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.
26 #include <sys/zfs_context.h>
28 #include <sys/spa_impl.h>
32 #include <sys/dmu_tx.h>
34 #include <sys/dsl_pool.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/zio_compress.h>
37 #include <sys/dsl_scan.h>
40 * Enable/disable prefetching of dedup-ed blocks which are going to be freed.
42 int zfs_dedup_prefetch = 1;
44 static const ddt_ops_t *ddt_ops[DDT_TYPES] = {
48 static const char *ddt_class_name[DDT_CLASSES] = {
55 ddt_object_create(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
58 spa_t *spa = ddt->ddt_spa;
59 objset_t *os = ddt->ddt_os;
60 uint64_t *objectp = &ddt->ddt_object[type][class];
61 boolean_t prehash = zio_checksum_table[ddt->ddt_checksum].ci_dedup;
62 char name[DDT_NAMELEN];
64 ddt_object_name(ddt, type, class, name);
66 ASSERT(*objectp == 0);
67 VERIFY(ddt_ops[type]->ddt_op_create(os, objectp, tx, prehash) == 0);
68 ASSERT(*objectp != 0);
70 VERIFY(zap_add(os, DMU_POOL_DIRECTORY_OBJECT, name,
71 sizeof (uint64_t), 1, objectp, tx) == 0);
73 VERIFY(zap_add(os, spa->spa_ddt_stat_object, name,
74 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
75 &ddt->ddt_histogram[type][class], tx) == 0);
79 ddt_object_destroy(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
82 spa_t *spa = ddt->ddt_spa;
83 objset_t *os = ddt->ddt_os;
84 uint64_t *objectp = &ddt->ddt_object[type][class];
85 char name[DDT_NAMELEN];
87 ddt_object_name(ddt, type, class, name);
89 ASSERT(*objectp != 0);
90 ASSERT(ddt_object_count(ddt, type, class) == 0);
91 ASSERT(ddt_histogram_empty(&ddt->ddt_histogram[type][class]));
92 VERIFY(zap_remove(os, DMU_POOL_DIRECTORY_OBJECT, name, tx) == 0);
93 VERIFY(zap_remove(os, spa->spa_ddt_stat_object, name, tx) == 0);
94 VERIFY(ddt_ops[type]->ddt_op_destroy(os, *objectp, tx) == 0);
95 bzero(&ddt->ddt_object_stats[type][class], sizeof (ddt_object_t));
101 ddt_object_load(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
103 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
104 dmu_object_info_t doi;
105 char name[DDT_NAMELEN];
108 ddt_object_name(ddt, type, class, name);
110 error = zap_lookup(ddt->ddt_os, DMU_POOL_DIRECTORY_OBJECT, name,
111 sizeof (uint64_t), 1, &ddt->ddt_object[type][class]);
116 error = zap_lookup(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
117 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
118 &ddt->ddt_histogram[type][class]);
121 * Seed the cached statistics.
123 error = ddt_object_info(ddt, type, class, &doi);
127 ddo->ddo_count = ddt_object_count(ddt, type, class);
128 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
129 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
136 ddt_object_sync(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
139 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
140 dmu_object_info_t doi;
141 char name[DDT_NAMELEN];
143 ddt_object_name(ddt, type, class, name);
145 VERIFY(zap_update(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
146 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
147 &ddt->ddt_histogram[type][class], tx) == 0);
150 * Cache DDT statistics; this is the only time they'll change.
152 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0);
154 ddo->ddo_count = ddt_object_count(ddt, type, class);
155 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
156 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
160 ddt_object_lookup(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
163 if (!ddt_object_exists(ddt, type, class))
166 return (ddt_ops[type]->ddt_op_lookup(ddt->ddt_os,
167 ddt->ddt_object[type][class], dde));
171 ddt_object_prefetch(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
174 if (!ddt_object_exists(ddt, type, class))
177 ddt_ops[type]->ddt_op_prefetch(ddt->ddt_os,
178 ddt->ddt_object[type][class], dde);
182 ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
183 ddt_entry_t *dde, dmu_tx_t *tx)
185 ASSERT(ddt_object_exists(ddt, type, class));
187 return (ddt_ops[type]->ddt_op_update(ddt->ddt_os,
188 ddt->ddt_object[type][class], dde, tx));
192 ddt_object_remove(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
193 ddt_entry_t *dde, dmu_tx_t *tx)
195 ASSERT(ddt_object_exists(ddt, type, class));
197 return (ddt_ops[type]->ddt_op_remove(ddt->ddt_os,
198 ddt->ddt_object[type][class], dde, tx));
202 ddt_object_walk(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
203 uint64_t *walk, ddt_entry_t *dde)
205 ASSERT(ddt_object_exists(ddt, type, class));
207 return (ddt_ops[type]->ddt_op_walk(ddt->ddt_os,
208 ddt->ddt_object[type][class], dde, walk));
212 ddt_object_count(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
214 ASSERT(ddt_object_exists(ddt, type, class));
216 return (ddt_ops[type]->ddt_op_count(ddt->ddt_os,
217 ddt->ddt_object[type][class]));
221 ddt_object_info(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
222 dmu_object_info_t *doi)
224 if (!ddt_object_exists(ddt, type, class))
227 return (dmu_object_info(ddt->ddt_os, ddt->ddt_object[type][class],
232 ddt_object_exists(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
234 return (!!ddt->ddt_object[type][class]);
238 ddt_object_name(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
241 (void) sprintf(name, DMU_POOL_DDT,
242 zio_checksum_table[ddt->ddt_checksum].ci_name,
243 ddt_ops[type]->ddt_op_name, ddt_class_name[class]);
247 ddt_bp_fill(const ddt_phys_t *ddp, blkptr_t *bp, uint64_t txg)
252 for (d = 0; d < SPA_DVAS_PER_BP; d++)
253 bp->blk_dva[d] = ddp->ddp_dva[d];
254 BP_SET_BIRTH(bp, txg, ddp->ddp_phys_birth);
258 ddt_bp_create(enum zio_checksum checksum,
259 const ddt_key_t *ddk, const ddt_phys_t *ddp, blkptr_t *bp)
264 ddt_bp_fill(ddp, bp, ddp->ddp_phys_birth);
266 bp->blk_cksum = ddk->ddk_cksum;
269 BP_SET_LSIZE(bp, DDK_GET_LSIZE(ddk));
270 BP_SET_PSIZE(bp, DDK_GET_PSIZE(ddk));
271 BP_SET_COMPRESS(bp, DDK_GET_COMPRESS(ddk));
272 BP_SET_CHECKSUM(bp, checksum);
273 BP_SET_TYPE(bp, DMU_OT_DEDUP);
276 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
280 ddt_key_fill(ddt_key_t *ddk, const blkptr_t *bp)
282 ddk->ddk_cksum = bp->blk_cksum;
285 DDK_SET_LSIZE(ddk, BP_GET_LSIZE(bp));
286 DDK_SET_PSIZE(ddk, BP_GET_PSIZE(bp));
287 DDK_SET_COMPRESS(ddk, BP_GET_COMPRESS(bp));
291 ddt_phys_fill(ddt_phys_t *ddp, const blkptr_t *bp)
294 ASSERT(ddp->ddp_phys_birth == 0);
296 for (d = 0; d < SPA_DVAS_PER_BP; d++)
297 ddp->ddp_dva[d] = bp->blk_dva[d];
298 ddp->ddp_phys_birth = BP_PHYSICAL_BIRTH(bp);
302 ddt_phys_clear(ddt_phys_t *ddp)
304 bzero(ddp, sizeof (*ddp));
308 ddt_phys_addref(ddt_phys_t *ddp)
314 ddt_phys_decref(ddt_phys_t *ddp)
316 ASSERT((int64_t)ddp->ddp_refcnt > 0);
321 ddt_phys_free(ddt_t *ddt, ddt_key_t *ddk, ddt_phys_t *ddp, uint64_t txg)
325 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
327 zio_free(ddt->ddt_spa, txg, &blk);
331 ddt_phys_select(const ddt_entry_t *dde, const blkptr_t *bp)
333 ddt_phys_t *ddp = (ddt_phys_t *)dde->dde_phys;
336 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
337 if (DVA_EQUAL(BP_IDENTITY(bp), &ddp->ddp_dva[0]) &&
338 BP_PHYSICAL_BIRTH(bp) == ddp->ddp_phys_birth)
345 ddt_phys_total_refcnt(const ddt_entry_t *dde)
350 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++)
351 refcnt += dde->dde_phys[p].ddp_refcnt;
357 ddt_stat_generate(ddt_t *ddt, ddt_entry_t *dde, ddt_stat_t *dds)
359 spa_t *spa = ddt->ddt_spa;
360 ddt_phys_t *ddp = dde->dde_phys;
361 ddt_key_t *ddk = &dde->dde_key;
362 uint64_t lsize = DDK_GET_LSIZE(ddk);
363 uint64_t psize = DDK_GET_PSIZE(ddk);
366 bzero(dds, sizeof (*dds));
368 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
370 uint64_t refcnt = ddp->ddp_refcnt;
372 if (ddp->ddp_phys_birth == 0)
375 for (d = 0; d < SPA_DVAS_PER_BP; d++)
376 dsize += dva_get_dsize_sync(spa, &ddp->ddp_dva[d]);
378 dds->dds_blocks += 1;
379 dds->dds_lsize += lsize;
380 dds->dds_psize += psize;
381 dds->dds_dsize += dsize;
383 dds->dds_ref_blocks += refcnt;
384 dds->dds_ref_lsize += lsize * refcnt;
385 dds->dds_ref_psize += psize * refcnt;
386 dds->dds_ref_dsize += dsize * refcnt;
391 ddt_stat_add(ddt_stat_t *dst, const ddt_stat_t *src, uint64_t neg)
393 const uint64_t *s = (const uint64_t *)src;
394 uint64_t *d = (uint64_t *)dst;
395 uint64_t *d_end = (uint64_t *)(dst + 1);
397 ASSERT(neg == 0 || neg == -1ULL); /* add or subtract */
400 *d++ += (*s++ ^ neg) - neg;
404 ddt_stat_update(ddt_t *ddt, ddt_entry_t *dde, uint64_t neg)
407 ddt_histogram_t *ddh;
410 ddt_stat_generate(ddt, dde, &dds);
412 bucket = highbit(dds.dds_ref_blocks) - 1;
415 ddh = &ddt->ddt_histogram[dde->dde_type][dde->dde_class];
417 ddt_stat_add(&ddh->ddh_stat[bucket], &dds, neg);
421 ddt_histogram_add(ddt_histogram_t *dst, const ddt_histogram_t *src)
425 for (h = 0; h < 64; h++)
426 ddt_stat_add(&dst->ddh_stat[h], &src->ddh_stat[h], 0);
430 ddt_histogram_stat(ddt_stat_t *dds, const ddt_histogram_t *ddh)
434 bzero(dds, sizeof (*dds));
436 for (h = 0; h < 64; h++)
437 ddt_stat_add(dds, &ddh->ddh_stat[h], 0);
441 ddt_histogram_empty(const ddt_histogram_t *ddh)
443 const uint64_t *s = (const uint64_t *)ddh;
444 const uint64_t *s_end = (const uint64_t *)(ddh + 1);
454 ddt_get_dedup_object_stats(spa_t *spa, ddt_object_t *ddo_total)
458 enum ddt_class class;
460 /* Sum the statistics we cached in ddt_object_sync(). */
461 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
462 ddt_t *ddt = spa->spa_ddt[c];
463 for (type = 0; type < DDT_TYPES; type++) {
464 for (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)
487 enum ddt_class class;
489 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
490 ddt_t *ddt = spa->spa_ddt[c];
491 for (type = 0; type < DDT_TYPES; type++) {
492 for (class = 0; class < DDT_CLASSES;
494 ddt_histogram_add(ddh,
495 &ddt->ddt_histogram_cache[type][class]);
502 ddt_get_dedup_stats(spa_t *spa, ddt_stat_t *dds_total)
504 ddt_histogram_t *ddh_total;
506 /* XXX: Move to a slab */
507 ddh_total = kmem_zalloc(sizeof (ddt_histogram_t), KM_PUSHPAGE);
508 ddt_get_dedup_histogram(spa, ddh_total);
509 ddt_histogram_stat(dds_total, ddh_total);
510 kmem_free(ddh_total, sizeof (ddt_histogram_t));
514 ddt_get_dedup_dspace(spa_t *spa)
516 ddt_stat_t dds_total = { 0 };
518 ddt_get_dedup_stats(spa, &dds_total);
519 return (dds_total.dds_ref_dsize - dds_total.dds_dsize);
523 ddt_get_pool_dedup_ratio(spa_t *spa)
525 ddt_stat_t dds_total = { 0 };
527 ddt_get_dedup_stats(spa, &dds_total);
528 if (dds_total.dds_dsize == 0)
531 return (dds_total.dds_ref_dsize * 100 / dds_total.dds_dsize);
535 ddt_ditto_copies_needed(ddt_t *ddt, ddt_entry_t *dde, ddt_phys_t *ddp_willref)
537 spa_t *spa = ddt->ddt_spa;
538 uint64_t total_refcnt = 0;
539 uint64_t ditto = spa->spa_dedup_ditto;
540 int total_copies = 0;
541 int desired_copies = 0;
544 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
545 ddt_phys_t *ddp = &dde->dde_phys[p];
546 zio_t *zio = dde->dde_lead_zio[p];
547 uint64_t refcnt = ddp->ddp_refcnt; /* committed refs */
549 refcnt += zio->io_parent_count; /* pending refs */
550 if (ddp == ddp_willref)
551 refcnt++; /* caller's ref */
553 total_refcnt += refcnt;
558 if (ditto == 0 || ditto > UINT32_MAX)
561 if (total_refcnt >= 1)
563 if (total_refcnt >= ditto)
565 if (total_refcnt >= ditto * ditto)
568 return (MAX(desired_copies, total_copies) - total_copies);
572 ddt_ditto_copies_present(ddt_entry_t *dde)
574 ddt_phys_t *ddp = &dde->dde_phys[DDT_PHYS_DITTO];
575 dva_t *dva = ddp->ddp_dva;
576 int copies = 0 - DVA_GET_GANG(dva);
579 for (d = 0; d < SPA_DVAS_PER_BP; d++, dva++)
580 if (DVA_IS_VALID(dva))
583 ASSERT(copies >= 0 && copies < SPA_DVAS_PER_BP);
589 ddt_compress(void *src, uchar_t *dst, size_t s_len, size_t d_len)
591 uchar_t *version = dst++;
592 int cpfunc = ZIO_COMPRESS_ZLE;
593 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
596 ASSERT(d_len >= s_len + 1); /* no compression plus version byte */
598 c_len = ci->ci_compress(src, dst, s_len, d_len - 1, ci->ci_level);
600 if (c_len == s_len) {
601 cpfunc = ZIO_COMPRESS_OFF;
602 bcopy(src, dst, s_len);
605 *version = (ZFS_HOST_BYTEORDER & DDT_COMPRESS_BYTEORDER_MASK) | cpfunc;
611 ddt_decompress(uchar_t *src, void *dst, size_t s_len, size_t d_len)
613 uchar_t version = *src++;
614 int cpfunc = version & DDT_COMPRESS_FUNCTION_MASK;
615 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
617 if (ci->ci_decompress != NULL)
618 (void) ci->ci_decompress(src, dst, s_len, d_len, ci->ci_level);
620 bcopy(src, dst, d_len);
622 if ((version ^ ZFS_HOST_BYTEORDER) & DDT_COMPRESS_BYTEORDER_MASK)
623 byteswap_uint64_array(dst, d_len);
627 ddt_select_by_checksum(spa_t *spa, enum zio_checksum c)
629 return (spa->spa_ddt[c]);
633 ddt_select(spa_t *spa, const blkptr_t *bp)
635 return (spa->spa_ddt[BP_GET_CHECKSUM(bp)]);
639 ddt_enter(ddt_t *ddt)
641 mutex_enter(&ddt->ddt_lock);
647 mutex_exit(&ddt->ddt_lock);
651 ddt_alloc(const ddt_key_t *ddk)
655 /* XXX: Move to a slab */
656 dde = kmem_zalloc(sizeof (ddt_entry_t), KM_SLEEP);
657 cv_init(&dde->dde_cv, NULL, CV_DEFAULT, NULL);
665 ddt_free(ddt_entry_t *dde)
669 ASSERT(!dde->dde_loading);
671 for (p = 0; p < DDT_PHYS_TYPES; p++)
672 ASSERT(dde->dde_lead_zio[p] == NULL);
674 if (dde->dde_repair_data != NULL)
675 zio_buf_free(dde->dde_repair_data,
676 DDK_GET_PSIZE(&dde->dde_key));
678 cv_destroy(&dde->dde_cv);
679 kmem_free(dde, sizeof (*dde));
683 ddt_remove(ddt_t *ddt, ddt_entry_t *dde)
685 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
687 avl_remove(&ddt->ddt_tree, dde);
692 ddt_lookup(ddt_t *ddt, const blkptr_t *bp, boolean_t add)
694 ddt_entry_t *dde, dde_search;
696 enum ddt_class class;
700 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
702 ddt_key_fill(&dde_search.dde_key, bp);
704 dde = avl_find(&ddt->ddt_tree, &dde_search, &where);
708 dde = ddt_alloc(&dde_search.dde_key);
709 avl_insert(&ddt->ddt_tree, dde, where);
712 while (dde->dde_loading)
713 cv_wait(&dde->dde_cv, &ddt->ddt_lock);
718 dde->dde_loading = B_TRUE;
724 for (type = 0; type < DDT_TYPES; type++) {
725 for (class = 0; class < DDT_CLASSES; class++) {
726 error = ddt_object_lookup(ddt, type, class, dde);
734 ASSERT(error == 0 || error == ENOENT);
738 ASSERT(dde->dde_loaded == B_FALSE);
739 ASSERT(dde->dde_loading == B_TRUE);
741 dde->dde_type = type; /* will be DDT_TYPES if no entry found */
742 dde->dde_class = class; /* will be DDT_CLASSES if no entry found */
743 dde->dde_loaded = B_TRUE;
744 dde->dde_loading = B_FALSE;
747 ddt_stat_update(ddt, dde, -1ULL);
749 cv_broadcast(&dde->dde_cv);
755 ddt_prefetch(spa_t *spa, const blkptr_t *bp)
760 enum ddt_class class;
762 if (!zfs_dedup_prefetch || bp == NULL || !BP_GET_DEDUP(bp))
766 * We only remove the DDT once all tables are empty and only
767 * prefetch dedup blocks when there are entries in the DDT.
768 * Thus no locking is required as the DDT can't disappear on us.
770 ddt = ddt_select(spa, bp);
771 ddt_key_fill(&dde.dde_key, bp);
773 for (type = 0; type < DDT_TYPES; type++) {
774 for (class = 0; class < DDT_CLASSES; class++) {
775 ddt_object_prefetch(ddt, type, class, &dde);
781 ddt_entry_compare(const void *x1, const void *x2)
783 const ddt_entry_t *dde1 = x1;
784 const ddt_entry_t *dde2 = x2;
785 const uint64_t *u1 = (const uint64_t *)&dde1->dde_key;
786 const uint64_t *u2 = (const uint64_t *)&dde2->dde_key;
789 for (i = 0; i < DDT_KEY_WORDS; i++) {
800 ddt_table_alloc(spa_t *spa, enum zio_checksum c)
804 /* XXX: Move to a slab */
805 ddt = kmem_zalloc(sizeof (*ddt), KM_SLEEP | KM_NODEBUG);
807 mutex_init(&ddt->ddt_lock, NULL, MUTEX_DEFAULT, NULL);
808 avl_create(&ddt->ddt_tree, ddt_entry_compare,
809 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
810 avl_create(&ddt->ddt_repair_tree, ddt_entry_compare,
811 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
812 ddt->ddt_checksum = c;
814 ddt->ddt_os = spa->spa_meta_objset;
820 ddt_table_free(ddt_t *ddt)
822 ASSERT(avl_numnodes(&ddt->ddt_tree) == 0);
823 ASSERT(avl_numnodes(&ddt->ddt_repair_tree) == 0);
824 avl_destroy(&ddt->ddt_tree);
825 avl_destroy(&ddt->ddt_repair_tree);
826 mutex_destroy(&ddt->ddt_lock);
827 kmem_free(ddt, sizeof (*ddt));
831 ddt_create(spa_t *spa)
835 spa->spa_dedup_checksum = ZIO_DEDUPCHECKSUM;
837 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++)
838 spa->spa_ddt[c] = ddt_table_alloc(spa, c);
846 enum ddt_class class;
851 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
852 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1,
853 &spa->spa_ddt_stat_object);
856 return (error == ENOENT ? 0 : error);
858 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
859 ddt_t *ddt = spa->spa_ddt[c];
860 for (type = 0; type < DDT_TYPES; type++) {
861 for (class = 0; class < DDT_CLASSES;
863 error = ddt_object_load(ddt, type, class);
864 if (error != 0 && error != ENOENT)
870 * Seed the cached histograms.
872 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
873 sizeof (ddt->ddt_histogram));
880 ddt_unload(spa_t *spa)
884 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
885 if (spa->spa_ddt[c]) {
886 ddt_table_free(spa->spa_ddt[c]);
887 spa->spa_ddt[c] = NULL;
893 ddt_class_contains(spa_t *spa, enum ddt_class max_class, const blkptr_t *bp)
898 enum ddt_class class;
900 if (!BP_GET_DEDUP(bp))
903 if (max_class == DDT_CLASS_UNIQUE)
906 ddt = spa->spa_ddt[BP_GET_CHECKSUM(bp)];
907 dde = kmem_alloc(sizeof(ddt_entry_t), KM_SLEEP);
909 ddt_key_fill(&(dde->dde_key), bp);
911 for (type = 0; type < DDT_TYPES; type++) {
912 for (class = 0; class <= max_class; class++) {
913 if (ddt_object_lookup(ddt, type, class, dde) == 0) {
914 kmem_free(dde, sizeof(ddt_entry_t));
920 kmem_free(dde, sizeof(ddt_entry_t));
925 ddt_repair_start(ddt_t *ddt, const blkptr_t *bp)
930 enum ddt_class class;
932 ddt_key_fill(&ddk, bp);
934 dde = ddt_alloc(&ddk);
936 for (type = 0; type < DDT_TYPES; type++) {
937 for (class = 0; class < DDT_CLASSES; class++) {
939 * We can only do repair if there are multiple copies
940 * of the block. For anything in the UNIQUE class,
941 * there's definitely only one copy, so don't even try.
943 if (class != DDT_CLASS_UNIQUE &&
944 ddt_object_lookup(ddt, type, class, dde) == 0)
949 bzero(dde->dde_phys, sizeof (dde->dde_phys));
955 ddt_repair_done(ddt_t *ddt, ddt_entry_t *dde)
961 if (dde->dde_repair_data != NULL && spa_writeable(ddt->ddt_spa) &&
962 avl_find(&ddt->ddt_repair_tree, dde, &where) == NULL)
963 avl_insert(&ddt->ddt_repair_tree, dde, where);
971 ddt_repair_entry_done(zio_t *zio)
973 ddt_entry_t *rdde = zio->io_private;
979 ddt_repair_entry(ddt_t *ddt, ddt_entry_t *dde, ddt_entry_t *rdde, zio_t *rio)
981 ddt_phys_t *ddp = dde->dde_phys;
982 ddt_phys_t *rddp = rdde->dde_phys;
983 ddt_key_t *ddk = &dde->dde_key;
984 ddt_key_t *rddk = &rdde->dde_key;
989 zio = zio_null(rio, rio->io_spa, NULL,
990 ddt_repair_entry_done, rdde, rio->io_flags);
992 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++, rddp++) {
993 if (ddp->ddp_phys_birth == 0 ||
994 ddp->ddp_phys_birth != rddp->ddp_phys_birth ||
995 bcmp(ddp->ddp_dva, rddp->ddp_dva, sizeof (ddp->ddp_dva)))
997 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
998 zio_nowait(zio_rewrite(zio, zio->io_spa, 0, &blk,
999 rdde->dde_repair_data, DDK_GET_PSIZE(rddk), NULL, NULL,
1000 ZIO_PRIORITY_SYNC_WRITE, ZIO_DDT_CHILD_FLAGS(zio), NULL));
1007 ddt_repair_table(ddt_t *ddt, zio_t *rio)
1009 spa_t *spa = ddt->ddt_spa;
1010 ddt_entry_t *dde, *rdde_next, *rdde;
1011 avl_tree_t *t = &ddt->ddt_repair_tree;
1014 if (spa_sync_pass(spa) > 1)
1018 for (rdde = avl_first(t); rdde != NULL; rdde = rdde_next) {
1019 rdde_next = AVL_NEXT(t, rdde);
1020 avl_remove(&ddt->ddt_repair_tree, rdde);
1022 ddt_bp_create(ddt->ddt_checksum, &rdde->dde_key, NULL, &blk);
1023 dde = ddt_repair_start(ddt, &blk);
1024 ddt_repair_entry(ddt, dde, rdde, rio);
1025 ddt_repair_done(ddt, dde);
1032 ddt_sync_entry(ddt_t *ddt, ddt_entry_t *dde, dmu_tx_t *tx, uint64_t txg)
1034 dsl_pool_t *dp = ddt->ddt_spa->spa_dsl_pool;
1035 ddt_phys_t *ddp = dde->dde_phys;
1036 ddt_key_t *ddk = &dde->dde_key;
1037 enum ddt_type otype = dde->dde_type;
1038 enum ddt_type ntype = DDT_TYPE_CURRENT;
1039 enum ddt_class oclass = dde->dde_class;
1040 enum ddt_class nclass;
1041 uint64_t total_refcnt = 0;
1044 ASSERT(dde->dde_loaded);
1045 ASSERT(!dde->dde_loading);
1047 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1048 ASSERT(dde->dde_lead_zio[p] == NULL);
1049 ASSERT((int64_t)ddp->ddp_refcnt >= 0);
1050 if (ddp->ddp_phys_birth == 0) {
1051 ASSERT(ddp->ddp_refcnt == 0);
1054 if (p == DDT_PHYS_DITTO) {
1055 if (ddt_ditto_copies_needed(ddt, dde, NULL) == 0)
1056 ddt_phys_free(ddt, ddk, ddp, txg);
1059 if (ddp->ddp_refcnt == 0)
1060 ddt_phys_free(ddt, ddk, ddp, txg);
1061 total_refcnt += ddp->ddp_refcnt;
1064 if (dde->dde_phys[DDT_PHYS_DITTO].ddp_phys_birth != 0)
1065 nclass = DDT_CLASS_DITTO;
1066 else if (total_refcnt > 1)
1067 nclass = DDT_CLASS_DUPLICATE;
1069 nclass = DDT_CLASS_UNIQUE;
1071 if (otype != DDT_TYPES &&
1072 (otype != ntype || oclass != nclass || total_refcnt == 0)) {
1073 VERIFY(ddt_object_remove(ddt, otype, oclass, dde, tx) == 0);
1074 ASSERT(ddt_object_lookup(ddt, otype, oclass, dde) == ENOENT);
1077 if (total_refcnt != 0) {
1078 dde->dde_type = ntype;
1079 dde->dde_class = nclass;
1080 ddt_stat_update(ddt, dde, 0);
1081 if (!ddt_object_exists(ddt, ntype, nclass))
1082 ddt_object_create(ddt, ntype, nclass, tx);
1083 VERIFY(ddt_object_update(ddt, ntype, nclass, dde, tx) == 0);
1086 * If the class changes, the order that we scan this bp
1087 * changes. If it decreases, we could miss it, so
1088 * scan it right now. (This covers both class changing
1089 * while we are doing ddt_walk(), and when we are
1092 if (nclass < oclass) {
1093 dsl_scan_ddt_entry(dp->dp_scan,
1094 ddt->ddt_checksum, dde, tx);
1100 ddt_sync_table(ddt_t *ddt, dmu_tx_t *tx, uint64_t txg)
1102 spa_t *spa = ddt->ddt_spa;
1104 void *cookie = NULL;
1106 enum ddt_class class;
1108 if (avl_numnodes(&ddt->ddt_tree) == 0)
1111 ASSERT(spa->spa_uberblock.ub_version >= SPA_VERSION_DEDUP);
1113 if (spa->spa_ddt_stat_object == 0) {
1114 spa->spa_ddt_stat_object = zap_create(ddt->ddt_os,
1115 DMU_OT_DDT_STATS, DMU_OT_NONE, 0, tx);
1116 VERIFY(zap_add(ddt->ddt_os, DMU_POOL_DIRECTORY_OBJECT,
1117 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1,
1118 &spa->spa_ddt_stat_object, tx) == 0);
1121 while ((dde = avl_destroy_nodes(&ddt->ddt_tree, &cookie)) != NULL) {
1122 ddt_sync_entry(ddt, dde, tx, txg);
1126 for (type = 0; type < DDT_TYPES; type++) {
1128 for (class = 0; class < DDT_CLASSES; class++) {
1129 if (ddt_object_exists(ddt, type, class)) {
1130 ddt_object_sync(ddt, type, class, tx);
1131 count += ddt_object_count(ddt, type, class);
1134 for (class = 0; class < DDT_CLASSES; class++) {
1135 if (count == 0 && ddt_object_exists(ddt, type, class))
1136 ddt_object_destroy(ddt, type, class, tx);
1140 bcopy(ddt->ddt_histogram, &ddt->ddt_histogram_cache,
1141 sizeof (ddt->ddt_histogram));
1145 ddt_sync(spa_t *spa, uint64_t txg)
1148 zio_t *rio = zio_root(spa, NULL, NULL,
1149 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
1150 enum zio_checksum c;
1152 ASSERT(spa_syncing_txg(spa) == txg);
1154 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1156 for (c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
1157 ddt_t *ddt = spa->spa_ddt[c];
1160 ddt_sync_table(ddt, tx, txg);
1161 ddt_repair_table(ddt, rio);
1164 (void) zio_wait(rio);
1170 ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde)
1175 ddt_t *ddt = spa->spa_ddt[ddb->ddb_checksum];
1177 if (ddt_object_exists(ddt, ddb->ddb_type,
1179 error = ddt_object_walk(ddt,
1180 ddb->ddb_type, ddb->ddb_class,
1181 &ddb->ddb_cursor, dde);
1183 dde->dde_type = ddb->ddb_type;
1184 dde->dde_class = ddb->ddb_class;
1187 if (error != ENOENT)
1189 ddb->ddb_cursor = 0;
1190 } while (++ddb->ddb_checksum < ZIO_CHECKSUM_FUNCTIONS);
1191 ddb->ddb_checksum = 0;
1192 } while (++ddb->ddb_type < DDT_TYPES);
1194 } while (++ddb->ddb_class < DDT_CLASSES);
1199 #if defined(_KERNEL) && defined(HAVE_SPL)
1200 module_param(zfs_dedup_prefetch, int, 0644);
1201 MODULE_PARM_DESC(zfs_dedup_prefetch,"Enable prefetching dedup-ed blks");