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]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
27 * Copyright (c) 2019, Klara Inc.
28 * Copyright (c) 2019, Allan Jude
29 * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
32 #include <sys/zfs_context.h>
35 #include <sys/dmu_send.h>
36 #include <sys/dmu_impl.h>
38 #include <sys/dmu_objset.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/dsl_dir.h>
41 #include <sys/dmu_tx.h>
44 #include <sys/dmu_zfetch.h>
46 #include <sys/sa_impl.h>
47 #include <sys/zfeature.h>
48 #include <sys/blkptr.h>
49 #include <sys/range_tree.h>
50 #include <sys/trace_zfs.h>
51 #include <sys/callb.h>
56 #include <sys/spa_impl.h>
57 #include <sys/wmsum.h>
58 #include <sys/vdev_impl.h>
60 static kstat_t
*dbuf_ksp
;
62 typedef struct dbuf_stats
{
64 * Various statistics about the size of the dbuf cache.
66 kstat_named_t cache_count
;
67 kstat_named_t cache_size_bytes
;
68 kstat_named_t cache_size_bytes_max
;
70 * Statistics regarding the bounds on the dbuf cache size.
72 kstat_named_t cache_target_bytes
;
73 kstat_named_t cache_lowater_bytes
;
74 kstat_named_t cache_hiwater_bytes
;
76 * Total number of dbuf cache evictions that have occurred.
78 kstat_named_t cache_total_evicts
;
80 * The distribution of dbuf levels in the dbuf cache and
81 * the total size of all dbufs at each level.
83 kstat_named_t cache_levels
[DN_MAX_LEVELS
];
84 kstat_named_t cache_levels_bytes
[DN_MAX_LEVELS
];
86 * Statistics about the dbuf hash table.
88 kstat_named_t hash_hits
;
89 kstat_named_t hash_misses
;
90 kstat_named_t hash_collisions
;
91 kstat_named_t hash_elements
;
92 kstat_named_t hash_elements_max
;
94 * Number of sublists containing more than one dbuf in the dbuf
95 * hash table. Keep track of the longest hash chain.
97 kstat_named_t hash_chains
;
98 kstat_named_t hash_chain_max
;
100 * Number of times a dbuf_create() discovers that a dbuf was
101 * already created and in the dbuf hash table.
103 kstat_named_t hash_insert_race
;
105 * Number of entries in the hash table dbuf and mutex arrays.
107 kstat_named_t hash_table_count
;
108 kstat_named_t hash_mutex_count
;
110 * Statistics about the size of the metadata dbuf cache.
112 kstat_named_t metadata_cache_count
;
113 kstat_named_t metadata_cache_size_bytes
;
114 kstat_named_t metadata_cache_size_bytes_max
;
116 * For diagnostic purposes, this is incremented whenever we can't add
117 * something to the metadata cache because it's full, and instead put
118 * the data in the regular dbuf cache.
120 kstat_named_t metadata_cache_overflow
;
123 dbuf_stats_t dbuf_stats
= {
124 { "cache_count", KSTAT_DATA_UINT64
},
125 { "cache_size_bytes", KSTAT_DATA_UINT64
},
126 { "cache_size_bytes_max", KSTAT_DATA_UINT64
},
127 { "cache_target_bytes", KSTAT_DATA_UINT64
},
128 { "cache_lowater_bytes", KSTAT_DATA_UINT64
},
129 { "cache_hiwater_bytes", KSTAT_DATA_UINT64
},
130 { "cache_total_evicts", KSTAT_DATA_UINT64
},
131 { { "cache_levels_N", KSTAT_DATA_UINT64
} },
132 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64
} },
133 { "hash_hits", KSTAT_DATA_UINT64
},
134 { "hash_misses", KSTAT_DATA_UINT64
},
135 { "hash_collisions", KSTAT_DATA_UINT64
},
136 { "hash_elements", KSTAT_DATA_UINT64
},
137 { "hash_elements_max", KSTAT_DATA_UINT64
},
138 { "hash_chains", KSTAT_DATA_UINT64
},
139 { "hash_chain_max", KSTAT_DATA_UINT64
},
140 { "hash_insert_race", KSTAT_DATA_UINT64
},
141 { "hash_table_count", KSTAT_DATA_UINT64
},
142 { "hash_mutex_count", KSTAT_DATA_UINT64
},
143 { "metadata_cache_count", KSTAT_DATA_UINT64
},
144 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64
},
145 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64
},
146 { "metadata_cache_overflow", KSTAT_DATA_UINT64
}
151 wmsum_t cache_total_evicts
;
152 wmsum_t cache_levels
[DN_MAX_LEVELS
];
153 wmsum_t cache_levels_bytes
[DN_MAX_LEVELS
];
156 wmsum_t hash_collisions
;
158 wmsum_t hash_insert_race
;
159 wmsum_t metadata_cache_count
;
160 wmsum_t metadata_cache_overflow
;
163 #define DBUF_STAT_INCR(stat, val) \
164 wmsum_add(&dbuf_sums.stat, val)
165 #define DBUF_STAT_DECR(stat, val) \
166 DBUF_STAT_INCR(stat, -(val))
167 #define DBUF_STAT_BUMP(stat) \
168 DBUF_STAT_INCR(stat, 1)
169 #define DBUF_STAT_BUMPDOWN(stat) \
170 DBUF_STAT_INCR(stat, -1)
171 #define DBUF_STAT_MAX(stat, v) { \
173 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
174 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
178 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
179 static void dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
);
182 * Global data structures and functions for the dbuf cache.
184 static kmem_cache_t
*dbuf_kmem_cache
;
185 static taskq_t
*dbu_evict_taskq
;
187 static kthread_t
*dbuf_cache_evict_thread
;
188 static kmutex_t dbuf_evict_lock
;
189 static kcondvar_t dbuf_evict_cv
;
190 static boolean_t dbuf_evict_thread_exit
;
193 * There are two dbuf caches; each dbuf can only be in one of them at a time.
195 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
196 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
197 * that represent the metadata that describes filesystems/snapshots/
198 * bookmarks/properties/etc. We only evict from this cache when we export a
199 * pool, to short-circuit as much I/O as possible for all administrative
200 * commands that need the metadata. There is no eviction policy for this
201 * cache, because we try to only include types in it which would occupy a
202 * very small amount of space per object but create a large impact on the
203 * performance of these commands. Instead, after it reaches a maximum size
204 * (which should only happen on very small memory systems with a very large
205 * number of filesystem objects), we stop taking new dbufs into the
206 * metadata cache, instead putting them in the normal dbuf cache.
208 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
209 * are not currently held but have been recently released. These dbufs
210 * are not eligible for arc eviction until they are aged out of the cache.
211 * Dbufs that are aged out of the cache will be immediately destroyed and
212 * become eligible for arc eviction.
214 * Dbufs are added to these caches once the last hold is released. If a dbuf is
215 * later accessed and still exists in the dbuf cache, then it will be removed
216 * from the cache and later re-added to the head of the cache.
218 * If a given dbuf meets the requirements for the metadata cache, it will go
219 * there, otherwise it will be considered for the generic LRU dbuf cache. The
220 * caches and the refcounts tracking their sizes are stored in an array indexed
221 * by those caches' matching enum values (from dbuf_cached_state_t).
223 typedef struct dbuf_cache
{
225 zfs_refcount_t size ____cacheline_aligned
;
227 dbuf_cache_t dbuf_caches
[DB_CACHE_MAX
];
229 /* Size limits for the caches */
230 static uint64_t dbuf_cache_max_bytes
= UINT64_MAX
;
231 static uint64_t dbuf_metadata_cache_max_bytes
= UINT64_MAX
;
233 /* Set the default sizes of the caches to log2 fraction of arc size */
234 static uint_t dbuf_cache_shift
= 5;
235 static uint_t dbuf_metadata_cache_shift
= 6;
237 /* Set the dbuf hash mutex count as log2 shift (dynamic by default) */
238 static uint_t dbuf_mutex_cache_shift
= 0;
240 static unsigned long dbuf_cache_target_bytes(void);
241 static unsigned long dbuf_metadata_cache_target_bytes(void);
244 * The LRU dbuf cache uses a three-stage eviction policy:
245 * - A low water marker designates when the dbuf eviction thread
246 * should stop evicting from the dbuf cache.
247 * - When we reach the maximum size (aka mid water mark), we
248 * signal the eviction thread to run.
249 * - The high water mark indicates when the eviction thread
250 * is unable to keep up with the incoming load and eviction must
251 * happen in the context of the calling thread.
255 * low water mid water hi water
256 * +----------------------------------------+----------+----------+
261 * +----------------------------------------+----------+----------+
263 * evicting eviction directly
266 * The high and low water marks indicate the operating range for the eviction
267 * thread. The low water mark is, by default, 90% of the total size of the
268 * cache and the high water mark is at 110% (both of these percentages can be
269 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
270 * respectively). The eviction thread will try to ensure that the cache remains
271 * within this range by waking up every second and checking if the cache is
272 * above the low water mark. The thread can also be woken up by callers adding
273 * elements into the cache if the cache is larger than the mid water (i.e max
274 * cache size). Once the eviction thread is woken up and eviction is required,
275 * it will continue evicting buffers until it's able to reduce the cache size
276 * to the low water mark. If the cache size continues to grow and hits the high
277 * water mark, then callers adding elements to the cache will begin to evict
278 * directly from the cache until the cache is no longer above the high water
283 * The percentage above and below the maximum cache size.
285 static uint_t dbuf_cache_hiwater_pct
= 10;
286 static uint_t dbuf_cache_lowater_pct
= 10;
289 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
291 (void) unused
, (void) kmflag
;
292 dmu_buf_impl_t
*db
= vdb
;
293 memset(db
, 0, sizeof (dmu_buf_impl_t
));
295 mutex_init(&db
->db_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
296 rw_init(&db
->db_rwlock
, NULL
, RW_NOLOCKDEP
, NULL
);
297 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
298 multilist_link_init(&db
->db_cache_link
);
299 zfs_refcount_create(&db
->db_holds
);
305 dbuf_dest(void *vdb
, void *unused
)
308 dmu_buf_impl_t
*db
= vdb
;
309 mutex_destroy(&db
->db_mtx
);
310 rw_destroy(&db
->db_rwlock
);
311 cv_destroy(&db
->db_changed
);
312 ASSERT(!multilist_link_active(&db
->db_cache_link
));
313 zfs_refcount_destroy(&db
->db_holds
);
317 * dbuf hash table routines
319 static dbuf_hash_table_t dbuf_hash_table
;
322 * We use Cityhash for this. It's fast, and has good hash properties without
323 * requiring any large static buffers.
326 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
328 return (cityhash4((uintptr_t)os
, obj
, (uint64_t)lvl
, blkid
));
331 #define DTRACE_SET_STATE(db, why) \
332 DTRACE_PROBE2(dbuf__state_change, dmu_buf_impl_t *, db, \
335 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
336 ((dbuf)->db.db_object == (obj) && \
337 (dbuf)->db_objset == (os) && \
338 (dbuf)->db_level == (level) && \
339 (dbuf)->db_blkid == (blkid))
342 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
,
345 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
350 hv
= dbuf_hash(os
, obj
, level
, blkid
);
351 idx
= hv
& h
->hash_table_mask
;
353 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
354 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
355 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
356 mutex_enter(&db
->db_mtx
);
357 if (db
->db_state
!= DB_EVICTING
) {
358 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
361 mutex_exit(&db
->db_mtx
);
364 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
365 if (hash_out
!= NULL
)
370 static dmu_buf_impl_t
*
371 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
374 dmu_buf_impl_t
*db
= NULL
;
376 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
377 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
378 if (dn
->dn_bonus
!= NULL
) {
380 mutex_enter(&db
->db_mtx
);
382 rw_exit(&dn
->dn_struct_rwlock
);
383 dnode_rele(dn
, FTAG
);
389 * Insert an entry into the hash table. If there is already an element
390 * equal to elem in the hash table, then the already existing element
391 * will be returned and the new element will not be inserted.
392 * Otherwise returns NULL.
394 static dmu_buf_impl_t
*
395 dbuf_hash_insert(dmu_buf_impl_t
*db
)
397 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
398 objset_t
*os
= db
->db_objset
;
399 uint64_t obj
= db
->db
.db_object
;
400 int level
= db
->db_level
;
405 blkid
= db
->db_blkid
;
406 ASSERT3U(dbuf_hash(os
, obj
, level
, blkid
), ==, db
->db_hash
);
407 idx
= db
->db_hash
& h
->hash_table_mask
;
409 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
410 for (dbf
= h
->hash_table
[idx
], i
= 0; dbf
!= NULL
;
411 dbf
= dbf
->db_hash_next
, i
++) {
412 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
413 mutex_enter(&dbf
->db_mtx
);
414 if (dbf
->db_state
!= DB_EVICTING
) {
415 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
418 mutex_exit(&dbf
->db_mtx
);
423 DBUF_STAT_BUMP(hash_collisions
);
425 DBUF_STAT_BUMP(hash_chains
);
427 DBUF_STAT_MAX(hash_chain_max
, i
);
430 mutex_enter(&db
->db_mtx
);
431 db
->db_hash_next
= h
->hash_table
[idx
];
432 h
->hash_table
[idx
] = db
;
433 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
434 uint64_t he
= atomic_inc_64_nv(&dbuf_stats
.hash_elements
.value
.ui64
);
435 DBUF_STAT_MAX(hash_elements_max
, he
);
441 * This returns whether this dbuf should be stored in the metadata cache, which
442 * is based on whether it's from one of the dnode types that store data related
443 * to traversing dataset hierarchies.
446 dbuf_include_in_metadata_cache(dmu_buf_impl_t
*db
)
449 dmu_object_type_t type
= DB_DNODE(db
)->dn_type
;
452 /* Check if this dbuf is one of the types we care about */
453 if (DMU_OT_IS_METADATA_CACHED(type
)) {
454 /* If we hit this, then we set something up wrong in dmu_ot */
455 ASSERT(DMU_OT_IS_METADATA(type
));
458 * Sanity check for small-memory systems: don't allocate too
459 * much memory for this purpose.
461 if (zfs_refcount_count(
462 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
) >
463 dbuf_metadata_cache_target_bytes()) {
464 DBUF_STAT_BUMP(metadata_cache_overflow
);
475 * Remove an entry from the hash table. It must be in the EVICTING state.
478 dbuf_hash_remove(dmu_buf_impl_t
*db
)
480 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
482 dmu_buf_impl_t
*dbf
, **dbp
;
484 ASSERT3U(dbuf_hash(db
->db_objset
, db
->db
.db_object
, db
->db_level
,
485 db
->db_blkid
), ==, db
->db_hash
);
486 idx
= db
->db_hash
& h
->hash_table_mask
;
489 * We mustn't hold db_mtx to maintain lock ordering:
490 * DBUF_HASH_MUTEX > db_mtx.
492 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
493 ASSERT(db
->db_state
== DB_EVICTING
);
494 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
496 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
497 dbp
= &h
->hash_table
[idx
];
498 while ((dbf
= *dbp
) != db
) {
499 dbp
= &dbf
->db_hash_next
;
502 *dbp
= db
->db_hash_next
;
503 db
->db_hash_next
= NULL
;
504 if (h
->hash_table
[idx
] &&
505 h
->hash_table
[idx
]->db_hash_next
== NULL
)
506 DBUF_STAT_BUMPDOWN(hash_chains
);
507 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
508 atomic_dec_64(&dbuf_stats
.hash_elements
.value
.ui64
);
514 } dbvu_verify_type_t
;
517 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
522 if (db
->db_user
== NULL
)
525 /* Only data blocks support the attachment of user data. */
526 ASSERT(db
->db_level
== 0);
528 /* Clients must resolve a dbuf before attaching user data. */
529 ASSERT(db
->db
.db_data
!= NULL
);
530 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
532 holds
= zfs_refcount_count(&db
->db_holds
);
533 if (verify_type
== DBVU_EVICTING
) {
535 * Immediate eviction occurs when holds == dirtycnt.
536 * For normal eviction buffers, holds is zero on
537 * eviction, except when dbuf_fix_old_data() calls
538 * dbuf_clear_data(). However, the hold count can grow
539 * during eviction even though db_mtx is held (see
540 * dmu_bonus_hold() for an example), so we can only
541 * test the generic invariant that holds >= dirtycnt.
543 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
545 if (db
->db_user_immediate_evict
== TRUE
)
546 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
548 ASSERT3U(holds
, >, 0);
554 dbuf_evict_user(dmu_buf_impl_t
*db
)
556 dmu_buf_user_t
*dbu
= db
->db_user
;
558 ASSERT(MUTEX_HELD(&db
->db_mtx
));
563 dbuf_verify_user(db
, DBVU_EVICTING
);
567 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
568 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
571 if (db
->db_caching_status
!= DB_NO_CACHE
) {
573 * This is a cached dbuf, so the size of the user data is
574 * included in its cached amount. We adjust it here because the
575 * user data has already been detached from the dbuf, and the
576 * sync functions are not supposed to touch it (the dbuf might
577 * not exist anymore by the time the sync functions run.
579 uint64_t size
= dbu
->dbu_size
;
580 (void) zfs_refcount_remove_many(
581 &dbuf_caches
[db
->db_caching_status
].size
, size
, dbu
);
582 if (db
->db_caching_status
== DB_DBUF_CACHE
)
583 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
], size
);
587 * There are two eviction callbacks - one that we call synchronously
588 * and one that we invoke via a taskq. The async one is useful for
589 * avoiding lock order reversals and limiting stack depth.
591 * Note that if we have a sync callback but no async callback,
592 * it's likely that the sync callback will free the structure
593 * containing the dbu. In that case we need to take care to not
594 * dereference dbu after calling the sync evict func.
596 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
598 if (dbu
->dbu_evict_func_sync
!= NULL
)
599 dbu
->dbu_evict_func_sync(dbu
);
602 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
603 dbu
, 0, &dbu
->dbu_tqent
);
608 dbuf_is_metadata(dmu_buf_impl_t
*db
)
611 * Consider indirect blocks and spill blocks to be meta data.
613 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
616 boolean_t is_metadata
;
619 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
622 return (is_metadata
);
627 * We want to exclude buffers that are on a special allocation class from
631 dbuf_is_l2cacheable(dmu_buf_impl_t
*db
)
633 if (db
->db_objset
->os_secondary_cache
== ZFS_CACHE_ALL
||
634 (db
->db_objset
->os_secondary_cache
==
635 ZFS_CACHE_METADATA
&& dbuf_is_metadata(db
))) {
636 if (l2arc_exclude_special
== 0)
639 blkptr_t
*bp
= db
->db_blkptr
;
640 if (bp
== NULL
|| BP_IS_HOLE(bp
))
642 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
643 vdev_t
*rvd
= db
->db_objset
->os_spa
->spa_root_vdev
;
646 if (vdev
< rvd
->vdev_children
)
647 vd
= rvd
->vdev_child
[vdev
];
652 if (vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
653 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
)
659 static inline boolean_t
660 dnode_level_is_l2cacheable(blkptr_t
*bp
, dnode_t
*dn
, int64_t level
)
662 if (dn
->dn_objset
->os_secondary_cache
== ZFS_CACHE_ALL
||
663 (dn
->dn_objset
->os_secondary_cache
== ZFS_CACHE_METADATA
&&
665 DMU_OT_IS_METADATA(dn
->dn_handle
->dnh_dnode
->dn_type
)))) {
666 if (l2arc_exclude_special
== 0)
669 if (bp
== NULL
|| BP_IS_HOLE(bp
))
671 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
672 vdev_t
*rvd
= dn
->dn_objset
->os_spa
->spa_root_vdev
;
675 if (vdev
< rvd
->vdev_children
)
676 vd
= rvd
->vdev_child
[vdev
];
681 if (vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
682 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
)
690 * This function *must* return indices evenly distributed between all
691 * sublists of the multilist. This is needed due to how the dbuf eviction
692 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
693 * distributed between all sublists and uses this assumption when
694 * deciding which sublist to evict from and how much to evict from it.
697 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
699 dmu_buf_impl_t
*db
= obj
;
702 * The assumption here, is the hash value for a given
703 * dmu_buf_impl_t will remain constant throughout it's lifetime
704 * (i.e. it's objset, object, level and blkid fields don't change).
705 * Thus, we don't need to store the dbuf's sublist index
706 * on insertion, as this index can be recalculated on removal.
708 * Also, the low order bits of the hash value are thought to be
709 * distributed evenly. Otherwise, in the case that the multilist
710 * has a power of two number of sublists, each sublists' usage
711 * would not be evenly distributed. In this context full 64bit
712 * division would be a waste of time, so limit it to 32 bits.
714 return ((unsigned int)dbuf_hash(db
->db_objset
, db
->db
.db_object
,
715 db
->db_level
, db
->db_blkid
) %
716 multilist_get_num_sublists(ml
));
720 * The target size of the dbuf cache can grow with the ARC target,
721 * unless limited by the tunable dbuf_cache_max_bytes.
723 static inline unsigned long
724 dbuf_cache_target_bytes(void)
726 return (MIN(dbuf_cache_max_bytes
,
727 arc_target_bytes() >> dbuf_cache_shift
));
731 * The target size of the dbuf metadata cache can grow with the ARC target,
732 * unless limited by the tunable dbuf_metadata_cache_max_bytes.
734 static inline unsigned long
735 dbuf_metadata_cache_target_bytes(void)
737 return (MIN(dbuf_metadata_cache_max_bytes
,
738 arc_target_bytes() >> dbuf_metadata_cache_shift
));
741 static inline uint64_t
742 dbuf_cache_hiwater_bytes(void)
744 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
745 return (dbuf_cache_target
+
746 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100);
749 static inline uint64_t
750 dbuf_cache_lowater_bytes(void)
752 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
753 return (dbuf_cache_target
-
754 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100);
757 static inline boolean_t
758 dbuf_cache_above_lowater(void)
760 return (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
761 dbuf_cache_lowater_bytes());
765 * Evict the oldest eligible dbuf from the dbuf cache.
770 int idx
= multilist_get_random_index(&dbuf_caches
[DB_DBUF_CACHE
].cache
);
771 multilist_sublist_t
*mls
= multilist_sublist_lock_idx(
772 &dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
774 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
776 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
777 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
778 db
= multilist_sublist_prev(mls
, db
);
781 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
782 multilist_sublist_t
*, mls
);
785 multilist_sublist_remove(mls
, db
);
786 multilist_sublist_unlock(mls
);
787 uint64_t size
= db
->db
.db_size
;
788 uint64_t usize
= dmu_buf_user_size(&db
->db
);
789 (void) zfs_refcount_remove_many(
790 &dbuf_caches
[DB_DBUF_CACHE
].size
, size
, db
);
791 (void) zfs_refcount_remove_many(
792 &dbuf_caches
[DB_DBUF_CACHE
].size
, usize
, db
->db_user
);
793 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
794 DBUF_STAT_BUMPDOWN(cache_count
);
795 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
], size
+ usize
);
796 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
797 db
->db_caching_status
= DB_NO_CACHE
;
799 DBUF_STAT_BUMP(cache_total_evicts
);
801 multilist_sublist_unlock(mls
);
806 * The dbuf evict thread is responsible for aging out dbufs from the
807 * cache. Once the cache has reached it's maximum size, dbufs are removed
808 * and destroyed. The eviction thread will continue running until the size
809 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
810 * out of the cache it is destroyed and becomes eligible for arc eviction.
812 static __attribute__((noreturn
)) void
813 dbuf_evict_thread(void *unused
)
818 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
820 mutex_enter(&dbuf_evict_lock
);
821 while (!dbuf_evict_thread_exit
) {
822 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
823 CALLB_CPR_SAFE_BEGIN(&cpr
);
824 (void) cv_timedwait_idle_hires(&dbuf_evict_cv
,
825 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
826 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
828 mutex_exit(&dbuf_evict_lock
);
831 * Keep evicting as long as we're above the low water mark
832 * for the cache. We do this without holding the locks to
833 * minimize lock contention.
835 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
839 mutex_enter(&dbuf_evict_lock
);
842 dbuf_evict_thread_exit
= B_FALSE
;
843 cv_broadcast(&dbuf_evict_cv
);
844 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
849 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
850 * If the dbuf cache is at its high water mark, then evict a dbuf from the
851 * dbuf cache using the caller's context.
854 dbuf_evict_notify(uint64_t size
)
857 * We check if we should evict without holding the dbuf_evict_lock,
858 * because it's OK to occasionally make the wrong decision here,
859 * and grabbing the lock results in massive lock contention.
861 if (size
> dbuf_cache_target_bytes()) {
862 if (size
> dbuf_cache_hiwater_bytes())
864 cv_signal(&dbuf_evict_cv
);
869 dbuf_kstat_update(kstat_t
*ksp
, int rw
)
871 dbuf_stats_t
*ds
= ksp
->ks_data
;
872 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
874 if (rw
== KSTAT_WRITE
)
875 return (SET_ERROR(EACCES
));
877 ds
->cache_count
.value
.ui64
=
878 wmsum_value(&dbuf_sums
.cache_count
);
879 ds
->cache_size_bytes
.value
.ui64
=
880 zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
);
881 ds
->cache_target_bytes
.value
.ui64
= dbuf_cache_target_bytes();
882 ds
->cache_hiwater_bytes
.value
.ui64
= dbuf_cache_hiwater_bytes();
883 ds
->cache_lowater_bytes
.value
.ui64
= dbuf_cache_lowater_bytes();
884 ds
->cache_total_evicts
.value
.ui64
=
885 wmsum_value(&dbuf_sums
.cache_total_evicts
);
886 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
887 ds
->cache_levels
[i
].value
.ui64
=
888 wmsum_value(&dbuf_sums
.cache_levels
[i
]);
889 ds
->cache_levels_bytes
[i
].value
.ui64
=
890 wmsum_value(&dbuf_sums
.cache_levels_bytes
[i
]);
892 ds
->hash_hits
.value
.ui64
=
893 wmsum_value(&dbuf_sums
.hash_hits
);
894 ds
->hash_misses
.value
.ui64
=
895 wmsum_value(&dbuf_sums
.hash_misses
);
896 ds
->hash_collisions
.value
.ui64
=
897 wmsum_value(&dbuf_sums
.hash_collisions
);
898 ds
->hash_chains
.value
.ui64
=
899 wmsum_value(&dbuf_sums
.hash_chains
);
900 ds
->hash_insert_race
.value
.ui64
=
901 wmsum_value(&dbuf_sums
.hash_insert_race
);
902 ds
->hash_table_count
.value
.ui64
= h
->hash_table_mask
+ 1;
903 ds
->hash_mutex_count
.value
.ui64
= h
->hash_mutex_mask
+ 1;
904 ds
->metadata_cache_count
.value
.ui64
=
905 wmsum_value(&dbuf_sums
.metadata_cache_count
);
906 ds
->metadata_cache_size_bytes
.value
.ui64
= zfs_refcount_count(
907 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
);
908 ds
->metadata_cache_overflow
.value
.ui64
=
909 wmsum_value(&dbuf_sums
.metadata_cache_overflow
);
916 uint64_t hmsize
, hsize
= 1ULL << 16;
917 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
920 * The hash table is big enough to fill one eighth of physical memory
921 * with an average block size of zfs_arc_average_blocksize (default 8K).
922 * By default, the table will take up
923 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
925 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory() / 8)
928 h
->hash_table
= NULL
;
929 while (h
->hash_table
== NULL
) {
930 h
->hash_table_mask
= hsize
- 1;
932 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
933 if (h
->hash_table
== NULL
)
936 ASSERT3U(hsize
, >=, 1ULL << 10);
940 * The hash table buckets are protected by an array of mutexes where
941 * each mutex is reponsible for protecting 128 buckets. A minimum
942 * array size of 8192 is targeted to avoid contention.
944 if (dbuf_mutex_cache_shift
== 0)
945 hmsize
= MAX(hsize
>> 7, 1ULL << 13);
947 hmsize
= 1ULL << MIN(dbuf_mutex_cache_shift
, 24);
949 h
->hash_mutexes
= NULL
;
950 while (h
->hash_mutexes
== NULL
) {
951 h
->hash_mutex_mask
= hmsize
- 1;
953 h
->hash_mutexes
= vmem_zalloc(hmsize
* sizeof (kmutex_t
),
955 if (h
->hash_mutexes
== NULL
)
959 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
960 sizeof (dmu_buf_impl_t
),
961 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
963 for (int i
= 0; i
< hmsize
; i
++)
964 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_NOLOCKDEP
, NULL
);
969 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
970 * configuration is not required.
972 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
974 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
975 multilist_create(&dbuf_caches
[dcs
].cache
,
976 sizeof (dmu_buf_impl_t
),
977 offsetof(dmu_buf_impl_t
, db_cache_link
),
978 dbuf_cache_multilist_index_func
);
979 zfs_refcount_create(&dbuf_caches
[dcs
].size
);
982 dbuf_evict_thread_exit
= B_FALSE
;
983 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
984 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
985 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
986 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
988 wmsum_init(&dbuf_sums
.cache_count
, 0);
989 wmsum_init(&dbuf_sums
.cache_total_evicts
, 0);
990 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
991 wmsum_init(&dbuf_sums
.cache_levels
[i
], 0);
992 wmsum_init(&dbuf_sums
.cache_levels_bytes
[i
], 0);
994 wmsum_init(&dbuf_sums
.hash_hits
, 0);
995 wmsum_init(&dbuf_sums
.hash_misses
, 0);
996 wmsum_init(&dbuf_sums
.hash_collisions
, 0);
997 wmsum_init(&dbuf_sums
.hash_chains
, 0);
998 wmsum_init(&dbuf_sums
.hash_insert_race
, 0);
999 wmsum_init(&dbuf_sums
.metadata_cache_count
, 0);
1000 wmsum_init(&dbuf_sums
.metadata_cache_overflow
, 0);
1002 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
1003 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
1004 KSTAT_FLAG_VIRTUAL
);
1005 if (dbuf_ksp
!= NULL
) {
1006 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
1007 snprintf(dbuf_stats
.cache_levels
[i
].name
,
1008 KSTAT_STRLEN
, "cache_level_%d", i
);
1009 dbuf_stats
.cache_levels
[i
].data_type
=
1011 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
1012 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
1013 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
1016 dbuf_ksp
->ks_data
= &dbuf_stats
;
1017 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
1018 kstat_install(dbuf_ksp
);
1025 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
1027 dbuf_stats_destroy();
1029 for (int i
= 0; i
< (h
->hash_mutex_mask
+ 1); i
++)
1030 mutex_destroy(&h
->hash_mutexes
[i
]);
1032 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
1033 vmem_free(h
->hash_mutexes
, (h
->hash_mutex_mask
+ 1) *
1036 kmem_cache_destroy(dbuf_kmem_cache
);
1037 taskq_destroy(dbu_evict_taskq
);
1039 mutex_enter(&dbuf_evict_lock
);
1040 dbuf_evict_thread_exit
= B_TRUE
;
1041 while (dbuf_evict_thread_exit
) {
1042 cv_signal(&dbuf_evict_cv
);
1043 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
1045 mutex_exit(&dbuf_evict_lock
);
1047 mutex_destroy(&dbuf_evict_lock
);
1048 cv_destroy(&dbuf_evict_cv
);
1050 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
1051 zfs_refcount_destroy(&dbuf_caches
[dcs
].size
);
1052 multilist_destroy(&dbuf_caches
[dcs
].cache
);
1055 if (dbuf_ksp
!= NULL
) {
1056 kstat_delete(dbuf_ksp
);
1060 wmsum_fini(&dbuf_sums
.cache_count
);
1061 wmsum_fini(&dbuf_sums
.cache_total_evicts
);
1062 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
1063 wmsum_fini(&dbuf_sums
.cache_levels
[i
]);
1064 wmsum_fini(&dbuf_sums
.cache_levels_bytes
[i
]);
1066 wmsum_fini(&dbuf_sums
.hash_hits
);
1067 wmsum_fini(&dbuf_sums
.hash_misses
);
1068 wmsum_fini(&dbuf_sums
.hash_collisions
);
1069 wmsum_fini(&dbuf_sums
.hash_chains
);
1070 wmsum_fini(&dbuf_sums
.hash_insert_race
);
1071 wmsum_fini(&dbuf_sums
.metadata_cache_count
);
1072 wmsum_fini(&dbuf_sums
.metadata_cache_overflow
);
1081 dbuf_verify(dmu_buf_impl_t
*db
)
1084 dbuf_dirty_record_t
*dr
;
1087 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1089 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
1092 ASSERT(db
->db_objset
!= NULL
);
1096 ASSERT(db
->db_parent
== NULL
);
1097 ASSERT(db
->db_blkptr
== NULL
);
1099 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
1100 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
1101 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
1102 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
1103 db
->db_blkid
== DMU_SPILL_BLKID
||
1104 !avl_is_empty(&dn
->dn_dbufs
));
1106 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1108 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1109 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
1110 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
1112 ASSERT0(db
->db
.db_offset
);
1114 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
1117 if ((dr
= list_head(&db
->db_dirty_records
)) != NULL
) {
1118 ASSERT(dr
->dr_dbuf
== db
);
1119 txg_prev
= dr
->dr_txg
;
1120 for (dr
= list_next(&db
->db_dirty_records
, dr
); dr
!= NULL
;
1121 dr
= list_next(&db
->db_dirty_records
, dr
)) {
1122 ASSERT(dr
->dr_dbuf
== db
);
1123 ASSERT(txg_prev
> dr
->dr_txg
);
1124 txg_prev
= dr
->dr_txg
;
1129 * We can't assert that db_size matches dn_datablksz because it
1130 * can be momentarily different when another thread is doing
1131 * dnode_set_blksz().
1133 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
1134 dr
= db
->db_data_pending
;
1136 * It should only be modified in syncing context, so
1137 * make sure we only have one copy of the data.
1139 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
1142 /* verify db->db_blkptr */
1143 if (db
->db_blkptr
) {
1144 if (db
->db_parent
== dn
->dn_dbuf
) {
1145 /* db is pointed to by the dnode */
1146 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1147 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
1148 ASSERT(db
->db_parent
== NULL
);
1150 ASSERT(db
->db_parent
!= NULL
);
1151 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
1152 ASSERT3P(db
->db_blkptr
, ==,
1153 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
1155 /* db is pointed to by an indirect block */
1156 int epb __maybe_unused
= db
->db_parent
->db
.db_size
>>
1158 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
1159 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
1162 * dnode_grow_indblksz() can make this fail if we don't
1163 * have the parent's rwlock. XXX indblksz no longer
1164 * grows. safe to do this now?
1166 if (RW_LOCK_HELD(&db
->db_parent
->db_rwlock
)) {
1167 ASSERT3P(db
->db_blkptr
, ==,
1168 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
1169 db
->db_blkid
% epb
));
1173 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
1174 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
1175 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
1176 db
->db_state
!= DB_FILL
&& (dn
== NULL
|| !dn
->dn_free_txg
)) {
1178 * If the blkptr isn't set but they have nonzero data,
1179 * it had better be dirty, otherwise we'll lose that
1180 * data when we evict this buffer.
1182 * There is an exception to this rule for indirect blocks; in
1183 * this case, if the indirect block is a hole, we fill in a few
1184 * fields on each of the child blocks (importantly, birth time)
1185 * to prevent hole birth times from being lost when you
1186 * partially fill in a hole.
1188 if (db
->db_dirtycnt
== 0) {
1189 if (db
->db_level
== 0) {
1190 uint64_t *buf
= db
->db
.db_data
;
1193 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
1194 ASSERT(buf
[i
] == 0);
1197 blkptr_t
*bps
= db
->db
.db_data
;
1198 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1201 * We want to verify that all the blkptrs in the
1202 * indirect block are holes, but we may have
1203 * automatically set up a few fields for them.
1204 * We iterate through each blkptr and verify
1205 * they only have those fields set.
1208 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1210 blkptr_t
*bp
= &bps
[i
];
1211 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1214 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1215 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1216 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1217 ASSERT0(bp
->blk_fill
);
1218 ASSERT0(bp
->blk_pad
[0]);
1219 ASSERT0(bp
->blk_pad
[1]);
1220 ASSERT(!BP_IS_EMBEDDED(bp
));
1221 ASSERT(BP_IS_HOLE(bp
));
1222 ASSERT0(BP_GET_PHYSICAL_BIRTH(bp
));
1232 dbuf_clear_data(dmu_buf_impl_t
*db
)
1234 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1235 dbuf_evict_user(db
);
1236 ASSERT3P(db
->db_buf
, ==, NULL
);
1237 db
->db
.db_data
= NULL
;
1238 if (db
->db_state
!= DB_NOFILL
) {
1239 db
->db_state
= DB_UNCACHED
;
1240 DTRACE_SET_STATE(db
, "clear data");
1245 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1247 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1248 ASSERT(buf
!= NULL
);
1251 ASSERT(buf
->b_data
!= NULL
);
1252 db
->db
.db_data
= buf
->b_data
;
1256 dbuf_alloc_arcbuf(dmu_buf_impl_t
*db
)
1258 spa_t
*spa
= db
->db_objset
->os_spa
;
1260 return (arc_alloc_buf(spa
, db
, DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
1264 * Loan out an arc_buf for read. Return the loaned arc_buf.
1267 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1271 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1272 mutex_enter(&db
->db_mtx
);
1273 if (arc_released(db
->db_buf
) || zfs_refcount_count(&db
->db_holds
) > 1) {
1274 int blksz
= db
->db
.db_size
;
1275 spa_t
*spa
= db
->db_objset
->os_spa
;
1277 mutex_exit(&db
->db_mtx
);
1278 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1279 memcpy(abuf
->b_data
, db
->db
.db_data
, blksz
);
1282 arc_loan_inuse_buf(abuf
, db
);
1284 dbuf_clear_data(db
);
1285 mutex_exit(&db
->db_mtx
);
1291 * Calculate which level n block references the data at the level 0 offset
1295 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1297 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1299 * The level n blkid is equal to the level 0 blkid divided by
1300 * the number of level 0s in a level n block.
1302 * The level 0 blkid is offset >> datablkshift =
1303 * offset / 2^datablkshift.
1305 * The number of level 0s in a level n is the number of block
1306 * pointers in an indirect block, raised to the power of level.
1307 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1308 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1310 * Thus, the level n blkid is: offset /
1311 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1312 * = offset / 2^(datablkshift + level *
1313 * (indblkshift - SPA_BLKPTRSHIFT))
1314 * = offset >> (datablkshift + level *
1315 * (indblkshift - SPA_BLKPTRSHIFT))
1318 const unsigned exp
= dn
->dn_datablkshift
+
1319 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1321 if (exp
>= 8 * sizeof (offset
)) {
1322 /* This only happens on the highest indirection level */
1323 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1327 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1329 return (offset
>> exp
);
1331 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1337 * This function is used to lock the parent of the provided dbuf. This should be
1338 * used when modifying or reading db_blkptr.
1341 dmu_buf_lock_parent(dmu_buf_impl_t
*db
, krw_t rw
, const void *tag
)
1343 enum db_lock_type ret
= DLT_NONE
;
1344 if (db
->db_parent
!= NULL
) {
1345 rw_enter(&db
->db_parent
->db_rwlock
, rw
);
1347 } else if (dmu_objset_ds(db
->db_objset
) != NULL
) {
1348 rrw_enter(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, rw
,
1353 * We only return a DLT_NONE lock when it's the top-most indirect block
1354 * of the meta-dnode of the MOS.
1360 * We need to pass the lock type in because it's possible that the block will
1361 * move from being the topmost indirect block in a dnode (and thus, have no
1362 * parent) to not the top-most via an indirection increase. This would cause a
1363 * panic if we didn't pass the lock type in.
1366 dmu_buf_unlock_parent(dmu_buf_impl_t
*db
, db_lock_type_t type
, const void *tag
)
1368 if (type
== DLT_PARENT
)
1369 rw_exit(&db
->db_parent
->db_rwlock
);
1370 else if (type
== DLT_OBJSET
)
1371 rrw_exit(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, tag
);
1375 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1376 arc_buf_t
*buf
, void *vdb
)
1378 (void) zb
, (void) bp
;
1379 dmu_buf_impl_t
*db
= vdb
;
1381 mutex_enter(&db
->db_mtx
);
1382 ASSERT3U(db
->db_state
, ==, DB_READ
);
1384 * All reads are synchronous, so we must have a hold on the dbuf
1386 ASSERT(zfs_refcount_count(&db
->db_holds
) > 0);
1387 ASSERT(db
->db_buf
== NULL
);
1388 ASSERT(db
->db
.db_data
== NULL
);
1391 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
1392 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1393 ASSERT3P(db
->db_buf
, ==, NULL
);
1394 db
->db_state
= DB_UNCACHED
;
1395 DTRACE_SET_STATE(db
, "i/o error");
1396 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1397 /* freed in flight */
1398 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1399 arc_release(buf
, db
);
1400 memset(buf
->b_data
, 0, db
->db
.db_size
);
1401 arc_buf_freeze(buf
);
1402 db
->db_freed_in_flight
= FALSE
;
1403 dbuf_set_data(db
, buf
);
1404 db
->db_state
= DB_CACHED
;
1405 DTRACE_SET_STATE(db
, "freed in flight");
1408 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1409 dbuf_set_data(db
, buf
);
1410 db
->db_state
= DB_CACHED
;
1411 DTRACE_SET_STATE(db
, "successful read");
1413 cv_broadcast(&db
->db_changed
);
1414 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
1418 * Shortcut for performing reads on bonus dbufs. Returns
1419 * an error if we fail to verify the dnode associated with
1420 * a decrypted block. Otherwise success.
1423 dbuf_read_bonus(dmu_buf_impl_t
*db
, dnode_t
*dn
)
1425 int bonuslen
, max_bonuslen
;
1427 bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1428 max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1429 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1430 ASSERT(DB_DNODE_HELD(db
));
1431 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1432 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1433 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1434 if (bonuslen
< max_bonuslen
)
1435 memset(db
->db
.db_data
, 0, max_bonuslen
);
1437 memcpy(db
->db
.db_data
, DN_BONUS(dn
->dn_phys
), bonuslen
);
1438 db
->db_state
= DB_CACHED
;
1439 DTRACE_SET_STATE(db
, "bonus buffer filled");
1444 dbuf_handle_indirect_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
, blkptr_t
*dbbp
)
1446 blkptr_t
*bps
= db
->db
.db_data
;
1447 uint32_t indbs
= 1ULL << dn
->dn_indblkshift
;
1448 int n_bps
= indbs
>> SPA_BLKPTRSHIFT
;
1450 for (int i
= 0; i
< n_bps
; i
++) {
1451 blkptr_t
*bp
= &bps
[i
];
1453 ASSERT3U(BP_GET_LSIZE(dbbp
), ==, indbs
);
1454 BP_SET_LSIZE(bp
, BP_GET_LEVEL(dbbp
) == 1 ?
1455 dn
->dn_datablksz
: BP_GET_LSIZE(dbbp
));
1456 BP_SET_TYPE(bp
, BP_GET_TYPE(dbbp
));
1457 BP_SET_LEVEL(bp
, BP_GET_LEVEL(dbbp
) - 1);
1458 BP_SET_BIRTH(bp
, BP_GET_LOGICAL_BIRTH(dbbp
), 0);
1463 * Handle reads on dbufs that are holes, if necessary. This function
1464 * requires that the dbuf's mutex is held. Returns success (0) if action
1465 * was taken, ENOENT if no action was taken.
1468 dbuf_read_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
, blkptr_t
*bp
)
1470 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1472 int is_hole
= bp
== NULL
|| BP_IS_HOLE(bp
);
1474 * For level 0 blocks only, if the above check fails:
1475 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1476 * processes the delete record and clears the bp while we are waiting
1477 * for the dn_mtx (resulting in a "no" from block_freed).
1479 if (!is_hole
&& db
->db_level
== 0)
1480 is_hole
= dnode_block_freed(dn
, db
->db_blkid
) || BP_IS_HOLE(bp
);
1483 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1484 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
1486 if (bp
!= NULL
&& db
->db_level
> 0 && BP_IS_HOLE(bp
) &&
1487 BP_GET_LOGICAL_BIRTH(bp
) != 0) {
1488 dbuf_handle_indirect_hole(db
, dn
, bp
);
1490 db
->db_state
= DB_CACHED
;
1491 DTRACE_SET_STATE(db
, "hole read satisfied");
1498 * This function ensures that, when doing a decrypting read of a block,
1499 * we make sure we have decrypted the dnode associated with it. We must do
1500 * this so that we ensure we are fully authenticating the checksum-of-MACs
1501 * tree from the root of the objset down to this block. Indirect blocks are
1502 * always verified against their secure checksum-of-MACs assuming that the
1503 * dnode containing them is correct. Now that we are doing a decrypting read,
1504 * we can be sure that the key is loaded and verify that assumption. This is
1505 * especially important considering that we always read encrypted dnode
1506 * blocks as raw data (without verifying their MACs) to start, and
1507 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1510 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, dnode_t
*dn
, uint32_t flags
)
1512 objset_t
*os
= db
->db_objset
;
1513 dmu_buf_impl_t
*dndb
;
1515 zbookmark_phys_t zb
;
1518 if ((flags
& DB_RF_NO_DECRYPT
) != 0 ||
1519 !os
->os_encrypted
|| os
->os_raw_receive
||
1520 (dndb
= dn
->dn_dbuf
) == NULL
)
1523 dnbuf
= dndb
->db_buf
;
1524 if (!arc_is_encrypted(dnbuf
))
1527 mutex_enter(&dndb
->db_mtx
);
1530 * Since dnode buffer is modified by sync process, there can be only
1531 * one copy of it. It means we can not modify (decrypt) it while it
1532 * is being written. I don't see how this may happen now, since
1533 * encrypted dnode writes by receive should be completed before any
1534 * plain-text reads due to txg wait, but better be safe than sorry.
1537 if (!arc_is_encrypted(dnbuf
)) {
1538 mutex_exit(&dndb
->db_mtx
);
1541 dbuf_dirty_record_t
*dr
= dndb
->db_data_pending
;
1542 if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= dnbuf
)
1544 cv_wait(&dndb
->db_changed
, &dndb
->db_mtx
);
1547 SET_BOOKMARK(&zb
, dmu_objset_id(os
),
1548 DMU_META_DNODE_OBJECT
, 0, dndb
->db_blkid
);
1549 err
= arc_untransform(dnbuf
, os
->os_spa
, &zb
, B_TRUE
);
1552 * An error code of EACCES tells us that the key is still not
1553 * available. This is ok if we are only reading authenticated
1554 * (and therefore non-encrypted) blocks.
1556 if (err
== EACCES
&& ((db
->db_blkid
!= DMU_BONUS_BLKID
&&
1557 !DMU_OT_IS_ENCRYPTED(dn
->dn_type
)) ||
1558 (db
->db_blkid
== DMU_BONUS_BLKID
&&
1559 !DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))))
1562 mutex_exit(&dndb
->db_mtx
);
1568 * Drops db_mtx and the parent lock specified by dblt and tag before
1572 dbuf_read_impl(dmu_buf_impl_t
*db
, dnode_t
*dn
, zio_t
*zio
, uint32_t flags
,
1573 db_lock_type_t dblt
, const void *tag
)
1575 zbookmark_phys_t zb
;
1576 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1578 blkptr_t bp
, *bpp
= NULL
;
1580 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1581 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1582 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1583 ASSERT(db
->db_buf
== NULL
);
1584 ASSERT(db
->db_parent
== NULL
||
1585 RW_LOCK_HELD(&db
->db_parent
->db_rwlock
));
1587 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1588 err
= dbuf_read_bonus(db
, dn
);
1593 * If we have a pending block clone, we don't want to read the
1594 * underlying block, but the content of the block being cloned,
1595 * pointed by the dirty record, so we have the most recent data.
1596 * If there is no dirty record, then we hit a race in a sync
1597 * process when the dirty record is already removed, while the
1598 * dbuf is not yet destroyed. Such case is equivalent to uncached.
1600 if (db
->db_state
== DB_NOFILL
) {
1601 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
1603 if (!dr
->dt
.dl
.dr_brtwrite
) {
1607 bp
= dr
->dt
.dl
.dr_overridden_by
;
1612 if (bpp
== NULL
&& db
->db_blkptr
!= NULL
) {
1613 bp
= *db
->db_blkptr
;
1617 err
= dbuf_read_hole(db
, dn
, bpp
);
1621 ASSERT(bpp
!= NULL
);
1624 * Any attempt to read a redacted block should result in an error. This
1625 * will never happen under normal conditions, but can be useful for
1626 * debugging purposes.
1628 if (BP_IS_REDACTED(bpp
)) {
1629 ASSERT(dsl_dataset_feature_is_active(
1630 db
->db_objset
->os_dsl_dataset
,
1631 SPA_FEATURE_REDACTED_DATASETS
));
1632 err
= SET_ERROR(EIO
);
1636 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1637 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1640 * All bps of an encrypted os should have the encryption bit set.
1641 * If this is not true it indicates tampering and we report an error.
1643 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(bpp
)) {
1644 spa_log_error(db
->db_objset
->os_spa
, &zb
,
1645 BP_GET_LOGICAL_BIRTH(bpp
));
1646 err
= SET_ERROR(EIO
);
1650 db
->db_state
= DB_READ
;
1651 DTRACE_SET_STATE(db
, "read issued");
1652 mutex_exit(&db
->db_mtx
);
1654 if (!DBUF_IS_CACHEABLE(db
))
1655 aflags
|= ARC_FLAG_UNCACHED
;
1656 else if (dbuf_is_l2cacheable(db
))
1657 aflags
|= ARC_FLAG_L2CACHE
;
1659 dbuf_add_ref(db
, NULL
);
1661 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1662 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1664 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1665 zio_flags
|= ZIO_FLAG_RAW
;
1667 * The zio layer will copy the provided blkptr later, but we have our
1668 * own copy so that we can release the parent's rwlock. We have to
1669 * do that so that if dbuf_read_done is called synchronously (on
1670 * an l1 cache hit) we don't acquire the db_mtx while holding the
1671 * parent's rwlock, which would be a lock ordering violation.
1673 dmu_buf_unlock_parent(db
, dblt
, tag
);
1674 return (arc_read(zio
, db
->db_objset
->os_spa
, bpp
,
1675 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1679 mutex_exit(&db
->db_mtx
);
1680 dmu_buf_unlock_parent(db
, dblt
, tag
);
1685 * This is our just-in-time copy function. It makes a copy of buffers that
1686 * have been modified in a previous transaction group before we access them in
1687 * the current active group.
1689 * This function is used in three places: when we are dirtying a buffer for the
1690 * first time in a txg, when we are freeing a range in a dnode that includes
1691 * this buffer, and when we are accessing a buffer which was received compressed
1692 * and later referenced in a WRITE_BYREF record.
1694 * Note that when we are called from dbuf_free_range() we do not put a hold on
1695 * the buffer, we just traverse the active dbuf list for the dnode.
1698 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1700 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
1702 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1703 ASSERT(db
->db
.db_data
!= NULL
);
1704 ASSERT(db
->db_level
== 0);
1705 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1708 (dr
->dt
.dl
.dr_data
!=
1709 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1713 * If the last dirty record for this dbuf has not yet synced
1714 * and its referencing the dbuf data, either:
1715 * reset the reference to point to a new copy,
1716 * or (if there a no active holders)
1717 * just null out the current db_data pointer.
1719 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1720 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1721 dnode_t
*dn
= DB_DNODE(db
);
1722 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1723 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1724 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1725 memcpy(dr
->dt
.dl
.dr_data
, db
->db
.db_data
, bonuslen
);
1726 } else if (zfs_refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1727 dnode_t
*dn
= DB_DNODE(db
);
1728 int size
= arc_buf_size(db
->db_buf
);
1729 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1730 spa_t
*spa
= db
->db_objset
->os_spa
;
1731 enum zio_compress compress_type
=
1732 arc_get_compression(db
->db_buf
);
1733 uint8_t complevel
= arc_get_complevel(db
->db_buf
);
1735 if (arc_is_encrypted(db
->db_buf
)) {
1736 boolean_t byteorder
;
1737 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1738 uint8_t iv
[ZIO_DATA_IV_LEN
];
1739 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1741 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1743 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1744 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1745 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1746 compress_type
, complevel
);
1747 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1748 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1749 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1750 size
, arc_buf_lsize(db
->db_buf
), compress_type
,
1753 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1755 memcpy(dr
->dt
.dl
.dr_data
->b_data
, db
->db
.db_data
, size
);
1758 dbuf_clear_data(db
);
1763 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*pio
, uint32_t flags
)
1766 boolean_t miss
= B_TRUE
, need_wait
= B_FALSE
, prefetch
;
1769 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1775 * Ensure that this block's dnode has been decrypted if the caller
1776 * has requested decrypted data.
1778 err
= dbuf_read_verify_dnode_crypt(db
, dn
, flags
);
1782 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1783 (flags
& DB_RF_NOPREFETCH
) == 0;
1785 mutex_enter(&db
->db_mtx
);
1786 if (flags
& DB_RF_PARTIAL_FIRST
)
1787 db
->db_partial_read
= B_TRUE
;
1788 else if (!(flags
& DB_RF_PARTIAL_MORE
))
1789 db
->db_partial_read
= B_FALSE
;
1790 miss
= (db
->db_state
!= DB_CACHED
);
1792 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1794 * Another reader came in while the dbuf was in flight between
1795 * UNCACHED and CACHED. Either a writer will finish filling
1796 * the buffer, sending the dbuf to CACHED, or the first reader's
1797 * request will reach the read_done callback and send the dbuf
1798 * to CACHED. Otherwise, a failure occurred and the dbuf will
1799 * be sent to UNCACHED.
1801 if (flags
& DB_RF_NEVERWAIT
) {
1802 mutex_exit(&db
->db_mtx
);
1807 ASSERT(db
->db_state
== DB_READ
||
1808 (flags
& DB_RF_HAVESTRUCT
) == 0);
1809 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*, db
,
1811 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1812 } while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
);
1813 if (db
->db_state
== DB_UNCACHED
) {
1814 err
= SET_ERROR(EIO
);
1815 mutex_exit(&db
->db_mtx
);
1821 if (db
->db_state
== DB_CACHED
) {
1823 * If the arc buf is compressed or encrypted and the caller
1824 * requested uncompressed data, we need to untransform it
1825 * before returning. We also call arc_untransform() on any
1826 * unauthenticated blocks, which will verify their MAC if
1827 * the key is now available.
1829 if ((flags
& DB_RF_NO_DECRYPT
) == 0 && db
->db_buf
!= NULL
&&
1830 (arc_is_encrypted(db
->db_buf
) ||
1831 arc_is_unauthenticated(db
->db_buf
) ||
1832 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1833 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1834 zbookmark_phys_t zb
;
1836 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1837 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1838 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1839 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1840 dbuf_set_data(db
, db
->db_buf
);
1842 mutex_exit(&db
->db_mtx
);
1844 ASSERT(db
->db_state
== DB_UNCACHED
||
1845 db
->db_state
== DB_NOFILL
);
1846 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
1847 if (pio
== NULL
&& (db
->db_state
== DB_NOFILL
||
1848 (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)))) {
1849 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1850 pio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1853 err
= dbuf_read_impl(db
, dn
, pio
, flags
, dblt
, FTAG
);
1854 /* dbuf_read_impl drops db_mtx and parent's rwlock. */
1855 miss
= (db
->db_state
!= DB_CACHED
);
1858 if (err
== 0 && prefetch
) {
1859 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
, miss
,
1860 flags
& DB_RF_HAVESTRUCT
);
1865 * If we created a zio we must execute it to avoid leaking it, even if
1866 * it isn't attached to any work due to an error in dbuf_read_impl().
1870 err
= zio_wait(pio
);
1872 (void) zio_wait(pio
);
1878 DBUF_STAT_BUMP(hash_misses
);
1880 DBUF_STAT_BUMP(hash_hits
);
1881 if (pio
&& err
!= 0) {
1882 zio_t
*zio
= zio_null(pio
, pio
->io_spa
, NULL
, NULL
, NULL
,
1884 zio
->io_error
= err
;
1892 dbuf_noread(dmu_buf_impl_t
*db
)
1894 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1895 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1896 mutex_enter(&db
->db_mtx
);
1897 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1898 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1899 if (db
->db_state
== DB_UNCACHED
) {
1900 ASSERT(db
->db_buf
== NULL
);
1901 ASSERT(db
->db
.db_data
== NULL
);
1902 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1903 db
->db_state
= DB_FILL
;
1904 DTRACE_SET_STATE(db
, "assigning filled buffer");
1905 } else if (db
->db_state
== DB_NOFILL
) {
1906 dbuf_clear_data(db
);
1908 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1910 mutex_exit(&db
->db_mtx
);
1914 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1916 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1917 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1918 uint64_t txg
= dr
->dr_txg
;
1920 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1922 * This assert is valid because dmu_sync() expects to be called by
1923 * a zilog's get_data while holding a range lock. This call only
1924 * comes from dbuf_dirty() callers who must also hold a range lock.
1926 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1927 ASSERT(db
->db_level
== 0);
1929 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1930 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1933 ASSERT(db
->db_data_pending
!= dr
);
1935 /* free this block */
1936 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1937 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1939 if (dr
->dt
.dl
.dr_brtwrite
) {
1940 ASSERT0P(dr
->dt
.dl
.dr_data
);
1941 dr
->dt
.dl
.dr_data
= db
->db_buf
;
1943 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1944 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1945 dr
->dt
.dl
.dr_brtwrite
= B_FALSE
;
1946 dr
->dt
.dl
.dr_has_raw_params
= B_FALSE
;
1949 * Release the already-written buffer, so we leave it in
1950 * a consistent dirty state. Note that all callers are
1951 * modifying the buffer, so they will immediately do
1952 * another (redundant) arc_release(). Therefore, leave
1953 * the buf thawed to save the effort of freezing &
1954 * immediately re-thawing it.
1956 if (dr
->dt
.dl
.dr_data
)
1957 arc_release(dr
->dt
.dl
.dr_data
, db
);
1961 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1962 * data blocks in the free range, so that any future readers will find
1966 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1969 dmu_buf_impl_t
*db_search
;
1970 dmu_buf_impl_t
*db
, *db_next
;
1971 uint64_t txg
= tx
->tx_txg
;
1973 dbuf_dirty_record_t
*dr
;
1975 if (end_blkid
> dn
->dn_maxblkid
&&
1976 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1977 end_blkid
= dn
->dn_maxblkid
;
1978 dprintf_dnode(dn
, "start=%llu end=%llu\n", (u_longlong_t
)start_blkid
,
1979 (u_longlong_t
)end_blkid
);
1981 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1982 db_search
->db_level
= 0;
1983 db_search
->db_blkid
= start_blkid
;
1984 db_search
->db_state
= DB_SEARCH
;
1986 mutex_enter(&dn
->dn_dbufs_mtx
);
1987 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1988 ASSERT3P(db
, ==, NULL
);
1990 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1992 for (; db
!= NULL
; db
= db_next
) {
1993 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1994 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1996 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1999 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
2001 /* found a level 0 buffer in the range */
2002 mutex_enter(&db
->db_mtx
);
2003 if (dbuf_undirty(db
, tx
)) {
2004 /* mutex has been dropped and dbuf destroyed */
2008 if (db
->db_state
== DB_UNCACHED
||
2009 db
->db_state
== DB_NOFILL
||
2010 db
->db_state
== DB_EVICTING
) {
2011 ASSERT(db
->db
.db_data
== NULL
);
2012 mutex_exit(&db
->db_mtx
);
2015 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
2016 /* will be handled in dbuf_read_done or dbuf_rele */
2017 db
->db_freed_in_flight
= TRUE
;
2018 mutex_exit(&db
->db_mtx
);
2021 if (zfs_refcount_count(&db
->db_holds
) == 0) {
2026 /* The dbuf is referenced */
2028 dr
= list_head(&db
->db_dirty_records
);
2030 if (dr
->dr_txg
== txg
) {
2032 * This buffer is "in-use", re-adjust the file
2033 * size to reflect that this buffer may
2034 * contain new data when we sync.
2036 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
2037 db
->db_blkid
> dn
->dn_maxblkid
)
2038 dn
->dn_maxblkid
= db
->db_blkid
;
2039 dbuf_unoverride(dr
);
2042 * This dbuf is not dirty in the open context.
2043 * Either uncache it (if its not referenced in
2044 * the open context) or reset its contents to
2047 dbuf_fix_old_data(db
, txg
);
2050 /* clear the contents if its cached */
2051 if (db
->db_state
== DB_CACHED
) {
2052 ASSERT(db
->db
.db_data
!= NULL
);
2053 arc_release(db
->db_buf
, db
);
2054 rw_enter(&db
->db_rwlock
, RW_WRITER
);
2055 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2056 rw_exit(&db
->db_rwlock
);
2057 arc_buf_freeze(db
->db_buf
);
2060 mutex_exit(&db
->db_mtx
);
2063 mutex_exit(&dn
->dn_dbufs_mtx
);
2064 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
2068 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
2070 arc_buf_t
*buf
, *old_buf
;
2071 dbuf_dirty_record_t
*dr
;
2072 int osize
= db
->db
.db_size
;
2073 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2076 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2082 * XXX we should be doing a dbuf_read, checking the return
2083 * value and returning that up to our callers
2085 dmu_buf_will_dirty(&db
->db
, tx
);
2087 /* create the data buffer for the new block */
2088 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
2090 /* copy old block data to the new block */
2091 old_buf
= db
->db_buf
;
2092 memcpy(buf
->b_data
, old_buf
->b_data
, MIN(osize
, size
));
2093 /* zero the remainder */
2095 memset((uint8_t *)buf
->b_data
+ osize
, 0, size
- osize
);
2097 mutex_enter(&db
->db_mtx
);
2098 dbuf_set_data(db
, buf
);
2099 arc_buf_destroy(old_buf
, db
);
2100 db
->db
.db_size
= size
;
2102 dr
= list_head(&db
->db_dirty_records
);
2103 /* dirty record added by dmu_buf_will_dirty() */
2105 if (db
->db_level
== 0)
2106 dr
->dt
.dl
.dr_data
= buf
;
2107 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2108 ASSERT3U(dr
->dr_accounted
, ==, osize
);
2109 dr
->dr_accounted
= size
;
2110 mutex_exit(&db
->db_mtx
);
2112 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
2117 dbuf_release_bp(dmu_buf_impl_t
*db
)
2119 objset_t
*os __maybe_unused
= db
->db_objset
;
2121 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
2122 ASSERT(arc_released(os
->os_phys_buf
) ||
2123 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
2124 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
2126 (void) arc_release(db
->db_buf
, db
);
2130 * We already have a dirty record for this TXG, and we are being
2134 dbuf_redirty(dbuf_dirty_record_t
*dr
)
2136 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2138 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2140 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
2142 * If this buffer has already been written out,
2143 * we now need to reset its state.
2145 dbuf_unoverride(dr
);
2146 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
2147 db
->db_state
!= DB_NOFILL
) {
2148 /* Already released on initial dirty, so just thaw. */
2149 ASSERT(arc_released(db
->db_buf
));
2150 arc_buf_thaw(db
->db_buf
);
2155 dbuf_dirty_record_t
*
2156 dbuf_dirty_lightweight(dnode_t
*dn
, uint64_t blkid
, dmu_tx_t
*tx
)
2158 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2159 IMPLY(dn
->dn_objset
->os_raw_receive
, dn
->dn_maxblkid
>= blkid
);
2160 dnode_new_blkid(dn
, blkid
, tx
, B_TRUE
, B_FALSE
);
2161 ASSERT(dn
->dn_maxblkid
>= blkid
);
2163 dbuf_dirty_record_t
*dr
= kmem_zalloc(sizeof (*dr
), KM_SLEEP
);
2164 list_link_init(&dr
->dr_dirty_node
);
2165 list_link_init(&dr
->dr_dbuf_node
);
2167 dr
->dr_txg
= tx
->tx_txg
;
2168 dr
->dt
.dll
.dr_blkid
= blkid
;
2169 dr
->dr_accounted
= dn
->dn_datablksz
;
2172 * There should not be any dbuf for the block that we're dirtying.
2173 * Otherwise the buffer contents could be inconsistent between the
2174 * dbuf and the lightweight dirty record.
2176 ASSERT3P(NULL
, ==, dbuf_find(dn
->dn_objset
, dn
->dn_object
, 0, blkid
,
2179 mutex_enter(&dn
->dn_mtx
);
2180 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2181 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2182 range_tree_clear(dn
->dn_free_ranges
[txgoff
], blkid
, 1);
2185 if (dn
->dn_nlevels
== 1) {
2186 ASSERT3U(blkid
, <, dn
->dn_nblkptr
);
2187 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2188 mutex_exit(&dn
->dn_mtx
);
2189 rw_exit(&dn
->dn_struct_rwlock
);
2190 dnode_setdirty(dn
, tx
);
2192 mutex_exit(&dn
->dn_mtx
);
2194 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2195 dmu_buf_impl_t
*parent_db
= dbuf_hold_level(dn
,
2196 1, blkid
>> epbs
, FTAG
);
2197 rw_exit(&dn
->dn_struct_rwlock
);
2198 if (parent_db
== NULL
) {
2199 kmem_free(dr
, sizeof (*dr
));
2202 int err
= dbuf_read(parent_db
, NULL
,
2203 (DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2205 dbuf_rele(parent_db
, FTAG
);
2206 kmem_free(dr
, sizeof (*dr
));
2210 dbuf_dirty_record_t
*parent_dr
= dbuf_dirty(parent_db
, tx
);
2211 dbuf_rele(parent_db
, FTAG
);
2212 mutex_enter(&parent_dr
->dt
.di
.dr_mtx
);
2213 ASSERT3U(parent_dr
->dr_txg
, ==, tx
->tx_txg
);
2214 list_insert_tail(&parent_dr
->dt
.di
.dr_children
, dr
);
2215 mutex_exit(&parent_dr
->dt
.di
.dr_mtx
);
2216 dr
->dr_parent
= parent_dr
;
2219 dmu_objset_willuse_space(dn
->dn_objset
, dr
->dr_accounted
, tx
);
2224 dbuf_dirty_record_t
*
2225 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2229 dbuf_dirty_record_t
*dr
, *dr_next
, *dr_head
;
2230 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2231 boolean_t drop_struct_rwlock
= B_FALSE
;
2233 ASSERT(tx
->tx_txg
!= 0);
2234 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2235 DMU_TX_DIRTY_BUF(tx
, db
);
2240 * Shouldn't dirty a regular buffer in syncing context. Private
2241 * objects may be dirtied in syncing context, but only if they
2242 * were already pre-dirtied in open context.
2245 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
2246 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
2249 ASSERT(!dmu_tx_is_syncing(tx
) ||
2250 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
2251 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2252 dn
->dn_objset
->os_dsl_dataset
== NULL
);
2253 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2254 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2257 * We make this assert for private objects as well, but after we
2258 * check if we're already dirty. They are allowed to re-dirty
2259 * in syncing context.
2261 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2262 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2263 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2265 mutex_enter(&db
->db_mtx
);
2267 * XXX make this true for indirects too? The problem is that
2268 * transactions created with dmu_tx_create_assigned() from
2269 * syncing context don't bother holding ahead.
2271 ASSERT(db
->db_level
!= 0 ||
2272 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
2273 db
->db_state
== DB_NOFILL
);
2275 mutex_enter(&dn
->dn_mtx
);
2276 dnode_set_dirtyctx(dn
, tx
, db
);
2277 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
2278 dn
->dn_dirty_txg
= tx
->tx_txg
;
2279 mutex_exit(&dn
->dn_mtx
);
2281 if (db
->db_blkid
== DMU_SPILL_BLKID
)
2282 dn
->dn_have_spill
= B_TRUE
;
2285 * If this buffer is already dirty, we're done.
2287 dr_head
= list_head(&db
->db_dirty_records
);
2288 ASSERT(dr_head
== NULL
|| dr_head
->dr_txg
<= tx
->tx_txg
||
2289 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
2290 dr_next
= dbuf_find_dirty_lte(db
, tx
->tx_txg
);
2291 if (dr_next
&& dr_next
->dr_txg
== tx
->tx_txg
) {
2294 dbuf_redirty(dr_next
);
2295 mutex_exit(&db
->db_mtx
);
2300 * Only valid if not already dirty.
2302 ASSERT(dn
->dn_object
== 0 ||
2303 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2304 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2306 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
2309 * We should only be dirtying in syncing context if it's the
2310 * mos or we're initializing the os or it's a special object.
2311 * However, we are allowed to dirty in syncing context provided
2312 * we already dirtied it in open context. Hence we must make
2313 * this assertion only if we're not already dirty.
2316 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
2318 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2319 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
2320 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2321 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
2322 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2323 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2325 ASSERT(db
->db
.db_size
!= 0);
2327 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2329 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& db
->db_state
!= DB_NOFILL
) {
2330 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
2334 * If this buffer is dirty in an old transaction group we need
2335 * to make a copy of it so that the changes we make in this
2336 * transaction group won't leak out when we sync the older txg.
2338 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
2339 list_link_init(&dr
->dr_dirty_node
);
2340 list_link_init(&dr
->dr_dbuf_node
);
2342 if (db
->db_level
== 0) {
2343 void *data_old
= db
->db_buf
;
2345 if (db
->db_state
!= DB_NOFILL
) {
2346 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2347 dbuf_fix_old_data(db
, tx
->tx_txg
);
2348 data_old
= db
->db
.db_data
;
2349 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
2351 * Release the data buffer from the cache so
2352 * that we can modify it without impacting
2353 * possible other users of this cached data
2354 * block. Note that indirect blocks and
2355 * private objects are not released until the
2356 * syncing state (since they are only modified
2359 arc_release(db
->db_buf
, db
);
2360 dbuf_fix_old_data(db
, tx
->tx_txg
);
2361 data_old
= db
->db_buf
;
2363 ASSERT(data_old
!= NULL
);
2365 dr
->dt
.dl
.dr_data
= data_old
;
2367 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
2368 list_create(&dr
->dt
.di
.dr_children
,
2369 sizeof (dbuf_dirty_record_t
),
2370 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
2372 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& db
->db_state
!= DB_NOFILL
) {
2373 dr
->dr_accounted
= db
->db
.db_size
;
2376 dr
->dr_txg
= tx
->tx_txg
;
2377 list_insert_before(&db
->db_dirty_records
, dr_next
, dr
);
2380 * We could have been freed_in_flight between the dbuf_noread
2381 * and dbuf_dirty. We win, as though the dbuf_noread() had
2382 * happened after the free.
2384 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2385 db
->db_blkid
!= DMU_SPILL_BLKID
) {
2386 mutex_enter(&dn
->dn_mtx
);
2387 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2388 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
2391 mutex_exit(&dn
->dn_mtx
);
2392 db
->db_freed_in_flight
= FALSE
;
2396 * This buffer is now part of this txg
2398 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
2399 db
->db_dirtycnt
+= 1;
2400 ASSERT3U(db
->db_dirtycnt
, <=, 3);
2402 mutex_exit(&db
->db_mtx
);
2404 if (db
->db_blkid
== DMU_BONUS_BLKID
||
2405 db
->db_blkid
== DMU_SPILL_BLKID
) {
2406 mutex_enter(&dn
->dn_mtx
);
2407 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2408 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2409 mutex_exit(&dn
->dn_mtx
);
2410 dnode_setdirty(dn
, tx
);
2415 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
2416 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2417 drop_struct_rwlock
= B_TRUE
;
2421 * If we are overwriting a dedup BP, then unless it is snapshotted,
2422 * when we get to syncing context we will need to decrement its
2423 * refcount in the DDT. Prefetch the relevant DDT block so that
2424 * syncing context won't have to wait for the i/o.
2426 if (db
->db_blkptr
!= NULL
) {
2427 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
2428 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
2429 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
2433 * We need to hold the dn_struct_rwlock to make this assertion,
2434 * because it protects dn_phys / dn_next_nlevels from changing.
2436 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
2437 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
2438 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
2439 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
2440 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
2443 if (db
->db_level
== 0) {
2444 ASSERT(!db
->db_objset
->os_raw_receive
||
2445 dn
->dn_maxblkid
>= db
->db_blkid
);
2446 dnode_new_blkid(dn
, db
->db_blkid
, tx
,
2447 drop_struct_rwlock
, B_FALSE
);
2448 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
2451 if (db
->db_level
+1 < dn
->dn_nlevels
) {
2452 dmu_buf_impl_t
*parent
= db
->db_parent
;
2453 dbuf_dirty_record_t
*di
;
2454 int parent_held
= FALSE
;
2456 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
2457 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2458 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2459 db
->db_blkid
>> epbs
, FTAG
);
2460 ASSERT(parent
!= NULL
);
2463 if (drop_struct_rwlock
)
2464 rw_exit(&dn
->dn_struct_rwlock
);
2465 ASSERT3U(db
->db_level
+ 1, ==, parent
->db_level
);
2466 di
= dbuf_dirty(parent
, tx
);
2468 dbuf_rele(parent
, FTAG
);
2470 mutex_enter(&db
->db_mtx
);
2472 * Since we've dropped the mutex, it's possible that
2473 * dbuf_undirty() might have changed this out from under us.
2475 if (list_head(&db
->db_dirty_records
) == dr
||
2476 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2477 mutex_enter(&di
->dt
.di
.dr_mtx
);
2478 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2479 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2480 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2481 mutex_exit(&di
->dt
.di
.dr_mtx
);
2484 mutex_exit(&db
->db_mtx
);
2486 ASSERT(db
->db_level
+ 1 == dn
->dn_nlevels
);
2487 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2488 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2489 mutex_enter(&dn
->dn_mtx
);
2490 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2491 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2492 mutex_exit(&dn
->dn_mtx
);
2493 if (drop_struct_rwlock
)
2494 rw_exit(&dn
->dn_struct_rwlock
);
2497 dnode_setdirty(dn
, tx
);
2503 dbuf_undirty_bonus(dbuf_dirty_record_t
*dr
)
2505 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2507 if (dr
->dt
.dl
.dr_data
!= db
->db
.db_data
) {
2508 struct dnode
*dn
= dr
->dr_dnode
;
2509 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
2511 kmem_free(dr
->dt
.dl
.dr_data
, max_bonuslen
);
2512 arc_space_return(max_bonuslen
, ARC_SPACE_BONUS
);
2514 db
->db_data_pending
= NULL
;
2515 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
2516 list_remove(&db
->db_dirty_records
, dr
);
2517 if (dr
->dr_dbuf
->db_level
!= 0) {
2518 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2519 list_destroy(&dr
->dt
.di
.dr_children
);
2521 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2522 ASSERT3U(db
->db_dirtycnt
, >, 0);
2523 db
->db_dirtycnt
-= 1;
2527 * Undirty a buffer in the transaction group referenced by the given
2528 * transaction. Return whether this evicted the dbuf.
2531 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2533 uint64_t txg
= tx
->tx_txg
;
2539 * Due to our use of dn_nlevels below, this can only be called
2540 * in open context, unless we are operating on the MOS.
2541 * From syncing context, dn_nlevels may be different from the
2542 * dn_nlevels used when dbuf was dirtied.
2544 ASSERT(db
->db_objset
==
2545 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2546 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2547 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2548 ASSERT0(db
->db_level
);
2549 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2552 * If this buffer is not dirty, we're done.
2554 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, txg
);
2557 ASSERT(dr
->dr_dbuf
== db
);
2559 brtwrite
= dr
->dt
.dl
.dr_brtwrite
;
2562 * We are freeing a block that we cloned in the same
2563 * transaction group.
2565 brt_pending_remove(dmu_objset_spa(db
->db_objset
),
2566 &dr
->dt
.dl
.dr_overridden_by
, tx
);
2569 dnode_t
*dn
= dr
->dr_dnode
;
2571 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2573 ASSERT(db
->db
.db_size
!= 0);
2575 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2576 dr
->dr_accounted
, txg
);
2578 list_remove(&db
->db_dirty_records
, dr
);
2581 * Note that there are three places in dbuf_dirty()
2582 * where this dirty record may be put on a list.
2583 * Make sure to do a list_remove corresponding to
2584 * every one of those list_insert calls.
2586 if (dr
->dr_parent
) {
2587 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2588 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2589 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2590 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2591 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2592 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2593 mutex_enter(&dn
->dn_mtx
);
2594 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2595 mutex_exit(&dn
->dn_mtx
);
2598 if (db
->db_state
!= DB_NOFILL
&& !brtwrite
) {
2599 dbuf_unoverride(dr
);
2601 ASSERT(db
->db_buf
!= NULL
);
2602 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2603 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2604 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2607 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2609 ASSERT(db
->db_dirtycnt
> 0);
2610 db
->db_dirtycnt
-= 1;
2612 if (zfs_refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2613 ASSERT(db
->db_state
== DB_NOFILL
|| brtwrite
||
2614 arc_released(db
->db_buf
));
2623 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2625 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2626 boolean_t undirty
= B_FALSE
;
2628 ASSERT(tx
->tx_txg
!= 0);
2629 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2632 * Quick check for dirtiness to improve performance for some workloads
2633 * (e.g. file deletion with indirect blocks cached).
2635 mutex_enter(&db
->db_mtx
);
2636 if (db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
) {
2638 * It's possible that the dbuf is already dirty but not cached,
2639 * because there are some calls to dbuf_dirty() that don't
2640 * go through dmu_buf_will_dirty().
2642 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2644 if (db
->db_level
== 0 &&
2645 dr
->dt
.dl
.dr_brtwrite
) {
2647 * Block cloning: If we are dirtying a cloned
2648 * level 0 block, we cannot simply redirty it,
2649 * because this dr has no associated data.
2650 * We will go through a full undirtying below,
2651 * before dirtying it again.
2655 /* This dbuf is already dirty and cached. */
2657 mutex_exit(&db
->db_mtx
);
2662 mutex_exit(&db
->db_mtx
);
2665 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2666 flags
|= DB_RF_HAVESTRUCT
;
2670 * Block cloning: Do the dbuf_read() before undirtying the dbuf, as we
2671 * want to make sure dbuf_read() will read the pending cloned block and
2672 * not the uderlying block that is being replaced. dbuf_undirty() will
2673 * do dbuf_unoverride(), so we will end up with cloned block content,
2674 * without overridden BP.
2676 (void) dbuf_read(db
, NULL
, flags
);
2678 mutex_enter(&db
->db_mtx
);
2679 VERIFY(!dbuf_undirty(db
, tx
));
2680 mutex_exit(&db
->db_mtx
);
2682 (void) dbuf_dirty(db
, tx
);
2686 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2688 dmu_buf_will_dirty_impl(db_fake
,
2689 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2693 dmu_buf_is_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2695 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2696 dbuf_dirty_record_t
*dr
;
2698 mutex_enter(&db
->db_mtx
);
2699 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2700 mutex_exit(&db
->db_mtx
);
2701 return (dr
!= NULL
);
2705 dmu_buf_will_clone(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2707 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2710 * Block cloning: We are going to clone into this block, so undirty
2711 * modifications done to this block so far in this txg. This includes
2712 * writes and clones into this block.
2714 mutex_enter(&db
->db_mtx
);
2716 VERIFY(!dbuf_undirty(db
, tx
));
2717 ASSERT0P(dbuf_find_dirty_eq(db
, tx
->tx_txg
));
2718 if (db
->db_buf
!= NULL
) {
2719 arc_buf_destroy(db
->db_buf
, db
);
2721 dbuf_clear_data(db
);
2724 db
->db_state
= DB_NOFILL
;
2725 DTRACE_SET_STATE(db
, "allocating NOFILL buffer for clone");
2728 mutex_exit(&db
->db_mtx
);
2731 (void) dbuf_dirty(db
, tx
);
2735 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2737 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2739 mutex_enter(&db
->db_mtx
);
2740 db
->db_state
= DB_NOFILL
;
2741 DTRACE_SET_STATE(db
, "allocating NOFILL buffer");
2742 mutex_exit(&db
->db_mtx
);
2745 (void) dbuf_dirty(db
, tx
);
2749 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
, boolean_t canfail
)
2751 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2753 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2754 ASSERT(tx
->tx_txg
!= 0);
2755 ASSERT(db
->db_level
== 0);
2756 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2758 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2759 dmu_tx_private_ok(tx
));
2761 mutex_enter(&db
->db_mtx
);
2762 if (db
->db_state
== DB_NOFILL
) {
2764 * Block cloning: We will be completely overwriting a block
2765 * cloned in this transaction group, so let's undirty the
2766 * pending clone and mark the block as uncached. This will be
2767 * as if the clone was never done. But if the fill can fail
2768 * we should have a way to return back to the cloned data.
2770 if (canfail
&& dbuf_find_dirty_eq(db
, tx
->tx_txg
) != NULL
) {
2771 mutex_exit(&db
->db_mtx
);
2772 dmu_buf_will_dirty(db_fake
, tx
);
2775 VERIFY(!dbuf_undirty(db
, tx
));
2776 db
->db_state
= DB_UNCACHED
;
2778 mutex_exit(&db
->db_mtx
);
2781 (void) dbuf_dirty(db
, tx
);
2785 * This function is effectively the same as dmu_buf_will_dirty(), but
2786 * indicates the caller expects raw encrypted data in the db, and provides
2787 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2788 * blkptr_t when this dbuf is written. This is only used for blocks of
2789 * dnodes, during raw receive.
2792 dmu_buf_set_crypt_params(dmu_buf_t
*db_fake
, boolean_t byteorder
,
2793 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
, dmu_tx_t
*tx
)
2795 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2796 dbuf_dirty_record_t
*dr
;
2799 * dr_has_raw_params is only processed for blocks of dnodes
2800 * (see dbuf_sync_dnode_leaf_crypt()).
2802 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
2803 ASSERT3U(db
->db_level
, ==, 0);
2804 ASSERT(db
->db_objset
->os_raw_receive
);
2806 dmu_buf_will_dirty_impl(db_fake
,
2807 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2809 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2811 ASSERT3P(dr
, !=, NULL
);
2813 dr
->dt
.dl
.dr_has_raw_params
= B_TRUE
;
2814 dr
->dt
.dl
.dr_byteorder
= byteorder
;
2815 memcpy(dr
->dt
.dl
.dr_salt
, salt
, ZIO_DATA_SALT_LEN
);
2816 memcpy(dr
->dt
.dl
.dr_iv
, iv
, ZIO_DATA_IV_LEN
);
2817 memcpy(dr
->dt
.dl
.dr_mac
, mac
, ZIO_DATA_MAC_LEN
);
2821 dbuf_override_impl(dmu_buf_impl_t
*db
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
2823 struct dirty_leaf
*dl
;
2824 dbuf_dirty_record_t
*dr
;
2826 dr
= list_head(&db
->db_dirty_records
);
2827 ASSERT3P(dr
, !=, NULL
);
2828 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2830 dl
->dr_overridden_by
= *bp
;
2831 dl
->dr_override_state
= DR_OVERRIDDEN
;
2832 BP_SET_LOGICAL_BIRTH(&dl
->dr_overridden_by
, dr
->dr_txg
);
2836 dmu_buf_fill_done(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
, boolean_t failed
)
2839 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2840 mutex_enter(&db
->db_mtx
);
2843 if (db
->db_state
== DB_FILL
) {
2844 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2845 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2846 /* we were freed while filling */
2847 /* XXX dbuf_undirty? */
2848 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2849 db
->db_freed_in_flight
= FALSE
;
2850 db
->db_state
= DB_CACHED
;
2851 DTRACE_SET_STATE(db
,
2852 "fill done handling freed in flight");
2854 } else if (failed
) {
2855 VERIFY(!dbuf_undirty(db
, tx
));
2857 dbuf_clear_data(db
);
2858 DTRACE_SET_STATE(db
, "fill failed");
2860 db
->db_state
= DB_CACHED
;
2861 DTRACE_SET_STATE(db
, "fill done");
2863 cv_broadcast(&db
->db_changed
);
2865 db
->db_state
= DB_CACHED
;
2868 mutex_exit(&db
->db_mtx
);
2873 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2874 bp_embedded_type_t etype
, enum zio_compress comp
,
2875 int uncompressed_size
, int compressed_size
, int byteorder
,
2878 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2879 struct dirty_leaf
*dl
;
2880 dmu_object_type_t type
;
2881 dbuf_dirty_record_t
*dr
;
2883 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2884 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2885 SPA_FEATURE_EMBEDDED_DATA
));
2889 type
= DB_DNODE(db
)->dn_type
;
2892 ASSERT0(db
->db_level
);
2893 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2895 dmu_buf_will_not_fill(dbuf
, tx
);
2897 dr
= list_head(&db
->db_dirty_records
);
2898 ASSERT3P(dr
, !=, NULL
);
2899 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2901 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2902 data
, comp
, uncompressed_size
, compressed_size
);
2903 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2904 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2905 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2906 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2908 dl
->dr_override_state
= DR_OVERRIDDEN
;
2909 BP_SET_LOGICAL_BIRTH(&dl
->dr_overridden_by
, dr
->dr_txg
);
2913 dmu_buf_redact(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2915 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2916 dmu_object_type_t type
;
2917 ASSERT(dsl_dataset_feature_is_active(db
->db_objset
->os_dsl_dataset
,
2918 SPA_FEATURE_REDACTED_DATASETS
));
2921 type
= DB_DNODE(db
)->dn_type
;
2924 ASSERT0(db
->db_level
);
2925 dmu_buf_will_not_fill(dbuf
, tx
);
2927 blkptr_t bp
= { { { {0} } } };
2928 BP_SET_TYPE(&bp
, type
);
2929 BP_SET_LEVEL(&bp
, 0);
2930 BP_SET_BIRTH(&bp
, tx
->tx_txg
, 0);
2931 BP_SET_REDACTED(&bp
);
2932 BPE_SET_LSIZE(&bp
, dbuf
->db_size
);
2934 dbuf_override_impl(db
, &bp
, tx
);
2938 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2939 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2942 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2944 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2945 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2946 ASSERT(db
->db_level
== 0);
2947 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2948 ASSERT(buf
!= NULL
);
2949 ASSERT3U(arc_buf_lsize(buf
), ==, db
->db
.db_size
);
2950 ASSERT(tx
->tx_txg
!= 0);
2952 arc_return_buf(buf
, db
);
2953 ASSERT(arc_released(buf
));
2955 mutex_enter(&db
->db_mtx
);
2957 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2958 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2960 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
||
2961 db
->db_state
== DB_NOFILL
);
2963 if (db
->db_state
== DB_CACHED
&&
2964 zfs_refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2966 * In practice, we will never have a case where we have an
2967 * encrypted arc buffer while additional holds exist on the
2968 * dbuf. We don't handle this here so we simply assert that
2971 ASSERT(!arc_is_encrypted(buf
));
2972 mutex_exit(&db
->db_mtx
);
2973 (void) dbuf_dirty(db
, tx
);
2974 memcpy(db
->db
.db_data
, buf
->b_data
, db
->db
.db_size
);
2975 arc_buf_destroy(buf
, db
);
2979 if (db
->db_state
== DB_CACHED
) {
2980 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
2982 ASSERT(db
->db_buf
!= NULL
);
2983 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2984 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2986 if (!arc_released(db
->db_buf
)) {
2987 ASSERT(dr
->dt
.dl
.dr_override_state
==
2989 arc_release(db
->db_buf
, db
);
2991 dr
->dt
.dl
.dr_data
= buf
;
2992 arc_buf_destroy(db
->db_buf
, db
);
2993 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2994 arc_release(db
->db_buf
, db
);
2995 arc_buf_destroy(db
->db_buf
, db
);
2998 } else if (db
->db_state
== DB_NOFILL
) {
3000 * We will be completely replacing the cloned block. In case
3001 * it was cloned in this transaction group, let's undirty the
3002 * pending clone and mark the block as uncached. This will be
3003 * as if the clone was never done.
3005 VERIFY(!dbuf_undirty(db
, tx
));
3006 db
->db_state
= DB_UNCACHED
;
3008 ASSERT(db
->db_buf
== NULL
);
3009 dbuf_set_data(db
, buf
);
3010 db
->db_state
= DB_FILL
;
3011 DTRACE_SET_STATE(db
, "filling assigned arcbuf");
3012 mutex_exit(&db
->db_mtx
);
3013 (void) dbuf_dirty(db
, tx
);
3014 dmu_buf_fill_done(&db
->db
, tx
, B_FALSE
);
3018 dbuf_destroy(dmu_buf_impl_t
*db
)
3021 dmu_buf_impl_t
*parent
= db
->db_parent
;
3022 dmu_buf_impl_t
*dndb
;
3024 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3025 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3027 if (db
->db_buf
!= NULL
) {
3028 arc_buf_destroy(db
->db_buf
, db
);
3032 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3033 int slots
= DB_DNODE(db
)->dn_num_slots
;
3034 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3035 if (db
->db
.db_data
!= NULL
) {
3036 kmem_free(db
->db
.db_data
, bonuslen
);
3037 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3038 db
->db_state
= DB_UNCACHED
;
3039 DTRACE_SET_STATE(db
, "buffer cleared");
3043 dbuf_clear_data(db
);
3045 if (multilist_link_active(&db
->db_cache_link
)) {
3046 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3047 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3049 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
3051 ASSERT0(dmu_buf_user_size(&db
->db
));
3052 (void) zfs_refcount_remove_many(
3053 &dbuf_caches
[db
->db_caching_status
].size
,
3054 db
->db
.db_size
, db
);
3056 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3057 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3059 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3060 DBUF_STAT_BUMPDOWN(cache_count
);
3061 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3064 db
->db_caching_status
= DB_NO_CACHE
;
3067 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
3068 ASSERT(db
->db_data_pending
== NULL
);
3069 ASSERT(list_is_empty(&db
->db_dirty_records
));
3071 db
->db_state
= DB_EVICTING
;
3072 DTRACE_SET_STATE(db
, "buffer eviction started");
3073 db
->db_blkptr
= NULL
;
3076 * Now that db_state is DB_EVICTING, nobody else can find this via
3077 * the hash table. We can now drop db_mtx, which allows us to
3078 * acquire the dn_dbufs_mtx.
3080 mutex_exit(&db
->db_mtx
);
3085 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
3086 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
3088 mutex_enter_nested(&dn
->dn_dbufs_mtx
,
3090 avl_remove(&dn
->dn_dbufs
, db
);
3094 mutex_exit(&dn
->dn_dbufs_mtx
);
3096 * Decrementing the dbuf count means that the hold corresponding
3097 * to the removed dbuf is no longer discounted in dnode_move(),
3098 * so the dnode cannot be moved until after we release the hold.
3099 * The membar_producer() ensures visibility of the decremented
3100 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
3103 mutex_enter(&dn
->dn_mtx
);
3104 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
3105 db
->db_dnode_handle
= NULL
;
3107 dbuf_hash_remove(db
);
3112 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3114 db
->db_parent
= NULL
;
3116 ASSERT(db
->db_buf
== NULL
);
3117 ASSERT(db
->db
.db_data
== NULL
);
3118 ASSERT(db
->db_hash_next
== NULL
);
3119 ASSERT(db
->db_blkptr
== NULL
);
3120 ASSERT(db
->db_data_pending
== NULL
);
3121 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
3122 ASSERT(!multilist_link_active(&db
->db_cache_link
));
3125 * If this dbuf is referenced from an indirect dbuf,
3126 * decrement the ref count on the indirect dbuf.
3128 if (parent
&& parent
!= dndb
) {
3129 mutex_enter(&parent
->db_mtx
);
3130 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
3133 kmem_cache_free(dbuf_kmem_cache
, db
);
3134 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3138 * Note: While bpp will always be updated if the function returns success,
3139 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
3140 * this happens when the dnode is the meta-dnode, or {user|group|project}used
3143 __attribute__((always_inline
))
3145 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
3146 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
3151 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3153 if (blkid
== DMU_SPILL_BLKID
) {
3154 mutex_enter(&dn
->dn_mtx
);
3155 if (dn
->dn_have_spill
&&
3156 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
3157 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3160 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3161 *parentp
= dn
->dn_dbuf
;
3162 mutex_exit(&dn
->dn_mtx
);
3167 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
3168 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3170 ASSERT3U(level
* epbs
, <, 64);
3171 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3173 * This assertion shouldn't trip as long as the max indirect block size
3174 * is less than 1M. The reason for this is that up to that point,
3175 * the number of levels required to address an entire object with blocks
3176 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
3177 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
3178 * (i.e. we can address the entire object), objects will all use at most
3179 * N-1 levels and the assertion won't overflow. However, once epbs is
3180 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
3181 * enough to address an entire object, so objects will have 5 levels,
3182 * but then this assertion will overflow.
3184 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
3185 * need to redo this logic to handle overflows.
3187 ASSERT(level
>= nlevels
||
3188 ((nlevels
- level
- 1) * epbs
) +
3189 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
3190 if (level
>= nlevels
||
3191 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
3192 ((nlevels
- level
- 1) * epbs
)) ||
3194 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
3195 /* the buffer has no parent yet */
3196 return (SET_ERROR(ENOENT
));
3197 } else if (level
< nlevels
-1) {
3198 /* this block is referenced from an indirect block */
3201 err
= dbuf_hold_impl(dn
, level
+ 1,
3202 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
3206 err
= dbuf_read(*parentp
, NULL
,
3207 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
3209 dbuf_rele(*parentp
, NULL
);
3213 rw_enter(&(*parentp
)->db_rwlock
, RW_READER
);
3214 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
3215 (blkid
& ((1ULL << epbs
) - 1));
3216 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
3217 ASSERT(BP_IS_HOLE(*bpp
));
3218 rw_exit(&(*parentp
)->db_rwlock
);
3221 /* the block is referenced from the dnode */
3222 ASSERT3U(level
, ==, nlevels
-1);
3223 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
3224 blkid
< dn
->dn_phys
->dn_nblkptr
);
3226 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3227 *parentp
= dn
->dn_dbuf
;
3229 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
3234 static dmu_buf_impl_t
*
3235 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3236 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
, uint64_t hash
)
3238 objset_t
*os
= dn
->dn_objset
;
3239 dmu_buf_impl_t
*db
, *odb
;
3241 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3242 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
3244 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
3246 list_create(&db
->db_dirty_records
, sizeof (dbuf_dirty_record_t
),
3247 offsetof(dbuf_dirty_record_t
, dr_dbuf_node
));
3250 db
->db
.db_object
= dn
->dn_object
;
3251 db
->db_level
= level
;
3252 db
->db_blkid
= blkid
;
3253 db
->db_dirtycnt
= 0;
3254 db
->db_dnode_handle
= dn
->dn_handle
;
3255 db
->db_parent
= parent
;
3256 db
->db_blkptr
= blkptr
;
3260 db
->db_user_immediate_evict
= FALSE
;
3261 db
->db_freed_in_flight
= FALSE
;
3262 db
->db_pending_evict
= FALSE
;
3264 if (blkid
== DMU_BONUS_BLKID
) {
3265 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
3266 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
3267 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
3268 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
3269 db
->db
.db_offset
= DMU_BONUS_BLKID
;
3270 db
->db_state
= DB_UNCACHED
;
3271 DTRACE_SET_STATE(db
, "bonus buffer created");
3272 db
->db_caching_status
= DB_NO_CACHE
;
3273 /* the bonus dbuf is not placed in the hash table */
3274 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3276 } else if (blkid
== DMU_SPILL_BLKID
) {
3277 db
->db
.db_size
= (blkptr
!= NULL
) ?
3278 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
3279 db
->db
.db_offset
= 0;
3282 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
3283 db
->db
.db_size
= blocksize
;
3284 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
3288 * Hold the dn_dbufs_mtx while we get the new dbuf
3289 * in the hash table *and* added to the dbufs list.
3290 * This prevents a possible deadlock with someone
3291 * trying to look up this dbuf before it's added to the
3294 mutex_enter(&dn
->dn_dbufs_mtx
);
3295 db
->db_state
= DB_EVICTING
; /* not worth logging this state change */
3296 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
3297 /* someone else inserted it first */
3298 mutex_exit(&dn
->dn_dbufs_mtx
);
3299 kmem_cache_free(dbuf_kmem_cache
, db
);
3300 DBUF_STAT_BUMP(hash_insert_race
);
3303 avl_add(&dn
->dn_dbufs
, db
);
3305 db
->db_state
= DB_UNCACHED
;
3306 DTRACE_SET_STATE(db
, "regular buffer created");
3307 db
->db_caching_status
= DB_NO_CACHE
;
3308 mutex_exit(&dn
->dn_dbufs_mtx
);
3309 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3311 if (parent
&& parent
!= dn
->dn_dbuf
)
3312 dbuf_add_ref(parent
, db
);
3314 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
3315 zfs_refcount_count(&dn
->dn_holds
) > 0);
3316 (void) zfs_refcount_add(&dn
->dn_holds
, db
);
3318 dprintf_dbuf(db
, "db=%p\n", db
);
3324 * This function returns a block pointer and information about the object,
3325 * given a dnode and a block. This is a publicly accessible version of
3326 * dbuf_findbp that only returns some information, rather than the
3327 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
3328 * should be locked as (at least) a reader.
3331 dbuf_dnode_findbp(dnode_t
*dn
, uint64_t level
, uint64_t blkid
,
3332 blkptr_t
*bp
, uint16_t *datablkszsec
, uint8_t *indblkshift
)
3334 dmu_buf_impl_t
*dbp
= NULL
;
3337 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3339 err
= dbuf_findbp(dn
, level
, blkid
, B_FALSE
, &dbp
, &bp2
);
3341 ASSERT3P(bp2
, !=, NULL
);
3344 dbuf_rele(dbp
, NULL
);
3345 if (datablkszsec
!= NULL
)
3346 *datablkszsec
= dn
->dn_phys
->dn_datablkszsec
;
3347 if (indblkshift
!= NULL
)
3348 *indblkshift
= dn
->dn_phys
->dn_indblkshift
;
3354 typedef struct dbuf_prefetch_arg
{
3355 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
3356 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
3357 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
3358 int dpa_curlevel
; /* The current level that we're reading */
3359 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
3360 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
3361 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
3362 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
3363 dbuf_prefetch_fn dpa_cb
; /* prefetch completion callback */
3364 void *dpa_arg
; /* prefetch completion arg */
3365 } dbuf_prefetch_arg_t
;
3368 dbuf_prefetch_fini(dbuf_prefetch_arg_t
*dpa
, boolean_t io_done
)
3370 if (dpa
->dpa_cb
!= NULL
) {
3371 dpa
->dpa_cb(dpa
->dpa_arg
, dpa
->dpa_zb
.zb_level
,
3372 dpa
->dpa_zb
.zb_blkid
, io_done
);
3374 kmem_free(dpa
, sizeof (*dpa
));
3378 dbuf_issue_final_prefetch_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3379 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3381 (void) zio
, (void) zb
, (void) iobp
;
3382 dbuf_prefetch_arg_t
*dpa
= private;
3385 arc_buf_destroy(abuf
, private);
3387 dbuf_prefetch_fini(dpa
, B_TRUE
);
3391 * Actually issue the prefetch read for the block given.
3394 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
3396 ASSERT(!BP_IS_REDACTED(bp
) ||
3397 dsl_dataset_feature_is_active(
3398 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3399 SPA_FEATURE_REDACTED_DATASETS
));
3401 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
) || BP_IS_REDACTED(bp
))
3402 return (dbuf_prefetch_fini(dpa
, B_FALSE
));
3404 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
3405 arc_flags_t aflags
=
3406 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
|
3409 /* dnodes are always read as raw and then converted later */
3410 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
3411 dpa
->dpa_curlevel
== 0)
3412 zio_flags
|= ZIO_FLAG_RAW
;
3414 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3415 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
3416 ASSERT(dpa
->dpa_zio
!= NULL
);
3417 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
,
3418 dbuf_issue_final_prefetch_done
, dpa
,
3419 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
3423 * Called when an indirect block above our prefetch target is read in. This
3424 * will either read in the next indirect block down the tree or issue the actual
3425 * prefetch if the next block down is our target.
3428 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3429 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3431 (void) zb
, (void) iobp
;
3432 dbuf_prefetch_arg_t
*dpa
= private;
3434 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
3435 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
3438 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
3439 dbuf_prefetch_fini(dpa
, B_TRUE
);
3442 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
3445 * The dpa_dnode is only valid if we are called with a NULL
3446 * zio. This indicates that the arc_read() returned without
3447 * first calling zio_read() to issue a physical read. Once
3448 * a physical read is made the dpa_dnode must be invalidated
3449 * as the locks guarding it may have been dropped. If the
3450 * dpa_dnode is still valid, then we want to add it to the dbuf
3451 * cache. To do so, we must hold the dbuf associated with the block
3452 * we just prefetched, read its contents so that we associate it
3453 * with an arc_buf_t, and then release it.
3456 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
3457 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
3458 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
3460 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
3462 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
3464 dpa
->dpa_dnode
= NULL
;
3465 } else if (dpa
->dpa_dnode
!= NULL
) {
3466 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
3467 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
3468 dpa
->dpa_zb
.zb_level
));
3469 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
3470 dpa
->dpa_curlevel
, curblkid
, FTAG
);
3472 arc_buf_destroy(abuf
, private);
3473 dbuf_prefetch_fini(dpa
, B_TRUE
);
3476 (void) dbuf_read(db
, NULL
,
3477 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
3478 dbuf_rele(db
, FTAG
);
3481 dpa
->dpa_curlevel
--;
3482 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
3483 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
3484 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
3485 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
3487 ASSERT(!BP_IS_REDACTED(bp
) || (dpa
->dpa_dnode
&&
3488 dsl_dataset_feature_is_active(
3489 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3490 SPA_FEATURE_REDACTED_DATASETS
)));
3491 if (BP_IS_HOLE(bp
) || BP_IS_REDACTED(bp
)) {
3492 arc_buf_destroy(abuf
, private);
3493 dbuf_prefetch_fini(dpa
, B_TRUE
);
3495 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
3496 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
3497 dbuf_issue_final_prefetch(dpa
, bp
);
3499 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3500 zbookmark_phys_t zb
;
3502 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3503 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
3504 iter_aflags
|= ARC_FLAG_L2CACHE
;
3506 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3508 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
3509 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
3511 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3512 bp
, dbuf_prefetch_indirect_done
, dpa
,
3513 ZIO_PRIORITY_SYNC_READ
,
3514 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3518 arc_buf_destroy(abuf
, private);
3522 * Issue prefetch reads for the given block on the given level. If the indirect
3523 * blocks above that block are not in memory, we will read them in
3524 * asynchronously. As a result, this call never blocks waiting for a read to
3525 * complete. Note that the prefetch might fail if the dataset is encrypted and
3526 * the encryption key is unmapped before the IO completes.
3529 dbuf_prefetch_impl(dnode_t
*dn
, int64_t level
, uint64_t blkid
,
3530 zio_priority_t prio
, arc_flags_t aflags
, dbuf_prefetch_fn cb
,
3534 int epbs
, nlevels
, curlevel
;
3537 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3538 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3540 if (blkid
> dn
->dn_maxblkid
)
3543 if (level
== 0 && dnode_block_freed(dn
, blkid
))
3547 * This dnode hasn't been written to disk yet, so there's nothing to
3550 nlevels
= dn
->dn_phys
->dn_nlevels
;
3551 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
3554 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3555 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
3558 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
3559 level
, blkid
, NULL
);
3561 mutex_exit(&db
->db_mtx
);
3563 * This dbuf already exists. It is either CACHED, or
3564 * (we assume) about to be read or filled.
3570 * Find the closest ancestor (indirect block) of the target block
3571 * that is present in the cache. In this indirect block, we will
3572 * find the bp that is at curlevel, curblkid.
3576 while (curlevel
< nlevels
- 1) {
3577 int parent_level
= curlevel
+ 1;
3578 uint64_t parent_blkid
= curblkid
>> epbs
;
3581 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
3582 FALSE
, TRUE
, FTAG
, &db
) == 0) {
3583 blkptr_t
*bpp
= db
->db_buf
->b_data
;
3584 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
3585 dbuf_rele(db
, FTAG
);
3589 curlevel
= parent_level
;
3590 curblkid
= parent_blkid
;
3593 if (curlevel
== nlevels
- 1) {
3594 /* No cached indirect blocks found. */
3595 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
3596 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
3598 ASSERT(!BP_IS_REDACTED(&bp
) ||
3599 dsl_dataset_feature_is_active(dn
->dn_objset
->os_dsl_dataset
,
3600 SPA_FEATURE_REDACTED_DATASETS
));
3601 if (BP_IS_HOLE(&bp
) || BP_IS_REDACTED(&bp
))
3604 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
3606 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
3609 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
3610 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
3611 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3612 dn
->dn_object
, level
, blkid
);
3613 dpa
->dpa_curlevel
= curlevel
;
3614 dpa
->dpa_prio
= prio
;
3615 dpa
->dpa_aflags
= aflags
;
3616 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
3617 dpa
->dpa_dnode
= dn
;
3618 dpa
->dpa_epbs
= epbs
;
3623 if (!DNODE_LEVEL_IS_CACHEABLE(dn
, level
))
3624 dpa
->dpa_aflags
|= ARC_FLAG_UNCACHED
;
3625 else if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3626 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
3629 * If we have the indirect just above us, no need to do the asynchronous
3630 * prefetch chain; we'll just run the last step ourselves. If we're at
3631 * a higher level, though, we want to issue the prefetches for all the
3632 * indirect blocks asynchronously, so we can go on with whatever we were
3635 if (curlevel
== level
) {
3636 ASSERT3U(curblkid
, ==, blkid
);
3637 dbuf_issue_final_prefetch(dpa
, &bp
);
3639 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3640 zbookmark_phys_t zb
;
3642 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3643 if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3644 iter_aflags
|= ARC_FLAG_L2CACHE
;
3646 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3647 dn
->dn_object
, curlevel
, curblkid
);
3648 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3649 &bp
, dbuf_prefetch_indirect_done
, dpa
,
3650 ZIO_PRIORITY_SYNC_READ
,
3651 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3655 * We use pio here instead of dpa_zio since it's possible that
3656 * dpa may have already been freed.
3662 cb(arg
, level
, blkid
, B_FALSE
);
3667 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
3671 return (dbuf_prefetch_impl(dn
, level
, blkid
, prio
, aflags
, NULL
, NULL
));
3675 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3676 * the case of encrypted, compressed and uncompressed buffers by
3677 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3678 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3680 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3682 noinline
static void
3683 dbuf_hold_copy(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3685 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3686 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
3687 enum zio_compress compress_type
= arc_get_compression(data
);
3688 uint8_t complevel
= arc_get_complevel(data
);
3690 if (arc_is_encrypted(data
)) {
3691 boolean_t byteorder
;
3692 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3693 uint8_t iv
[ZIO_DATA_IV_LEN
];
3694 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3696 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
3697 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
3698 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
3699 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
3700 compress_type
, complevel
));
3701 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3702 dbuf_set_data(db
, arc_alloc_compressed_buf(
3703 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
3704 arc_buf_lsize(data
), compress_type
, complevel
));
3706 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
3707 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
3710 rw_enter(&db
->db_rwlock
, RW_WRITER
);
3711 memcpy(db
->db
.db_data
, data
->b_data
, arc_buf_size(data
));
3712 rw_exit(&db
->db_rwlock
);
3716 * Returns with db_holds incremented, and db_mtx not held.
3717 * Note: dn_struct_rwlock must be held.
3720 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3721 boolean_t fail_sparse
, boolean_t fail_uncached
,
3722 const void *tag
, dmu_buf_impl_t
**dbp
)
3724 dmu_buf_impl_t
*db
, *parent
= NULL
;
3727 /* If the pool has been created, verify the tx_sync_lock is not held */
3728 spa_t
*spa
= dn
->dn_objset
->os_spa
;
3729 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
3731 ASSERT(!MUTEX_HELD(&dp
->dp_tx
.tx_sync_lock
));
3734 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3735 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3736 ASSERT3U(dn
->dn_nlevels
, >, level
);
3740 /* dbuf_find() returns with db_mtx held */
3741 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
, &hv
);
3744 blkptr_t
*bp
= NULL
;
3748 return (SET_ERROR(ENOENT
));
3750 ASSERT3P(parent
, ==, NULL
);
3751 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
3753 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
3754 err
= SET_ERROR(ENOENT
);
3757 dbuf_rele(parent
, NULL
);
3761 if (err
&& err
!= ENOENT
)
3763 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
, hv
);
3766 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
3767 mutex_exit(&db
->db_mtx
);
3768 return (SET_ERROR(ENOENT
));
3771 if (db
->db_buf
!= NULL
) {
3772 arc_buf_access(db
->db_buf
);
3773 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
3776 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
3779 * If this buffer is currently syncing out, and we are
3780 * still referencing it from db_data, we need to make a copy
3781 * of it in case we decide we want to dirty it again in this txg.
3783 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
3784 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3785 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
3786 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3787 if (dr
->dt
.dl
.dr_data
== db
->db_buf
) {
3788 ASSERT3P(db
->db_buf
, !=, NULL
);
3789 dbuf_hold_copy(dn
, db
);
3793 if (multilist_link_active(&db
->db_cache_link
)) {
3794 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3795 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3796 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3798 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
3800 uint64_t size
= db
->db
.db_size
;
3801 uint64_t usize
= dmu_buf_user_size(&db
->db
);
3802 (void) zfs_refcount_remove_many(
3803 &dbuf_caches
[db
->db_caching_status
].size
, size
, db
);
3804 (void) zfs_refcount_remove_many(
3805 &dbuf_caches
[db
->db_caching_status
].size
, usize
,
3808 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3809 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3811 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3812 DBUF_STAT_BUMPDOWN(cache_count
);
3813 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3816 db
->db_caching_status
= DB_NO_CACHE
;
3818 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3820 mutex_exit(&db
->db_mtx
);
3822 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3824 dbuf_rele(parent
, NULL
);
3826 ASSERT3P(DB_DNODE(db
), ==, dn
);
3827 ASSERT3U(db
->db_blkid
, ==, blkid
);
3828 ASSERT3U(db
->db_level
, ==, level
);
3835 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, const void *tag
)
3837 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3841 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, const void *tag
)
3844 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3845 return (err
? NULL
: db
);
3849 dbuf_create_bonus(dnode_t
*dn
)
3851 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3853 ASSERT(dn
->dn_bonus
== NULL
);
3854 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
,
3855 dbuf_hash(dn
->dn_objset
, dn
->dn_object
, 0, DMU_BONUS_BLKID
));
3859 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3861 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3863 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3864 return (SET_ERROR(ENOTSUP
));
3866 blksz
= SPA_MINBLOCKSIZE
;
3867 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3868 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3870 dbuf_new_size(db
, blksz
, tx
);
3876 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3878 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3881 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3883 dbuf_add_ref(dmu_buf_impl_t
*db
, const void *tag
)
3885 int64_t holds
= zfs_refcount_add(&db
->db_holds
, tag
);
3886 VERIFY3S(holds
, >, 1);
3889 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3891 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3894 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3895 dmu_buf_impl_t
*found_db
;
3896 boolean_t result
= B_FALSE
;
3898 if (blkid
== DMU_BONUS_BLKID
)
3899 found_db
= dbuf_find_bonus(os
, obj
);
3901 found_db
= dbuf_find(os
, obj
, 0, blkid
, NULL
);
3903 if (found_db
!= NULL
) {
3904 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3905 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3908 mutex_exit(&found_db
->db_mtx
);
3914 * If you call dbuf_rele() you had better not be referencing the dnode handle
3915 * unless you have some other direct or indirect hold on the dnode. (An indirect
3916 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3917 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3918 * dnode's parent dbuf evicting its dnode handles.
3921 dbuf_rele(dmu_buf_impl_t
*db
, const void *tag
)
3923 mutex_enter(&db
->db_mtx
);
3924 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
3928 dmu_buf_rele(dmu_buf_t
*db
, const void *tag
)
3930 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3934 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3935 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3936 * argument should be set if we are already in the dbuf-evicting code
3937 * path, in which case we don't want to recursively evict. This allows us to
3938 * avoid deeply nested stacks that would have a call flow similar to this:
3940 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3943 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3947 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, const void *tag
, boolean_t evicting
)
3952 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3956 * Remove the reference to the dbuf before removing its hold on the
3957 * dnode so we can guarantee in dnode_move() that a referenced bonus
3958 * buffer has a corresponding dnode hold.
3960 holds
= zfs_refcount_remove(&db
->db_holds
, tag
);
3964 * We can't freeze indirects if there is a possibility that they
3965 * may be modified in the current syncing context.
3967 if (db
->db_buf
!= NULL
&&
3968 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3969 arc_buf_freeze(db
->db_buf
);
3972 if (holds
== db
->db_dirtycnt
&&
3973 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3974 dbuf_evict_user(db
);
3977 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3979 boolean_t evict_dbuf
= db
->db_pending_evict
;
3982 * If the dnode moves here, we cannot cross this
3983 * barrier until the move completes.
3988 atomic_dec_32(&dn
->dn_dbufs_count
);
3991 * Decrementing the dbuf count means that the bonus
3992 * buffer's dnode hold is no longer discounted in
3993 * dnode_move(). The dnode cannot move until after
3994 * the dnode_rele() below.
3999 * Do not reference db after its lock is dropped.
4000 * Another thread may evict it.
4002 mutex_exit(&db
->db_mtx
);
4005 dnode_evict_bonus(dn
);
4008 } else if (db
->db_buf
== NULL
) {
4010 * This is a special case: we never associated this
4011 * dbuf with any data allocated from the ARC.
4013 ASSERT(db
->db_state
== DB_UNCACHED
||
4014 db
->db_state
== DB_NOFILL
);
4016 } else if (arc_released(db
->db_buf
)) {
4018 * This dbuf has anonymous data associated with it.
4021 } else if (!(DBUF_IS_CACHEABLE(db
) || db
->db_partial_read
) ||
4022 db
->db_pending_evict
) {
4024 } else if (!multilist_link_active(&db
->db_cache_link
)) {
4025 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
4027 dbuf_cached_state_t dcs
=
4028 dbuf_include_in_metadata_cache(db
) ?
4029 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
4030 db
->db_caching_status
= dcs
;
4032 multilist_insert(&dbuf_caches
[dcs
].cache
, db
);
4033 uint64_t db_size
= db
->db
.db_size
;
4034 uint64_t dbu_size
= dmu_buf_user_size(&db
->db
);
4035 (void) zfs_refcount_add_many(
4036 &dbuf_caches
[dcs
].size
, db_size
, db
);
4037 size
= zfs_refcount_add_many(
4038 &dbuf_caches
[dcs
].size
, dbu_size
, db
->db_user
);
4039 uint8_t db_level
= db
->db_level
;
4040 mutex_exit(&db
->db_mtx
);
4042 if (dcs
== DB_DBUF_METADATA_CACHE
) {
4043 DBUF_STAT_BUMP(metadata_cache_count
);
4044 DBUF_STAT_MAX(metadata_cache_size_bytes_max
,
4047 DBUF_STAT_BUMP(cache_count
);
4048 DBUF_STAT_MAX(cache_size_bytes_max
, size
);
4049 DBUF_STAT_BUMP(cache_levels
[db_level
]);
4050 DBUF_STAT_INCR(cache_levels_bytes
[db_level
],
4051 db_size
+ dbu_size
);
4054 if (dcs
== DB_DBUF_CACHE
&& !evicting
)
4055 dbuf_evict_notify(size
);
4058 mutex_exit(&db
->db_mtx
);
4063 #pragma weak dmu_buf_refcount = dbuf_refcount
4065 dbuf_refcount(dmu_buf_impl_t
*db
)
4067 return (zfs_refcount_count(&db
->db_holds
));
4071 dmu_buf_user_refcount(dmu_buf_t
*db_fake
)
4074 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4076 mutex_enter(&db
->db_mtx
);
4077 ASSERT3U(zfs_refcount_count(&db
->db_holds
), >=, db
->db_dirtycnt
);
4078 holds
= zfs_refcount_count(&db
->db_holds
) - db
->db_dirtycnt
;
4079 mutex_exit(&db
->db_mtx
);
4085 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
4086 dmu_buf_user_t
*new_user
)
4088 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4090 mutex_enter(&db
->db_mtx
);
4091 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4092 if (db
->db_user
== old_user
)
4093 db
->db_user
= new_user
;
4095 old_user
= db
->db_user
;
4096 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4097 mutex_exit(&db
->db_mtx
);
4103 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4105 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
4109 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4111 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4113 db
->db_user_immediate_evict
= TRUE
;
4114 return (dmu_buf_set_user(db_fake
, user
));
4118 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4120 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
4124 dmu_buf_get_user(dmu_buf_t
*db_fake
)
4126 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4128 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4129 return (db
->db_user
);
4133 dmu_buf_user_size(dmu_buf_t
*db_fake
)
4135 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4136 if (db
->db_user
== NULL
)
4138 return (atomic_load_64(&db
->db_user
->dbu_size
));
4142 dmu_buf_add_user_size(dmu_buf_t
*db_fake
, uint64_t nadd
)
4144 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4145 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
4146 ASSERT3P(db
->db_user
, !=, NULL
);
4147 ASSERT3U(atomic_load_64(&db
->db_user
->dbu_size
), <, UINT64_MAX
- nadd
);
4148 atomic_add_64(&db
->db_user
->dbu_size
, nadd
);
4152 dmu_buf_sub_user_size(dmu_buf_t
*db_fake
, uint64_t nsub
)
4154 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4155 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
4156 ASSERT3P(db
->db_user
, !=, NULL
);
4157 ASSERT3U(atomic_load_64(&db
->db_user
->dbu_size
), >=, nsub
);
4158 atomic_sub_64(&db
->db_user
->dbu_size
, nsub
);
4162 dmu_buf_user_evict_wait(void)
4164 taskq_wait(dbu_evict_taskq
);
4168 dmu_buf_get_blkptr(dmu_buf_t
*db
)
4170 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4171 return (dbi
->db_blkptr
);
4175 dmu_buf_get_objset(dmu_buf_t
*db
)
4177 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4178 return (dbi
->db_objset
);
4182 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
4184 /* ASSERT(dmu_tx_is_syncing(tx) */
4185 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4187 if (db
->db_blkptr
!= NULL
)
4190 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4191 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
4192 BP_ZERO(db
->db_blkptr
);
4195 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
4197 * This buffer was allocated at a time when there was
4198 * no available blkptrs from the dnode, or it was
4199 * inappropriate to hook it in (i.e., nlevels mismatch).
4201 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
4202 ASSERT(db
->db_parent
== NULL
);
4203 db
->db_parent
= dn
->dn_dbuf
;
4204 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
4207 dmu_buf_impl_t
*parent
= db
->db_parent
;
4208 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4210 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
4211 if (parent
== NULL
) {
4212 mutex_exit(&db
->db_mtx
);
4213 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
4214 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
4215 db
->db_blkid
>> epbs
, db
);
4216 rw_exit(&dn
->dn_struct_rwlock
);
4217 mutex_enter(&db
->db_mtx
);
4218 db
->db_parent
= parent
;
4220 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
4221 (db
->db_blkid
& ((1ULL << epbs
) - 1));
4227 dbuf_sync_bonus(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4229 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4230 void *data
= dr
->dt
.dl
.dr_data
;
4232 ASSERT0(db
->db_level
);
4233 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4234 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
);
4235 ASSERT(data
!= NULL
);
4237 dnode_t
*dn
= dr
->dr_dnode
;
4238 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
4239 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
4240 memcpy(DN_BONUS(dn
->dn_phys
), data
, DN_MAX_BONUS_LEN(dn
->dn_phys
));
4242 dbuf_sync_leaf_verify_bonus_dnode(dr
);
4244 dbuf_undirty_bonus(dr
);
4245 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4249 * When syncing out a blocks of dnodes, adjust the block to deal with
4250 * encryption. Normally, we make sure the block is decrypted before writing
4251 * it. If we have crypt params, then we are writing a raw (encrypted) block,
4252 * from a raw receive. In this case, set the ARC buf's crypt params so
4253 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
4256 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t
*dr
)
4259 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4261 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4262 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
4263 ASSERT3U(db
->db_level
, ==, 0);
4265 if (!db
->db_objset
->os_raw_receive
&& arc_is_encrypted(db
->db_buf
)) {
4266 zbookmark_phys_t zb
;
4269 * Unfortunately, there is currently no mechanism for
4270 * syncing context to handle decryption errors. An error
4271 * here is only possible if an attacker maliciously
4272 * changed a dnode block and updated the associated
4273 * checksums going up the block tree.
4275 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
4276 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4277 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
4280 panic("Invalid dnode block MAC");
4281 } else if (dr
->dt
.dl
.dr_has_raw_params
) {
4282 (void) arc_release(dr
->dt
.dl
.dr_data
, db
);
4283 arc_convert_to_raw(dr
->dt
.dl
.dr_data
,
4284 dmu_objset_id(db
->db_objset
),
4285 dr
->dt
.dl
.dr_byteorder
, DMU_OT_DNODE
,
4286 dr
->dt
.dl
.dr_salt
, dr
->dt
.dl
.dr_iv
, dr
->dt
.dl
.dr_mac
);
4291 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
4292 * is critical the we not allow the compiler to inline this function in to
4293 * dbuf_sync_list() thereby drastically bloating the stack usage.
4295 noinline
static void
4296 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4298 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4299 dnode_t
*dn
= dr
->dr_dnode
;
4301 ASSERT(dmu_tx_is_syncing(tx
));
4303 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4305 mutex_enter(&db
->db_mtx
);
4307 ASSERT(db
->db_level
> 0);
4310 /* Read the block if it hasn't been read yet. */
4311 if (db
->db_buf
== NULL
) {
4312 mutex_exit(&db
->db_mtx
);
4313 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
4314 mutex_enter(&db
->db_mtx
);
4316 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
4317 ASSERT(db
->db_buf
!= NULL
);
4319 /* Indirect block size must match what the dnode thinks it is. */
4320 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4321 dbuf_check_blkptr(dn
, db
);
4323 /* Provide the pending dirty record to child dbufs */
4324 db
->db_data_pending
= dr
;
4326 mutex_exit(&db
->db_mtx
);
4328 dbuf_write(dr
, db
->db_buf
, tx
);
4330 zio_t
*zio
= dr
->dr_zio
;
4331 mutex_enter(&dr
->dt
.di
.dr_mtx
);
4332 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
4333 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4334 mutex_exit(&dr
->dt
.di
.dr_mtx
);
4339 * Verify that the size of the data in our bonus buffer does not exceed
4340 * its recorded size.
4342 * The purpose of this verification is to catch any cases in development
4343 * where the size of a phys structure (i.e space_map_phys_t) grows and,
4344 * due to incorrect feature management, older pools expect to read more
4345 * data even though they didn't actually write it to begin with.
4347 * For a example, this would catch an error in the feature logic where we
4348 * open an older pool and we expect to write the space map histogram of
4349 * a space map with size SPACE_MAP_SIZE_V0.
4352 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
)
4355 dnode_t
*dn
= dr
->dr_dnode
;
4358 * Encrypted bonus buffers can have data past their bonuslen.
4359 * Skip the verification of these blocks.
4361 if (DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))
4364 uint16_t bonuslen
= dn
->dn_phys
->dn_bonuslen
;
4365 uint16_t maxbonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
4366 ASSERT3U(bonuslen
, <=, maxbonuslen
);
4368 arc_buf_t
*datap
= dr
->dt
.dl
.dr_data
;
4369 char *datap_end
= ((char *)datap
) + bonuslen
;
4370 char *datap_max
= ((char *)datap
) + maxbonuslen
;
4372 /* ensure that everything is zero after our data */
4373 for (; datap_end
< datap_max
; datap_end
++)
4374 ASSERT(*datap_end
== 0);
4379 dbuf_lightweight_bp(dbuf_dirty_record_t
*dr
)
4381 /* This must be a lightweight dirty record. */
4382 ASSERT3P(dr
->dr_dbuf
, ==, NULL
);
4383 dnode_t
*dn
= dr
->dr_dnode
;
4385 if (dn
->dn_phys
->dn_nlevels
== 1) {
4386 VERIFY3U(dr
->dt
.dll
.dr_blkid
, <, dn
->dn_phys
->dn_nblkptr
);
4387 return (&dn
->dn_phys
->dn_blkptr
[dr
->dt
.dll
.dr_blkid
]);
4389 dmu_buf_impl_t
*parent_db
= dr
->dr_parent
->dr_dbuf
;
4390 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4391 VERIFY3U(parent_db
->db_level
, ==, 1);
4392 VERIFY3P(parent_db
->db_dnode_handle
->dnh_dnode
, ==, dn
);
4393 VERIFY3U(dr
->dt
.dll
.dr_blkid
>> epbs
, ==, parent_db
->db_blkid
);
4394 blkptr_t
*bp
= parent_db
->db
.db_data
;
4395 return (&bp
[dr
->dt
.dll
.dr_blkid
& ((1 << epbs
) - 1)]);
4400 dbuf_lightweight_ready(zio_t
*zio
)
4402 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4403 blkptr_t
*bp
= zio
->io_bp
;
4405 if (zio
->io_error
!= 0)
4408 dnode_t
*dn
= dr
->dr_dnode
;
4410 blkptr_t
*bp_orig
= dbuf_lightweight_bp(dr
);
4411 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4412 int64_t delta
= bp_get_dsize_sync(spa
, bp
) -
4413 bp_get_dsize_sync(spa
, bp_orig
);
4414 dnode_diduse_space(dn
, delta
);
4416 uint64_t blkid
= dr
->dt
.dll
.dr_blkid
;
4417 mutex_enter(&dn
->dn_mtx
);
4418 if (blkid
> dn
->dn_phys
->dn_maxblkid
) {
4419 ASSERT0(dn
->dn_objset
->os_raw_receive
);
4420 dn
->dn_phys
->dn_maxblkid
= blkid
;
4422 mutex_exit(&dn
->dn_mtx
);
4424 if (!BP_IS_EMBEDDED(bp
)) {
4425 uint64_t fill
= BP_IS_HOLE(bp
) ? 0 : 1;
4426 BP_SET_FILL(bp
, fill
);
4429 dmu_buf_impl_t
*parent_db
;
4430 EQUIV(dr
->dr_parent
== NULL
, dn
->dn_phys
->dn_nlevels
== 1);
4431 if (dr
->dr_parent
== NULL
) {
4432 parent_db
= dn
->dn_dbuf
;
4434 parent_db
= dr
->dr_parent
->dr_dbuf
;
4436 rw_enter(&parent_db
->db_rwlock
, RW_WRITER
);
4438 rw_exit(&parent_db
->db_rwlock
);
4442 dbuf_lightweight_done(zio_t
*zio
)
4444 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4446 VERIFY0(zio
->io_error
);
4448 objset_t
*os
= dr
->dr_dnode
->dn_objset
;
4449 dmu_tx_t
*tx
= os
->os_synctx
;
4451 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4452 ASSERT(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4454 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4455 (void) dsl_dataset_block_kill(ds
, &zio
->io_bp_orig
, tx
, B_TRUE
);
4456 dsl_dataset_block_born(ds
, zio
->io_bp
, tx
);
4459 dsl_pool_undirty_space(dmu_objset_pool(os
), dr
->dr_accounted
,
4462 abd_free(dr
->dt
.dll
.dr_abd
);
4463 kmem_free(dr
, sizeof (*dr
));
4466 noinline
static void
4467 dbuf_sync_lightweight(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4469 dnode_t
*dn
= dr
->dr_dnode
;
4471 if (dn
->dn_phys
->dn_nlevels
== 1) {
4474 pio
= dr
->dr_parent
->dr_zio
;
4477 zbookmark_phys_t zb
= {
4478 .zb_objset
= dmu_objset_id(dn
->dn_objset
),
4479 .zb_object
= dn
->dn_object
,
4481 .zb_blkid
= dr
->dt
.dll
.dr_blkid
,
4485 * See comment in dbuf_write(). This is so that zio->io_bp_orig
4486 * will have the old BP in dbuf_lightweight_done().
4488 dr
->dr_bp_copy
= *dbuf_lightweight_bp(dr
);
4490 dr
->dr_zio
= zio_write(pio
, dmu_objset_spa(dn
->dn_objset
),
4491 dmu_tx_get_txg(tx
), &dr
->dr_bp_copy
, dr
->dt
.dll
.dr_abd
,
4492 dn
->dn_datablksz
, abd_get_size(dr
->dt
.dll
.dr_abd
),
4493 &dr
->dt
.dll
.dr_props
, dbuf_lightweight_ready
, NULL
,
4494 dbuf_lightweight_done
, dr
, ZIO_PRIORITY_ASYNC_WRITE
,
4495 ZIO_FLAG_MUSTSUCCEED
| dr
->dt
.dll
.dr_flags
, &zb
);
4497 zio_nowait(dr
->dr_zio
);
4501 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
4502 * critical the we not allow the compiler to inline this function in to
4503 * dbuf_sync_list() thereby drastically bloating the stack usage.
4505 noinline
static void
4506 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4508 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
4509 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4510 dnode_t
*dn
= dr
->dr_dnode
;
4512 uint64_t txg
= tx
->tx_txg
;
4514 ASSERT(dmu_tx_is_syncing(tx
));
4516 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4518 mutex_enter(&db
->db_mtx
);
4520 * To be synced, we must be dirtied. But we
4521 * might have been freed after the dirty.
4523 if (db
->db_state
== DB_UNCACHED
) {
4524 /* This buffer has been freed since it was dirtied */
4525 ASSERT(db
->db
.db_data
== NULL
);
4526 } else if (db
->db_state
== DB_FILL
) {
4527 /* This buffer was freed and is now being re-filled */
4528 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
4529 } else if (db
->db_state
== DB_READ
) {
4531 * This buffer has a clone we need to write, and an in-flight
4532 * read on the BP we're about to clone. Its safe to issue the
4533 * write here because the read has already been issued and the
4534 * contents won't change.
4536 ASSERT(dr
->dt
.dl
.dr_brtwrite
&&
4537 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
);
4539 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
4543 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4544 mutex_enter(&dn
->dn_mtx
);
4545 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
4547 * In the previous transaction group, the bonus buffer
4548 * was entirely used to store the attributes for the
4549 * dnode which overrode the dn_spill field. However,
4550 * when adding more attributes to the file a spill
4551 * block was required to hold the extra attributes.
4553 * Make sure to clear the garbage left in the dn_spill
4554 * field from the previous attributes in the bonus
4555 * buffer. Otherwise, after writing out the spill
4556 * block to the new allocated dva, it will free
4557 * the old block pointed to by the invalid dn_spill.
4559 db
->db_blkptr
= NULL
;
4561 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
4562 mutex_exit(&dn
->dn_mtx
);
4566 * If this is a bonus buffer, simply copy the bonus data into the
4567 * dnode. It will be written out when the dnode is synced (and it
4568 * will be synced, since it must have been dirty for dbuf_sync to
4571 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
4572 ASSERT(dr
->dr_dbuf
== db
);
4573 dbuf_sync_bonus(dr
, tx
);
4580 * This function may have dropped the db_mtx lock allowing a dmu_sync
4581 * operation to sneak in. As a result, we need to ensure that we
4582 * don't check the dr_override_state until we have returned from
4583 * dbuf_check_blkptr.
4585 dbuf_check_blkptr(dn
, db
);
4588 * If this buffer is in the middle of an immediate write,
4589 * wait for the synchronous IO to complete.
4591 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
4592 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
4593 cv_wait(&db
->db_changed
, &db
->db_mtx
);
4597 * If this is a dnode block, ensure it is appropriately encrypted
4598 * or decrypted, depending on what we are writing to it this txg.
4600 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
4601 dbuf_prepare_encrypted_dnode_leaf(dr
);
4603 if (*datap
!= NULL
&& *datap
== db
->db_buf
&&
4604 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
4605 zfs_refcount_count(&db
->db_holds
) > 1 &&
4606 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
) {
4608 * If this buffer is currently "in use" (i.e., there
4609 * are active holds and db_data still references it),
4610 * then make a copy before we start the write so that
4611 * any modifications from the open txg will not leak
4614 * NOTE: this copy does not need to be made for
4615 * objects only modified in the syncing context (e.g.
4616 * DNONE_DNODE blocks).
4618 int psize
= arc_buf_size(*datap
);
4619 int lsize
= arc_buf_lsize(*datap
);
4620 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
4621 enum zio_compress compress_type
= arc_get_compression(*datap
);
4622 uint8_t complevel
= arc_get_complevel(*datap
);
4624 if (arc_is_encrypted(*datap
)) {
4625 boolean_t byteorder
;
4626 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4627 uint8_t iv
[ZIO_DATA_IV_LEN
];
4628 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4630 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
4631 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
4632 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
4633 dn
->dn_type
, psize
, lsize
, compress_type
,
4635 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
4636 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
4637 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
4638 psize
, lsize
, compress_type
, complevel
);
4640 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
4642 memcpy((*datap
)->b_data
, db
->db
.db_data
, psize
);
4644 db
->db_data_pending
= dr
;
4646 mutex_exit(&db
->db_mtx
);
4648 dbuf_write(dr
, *datap
, tx
);
4650 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4651 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
4652 list_insert_tail(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
4654 zio_nowait(dr
->dr_zio
);
4659 * Syncs out a range of dirty records for indirect or leaf dbufs. May be
4660 * called recursively from dbuf_sync_indirect().
4663 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
4665 dbuf_dirty_record_t
*dr
;
4667 while ((dr
= list_head(list
))) {
4668 if (dr
->dr_zio
!= NULL
) {
4670 * If we find an already initialized zio then we
4671 * are processing the meta-dnode, and we have finished.
4672 * The dbufs for all dnodes are put back on the list
4673 * during processing, so that we can zio_wait()
4674 * these IOs after initiating all child IOs.
4676 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
4677 DMU_META_DNODE_OBJECT
);
4680 list_remove(list
, dr
);
4681 if (dr
->dr_dbuf
== NULL
) {
4682 dbuf_sync_lightweight(dr
, tx
);
4684 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
4685 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
4686 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
4688 if (dr
->dr_dbuf
->db_level
> 0)
4689 dbuf_sync_indirect(dr
, tx
);
4691 dbuf_sync_leaf(dr
, tx
);
4697 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4700 dmu_buf_impl_t
*db
= vdb
;
4702 blkptr_t
*bp
= zio
->io_bp
;
4703 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4704 spa_t
*spa
= zio
->io_spa
;
4709 ASSERT3P(db
->db_blkptr
, !=, NULL
);
4710 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
4714 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
4715 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
4716 zio
->io_prev_space_delta
= delta
;
4718 if (BP_GET_LOGICAL_BIRTH(bp
) != 0) {
4719 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
4720 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
4721 (db
->db_blkid
== DMU_SPILL_BLKID
&&
4722 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
4723 BP_IS_EMBEDDED(bp
));
4724 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
4727 mutex_enter(&db
->db_mtx
);
4730 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4731 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4732 ASSERT(!(BP_IS_HOLE(bp
)) &&
4733 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4737 if (db
->db_level
== 0) {
4738 mutex_enter(&dn
->dn_mtx
);
4739 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
4740 db
->db_blkid
!= DMU_SPILL_BLKID
) {
4741 ASSERT0(db
->db_objset
->os_raw_receive
);
4742 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
4744 mutex_exit(&dn
->dn_mtx
);
4746 if (dn
->dn_type
== DMU_OT_DNODE
) {
4748 while (i
< db
->db
.db_size
) {
4750 (void *)(((char *)db
->db
.db_data
) + i
);
4752 i
+= DNODE_MIN_SIZE
;
4753 if (dnp
->dn_type
!= DMU_OT_NONE
) {
4755 for (int j
= 0; j
< dnp
->dn_nblkptr
;
4757 (void) zfs_blkptr_verify(spa
,
4763 DNODE_FLAG_SPILL_BLKPTR
) {
4764 (void) zfs_blkptr_verify(spa
,
4765 DN_SPILL_BLKPTR(dnp
),
4769 i
+= dnp
->dn_extra_slots
*
4774 if (BP_IS_HOLE(bp
)) {
4781 blkptr_t
*ibp
= db
->db
.db_data
;
4782 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4783 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
4784 if (BP_IS_HOLE(ibp
))
4786 (void) zfs_blkptr_verify(spa
, ibp
,
4787 BLK_CONFIG_SKIP
, BLK_VERIFY_HALT
);
4788 fill
+= BP_GET_FILL(ibp
);
4793 if (!BP_IS_EMBEDDED(bp
))
4794 BP_SET_FILL(bp
, fill
);
4796 mutex_exit(&db
->db_mtx
);
4798 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_WRITER
, FTAG
);
4799 *db
->db_blkptr
= *bp
;
4800 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
4804 * This function gets called just prior to running through the compression
4805 * stage of the zio pipeline. If we're an indirect block comprised of only
4806 * holes, then we want this indirect to be compressed away to a hole. In
4807 * order to do that we must zero out any information about the holes that
4808 * this indirect points to prior to before we try to compress it.
4811 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4813 (void) zio
, (void) buf
;
4814 dmu_buf_impl_t
*db
= vdb
;
4817 unsigned int epbs
, i
;
4819 ASSERT3U(db
->db_level
, >, 0);
4822 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4823 ASSERT3U(epbs
, <, 31);
4825 /* Determine if all our children are holes */
4826 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
4827 if (!BP_IS_HOLE(bp
))
4832 * If all the children are holes, then zero them all out so that
4833 * we may get compressed away.
4835 if (i
== 1ULL << epbs
) {
4837 * We only found holes. Grab the rwlock to prevent
4838 * anybody from reading the blocks we're about to
4841 rw_enter(&db
->db_rwlock
, RW_WRITER
);
4842 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
4843 rw_exit(&db
->db_rwlock
);
4849 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4852 dmu_buf_impl_t
*db
= vdb
;
4853 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4854 blkptr_t
*bp
= db
->db_blkptr
;
4855 objset_t
*os
= db
->db_objset
;
4856 dmu_tx_t
*tx
= os
->os_synctx
;
4858 ASSERT0(zio
->io_error
);
4859 ASSERT(db
->db_blkptr
== bp
);
4862 * For nopwrites and rewrites we ensure that the bp matches our
4863 * original and bypass all the accounting.
4865 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4866 ASSERT(BP_EQUAL(bp
, bp_orig
));
4868 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4869 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
4870 dsl_dataset_block_born(ds
, bp
, tx
);
4873 mutex_enter(&db
->db_mtx
);
4877 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
4878 dnode_t
*dn
= dr
->dr_dnode
;
4879 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4880 ASSERT(dr
->dr_dbuf
== db
);
4881 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
4882 list_remove(&db
->db_dirty_records
, dr
);
4885 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4886 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4887 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
4888 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4892 if (db
->db_level
== 0) {
4893 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
4894 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
4895 if (dr
->dt
.dl
.dr_data
!= NULL
&&
4896 dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
4897 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
4900 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4901 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
4902 if (!BP_IS_HOLE(db
->db_blkptr
)) {
4903 int epbs __maybe_unused
= dn
->dn_phys
->dn_indblkshift
-
4905 ASSERT3U(db
->db_blkid
, <=,
4906 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
4907 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4910 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4911 list_destroy(&dr
->dt
.di
.dr_children
);
4914 cv_broadcast(&db
->db_changed
);
4915 ASSERT(db
->db_dirtycnt
> 0);
4916 db
->db_dirtycnt
-= 1;
4917 db
->db_data_pending
= NULL
;
4918 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4920 dsl_pool_undirty_space(dmu_objset_pool(os
), dr
->dr_accounted
,
4923 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4927 dbuf_write_nofill_ready(zio_t
*zio
)
4929 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4933 dbuf_write_nofill_done(zio_t
*zio
)
4935 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4939 dbuf_write_override_ready(zio_t
*zio
)
4941 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4942 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4944 dbuf_write_ready(zio
, NULL
, db
);
4948 dbuf_write_override_done(zio_t
*zio
)
4950 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4951 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4952 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4954 mutex_enter(&db
->db_mtx
);
4955 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4956 if (!BP_IS_HOLE(obp
))
4957 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4958 arc_release(dr
->dt
.dl
.dr_data
, db
);
4960 mutex_exit(&db
->db_mtx
);
4962 dbuf_write_done(zio
, NULL
, db
);
4964 if (zio
->io_abd
!= NULL
)
4965 abd_free(zio
->io_abd
);
4968 typedef struct dbuf_remap_impl_callback_arg
{
4970 uint64_t drica_blk_birth
;
4972 } dbuf_remap_impl_callback_arg_t
;
4975 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
4978 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
4979 objset_t
*os
= drica
->drica_os
;
4980 spa_t
*spa
= dmu_objset_spa(os
);
4981 dmu_tx_t
*tx
= drica
->drica_tx
;
4983 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4985 if (os
== spa_meta_objset(spa
)) {
4986 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
4988 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
4989 size
, drica
->drica_blk_birth
, tx
);
4994 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, krwlock_t
*rw
, dmu_tx_t
*tx
)
4996 blkptr_t bp_copy
= *bp
;
4997 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4998 dbuf_remap_impl_callback_arg_t drica
;
5000 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
5002 drica
.drica_os
= dn
->dn_objset
;
5003 drica
.drica_blk_birth
= BP_GET_LOGICAL_BIRTH(bp
);
5004 drica
.drica_tx
= tx
;
5005 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
5008 * If the blkptr being remapped is tracked by a livelist,
5009 * then we need to make sure the livelist reflects the update.
5010 * First, cancel out the old blkptr by appending a 'FREE'
5011 * entry. Next, add an 'ALLOC' to track the new version. This
5012 * way we avoid trying to free an inaccurate blkptr at delete.
5013 * Note that embedded blkptrs are not tracked in livelists.
5015 if (dn
->dn_objset
!= spa_meta_objset(spa
)) {
5016 dsl_dataset_t
*ds
= dmu_objset_ds(dn
->dn_objset
);
5017 if (dsl_deadlist_is_open(&ds
->ds_dir
->dd_livelist
) &&
5018 BP_GET_LOGICAL_BIRTH(bp
) >
5019 ds
->ds_dir
->dd_origin_txg
) {
5020 ASSERT(!BP_IS_EMBEDDED(bp
));
5021 ASSERT(dsl_dir_is_clone(ds
->ds_dir
));
5022 ASSERT(spa_feature_is_enabled(spa
,
5023 SPA_FEATURE_LIVELIST
));
5024 bplist_append(&ds
->ds_dir
->dd_pending_frees
,
5026 bplist_append(&ds
->ds_dir
->dd_pending_allocs
,
5032 * The db_rwlock prevents dbuf_read_impl() from
5033 * dereferencing the BP while we are changing it. To
5034 * avoid lock contention, only grab it when we are actually
5038 rw_enter(rw
, RW_WRITER
);
5046 * Remap any existing BP's to concrete vdevs, if possible.
5049 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
5051 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
5052 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
5054 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
5057 if (db
->db_level
> 0) {
5058 blkptr_t
*bp
= db
->db
.db_data
;
5059 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
5060 dbuf_remap_impl(dn
, &bp
[i
], &db
->db_rwlock
, tx
);
5062 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
5063 dnode_phys_t
*dnp
= db
->db
.db_data
;
5064 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
5066 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
;
5067 i
+= dnp
[i
].dn_extra_slots
+ 1) {
5068 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
5069 krwlock_t
*lock
= (dn
->dn_dbuf
== NULL
? NULL
:
5070 &dn
->dn_dbuf
->db_rwlock
);
5071 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], lock
,
5080 * Populate dr->dr_zio with a zio to commit a dirty buffer to disk.
5081 * Caller is responsible for issuing the zio_[no]wait(dr->dr_zio).
5084 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
5086 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
5087 dnode_t
*dn
= dr
->dr_dnode
;
5089 dmu_buf_impl_t
*parent
= db
->db_parent
;
5090 uint64_t txg
= tx
->tx_txg
;
5091 zbookmark_phys_t zb
;
5093 zio_t
*pio
; /* parent I/O */
5096 ASSERT(dmu_tx_is_syncing(tx
));
5100 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
5102 * Private object buffers are released here rather than in
5103 * dbuf_dirty() since they are only modified in the syncing
5104 * context and we don't want the overhead of making multiple
5105 * copies of the data.
5107 if (BP_IS_HOLE(db
->db_blkptr
))
5110 dbuf_release_bp(db
);
5111 dbuf_remap(dn
, db
, tx
);
5114 if (parent
!= dn
->dn_dbuf
) {
5115 /* Our parent is an indirect block. */
5116 /* We have a dirty parent that has been scheduled for write. */
5117 ASSERT(parent
&& parent
->db_data_pending
);
5118 /* Our parent's buffer is one level closer to the dnode. */
5119 ASSERT(db
->db_level
== parent
->db_level
-1);
5121 * We're about to modify our parent's db_data by modifying
5122 * our block pointer, so the parent must be released.
5124 ASSERT(arc_released(parent
->db_buf
));
5125 pio
= parent
->db_data_pending
->dr_zio
;
5127 /* Our parent is the dnode itself. */
5128 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
5129 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
5130 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
5131 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
5132 ASSERT3P(db
->db_blkptr
, ==,
5133 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
5137 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
5138 ASSERT3U(BP_GET_LOGICAL_BIRTH(db
->db_blkptr
), <=, txg
);
5141 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
5142 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
5143 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
5145 if (db
->db_blkid
== DMU_SPILL_BLKID
)
5147 wp_flag
|= (data
== NULL
) ? WP_NOFILL
: 0;
5149 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
5152 * We copy the blkptr now (rather than when we instantiate the dirty
5153 * record), because its value can change between open context and
5154 * syncing context. We do not need to hold dn_struct_rwlock to read
5155 * db_blkptr because we are in syncing context.
5157 dr
->dr_bp_copy
= *db
->db_blkptr
;
5159 if (db
->db_level
== 0 &&
5160 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
5162 * The BP for this block has been provided by open context
5163 * (by dmu_sync() or dmu_buf_write_embedded()).
5165 abd_t
*contents
= (data
!= NULL
) ?
5166 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
5168 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
, &dr
->dr_bp_copy
,
5169 contents
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5170 dbuf_write_override_ready
, NULL
,
5171 dbuf_write_override_done
,
5172 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
5173 mutex_enter(&db
->db_mtx
);
5174 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
5175 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
5176 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
,
5177 dr
->dt
.dl
.dr_brtwrite
);
5178 mutex_exit(&db
->db_mtx
);
5179 } else if (data
== NULL
) {
5180 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
5181 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
5182 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
,
5183 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5184 dbuf_write_nofill_ready
, NULL
,
5185 dbuf_write_nofill_done
, db
,
5186 ZIO_PRIORITY_ASYNC_WRITE
,
5187 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
5189 ASSERT(arc_released(data
));
5192 * For indirect blocks, we want to setup the children
5193 * ready callback so that we can properly handle an indirect
5194 * block that only contains holes.
5196 arc_write_done_func_t
*children_ready_cb
= NULL
;
5197 if (db
->db_level
!= 0)
5198 children_ready_cb
= dbuf_write_children_ready
;
5200 dr
->dr_zio
= arc_write(pio
, os
->os_spa
, txg
,
5201 &dr
->dr_bp_copy
, data
, !DBUF_IS_CACHEABLE(db
),
5202 dbuf_is_l2cacheable(db
), &zp
, dbuf_write_ready
,
5203 children_ready_cb
, dbuf_write_done
, db
,
5204 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
5208 EXPORT_SYMBOL(dbuf_find
);
5209 EXPORT_SYMBOL(dbuf_is_metadata
);
5210 EXPORT_SYMBOL(dbuf_destroy
);
5211 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
5212 EXPORT_SYMBOL(dbuf_whichblock
);
5213 EXPORT_SYMBOL(dbuf_read
);
5214 EXPORT_SYMBOL(dbuf_unoverride
);
5215 EXPORT_SYMBOL(dbuf_free_range
);
5216 EXPORT_SYMBOL(dbuf_new_size
);
5217 EXPORT_SYMBOL(dbuf_release_bp
);
5218 EXPORT_SYMBOL(dbuf_dirty
);
5219 EXPORT_SYMBOL(dmu_buf_set_crypt_params
);
5220 EXPORT_SYMBOL(dmu_buf_will_dirty
);
5221 EXPORT_SYMBOL(dmu_buf_is_dirty
);
5222 EXPORT_SYMBOL(dmu_buf_will_clone
);
5223 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
5224 EXPORT_SYMBOL(dmu_buf_will_fill
);
5225 EXPORT_SYMBOL(dmu_buf_fill_done
);
5226 EXPORT_SYMBOL(dmu_buf_rele
);
5227 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
5228 EXPORT_SYMBOL(dbuf_prefetch
);
5229 EXPORT_SYMBOL(dbuf_hold_impl
);
5230 EXPORT_SYMBOL(dbuf_hold
);
5231 EXPORT_SYMBOL(dbuf_hold_level
);
5232 EXPORT_SYMBOL(dbuf_create_bonus
);
5233 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
5234 EXPORT_SYMBOL(dbuf_rm_spill
);
5235 EXPORT_SYMBOL(dbuf_add_ref
);
5236 EXPORT_SYMBOL(dbuf_rele
);
5237 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
5238 EXPORT_SYMBOL(dbuf_refcount
);
5239 EXPORT_SYMBOL(dbuf_sync_list
);
5240 EXPORT_SYMBOL(dmu_buf_set_user
);
5241 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
5242 EXPORT_SYMBOL(dmu_buf_get_user
);
5243 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
5245 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, max_bytes
, U64
, ZMOD_RW
,
5246 "Maximum size in bytes of the dbuf cache.");
5248 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, hiwater_pct
, UINT
, ZMOD_RW
,
5249 "Percentage over dbuf_cache_max_bytes for direct dbuf eviction.");
5251 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, lowater_pct
, UINT
, ZMOD_RW
,
5252 "Percentage below dbuf_cache_max_bytes when dbuf eviction stops.");
5254 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_max_bytes
, U64
, ZMOD_RW
,
5255 "Maximum size in bytes of dbuf metadata cache.");
5257 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, cache_shift
, UINT
, ZMOD_RW
,
5258 "Set size of dbuf cache to log2 fraction of arc size.");
5260 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_shift
, UINT
, ZMOD_RW
,
5261 "Set size of dbuf metadata cache to log2 fraction of arc size.");
5263 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, mutex_cache_shift
, UINT
, ZMOD_RD
,
5264 "Set size of dbuf cache mutex array as log2 shift.");