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, 2016 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 static kmem_cache_t
*ddt_cache
;
42 static kmem_cache_t
*ddt_entry_cache
;
45 * Enable/disable prefetching of dedup-ed blocks which are going to be freed.
47 int zfs_dedup_prefetch
= 0;
49 static const ddt_ops_t
*ddt_ops
[DDT_TYPES
] = {
53 static const char *ddt_class_name
[DDT_CLASSES
] = {
60 ddt_object_create(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
63 spa_t
*spa
= ddt
->ddt_spa
;
64 objset_t
*os
= ddt
->ddt_os
;
65 uint64_t *objectp
= &ddt
->ddt_object
[type
][class];
66 boolean_t prehash
= zio_checksum_table
[ddt
->ddt_checksum
].ci_flags
&
68 char name
[DDT_NAMELEN
];
70 ddt_object_name(ddt
, type
, class, name
);
72 ASSERT(*objectp
== 0);
73 VERIFY(ddt_ops
[type
]->ddt_op_create(os
, objectp
, tx
, prehash
) == 0);
74 ASSERT(*objectp
!= 0);
76 VERIFY(zap_add(os
, DMU_POOL_DIRECTORY_OBJECT
, name
,
77 sizeof (uint64_t), 1, objectp
, tx
) == 0);
79 VERIFY(zap_add(os
, spa
->spa_ddt_stat_object
, name
,
80 sizeof (uint64_t), sizeof (ddt_histogram_t
) / sizeof (uint64_t),
81 &ddt
->ddt_histogram
[type
][class], tx
) == 0);
85 ddt_object_destroy(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
88 spa_t
*spa
= ddt
->ddt_spa
;
89 objset_t
*os
= ddt
->ddt_os
;
90 uint64_t *objectp
= &ddt
->ddt_object
[type
][class];
92 char name
[DDT_NAMELEN
];
94 ddt_object_name(ddt
, type
, class, name
);
96 ASSERT(*objectp
!= 0);
97 ASSERT(ddt_histogram_empty(&ddt
->ddt_histogram
[type
][class]));
98 VERIFY(ddt_object_count(ddt
, type
, class, &count
) == 0 && count
== 0);
99 VERIFY(zap_remove(os
, DMU_POOL_DIRECTORY_OBJECT
, name
, tx
) == 0);
100 VERIFY(zap_remove(os
, spa
->spa_ddt_stat_object
, name
, tx
) == 0);
101 VERIFY(ddt_ops
[type
]->ddt_op_destroy(os
, *objectp
, tx
) == 0);
102 bzero(&ddt
->ddt_object_stats
[type
][class], sizeof (ddt_object_t
));
108 ddt_object_load(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class)
110 ddt_object_t
*ddo
= &ddt
->ddt_object_stats
[type
][class];
111 dmu_object_info_t doi
;
113 char name
[DDT_NAMELEN
];
116 ddt_object_name(ddt
, type
, class, name
);
118 error
= zap_lookup(ddt
->ddt_os
, DMU_POOL_DIRECTORY_OBJECT
, name
,
119 sizeof (uint64_t), 1, &ddt
->ddt_object
[type
][class]);
123 error
= zap_lookup(ddt
->ddt_os
, ddt
->ddt_spa
->spa_ddt_stat_object
, name
,
124 sizeof (uint64_t), sizeof (ddt_histogram_t
) / sizeof (uint64_t),
125 &ddt
->ddt_histogram
[type
][class]);
130 * Seed the cached statistics.
132 error
= ddt_object_info(ddt
, type
, class, &doi
);
136 error
= ddt_object_count(ddt
, type
, class, &count
);
140 ddo
->ddo_count
= count
;
141 ddo
->ddo_dspace
= doi
.doi_physical_blocks_512
<< 9;
142 ddo
->ddo_mspace
= doi
.doi_fill_count
* doi
.doi_data_block_size
;
148 ddt_object_sync(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
151 ddt_object_t
*ddo
= &ddt
->ddt_object_stats
[type
][class];
152 dmu_object_info_t doi
;
154 char name
[DDT_NAMELEN
];
156 ddt_object_name(ddt
, type
, class, name
);
158 VERIFY(zap_update(ddt
->ddt_os
, ddt
->ddt_spa
->spa_ddt_stat_object
, name
,
159 sizeof (uint64_t), sizeof (ddt_histogram_t
) / sizeof (uint64_t),
160 &ddt
->ddt_histogram
[type
][class], tx
) == 0);
163 * Cache DDT statistics; this is the only time they'll change.
165 VERIFY(ddt_object_info(ddt
, type
, class, &doi
) == 0);
166 VERIFY(ddt_object_count(ddt
, type
, class, &count
) == 0);
168 ddo
->ddo_count
= count
;
169 ddo
->ddo_dspace
= doi
.doi_physical_blocks_512
<< 9;
170 ddo
->ddo_mspace
= doi
.doi_fill_count
* doi
.doi_data_block_size
;
174 ddt_object_lookup(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
177 if (!ddt_object_exists(ddt
, type
, class))
178 return (SET_ERROR(ENOENT
));
180 return (ddt_ops
[type
]->ddt_op_lookup(ddt
->ddt_os
,
181 ddt
->ddt_object
[type
][class], dde
));
185 ddt_object_prefetch(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
188 if (!ddt_object_exists(ddt
, type
, class))
191 ddt_ops
[type
]->ddt_op_prefetch(ddt
->ddt_os
,
192 ddt
->ddt_object
[type
][class], dde
);
196 ddt_object_update(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
197 ddt_entry_t
*dde
, dmu_tx_t
*tx
)
199 ASSERT(ddt_object_exists(ddt
, type
, class));
201 return (ddt_ops
[type
]->ddt_op_update(ddt
->ddt_os
,
202 ddt
->ddt_object
[type
][class], dde
, tx
));
206 ddt_object_remove(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
207 ddt_entry_t
*dde
, dmu_tx_t
*tx
)
209 ASSERT(ddt_object_exists(ddt
, type
, class));
211 return (ddt_ops
[type
]->ddt_op_remove(ddt
->ddt_os
,
212 ddt
->ddt_object
[type
][class], dde
, tx
));
216 ddt_object_walk(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
217 uint64_t *walk
, ddt_entry_t
*dde
)
219 ASSERT(ddt_object_exists(ddt
, type
, class));
221 return (ddt_ops
[type
]->ddt_op_walk(ddt
->ddt_os
,
222 ddt
->ddt_object
[type
][class], dde
, walk
));
226 ddt_object_count(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
229 ASSERT(ddt_object_exists(ddt
, type
, class));
231 return (ddt_ops
[type
]->ddt_op_count(ddt
->ddt_os
,
232 ddt
->ddt_object
[type
][class], count
));
236 ddt_object_info(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
237 dmu_object_info_t
*doi
)
239 if (!ddt_object_exists(ddt
, type
, class))
240 return (SET_ERROR(ENOENT
));
242 return (dmu_object_info(ddt
->ddt_os
, ddt
->ddt_object
[type
][class],
247 ddt_object_exists(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class)
249 return (!!ddt
->ddt_object
[type
][class]);
253 ddt_object_name(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
256 (void) sprintf(name
, DMU_POOL_DDT
,
257 zio_checksum_table
[ddt
->ddt_checksum
].ci_name
,
258 ddt_ops
[type
]->ddt_op_name
, ddt_class_name
[class]);
262 ddt_bp_fill(const ddt_phys_t
*ddp
, blkptr_t
*bp
, uint64_t txg
)
267 for (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 ddt_bp_create(enum zio_checksum checksum
,
274 const ddt_key_t
*ddk
, const ddt_phys_t
*ddp
, blkptr_t
*bp
)
279 ddt_bp_fill(ddp
, bp
, ddp
->ddp_phys_birth
);
281 bp
->blk_cksum
= ddk
->ddk_cksum
;
284 BP_SET_LSIZE(bp
, DDK_GET_LSIZE(ddk
));
285 BP_SET_PSIZE(bp
, DDK_GET_PSIZE(ddk
));
286 BP_SET_COMPRESS(bp
, DDK_GET_COMPRESS(ddk
));
287 BP_SET_CHECKSUM(bp
, checksum
);
288 BP_SET_TYPE(bp
, DMU_OT_DEDUP
);
291 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
295 ddt_key_fill(ddt_key_t
*ddk
, const blkptr_t
*bp
)
297 ddk
->ddk_cksum
= bp
->blk_cksum
;
300 DDK_SET_LSIZE(ddk
, BP_GET_LSIZE(bp
));
301 DDK_SET_PSIZE(ddk
, BP_GET_PSIZE(bp
));
302 DDK_SET_COMPRESS(ddk
, BP_GET_COMPRESS(bp
));
306 ddt_phys_fill(ddt_phys_t
*ddp
, const blkptr_t
*bp
)
309 ASSERT(ddp
->ddp_phys_birth
== 0);
311 for (d
= 0; d
< SPA_DVAS_PER_BP
; d
++)
312 ddp
->ddp_dva
[d
] = bp
->blk_dva
[d
];
313 ddp
->ddp_phys_birth
= BP_PHYSICAL_BIRTH(bp
);
317 ddt_phys_clear(ddt_phys_t
*ddp
)
319 bzero(ddp
, sizeof (*ddp
));
323 ddt_phys_addref(ddt_phys_t
*ddp
)
329 ddt_phys_decref(ddt_phys_t
*ddp
)
332 ASSERT(ddp
->ddp_refcnt
> 0);
338 ddt_phys_free(ddt_t
*ddt
, ddt_key_t
*ddk
, ddt_phys_t
*ddp
, uint64_t txg
)
342 ddt_bp_create(ddt
->ddt_checksum
, ddk
, ddp
, &blk
);
344 zio_free(ddt
->ddt_spa
, txg
, &blk
);
348 ddt_phys_select(const ddt_entry_t
*dde
, const blkptr_t
*bp
)
350 ddt_phys_t
*ddp
= (ddt_phys_t
*)dde
->dde_phys
;
353 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
354 if (DVA_EQUAL(BP_IDENTITY(bp
), &ddp
->ddp_dva
[0]) &&
355 BP_PHYSICAL_BIRTH(bp
) == ddp
->ddp_phys_birth
)
362 ddt_phys_total_refcnt(const ddt_entry_t
*dde
)
367 for (p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++)
368 refcnt
+= dde
->dde_phys
[p
].ddp_refcnt
;
374 ddt_stat_generate(ddt_t
*ddt
, ddt_entry_t
*dde
, ddt_stat_t
*dds
)
376 spa_t
*spa
= ddt
->ddt_spa
;
377 ddt_phys_t
*ddp
= dde
->dde_phys
;
378 ddt_key_t
*ddk
= &dde
->dde_key
;
379 uint64_t lsize
= DDK_GET_LSIZE(ddk
);
380 uint64_t psize
= DDK_GET_PSIZE(ddk
);
383 bzero(dds
, sizeof (*dds
));
385 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
387 uint64_t refcnt
= ddp
->ddp_refcnt
;
389 if (ddp
->ddp_phys_birth
== 0)
392 for (d
= 0; d
< SPA_DVAS_PER_BP
; d
++)
393 dsize
+= dva_get_dsize_sync(spa
, &ddp
->ddp_dva
[d
]);
395 dds
->dds_blocks
+= 1;
396 dds
->dds_lsize
+= lsize
;
397 dds
->dds_psize
+= psize
;
398 dds
->dds_dsize
+= dsize
;
400 dds
->dds_ref_blocks
+= refcnt
;
401 dds
->dds_ref_lsize
+= lsize
* refcnt
;
402 dds
->dds_ref_psize
+= psize
* refcnt
;
403 dds
->dds_ref_dsize
+= dsize
* refcnt
;
408 ddt_stat_add(ddt_stat_t
*dst
, const ddt_stat_t
*src
, uint64_t neg
)
410 const uint64_t *s
= (const uint64_t *)src
;
411 uint64_t *d
= (uint64_t *)dst
;
412 uint64_t *d_end
= (uint64_t *)(dst
+ 1);
414 ASSERT(neg
== 0 || neg
== -1ULL); /* add or subtract */
417 *d
++ += (*s
++ ^ neg
) - neg
;
421 ddt_stat_update(ddt_t
*ddt
, ddt_entry_t
*dde
, uint64_t neg
)
424 ddt_histogram_t
*ddh
;
427 ddt_stat_generate(ddt
, dde
, &dds
);
429 bucket
= highbit64(dds
.dds_ref_blocks
) - 1;
432 ddh
= &ddt
->ddt_histogram
[dde
->dde_type
][dde
->dde_class
];
434 ddt_stat_add(&ddh
->ddh_stat
[bucket
], &dds
, neg
);
438 ddt_histogram_add(ddt_histogram_t
*dst
, const ddt_histogram_t
*src
)
442 for (h
= 0; h
< 64; h
++)
443 ddt_stat_add(&dst
->ddh_stat
[h
], &src
->ddh_stat
[h
], 0);
447 ddt_histogram_stat(ddt_stat_t
*dds
, const ddt_histogram_t
*ddh
)
451 bzero(dds
, sizeof (*dds
));
453 for (h
= 0; h
< 64; h
++)
454 ddt_stat_add(dds
, &ddh
->ddh_stat
[h
], 0);
458 ddt_histogram_empty(const ddt_histogram_t
*ddh
)
460 const uint64_t *s
= (const uint64_t *)ddh
;
461 const uint64_t *s_end
= (const uint64_t *)(ddh
+ 1);
471 ddt_get_dedup_object_stats(spa_t
*spa
, ddt_object_t
*ddo_total
)
475 enum ddt_class
class;
477 /* Sum the statistics we cached in ddt_object_sync(). */
478 for (c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
479 ddt_t
*ddt
= spa
->spa_ddt
[c
];
480 for (type
= 0; type
< DDT_TYPES
; type
++) {
481 for (class = 0; class < DDT_CLASSES
;
484 &ddt
->ddt_object_stats
[type
][class];
485 ddo_total
->ddo_count
+= ddo
->ddo_count
;
486 ddo_total
->ddo_dspace
+= ddo
->ddo_dspace
;
487 ddo_total
->ddo_mspace
+= ddo
->ddo_mspace
;
492 /* ... and compute the averages. */
493 if (ddo_total
->ddo_count
!= 0) {
494 ddo_total
->ddo_dspace
/= ddo_total
->ddo_count
;
495 ddo_total
->ddo_mspace
/= ddo_total
->ddo_count
;
500 ddt_get_dedup_histogram(spa_t
*spa
, ddt_histogram_t
*ddh
)
504 enum ddt_class
class;
506 for (c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
507 ddt_t
*ddt
= spa
->spa_ddt
[c
];
508 for (type
= 0; type
< DDT_TYPES
; type
++) {
509 for (class = 0; class < DDT_CLASSES
;
511 ddt_histogram_add(ddh
,
512 &ddt
->ddt_histogram_cache
[type
][class]);
519 ddt_get_dedup_stats(spa_t
*spa
, ddt_stat_t
*dds_total
)
521 ddt_histogram_t
*ddh_total
;
523 ddh_total
= kmem_zalloc(sizeof (ddt_histogram_t
), KM_SLEEP
);
524 ddt_get_dedup_histogram(spa
, ddh_total
);
525 ddt_histogram_stat(dds_total
, ddh_total
);
526 kmem_free(ddh_total
, sizeof (ddt_histogram_t
));
530 ddt_get_dedup_dspace(spa_t
*spa
)
532 ddt_stat_t dds_total
;
534 if (spa
->spa_dedup_dspace
!= ~0ULL)
535 return (spa
->spa_dedup_dspace
);
537 bzero(&dds_total
, sizeof (ddt_stat_t
));
539 /* Calculate and cache the stats */
540 ddt_get_dedup_stats(spa
, &dds_total
);
541 spa
->spa_dedup_dspace
= dds_total
.dds_ref_dsize
- dds_total
.dds_dsize
;
542 return (spa
->spa_dedup_dspace
);
546 ddt_get_pool_dedup_ratio(spa_t
*spa
)
548 ddt_stat_t dds_total
= { 0 };
550 ddt_get_dedup_stats(spa
, &dds_total
);
551 if (dds_total
.dds_dsize
== 0)
554 return (dds_total
.dds_ref_dsize
* 100 / dds_total
.dds_dsize
);
558 ddt_ditto_copies_needed(ddt_t
*ddt
, ddt_entry_t
*dde
, ddt_phys_t
*ddp_willref
)
560 spa_t
*spa
= ddt
->ddt_spa
;
561 uint64_t total_refcnt
= 0;
562 uint64_t ditto
= spa
->spa_dedup_ditto
;
563 int total_copies
= 0;
564 int desired_copies
= 0;
567 for (p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
568 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
569 zio_t
*zio
= dde
->dde_lead_zio
[p
];
570 uint64_t refcnt
= ddp
->ddp_refcnt
; /* committed refs */
572 refcnt
+= zio
->io_parent_count
; /* pending refs */
573 if (ddp
== ddp_willref
)
574 refcnt
++; /* caller's ref */
576 total_refcnt
+= refcnt
;
581 if (ditto
== 0 || ditto
> UINT32_MAX
)
584 if (total_refcnt
>= 1)
586 if (total_refcnt
>= ditto
)
588 if (total_refcnt
>= ditto
* ditto
)
591 return (MAX(desired_copies
, total_copies
) - total_copies
);
595 ddt_ditto_copies_present(ddt_entry_t
*dde
)
597 ddt_phys_t
*ddp
= &dde
->dde_phys
[DDT_PHYS_DITTO
];
598 dva_t
*dva
= ddp
->ddp_dva
;
599 int copies
= 0 - DVA_GET_GANG(dva
);
602 for (d
= 0; d
< SPA_DVAS_PER_BP
; d
++, dva
++)
603 if (DVA_IS_VALID(dva
))
606 ASSERT(copies
>= 0 && copies
< SPA_DVAS_PER_BP
);
612 ddt_compress(void *src
, uchar_t
*dst
, size_t s_len
, size_t d_len
)
614 uchar_t
*version
= dst
++;
615 int cpfunc
= ZIO_COMPRESS_ZLE
;
616 zio_compress_info_t
*ci
= &zio_compress_table
[cpfunc
];
619 ASSERT(d_len
>= s_len
+ 1); /* no compression plus version byte */
621 c_len
= ci
->ci_compress(src
, dst
, s_len
, d_len
- 1, ci
->ci_level
);
623 if (c_len
== s_len
) {
624 cpfunc
= ZIO_COMPRESS_OFF
;
625 bcopy(src
, dst
, s_len
);
630 if (ZFS_HOST_BYTEORDER
)
631 *version
|= DDT_COMPRESS_BYTEORDER_MASK
;
637 ddt_decompress(uchar_t
*src
, void *dst
, size_t s_len
, size_t d_len
)
639 uchar_t version
= *src
++;
640 int cpfunc
= version
& DDT_COMPRESS_FUNCTION_MASK
;
641 zio_compress_info_t
*ci
= &zio_compress_table
[cpfunc
];
643 if (ci
->ci_decompress
!= NULL
)
644 (void) ci
->ci_decompress(src
, dst
, s_len
, d_len
, ci
->ci_level
);
646 bcopy(src
, dst
, d_len
);
648 if (((version
& DDT_COMPRESS_BYTEORDER_MASK
) != 0) !=
649 (ZFS_HOST_BYTEORDER
!= 0))
650 byteswap_uint64_array(dst
, d_len
);
654 ddt_select_by_checksum(spa_t
*spa
, enum zio_checksum c
)
656 return (spa
->spa_ddt
[c
]);
660 ddt_select(spa_t
*spa
, const blkptr_t
*bp
)
662 return (spa
->spa_ddt
[BP_GET_CHECKSUM(bp
)]);
666 ddt_enter(ddt_t
*ddt
)
668 mutex_enter(&ddt
->ddt_lock
);
674 mutex_exit(&ddt
->ddt_lock
);
680 ddt_cache
= kmem_cache_create("ddt_cache",
681 sizeof (ddt_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
682 ddt_entry_cache
= kmem_cache_create("ddt_entry_cache",
683 sizeof (ddt_entry_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
689 kmem_cache_destroy(ddt_entry_cache
);
690 kmem_cache_destroy(ddt_cache
);
694 ddt_alloc(const ddt_key_t
*ddk
)
698 dde
= kmem_cache_alloc(ddt_entry_cache
, KM_SLEEP
);
699 bzero(dde
, sizeof (ddt_entry_t
));
700 cv_init(&dde
->dde_cv
, NULL
, CV_DEFAULT
, NULL
);
708 ddt_free(ddt_entry_t
*dde
)
712 ASSERT(!dde
->dde_loading
);
714 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++)
715 ASSERT(dde
->dde_lead_zio
[p
] == NULL
);
717 if (dde
->dde_repair_abd
!= NULL
)
718 abd_free(dde
->dde_repair_abd
);
720 cv_destroy(&dde
->dde_cv
);
721 kmem_cache_free(ddt_entry_cache
, dde
);
725 ddt_remove(ddt_t
*ddt
, ddt_entry_t
*dde
)
727 ASSERT(MUTEX_HELD(&ddt
->ddt_lock
));
729 avl_remove(&ddt
->ddt_tree
, dde
);
734 ddt_lookup(ddt_t
*ddt
, const blkptr_t
*bp
, boolean_t add
)
736 ddt_entry_t
*dde
, dde_search
;
738 enum ddt_class
class;
742 ASSERT(MUTEX_HELD(&ddt
->ddt_lock
));
744 ddt_key_fill(&dde_search
.dde_key
, bp
);
746 dde
= avl_find(&ddt
->ddt_tree
, &dde_search
, &where
);
750 dde
= ddt_alloc(&dde_search
.dde_key
);
751 avl_insert(&ddt
->ddt_tree
, dde
, where
);
754 while (dde
->dde_loading
)
755 cv_wait(&dde
->dde_cv
, &ddt
->ddt_lock
);
760 dde
->dde_loading
= B_TRUE
;
766 for (type
= 0; type
< DDT_TYPES
; type
++) {
767 for (class = 0; class < DDT_CLASSES
; class++) {
768 error
= ddt_object_lookup(ddt
, type
, class, dde
);
776 ASSERT(error
== 0 || error
== ENOENT
);
780 ASSERT(dde
->dde_loaded
== B_FALSE
);
781 ASSERT(dde
->dde_loading
== B_TRUE
);
783 dde
->dde_type
= type
; /* will be DDT_TYPES if no entry found */
784 dde
->dde_class
= class; /* will be DDT_CLASSES if no entry found */
785 dde
->dde_loaded
= B_TRUE
;
786 dde
->dde_loading
= B_FALSE
;
789 ddt_stat_update(ddt
, dde
, -1ULL);
791 cv_broadcast(&dde
->dde_cv
);
797 ddt_prefetch(spa_t
*spa
, const blkptr_t
*bp
)
802 enum ddt_class
class;
804 if (!zfs_dedup_prefetch
|| bp
== NULL
|| !BP_GET_DEDUP(bp
))
808 * We only remove the DDT once all tables are empty and only
809 * prefetch dedup blocks when there are entries in the DDT.
810 * Thus no locking is required as the DDT can't disappear on us.
812 ddt
= ddt_select(spa
, bp
);
813 ddt_key_fill(&dde
.dde_key
, bp
);
815 for (type
= 0; type
< DDT_TYPES
; type
++) {
816 for (class = 0; class < DDT_CLASSES
; class++) {
817 ddt_object_prefetch(ddt
, type
, class, &dde
);
823 * Opaque struct used for ddt_key comparison
825 #define DDT_KEY_CMP_LEN (sizeof (ddt_key_t) / sizeof (uint16_t))
827 typedef struct ddt_key_cmp
{
828 uint16_t u16
[DDT_KEY_CMP_LEN
];
832 ddt_entry_compare(const void *x1
, const void *x2
)
834 const ddt_entry_t
*dde1
= x1
;
835 const ddt_entry_t
*dde2
= x2
;
836 const ddt_key_cmp_t
*k1
= (const ddt_key_cmp_t
*)&dde1
->dde_key
;
837 const ddt_key_cmp_t
*k2
= (const ddt_key_cmp_t
*)&dde2
->dde_key
;
841 for (i
= 0; i
< DDT_KEY_CMP_LEN
; i
++) {
842 cmp
= (int32_t)k1
->u16
[i
] - (int32_t)k2
->u16
[i
];
847 return (AVL_ISIGN(cmp
));
851 ddt_table_alloc(spa_t
*spa
, enum zio_checksum c
)
855 ddt
= kmem_cache_alloc(ddt_cache
, KM_SLEEP
);
856 bzero(ddt
, sizeof (ddt_t
));
858 mutex_init(&ddt
->ddt_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
859 avl_create(&ddt
->ddt_tree
, ddt_entry_compare
,
860 sizeof (ddt_entry_t
), offsetof(ddt_entry_t
, dde_node
));
861 avl_create(&ddt
->ddt_repair_tree
, ddt_entry_compare
,
862 sizeof (ddt_entry_t
), offsetof(ddt_entry_t
, dde_node
));
863 ddt
->ddt_checksum
= c
;
865 ddt
->ddt_os
= spa
->spa_meta_objset
;
871 ddt_table_free(ddt_t
*ddt
)
873 ASSERT(avl_numnodes(&ddt
->ddt_tree
) == 0);
874 ASSERT(avl_numnodes(&ddt
->ddt_repair_tree
) == 0);
875 avl_destroy(&ddt
->ddt_tree
);
876 avl_destroy(&ddt
->ddt_repair_tree
);
877 mutex_destroy(&ddt
->ddt_lock
);
878 kmem_cache_free(ddt_cache
, ddt
);
882 ddt_create(spa_t
*spa
)
886 spa
->spa_dedup_checksum
= ZIO_DEDUPCHECKSUM
;
888 for (c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++)
889 spa
->spa_ddt
[c
] = ddt_table_alloc(spa
, c
);
897 enum ddt_class
class;
902 error
= zap_lookup(spa
->spa_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
903 DMU_POOL_DDT_STATS
, sizeof (uint64_t), 1,
904 &spa
->spa_ddt_stat_object
);
907 return (error
== ENOENT
? 0 : error
);
909 for (c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
910 ddt_t
*ddt
= spa
->spa_ddt
[c
];
911 for (type
= 0; type
< DDT_TYPES
; type
++) {
912 for (class = 0; class < DDT_CLASSES
;
914 error
= ddt_object_load(ddt
, type
, class);
915 if (error
!= 0 && error
!= ENOENT
)
921 * Seed the cached histograms.
923 bcopy(ddt
->ddt_histogram
, &ddt
->ddt_histogram_cache
,
924 sizeof (ddt
->ddt_histogram
));
925 spa
->spa_dedup_dspace
= ~0ULL;
932 ddt_unload(spa_t
*spa
)
936 for (c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
937 if (spa
->spa_ddt
[c
]) {
938 ddt_table_free(spa
->spa_ddt
[c
]);
939 spa
->spa_ddt
[c
] = NULL
;
945 ddt_class_contains(spa_t
*spa
, enum ddt_class max_class
, const blkptr_t
*bp
)
950 enum ddt_class
class;
952 if (!BP_GET_DEDUP(bp
))
955 if (max_class
== DDT_CLASS_UNIQUE
)
958 ddt
= spa
->spa_ddt
[BP_GET_CHECKSUM(bp
)];
959 dde
= kmem_cache_alloc(ddt_entry_cache
, KM_SLEEP
);
961 ddt_key_fill(&(dde
->dde_key
), bp
);
963 for (type
= 0; type
< DDT_TYPES
; type
++) {
964 for (class = 0; class <= max_class
; class++) {
965 if (ddt_object_lookup(ddt
, type
, class, dde
) == 0) {
966 kmem_cache_free(ddt_entry_cache
, dde
);
972 kmem_cache_free(ddt_entry_cache
, dde
);
977 ddt_repair_start(ddt_t
*ddt
, const blkptr_t
*bp
)
982 enum ddt_class
class;
984 ddt_key_fill(&ddk
, bp
);
986 dde
= ddt_alloc(&ddk
);
988 for (type
= 0; type
< DDT_TYPES
; type
++) {
989 for (class = 0; class < DDT_CLASSES
; class++) {
991 * We can only do repair if there are multiple copies
992 * of the block. For anything in the UNIQUE class,
993 * there's definitely only one copy, so don't even try.
995 if (class != DDT_CLASS_UNIQUE
&&
996 ddt_object_lookup(ddt
, type
, class, dde
) == 0)
1001 bzero(dde
->dde_phys
, sizeof (dde
->dde_phys
));
1007 ddt_repair_done(ddt_t
*ddt
, ddt_entry_t
*dde
)
1013 if (dde
->dde_repair_abd
!= NULL
&& spa_writeable(ddt
->ddt_spa
) &&
1014 avl_find(&ddt
->ddt_repair_tree
, dde
, &where
) == NULL
)
1015 avl_insert(&ddt
->ddt_repair_tree
, dde
, where
);
1023 ddt_repair_entry_done(zio_t
*zio
)
1025 ddt_entry_t
*rdde
= zio
->io_private
;
1031 ddt_repair_entry(ddt_t
*ddt
, ddt_entry_t
*dde
, ddt_entry_t
*rdde
, zio_t
*rio
)
1033 ddt_phys_t
*ddp
= dde
->dde_phys
;
1034 ddt_phys_t
*rddp
= rdde
->dde_phys
;
1035 ddt_key_t
*ddk
= &dde
->dde_key
;
1036 ddt_key_t
*rddk
= &rdde
->dde_key
;
1041 zio
= zio_null(rio
, rio
->io_spa
, NULL
,
1042 ddt_repair_entry_done
, rdde
, rio
->io_flags
);
1044 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++, rddp
++) {
1045 if (ddp
->ddp_phys_birth
== 0 ||
1046 ddp
->ddp_phys_birth
!= rddp
->ddp_phys_birth
||
1047 bcmp(ddp
->ddp_dva
, rddp
->ddp_dva
, sizeof (ddp
->ddp_dva
)))
1049 ddt_bp_create(ddt
->ddt_checksum
, ddk
, ddp
, &blk
);
1050 zio_nowait(zio_rewrite(zio
, zio
->io_spa
, 0, &blk
,
1051 rdde
->dde_repair_abd
, DDK_GET_PSIZE(rddk
), NULL
, NULL
,
1052 ZIO_PRIORITY_SYNC_WRITE
, ZIO_DDT_CHILD_FLAGS(zio
), NULL
));
1059 ddt_repair_table(ddt_t
*ddt
, zio_t
*rio
)
1061 spa_t
*spa
= ddt
->ddt_spa
;
1062 ddt_entry_t
*dde
, *rdde_next
, *rdde
;
1063 avl_tree_t
*t
= &ddt
->ddt_repair_tree
;
1066 if (spa_sync_pass(spa
) > 1)
1070 for (rdde
= avl_first(t
); rdde
!= NULL
; rdde
= rdde_next
) {
1071 rdde_next
= AVL_NEXT(t
, rdde
);
1072 avl_remove(&ddt
->ddt_repair_tree
, rdde
);
1074 ddt_bp_create(ddt
->ddt_checksum
, &rdde
->dde_key
, NULL
, &blk
);
1075 dde
= ddt_repair_start(ddt
, &blk
);
1076 ddt_repair_entry(ddt
, dde
, rdde
, rio
);
1077 ddt_repair_done(ddt
, dde
);
1084 ddt_sync_entry(ddt_t
*ddt
, ddt_entry_t
*dde
, dmu_tx_t
*tx
, uint64_t txg
)
1086 dsl_pool_t
*dp
= ddt
->ddt_spa
->spa_dsl_pool
;
1087 ddt_phys_t
*ddp
= dde
->dde_phys
;
1088 ddt_key_t
*ddk
= &dde
->dde_key
;
1089 enum ddt_type otype
= dde
->dde_type
;
1090 enum ddt_type ntype
= DDT_TYPE_CURRENT
;
1091 enum ddt_class oclass
= dde
->dde_class
;
1092 enum ddt_class nclass
;
1093 uint64_t total_refcnt
= 0;
1096 ASSERT(dde
->dde_loaded
);
1097 ASSERT(!dde
->dde_loading
);
1099 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
1100 ASSERT(dde
->dde_lead_zio
[p
] == NULL
);
1101 if (ddp
->ddp_phys_birth
== 0) {
1102 ASSERT(ddp
->ddp_refcnt
== 0);
1105 if (p
== DDT_PHYS_DITTO
) {
1106 if (ddt_ditto_copies_needed(ddt
, dde
, NULL
) == 0)
1107 ddt_phys_free(ddt
, ddk
, ddp
, txg
);
1110 if (ddp
->ddp_refcnt
== 0)
1111 ddt_phys_free(ddt
, ddk
, ddp
, txg
);
1112 total_refcnt
+= ddp
->ddp_refcnt
;
1115 if (dde
->dde_phys
[DDT_PHYS_DITTO
].ddp_phys_birth
!= 0)
1116 nclass
= DDT_CLASS_DITTO
;
1117 else if (total_refcnt
> 1)
1118 nclass
= DDT_CLASS_DUPLICATE
;
1120 nclass
= DDT_CLASS_UNIQUE
;
1122 if (otype
!= DDT_TYPES
&&
1123 (otype
!= ntype
|| oclass
!= nclass
|| total_refcnt
== 0)) {
1124 VERIFY(ddt_object_remove(ddt
, otype
, oclass
, dde
, tx
) == 0);
1125 ASSERT(ddt_object_lookup(ddt
, otype
, oclass
, dde
) == ENOENT
);
1128 if (total_refcnt
!= 0) {
1129 dde
->dde_type
= ntype
;
1130 dde
->dde_class
= nclass
;
1131 ddt_stat_update(ddt
, dde
, 0);
1132 if (!ddt_object_exists(ddt
, ntype
, nclass
))
1133 ddt_object_create(ddt
, ntype
, nclass
, tx
);
1134 VERIFY(ddt_object_update(ddt
, ntype
, nclass
, dde
, tx
) == 0);
1137 * If the class changes, the order that we scan this bp
1138 * changes. If it decreases, we could miss it, so
1139 * scan it right now. (This covers both class changing
1140 * while we are doing ddt_walk(), and when we are
1143 if (nclass
< oclass
) {
1144 dsl_scan_ddt_entry(dp
->dp_scan
,
1145 ddt
->ddt_checksum
, dde
, tx
);
1151 ddt_sync_table(ddt_t
*ddt
, dmu_tx_t
*tx
, uint64_t txg
)
1153 spa_t
*spa
= ddt
->ddt_spa
;
1155 void *cookie
= NULL
;
1157 enum ddt_class
class;
1159 if (avl_numnodes(&ddt
->ddt_tree
) == 0)
1162 ASSERT(spa
->spa_uberblock
.ub_version
>= SPA_VERSION_DEDUP
);
1164 if (spa
->spa_ddt_stat_object
== 0) {
1165 spa
->spa_ddt_stat_object
= zap_create_link(ddt
->ddt_os
,
1166 DMU_OT_DDT_STATS
, DMU_POOL_DIRECTORY_OBJECT
,
1167 DMU_POOL_DDT_STATS
, tx
);
1170 while ((dde
= avl_destroy_nodes(&ddt
->ddt_tree
, &cookie
)) != NULL
) {
1171 ddt_sync_entry(ddt
, dde
, tx
, txg
);
1175 for (type
= 0; type
< DDT_TYPES
; type
++) {
1176 uint64_t add
, count
= 0;
1177 for (class = 0; class < DDT_CLASSES
; class++) {
1178 if (ddt_object_exists(ddt
, type
, class)) {
1179 ddt_object_sync(ddt
, type
, class, tx
);
1180 VERIFY(ddt_object_count(ddt
, type
, class,
1185 for (class = 0; class < DDT_CLASSES
; class++) {
1186 if (count
== 0 && ddt_object_exists(ddt
, type
, class))
1187 ddt_object_destroy(ddt
, type
, class, tx
);
1191 bcopy(ddt
->ddt_histogram
, &ddt
->ddt_histogram_cache
,
1192 sizeof (ddt
->ddt_histogram
));
1193 spa
->spa_dedup_dspace
= ~0ULL;
1197 ddt_sync(spa_t
*spa
, uint64_t txg
)
1200 zio_t
*rio
= zio_root(spa
, NULL
, NULL
,
1201 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
);
1202 enum zio_checksum c
;
1204 ASSERT(spa_syncing_txg(spa
) == txg
);
1206 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1208 for (c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
1209 ddt_t
*ddt
= spa
->spa_ddt
[c
];
1212 ddt_sync_table(ddt
, tx
, txg
);
1213 ddt_repair_table(ddt
, rio
);
1216 (void) zio_wait(rio
);
1222 ddt_walk(spa_t
*spa
, ddt_bookmark_t
*ddb
, ddt_entry_t
*dde
)
1227 ddt_t
*ddt
= spa
->spa_ddt
[ddb
->ddb_checksum
];
1229 if (ddt_object_exists(ddt
, ddb
->ddb_type
,
1231 error
= ddt_object_walk(ddt
,
1232 ddb
->ddb_type
, ddb
->ddb_class
,
1233 &ddb
->ddb_cursor
, dde
);
1235 dde
->dde_type
= ddb
->ddb_type
;
1236 dde
->dde_class
= ddb
->ddb_class
;
1239 if (error
!= ENOENT
)
1241 ddb
->ddb_cursor
= 0;
1242 } while (++ddb
->ddb_checksum
< ZIO_CHECKSUM_FUNCTIONS
);
1243 ddb
->ddb_checksum
= 0;
1244 } while (++ddb
->ddb_type
< DDT_TYPES
);
1246 } while (++ddb
->ddb_class
< DDT_CLASSES
);
1248 return (SET_ERROR(ENOENT
));
1251 #if defined(_KERNEL) && defined(HAVE_SPL)
1252 module_param(zfs_dedup_prefetch
, int, 0644);
1253 MODULE_PARM_DESC(zfs_dedup_prefetch
, "Enable prefetching dedup-ed blks");