BRT should return EOPNOTSUPP
[zfs.git] / module / zfs / ddt.c
blob1fb1982199042e68ecb38c2312b51a9d69f13d72
1 /*
2 * CDDL HEADER START
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 https://opensource.org/licenses/CDDL-1.0.
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]
19 * CDDL HEADER END
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
25 * Copyright (c) 2022 by Pawel Jakub Dawidek
28 #include <sys/zfs_context.h>
29 #include <sys/spa.h>
30 #include <sys/spa_impl.h>
31 #include <sys/zio.h>
32 #include <sys/ddt.h>
33 #include <sys/zap.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/arc.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/zio_compress.h>
39 #include <sys/dsl_scan.h>
40 #include <sys/abd.h>
42 static kmem_cache_t *ddt_cache;
43 static kmem_cache_t *ddt_entry_cache;
46 * Enable/disable prefetching of dedup-ed blocks which are going to be freed.
48 int zfs_dedup_prefetch = 0;
50 static const ddt_ops_t *const ddt_ops[DDT_TYPES] = {
51 &ddt_zap_ops,
54 static const char *const ddt_class_name[DDT_CLASSES] = {
55 "ditto",
56 "duplicate",
57 "unique",
60 static void
61 ddt_object_create(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
62 dmu_tx_t *tx)
64 spa_t *spa = ddt->ddt_spa;
65 objset_t *os = ddt->ddt_os;
66 uint64_t *objectp = &ddt->ddt_object[type][class];
67 boolean_t prehash = zio_checksum_table[ddt->ddt_checksum].ci_flags &
68 ZCHECKSUM_FLAG_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);
85 static void
86 ddt_object_destroy(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
87 dmu_tx_t *tx)
89 spa_t *spa = ddt->ddt_spa;
90 objset_t *os = ddt->ddt_os;
91 uint64_t *objectp = &ddt->ddt_object[type][class];
92 uint64_t count;
93 char name[DDT_NAMELEN];
95 ddt_object_name(ddt, type, class, name);
97 ASSERT(*objectp != 0);
98 ASSERT(ddt_histogram_empty(&ddt->ddt_histogram[type][class]));
99 VERIFY(ddt_object_count(ddt, type, class, &count) == 0 && count == 0);
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 memset(&ddt->ddt_object_stats[type][class], 0, sizeof (ddt_object_t));
105 *objectp = 0;
108 static int
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;
113 uint64_t count;
114 char name[DDT_NAMELEN];
115 int error;
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]);
121 if (error != 0)
122 return (error);
124 error = zap_lookup(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
125 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
126 &ddt->ddt_histogram[type][class]);
127 if (error != 0)
128 return (error);
131 * Seed the cached statistics.
133 error = ddt_object_info(ddt, type, class, &doi);
134 if (error)
135 return (error);
137 error = ddt_object_count(ddt, type, class, &count);
138 if (error)
139 return (error);
141 ddo->ddo_count = count;
142 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
143 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
145 return (0);
148 static void
149 ddt_object_sync(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
150 dmu_tx_t *tx)
152 ddt_object_t *ddo = &ddt->ddt_object_stats[type][class];
153 dmu_object_info_t doi;
154 uint64_t count;
155 char name[DDT_NAMELEN];
157 ddt_object_name(ddt, type, class, name);
159 VERIFY(zap_update(ddt->ddt_os, ddt->ddt_spa->spa_ddt_stat_object, name,
160 sizeof (uint64_t), sizeof (ddt_histogram_t) / sizeof (uint64_t),
161 &ddt->ddt_histogram[type][class], tx) == 0);
164 * Cache DDT statistics; this is the only time they'll change.
166 VERIFY(ddt_object_info(ddt, type, class, &doi) == 0);
167 VERIFY(ddt_object_count(ddt, type, class, &count) == 0);
169 ddo->ddo_count = count;
170 ddo->ddo_dspace = doi.doi_physical_blocks_512 << 9;
171 ddo->ddo_mspace = doi.doi_fill_count * doi.doi_data_block_size;
174 static int
175 ddt_object_lookup(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
176 ddt_entry_t *dde)
178 if (!ddt_object_exists(ddt, type, class))
179 return (SET_ERROR(ENOENT));
181 return (ddt_ops[type]->ddt_op_lookup(ddt->ddt_os,
182 ddt->ddt_object[type][class], dde));
185 static void
186 ddt_object_prefetch(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
187 ddt_entry_t *dde)
189 if (!ddt_object_exists(ddt, type, class))
190 return;
192 ddt_ops[type]->ddt_op_prefetch(ddt->ddt_os,
193 ddt->ddt_object[type][class], dde);
197 ddt_object_update(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
198 ddt_entry_t *dde, dmu_tx_t *tx)
200 ASSERT(ddt_object_exists(ddt, type, class));
202 return (ddt_ops[type]->ddt_op_update(ddt->ddt_os,
203 ddt->ddt_object[type][class], dde, tx));
206 static int
207 ddt_object_remove(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
208 ddt_entry_t *dde, dmu_tx_t *tx)
210 ASSERT(ddt_object_exists(ddt, type, class));
212 return (ddt_ops[type]->ddt_op_remove(ddt->ddt_os,
213 ddt->ddt_object[type][class], dde, tx));
217 ddt_object_walk(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
218 uint64_t *walk, ddt_entry_t *dde)
220 ASSERT(ddt_object_exists(ddt, type, class));
222 return (ddt_ops[type]->ddt_op_walk(ddt->ddt_os,
223 ddt->ddt_object[type][class], dde, walk));
227 ddt_object_count(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
228 uint64_t *count)
230 ASSERT(ddt_object_exists(ddt, type, class));
232 return (ddt_ops[type]->ddt_op_count(ddt->ddt_os,
233 ddt->ddt_object[type][class], count));
237 ddt_object_info(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
238 dmu_object_info_t *doi)
240 if (!ddt_object_exists(ddt, type, class))
241 return (SET_ERROR(ENOENT));
243 return (dmu_object_info(ddt->ddt_os, ddt->ddt_object[type][class],
244 doi));
247 boolean_t
248 ddt_object_exists(ddt_t *ddt, enum ddt_type type, enum ddt_class class)
250 return (!!ddt->ddt_object[type][class]);
253 void
254 ddt_object_name(ddt_t *ddt, enum ddt_type type, enum ddt_class class,
255 char *name)
257 (void) snprintf(name, DDT_NAMELEN, DMU_POOL_DDT,
258 zio_checksum_table[ddt->ddt_checksum].ci_name,
259 ddt_ops[type]->ddt_op_name, ddt_class_name[class]);
262 void
263 ddt_bp_fill(const ddt_phys_t *ddp, blkptr_t *bp, uint64_t txg)
265 ASSERT(txg != 0);
267 for (int d = 0; d < SPA_DVAS_PER_BP; d++)
268 bp->blk_dva[d] = ddp->ddp_dva[d];
269 BP_SET_BIRTH(bp, txg, ddp->ddp_phys_birth);
273 * The bp created via this function may be used for repairs and scrub, but it
274 * will be missing the salt / IV required to do a full decrypting read.
276 void
277 ddt_bp_create(enum zio_checksum checksum,
278 const ddt_key_t *ddk, const ddt_phys_t *ddp, blkptr_t *bp)
280 BP_ZERO(bp);
282 if (ddp != NULL)
283 ddt_bp_fill(ddp, bp, ddp->ddp_phys_birth);
285 bp->blk_cksum = ddk->ddk_cksum;
287 BP_SET_LSIZE(bp, DDK_GET_LSIZE(ddk));
288 BP_SET_PSIZE(bp, DDK_GET_PSIZE(ddk));
289 BP_SET_COMPRESS(bp, DDK_GET_COMPRESS(ddk));
290 BP_SET_CRYPT(bp, DDK_GET_CRYPT(ddk));
291 BP_SET_FILL(bp, 1);
292 BP_SET_CHECKSUM(bp, checksum);
293 BP_SET_TYPE(bp, DMU_OT_DEDUP);
294 BP_SET_LEVEL(bp, 0);
295 BP_SET_DEDUP(bp, 1);
296 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
299 void
300 ddt_key_fill(ddt_key_t *ddk, const blkptr_t *bp)
302 ddk->ddk_cksum = bp->blk_cksum;
303 ddk->ddk_prop = 0;
305 ASSERT(BP_IS_ENCRYPTED(bp) || !BP_USES_CRYPT(bp));
307 DDK_SET_LSIZE(ddk, BP_GET_LSIZE(bp));
308 DDK_SET_PSIZE(ddk, BP_GET_PSIZE(bp));
309 DDK_SET_COMPRESS(ddk, BP_GET_COMPRESS(bp));
310 DDK_SET_CRYPT(ddk, BP_USES_CRYPT(bp));
313 void
314 ddt_phys_fill(ddt_phys_t *ddp, const blkptr_t *bp)
316 ASSERT(ddp->ddp_phys_birth == 0);
318 for (int d = 0; d < SPA_DVAS_PER_BP; d++)
319 ddp->ddp_dva[d] = bp->blk_dva[d];
320 ddp->ddp_phys_birth = BP_PHYSICAL_BIRTH(bp);
323 void
324 ddt_phys_clear(ddt_phys_t *ddp)
326 memset(ddp, 0, sizeof (*ddp));
329 void
330 ddt_phys_addref(ddt_phys_t *ddp)
332 ddp->ddp_refcnt++;
335 void
336 ddt_phys_decref(ddt_phys_t *ddp)
338 if (ddp) {
339 ASSERT(ddp->ddp_refcnt > 0);
340 ddp->ddp_refcnt--;
344 void
345 ddt_phys_free(ddt_t *ddt, ddt_key_t *ddk, ddt_phys_t *ddp, uint64_t txg)
347 blkptr_t blk;
349 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
352 * We clear the dedup bit so that zio_free() will actually free the
353 * space, rather than just decrementing the refcount in the DDT.
355 BP_SET_DEDUP(&blk, 0);
357 ddt_phys_clear(ddp);
358 zio_free(ddt->ddt_spa, txg, &blk);
361 ddt_phys_t *
362 ddt_phys_select(const ddt_entry_t *dde, const blkptr_t *bp)
364 ddt_phys_t *ddp = (ddt_phys_t *)dde->dde_phys;
366 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
367 if (DVA_EQUAL(BP_IDENTITY(bp), &ddp->ddp_dva[0]) &&
368 BP_PHYSICAL_BIRTH(bp) == ddp->ddp_phys_birth)
369 return (ddp);
371 return (NULL);
374 uint64_t
375 ddt_phys_total_refcnt(const ddt_entry_t *dde)
377 uint64_t refcnt = 0;
379 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++)
380 refcnt += dde->dde_phys[p].ddp_refcnt;
382 return (refcnt);
385 static void
386 ddt_stat_generate(ddt_t *ddt, ddt_entry_t *dde, ddt_stat_t *dds)
388 spa_t *spa = ddt->ddt_spa;
389 ddt_phys_t *ddp = dde->dde_phys;
390 ddt_key_t *ddk = &dde->dde_key;
391 uint64_t lsize = DDK_GET_LSIZE(ddk);
392 uint64_t psize = DDK_GET_PSIZE(ddk);
394 memset(dds, 0, sizeof (*dds));
396 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
397 uint64_t dsize = 0;
398 uint64_t refcnt = ddp->ddp_refcnt;
400 if (ddp->ddp_phys_birth == 0)
401 continue;
403 for (int d = 0; d < DDE_GET_NDVAS(dde); d++)
404 dsize += dva_get_dsize_sync(spa, &ddp->ddp_dva[d]);
406 dds->dds_blocks += 1;
407 dds->dds_lsize += lsize;
408 dds->dds_psize += psize;
409 dds->dds_dsize += dsize;
411 dds->dds_ref_blocks += refcnt;
412 dds->dds_ref_lsize += lsize * refcnt;
413 dds->dds_ref_psize += psize * refcnt;
414 dds->dds_ref_dsize += dsize * refcnt;
418 void
419 ddt_stat_add(ddt_stat_t *dst, const ddt_stat_t *src, uint64_t neg)
421 const uint64_t *s = (const uint64_t *)src;
422 uint64_t *d = (uint64_t *)dst;
423 uint64_t *d_end = (uint64_t *)(dst + 1);
425 ASSERT(neg == 0 || neg == -1ULL); /* add or subtract */
427 for (int i = 0; i < d_end - d; i++)
428 d[i] += (s[i] ^ neg) - neg;
431 static void
432 ddt_stat_update(ddt_t *ddt, ddt_entry_t *dde, uint64_t neg)
434 ddt_stat_t dds;
435 ddt_histogram_t *ddh;
436 int bucket;
438 ddt_stat_generate(ddt, dde, &dds);
440 bucket = highbit64(dds.dds_ref_blocks) - 1;
441 ASSERT(bucket >= 0);
443 ddh = &ddt->ddt_histogram[dde->dde_type][dde->dde_class];
445 ddt_stat_add(&ddh->ddh_stat[bucket], &dds, neg);
448 void
449 ddt_histogram_add(ddt_histogram_t *dst, const ddt_histogram_t *src)
451 for (int h = 0; h < 64; h++)
452 ddt_stat_add(&dst->ddh_stat[h], &src->ddh_stat[h], 0);
455 void
456 ddt_histogram_stat(ddt_stat_t *dds, const ddt_histogram_t *ddh)
458 memset(dds, 0, sizeof (*dds));
460 for (int h = 0; h < 64; h++)
461 ddt_stat_add(dds, &ddh->ddh_stat[h], 0);
464 boolean_t
465 ddt_histogram_empty(const ddt_histogram_t *ddh)
467 const uint64_t *s = (const uint64_t *)ddh;
468 const uint64_t *s_end = (const uint64_t *)(ddh + 1);
470 while (s < s_end)
471 if (*s++ != 0)
472 return (B_FALSE);
474 return (B_TRUE);
477 void
478 ddt_get_dedup_object_stats(spa_t *spa, ddt_object_t *ddo_total)
480 /* Sum the statistics we cached in ddt_object_sync(). */
481 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
482 ddt_t *ddt = spa->spa_ddt[c];
483 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
484 for (enum ddt_class class = 0; class < DDT_CLASSES;
485 class++) {
486 ddt_object_t *ddo =
487 &ddt->ddt_object_stats[type][class];
488 ddo_total->ddo_count += ddo->ddo_count;
489 ddo_total->ddo_dspace += ddo->ddo_dspace;
490 ddo_total->ddo_mspace += ddo->ddo_mspace;
495 /* ... and compute the averages. */
496 if (ddo_total->ddo_count != 0) {
497 ddo_total->ddo_dspace /= ddo_total->ddo_count;
498 ddo_total->ddo_mspace /= ddo_total->ddo_count;
502 void
503 ddt_get_dedup_histogram(spa_t *spa, ddt_histogram_t *ddh)
505 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
506 ddt_t *ddt = spa->spa_ddt[c];
507 for (enum ddt_type type = 0; type < DDT_TYPES && ddt; type++) {
508 for (enum ddt_class class = 0; class < DDT_CLASSES;
509 class++) {
510 ddt_histogram_add(ddh,
511 &ddt->ddt_histogram_cache[type][class]);
517 void
518 ddt_get_dedup_stats(spa_t *spa, ddt_stat_t *dds_total)
520 ddt_histogram_t *ddh_total;
522 ddh_total = kmem_zalloc(sizeof (ddt_histogram_t), KM_SLEEP);
523 ddt_get_dedup_histogram(spa, ddh_total);
524 ddt_histogram_stat(dds_total, ddh_total);
525 kmem_free(ddh_total, sizeof (ddt_histogram_t));
528 uint64_t
529 ddt_get_dedup_dspace(spa_t *spa)
531 ddt_stat_t dds_total;
533 if (spa->spa_dedup_dspace != ~0ULL)
534 return (spa->spa_dedup_dspace);
536 memset(&dds_total, 0, sizeof (ddt_stat_t));
538 /* Calculate and cache the stats */
539 ddt_get_dedup_stats(spa, &dds_total);
540 spa->spa_dedup_dspace = dds_total.dds_ref_dsize - dds_total.dds_dsize;
541 return (spa->spa_dedup_dspace);
544 uint64_t
545 ddt_get_pool_dedup_ratio(spa_t *spa)
547 ddt_stat_t dds_total = { 0 };
549 ddt_get_dedup_stats(spa, &dds_total);
550 if (dds_total.dds_dsize == 0)
551 return (100);
553 return (dds_total.dds_ref_dsize * 100 / dds_total.dds_dsize);
556 size_t
557 ddt_compress(void *src, uchar_t *dst, size_t s_len, size_t d_len)
559 uchar_t *version = dst++;
560 int cpfunc = ZIO_COMPRESS_ZLE;
561 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
562 size_t c_len;
564 ASSERT(d_len >= s_len + 1); /* no compression plus version byte */
566 c_len = ci->ci_compress(src, dst, s_len, d_len - 1, ci->ci_level);
568 if (c_len == s_len) {
569 cpfunc = ZIO_COMPRESS_OFF;
570 memcpy(dst, src, s_len);
573 *version = cpfunc;
574 if (ZFS_HOST_BYTEORDER)
575 *version |= DDT_COMPRESS_BYTEORDER_MASK;
577 return (c_len + 1);
580 void
581 ddt_decompress(uchar_t *src, void *dst, size_t s_len, size_t d_len)
583 uchar_t version = *src++;
584 int cpfunc = version & DDT_COMPRESS_FUNCTION_MASK;
585 zio_compress_info_t *ci = &zio_compress_table[cpfunc];
587 if (ci->ci_decompress != NULL)
588 (void) ci->ci_decompress(src, dst, s_len, d_len, ci->ci_level);
589 else
590 memcpy(dst, src, d_len);
592 if (((version & DDT_COMPRESS_BYTEORDER_MASK) != 0) !=
593 (ZFS_HOST_BYTEORDER != 0))
594 byteswap_uint64_array(dst, d_len);
597 ddt_t *
598 ddt_select(spa_t *spa, const blkptr_t *bp)
600 return (spa->spa_ddt[BP_GET_CHECKSUM(bp)]);
603 void
604 ddt_enter(ddt_t *ddt)
606 mutex_enter(&ddt->ddt_lock);
609 void
610 ddt_exit(ddt_t *ddt)
612 mutex_exit(&ddt->ddt_lock);
615 void
616 ddt_init(void)
618 ddt_cache = kmem_cache_create("ddt_cache",
619 sizeof (ddt_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
620 ddt_entry_cache = kmem_cache_create("ddt_entry_cache",
621 sizeof (ddt_entry_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
624 void
625 ddt_fini(void)
627 kmem_cache_destroy(ddt_entry_cache);
628 kmem_cache_destroy(ddt_cache);
631 static ddt_entry_t *
632 ddt_alloc(const ddt_key_t *ddk)
634 ddt_entry_t *dde;
636 dde = kmem_cache_alloc(ddt_entry_cache, KM_SLEEP);
637 memset(dde, 0, sizeof (ddt_entry_t));
638 cv_init(&dde->dde_cv, NULL, CV_DEFAULT, NULL);
640 dde->dde_key = *ddk;
642 return (dde);
645 static void
646 ddt_free(ddt_entry_t *dde)
648 ASSERT(!dde->dde_loading);
650 for (int p = 0; p < DDT_PHYS_TYPES; p++)
651 ASSERT(dde->dde_lead_zio[p] == NULL);
653 if (dde->dde_repair_abd != NULL)
654 abd_free(dde->dde_repair_abd);
656 cv_destroy(&dde->dde_cv);
657 kmem_cache_free(ddt_entry_cache, dde);
660 void
661 ddt_remove(ddt_t *ddt, ddt_entry_t *dde)
663 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
665 avl_remove(&ddt->ddt_tree, dde);
666 ddt_free(dde);
669 ddt_entry_t *
670 ddt_lookup(ddt_t *ddt, const blkptr_t *bp, boolean_t add)
672 ddt_entry_t *dde, dde_search;
673 enum ddt_type type;
674 enum ddt_class class;
675 avl_index_t where;
676 int error;
678 ASSERT(MUTEX_HELD(&ddt->ddt_lock));
680 ddt_key_fill(&dde_search.dde_key, bp);
682 dde = avl_find(&ddt->ddt_tree, &dde_search, &where);
683 if (dde == NULL) {
684 if (!add)
685 return (NULL);
686 dde = ddt_alloc(&dde_search.dde_key);
687 avl_insert(&ddt->ddt_tree, dde, where);
690 while (dde->dde_loading)
691 cv_wait(&dde->dde_cv, &ddt->ddt_lock);
693 if (dde->dde_loaded)
694 return (dde);
696 dde->dde_loading = B_TRUE;
698 ddt_exit(ddt);
700 error = ENOENT;
702 for (type = 0; type < DDT_TYPES; type++) {
703 for (class = 0; class < DDT_CLASSES; class++) {
704 error = ddt_object_lookup(ddt, type, class, dde);
705 if (error != ENOENT) {
706 ASSERT0(error);
707 break;
710 if (error != ENOENT)
711 break;
714 ddt_enter(ddt);
716 ASSERT(dde->dde_loaded == B_FALSE);
717 ASSERT(dde->dde_loading == B_TRUE);
719 dde->dde_type = type; /* will be DDT_TYPES if no entry found */
720 dde->dde_class = class; /* will be DDT_CLASSES if no entry found */
721 dde->dde_loaded = B_TRUE;
722 dde->dde_loading = B_FALSE;
724 if (error == 0)
725 ddt_stat_update(ddt, dde, -1ULL);
727 cv_broadcast(&dde->dde_cv);
729 return (dde);
732 void
733 ddt_prefetch(spa_t *spa, const blkptr_t *bp)
735 ddt_t *ddt;
736 ddt_entry_t dde;
738 if (!zfs_dedup_prefetch || bp == NULL || !BP_GET_DEDUP(bp))
739 return;
742 * We only remove the DDT once all tables are empty and only
743 * prefetch dedup blocks when there are entries in the DDT.
744 * Thus no locking is required as the DDT can't disappear on us.
746 ddt = ddt_select(spa, bp);
747 ddt_key_fill(&dde.dde_key, bp);
749 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
750 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
751 ddt_object_prefetch(ddt, type, class, &dde);
757 * Opaque struct used for ddt_key comparison
759 #define DDT_KEY_CMP_LEN (sizeof (ddt_key_t) / sizeof (uint16_t))
761 typedef struct ddt_key_cmp {
762 uint16_t u16[DDT_KEY_CMP_LEN];
763 } ddt_key_cmp_t;
766 ddt_entry_compare(const void *x1, const void *x2)
768 const ddt_entry_t *dde1 = x1;
769 const ddt_entry_t *dde2 = x2;
770 const ddt_key_cmp_t *k1 = (const ddt_key_cmp_t *)&dde1->dde_key;
771 const ddt_key_cmp_t *k2 = (const ddt_key_cmp_t *)&dde2->dde_key;
772 int32_t cmp = 0;
774 for (int i = 0; i < DDT_KEY_CMP_LEN; i++) {
775 cmp = (int32_t)k1->u16[i] - (int32_t)k2->u16[i];
776 if (likely(cmp))
777 break;
780 return (TREE_ISIGN(cmp));
783 static ddt_t *
784 ddt_table_alloc(spa_t *spa, enum zio_checksum c)
786 ddt_t *ddt;
788 ddt = kmem_cache_alloc(ddt_cache, KM_SLEEP);
789 memset(ddt, 0, sizeof (ddt_t));
791 mutex_init(&ddt->ddt_lock, NULL, MUTEX_DEFAULT, NULL);
792 avl_create(&ddt->ddt_tree, ddt_entry_compare,
793 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
794 avl_create(&ddt->ddt_repair_tree, ddt_entry_compare,
795 sizeof (ddt_entry_t), offsetof(ddt_entry_t, dde_node));
796 ddt->ddt_checksum = c;
797 ddt->ddt_spa = spa;
798 ddt->ddt_os = spa->spa_meta_objset;
800 return (ddt);
803 static void
804 ddt_table_free(ddt_t *ddt)
806 ASSERT(avl_numnodes(&ddt->ddt_tree) == 0);
807 ASSERT(avl_numnodes(&ddt->ddt_repair_tree) == 0);
808 avl_destroy(&ddt->ddt_tree);
809 avl_destroy(&ddt->ddt_repair_tree);
810 mutex_destroy(&ddt->ddt_lock);
811 kmem_cache_free(ddt_cache, ddt);
814 void
815 ddt_create(spa_t *spa)
817 spa->spa_dedup_checksum = ZIO_DEDUPCHECKSUM;
819 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++)
820 spa->spa_ddt[c] = ddt_table_alloc(spa, c);
824 ddt_load(spa_t *spa)
826 int error;
828 ddt_create(spa);
830 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
831 DMU_POOL_DDT_STATS, sizeof (uint64_t), 1,
832 &spa->spa_ddt_stat_object);
834 if (error)
835 return (error == ENOENT ? 0 : error);
837 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
838 ddt_t *ddt = spa->spa_ddt[c];
839 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
840 for (enum ddt_class class = 0; class < DDT_CLASSES;
841 class++) {
842 error = ddt_object_load(ddt, type, class);
843 if (error != 0 && error != ENOENT)
844 return (error);
849 * Seed the cached histograms.
851 memcpy(&ddt->ddt_histogram_cache, ddt->ddt_histogram,
852 sizeof (ddt->ddt_histogram));
853 spa->spa_dedup_dspace = ~0ULL;
856 return (0);
859 void
860 ddt_unload(spa_t *spa)
862 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
863 if (spa->spa_ddt[c]) {
864 ddt_table_free(spa->spa_ddt[c]);
865 spa->spa_ddt[c] = NULL;
870 boolean_t
871 ddt_class_contains(spa_t *spa, enum ddt_class max_class, const blkptr_t *bp)
873 ddt_t *ddt;
874 ddt_entry_t *dde;
876 if (!BP_GET_DEDUP(bp))
877 return (B_FALSE);
879 if (max_class == DDT_CLASS_UNIQUE)
880 return (B_TRUE);
882 ddt = spa->spa_ddt[BP_GET_CHECKSUM(bp)];
883 dde = kmem_cache_alloc(ddt_entry_cache, KM_SLEEP);
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) {
890 kmem_cache_free(ddt_entry_cache, dde);
891 return (B_TRUE);
896 kmem_cache_free(ddt_entry_cache, dde);
897 return (B_FALSE);
900 ddt_entry_t *
901 ddt_repair_start(ddt_t *ddt, const blkptr_t *bp)
903 ddt_key_t ddk;
904 ddt_entry_t *dde;
906 ddt_key_fill(&ddk, bp);
908 dde = ddt_alloc(&ddk);
910 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
911 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
913 * We can only do repair if there are multiple copies
914 * of the block. For anything in the UNIQUE class,
915 * there's definitely only one copy, so don't even try.
917 if (class != DDT_CLASS_UNIQUE &&
918 ddt_object_lookup(ddt, type, class, dde) == 0)
919 return (dde);
923 memset(dde->dde_phys, 0, sizeof (dde->dde_phys));
925 return (dde);
928 void
929 ddt_repair_done(ddt_t *ddt, ddt_entry_t *dde)
931 avl_index_t where;
933 ddt_enter(ddt);
935 if (dde->dde_repair_abd != NULL && spa_writeable(ddt->ddt_spa) &&
936 avl_find(&ddt->ddt_repair_tree, dde, &where) == NULL)
937 avl_insert(&ddt->ddt_repair_tree, dde, where);
938 else
939 ddt_free(dde);
941 ddt_exit(ddt);
944 static void
945 ddt_repair_entry_done(zio_t *zio)
947 ddt_entry_t *rdde = zio->io_private;
949 ddt_free(rdde);
952 static void
953 ddt_repair_entry(ddt_t *ddt, ddt_entry_t *dde, ddt_entry_t *rdde, zio_t *rio)
955 ddt_phys_t *ddp = dde->dde_phys;
956 ddt_phys_t *rddp = rdde->dde_phys;
957 ddt_key_t *ddk = &dde->dde_key;
958 ddt_key_t *rddk = &rdde->dde_key;
959 zio_t *zio;
960 blkptr_t blk;
962 zio = zio_null(rio, rio->io_spa, NULL,
963 ddt_repair_entry_done, rdde, rio->io_flags);
965 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++, rddp++) {
966 if (ddp->ddp_phys_birth == 0 ||
967 ddp->ddp_phys_birth != rddp->ddp_phys_birth ||
968 memcmp(ddp->ddp_dva, rddp->ddp_dva, sizeof (ddp->ddp_dva)))
969 continue;
970 ddt_bp_create(ddt->ddt_checksum, ddk, ddp, &blk);
971 zio_nowait(zio_rewrite(zio, zio->io_spa, 0, &blk,
972 rdde->dde_repair_abd, DDK_GET_PSIZE(rddk), NULL, NULL,
973 ZIO_PRIORITY_SYNC_WRITE, ZIO_DDT_CHILD_FLAGS(zio), NULL));
976 zio_nowait(zio);
979 static void
980 ddt_repair_table(ddt_t *ddt, zio_t *rio)
982 spa_t *spa = ddt->ddt_spa;
983 ddt_entry_t *dde, *rdde_next, *rdde;
984 avl_tree_t *t = &ddt->ddt_repair_tree;
985 blkptr_t blk;
987 if (spa_sync_pass(spa) > 1)
988 return;
990 ddt_enter(ddt);
991 for (rdde = avl_first(t); rdde != NULL; rdde = rdde_next) {
992 rdde_next = AVL_NEXT(t, rdde);
993 avl_remove(&ddt->ddt_repair_tree, rdde);
994 ddt_exit(ddt);
995 ddt_bp_create(ddt->ddt_checksum, &rdde->dde_key, NULL, &blk);
996 dde = ddt_repair_start(ddt, &blk);
997 ddt_repair_entry(ddt, dde, rdde, rio);
998 ddt_repair_done(ddt, dde);
999 ddt_enter(ddt);
1001 ddt_exit(ddt);
1004 static void
1005 ddt_sync_entry(ddt_t *ddt, ddt_entry_t *dde, dmu_tx_t *tx, uint64_t txg)
1007 dsl_pool_t *dp = ddt->ddt_spa->spa_dsl_pool;
1008 ddt_phys_t *ddp = dde->dde_phys;
1009 ddt_key_t *ddk = &dde->dde_key;
1010 enum ddt_type otype = dde->dde_type;
1011 enum ddt_type ntype = DDT_TYPE_CURRENT;
1012 enum ddt_class oclass = dde->dde_class;
1013 enum ddt_class nclass;
1014 uint64_t total_refcnt = 0;
1016 ASSERT(dde->dde_loaded);
1017 ASSERT(!dde->dde_loading);
1019 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1020 ASSERT(dde->dde_lead_zio[p] == NULL);
1021 if (ddp->ddp_phys_birth == 0) {
1022 ASSERT(ddp->ddp_refcnt == 0);
1023 continue;
1025 if (p == DDT_PHYS_DITTO) {
1027 * Note, we no longer create DDT-DITTO blocks, but we
1028 * don't want to leak any written by older software.
1030 ddt_phys_free(ddt, ddk, ddp, txg);
1031 continue;
1033 if (ddp->ddp_refcnt == 0)
1034 ddt_phys_free(ddt, ddk, ddp, txg);
1035 total_refcnt += ddp->ddp_refcnt;
1038 /* We do not create new DDT-DITTO blocks. */
1039 ASSERT0(dde->dde_phys[DDT_PHYS_DITTO].ddp_phys_birth);
1040 if (total_refcnt > 1)
1041 nclass = DDT_CLASS_DUPLICATE;
1042 else
1043 nclass = DDT_CLASS_UNIQUE;
1045 if (otype != DDT_TYPES &&
1046 (otype != ntype || oclass != nclass || total_refcnt == 0)) {
1047 VERIFY(ddt_object_remove(ddt, otype, oclass, dde, tx) == 0);
1048 ASSERT(ddt_object_lookup(ddt, otype, oclass, dde) == ENOENT);
1051 if (total_refcnt != 0) {
1052 dde->dde_type = ntype;
1053 dde->dde_class = nclass;
1054 ddt_stat_update(ddt, dde, 0);
1055 if (!ddt_object_exists(ddt, ntype, nclass))
1056 ddt_object_create(ddt, ntype, nclass, tx);
1057 VERIFY(ddt_object_update(ddt, ntype, nclass, dde, tx) == 0);
1060 * If the class changes, the order that we scan this bp
1061 * changes. If it decreases, we could miss it, so
1062 * scan it right now. (This covers both class changing
1063 * while we are doing ddt_walk(), and when we are
1064 * traversing.)
1066 if (nclass < oclass) {
1067 dsl_scan_ddt_entry(dp->dp_scan,
1068 ddt->ddt_checksum, dde, tx);
1073 static void
1074 ddt_sync_table(ddt_t *ddt, dmu_tx_t *tx, uint64_t txg)
1076 spa_t *spa = ddt->ddt_spa;
1077 ddt_entry_t *dde;
1078 void *cookie = NULL;
1080 if (avl_numnodes(&ddt->ddt_tree) == 0)
1081 return;
1083 ASSERT(spa->spa_uberblock.ub_version >= SPA_VERSION_DEDUP);
1085 if (spa->spa_ddt_stat_object == 0) {
1086 spa->spa_ddt_stat_object = zap_create_link(ddt->ddt_os,
1087 DMU_OT_DDT_STATS, DMU_POOL_DIRECTORY_OBJECT,
1088 DMU_POOL_DDT_STATS, tx);
1091 while ((dde = avl_destroy_nodes(&ddt->ddt_tree, &cookie)) != NULL) {
1092 ddt_sync_entry(ddt, dde, tx, txg);
1093 ddt_free(dde);
1096 for (enum ddt_type type = 0; type < DDT_TYPES; type++) {
1097 uint64_t add, count = 0;
1098 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
1099 if (ddt_object_exists(ddt, type, class)) {
1100 ddt_object_sync(ddt, type, class, tx);
1101 VERIFY(ddt_object_count(ddt, type, class,
1102 &add) == 0);
1103 count += add;
1106 for (enum ddt_class class = 0; class < DDT_CLASSES; class++) {
1107 if (count == 0 && ddt_object_exists(ddt, type, class))
1108 ddt_object_destroy(ddt, type, class, tx);
1112 memcpy(&ddt->ddt_histogram_cache, ddt->ddt_histogram,
1113 sizeof (ddt->ddt_histogram));
1114 spa->spa_dedup_dspace = ~0ULL;
1117 void
1118 ddt_sync(spa_t *spa, uint64_t txg)
1120 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1121 dmu_tx_t *tx;
1122 zio_t *rio;
1124 ASSERT(spa_syncing_txg(spa) == txg);
1126 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1128 rio = zio_root(spa, NULL, NULL,
1129 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SELF_HEAL);
1132 * This function may cause an immediate scan of ddt blocks (see
1133 * the comment above dsl_scan_ddt() for details). We set the
1134 * scan's root zio here so that we can wait for any scan IOs in
1135 * addition to the regular ddt IOs.
1137 ASSERT3P(scn->scn_zio_root, ==, NULL);
1138 scn->scn_zio_root = rio;
1140 for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
1141 ddt_t *ddt = spa->spa_ddt[c];
1142 if (ddt == NULL)
1143 continue;
1144 ddt_sync_table(ddt, tx, txg);
1145 ddt_repair_table(ddt, rio);
1148 (void) zio_wait(rio);
1149 scn->scn_zio_root = NULL;
1151 dmu_tx_commit(tx);
1155 ddt_walk(spa_t *spa, ddt_bookmark_t *ddb, ddt_entry_t *dde)
1157 do {
1158 do {
1159 do {
1160 ddt_t *ddt = spa->spa_ddt[ddb->ddb_checksum];
1161 int error = ENOENT;
1162 if (ddt_object_exists(ddt, ddb->ddb_type,
1163 ddb->ddb_class)) {
1164 error = ddt_object_walk(ddt,
1165 ddb->ddb_type, ddb->ddb_class,
1166 &ddb->ddb_cursor, dde);
1168 dde->dde_type = ddb->ddb_type;
1169 dde->dde_class = ddb->ddb_class;
1170 if (error == 0)
1171 return (0);
1172 if (error != ENOENT)
1173 return (error);
1174 ddb->ddb_cursor = 0;
1175 } while (++ddb->ddb_checksum < ZIO_CHECKSUM_FUNCTIONS);
1176 ddb->ddb_checksum = 0;
1177 } while (++ddb->ddb_type < DDT_TYPES);
1178 ddb->ddb_type = 0;
1179 } while (++ddb->ddb_class < DDT_CLASSES);
1181 return (SET_ERROR(ENOENT));
1185 * This function is used by Block Cloning (brt.c) to increase reference
1186 * counter for the DDT entry if the block is already in DDT.
1188 * Return false if the block, despite having the D bit set, is not present
1189 * in the DDT. Currently this is not possible but might be in the future.
1190 * See the comment below.
1192 boolean_t
1193 ddt_addref(spa_t *spa, const blkptr_t *bp)
1195 ddt_t *ddt;
1196 ddt_entry_t *dde;
1197 boolean_t result;
1199 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1200 ddt = ddt_select(spa, bp);
1201 ddt_enter(ddt);
1203 dde = ddt_lookup(ddt, bp, B_TRUE);
1204 ASSERT(dde != NULL);
1206 if (dde->dde_type < DDT_TYPES) {
1207 ddt_phys_t *ddp;
1209 ASSERT3S(dde->dde_class, <, DDT_CLASSES);
1211 ddp = &dde->dde_phys[BP_GET_NDVAS(bp)];
1214 * This entry already existed (dde_type is real), so it must
1215 * have refcnt >0 at the start of this txg. We are called from
1216 * brt_pending_apply(), before frees are issued, so the refcnt
1217 * can't be lowered yet. Therefore, it must be >0. We assert
1218 * this because if the order of BRT and DDT interactions were
1219 * ever to change and the refcnt was ever zero here, then
1220 * likely further action is required to fill out the DDT entry,
1221 * and this is a place that is likely to be missed in testing.
1223 ASSERT3U(ddp->ddp_refcnt, >, 0);
1225 ddt_phys_addref(ddp);
1226 result = B_TRUE;
1227 } else {
1229 * At the time of implementating this if the block has the
1230 * DEDUP flag set it must exist in the DEDUP table, but
1231 * there are many advocates that want ability to remove
1232 * entries from DDT with refcnt=1. If this will happen,
1233 * we may have a block with the DEDUP set, but which doesn't
1234 * have a corresponding entry in the DDT. Be ready.
1236 ASSERT3S(dde->dde_class, ==, DDT_CLASSES);
1237 ddt_remove(ddt, dde);
1238 result = B_FALSE;
1241 ddt_exit(ddt);
1242 spa_config_exit(spa, SCL_ZIO, FTAG);
1244 return (result);
1247 ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, prefetch, INT, ZMOD_RW,
1248 "Enable prefetching dedup-ed blks");