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) 2012 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 static const ddt_ops_t *ddt_ops[DDT_TYPES] = {
49 static const char *ddt_class_name[DDT_CLASSES] = {
56 ddt_object_create(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
59 spa_t *spa = ddt->ddt_spa;
60 objset_t *os = ddt->ddt_os;
61 uint64_t *objectp = &ddt->ddt_object[type][class];
62 boolean_t prehash = zio_checksum_table[ddt->ddt_checksum].ci_dedup;
63 char name[DDT_NAMELEN];
65 ddt_object_name(ddt, type, class, name);
67 ASSERT(*objectp == 0);
68 VERIFY(ddt_ops[type]->ddt_op_create(os, objectp, tx, prehash) == 0);
69 ASSERT(*objectp != 0);
71 VERIFY(zap_add(os, DMU_POOL_DIRECTORY_OBJECT, name,
72 sizeof (uint64_t), 1, objectp, tx) == 0);
74 VERIFY(zap_add(os, spa->spa_ddt_stat_object, name,
75 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
76 &ddt->ddt_histogram[type][class], tx) == 0);
80 ddt_object_destroy(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
83 spa_t *spa = ddt->ddt_spa;
84 objset_t *os = ddt->ddt_os;
85 uint64_t *objectp = &ddt->ddt_object[type][class];
87 char name[DDT_NAMELEN];
89 ddt_object_name(ddt, type, class, name);
91 ASSERT(*objectp != 0);
92 ASSERT(ddt_histogram_empty(&ddt->ddt_histogram[type][class]));
93 VERIFY(ddt_object_count(ddt, type, class, &count) == 0 && count == 0);
94 VERIFY(zap_remove(os, DMU_POOL_DIRECTORY_OBJECT, name, tx) == 0);
95 VERIFY(zap_remove(os, spa->spa_ddt_stat_object, name, tx) == 0);
96 VERIFY(ddt_ops[type]->ddt_op_destroy(os, *objectp, tx) == 0);
97 bzero(&ddt->ddt_object_stats[type][class], sizeof (ddt_object_t));
103 ddt_object_load(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
105 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
106 dmu_object_info_t doi;
108 char name[DDT_NAMELEN];
111 ddt_object_name(ddt, type, class, name);
113 error = zap_lookup(ddt->ddt_os, DMU_POOL_DIRECTORY_OBJECT, name,
114 sizeof (uint64_t), 1, &ddt->ddt_object[type][class]);
119 error = zap_lookup(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
120 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
121 &ddt->ddt_histogram[type][class]);
124 * Seed the cached statistics.
126 error = ddt_object_info(ddt, type, class, &doi);
130 error = ddt_object_count(ddt, type, class, &count);
134 ddo->ddo_count = count;
135 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
136 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
143 ddt_object_sync(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
146 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
147 dmu_object_info_t doi;
149 char name[DDT_NAMELEN];
151 ddt_object_name(ddt, type, class, name);
153 VERIFY(zap_update(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
154 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
155 &ddt->ddt_histogram[type][class], tx) == 0);
158 * Cache DDT statistics; this is the only time they'll change.
160 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0);
161 VERIFY(ddt_object_count(ddt, type, class, &count) == 0);
163 ddo->ddo_count = count;
164 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
165 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
169 ddt_object_lookup(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
172 if (!ddt_object_exists(ddt, type, class))
175 return (ddt_ops[type]->ddt_op_lookup(ddt->ddt_os,
176 ddt->ddt_object[type][class], dde));
180 ddt_object_prefetch(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
183 if (!ddt_object_exists(ddt, type, class))
186 ddt_ops[type]->ddt_op_prefetch(ddt->ddt_os,
187 ddt->ddt_object[type][class], dde);
191 ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
192 ddt_entry_t *dde, dmu_tx_t *tx)
194 ASSERT(ddt_object_exists(ddt, type, class));
196 return (ddt_ops[type]->ddt_op_update(ddt->ddt_os,
197 ddt->ddt_object[type][class], dde, tx));
201 ddt_object_remove(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
202 ddt_entry_t *dde, dmu_tx_t *tx)
204 ASSERT(ddt_object_exists(ddt, type, class));
206 return (ddt_ops[type]->ddt_op_remove(ddt->ddt_os,
207 ddt->ddt_object[type][class], dde, tx));
211 ddt_object_walk(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
212 uint64_t *walk, ddt_entry_t *dde)
214 ASSERT(ddt_object_exists(ddt, type, class));
216 return (ddt_ops[type]->ddt_op_walk(ddt->ddt_os,
217 ddt->ddt_object[type][class], dde, walk));
221 ddt_object_count(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
224 ASSERT(ddt_object_exists(ddt, type, class));
226 return (ddt_ops[type]->ddt_op_count(ddt->ddt_os,
227 ddt->ddt_object[type][class], count));
231 ddt_object_info(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
232 dmu_object_info_t *doi)
234 if (!ddt_object_exists(ddt, type, class))
237 return (dmu_object_info(ddt->ddt_os, ddt->ddt_object[type][class],
242 ddt_object_exists(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
244 return (!!ddt->ddt_object[type][class]);
248 ddt_object_name(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
251 (void) sprintf(name, DMU_POOL_DDT,
252 zio_checksum_table[ddt->ddt_checksum].ci_name,
253 ddt_ops[type]->ddt_op_name, ddt_class_name[class]);
257 ddt_bp_fill(const ddt_phys_t *ddp, blkptr_t *bp, uint64_t txg)
262 for (d = 0; d < SPA_DVAS_PER_BP; d++)
263 bp->blk_dva[d] = ddp->ddp_dva[d];
264 BP_SET_BIRTH(bp, txg, ddp->ddp_phys_birth);
268 ddt_bp_create(enum zio_checksum checksum,
269 const ddt_key_t *ddk, const ddt_phys_t *ddp, blkptr_t *bp)
274 ddt_bp_fill(ddp, bp, ddp->ddp_phys_birth);
276 bp->blk_cksum = ddk->ddk_cksum;
279 BP_SET_LSIZE(bp, DDK_GET_LSIZE(ddk));
280 BP_SET_PSIZE(bp, DDK_GET_PSIZE(ddk));
281 BP_SET_COMPRESS(bp, DDK_GET_COMPRESS(ddk));
282 BP_SET_CHECKSUM(bp, checksum);
283 BP_SET_TYPE(bp, DMU_OT_DEDUP);
286 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
290 ddt_key_fill(ddt_key_t *ddk, const blkptr_t *bp)
292 ddk->ddk_cksum = bp->blk_cksum;
295 DDK_SET_LSIZE(ddk, BP_GET_LSIZE(bp));
296 DDK_SET_PSIZE(ddk, BP_GET_PSIZE(bp));
297 DDK_SET_COMPRESS(ddk, BP_GET_COMPRESS(bp));
301 ddt_phys_fill(ddt_phys_t *ddp, const blkptr_t *bp)
304 ASSERT(ddp->ddp_phys_birth == 0);
306 for (d = 0; d < SPA_DVAS_PER_BP; d++)
307 ddp->ddp_dva[d] = bp->blk_dva[d];
308 ddp->ddp_phys_birth = BP_PHYSICAL_BIRTH(bp);
312 ddt_phys_clear(ddt_phys_t *ddp)
314 bzero(ddp, sizeof (*ddp));
318 ddt_phys_addref(ddt_phys_t *ddp)
324 ddt_phys_decref(ddt_phys_t *ddp)
326 ASSERT((int64_t)ddp->ddp_refcnt > 0);
331 ddt_phys_free(ddt_t *ddt, ddt_key_t *ddk, ddt_phys_t *ddp, uint64_t txg)
335 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
337 zio_free(ddt->ddt_spa, txg, &blk);
341 ddt_phys_select(const ddt_entry_t *dde, const blkptr_t *bp)
343 ddt_phys_t *ddp = (ddt_phys_t *)dde->dde_phys;
346 for (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)
360 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++)
361 refcnt += dde->dde_phys[p].ddp_refcnt;
367 ddt_stat_generate(ddt_t *ddt, ddt_entry_t *dde, ddt_stat_t *dds)
369 spa_t *spa = ddt->ddt_spa;
370 ddt_phys_t *ddp = dde->dde_phys;
371 ddt_key_t *ddk = &dde->dde_key;
372 uint64_t lsize = DDK_GET_LSIZE(ddk);
373 uint64_t psize = DDK_GET_PSIZE(ddk);
376 bzero(dds, sizeof (*dds));
378 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
380 uint64_t refcnt = ddp->ddp_refcnt;
382 if (ddp->ddp_phys_birth == 0)
385 for (d = 0; d < SPA_DVAS_PER_BP; d++)
386 dsize += dva_get_dsize_sync(spa, &ddp->ddp_dva[d]);
388 dds->dds_blocks += 1;
389 dds->dds_lsize += lsize;
390 dds->dds_psize += psize;
391 dds->dds_dsize += dsize;
393 dds->dds_ref_blocks += refcnt;
394 dds->dds_ref_lsize += lsize * refcnt;
395 dds->dds_ref_psize += psize * refcnt;
396 dds->dds_ref_dsize += dsize * refcnt;
401 ddt_stat_add(ddt_stat_t *dst, const ddt_stat_t *src, uint64_t neg)
403 const uint64_t *s = (const uint64_t *)src;
404 uint64_t *d = (uint64_t *)dst;
405 uint64_t *d_end = (uint64_t *)(dst + 1);
407 ASSERT(neg == 0 || neg == -1ULL); /* add or subtract */
410 *d++ += (*s++ ^ neg) - neg;
414 ddt_stat_update(ddt_t *ddt, ddt_entry_t *dde, uint64_t neg)
417 ddt_histogram_t *ddh;
420 ddt_stat_generate(ddt, dde, &dds);
422 bucket = highbit(dds.dds_ref_blocks) - 1;
425 ddh = &ddt->ddt_histogram[dde->dde_type][dde->dde_class];
427 ddt_stat_add(&ddh->ddh_stat[bucket], &dds, neg);
431 ddt_histogram_add(ddt_histogram_t *dst, const ddt_histogram_t *src)
435 for (h = 0; h < 64; h++)
436 ddt_stat_add(&dst->ddh_stat[h], &src->ddh_stat[h], 0);
440 ddt_histogram_stat(ddt_stat_t *dds, const ddt_histogram_t *ddh)
444 bzero(dds, sizeof (*dds));
446 for (h = 0; h < 64; h++)
447 ddt_stat_add(dds, &ddh->ddh_stat[h], 0);
451 ddt_histogram_empty(const ddt_histogram_t *ddh)
453 const uint64_t *s = (const uint64_t *)ddh;
454 const uint64_t *s_end = (const uint64_t *)(ddh + 1);
464 ddt_get_dedup_object_stats(spa_t *spa, ddt_object_t *ddo_total)
468 enum ddt_class class;
470 /* Sum the statistics we cached in ddt_object_sync(). */
471 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
472 ddt_t *ddt = spa->spa_ddt[c];
473 for (type = 0; type < DDT_TYPES; type++) {
474 for (class = 0; class < DDT_CLASSES;
477 &ddt->ddt_object_stats[type][class];
478 ddo_total->ddo_count += ddo->ddo_count;
479 ddo_total->ddo_dspace += ddo->ddo_dspace;
480 ddo_total->ddo_mspace += ddo->ddo_mspace;
485 /* ... and compute the averages. */
486 if (ddo_total->ddo_count != 0) {
487 ddo_total->ddo_dspace /= ddo_total->ddo_count;
488 ddo_total->ddo_mspace /= ddo_total->ddo_count;
493 ddt_get_dedup_histogram(spa_t *spa, ddt_histogram_t *ddh)
497 enum ddt_class class;
499 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
500 ddt_t *ddt = spa->spa_ddt[c];
501 for (type = 0; type < DDT_TYPES; type++) {
502 for (class = 0; class < DDT_CLASSES;
504 ddt_histogram_add(ddh,
505 &ddt->ddt_histogram_cache[type][class]);
512 ddt_get_dedup_stats(spa_t *spa, ddt_stat_t *dds_total)
514 ddt_histogram_t *ddh_total;
516 /* XXX: Move to a slab */
517 ddh_total = kmem_zalloc(sizeof (ddt_histogram_t), KM_PUSHPAGE);
518 ddt_get_dedup_histogram(spa, ddh_total);
519 ddt_histogram_stat(dds_total, ddh_total);
520 kmem_free(ddh_total, sizeof (ddt_histogram_t));
524 ddt_get_dedup_dspace(spa_t *spa)
526 ddt_stat_t dds_total = { 0 };
528 ddt_get_dedup_stats(spa, &dds_total);
529 return (dds_total.dds_ref_dsize - dds_total.dds_dsize);
533 ddt_get_pool_dedup_ratio(spa_t *spa)
535 ddt_stat_t dds_total = { 0 };
537 ddt_get_dedup_stats(spa, &dds_total);
538 if (dds_total.dds_dsize == 0)
541 return (dds_total.dds_ref_dsize * 100 / dds_total.dds_dsize);
545 ddt_ditto_copies_needed(ddt_t *ddt, ddt_entry_t *dde, ddt_phys_t *ddp_willref)
547 spa_t *spa = ddt->ddt_spa;
548 uint64_t total_refcnt = 0;
549 uint64_t ditto = spa->spa_dedup_ditto;
550 int total_copies = 0;
551 int desired_copies = 0;
554 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
555 ddt_phys_t *ddp = &dde->dde_phys[p];
556 zio_t *zio = dde->dde_lead_zio[p];
557 uint64_t refcnt = ddp->ddp_refcnt; /* committed refs */
559 refcnt += zio->io_parent_count; /* pending refs */
560 if (ddp == ddp_willref)
561 refcnt++; /* caller's ref */
563 total_refcnt += refcnt;
568 if (ditto == 0 || ditto > UINT32_MAX)
571 if (total_refcnt >= 1)
573 if (total_refcnt >= ditto)
575 if (total_refcnt >= ditto * ditto)
578 return (MAX(desired_copies, total_copies) - total_copies);
582 ddt_ditto_copies_present(ddt_entry_t *dde)
584 ddt_phys_t *ddp = &dde->dde_phys[DDT_PHYS_DITTO];
585 dva_t *dva = ddp->ddp_dva;
586 int copies = 0 - DVA_GET_GANG(dva);
589 for (d = 0; d < SPA_DVAS_PER_BP; d++, dva++)
590 if (DVA_IS_VALID(dva))
593 ASSERT(copies >= 0 && copies < SPA_DVAS_PER_BP);
599 ddt_compress(void *src, uchar_t *dst, size_t s_len, size_t d_len)
601 uchar_t *version = dst++;
602 int cpfunc = ZIO_COMPRESS_ZLE;
603 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
606 ASSERT(d_len >= s_len + 1); /* no compression plus version byte */
608 c_len = ci->ci_compress(src, dst, s_len, d_len - 1, ci->ci_level);
610 if (c_len == s_len) {
611 cpfunc = ZIO_COMPRESS_OFF;
612 bcopy(src, dst, s_len);
615 *version = (ZFS_HOST_BYTEORDER & DDT_COMPRESS_BYTEORDER_MASK) | cpfunc;
621 ddt_decompress(uchar_t *src, void *dst, size_t s_len, size_t d_len)
623 uchar_t version = *src++;
624 int cpfunc = version & DDT_COMPRESS_FUNCTION_MASK;
625 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
627 if (ci->ci_decompress != NULL)
628 (void) ci->ci_decompress(src, dst, s_len, d_len, ci->ci_level);
630 bcopy(src, dst, d_len);
632 if ((version ^ ZFS_HOST_BYTEORDER) & DDT_COMPRESS_BYTEORDER_MASK)
633 byteswap_uint64_array(dst, d_len);
637 ddt_select_by_checksum(spa_t *spa, enum zio_checksum c)
639 return (spa->spa_ddt[c]);
643 ddt_select(spa_t *spa, const blkptr_t *bp)
645 return (spa->spa_ddt[BP_GET_CHECKSUM(bp)]);
649 ddt_enter(ddt_t *ddt)
651 mutex_enter(&ddt->ddt_lock);
657 mutex_exit(&ddt->ddt_lock);
661 ddt_alloc(const ddt_key_t *ddk)
665 /* XXX: Move to a slab */
666 dde = kmem_zalloc(sizeof (ddt_entry_t), KM_PUSHPAGE);
667 cv_init(&dde->dde_cv, NULL, CV_DEFAULT, NULL);
675 ddt_free(ddt_entry_t *dde)
679 ASSERT(!dde->dde_loading);
681 for (p = 0; p < DDT_PHYS_TYPES; p++)
682 ASSERT(dde->dde_lead_zio[p] == NULL);
684 if (dde->dde_repair_data != NULL)
685 zio_buf_free(dde->dde_repair_data,
686 DDK_GET_PSIZE(&dde->dde_key));
688 cv_destroy(&dde->dde_cv);
689 kmem_free(dde, sizeof (*dde));
693 ddt_remove(ddt_t *ddt, ddt_entry_t *dde)
695 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
697 avl_remove(&ddt->ddt_tree, dde);
702 ddt_lookup(ddt_t *ddt, const blkptr_t *bp, boolean_t add)
704 ddt_entry_t *dde, dde_search;
706 enum ddt_class class;
710 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
712 ddt_key_fill(&dde_search.dde_key, bp);
714 dde = avl_find(&ddt->ddt_tree, &dde_search, &where);
718 dde = ddt_alloc(&dde_search.dde_key);
719 avl_insert(&ddt->ddt_tree, dde, where);
722 while (dde->dde_loading)
723 cv_wait(&dde->dde_cv, &ddt->ddt_lock);
728 dde->dde_loading = B_TRUE;
734 for (type = 0; type < DDT_TYPES; type++) {
735 for (class = 0; class < DDT_CLASSES; class++) {
736 error = ddt_object_lookup(ddt, type, class, dde);
744 ASSERT(error == 0 || error == ENOENT);
748 ASSERT(dde->dde_loaded == B_FALSE);
749 ASSERT(dde->dde_loading == B_TRUE);
751 dde->dde_type = type; /* will be DDT_TYPES if no entry found */
752 dde->dde_class = class; /* will be DDT_CLASSES if no entry found */
753 dde->dde_loaded = B_TRUE;
754 dde->dde_loading = B_FALSE;
757 ddt_stat_update(ddt, dde, -1ULL);
759 cv_broadcast(&dde->dde_cv);
765 ddt_prefetch(spa_t *spa, const blkptr_t *bp)
770 enum ddt_class class;
772 if (!zfs_dedup_prefetch || bp == NULL || !BP_GET_DEDUP(bp))
776 * We only remove the DDT once all tables are empty and only
777 * prefetch dedup blocks when there are entries in the DDT.
778 * Thus no locking is required as the DDT can't disappear on us.
780 ddt = ddt_select(spa, bp);
781 ddt_key_fill(&dde.dde_key, bp);
783 for (type = 0; type < DDT_TYPES; type++) {
784 for (class = 0; class < DDT_CLASSES; class++) {
785 ddt_object_prefetch(ddt, type, class, &dde);
791 ddt_entry_compare(const void *x1, const void *x2)
793 const ddt_entry_t *dde1 = x1;
794 const ddt_entry_t *dde2 = x2;
795 const uint64_t *u1 = (const uint64_t *)&dde1->dde_key;
796 const uint64_t *u2 = (const uint64_t *)&dde2->dde_key;
799 for (i = 0; i < DDT_KEY_WORDS; i++) {
810 ddt_table_alloc(spa_t *spa, enum zio_checksum c)
814 /* XXX: Move to a slab */
815 ddt = kmem_zalloc(sizeof (*ddt), KM_PUSHPAGE | KM_NODEBUG);
817 mutex_init(&ddt->ddt_lock, NULL, MUTEX_DEFAULT, NULL);
818 avl_create(&ddt->ddt_tree, ddt_entry_compare,
819 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
820 avl_create(&ddt->ddt_repair_tree, ddt_entry_compare,
821 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
822 ddt->ddt_checksum = c;
824 ddt->ddt_os = spa->spa_meta_objset;
830 ddt_table_free(ddt_t *ddt)
832 ASSERT(avl_numnodes(&ddt->ddt_tree) == 0);
833 ASSERT(avl_numnodes(&ddt->ddt_repair_tree) == 0);
834 avl_destroy(&ddt->ddt_tree);
835 avl_destroy(&ddt->ddt_repair_tree);
836 mutex_destroy(&ddt->ddt_lock);
837 kmem_free(ddt, sizeof (*ddt));
841 ddt_create(spa_t *spa)
845 spa->spa_dedup_checksum = ZIO_DEDUPCHECKSUM;
847 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++)
848 spa->spa_ddt[c] = ddt_table_alloc(spa, c);
856 enum ddt_class class;
861 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
862 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1,
863 &spa->spa_ddt_stat_object);
866 return (error == ENOENT ? 0 : error);
868 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
869 ddt_t *ddt = spa->spa_ddt[c];
870 for (type = 0; type < DDT_TYPES; type++) {
871 for (class = 0; class < DDT_CLASSES;
873 error = ddt_object_load(ddt, type, class);
874 if (error != 0 && error != ENOENT)
880 * Seed the cached histograms.
882 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
883 sizeof (ddt->ddt_histogram));
890 ddt_unload(spa_t *spa)
894 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
895 if (spa->spa_ddt[c]) {
896 ddt_table_free(spa->spa_ddt[c]);
897 spa->spa_ddt[c] = NULL;
903 ddt_class_contains(spa_t *spa, enum ddt_class max_class, const blkptr_t *bp)
908 enum ddt_class class;
910 if (!BP_GET_DEDUP(bp))
913 if (max_class == DDT_CLASS_UNIQUE)
916 ddt = spa->spa_ddt[BP_GET_CHECKSUM(bp)];
917 dde = kmem_alloc(sizeof(ddt_entry_t), KM_PUSHPAGE);
919 ddt_key_fill(&(dde->dde_key), bp);
921 for (type = 0; type < DDT_TYPES; type++) {
922 for (class = 0; class <= max_class; class++) {
923 if (ddt_object_lookup(ddt, type, class, dde) == 0) {
924 kmem_free(dde, sizeof(ddt_entry_t));
930 kmem_free(dde, sizeof(ddt_entry_t));
935 ddt_repair_start(ddt_t *ddt, const blkptr_t *bp)
940 enum ddt_class class;
942 ddt_key_fill(&ddk, bp);
944 dde = ddt_alloc(&ddk);
946 for (type = 0; type < DDT_TYPES; type++) {
947 for (class = 0; class < DDT_CLASSES; class++) {
949 * We can only do repair if there are multiple copies
950 * of the block. For anything in the UNIQUE class,
951 * there's definitely only one copy, so don't even try.
953 if (class != DDT_CLASS_UNIQUE &&
954 ddt_object_lookup(ddt, type, class, dde) == 0)
959 bzero(dde->dde_phys, sizeof (dde->dde_phys));
965 ddt_repair_done(ddt_t *ddt, ddt_entry_t *dde)
971 if (dde->dde_repair_data != NULL && spa_writeable(ddt->ddt_spa) &&
972 avl_find(&ddt->ddt_repair_tree, dde, &where) == NULL)
973 avl_insert(&ddt->ddt_repair_tree, dde, where);
981 ddt_repair_entry_done(zio_t *zio)
983 ddt_entry_t *rdde = zio->io_private;
989 ddt_repair_entry(ddt_t *ddt, ddt_entry_t *dde, ddt_entry_t *rdde, zio_t *rio)
991 ddt_phys_t *ddp = dde->dde_phys;
992 ddt_phys_t *rddp = rdde->dde_phys;
993 ddt_key_t *ddk = &dde->dde_key;
994 ddt_key_t *rddk = &rdde->dde_key;
999 zio = zio_null(rio, rio->io_spa, NULL,
1000 ddt_repair_entry_done, rdde, rio->io_flags);
1002 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++, rddp++) {
1003 if (ddp->ddp_phys_birth == 0 ||
1004 ddp->ddp_phys_birth != rddp->ddp_phys_birth ||
1005 bcmp(ddp->ddp_dva, rddp->ddp_dva, sizeof (ddp->ddp_dva)))
1007 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
1008 zio_nowait(zio_rewrite(zio, zio->io_spa, 0, &blk,
1009 rdde->dde_repair_data, DDK_GET_PSIZE(rddk), NULL, NULL,
1010 ZIO_PRIORITY_SYNC_WRITE, ZIO_DDT_CHILD_FLAGS(zio), NULL));
1017 ddt_repair_table(ddt_t *ddt, zio_t *rio)
1019 spa_t *spa = ddt->ddt_spa;
1020 ddt_entry_t *dde, *rdde_next, *rdde;
1021 avl_tree_t *t = &ddt->ddt_repair_tree;
1024 if (spa_sync_pass(spa) > 1)
1028 for (rdde = avl_first(t); rdde != NULL; rdde = rdde_next) {
1029 rdde_next = AVL_NEXT(t, rdde);
1030 avl_remove(&ddt->ddt_repair_tree, rdde);
1032 ddt_bp_create(ddt->ddt_checksum, &rdde->dde_key, NULL, &blk);
1033 dde = ddt_repair_start(ddt, &blk);
1034 ddt_repair_entry(ddt, dde, rdde, rio);
1035 ddt_repair_done(ddt, dde);
1042 ddt_sync_entry(ddt_t *ddt, ddt_entry_t *dde, dmu_tx_t *tx, uint64_t txg)
1044 dsl_pool_t *dp = ddt->ddt_spa->spa_dsl_pool;
1045 ddt_phys_t *ddp = dde->dde_phys;
1046 ddt_key_t *ddk = &dde->dde_key;
1047 enum ddt_type otype = dde->dde_type;
1048 enum ddt_type ntype = DDT_TYPE_CURRENT;
1049 enum ddt_class oclass = dde->dde_class;
1050 enum ddt_class nclass;
1051 uint64_t total_refcnt = 0;
1054 ASSERT(dde->dde_loaded);
1055 ASSERT(!dde->dde_loading);
1057 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1058 ASSERT(dde->dde_lead_zio[p] == NULL);
1059 ASSERT((int64_t)ddp->ddp_refcnt >= 0);
1060 if (ddp->ddp_phys_birth == 0) {
1061 ASSERT(ddp->ddp_refcnt == 0);
1064 if (p == DDT_PHYS_DITTO) {
1065 if (ddt_ditto_copies_needed(ddt, dde, NULL) == 0)
1066 ddt_phys_free(ddt, ddk, ddp, txg);
1069 if (ddp->ddp_refcnt == 0)
1070 ddt_phys_free(ddt, ddk, ddp, txg);
1071 total_refcnt += ddp->ddp_refcnt;
1074 if (dde->dde_phys[DDT_PHYS_DITTO].ddp_phys_birth != 0)
1075 nclass = DDT_CLASS_DITTO;
1076 else if (total_refcnt > 1)
1077 nclass = DDT_CLASS_DUPLICATE;
1079 nclass = DDT_CLASS_UNIQUE;
1081 if (otype != DDT_TYPES &&
1082 (otype != ntype || oclass != nclass || total_refcnt == 0)) {
1083 VERIFY(ddt_object_remove(ddt, otype, oclass, dde, tx) == 0);
1084 ASSERT(ddt_object_lookup(ddt, otype, oclass, dde) == ENOENT);
1087 if (total_refcnt != 0) {
1088 dde->dde_type = ntype;
1089 dde->dde_class = nclass;
1090 ddt_stat_update(ddt, dde, 0);
1091 if (!ddt_object_exists(ddt, ntype, nclass))
1092 ddt_object_create(ddt, ntype, nclass, tx);
1093 VERIFY(ddt_object_update(ddt, ntype, nclass, dde, tx) == 0);
1096 * If the class changes, the order that we scan this bp
1097 * changes. If it decreases, we could miss it, so
1098 * scan it right now. (This covers both class changing
1099 * while we are doing ddt_walk(), and when we are
1102 if (nclass < oclass) {
1103 dsl_scan_ddt_entry(dp->dp_scan,
1104 ddt->ddt_checksum, dde, tx);
1110 ddt_sync_table(ddt_t *ddt, dmu_tx_t *tx, uint64_t txg)
1112 spa_t *spa = ddt->ddt_spa;
1114 void *cookie = NULL;
1116 enum ddt_class class;
1118 if (avl_numnodes(&ddt->ddt_tree) == 0)
1121 ASSERT(spa->spa_uberblock.ub_version >= SPA_VERSION_DEDUP);
1123 if (spa->spa_ddt_stat_object == 0) {
1124 spa->spa_ddt_stat_object = zap_create_link(ddt->ddt_os,
1125 DMU_OT_DDT_STATS, DMU_POOL_DIRECTORY_OBJECT,
1126 DMU_POOL_DDT_STATS, tx);
1129 while ((dde = avl_destroy_nodes(&ddt->ddt_tree, &cookie)) != NULL) {
1130 ddt_sync_entry(ddt, dde, tx, txg);
1134 for (type = 0; type < DDT_TYPES; type++) {
1135 uint64_t add, count = 0;
1136 for (class = 0; class < DDT_CLASSES; class++) {
1137 if (ddt_object_exists(ddt, type, class)) {
1138 ddt_object_sync(ddt, type, class, tx);
1139 VERIFY(ddt_object_count(ddt, type, class,
1144 for (class = 0; class < DDT_CLASSES; class++) {
1145 if (count == 0 && ddt_object_exists(ddt, type, class))
1146 ddt_object_destroy(ddt, type, class, tx);
1150 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
1151 sizeof (ddt->ddt_histogram));
1155 ddt_sync(spa_t *spa, uint64_t txg)
1158 zio_t *rio = zio_root(spa, NULL, NULL,
1159 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
1160 enum zio_checksum c;
1162 ASSERT(spa_syncing_txg(spa) == txg);
1164 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1166 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
1167 ddt_t *ddt = spa->spa_ddt[c];
1170 ddt_sync_table(ddt, tx, txg);
1171 ddt_repair_table(ddt, rio);
1174 (void) zio_wait(rio);
1180 ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde)
1185 ddt_t *ddt = spa->spa_ddt[ddb->ddb_checksum];
1187 if (ddt_object_exists(ddt, ddb->ddb_type,
1189 error = ddt_object_walk(ddt,
1190 ddb->ddb_type, ddb->ddb_class,
1191 &ddb->ddb_cursor, dde);
1193 dde->dde_type = ddb->ddb_type;
1194 dde->dde_class = ddb->ddb_class;
1197 if (error != ENOENT)
1199 ddb->ddb_cursor = 0;
1200 } while (++ddb->ddb_checksum < ZIO_CHECKSUM_FUNCTIONS);
1201 ddb->ddb_checksum = 0;
1202 } while (++ddb->ddb_type < DDT_TYPES);
1204 } while (++ddb->ddb_class < DDT_CLASSES);
1209 #if defined(_KERNEL) && defined(HAVE_SPL)
1210 module_param(zfs_dedup_prefetch, int, 0644);
1211 MODULE_PARM_DESC(zfs_dedup_prefetch,"Enable prefetching dedup-ed blks");