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>
39 static const ddt_ops_t
*ddt_ops
[DDT_TYPES
] = {
43 static const char *ddt_class_name
[DDT_CLASSES
] = {
50 ddt_object_create(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
53 spa_t
*spa
= ddt
->ddt_spa
;
54 objset_t
*os
= ddt
->ddt_os
;
55 uint64_t *objectp
= &ddt
->ddt_object
[type
][class];
56 boolean_t prehash
= zio_checksum_table
[ddt
->ddt_checksum
].ci_dedup
;
57 char name
[DDT_NAMELEN
];
59 ddt_object_name(ddt
, type
, class, name
);
61 ASSERT(*objectp
== 0);
62 VERIFY(ddt_ops
[type
]->ddt_op_create(os
, objectp
, tx
, prehash
) == 0);
63 ASSERT(*objectp
!= 0);
65 VERIFY(zap_add(os
, DMU_POOL_DIRECTORY_OBJECT
, name
,
66 sizeof (uint64_t), 1, objectp
, tx
) == 0);
68 VERIFY(zap_add(os
, spa
->spa_ddt_stat_object
, name
,
69 sizeof (uint64_t), sizeof (ddt_histogram_t
) / sizeof (uint64_t),
70 &ddt
->ddt_histogram
[type
][class], tx
) == 0);
74 ddt_object_destroy(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
77 spa_t
*spa
= ddt
->ddt_spa
;
78 objset_t
*os
= ddt
->ddt_os
;
79 uint64_t *objectp
= &ddt
->ddt_object
[type
][class];
80 char name
[DDT_NAMELEN
];
82 ddt_object_name(ddt
, type
, class, name
);
84 ASSERT(*objectp
!= 0);
85 ASSERT(ddt_object_count(ddt
, type
, class) == 0);
86 ASSERT(ddt_histogram_empty(&ddt
->ddt_histogram
[type
][class]));
87 VERIFY(zap_remove(os
, DMU_POOL_DIRECTORY_OBJECT
, name
, tx
) == 0);
88 VERIFY(zap_remove(os
, spa
->spa_ddt_stat_object
, name
, tx
) == 0);
89 VERIFY(ddt_ops
[type
]->ddt_op_destroy(os
, *objectp
, tx
) == 0);
90 bzero(&ddt
->ddt_object_stats
[type
][class], sizeof (ddt_object_t
));
96 ddt_object_load(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class)
98 ddt_object_t
*ddo
= &ddt
->ddt_object_stats
[type
][class];
99 dmu_object_info_t doi
;
100 char name
[DDT_NAMELEN
];
103 ddt_object_name(ddt
, type
, class, name
);
105 error
= zap_lookup(ddt
->ddt_os
, DMU_POOL_DIRECTORY_OBJECT
, name
,
106 sizeof (uint64_t), 1, &ddt
->ddt_object
[type
][class]);
111 error
= zap_lookup(ddt
->ddt_os
, ddt
->ddt_spa
->spa_ddt_stat_object
, name
,
112 sizeof (uint64_t), sizeof (ddt_histogram_t
) / sizeof (uint64_t),
113 &ddt
->ddt_histogram
[type
][class]);
116 * Seed the cached statistics.
118 VERIFY(ddt_object_info(ddt
, type
, class, &doi
) == 0);
120 ddo
->ddo_count
= ddt_object_count(ddt
, type
, class);
121 ddo
->ddo_dspace
= doi
.doi_physical_blocks_512
<< 9;
122 ddo
->ddo_mspace
= doi
.doi_fill_count
* doi
.doi_data_block_size
;
129 ddt_object_sync(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
132 ddt_object_t
*ddo
= &ddt
->ddt_object_stats
[type
][class];
133 dmu_object_info_t doi
;
134 char name
[DDT_NAMELEN
];
136 ddt_object_name(ddt
, type
, class, name
);
138 VERIFY(zap_update(ddt
->ddt_os
, ddt
->ddt_spa
->spa_ddt_stat_object
, name
,
139 sizeof (uint64_t), sizeof (ddt_histogram_t
) / sizeof (uint64_t),
140 &ddt
->ddt_histogram
[type
][class], tx
) == 0);
143 * Cache DDT statistics; this is the only time they'll change.
145 VERIFY(ddt_object_info(ddt
, type
, class, &doi
) == 0);
147 ddo
->ddo_count
= ddt_object_count(ddt
, type
, class);
148 ddo
->ddo_dspace
= doi
.doi_physical_blocks_512
<< 9;
149 ddo
->ddo_mspace
= doi
.doi_fill_count
* doi
.doi_data_block_size
;
153 ddt_object_lookup(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
156 if (!ddt_object_exists(ddt
, type
, class))
159 return (ddt_ops
[type
]->ddt_op_lookup(ddt
->ddt_os
,
160 ddt
->ddt_object
[type
][class], dde
));
164 ddt_object_prefetch(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
167 if (!ddt_object_exists(ddt
, type
, class))
170 ddt_ops
[type
]->ddt_op_prefetch(ddt
->ddt_os
,
171 ddt
->ddt_object
[type
][class], dde
);
175 ddt_object_update(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
176 ddt_entry_t
*dde
, dmu_tx_t
*tx
)
178 ASSERT(ddt_object_exists(ddt
, type
, class));
180 return (ddt_ops
[type
]->ddt_op_update(ddt
->ddt_os
,
181 ddt
->ddt_object
[type
][class], dde
, tx
));
185 ddt_object_remove(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
186 ddt_entry_t
*dde
, dmu_tx_t
*tx
)
188 ASSERT(ddt_object_exists(ddt
, type
, class));
190 return (ddt_ops
[type
]->ddt_op_remove(ddt
->ddt_os
,
191 ddt
->ddt_object
[type
][class], dde
, tx
));
195 ddt_object_walk(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
196 uint64_t *walk
, ddt_entry_t
*dde
)
198 ASSERT(ddt_object_exists(ddt
, type
, class));
200 return (ddt_ops
[type
]->ddt_op_walk(ddt
->ddt_os
,
201 ddt
->ddt_object
[type
][class], dde
, walk
));
205 ddt_object_count(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class)
207 ASSERT(ddt_object_exists(ddt
, type
, class));
209 return (ddt_ops
[type
]->ddt_op_count(ddt
->ddt_os
,
210 ddt
->ddt_object
[type
][class]));
214 ddt_object_info(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
215 dmu_object_info_t
*doi
)
217 if (!ddt_object_exists(ddt
, type
, class))
220 return (dmu_object_info(ddt
->ddt_os
, ddt
->ddt_object
[type
][class],
225 ddt_object_exists(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class)
227 return (!!ddt
->ddt_object
[type
][class]);
231 ddt_object_name(ddt_t
*ddt
, enum ddt_type type
, enum ddt_class
class,
234 (void) sprintf(name
, DMU_POOL_DDT
,
235 zio_checksum_table
[ddt
->ddt_checksum
].ci_name
,
236 ddt_ops
[type
]->ddt_op_name
, ddt_class_name
[class]);
240 ddt_bp_fill(const ddt_phys_t
*ddp
, blkptr_t
*bp
, uint64_t txg
)
244 for (int d
= 0; d
< SPA_DVAS_PER_BP
; d
++)
245 bp
->blk_dva
[d
] = ddp
->ddp_dva
[d
];
246 BP_SET_BIRTH(bp
, txg
, ddp
->ddp_phys_birth
);
250 ddt_bp_create(enum zio_checksum checksum
,
251 const ddt_key_t
*ddk
, const ddt_phys_t
*ddp
, blkptr_t
*bp
)
256 ddt_bp_fill(ddp
, bp
, ddp
->ddp_phys_birth
);
258 bp
->blk_cksum
= ddk
->ddk_cksum
;
261 BP_SET_LSIZE(bp
, DDK_GET_LSIZE(ddk
));
262 BP_SET_PSIZE(bp
, DDK_GET_PSIZE(ddk
));
263 BP_SET_COMPRESS(bp
, DDK_GET_COMPRESS(ddk
));
264 BP_SET_CHECKSUM(bp
, checksum
);
265 BP_SET_TYPE(bp
, DMU_OT_DEDUP
);
268 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
272 ddt_key_fill(ddt_key_t
*ddk
, const blkptr_t
*bp
)
274 ddk
->ddk_cksum
= bp
->blk_cksum
;
277 DDK_SET_LSIZE(ddk
, BP_GET_LSIZE(bp
));
278 DDK_SET_PSIZE(ddk
, BP_GET_PSIZE(bp
));
279 DDK_SET_COMPRESS(ddk
, BP_GET_COMPRESS(bp
));
283 ddt_phys_fill(ddt_phys_t
*ddp
, const blkptr_t
*bp
)
285 ASSERT(ddp
->ddp_phys_birth
== 0);
287 for (int d
= 0; d
< SPA_DVAS_PER_BP
; d
++)
288 ddp
->ddp_dva
[d
] = bp
->blk_dva
[d
];
289 ddp
->ddp_phys_birth
= BP_PHYSICAL_BIRTH(bp
);
293 ddt_phys_clear(ddt_phys_t
*ddp
)
295 bzero(ddp
, sizeof (*ddp
));
299 ddt_phys_addref(ddt_phys_t
*ddp
)
305 ddt_phys_decref(ddt_phys_t
*ddp
)
307 ASSERT((int64_t)ddp
->ddp_refcnt
> 0);
312 ddt_phys_free(ddt_t
*ddt
, ddt_key_t
*ddk
, ddt_phys_t
*ddp
, uint64_t txg
)
316 ddt_bp_create(ddt
->ddt_checksum
, ddk
, ddp
, &blk
);
318 zio_free(ddt
->ddt_spa
, txg
, &blk
);
322 ddt_phys_select(const ddt_entry_t
*dde
, const blkptr_t
*bp
)
324 ddt_phys_t
*ddp
= (ddt_phys_t
*)dde
->dde_phys
;
326 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
327 if (DVA_EQUAL(BP_IDENTITY(bp
), &ddp
->ddp_dva
[0]) &&
328 BP_PHYSICAL_BIRTH(bp
) == ddp
->ddp_phys_birth
)
335 ddt_phys_total_refcnt(const ddt_entry_t
*dde
)
339 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++)
340 refcnt
+= dde
->dde_phys
[p
].ddp_refcnt
;
346 ddt_stat_generate(ddt_t
*ddt
, ddt_entry_t
*dde
, ddt_stat_t
*dds
)
348 spa_t
*spa
= ddt
->ddt_spa
;
349 ddt_phys_t
*ddp
= dde
->dde_phys
;
350 ddt_key_t
*ddk
= &dde
->dde_key
;
351 uint64_t lsize
= DDK_GET_LSIZE(ddk
);
352 uint64_t psize
= DDK_GET_PSIZE(ddk
);
354 bzero(dds
, sizeof (*dds
));
356 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
358 uint64_t refcnt
= ddp
->ddp_refcnt
;
360 if (ddp
->ddp_phys_birth
== 0)
363 for (int d
= 0; d
< SPA_DVAS_PER_BP
; d
++)
364 dsize
+= dva_get_dsize_sync(spa
, &ddp
->ddp_dva
[d
]);
366 dds
->dds_blocks
+= 1;
367 dds
->dds_lsize
+= lsize
;
368 dds
->dds_psize
+= psize
;
369 dds
->dds_dsize
+= dsize
;
371 dds
->dds_ref_blocks
+= refcnt
;
372 dds
->dds_ref_lsize
+= lsize
* refcnt
;
373 dds
->dds_ref_psize
+= psize
* refcnt
;
374 dds
->dds_ref_dsize
+= dsize
* refcnt
;
379 ddt_stat_add(ddt_stat_t
*dst
, const ddt_stat_t
*src
, uint64_t neg
)
381 const uint64_t *s
= (const uint64_t *)src
;
382 uint64_t *d
= (uint64_t *)dst
;
383 uint64_t *d_end
= (uint64_t *)(dst
+ 1);
385 ASSERT(neg
== 0 || neg
== -1ULL); /* add or subtract */
388 *d
++ += (*s
++ ^ neg
) - neg
;
392 ddt_stat_update(ddt_t
*ddt
, ddt_entry_t
*dde
, uint64_t neg
)
395 ddt_histogram_t
*ddh
;
398 ddt_stat_generate(ddt
, dde
, &dds
);
400 bucket
= highbit(dds
.dds_ref_blocks
) - 1;
403 ddh
= &ddt
->ddt_histogram
[dde
->dde_type
][dde
->dde_class
];
405 ddt_stat_add(&ddh
->ddh_stat
[bucket
], &dds
, neg
);
409 ddt_histogram_add(ddt_histogram_t
*dst
, const ddt_histogram_t
*src
)
411 for (int h
= 0; h
< 64; h
++)
412 ddt_stat_add(&dst
->ddh_stat
[h
], &src
->ddh_stat
[h
], 0);
416 ddt_histogram_stat(ddt_stat_t
*dds
, const ddt_histogram_t
*ddh
)
418 bzero(dds
, sizeof (*dds
));
420 for (int h
= 0; h
< 64; h
++)
421 ddt_stat_add(dds
, &ddh
->ddh_stat
[h
], 0);
425 ddt_histogram_empty(const ddt_histogram_t
*ddh
)
427 const uint64_t *s
= (const uint64_t *)ddh
;
428 const uint64_t *s_end
= (const uint64_t *)(ddh
+ 1);
438 ddt_get_dedup_object_stats(spa_t
*spa
, ddt_object_t
*ddo_total
)
440 /* Sum the statistics we cached in ddt_object_sync(). */
441 for (enum zio_checksum c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
442 ddt_t
*ddt
= spa
->spa_ddt
[c
];
443 for (enum ddt_type type
= 0; type
< DDT_TYPES
; type
++) {
444 for (enum ddt_class
class = 0; class < DDT_CLASSES
;
447 &ddt
->ddt_object_stats
[type
][class];
448 ddo_total
->ddo_count
+= ddo
->ddo_count
;
449 ddo_total
->ddo_dspace
+= ddo
->ddo_dspace
;
450 ddo_total
->ddo_mspace
+= ddo
->ddo_mspace
;
455 /* ... and compute the averages. */
456 if (ddo_total
->ddo_count
!= 0) {
457 ddo_total
->ddo_dspace
/= ddo_total
->ddo_count
;
458 ddo_total
->ddo_mspace
/= ddo_total
->ddo_count
;
460 ASSERT(ddo_total
->ddo_dspace
== 0);
461 ASSERT(ddo_total
->ddo_mspace
== 0);
466 ddt_get_dedup_histogram(spa_t
*spa
, ddt_histogram_t
*ddh
)
468 for (enum zio_checksum c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
469 ddt_t
*ddt
= spa
->spa_ddt
[c
];
470 for (enum ddt_type type
= 0; type
< DDT_TYPES
; type
++) {
471 for (enum ddt_class
class = 0; class < DDT_CLASSES
;
473 ddt_histogram_add(ddh
,
474 &ddt
->ddt_histogram_cache
[type
][class]);
481 ddt_get_dedup_stats(spa_t
*spa
, ddt_stat_t
*dds_total
)
483 ddt_histogram_t
*ddh_total
;
485 ddh_total
= kmem_zalloc(sizeof (ddt_histogram_t
), KM_SLEEP
);
486 ddt_get_dedup_histogram(spa
, ddh_total
);
487 ddt_histogram_stat(dds_total
, ddh_total
);
488 kmem_free(ddh_total
, sizeof (ddt_histogram_t
));
492 ddt_get_dedup_dspace(spa_t
*spa
)
494 ddt_stat_t dds_total
= { 0 };
496 ddt_get_dedup_stats(spa
, &dds_total
);
497 return (dds_total
.dds_ref_dsize
- dds_total
.dds_dsize
);
501 ddt_get_pool_dedup_ratio(spa_t
*spa
)
503 ddt_stat_t dds_total
= { 0 };
505 ddt_get_dedup_stats(spa
, &dds_total
);
506 if (dds_total
.dds_dsize
== 0)
509 return (dds_total
.dds_ref_dsize
* 100 / dds_total
.dds_dsize
);
513 ddt_ditto_copies_needed(ddt_t
*ddt
, ddt_entry_t
*dde
, ddt_phys_t
*ddp_willref
)
515 spa_t
*spa
= ddt
->ddt_spa
;
516 uint64_t total_refcnt
= 0;
517 uint64_t ditto
= spa
->spa_dedup_ditto
;
518 int total_copies
= 0;
519 int desired_copies
= 0;
521 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
522 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
523 zio_t
*zio
= dde
->dde_lead_zio
[p
];
524 uint64_t refcnt
= ddp
->ddp_refcnt
; /* committed refs */
526 refcnt
+= zio
->io_parent_count
; /* pending refs */
527 if (ddp
== ddp_willref
)
528 refcnt
++; /* caller's ref */
530 total_refcnt
+= refcnt
;
535 if (ditto
== 0 || ditto
> UINT32_MAX
)
538 if (total_refcnt
>= 1)
540 if (total_refcnt
>= ditto
)
542 if (total_refcnt
>= ditto
* ditto
)
545 return (MAX(desired_copies
, total_copies
) - total_copies
);
549 ddt_ditto_copies_present(ddt_entry_t
*dde
)
551 ddt_phys_t
*ddp
= &dde
->dde_phys
[DDT_PHYS_DITTO
];
552 dva_t
*dva
= ddp
->ddp_dva
;
553 int copies
= 0 - DVA_GET_GANG(dva
);
555 for (int d
= 0; d
< SPA_DVAS_PER_BP
; d
++, dva
++)
556 if (DVA_IS_VALID(dva
))
559 ASSERT(copies
>= 0 && copies
< SPA_DVAS_PER_BP
);
565 ddt_compress(void *src
, uchar_t
*dst
, size_t s_len
, size_t d_len
)
567 uchar_t
*version
= dst
++;
568 int cpfunc
= ZIO_COMPRESS_ZLE
;
569 zio_compress_info_t
*ci
= &zio_compress_table
[cpfunc
];
572 ASSERT(d_len
>= s_len
+ 1); /* no compression plus version byte */
574 c_len
= ci
->ci_compress(src
, dst
, s_len
, d_len
- 1, ci
->ci_level
);
576 if (c_len
== s_len
) {
577 cpfunc
= ZIO_COMPRESS_OFF
;
578 bcopy(src
, dst
, s_len
);
581 *version
= (ZFS_HOST_BYTEORDER
& DDT_COMPRESS_BYTEORDER_MASK
) | cpfunc
;
587 ddt_decompress(uchar_t
*src
, void *dst
, size_t s_len
, size_t d_len
)
589 uchar_t version
= *src
++;
590 int cpfunc
= version
& DDT_COMPRESS_FUNCTION_MASK
;
591 zio_compress_info_t
*ci
= &zio_compress_table
[cpfunc
];
593 if (ci
->ci_decompress
!= NULL
)
594 (void) ci
->ci_decompress(src
, dst
, s_len
, d_len
, ci
->ci_level
);
596 bcopy(src
, dst
, d_len
);
598 if ((version
^ ZFS_HOST_BYTEORDER
) & DDT_COMPRESS_BYTEORDER_MASK
)
599 byteswap_uint64_array(dst
, d_len
);
603 ddt_select_by_checksum(spa_t
*spa
, enum zio_checksum c
)
605 return (spa
->spa_ddt
[c
]);
609 ddt_select(spa_t
*spa
, const blkptr_t
*bp
)
611 return (spa
->spa_ddt
[BP_GET_CHECKSUM(bp
)]);
615 ddt_enter(ddt_t
*ddt
)
617 mutex_enter(&ddt
->ddt_lock
);
623 mutex_exit(&ddt
->ddt_lock
);
627 ddt_alloc(const ddt_key_t
*ddk
)
631 dde
= kmem_zalloc(sizeof (ddt_entry_t
), KM_SLEEP
);
632 cv_init(&dde
->dde_cv
, NULL
, CV_DEFAULT
, NULL
);
640 ddt_free(ddt_entry_t
*dde
)
642 ASSERT(!dde
->dde_loading
);
644 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++)
645 ASSERT(dde
->dde_lead_zio
[p
] == NULL
);
647 if (dde
->dde_repair_data
!= NULL
)
648 zio_buf_free(dde
->dde_repair_data
,
649 DDK_GET_PSIZE(&dde
->dde_key
));
651 cv_destroy(&dde
->dde_cv
);
652 kmem_free(dde
, sizeof (*dde
));
656 ddt_remove(ddt_t
*ddt
, ddt_entry_t
*dde
)
658 ASSERT(MUTEX_HELD(&ddt
->ddt_lock
));
660 avl_remove(&ddt
->ddt_tree
, dde
);
665 ddt_lookup(ddt_t
*ddt
, const blkptr_t
*bp
, boolean_t add
)
667 ddt_entry_t
*dde
, dde_search
;
669 enum ddt_class
class;
673 ASSERT(MUTEX_HELD(&ddt
->ddt_lock
));
675 ddt_key_fill(&dde_search
.dde_key
, bp
);
677 dde
= avl_find(&ddt
->ddt_tree
, &dde_search
, &where
);
681 dde
= ddt_alloc(&dde_search
.dde_key
);
682 avl_insert(&ddt
->ddt_tree
, dde
, where
);
685 while (dde
->dde_loading
)
686 cv_wait(&dde
->dde_cv
, &ddt
->ddt_lock
);
691 dde
->dde_loading
= B_TRUE
;
697 for (type
= 0; type
< DDT_TYPES
; type
++) {
698 for (class = 0; class < DDT_CLASSES
; class++) {
699 error
= ddt_object_lookup(ddt
, type
, class, dde
);
707 ASSERT(error
== 0 || error
== ENOENT
);
711 ASSERT(dde
->dde_loaded
== B_FALSE
);
712 ASSERT(dde
->dde_loading
== B_TRUE
);
714 dde
->dde_type
= type
; /* will be DDT_TYPES if no entry found */
715 dde
->dde_class
= class; /* will be DDT_CLASSES if no entry found */
716 dde
->dde_loaded
= B_TRUE
;
717 dde
->dde_loading
= B_FALSE
;
720 ddt_stat_update(ddt
, dde
, -1ULL);
722 cv_broadcast(&dde
->dde_cv
);
728 ddt_prefetch(spa_t
*spa
, const blkptr_t
*bp
)
733 if (!BP_GET_DEDUP(bp
))
737 * We remove the DDT once it's empty and only prefetch dedup blocks
738 * when there are entries in the DDT. Thus no locking is required
739 * as the DDT can't disappear on us.
741 ddt
= ddt_select(spa
, bp
);
742 ddt_key_fill(&dde
.dde_key
, bp
);
744 for (enum ddt_type type
= 0; type
< DDT_TYPES
; type
++) {
745 for (enum ddt_class
class = 0; class < DDT_CLASSES
; class++) {
746 ddt_object_prefetch(ddt
, type
, class, &dde
);
752 ddt_entry_compare(const void *x1
, const void *x2
)
754 const ddt_entry_t
*dde1
= x1
;
755 const ddt_entry_t
*dde2
= x2
;
756 const uint64_t *u1
= (const uint64_t *)&dde1
->dde_key
;
757 const uint64_t *u2
= (const uint64_t *)&dde2
->dde_key
;
759 for (int i
= 0; i
< DDT_KEY_WORDS
; i
++) {
770 ddt_table_alloc(spa_t
*spa
, enum zio_checksum c
)
774 ddt
= kmem_zalloc(sizeof (*ddt
), KM_SLEEP
);
776 mutex_init(&ddt
->ddt_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
777 avl_create(&ddt
->ddt_tree
, ddt_entry_compare
,
778 sizeof (ddt_entry_t
), offsetof(ddt_entry_t
, dde_node
));
779 avl_create(&ddt
->ddt_repair_tree
, ddt_entry_compare
,
780 sizeof (ddt_entry_t
), offsetof(ddt_entry_t
, dde_node
));
781 ddt
->ddt_checksum
= c
;
783 ddt
->ddt_os
= spa
->spa_meta_objset
;
789 ddt_table_free(ddt_t
*ddt
)
791 ASSERT(avl_numnodes(&ddt
->ddt_tree
) == 0);
792 ASSERT(avl_numnodes(&ddt
->ddt_repair_tree
) == 0);
793 avl_destroy(&ddt
->ddt_tree
);
794 avl_destroy(&ddt
->ddt_repair_tree
);
795 mutex_destroy(&ddt
->ddt_lock
);
796 kmem_free(ddt
, sizeof (*ddt
));
800 ddt_create(spa_t
*spa
)
802 spa
->spa_dedup_checksum
= ZIO_DEDUPCHECKSUM
;
804 for (enum zio_checksum c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++)
805 spa
->spa_ddt
[c
] = ddt_table_alloc(spa
, c
);
815 error
= zap_lookup(spa
->spa_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
816 DMU_POOL_DDT_STATS
, sizeof (uint64_t), 1,
817 &spa
->spa_ddt_stat_object
);
820 return (error
== ENOENT
? 0 : error
);
822 for (enum zio_checksum c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
823 ddt_t
*ddt
= spa
->spa_ddt
[c
];
824 for (enum ddt_type type
= 0; type
< DDT_TYPES
; type
++) {
825 for (enum ddt_class
class = 0; class < DDT_CLASSES
;
827 error
= ddt_object_load(ddt
, type
, class);
828 if (error
!= 0 && error
!= ENOENT
)
834 * Seed the cached histograms.
836 bcopy(ddt
->ddt_histogram
, &ddt
->ddt_histogram_cache
,
837 sizeof (ddt
->ddt_histogram
));
844 ddt_unload(spa_t
*spa
)
846 for (enum zio_checksum c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
847 if (spa
->spa_ddt
[c
]) {
848 ddt_table_free(spa
->spa_ddt
[c
]);
849 spa
->spa_ddt
[c
] = NULL
;
855 ddt_class_contains(spa_t
*spa
, enum ddt_class max_class
, const blkptr_t
*bp
)
860 if (!BP_GET_DEDUP(bp
))
863 if (max_class
== DDT_CLASS_UNIQUE
)
866 ddt
= spa
->spa_ddt
[BP_GET_CHECKSUM(bp
)];
868 ddt_key_fill(&dde
.dde_key
, bp
);
870 for (enum ddt_type type
= 0; type
< DDT_TYPES
; type
++)
871 for (enum ddt_class
class = 0; class <= max_class
; class++)
872 if (ddt_object_lookup(ddt
, type
, class, &dde
) == 0)
879 ddt_repair_start(ddt_t
*ddt
, const blkptr_t
*bp
)
884 ddt_key_fill(&ddk
, bp
);
886 dde
= ddt_alloc(&ddk
);
888 for (enum ddt_type type
= 0; type
< DDT_TYPES
; type
++) {
889 for (enum ddt_class
class = 0; class < DDT_CLASSES
; class++) {
891 * We can only do repair if there are multiple copies
892 * of the block. For anything in the UNIQUE class,
893 * there's definitely only one copy, so don't even try.
895 if (class != DDT_CLASS_UNIQUE
&&
896 ddt_object_lookup(ddt
, type
, class, dde
) == 0)
901 bzero(dde
->dde_phys
, sizeof (dde
->dde_phys
));
907 ddt_repair_done(ddt_t
*ddt
, ddt_entry_t
*dde
)
913 if (dde
->dde_repair_data
!= NULL
&& spa_writeable(ddt
->ddt_spa
) &&
914 avl_find(&ddt
->ddt_repair_tree
, dde
, &where
) == NULL
)
915 avl_insert(&ddt
->ddt_repair_tree
, dde
, where
);
923 ddt_repair_entry_done(zio_t
*zio
)
925 ddt_entry_t
*rdde
= zio
->io_private
;
931 ddt_repair_entry(ddt_t
*ddt
, ddt_entry_t
*dde
, ddt_entry_t
*rdde
, zio_t
*rio
)
933 ddt_phys_t
*ddp
= dde
->dde_phys
;
934 ddt_phys_t
*rddp
= rdde
->dde_phys
;
935 ddt_key_t
*ddk
= &dde
->dde_key
;
936 ddt_key_t
*rddk
= &rdde
->dde_key
;
940 zio
= zio_null(rio
, rio
->io_spa
, NULL
,
941 ddt_repair_entry_done
, rdde
, rio
->io_flags
);
943 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++, rddp
++) {
944 if (ddp
->ddp_phys_birth
== 0 ||
945 ddp
->ddp_phys_birth
!= rddp
->ddp_phys_birth
||
946 bcmp(ddp
->ddp_dva
, rddp
->ddp_dva
, sizeof (ddp
->ddp_dva
)))
948 ddt_bp_create(ddt
->ddt_checksum
, ddk
, ddp
, &blk
);
949 zio_nowait(zio_rewrite(zio
, zio
->io_spa
, 0, &blk
,
950 rdde
->dde_repair_data
, DDK_GET_PSIZE(rddk
), NULL
, NULL
,
951 ZIO_PRIORITY_SYNC_WRITE
, ZIO_DDT_CHILD_FLAGS(zio
), NULL
));
958 ddt_repair_table(ddt_t
*ddt
, zio_t
*rio
)
960 spa_t
*spa
= ddt
->ddt_spa
;
961 ddt_entry_t
*dde
, *rdde_next
, *rdde
;
962 avl_tree_t
*t
= &ddt
->ddt_repair_tree
;
965 if (spa_sync_pass(spa
) > 1)
969 for (rdde
= avl_first(t
); rdde
!= NULL
; rdde
= rdde_next
) {
970 rdde_next
= AVL_NEXT(t
, rdde
);
971 avl_remove(&ddt
->ddt_repair_tree
, rdde
);
973 ddt_bp_create(ddt
->ddt_checksum
, &rdde
->dde_key
, NULL
, &blk
);
974 dde
= ddt_repair_start(ddt
, &blk
);
975 ddt_repair_entry(ddt
, dde
, rdde
, rio
);
976 ddt_repair_done(ddt
, dde
);
983 ddt_sync_entry(ddt_t
*ddt
, ddt_entry_t
*dde
, dmu_tx_t
*tx
, uint64_t txg
)
985 dsl_pool_t
*dp
= ddt
->ddt_spa
->spa_dsl_pool
;
986 ddt_phys_t
*ddp
= dde
->dde_phys
;
987 ddt_key_t
*ddk
= &dde
->dde_key
;
988 enum ddt_type otype
= dde
->dde_type
;
989 enum ddt_type ntype
= DDT_TYPE_CURRENT
;
990 enum ddt_class oclass
= dde
->dde_class
;
991 enum ddt_class nclass
;
992 uint64_t total_refcnt
= 0;
994 ASSERT(dde
->dde_loaded
);
995 ASSERT(!dde
->dde_loading
);
997 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
998 ASSERT(dde
->dde_lead_zio
[p
] == NULL
);
999 ASSERT((int64_t)ddp
->ddp_refcnt
>= 0);
1000 if (ddp
->ddp_phys_birth
== 0) {
1001 ASSERT(ddp
->ddp_refcnt
== 0);
1004 if (p
== DDT_PHYS_DITTO
) {
1005 if (ddt_ditto_copies_needed(ddt
, dde
, NULL
) == 0)
1006 ddt_phys_free(ddt
, ddk
, ddp
, txg
);
1009 if (ddp
->ddp_refcnt
== 0)
1010 ddt_phys_free(ddt
, ddk
, ddp
, txg
);
1011 total_refcnt
+= ddp
->ddp_refcnt
;
1014 if (dde
->dde_phys
[DDT_PHYS_DITTO
].ddp_phys_birth
!= 0)
1015 nclass
= DDT_CLASS_DITTO
;
1016 else if (total_refcnt
> 1)
1017 nclass
= DDT_CLASS_DUPLICATE
;
1019 nclass
= DDT_CLASS_UNIQUE
;
1021 if (otype
!= DDT_TYPES
&&
1022 (otype
!= ntype
|| oclass
!= nclass
|| total_refcnt
== 0)) {
1023 VERIFY(ddt_object_remove(ddt
, otype
, oclass
, dde
, tx
) == 0);
1024 ASSERT(ddt_object_lookup(ddt
, otype
, oclass
, dde
) == ENOENT
);
1027 if (total_refcnt
!= 0) {
1028 dde
->dde_type
= ntype
;
1029 dde
->dde_class
= nclass
;
1030 ddt_stat_update(ddt
, dde
, 0);
1031 if (!ddt_object_exists(ddt
, ntype
, nclass
))
1032 ddt_object_create(ddt
, ntype
, nclass
, tx
);
1033 VERIFY(ddt_object_update(ddt
, ntype
, nclass
, dde
, tx
) == 0);
1036 * If the class changes, the order that we scan this bp
1037 * changes. If it decreases, we could miss it, so
1038 * scan it right now. (This covers both class changing
1039 * while we are doing ddt_walk(), and when we are
1042 if (nclass
< oclass
) {
1043 dsl_scan_ddt_entry(dp
->dp_scan
,
1044 ddt
->ddt_checksum
, dde
, tx
);
1050 ddt_sync_table(ddt_t
*ddt
, dmu_tx_t
*tx
, uint64_t txg
)
1052 spa_t
*spa
= ddt
->ddt_spa
;
1054 void *cookie
= NULL
;
1056 if (avl_numnodes(&ddt
->ddt_tree
) == 0)
1059 ASSERT(spa
->spa_uberblock
.ub_version
>= SPA_VERSION_DEDUP
);
1061 if (spa
->spa_ddt_stat_object
== 0) {
1062 spa
->spa_ddt_stat_object
= zap_create(ddt
->ddt_os
,
1063 DMU_OT_DDT_STATS
, DMU_OT_NONE
, 0, tx
);
1064 VERIFY(zap_add(ddt
->ddt_os
, DMU_POOL_DIRECTORY_OBJECT
,
1065 DMU_POOL_DDT_STATS
, sizeof (uint64_t), 1,
1066 &spa
->spa_ddt_stat_object
, tx
) == 0);
1069 while ((dde
= avl_destroy_nodes(&ddt
->ddt_tree
, &cookie
)) != NULL
) {
1070 ddt_sync_entry(ddt
, dde
, tx
, txg
);
1074 for (enum ddt_type type
= 0; type
< DDT_TYPES
; type
++) {
1075 for (enum ddt_class
class = 0; class < DDT_CLASSES
; class++) {
1076 if (!ddt_object_exists(ddt
, type
, class))
1078 ddt_object_sync(ddt
, type
, class, tx
);
1079 if (ddt_object_count(ddt
, type
, class) == 0)
1080 ddt_object_destroy(ddt
, type
, class, tx
);
1084 bcopy(ddt
->ddt_histogram
, &ddt
->ddt_histogram_cache
,
1085 sizeof (ddt
->ddt_histogram
));
1089 ddt_sync(spa_t
*spa
, uint64_t txg
)
1092 zio_t
*rio
= zio_root(spa
, NULL
, NULL
,
1093 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
);
1095 ASSERT(spa_syncing_txg(spa
) == txg
);
1097 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
1099 for (enum zio_checksum c
= 0; c
< ZIO_CHECKSUM_FUNCTIONS
; c
++) {
1100 ddt_t
*ddt
= spa
->spa_ddt
[c
];
1103 ddt_sync_table(ddt
, tx
, txg
);
1104 ddt_repair_table(ddt
, rio
);
1107 (void) zio_wait(rio
);
1113 ddt_walk(spa_t
*spa
, ddt_bookmark_t
*ddb
, ddt_entry_t
*dde
)
1118 ddt_t
*ddt
= spa
->spa_ddt
[ddb
->ddb_checksum
];
1120 if (ddt_object_exists(ddt
, ddb
->ddb_type
,
1122 error
= ddt_object_walk(ddt
,
1123 ddb
->ddb_type
, ddb
->ddb_class
,
1124 &ddb
->ddb_cursor
, dde
);
1126 dde
->dde_type
= ddb
->ddb_type
;
1127 dde
->dde_class
= ddb
->ddb_class
;
1130 if (error
!= ENOENT
)
1132 ddb
->ddb_cursor
= 0;
1133 } while (++ddb
->ddb_checksum
< ZIO_CHECKSUM_FUNCTIONS
);
1134 ddb
->ddb_checksum
= 0;
1135 } while (++ddb
->ddb_type
< DDT_TYPES
);
1137 } while (++ddb
->ddb_class
< DDT_CLASSES
);