_needed(ddt_t *ddt, ddt_entry_t *dde, ddt_phys_t *ddp_willref)
532 spa_t *spa = ddt->ddt_spa;
533 uint64_t total_refcnt = 0;
534 uint64_t ditto = spa->spa_dedup_ditto;
535 int total_copies = 0;
536 int desired_copies = 0;
538 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
539 ddt_phys_t *ddp = &dde->dde_phys[p];
540 zio_t *zio = dde->dde_lead_zio[p];
541 uint64_t refcnt = ddp->ddp_refcnt; /* committed refs */
543 refcnt += zio->io_parent_count; /* pending refs */
544 if (ddp == ddp_willref)
545 refcnt++; /* caller's ref */
547 total_refcnt += refcnt;
552 if (ditto == 0 || ditto > UINT32_MAX)
555 if (total_refcnt >= 1)
557 if (total_refcnt >= ditto)
559 if (total_refcnt >= ditto * ditto)
562 return (MAX(desired_copies, total_copies) - total_copies);
566 ddt_ditto_copies_present(ddt_entry_t *dde)
568 ddt_phys_t *ddp = &dde->dde_phys[DDT_PHYS_DITTO];
569 dva_t *dva = ddp->ddp_dva;
570 int copies = 0 - DVA_GET_GANG(dva);
572 for (int d = 0; d < SPA_DVAS_PER_BP; d++, dva++)
573 if (DVA_IS_VALID(dva))
576 ASSERT(copies >= 0 && copies < SPA_DVAS_PER_BP);
582 ddt_compress(void *src, uchar_t *dst, size_t s_len, size_t d_len)
584 uchar_t *version = dst++;
585 int cpfunc = ZIO_COMPRESS_ZLE;
586 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
589 ASSERT(d_len >= s_len + 1); /* no compression plus version byte */
591 c_len = ci->ci_compress(src, dst, s_len, d_len - 1, ci->ci_level);
593 if (c_len == s_len) {
594 cpfunc = ZIO_COMPRESS_OFF;
595 bcopy(src, dst, s_len);
598 *version = (ZFS_HOST_BYTEORDER & DDT_COMPRESS_BYTEORDER_MASK) | cpfunc;
604 ddt_decompress(uchar_t *src, void *dst, size_t s_len, size_t d_len)
606 uchar_t version = *src++;
607 int cpfunc = version & DDT_COMPRESS_FUNCTION_MASK;
608 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
610 if (ci->ci_decompress != NULL)
611 (void) ci->ci_decompress(src, dst, s_len, d_len, ci->ci_level);
613 bcopy(src, dst, d_len);
615 if ((version ^ ZFS_HOST_BYTEORDER) & DDT_COMPRESS_BYTEORDER_MASK)
616 byteswap_uint64_array(dst, d_len);
620 ddt_select_by_checksum(spa_t *spa, enum zio_checksum c)
622 return (spa->spa_ddt[c]);
626 ddt_select(spa_t *spa, const blkptr_t *bp)
628 return (spa->spa_ddt[BP_GET_CHECKSUM(bp)]);
632 ddt_enter(ddt_t *ddt)
634 mutex_enter(&ddt->ddt_lock);
640 mutex_exit(&ddt->ddt_lock);
644 ddt_alloc(const ddt_key_t *ddk)
648 dde = kmem_zalloc(sizeof (ddt_entry_t), KM_SLEEP);
649 cv_init(&dde->dde_cv, NULL, CV_DEFAULT, NULL);
657 ddt_free(ddt_entry_t *dde)
659 ASSERT(!dde->dde_loading);
661 for (int p = 0; p < DDT_PHYS_TYPES; p++)
662 ASSERT(dde->dde_lead_zio[p] == NULL);
664 if (dde->dde_repair_data != NULL)
665 zio_buf_free(dde->dde_repair_data,
666 DDK_GET_PSIZE(&dde->dde_key));
668 cv_destroy(&dde->dde_cv);
669 kmem_free(dde, sizeof (*dde));
673 ddt_remove(ddt_t *ddt, ddt_entry_t *dde)
675 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
677 avl_remove(&ddt->ddt_tree, dde);
682 ddt_lookup(ddt_t *ddt, const blkptr_t *bp, boolean_t add)
684 ddt_entry_t *dde, dde_search;
686 enum ddt_class class;
690 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
692 ddt_key_fill(&dde_search.dde_key, bp);
694 dde = avl_find(&ddt->ddt_tree, &dde_search, &where);
698 dde = ddt_alloc(&dde_search.dde_key);
699 avl_insert(&ddt->ddt_tree, dde, where);
702 while (dde->dde_loading)
703 cv_wait(&dde->dde_cv, &ddt->ddt_lock);
708 dde->dde_loading = B_TRUE;
714 for (type = 0; type < DDT_TYPES; type++) {
715 for (class = 0; class < DDT_CLASSES; class++) {
716 error = ddt_object_lookup(ddt, type, class, dde);
724 ASSERT(error == 0 || error == ENOENT);
728 ASSERT(dde->dde_loaded == B_FALSE);
729 ASSERT(dde->dde_loading == B_TRUE);
731 dde->dde_type = type; /* will be DDT_TYPES if no entry found */
732 dde->dde_class = class; /* will be DDT_CLASSES if no entry found */
733 dde->dde_loaded = B_TRUE;
734 dde->dde_loading = B_FALSE;
737 ddt_stat_update(ddt, dde, -1ULL);
739 cv_broadcast(&dde->dde_cv);
745 ddt_prefetch(spa_t *spa, const blkptr_t *bp)
750 if (!zfs_dedup_prefetch || bp == NULL || !BP_GET_DEDUP(bp))
754 * We only remove the DDT once all tables are empty and only
755 * prefetch dedup blocks when there are entries in the DDT.
756 * Thus no locking is required as the DDT can't disappear on us.
758 ddt = ddt_select(spa, bp);
759 ddt_key_fill(&dde.dde_key, bp);
761 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
762 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
763 ddt_object_prefetch(ddt, type, class, &dde);
769 ddt_entry_compare(const void *x1, const void *x2)
771 const ddt_entry_t *dde1 = x1;
772 const ddt_entry_t *dde2 = x2;
773 const uint64_t *u1 = (const uint64_t *)&dde1->dde_key;
774 const uint64_t *u2 = (const uint64_t *)&dde2->dde_key;
776 for (int i = 0; i < DDT_KEY_WORDS; i++) {
787 ddt_table_alloc(spa_t *spa, enum zio_checksum c)
791 ddt = kmem_zalloc(sizeof (*ddt), KM_SLEEP);
793 mutex_init(&ddt->ddt_lock, NULL, MUTEX_DEFAULT, NULL);
794 avl_create(&ddt->ddt_tree, ddt_entry_compare,
795 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
796 avl_create(&ddt->ddt_repair_tree, ddt_entry_compare,
797 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
798 ddt->ddt_checksum = c;
800 ddt->ddt_os = spa->spa_meta_objset;
806 ddt_table_free(ddt_t *ddt)
808 ASSERT(avl_numnodes(&ddt->ddt_tree) == 0);
809 ASSERT(avl_numnodes(&ddt->ddt_repair_tree) == 0);
810 avl_destroy(&ddt->ddt_tree);
811 avl_destroy(&ddt->ddt_repair_tree);
812 mutex_destroy(&ddt->ddt_lock);
813 kmem_free(ddt, sizeof (*ddt));
817 ddt_create(spa_t *spa)
819 spa->spa_dedup_checksum = ZIO_DEDUPCHECKSUM;
821 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++)
822 spa->spa_ddt[c] = ddt_table_alloc(spa, c);
832 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
833 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1,
834 &spa->spa_ddt_stat_object);
837 return (error == ENOENT ? 0 : error);
839 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
840 ddt_t *ddt = spa->spa_ddt[c];
841 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
842 for (enum ddt_class class = 0; class < DDT_CLASSES;
844 error = ddt_object_load(ddt, type, class);
845 if (error != 0 && error != ENOENT)
851 * Seed the cached histograms.
853 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
854 sizeof (ddt->ddt_histogram));
861 ddt_unload(spa_t *spa)
863 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
864 if (spa->spa_ddt[c]) {
865 ddt_table_free(spa->spa_ddt[c]);
866 spa->spa_ddt[c] = NULL;
872 ddt_class_contains(spa_t *spa, enum ddt_class max_class, const blkptr_t *bp)
877 if (!BP_GET_DEDUP(bp))
880 if (max_class == DDT_CLASS_UNIQUE)
883 ddt = spa->spa_ddt[BP_GET_CHECKSUM(bp)];
885 ddt_key_fill(&dde.dde_key, bp);
887 for (enum ddt_type type = 0; type < DDT_TYPES; type++)
888 for (enum ddt_class class = 0; class <= max_class; class++)
889 if (ddt_object_lookup(ddt, type, class, &dde) == 0)
896 ddt_repair_start(ddt_t *ddt, const blkptr_t *bp)
901 ddt_key_fill(&ddk, bp);
903 dde = ddt_alloc(&ddk);
905 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
906 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
908 * We can only do repair if there are multiple copies
909 * of the block. For anything in the UNIQUE class,
910 * there's definitely only one copy, so don't even try.
912 if (class != DDT_CLASS_UNIQUE &&
913 ddt_object_lookup(ddt, type, class, dde) == 0)
918 bzero(dde->dde_phys, sizeof (dde->dde_phys));
924 ddt_repair_done(ddt_t *ddt, ddt_entry_t *dde)
930 if (dde->dde_repair_data != NULL && spa_writeable(ddt->ddt_spa) &&
931 avl_find(&ddt->ddt_repair_tree, dde, &where) == NULL)
932 avl_insert(&ddt->ddt_repair_tree, dde, where);
940 ddt_repair_entry_done(zio_t *zio)
942 ddt_entry_t *rdde = zio->io_private;
948 ddt_repair_entry(ddt_t *ddt, ddt_entry_t *dde, ddt_entry_t *rdde, zio_t *rio)
950 ddt_phys_t *ddp = dde->dde_phys;
951 ddt_phys_t *rddp = rdde->dde_phys;
952 ddt_key_t *ddk = &dde->dde_key;
953 ddt_key_t *rddk = &rdde->dde_key;
957 zio = zio_null(rio, rio->io_spa, NULL,
958 ddt_repair_entry_done, rdde, rio->io_flags);
960 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++, rddp++) {
961 if (ddp->ddp_phys_birth == 0 ||
962 ddp->ddp_phys_birth != rddp->ddp_phys_birth ||
963 bcmp(ddp->ddp_dva, rddp->ddp_dva, sizeof (ddp->ddp_dva)))
965 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
966 zio_nowait(zio_rewrite(zio, zio->io_spa, 0, &blk,
967 rdde->dde_repair_data, DDK_GET_PSIZE(rddk), NULL, NULL,
968 ZIO_PRIORITY_SYNC_WRITE, ZIO_DDT_CHILD_FLAGS(zio), NULL));
975 ddt_repair_table(ddt_t *ddt, zio_t *rio)
977 spa_t *spa = ddt->ddt_spa;
978 ddt_entry_t *dde, *rdde_next, *rdde;
979 avl_tree_t *t = &ddt->ddt_repair_tree;
982 if (spa_sync_pass(spa) > 1)
986 for (rdde = avl_first(t); rdde != NULL; rdde = rdde_next) {
987 rdde_next = AVL_NEXT(t, rdde);
988 avl_remove(&ddt->ddt_repair_tree, rdde);
990 ddt_bp_create(ddt->ddt_checksum, &rdde->dde_key, NULL, &blk);
991 dde = ddt_repair_start(ddt, &blk);
992 ddt_repair_entry(ddt, dde, rdde, rio);
993 ddt_repair_done(ddt, dde);
1000 ddt_sync_entry(ddt_t *ddt, ddt_entry_t *dde, dmu_tx_t *tx, uint64_t txg)
1002 dsl_pool_t *dp = ddt->ddt_spa->spa_dsl_pool;
1003 ddt_phys_t *ddp = dde->dde_phys;
1004 ddt_key_t *ddk = &dde->dde_key;
1005 enum ddt_type otype = dde->dde_type;
1006 enum ddt_type ntype = DDT_TYPE_CURRENT;
1007 enum ddt_class oclass = dde->dde_class;
1008 enum ddt_class nclass;
1009 uint64_t total_refcnt = 0;
1011 ASSERT(dde->dde_loaded);
1012 ASSERT(!dde->dde_loading);
1014 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1015 ASSERT(dde->dde_lead_zio[p] == NULL);
1016 ASSERT((int64_t)ddp->ddp_refcnt >= 0);
1017 if (ddp->ddp_phys_birth == 0) {
1018 ASSERT(ddp->ddp_refcnt == 0);
1021 if (p == DDT_PHYS_DITTO) {
1022 if (ddt_ditto_copies_needed(ddt, dde, NULL) == 0)
1023 ddt_phys_free(ddt, ddk, ddp, txg);
1026 if (ddp->ddp_refcnt == 0)
1027 ddt_phys_free(ddt, ddk, ddp, txg);
1028 total_refcnt += ddp->ddp_refcnt;
1031 if (dde->dde_phys[DDT_PHYS_DITTO].ddp_phys_birth != 0)
1032 nclass = DDT_CLASS_DITTO;
1033 else if (total_refcnt > 1)
1034 nclass = DDT_CLASS_DUPLICATE;
1036 nclass = DDT_CLASS_UNIQUE;
1038 if (otype != DDT_TYPES &&
1039 (otype != ntype || oclass != nclass || total_refcnt == 0)) {
1040 VERIFY(ddt_object_remove(ddt, otype, oclass, dde, tx) == 0);
1041 ASSERT(ddt_object_lookup(ddt, otype, oclass, dde) == ENOENT);
1044 if (total_refcnt != 0) {
1045 dde->dde_type = ntype;
1046 dde->dde_class = nclass;
1047 ddt_stat_update(ddt, dde, 0);
1048 if (!ddt_object_exists(ddt, ntype, nclass))
1049 ddt_object_create(ddt, ntype, nclass, tx);
1050 VERIFY(ddt_object_update(ddt, ntype, nclass, dde, tx) == 0);
1053 * If the class changes, the order that we scan this bp
1054 * changes. If it decreases, we could miss it, so
1055 * scan it right now. (This covers both class changing
1056 * while we are doing ddt_walk(), and when we are
1059 if (nclass < oclass) {
1060 dsl_scan_ddt_entry(dp->dp_scan,
1061 ddt->ddt_checksum, dde, tx);
1067 ddt_sync_table(ddt_t *ddt, dmu_tx_t *tx, uint64_t txg)
1069 spa_t *spa = ddt->ddt_spa;
1071 void *cookie = NULL;
1073 if (avl_numnodes(&ddt->ddt_tree) == 0)
1076 ASSERT(spa->spa_uberblock.ub_version >= SPA_VERSION_DEDUP);
1078 if (spa->spa_ddt_stat_object == 0) {
1079 spa->spa_ddt_stat_object = zap_create_link(ddt->ddt_os,
1080 DMU_OT_DDT_STATS, DMU_POOL_DIRECTORY_OBJECT,
1081 DMU_POOL_DDT_STATS, tx);
1084 while ((dde = avl_destroy_nodes(&ddt->ddt_tree, &cookie)) != NULL) {
1085 ddt_sync_entry(ddt, dde, tx, txg);
1089 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
1090 uint64_t add, count = 0;
1091 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
1092 if (ddt_object_exists(ddt, type, class)) {
1093 ddt_object_sync(ddt, type, class, tx);
1094 VERIFY(ddt_object_count(ddt, type, class,
1099 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
1100 if (count == 0 && ddt_object_exists(ddt, type, class))
1101 ddt_object_destroy(ddt, type, class, tx);
1105 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
1106 sizeof (ddt->ddt_histogram));
1110 ddt_sync(spa_t *spa, uint64_t txg)
1113 zio_t *rio = zio_root(spa, NULL, NULL,
1114 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
1116 ASSERT(spa_syncing_txg(spa) == txg);
1118 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1120 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
1121 ddt_t *ddt = spa->spa_ddt[c];
1124 ddt_sync_table(ddt, tx, txg);
1125 ddt_repair_table(ddt, rio);
1128 (void) zio_wait(rio);
1134 ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde)
1139 ddt_t *ddt = spa->spa_ddt[ddb->ddb_checksum];
1141 if (ddt_object_exists(ddt, ddb->ddb_type,
1143 error = ddt_object_walk(ddt,
1144 ddb->ddb_type, ddb->ddb_class,
1145 &ddb->ddb_cursor, dde);
1147 dde->dde_type = ddb->ddb_type;
1148 dde->dde_class = ddb->ddb_class;
1151 if (error != ENOENT)
1153 ddb->ddb_cursor = 0;
1154 } while (++ddb->ddb_checksum < ZIO_CHECKSUM_FUNCTIONS);
1155 ddb->ddb_checksum = 0;
1156 } while (++ddb->ddb_type < DDT_TYPES);
1158 } while (++ddb->ddb_class < DDT_CLASSES);
1160 return (SET_ERROR(ENOENT));