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>
40 static kmem_cache_t *ddt_cache;
41 static kmem_cache_t *ddt_entry_cache;
44 * Enable/disable prefetching of dedup-ed blocks which are going to be freed.
46 int zfs_dedup_prefetch = 1;
48 static const ddt_ops_t *ddt_ops[DDT_TYPES] = {
52 static const char *ddt_class_name[DDT_CLASSES] = {
59 ddt_object_create(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
62 spa_t *spa = ddt->ddt_spa;
63 objset_t *os = ddt->ddt_os;
64 uint64_t *objectp = &ddt->ddt_object[type][class];
65 boolean_t prehash = zio_checksum_table[ddt->ddt_checksum].ci_dedup;
66 char name[DDT_NAMELEN];
68 ddt_object_name(ddt, type, class, name);
70 ASSERT(*objectp == 0);
71 VERIFY(ddt_ops[type]->ddt_op_create(os, objectp, tx, prehash) == 0);
72 ASSERT(*objectp != 0);
74 VERIFY(zap_add(os, DMU_POOL_DIRECTORY_OBJECT, name,
75 sizeof (uint64_t), 1, objectp, tx) == 0);
77 VERIFY(zap_add(os, spa->spa_ddt_stat_object, name,
78 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
79 &ddt->ddt_histogram[type][class], tx) == 0);
83 ddt_object_destroy(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
86 spa_t *spa = ddt->ddt_spa;
87 objset_t *os = ddt->ddt_os;
88 uint64_t *objectp = &ddt->ddt_object[type][class];
90 char name[DDT_NAMELEN];
92 ddt_object_name(ddt, type, class, name);
94 ASSERT(*objectp != 0);
95 ASSERT(ddt_histogram_empty(&ddt->ddt_histogram[type][class]));
96 VERIFY(ddt_object_count(ddt, type, class, &count) == 0 && count == 0);
97 VERIFY(zap_remove(os, DMU_POOL_DIRECTORY_OBJECT, name, tx) == 0);
98 VERIFY(zap_remove(os, spa->spa_ddt_stat_object, name, tx) == 0);
99 VERIFY(ddt_ops[type]->ddt_op_destroy(os, *objectp, tx) == 0);
100 bzero(&ddt->ddt_object_stats[type][class], sizeof (ddt_object_t));
106 ddt_object_load(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
108 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
109 dmu_object_info_t doi;
111 char name[DDT_NAMELEN];
114 ddt_object_name(ddt, type, class, name);
116 error = zap_lookup(ddt->ddt_os, DMU_POOL_DIRECTORY_OBJECT, name,
117 sizeof (uint64_t), 1, &ddt->ddt_object[type][class]);
122 error = zap_lookup(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
123 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
124 &ddt->ddt_histogram[type][class]);
127 * Seed the cached statistics.
129 error = ddt_object_info(ddt, type, class, &doi);
133 error = ddt_object_count(ddt, type, class, &count);
137 ddo->ddo_count = count;
138 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
139 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
146 ddt_object_sync(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
149 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
150 dmu_object_info_t doi;
152 char name[DDT_NAMELEN];
154 ddt_object_name(ddt, type, class, name);
156 VERIFY(zap_update(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
157 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
158 &ddt->ddt_histogram[type][class], tx) == 0);
161 * Cache DDT statistics; this is the only time they'll change.
163 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0);
164 VERIFY(ddt_object_count(ddt, type, class, &count) == 0);
166 ddo->ddo_count = count;
167 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
168 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
172 ddt_object_lookup(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
175 if (!ddt_object_exists(ddt, type, class))
176 return (SET_ERROR(ENOENT));
178 return (ddt_ops[type]->ddt_op_lookup(ddt->ddt_os,
179 ddt->ddt_object[type][class], dde));
183 ddt_object_prefetch(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
186 if (!ddt_object_exists(ddt, type, class))
189 ddt_ops[type]->ddt_op_prefetch(ddt->ddt_os,
190 ddt->ddt_object[type][class], dde);
194 ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
195 ddt_entry_t *dde, dmu_tx_t *tx)
197 ASSERT(ddt_object_exists(ddt, type, class));
199 return (ddt_ops[type]->ddt_op_update(ddt->ddt_os,
200 ddt->ddt_object[type][class], dde, tx));
204 ddt_object_remove(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
205 ddt_entry_t *dde, dmu_tx_t *tx)
207 ASSERT(ddt_object_exists(ddt, type, class));
209 return (ddt_ops[type]->ddt_op_remove(ddt->ddt_os,
210 ddt->ddt_object[type][class], dde, tx));
214 ddt_object_walk(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
215 uint64_t *walk, ddt_entry_t *dde)
217 ASSERT(ddt_object_exists(ddt, type, class));
219 return (ddt_ops[type]->ddt_op_walk(ddt->ddt_os,
220 ddt->ddt_object[type][class], dde, walk));
224 ddt_object_count(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
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)
265 for (d = 0; d < SPA_DVAS_PER_BP; d++)
266 bp->blk_dva[d] = ddp->ddp_dva[d];
267 BP_SET_BIRTH(bp, txg, ddp->ddp_phys_birth);
271 ddt_bp_create(enum zio_checksum checksum,
272 const ddt_key_t *ddk, const ddt_phys_t *ddp, blkptr_t *bp)
277 ddt_bp_fill(ddp, bp, ddp->ddp_phys_birth);
279 bp->blk_cksum = ddk->ddk_cksum;
282 BP_SET_LSIZE(bp, DDK_GET_LSIZE(ddk));
283 BP_SET_PSIZE(bp, DDK_GET_PSIZE(ddk));
284 BP_SET_COMPRESS(bp, DDK_GET_COMPRESS(ddk));
285 BP_SET_CHECKSUM(bp, checksum);
286 BP_SET_TYPE(bp, DMU_OT_DEDUP);
289 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
293 ddt_key_fill(ddt_key_t *ddk, const blkptr_t *bp)
295 ddk->ddk_cksum = bp->blk_cksum;
298 DDK_SET_LSIZE(ddk, BP_GET_LSIZE(bp));
299 DDK_SET_PSIZE(ddk, BP_GET_PSIZE(bp));
300 DDK_SET_COMPRESS(ddk, BP_GET_COMPRESS(bp));
304 ddt_phys_fill(ddt_phys_t *ddp, const blkptr_t *bp)
307 ASSERT(ddp->ddp_phys_birth == 0);
309 for (d = 0; d < SPA_DVAS_PER_BP; d++)
310 ddp->ddp_dva[d] = bp->blk_dva[d];
311 ddp->ddp_phys_birth = BP_PHYSICAL_BIRTH(bp);
315 ddt_phys_clear(ddt_phys_t *ddp)
317 bzero(ddp, sizeof (*ddp));
321 ddt_phys_addref(ddt_phys_t *ddp)
327 ddt_phys_decref(ddt_phys_t *ddp)
330 ASSERT(ddp->ddp_refcnt > 0);
336 ddt_phys_free(ddt_t *ddt, ddt_key_t *ddk, ddt_phys_t *ddp, uint64_t txg)
340 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
342 zio_free(ddt->ddt_spa, txg, &blk);
346 ddt_phys_select(const ddt_entry_t *dde, const blkptr_t *bp)
348 ddt_phys_t *ddp = (ddt_phys_t *)dde->dde_phys;
351 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
352 if (DVA_EQUAL(BP_IDENTITY(bp), &ddp->ddp_dva[0]) &&
353 BP_PHYSICAL_BIRTH(bp) == ddp->ddp_phys_birth)
360 ddt_phys_total_refcnt(const ddt_entry_t *dde)
365 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++)
366 refcnt += dde->dde_phys[p].ddp_refcnt;
372 ddt_stat_generate(ddt_t *ddt, ddt_entry_t *dde, ddt_stat_t *dds)
374 spa_t *spa = ddt->ddt_spa;
375 ddt_phys_t *ddp = dde->dde_phys;
376 ddt_key_t *ddk = &dde->dde_key;
377 uint64_t lsize = DDK_GET_LSIZE(ddk);
378 uint64_t psize = DDK_GET_PSIZE(ddk);
381 bzero(dds, sizeof (*dds));
383 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
385 uint64_t refcnt = ddp->ddp_refcnt;
387 if (ddp->ddp_phys_birth == 0)
390 for (d = 0; d < SPA_DVAS_PER_BP; d++)
391 dsize += dva_get_dsize_sync(spa, &ddp->ddp_dva[d]);
393 dds->dds_blocks += 1;
394 dds->dds_lsize += lsize;
395 dds->dds_psize += psize;
396 dds->dds_dsize += dsize;
398 dds->dds_ref_blocks += refcnt;
399 dds->dds_ref_lsize += lsize * refcnt;
400 dds->dds_ref_psize += psize * refcnt;
401 dds->dds_ref_dsize += dsize * refcnt;
406 ddt_stat_add(ddt_stat_t *dst, const ddt_stat_t *src, uint64_t neg)
408 const uint64_t *s = (const uint64_t *)src;
409 uint64_t *d = (uint64_t *)dst;
410 uint64_t *d_end = (uint64_t *)(dst + 1);
412 ASSERT(neg == 0 || neg == -1ULL); /* add or subtract */
415 *d++ += (*s++ ^ neg) - neg;
419 ddt_stat_update(ddt_t *ddt, ddt_entry_t *dde, uint64_t neg)
422 ddt_histogram_t *ddh;
425 ddt_stat_generate(ddt, dde, &dds);
427 bucket = highbit(dds.dds_ref_blocks) - 1;
430 ddh = &ddt->ddt_histogram[dde->dde_type][dde->dde_class];
432 ddt_stat_add(&ddh->ddh_stat[bucket], &dds, neg);
436 ddt_histogram_add(ddt_histogram_t *dst, const ddt_histogram_t *src)
440 for (h = 0; h < 64; h++)
441 ddt_stat_add(&dst->ddh_stat[h], &src->ddh_stat[h], 0);
445 ddt_histogram_stat(ddt_stat_t *dds, const ddt_histogram_t *ddh)
449 bzero(dds, sizeof (*dds));
451 for (h = 0; h < 64; h++)
452 ddt_stat_add(dds, &ddh->ddh_stat[h], 0);
456 ddt_histogram_empty(const ddt_histogram_t *ddh)
458 const uint64_t *s = (const uint64_t *)ddh;
459 const uint64_t *s_end = (const uint64_t *)(ddh + 1);
469 ddt_get_dedup_object_stats(spa_t *spa, ddt_object_t *ddo_total)
473 enum ddt_class class;
475 /* Sum the statistics we cached in ddt_object_sync(). */
476 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
477 ddt_t *ddt = spa->spa_ddt[c];
478 for (type = 0; type < DDT_TYPES; type++) {
479 for (class = 0; class < DDT_CLASSES;
482 &ddt->ddt_object_stats[type][class];
483 ddo_total->ddo_count += ddo->ddo_count;
484 ddo_total->ddo_dspace += ddo->ddo_dspace;
485 ddo_total->ddo_mspace += ddo->ddo_mspace;
490 /* ... and compute the averages. */
491 if (ddo_total->ddo_count != 0) {
492 ddo_total->ddo_dspace /= ddo_total->ddo_count;
493 ddo_total->ddo_mspace /= ddo_total->ddo_count;
498 ddt_get_dedup_histogram(spa_t *spa, ddt_histogram_t *ddh)
502 enum ddt_class class;
504 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
505 ddt_t *ddt = spa->spa_ddt[c];
506 for (type = 0; type < DDT_TYPES; type++) {
507 for (class = 0; class < DDT_CLASSES;
509 ddt_histogram_add(ddh,
510 &ddt->ddt_histogram_cache[type][class]);
517 ddt_get_dedup_stats(spa_t *spa, ddt_stat_t *dds_total)
519 ddt_histogram_t *ddh_total;
521 ddh_total = kmem_zalloc(sizeof (ddt_histogram_t), KM_PUSHPAGE);
522 ddt_get_dedup_histogram(spa, ddh_total);
523 ddt_histogram_stat(dds_total, ddh_total);
524 kmem_free(ddh_total, sizeof (ddt_histogram_t));
528 ddt_get_dedup_dspace(spa_t *spa)
530 ddt_stat_t dds_total = { 0 };
532 ddt_get_dedup_stats(spa, &dds_total);
533 return (dds_total.dds_ref_dsize - dds_total.dds_dsize);
537 ddt_get_pool_dedup_ratio(spa_t *spa)
539 ddt_stat_t dds_total = { 0 };
541 ddt_get_dedup_stats(spa, &dds_total);
542 if (dds_total.dds_dsize == 0)
545 return (dds_total.dds_ref_dsize * 100 / dds_total.dds_dsize);
549 ddt_ditto_copies_needed(ddt_t *ddt, ddt_entry_t *dde, ddt_phys_t *ddp_willref)
551 spa_t *spa = ddt->ddt_spa;
552 uint64_t total_refcnt = 0;
553 uint64_t ditto = spa->spa_dedup_ditto;
554 int total_copies = 0;
555 int desired_copies = 0;
558 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
559 ddt_phys_t *ddp = &dde->dde_phys[p];
560 zio_t *zio = dde->dde_lead_zio[p];
561 uint64_t refcnt = ddp->ddp_refcnt; /* committed refs */
563 refcnt += zio->io_parent_count; /* pending refs */
564 if (ddp == ddp_willref)
565 refcnt++; /* caller's ref */
567 total_refcnt += refcnt;
572 if (ditto == 0 || ditto > UINT32_MAX)
575 if (total_refcnt >= 1)
577 if (total_refcnt >= ditto)
579 if (total_refcnt >= ditto * ditto)
582 return (MAX(desired_copies, total_copies) - total_copies);
586 ddt_ditto_copies_present(ddt_entry_t *dde)
588 ddt_phys_t *ddp = &dde->dde_phys[DDT_PHYS_DITTO];
589 dva_t *dva = ddp->ddp_dva;
590 int copies = 0 - DVA_GET_GANG(dva);
593 for (d = 0; d < SPA_DVAS_PER_BP; d++, dva++)
594 if (DVA_IS_VALID(dva))
597 ASSERT(copies >= 0 && copies < SPA_DVAS_PER_BP);
603 ddt_compress(void *src, uchar_t *dst, size_t s_len, size_t d_len)
605 uchar_t *version = dst++;
606 int cpfunc = ZIO_COMPRESS_ZLE;
607 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
610 ASSERT(d_len >= s_len + 1); /* no compression plus version byte */
612 c_len = ci->ci_compress(src, dst, s_len, d_len - 1, ci->ci_level);
614 if (c_len == s_len) {
615 cpfunc = ZIO_COMPRESS_OFF;
616 bcopy(src, dst, s_len);
619 *version = (ZFS_HOST_BYTEORDER & DDT_COMPRESS_BYTEORDER_MASK) | cpfunc;
625 ddt_decompress(uchar_t *src, void *dst, size_t s_len, size_t d_len)
627 uchar_t version = *src++;
628 int cpfunc = version & DDT_COMPRESS_FUNCTION_MASK;
629 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
631 if (ci->ci_decompress != NULL)
632 (void) ci->ci_decompress(src, dst, s_len, d_len, ci->ci_level);
634 bcopy(src, dst, d_len);
636 if ((version ^ ZFS_HOST_BYTEORDER) & DDT_COMPRESS_BYTEORDER_MASK)
637 byteswap_uint64_array(dst, d_len);
641 ddt_select_by_checksum(spa_t *spa, enum zio_checksum c)
643 return (spa->spa_ddt[c]);
647 ddt_select(spa_t *spa, const blkptr_t *bp)
649 return (spa->spa_ddt[BP_GET_CHECKSUM(bp)]);
653 ddt_enter(ddt_t *ddt)
655 mutex_enter(&ddt->ddt_lock);
661 mutex_exit(&ddt->ddt_lock);
667 ddt_cache = kmem_cache_create("ddt_cache",
668 sizeof (ddt_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
669 ddt_entry_cache = kmem_cache_create("ddt_entry_cache",
670 sizeof (ddt_entry_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
676 kmem_cache_destroy(ddt_entry_cache);
677 kmem_cache_destroy(ddt_cache);
681 ddt_alloc(const ddt_key_t *ddk)
685 dde = kmem_cache_alloc(ddt_entry_cache, KM_PUSHPAGE);
686 bzero(dde, sizeof (ddt_entry_t));
687 cv_init(&dde->dde_cv, NULL, CV_DEFAULT, NULL);
695 ddt_free(ddt_entry_t *dde)
699 ASSERT(!dde->dde_loading);
701 for (p = 0; p < DDT_PHYS_TYPES; p++)
702 ASSERT(dde->dde_lead_zio[p] == NULL);
704 if (dde->dde_repair_data != NULL)
705 zio_buf_free(dde->dde_repair_data,
706 DDK_GET_PSIZE(&dde->dde_key));
708 cv_destroy(&dde->dde_cv);
709 kmem_cache_free(ddt_entry_cache, dde);
713 ddt_remove(ddt_t *ddt, ddt_entry_t *dde)
715 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
717 avl_remove(&ddt->ddt_tree, dde);
722 ddt_lookup(ddt_t *ddt, const blkptr_t *bp, boolean_t add)
724 ddt_entry_t *dde, dde_search;
726 enum ddt_class class;
730 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
732 ddt_key_fill(&dde_search.dde_key, bp);
734 dde = avl_find(&ddt->ddt_tree, &dde_search, &where);
738 dde = ddt_alloc(&dde_search.dde_key);
739 avl_insert(&ddt->ddt_tree, dde, where);
742 while (dde->dde_loading)
743 cv_wait(&dde->dde_cv, &ddt->ddt_lock);
748 dde->dde_loading = B_TRUE;
754 for (type = 0; type < DDT_TYPES; type++) {
755 for (class = 0; class < DDT_CLASSES; class++) {
756 error = ddt_object_lookup(ddt, type, class, dde);
764 ASSERT(error == 0 || error == ENOENT);
768 ASSERT(dde->dde_loaded == B_FALSE);
769 ASSERT(dde->dde_loading == B_TRUE);
771 dde->dde_type = type; /* will be DDT_TYPES if no entry found */
772 dde->dde_class = class; /* will be DDT_CLASSES if no entry found */
773 dde->dde_loaded = B_TRUE;
774 dde->dde_loading = B_FALSE;
777 ddt_stat_update(ddt, dde, -1ULL);
779 cv_broadcast(&dde->dde_cv);
785 ddt_prefetch(spa_t *spa, const blkptr_t *bp)
790 enum ddt_class class;
792 if (!zfs_dedup_prefetch || bp == NULL || !BP_GET_DEDUP(bp))
796 * We only remove the DDT once all tables are empty and only
797 * prefetch dedup blocks when there are entries in the DDT.
798 * Thus no locking is required as the DDT can't disappear on us.
800 ddt = ddt_select(spa, bp);
801 ddt_key_fill(&dde.dde_key, bp);
803 for (type = 0; type < DDT_TYPES; type++) {
804 for (class = 0; class < DDT_CLASSES; class++) {
805 ddt_object_prefetch(ddt, type, class, &dde);
811 ddt_entry_compare(const void *x1, const void *x2)
813 const ddt_entry_t *dde1 = x1;
814 const ddt_entry_t *dde2 = x2;
815 const uint64_t *u1 = (const uint64_t *)&dde1->dde_key;
816 const uint64_t *u2 = (const uint64_t *)&dde2->dde_key;
819 for (i = 0; i < DDT_KEY_WORDS; i++) {
830 ddt_table_alloc(spa_t *spa, enum zio_checksum c)
834 ddt = kmem_cache_alloc(ddt_cache, KM_PUSHPAGE | KM_NODEBUG);
835 bzero(ddt, sizeof (ddt_t));
837 mutex_init(&ddt->ddt_lock, NULL, MUTEX_DEFAULT, NULL);
838 avl_create(&ddt->ddt_tree, ddt_entry_compare,
839 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
840 avl_create(&ddt->ddt_repair_tree, ddt_entry_compare,
841 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
842 ddt->ddt_checksum = c;
844 ddt->ddt_os = spa->spa_meta_objset;
850 ddt_table_free(ddt_t *ddt)
852 ASSERT(avl_numnodes(&ddt->ddt_tree) == 0);
853 ASSERT(avl_numnodes(&ddt->ddt_repair_tree) == 0);
854 avl_destroy(&ddt->ddt_tree);
855 avl_destroy(&ddt->ddt_repair_tree);
856 mutex_destroy(&ddt->ddt_lock);
857 kmem_cache_free(ddt_cache, ddt);
861 ddt_create(spa_t *spa)
865 spa->spa_dedup_checksum = ZIO_DEDUPCHECKSUM;
867 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++)
868 spa->spa_ddt[c] = ddt_table_alloc(spa, c);
876 enum ddt_class class;
881 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
882 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1,
883 &spa->spa_ddt_stat_object);
886 return (error == ENOENT ? 0 : error);
888 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
889 ddt_t *ddt = spa->spa_ddt[c];
890 for (type = 0; type < DDT_TYPES; type++) {
891 for (class = 0; class < DDT_CLASSES;
893 error = ddt_object_load(ddt, type, class);
894 if (error != 0 && error != ENOENT)
900 * Seed the cached histograms.
902 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
903 sizeof (ddt->ddt_histogram));
910 ddt_unload(spa_t *spa)
914 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
915 if (spa->spa_ddt[c]) {
916 ddt_table_free(spa->spa_ddt[c]);
917 spa->spa_ddt[c] = NULL;
923 ddt_class_contains(spa_t *spa, enum ddt_class max_class, const blkptr_t *bp)
928 enum ddt_class class;
930 if (!BP_GET_DEDUP(bp))
933 if (max_class == DDT_CLASS_UNIQUE)
936 ddt = spa->spa_ddt[BP_GET_CHECKSUM(bp)];
937 dde = kmem_cache_alloc(ddt_entry_cache, KM_PUSHPAGE);
939 ddt_key_fill(&(dde->dde_key), bp);
941 for (type = 0; type < DDT_TYPES; type++) {
942 for (class = 0; class <= max_class; class++) {
943 if (ddt_object_lookup(ddt, type, class, dde) == 0) {
944 kmem_cache_free(ddt_entry_cache, dde);
950 kmem_cache_free(ddt_entry_cache, dde);
955 ddt_repair_start(ddt_t *ddt, const blkptr_t *bp)
960 enum ddt_class class;
962 ddt_key_fill(&ddk, bp);
964 dde = ddt_alloc(&ddk);
966 for (type = 0; type < DDT_TYPES; type++) {
967 for (class = 0; class < DDT_CLASSES; class++) {
969 * We can only do repair if there are multiple copies
970 * of the block. For anything in the UNIQUE class,
971 * there's definitely only one copy, so don't even try.
973 if (class != DDT_CLASS_UNIQUE &&
974 ddt_object_lookup(ddt, type, class, dde) == 0)
979 bzero(dde->dde_phys, sizeof (dde->dde_phys));
985 ddt_repair_done(ddt_t *ddt, ddt_entry_t *dde)
991 if (dde->dde_repair_data != NULL && spa_writeable(ddt->ddt_spa) &&
992 avl_find(&ddt->ddt_repair_tree, dde, &where) == NULL)
993 avl_insert(&ddt->ddt_repair_tree, dde, where);
1001 ddt_repair_entry_done(zio_t *zio)
1003 ddt_entry_t *rdde = zio->io_private;
1009 ddt_repair_entry(ddt_t *ddt, ddt_entry_t *dde, ddt_entry_t *rdde, zio_t *rio)
1011 ddt_phys_t *ddp = dde->dde_phys;
1012 ddt_phys_t *rddp = rdde->dde_phys;
1013 ddt_key_t *ddk = &dde->dde_key;
1014 ddt_key_t *rddk = &rdde->dde_key;
1019 zio = zio_null(rio, rio->io_spa, NULL,
1020 ddt_repair_entry_done, rdde, rio->io_flags);
1022 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++, rddp++) {
1023 if (ddp->ddp_phys_birth == 0 ||
1024 ddp->ddp_phys_birth != rddp->ddp_phys_birth ||
1025 bcmp(ddp->ddp_dva, rddp->ddp_dva, sizeof (ddp->ddp_dva)))
1027 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
1028 zio_nowait(zio_rewrite(zio, zio->io_spa, 0, &blk,
1029 rdde->dde_repair_data, DDK_GET_PSIZE(rddk), NULL, NULL,
1030 ZIO_PRIORITY_SYNC_WRITE, ZIO_DDT_CHILD_FLAGS(zio), NULL));
1037 ddt_repair_table(ddt_t *ddt, zio_t *rio)
1039 spa_t *spa = ddt->ddt_spa;
1040 ddt_entry_t *dde, *rdde_next, *rdde;
1041 avl_tree_t *t = &ddt->ddt_repair_tree;
1044 if (spa_sync_pass(spa) > 1)
1048 for (rdde = avl_first(t); rdde != NULL; rdde = rdde_next) {
1049 rdde_next = AVL_NEXT(t, rdde);
1050 avl_remove(&ddt->ddt_repair_tree, rdde);
1052 ddt_bp_create(ddt->ddt_checksum, &rdde->dde_key, NULL, &blk);
1053 dde = ddt_repair_start(ddt, &blk);
1054 ddt_repair_entry(ddt, dde, rdde, rio);
1055 ddt_repair_done(ddt, dde);
1062 ddt_sync_entry(ddt_t *ddt, ddt_entry_t *dde, dmu_tx_t *tx, uint64_t txg)
1064 dsl_pool_t *dp = ddt->ddt_spa->spa_dsl_pool;
1065 ddt_phys_t *ddp = dde->dde_phys;
1066 ddt_key_t *ddk = &dde->dde_key;
1067 enum ddt_type otype = dde->dde_type;
1068 enum ddt_type ntype = DDT_TYPE_CURRENT;
1069 enum ddt_class oclass = dde->dde_class;
1070 enum ddt_class nclass;
1071 uint64_t total_refcnt = 0;
1074 ASSERT(dde->dde_loaded);
1075 ASSERT(!dde->dde_loading);
1077 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1078 ASSERT(dde->dde_lead_zio[p] == NULL);
1079 if (ddp->ddp_phys_birth == 0) {
1080 ASSERT(ddp->ddp_refcnt == 0);
1083 if (p == DDT_PHYS_DITTO) {
1084 if (ddt_ditto_copies_needed(ddt, dde, NULL) == 0)
1085 ddt_phys_free(ddt, ddk, ddp, txg);
1088 if (ddp->ddp_refcnt == 0)
1089 ddt_phys_free(ddt, ddk, ddp, txg);
1090 total_refcnt += ddp->ddp_refcnt;
1093 if (dde->dde_phys[DDT_PHYS_DITTO].ddp_phys_birth != 0)
1094 nclass = DDT_CLASS_DITTO;
1095 else if (total_refcnt > 1)
1096 nclass = DDT_CLASS_DUPLICATE;
1098 nclass = DDT_CLASS_UNIQUE;
1100 if (otype != DDT_TYPES &&
1101 (otype != ntype || oclass != nclass || total_refcnt == 0)) {
1102 VERIFY(ddt_object_remove(ddt, otype, oclass, dde, tx) == 0);
1103 ASSERT(ddt_object_lookup(ddt, otype, oclass, dde) == ENOENT);
1106 if (total_refcnt != 0) {
1107 dde->dde_type = ntype;
1108 dde->dde_class = nclass;
1109 ddt_stat_update(ddt, dde, 0);
1110 if (!ddt_object_exists(ddt, ntype, nclass))
1111 ddt_object_create(ddt, ntype, nclass, tx);
1112 VERIFY(ddt_object_update(ddt, ntype, nclass, dde, tx) == 0);
1115 * If the class changes, the order that we scan this bp
1116 * changes. If it decreases, we could miss it, so
1117 * scan it right now. (This covers both class changing
1118 * while we are doing ddt_walk(), and when we are
1121 if (nclass < oclass) {
1122 dsl_scan_ddt_entry(dp->dp_scan,
1123 ddt->ddt_checksum, dde, tx);
1129 ddt_sync_table(ddt_t *ddt, dmu_tx_t *tx, uint64_t txg)
1131 spa_t *spa = ddt->ddt_spa;
1133 void *cookie = NULL;
1135 enum ddt_class class;
1137 if (avl_numnodes(&ddt->ddt_tree) == 0)
1140 ASSERT(spa->spa_uberblock.ub_version >= SPA_VERSION_DEDUP);
1142 if (spa->spa_ddt_stat_object == 0) {
1143 spa->spa_ddt_stat_object = zap_create_link(ddt->ddt_os,
1144 DMU_OT_DDT_STATS, DMU_POOL_DIRECTORY_OBJECT,
1145 DMU_POOL_DDT_STATS, tx);
1148 while ((dde = avl_destroy_nodes(&ddt->ddt_tree, &cookie)) != NULL) {
1149 ddt_sync_entry(ddt, dde, tx, txg);
1153 for (type = 0; type < DDT_TYPES; type++) {
1154 uint64_t add, count = 0;
1155 for (class = 0; class < DDT_CLASSES; class++) {
1156 if (ddt_object_exists(ddt, type, class)) {
1157 ddt_object_sync(ddt, type, class, tx);
1158 VERIFY(ddt_object_count(ddt, type, class,
1163 for (class = 0; class < DDT_CLASSES; class++) {
1164 if (count == 0 && ddt_object_exists(ddt, type, class))
1165 ddt_object_destroy(ddt, type, class, tx);
1169 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
1170 sizeof (ddt->ddt_histogram));
1174 ddt_sync(spa_t *spa, uint64_t txg)
1177 zio_t *rio = zio_root(spa, NULL, NULL,
1178 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
1179 enum zio_checksum c;
1181 ASSERT(spa_syncing_txg(spa) == txg);
1183 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1185 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
1186 ddt_t *ddt = spa->spa_ddt[c];
1189 ddt_sync_table(ddt, tx, txg);
1190 ddt_repair_table(ddt, rio);
1193 (void) zio_wait(rio);
1199 ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde)
1204 ddt_t *ddt = spa->spa_ddt[ddb->ddb_checksum];
1206 if (ddt_object_exists(ddt, ddb->ddb_type,
1208 error = ddt_object_walk(ddt,
1209 ddb->ddb_type, ddb->ddb_class,
1210 &ddb->ddb_cursor, dde);
1212 dde->dde_type = ddb->ddb_type;
1213 dde->dde_class = ddb->ddb_class;
1216 if (error != ENOENT)
1218 ddb->ddb_cursor = 0;
1219 } while (++ddb->ddb_checksum < ZIO_CHECKSUM_FUNCTIONS);
1220 ddb->ddb_checksum = 0;
1221 } while (++ddb->ddb_type < DDT_TYPES);
1223 } while (++ddb->ddb_class < DDT_CLASSES);
1225 return (SET_ERROR(ENOENT));
1228 #if defined(_KERNEL) && defined(HAVE_SPL)
1229 module_param(zfs_dedup_prefetch, int, 0644);
1230 MODULE_PARM_DESC(zfs_dedup_prefetch, "Enable prefetching dedup-ed blks");