4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
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, 2019 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.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_zfs.h>
48 #include <sys/callb.h>
51 #include <sys/cityhash.h>
52 #include <sys/spa_impl.h>
56 typedef struct dbuf_stats
{
58 * Various statistics about the size of the dbuf cache.
60 kstat_named_t cache_count
;
61 kstat_named_t cache_size_bytes
;
62 kstat_named_t cache_size_bytes_max
;
64 * Statistics regarding the bounds on the dbuf cache size.
66 kstat_named_t cache_target_bytes
;
67 kstat_named_t cache_lowater_bytes
;
68 kstat_named_t cache_hiwater_bytes
;
70 * Total number of dbuf cache evictions that have occurred.
72 kstat_named_t cache_total_evicts
;
74 * The distribution of dbuf levels in the dbuf cache and
75 * the total size of all dbufs at each level.
77 kstat_named_t cache_levels
[DN_MAX_LEVELS
];
78 kstat_named_t cache_levels_bytes
[DN_MAX_LEVELS
];
80 * Statistics about the dbuf hash table.
82 kstat_named_t hash_hits
;
83 kstat_named_t hash_misses
;
84 kstat_named_t hash_collisions
;
85 kstat_named_t hash_elements
;
86 kstat_named_t hash_elements_max
;
88 * Number of sublists containing more than one dbuf in the dbuf
89 * hash table. Keep track of the longest hash chain.
91 kstat_named_t hash_chains
;
92 kstat_named_t hash_chain_max
;
94 * Number of times a dbuf_create() discovers that a dbuf was
95 * already created and in the dbuf hash table.
97 kstat_named_t hash_insert_race
;
99 * Statistics about the size of the metadata dbuf cache.
101 kstat_named_t metadata_cache_count
;
102 kstat_named_t metadata_cache_size_bytes
;
103 kstat_named_t metadata_cache_size_bytes_max
;
105 * For diagnostic purposes, this is incremented whenever we can't add
106 * something to the metadata cache because it's full, and instead put
107 * the data in the regular dbuf cache.
109 kstat_named_t metadata_cache_overflow
;
112 dbuf_stats_t dbuf_stats
= {
113 { "cache_count", KSTAT_DATA_UINT64
},
114 { "cache_size_bytes", KSTAT_DATA_UINT64
},
115 { "cache_size_bytes_max", KSTAT_DATA_UINT64
},
116 { "cache_target_bytes", KSTAT_DATA_UINT64
},
117 { "cache_lowater_bytes", KSTAT_DATA_UINT64
},
118 { "cache_hiwater_bytes", KSTAT_DATA_UINT64
},
119 { "cache_total_evicts", KSTAT_DATA_UINT64
},
120 { { "cache_levels_N", KSTAT_DATA_UINT64
} },
121 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64
} },
122 { "hash_hits", KSTAT_DATA_UINT64
},
123 { "hash_misses", KSTAT_DATA_UINT64
},
124 { "hash_collisions", KSTAT_DATA_UINT64
},
125 { "hash_elements", KSTAT_DATA_UINT64
},
126 { "hash_elements_max", KSTAT_DATA_UINT64
},
127 { "hash_chains", KSTAT_DATA_UINT64
},
128 { "hash_chain_max", KSTAT_DATA_UINT64
},
129 { "hash_insert_race", KSTAT_DATA_UINT64
},
130 { "metadata_cache_count", KSTAT_DATA_UINT64
},
131 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64
},
132 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64
},
133 { "metadata_cache_overflow", KSTAT_DATA_UINT64
}
136 #define DBUF_STAT_INCR(stat, val) \
137 atomic_add_64(&dbuf_stats.stat.value.ui64, (val));
138 #define DBUF_STAT_DECR(stat, val) \
139 DBUF_STAT_INCR(stat, -(val));
140 #define DBUF_STAT_BUMP(stat) \
141 DBUF_STAT_INCR(stat, 1);
142 #define DBUF_STAT_BUMPDOWN(stat) \
143 DBUF_STAT_INCR(stat, -1);
144 #define DBUF_STAT_MAX(stat, v) { \
146 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
147 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
151 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
152 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
154 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
155 dmu_buf_evict_func_t
*evict_func_sync
,
156 dmu_buf_evict_func_t
*evict_func_async
,
157 dmu_buf_t
**clear_on_evict_dbufp
);
160 * Global data structures and functions for the dbuf cache.
162 static kmem_cache_t
*dbuf_kmem_cache
;
163 static taskq_t
*dbu_evict_taskq
;
165 static kthread_t
*dbuf_cache_evict_thread
;
166 static kmutex_t dbuf_evict_lock
;
167 static kcondvar_t dbuf_evict_cv
;
168 static boolean_t dbuf_evict_thread_exit
;
171 * There are two dbuf caches; each dbuf can only be in one of them at a time.
173 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
174 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
175 * that represent the metadata that describes filesystems/snapshots/
176 * bookmarks/properties/etc. We only evict from this cache when we export a
177 * pool, to short-circuit as much I/O as possible for all administrative
178 * commands that need the metadata. There is no eviction policy for this
179 * cache, because we try to only include types in it which would occupy a
180 * very small amount of space per object but create a large impact on the
181 * performance of these commands. Instead, after it reaches a maximum size
182 * (which should only happen on very small memory systems with a very large
183 * number of filesystem objects), we stop taking new dbufs into the
184 * metadata cache, instead putting them in the normal dbuf cache.
186 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
187 * are not currently held but have been recently released. These dbufs
188 * are not eligible for arc eviction until they are aged out of the cache.
189 * Dbufs that are aged out of the cache will be immediately destroyed and
190 * become eligible for arc eviction.
192 * Dbufs are added to these caches once the last hold is released. If a dbuf is
193 * later accessed and still exists in the dbuf cache, then it will be removed
194 * from the cache and later re-added to the head of the cache.
196 * If a given dbuf meets the requirements for the metadata cache, it will go
197 * there, otherwise it will be considered for the generic LRU dbuf cache. The
198 * caches and the refcounts tracking their sizes are stored in an array indexed
199 * by those caches' matching enum values (from dbuf_cached_state_t).
201 typedef struct dbuf_cache
{
205 dbuf_cache_t dbuf_caches
[DB_CACHE_MAX
];
207 /* Size limits for the caches */
208 unsigned long dbuf_cache_max_bytes
= 0;
209 unsigned long dbuf_metadata_cache_max_bytes
= 0;
210 /* Set the default sizes of the caches to log2 fraction of arc size */
211 int dbuf_cache_shift
= 5;
212 int dbuf_metadata_cache_shift
= 6;
215 * The LRU dbuf cache uses a three-stage eviction policy:
216 * - A low water marker designates when the dbuf eviction thread
217 * should stop evicting from the dbuf cache.
218 * - When we reach the maximum size (aka mid water mark), we
219 * signal the eviction thread to run.
220 * - The high water mark indicates when the eviction thread
221 * is unable to keep up with the incoming load and eviction must
222 * happen in the context of the calling thread.
226 * low water mid water hi water
227 * +----------------------------------------+----------+----------+
232 * +----------------------------------------+----------+----------+
234 * evicting eviction directly
237 * The high and low water marks indicate the operating range for the eviction
238 * thread. The low water mark is, by default, 90% of the total size of the
239 * cache and the high water mark is at 110% (both of these percentages can be
240 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
241 * respectively). The eviction thread will try to ensure that the cache remains
242 * within this range by waking up every second and checking if the cache is
243 * above the low water mark. The thread can also be woken up by callers adding
244 * elements into the cache if the cache is larger than the mid water (i.e max
245 * cache size). Once the eviction thread is woken up and eviction is required,
246 * it will continue evicting buffers until it's able to reduce the cache size
247 * to the low water mark. If the cache size continues to grow and hits the high
248 * water mark, then callers adding elements to the cache will begin to evict
249 * directly from the cache until the cache is no longer above the high water
254 * The percentage above and below the maximum cache size.
256 uint_t dbuf_cache_hiwater_pct
= 10;
257 uint_t dbuf_cache_lowater_pct
= 10;
261 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
263 dmu_buf_impl_t
*db
= vdb
;
264 bzero(db
, sizeof (dmu_buf_impl_t
));
266 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
267 rw_init(&db
->db_rwlock
, NULL
, RW_DEFAULT
, NULL
);
268 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
269 multilist_link_init(&db
->db_cache_link
);
270 zfs_refcount_create(&db
->db_holds
);
277 dbuf_dest(void *vdb
, void *unused
)
279 dmu_buf_impl_t
*db
= vdb
;
280 mutex_destroy(&db
->db_mtx
);
281 rw_destroy(&db
->db_rwlock
);
282 cv_destroy(&db
->db_changed
);
283 ASSERT(!multilist_link_active(&db
->db_cache_link
));
284 zfs_refcount_destroy(&db
->db_holds
);
288 * dbuf hash table routines
290 static dbuf_hash_table_t dbuf_hash_table
;
292 static uint64_t dbuf_hash_count
;
295 * We use Cityhash for this. It's fast, and has good hash properties without
296 * requiring any large static buffers.
299 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
301 return (cityhash4((uintptr_t)os
, obj
, (uint64_t)lvl
, blkid
));
304 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
305 ((dbuf)->db.db_object == (obj) && \
306 (dbuf)->db_objset == (os) && \
307 (dbuf)->db_level == (level) && \
308 (dbuf)->db_blkid == (blkid))
311 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
313 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
318 hv
= dbuf_hash(os
, obj
, level
, blkid
);
319 idx
= hv
& h
->hash_table_mask
;
321 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
322 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
323 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
324 mutex_enter(&db
->db_mtx
);
325 if (db
->db_state
!= DB_EVICTING
) {
326 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
329 mutex_exit(&db
->db_mtx
);
332 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
336 static dmu_buf_impl_t
*
337 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
340 dmu_buf_impl_t
*db
= NULL
;
342 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
343 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
344 if (dn
->dn_bonus
!= NULL
) {
346 mutex_enter(&db
->db_mtx
);
348 rw_exit(&dn
->dn_struct_rwlock
);
349 dnode_rele(dn
, FTAG
);
355 * Insert an entry into the hash table. If there is already an element
356 * equal to elem in the hash table, then the already existing element
357 * will be returned and the new element will not be inserted.
358 * Otherwise returns NULL.
360 static dmu_buf_impl_t
*
361 dbuf_hash_insert(dmu_buf_impl_t
*db
)
363 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
364 objset_t
*os
= db
->db_objset
;
365 uint64_t obj
= db
->db
.db_object
;
366 int level
= db
->db_level
;
367 uint64_t blkid
, hv
, idx
;
371 blkid
= db
->db_blkid
;
372 hv
= dbuf_hash(os
, obj
, level
, blkid
);
373 idx
= hv
& h
->hash_table_mask
;
375 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
376 for (dbf
= h
->hash_table
[idx
], i
= 0; dbf
!= NULL
;
377 dbf
= dbf
->db_hash_next
, i
++) {
378 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
379 mutex_enter(&dbf
->db_mtx
);
380 if (dbf
->db_state
!= DB_EVICTING
) {
381 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
384 mutex_exit(&dbf
->db_mtx
);
389 DBUF_STAT_BUMP(hash_collisions
);
391 DBUF_STAT_BUMP(hash_chains
);
393 DBUF_STAT_MAX(hash_chain_max
, i
);
396 mutex_enter(&db
->db_mtx
);
397 db
->db_hash_next
= h
->hash_table
[idx
];
398 h
->hash_table
[idx
] = db
;
399 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
400 atomic_inc_64(&dbuf_hash_count
);
401 DBUF_STAT_MAX(hash_elements_max
, dbuf_hash_count
);
407 * This returns whether this dbuf should be stored in the metadata cache, which
408 * is based on whether it's from one of the dnode types that store data related
409 * to traversing dataset hierarchies.
412 dbuf_include_in_metadata_cache(dmu_buf_impl_t
*db
)
415 dmu_object_type_t type
= DB_DNODE(db
)->dn_type
;
418 /* Check if this dbuf is one of the types we care about */
419 if (DMU_OT_IS_METADATA_CACHED(type
)) {
420 /* If we hit this, then we set something up wrong in dmu_ot */
421 ASSERT(DMU_OT_IS_METADATA(type
));
424 * Sanity check for small-memory systems: don't allocate too
425 * much memory for this purpose.
427 if (zfs_refcount_count(
428 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
) >
429 dbuf_metadata_cache_max_bytes
) {
430 DBUF_STAT_BUMP(metadata_cache_overflow
);
441 * Remove an entry from the hash table. It must be in the EVICTING state.
444 dbuf_hash_remove(dmu_buf_impl_t
*db
)
446 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
448 dmu_buf_impl_t
*dbf
, **dbp
;
450 hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
451 db
->db_level
, db
->db_blkid
);
452 idx
= hv
& h
->hash_table_mask
;
455 * We mustn't hold db_mtx to maintain lock ordering:
456 * DBUF_HASH_MUTEX > db_mtx.
458 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
459 ASSERT(db
->db_state
== DB_EVICTING
);
460 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
462 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
463 dbp
= &h
->hash_table
[idx
];
464 while ((dbf
= *dbp
) != db
) {
465 dbp
= &dbf
->db_hash_next
;
468 *dbp
= db
->db_hash_next
;
469 db
->db_hash_next
= NULL
;
470 if (h
->hash_table
[idx
] &&
471 h
->hash_table
[idx
]->db_hash_next
== NULL
)
472 DBUF_STAT_BUMPDOWN(hash_chains
);
473 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
474 atomic_dec_64(&dbuf_hash_count
);
480 } dbvu_verify_type_t
;
483 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
488 if (db
->db_user
== NULL
)
491 /* Only data blocks support the attachment of user data. */
492 ASSERT(db
->db_level
== 0);
494 /* Clients must resolve a dbuf before attaching user data. */
495 ASSERT(db
->db
.db_data
!= NULL
);
496 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
498 holds
= zfs_refcount_count(&db
->db_holds
);
499 if (verify_type
== DBVU_EVICTING
) {
501 * Immediate eviction occurs when holds == dirtycnt.
502 * For normal eviction buffers, holds is zero on
503 * eviction, except when dbuf_fix_old_data() calls
504 * dbuf_clear_data(). However, the hold count can grow
505 * during eviction even though db_mtx is held (see
506 * dmu_bonus_hold() for an example), so we can only
507 * test the generic invariant that holds >= dirtycnt.
509 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
511 if (db
->db_user_immediate_evict
== TRUE
)
512 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
514 ASSERT3U(holds
, >, 0);
520 dbuf_evict_user(dmu_buf_impl_t
*db
)
522 dmu_buf_user_t
*dbu
= db
->db_user
;
524 ASSERT(MUTEX_HELD(&db
->db_mtx
));
529 dbuf_verify_user(db
, DBVU_EVICTING
);
533 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
534 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
538 * There are two eviction callbacks - one that we call synchronously
539 * and one that we invoke via a taskq. The async one is useful for
540 * avoiding lock order reversals and limiting stack depth.
542 * Note that if we have a sync callback but no async callback,
543 * it's likely that the sync callback will free the structure
544 * containing the dbu. In that case we need to take care to not
545 * dereference dbu after calling the sync evict func.
547 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
549 if (dbu
->dbu_evict_func_sync
!= NULL
)
550 dbu
->dbu_evict_func_sync(dbu
);
553 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
554 dbu
, 0, &dbu
->dbu_tqent
);
559 dbuf_is_metadata(dmu_buf_impl_t
*db
)
562 * Consider indirect blocks and spill blocks to be meta data.
564 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
567 boolean_t is_metadata
;
570 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
573 return (is_metadata
);
579 * This function *must* return indices evenly distributed between all
580 * sublists of the multilist. This is needed due to how the dbuf eviction
581 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
582 * distributed between all sublists and uses this assumption when
583 * deciding which sublist to evict from and how much to evict from it.
586 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
588 dmu_buf_impl_t
*db
= obj
;
591 * The assumption here, is the hash value for a given
592 * dmu_buf_impl_t will remain constant throughout it's lifetime
593 * (i.e. it's objset, object, level and blkid fields don't change).
594 * Thus, we don't need to store the dbuf's sublist index
595 * on insertion, as this index can be recalculated on removal.
597 * Also, the low order bits of the hash value are thought to be
598 * distributed evenly. Otherwise, in the case that the multilist
599 * has a power of two number of sublists, each sublists' usage
600 * would not be evenly distributed.
602 return (dbuf_hash(db
->db_objset
, db
->db
.db_object
,
603 db
->db_level
, db
->db_blkid
) %
604 multilist_get_num_sublists(ml
));
607 static inline unsigned long
608 dbuf_cache_target_bytes(void)
610 return MIN(dbuf_cache_max_bytes
,
611 arc_target_bytes() >> dbuf_cache_shift
);
614 static inline uint64_t
615 dbuf_cache_hiwater_bytes(void)
617 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
618 return (dbuf_cache_target
+
619 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100);
622 static inline uint64_t
623 dbuf_cache_lowater_bytes(void)
625 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
626 return (dbuf_cache_target
-
627 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100);
630 static inline boolean_t
631 dbuf_cache_above_hiwater(void)
633 return (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
634 dbuf_cache_hiwater_bytes());
637 static inline boolean_t
638 dbuf_cache_above_lowater(void)
640 return (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
641 dbuf_cache_lowater_bytes());
645 * Evict the oldest eligible dbuf from the dbuf cache.
650 int idx
= multilist_get_random_index(dbuf_caches
[DB_DBUF_CACHE
].cache
);
651 multilist_sublist_t
*mls
= multilist_sublist_lock(
652 dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
654 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
656 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
657 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
658 db
= multilist_sublist_prev(mls
, db
);
661 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
662 multilist_sublist_t
*, mls
);
665 multilist_sublist_remove(mls
, db
);
666 multilist_sublist_unlock(mls
);
667 (void) zfs_refcount_remove_many(
668 &dbuf_caches
[DB_DBUF_CACHE
].size
, db
->db
.db_size
, db
);
669 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
670 DBUF_STAT_BUMPDOWN(cache_count
);
671 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
673 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
674 db
->db_caching_status
= DB_NO_CACHE
;
676 DBUF_STAT_MAX(cache_size_bytes_max
,
677 zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
));
678 DBUF_STAT_BUMP(cache_total_evicts
);
680 multilist_sublist_unlock(mls
);
685 * The dbuf evict thread is responsible for aging out dbufs from the
686 * cache. Once the cache has reached it's maximum size, dbufs are removed
687 * and destroyed. The eviction thread will continue running until the size
688 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
689 * out of the cache it is destroyed and becomes eligible for arc eviction.
693 dbuf_evict_thread(void *unused
)
697 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
699 mutex_enter(&dbuf_evict_lock
);
700 while (!dbuf_evict_thread_exit
) {
701 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
702 CALLB_CPR_SAFE_BEGIN(&cpr
);
703 (void) cv_timedwait_sig_hires(&dbuf_evict_cv
,
704 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
705 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
707 mutex_exit(&dbuf_evict_lock
);
710 * Keep evicting as long as we're above the low water mark
711 * for the cache. We do this without holding the locks to
712 * minimize lock contention.
714 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
718 mutex_enter(&dbuf_evict_lock
);
721 dbuf_evict_thread_exit
= B_FALSE
;
722 cv_broadcast(&dbuf_evict_cv
);
723 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
728 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
729 * If the dbuf cache is at its high water mark, then evict a dbuf from the
730 * dbuf cache using the callers context.
733 dbuf_evict_notify(void)
736 * We check if we should evict without holding the dbuf_evict_lock,
737 * because it's OK to occasionally make the wrong decision here,
738 * and grabbing the lock results in massive lock contention.
740 if (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
741 dbuf_cache_target_bytes()) {
742 if (dbuf_cache_above_hiwater())
744 cv_signal(&dbuf_evict_cv
);
749 dbuf_kstat_update(kstat_t
*ksp
, int rw
)
751 dbuf_stats_t
*ds
= ksp
->ks_data
;
753 if (rw
== KSTAT_WRITE
) {
754 return (SET_ERROR(EACCES
));
756 ds
->metadata_cache_size_bytes
.value
.ui64
= zfs_refcount_count(
757 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
);
758 ds
->cache_size_bytes
.value
.ui64
=
759 zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
);
760 ds
->cache_target_bytes
.value
.ui64
= dbuf_cache_target_bytes();
761 ds
->cache_hiwater_bytes
.value
.ui64
= dbuf_cache_hiwater_bytes();
762 ds
->cache_lowater_bytes
.value
.ui64
= dbuf_cache_lowater_bytes();
763 ds
->hash_elements
.value
.ui64
= dbuf_hash_count
;
772 uint64_t hsize
= 1ULL << 16;
773 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
777 * The hash table is big enough to fill all of physical memory
778 * with an average block size of zfs_arc_average_blocksize (default 8K).
779 * By default, the table will take up
780 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
782 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
786 h
->hash_table_mask
= hsize
- 1;
789 * Large allocations which do not require contiguous pages
790 * should be using vmem_alloc() in the linux kernel
792 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
794 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
796 if (h
->hash_table
== NULL
) {
797 /* XXX - we should really return an error instead of assert */
798 ASSERT(hsize
> (1ULL << 10));
803 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
804 sizeof (dmu_buf_impl_t
),
805 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
807 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
808 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
813 * Setup the parameters for the dbuf caches. We set the sizes of the
814 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
815 * of the target size of the ARC. If the values has been specified as
816 * a module option and they're not greater than the target size of the
817 * ARC, then we honor that value.
819 if (dbuf_cache_max_bytes
== 0 ||
820 dbuf_cache_max_bytes
>= arc_target_bytes()) {
821 dbuf_cache_max_bytes
= arc_target_bytes() >> dbuf_cache_shift
;
823 if (dbuf_metadata_cache_max_bytes
== 0 ||
824 dbuf_metadata_cache_max_bytes
>= arc_target_bytes()) {
825 dbuf_metadata_cache_max_bytes
=
826 arc_target_bytes() >> dbuf_metadata_cache_shift
;
830 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
831 * configuration is not required.
833 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
835 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
836 dbuf_caches
[dcs
].cache
=
837 multilist_create(sizeof (dmu_buf_impl_t
),
838 offsetof(dmu_buf_impl_t
, db_cache_link
),
839 dbuf_cache_multilist_index_func
);
840 zfs_refcount_create(&dbuf_caches
[dcs
].size
);
843 dbuf_evict_thread_exit
= B_FALSE
;
844 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
845 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
846 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
847 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
849 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
850 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
852 if (dbuf_ksp
!= NULL
) {
853 for (i
= 0; i
< DN_MAX_LEVELS
; i
++) {
854 snprintf(dbuf_stats
.cache_levels
[i
].name
,
855 KSTAT_STRLEN
, "cache_level_%d", i
);
856 dbuf_stats
.cache_levels
[i
].data_type
=
858 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
859 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
860 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
863 dbuf_ksp
->ks_data
= &dbuf_stats
;
864 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
865 kstat_install(dbuf_ksp
);
872 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
875 dbuf_stats_destroy();
877 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
878 mutex_destroy(&h
->hash_mutexes
[i
]);
881 * Large allocations which do not require contiguous pages
882 * should be using vmem_free() in the linux kernel
884 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
886 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
888 kmem_cache_destroy(dbuf_kmem_cache
);
889 taskq_destroy(dbu_evict_taskq
);
891 mutex_enter(&dbuf_evict_lock
);
892 dbuf_evict_thread_exit
= B_TRUE
;
893 while (dbuf_evict_thread_exit
) {
894 cv_signal(&dbuf_evict_cv
);
895 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
897 mutex_exit(&dbuf_evict_lock
);
899 mutex_destroy(&dbuf_evict_lock
);
900 cv_destroy(&dbuf_evict_cv
);
902 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
903 zfs_refcount_destroy(&dbuf_caches
[dcs
].size
);
904 multilist_destroy(dbuf_caches
[dcs
].cache
);
907 if (dbuf_ksp
!= NULL
) {
908 kstat_delete(dbuf_ksp
);
919 dbuf_verify(dmu_buf_impl_t
*db
)
922 dbuf_dirty_record_t
*dr
;
924 ASSERT(MUTEX_HELD(&db
->db_mtx
));
926 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
929 ASSERT(db
->db_objset
!= NULL
);
933 ASSERT(db
->db_parent
== NULL
);
934 ASSERT(db
->db_blkptr
== NULL
);
936 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
937 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
938 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
939 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
940 db
->db_blkid
== DMU_SPILL_BLKID
||
941 !avl_is_empty(&dn
->dn_dbufs
));
943 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
945 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
946 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
947 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
949 ASSERT0(db
->db
.db_offset
);
951 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
954 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
955 ASSERT(dr
->dr_dbuf
== db
);
957 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
958 ASSERT(dr
->dr_dbuf
== db
);
961 * We can't assert that db_size matches dn_datablksz because it
962 * can be momentarily different when another thread is doing
965 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
966 dr
= db
->db_data_pending
;
968 * It should only be modified in syncing context, so
969 * make sure we only have one copy of the data.
971 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
974 /* verify db->db_blkptr */
976 if (db
->db_parent
== dn
->dn_dbuf
) {
977 /* db is pointed to by the dnode */
978 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
979 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
980 ASSERT(db
->db_parent
== NULL
);
982 ASSERT(db
->db_parent
!= NULL
);
983 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
984 ASSERT3P(db
->db_blkptr
, ==,
985 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
987 /* db is pointed to by an indirect block */
988 int epb __maybe_unused
= db
->db_parent
->db
.db_size
>>
990 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
991 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
994 * dnode_grow_indblksz() can make this fail if we don't
995 * have the parent's rwlock. XXX indblksz no longer
996 * grows. safe to do this now?
998 if (RW_LOCK_HELD(&db
->db_parent
->db_rwlock
)) {
999 ASSERT3P(db
->db_blkptr
, ==,
1000 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
1001 db
->db_blkid
% epb
));
1005 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
1006 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
1007 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
1008 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
1010 * If the blkptr isn't set but they have nonzero data,
1011 * it had better be dirty, otherwise we'll lose that
1012 * data when we evict this buffer.
1014 * There is an exception to this rule for indirect blocks; in
1015 * this case, if the indirect block is a hole, we fill in a few
1016 * fields on each of the child blocks (importantly, birth time)
1017 * to prevent hole birth times from being lost when you
1018 * partially fill in a hole.
1020 if (db
->db_dirtycnt
== 0) {
1021 if (db
->db_level
== 0) {
1022 uint64_t *buf
= db
->db
.db_data
;
1025 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
1026 ASSERT(buf
[i
] == 0);
1029 blkptr_t
*bps
= db
->db
.db_data
;
1030 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1033 * We want to verify that all the blkptrs in the
1034 * indirect block are holes, but we may have
1035 * automatically set up a few fields for them.
1036 * We iterate through each blkptr and verify
1037 * they only have those fields set.
1040 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1042 blkptr_t
*bp
= &bps
[i
];
1043 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1046 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1047 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1048 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1049 ASSERT0(bp
->blk_fill
);
1050 ASSERT0(bp
->blk_pad
[0]);
1051 ASSERT0(bp
->blk_pad
[1]);
1052 ASSERT(!BP_IS_EMBEDDED(bp
));
1053 ASSERT(BP_IS_HOLE(bp
));
1054 ASSERT0(bp
->blk_phys_birth
);
1064 dbuf_clear_data(dmu_buf_impl_t
*db
)
1066 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1067 dbuf_evict_user(db
);
1068 ASSERT3P(db
->db_buf
, ==, NULL
);
1069 db
->db
.db_data
= NULL
;
1070 if (db
->db_state
!= DB_NOFILL
)
1071 db
->db_state
= DB_UNCACHED
;
1075 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1077 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1078 ASSERT(buf
!= NULL
);
1081 ASSERT(buf
->b_data
!= NULL
);
1082 db
->db
.db_data
= buf
->b_data
;
1086 * Loan out an arc_buf for read. Return the loaned arc_buf.
1089 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1093 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1094 mutex_enter(&db
->db_mtx
);
1095 if (arc_released(db
->db_buf
) || zfs_refcount_count(&db
->db_holds
) > 1) {
1096 int blksz
= db
->db
.db_size
;
1097 spa_t
*spa
= db
->db_objset
->os_spa
;
1099 mutex_exit(&db
->db_mtx
);
1100 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1101 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
1104 arc_loan_inuse_buf(abuf
, db
);
1106 dbuf_clear_data(db
);
1107 mutex_exit(&db
->db_mtx
);
1113 * Calculate which level n block references the data at the level 0 offset
1117 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1119 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1121 * The level n blkid is equal to the level 0 blkid divided by
1122 * the number of level 0s in a level n block.
1124 * The level 0 blkid is offset >> datablkshift =
1125 * offset / 2^datablkshift.
1127 * The number of level 0s in a level n is the number of block
1128 * pointers in an indirect block, raised to the power of level.
1129 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1130 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1132 * Thus, the level n blkid is: offset /
1133 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1134 * = offset / 2^(datablkshift + level *
1135 * (indblkshift - SPA_BLKPTRSHIFT))
1136 * = offset >> (datablkshift + level *
1137 * (indblkshift - SPA_BLKPTRSHIFT))
1140 const unsigned exp
= dn
->dn_datablkshift
+
1141 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1143 if (exp
>= 8 * sizeof (offset
)) {
1144 /* This only happens on the highest indirection level */
1145 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1149 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1151 return (offset
>> exp
);
1153 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1159 * This function is used to lock the parent of the provided dbuf. This should be
1160 * used when modifying or reading db_blkptr.
1163 dmu_buf_lock_parent(dmu_buf_impl_t
*db
, krw_t rw
, void *tag
)
1165 enum db_lock_type ret
= DLT_NONE
;
1166 if (db
->db_parent
!= NULL
) {
1167 rw_enter(&db
->db_parent
->db_rwlock
, rw
);
1169 } else if (dmu_objset_ds(db
->db_objset
) != NULL
) {
1170 rrw_enter(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, rw
,
1175 * We only return a DLT_NONE lock when it's the top-most indirect block
1176 * of the meta-dnode of the MOS.
1182 * We need to pass the lock type in because it's possible that the block will
1183 * move from being the topmost indirect block in a dnode (and thus, have no
1184 * parent) to not the top-most via an indirection increase. This would cause a
1185 * panic if we didn't pass the lock type in.
1188 dmu_buf_unlock_parent(dmu_buf_impl_t
*db
, db_lock_type_t type
, void *tag
)
1190 if (type
== DLT_PARENT
)
1191 rw_exit(&db
->db_parent
->db_rwlock
);
1192 else if (type
== DLT_OBJSET
)
1193 rrw_exit(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, tag
);
1197 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1198 arc_buf_t
*buf
, void *vdb
)
1200 dmu_buf_impl_t
*db
= vdb
;
1202 mutex_enter(&db
->db_mtx
);
1203 ASSERT3U(db
->db_state
, ==, DB_READ
);
1205 * All reads are synchronous, so we must have a hold on the dbuf
1207 ASSERT(zfs_refcount_count(&db
->db_holds
) > 0);
1208 ASSERT(db
->db_buf
== NULL
);
1209 ASSERT(db
->db
.db_data
== NULL
);
1212 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
1213 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1214 ASSERT3P(db
->db_buf
, ==, NULL
);
1215 db
->db_state
= DB_UNCACHED
;
1216 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1217 /* freed in flight */
1218 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1219 arc_release(buf
, db
);
1220 bzero(buf
->b_data
, db
->db
.db_size
);
1221 arc_buf_freeze(buf
);
1222 db
->db_freed_in_flight
= FALSE
;
1223 dbuf_set_data(db
, buf
);
1224 db
->db_state
= DB_CACHED
;
1227 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1228 dbuf_set_data(db
, buf
);
1229 db
->db_state
= DB_CACHED
;
1231 cv_broadcast(&db
->db_changed
);
1232 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
1237 * This function ensures that, when doing a decrypting read of a block,
1238 * we make sure we have decrypted the dnode associated with it. We must do
1239 * this so that we ensure we are fully authenticating the checksum-of-MACs
1240 * tree from the root of the objset down to this block. Indirect blocks are
1241 * always verified against their secure checksum-of-MACs assuming that the
1242 * dnode containing them is correct. Now that we are doing a decrypting read,
1243 * we can be sure that the key is loaded and verify that assumption. This is
1244 * especially important considering that we always read encrypted dnode
1245 * blocks as raw data (without verifying their MACs) to start, and
1246 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1249 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
)
1252 objset_t
*os
= db
->db_objset
;
1253 arc_buf_t
*dnode_abuf
;
1255 zbookmark_phys_t zb
;
1257 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1259 if (!os
->os_encrypted
|| os
->os_raw_receive
||
1260 (flags
& DB_RF_NO_DECRYPT
) != 0)
1265 dnode_abuf
= (dn
->dn_dbuf
!= NULL
) ? dn
->dn_dbuf
->db_buf
: NULL
;
1267 if (dnode_abuf
== NULL
|| !arc_is_encrypted(dnode_abuf
)) {
1272 SET_BOOKMARK(&zb
, dmu_objset_id(os
),
1273 DMU_META_DNODE_OBJECT
, 0, dn
->dn_dbuf
->db_blkid
);
1274 err
= arc_untransform(dnode_abuf
, os
->os_spa
, &zb
, B_TRUE
);
1277 * An error code of EACCES tells us that the key is still not
1278 * available. This is ok if we are only reading authenticated
1279 * (and therefore non-encrypted) blocks.
1281 if (err
== EACCES
&& ((db
->db_blkid
!= DMU_BONUS_BLKID
&&
1282 !DMU_OT_IS_ENCRYPTED(dn
->dn_type
)) ||
1283 (db
->db_blkid
== DMU_BONUS_BLKID
&&
1284 !DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))))
1293 * Drops db_mtx and the parent lock specified by dblt and tag before
1297 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
,
1298 db_lock_type_t dblt
, void *tag
)
1301 zbookmark_phys_t zb
;
1302 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1303 int err
, zio_flags
= 0;
1307 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1308 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1309 ASSERT(db
->db_state
== DB_UNCACHED
);
1310 ASSERT(db
->db_buf
== NULL
);
1311 ASSERT(db
->db_parent
== NULL
||
1312 RW_LOCK_HELD(&db
->db_parent
->db_rwlock
));
1314 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1316 * The bonus length stored in the dnode may be less than
1317 * the maximum available space in the bonus buffer.
1319 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1320 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1322 /* if the underlying dnode block is encrypted, decrypt it */
1323 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1326 mutex_exit(&db
->db_mtx
);
1330 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1331 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1332 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1333 if (bonuslen
< max_bonuslen
)
1334 bzero(db
->db
.db_data
, max_bonuslen
);
1336 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1338 db
->db_state
= DB_CACHED
;
1339 mutex_exit(&db
->db_mtx
);
1340 dmu_buf_unlock_parent(db
, dblt
, tag
);
1345 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1346 * processes the delete record and clears the bp while we are waiting
1347 * for the dn_mtx (resulting in a "no" from block_freed).
1349 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1350 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1351 BP_IS_HOLE(db
->db_blkptr
)))) {
1352 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1354 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1356 bzero(db
->db
.db_data
, db
->db
.db_size
);
1358 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1359 BP_IS_HOLE(db
->db_blkptr
) &&
1360 db
->db_blkptr
->blk_birth
!= 0) {
1361 blkptr_t
*bps
= db
->db
.db_data
;
1362 for (int i
= 0; i
< ((1 <<
1363 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1365 blkptr_t
*bp
= &bps
[i
];
1366 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1367 1 << dn
->dn_indblkshift
);
1369 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1371 BP_GET_LSIZE(db
->db_blkptr
));
1372 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1374 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1375 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1379 db
->db_state
= DB_CACHED
;
1380 mutex_exit(&db
->db_mtx
);
1381 dmu_buf_unlock_parent(db
, dblt
, tag
);
1386 * Any attempt to read a redacted block should result in an error. This
1387 * will never happen under normal conditions, but can be useful for
1388 * debugging purposes.
1390 if (BP_IS_REDACTED(db
->db_blkptr
)) {
1391 ASSERT(dsl_dataset_feature_is_active(
1392 db
->db_objset
->os_dsl_dataset
,
1393 SPA_FEATURE_REDACTED_DATASETS
));
1395 mutex_exit(&db
->db_mtx
);
1396 return (SET_ERROR(EIO
));
1400 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1401 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1404 * All bps of an encrypted os should have the encryption bit set.
1405 * If this is not true it indicates tampering and we report an error.
1407 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(db
->db_blkptr
)) {
1408 spa_log_error(db
->db_objset
->os_spa
, &zb
);
1409 zfs_panic_recover("unencrypted block in encrypted "
1410 "object set %llu", dmu_objset_id(db
->db_objset
));
1412 mutex_exit(&db
->db_mtx
);
1413 dmu_buf_unlock_parent(db
, dblt
, tag
);
1414 return (SET_ERROR(EIO
));
1417 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1420 dmu_buf_unlock_parent(db
, dblt
, tag
);
1421 mutex_exit(&db
->db_mtx
);
1427 db
->db_state
= DB_READ
;
1428 mutex_exit(&db
->db_mtx
);
1430 if (DBUF_IS_L2CACHEABLE(db
))
1431 aflags
|= ARC_FLAG_L2CACHE
;
1433 dbuf_add_ref(db
, NULL
);
1435 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1436 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1438 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1439 zio_flags
|= ZIO_FLAG_RAW
;
1441 * The zio layer will copy the provided blkptr later, but we need to
1442 * do this now so that we can release the parent's rwlock. We have to
1443 * do that now so that if dbuf_read_done is called synchronously (on
1444 * an l1 cache hit) we don't acquire the db_mtx while holding the
1445 * parent's rwlock, which would be a lock ordering violation.
1447 blkptr_t bp
= *db
->db_blkptr
;
1448 dmu_buf_unlock_parent(db
, dblt
, tag
);
1449 (void) arc_read(zio
, db
->db_objset
->os_spa
, &bp
,
1450 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1456 * This is our just-in-time copy function. It makes a copy of buffers that
1457 * have been modified in a previous transaction group before we access them in
1458 * the current active group.
1460 * This function is used in three places: when we are dirtying a buffer for the
1461 * first time in a txg, when we are freeing a range in a dnode that includes
1462 * this buffer, and when we are accessing a buffer which was received compressed
1463 * and later referenced in a WRITE_BYREF record.
1465 * Note that when we are called from dbuf_free_range() we do not put a hold on
1466 * the buffer, we just traverse the active dbuf list for the dnode.
1469 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1471 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1473 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1474 ASSERT(db
->db
.db_data
!= NULL
);
1475 ASSERT(db
->db_level
== 0);
1476 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1479 (dr
->dt
.dl
.dr_data
!=
1480 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1484 * If the last dirty record for this dbuf has not yet synced
1485 * and its referencing the dbuf data, either:
1486 * reset the reference to point to a new copy,
1487 * or (if there a no active holders)
1488 * just null out the current db_data pointer.
1490 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1491 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1492 dnode_t
*dn
= DB_DNODE(db
);
1493 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1494 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1495 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1496 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1497 } else if (zfs_refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1498 dnode_t
*dn
= DB_DNODE(db
);
1499 int size
= arc_buf_size(db
->db_buf
);
1500 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1501 spa_t
*spa
= db
->db_objset
->os_spa
;
1502 enum zio_compress compress_type
=
1503 arc_get_compression(db
->db_buf
);
1505 if (arc_is_encrypted(db
->db_buf
)) {
1506 boolean_t byteorder
;
1507 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1508 uint8_t iv
[ZIO_DATA_IV_LEN
];
1509 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1511 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1513 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1514 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1515 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1517 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1518 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1519 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1520 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1522 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1524 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1527 dbuf_clear_data(db
);
1532 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1539 * We don't have to hold the mutex to check db_state because it
1540 * can't be freed while we have a hold on the buffer.
1542 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1544 if (db
->db_state
== DB_NOFILL
)
1545 return (SET_ERROR(EIO
));
1550 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1551 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1552 DBUF_IS_CACHEABLE(db
);
1554 mutex_enter(&db
->db_mtx
);
1555 if (db
->db_state
== DB_CACHED
) {
1556 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1559 * Ensure that this block's dnode has been decrypted if
1560 * the caller has requested decrypted data.
1562 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1565 * If the arc buf is compressed or encrypted and the caller
1566 * requested uncompressed data, we need to untransform it
1567 * before returning. We also call arc_untransform() on any
1568 * unauthenticated blocks, which will verify their MAC if
1569 * the key is now available.
1571 if (err
== 0 && db
->db_buf
!= NULL
&&
1572 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1573 (arc_is_encrypted(db
->db_buf
) ||
1574 arc_is_unauthenticated(db
->db_buf
) ||
1575 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1576 zbookmark_phys_t zb
;
1578 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1579 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1580 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1581 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1582 dbuf_set_data(db
, db
->db_buf
);
1584 mutex_exit(&db
->db_mtx
);
1585 if (err
== 0 && prefetch
) {
1586 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1587 flags
& DB_RF_HAVESTRUCT
);
1590 DBUF_STAT_BUMP(hash_hits
);
1591 } else if (db
->db_state
== DB_UNCACHED
) {
1592 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1593 boolean_t need_wait
= B_FALSE
;
1595 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
1598 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1599 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1602 err
= dbuf_read_impl(db
, zio
, flags
, dblt
, FTAG
);
1604 * dbuf_read_impl has dropped db_mtx and our parent's rwlock
1607 if (!err
&& prefetch
) {
1608 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1609 flags
& DB_RF_HAVESTRUCT
);
1613 DBUF_STAT_BUMP(hash_misses
);
1616 * If we created a zio_root we must execute it to avoid
1617 * leaking it, even if it isn't attached to any work due
1618 * to an error in dbuf_read_impl().
1622 err
= zio_wait(zio
);
1624 VERIFY0(zio_wait(zio
));
1628 * Another reader came in while the dbuf was in flight
1629 * between UNCACHED and CACHED. Either a writer will finish
1630 * writing the buffer (sending the dbuf to CACHED) or the
1631 * first reader's request will reach the read_done callback
1632 * and send the dbuf to CACHED. Otherwise, a failure
1633 * occurred and the dbuf went to UNCACHED.
1635 mutex_exit(&db
->db_mtx
);
1637 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1638 flags
& DB_RF_HAVESTRUCT
);
1641 DBUF_STAT_BUMP(hash_misses
);
1643 /* Skip the wait per the caller's request. */
1644 mutex_enter(&db
->db_mtx
);
1645 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1646 while (db
->db_state
== DB_READ
||
1647 db
->db_state
== DB_FILL
) {
1648 ASSERT(db
->db_state
== DB_READ
||
1649 (flags
& DB_RF_HAVESTRUCT
) == 0);
1650 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1652 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1654 if (db
->db_state
== DB_UNCACHED
)
1655 err
= SET_ERROR(EIO
);
1657 mutex_exit(&db
->db_mtx
);
1664 dbuf_noread(dmu_buf_impl_t
*db
)
1666 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1667 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1668 mutex_enter(&db
->db_mtx
);
1669 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1670 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1671 if (db
->db_state
== DB_UNCACHED
) {
1672 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1673 spa_t
*spa
= db
->db_objset
->os_spa
;
1675 ASSERT(db
->db_buf
== NULL
);
1676 ASSERT(db
->db
.db_data
== NULL
);
1677 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1678 db
->db_state
= DB_FILL
;
1679 } else if (db
->db_state
== DB_NOFILL
) {
1680 dbuf_clear_data(db
);
1682 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1684 mutex_exit(&db
->db_mtx
);
1688 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1690 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1691 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1692 uint64_t txg
= dr
->dr_txg
;
1694 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1696 * This assert is valid because dmu_sync() expects to be called by
1697 * a zilog's get_data while holding a range lock. This call only
1698 * comes from dbuf_dirty() callers who must also hold a range lock.
1700 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1701 ASSERT(db
->db_level
== 0);
1703 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1704 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1707 ASSERT(db
->db_data_pending
!= dr
);
1709 /* free this block */
1710 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1711 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1713 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1714 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1715 dr
->dt
.dl
.dr_has_raw_params
= B_FALSE
;
1718 * Release the already-written buffer, so we leave it in
1719 * a consistent dirty state. Note that all callers are
1720 * modifying the buffer, so they will immediately do
1721 * another (redundant) arc_release(). Therefore, leave
1722 * the buf thawed to save the effort of freezing &
1723 * immediately re-thawing it.
1725 arc_release(dr
->dt
.dl
.dr_data
, db
);
1729 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1730 * data blocks in the free range, so that any future readers will find
1734 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1737 dmu_buf_impl_t
*db_search
;
1738 dmu_buf_impl_t
*db
, *db_next
;
1739 uint64_t txg
= tx
->tx_txg
;
1742 if (end_blkid
> dn
->dn_maxblkid
&&
1743 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1744 end_blkid
= dn
->dn_maxblkid
;
1745 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1747 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1748 db_search
->db_level
= 0;
1749 db_search
->db_blkid
= start_blkid
;
1750 db_search
->db_state
= DB_SEARCH
;
1752 mutex_enter(&dn
->dn_dbufs_mtx
);
1753 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1754 ASSERT3P(db
, ==, NULL
);
1756 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1758 for (; db
!= NULL
; db
= db_next
) {
1759 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1760 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1762 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1765 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1767 /* found a level 0 buffer in the range */
1768 mutex_enter(&db
->db_mtx
);
1769 if (dbuf_undirty(db
, tx
)) {
1770 /* mutex has been dropped and dbuf destroyed */
1774 if (db
->db_state
== DB_UNCACHED
||
1775 db
->db_state
== DB_NOFILL
||
1776 db
->db_state
== DB_EVICTING
) {
1777 ASSERT(db
->db
.db_data
== NULL
);
1778 mutex_exit(&db
->db_mtx
);
1781 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1782 /* will be handled in dbuf_read_done or dbuf_rele */
1783 db
->db_freed_in_flight
= TRUE
;
1784 mutex_exit(&db
->db_mtx
);
1787 if (zfs_refcount_count(&db
->db_holds
) == 0) {
1792 /* The dbuf is referenced */
1794 if (db
->db_last_dirty
!= NULL
) {
1795 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1797 if (dr
->dr_txg
== txg
) {
1799 * This buffer is "in-use", re-adjust the file
1800 * size to reflect that this buffer may
1801 * contain new data when we sync.
1803 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1804 db
->db_blkid
> dn
->dn_maxblkid
)
1805 dn
->dn_maxblkid
= db
->db_blkid
;
1806 dbuf_unoverride(dr
);
1809 * This dbuf is not dirty in the open context.
1810 * Either uncache it (if its not referenced in
1811 * the open context) or reset its contents to
1814 dbuf_fix_old_data(db
, txg
);
1817 /* clear the contents if its cached */
1818 if (db
->db_state
== DB_CACHED
) {
1819 ASSERT(db
->db
.db_data
!= NULL
);
1820 arc_release(db
->db_buf
, db
);
1821 rw_enter(&db
->db_rwlock
, RW_WRITER
);
1822 bzero(db
->db
.db_data
, db
->db
.db_size
);
1823 rw_exit(&db
->db_rwlock
);
1824 arc_buf_freeze(db
->db_buf
);
1827 mutex_exit(&db
->db_mtx
);
1830 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1831 mutex_exit(&dn
->dn_dbufs_mtx
);
1835 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1837 arc_buf_t
*buf
, *obuf
;
1838 int osize
= db
->db
.db_size
;
1839 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1842 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1848 * XXX we should be doing a dbuf_read, checking the return
1849 * value and returning that up to our callers
1851 dmu_buf_will_dirty(&db
->db
, tx
);
1853 /* create the data buffer for the new block */
1854 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1856 /* copy old block data to the new block */
1858 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1859 /* zero the remainder */
1861 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1863 mutex_enter(&db
->db_mtx
);
1864 dbuf_set_data(db
, buf
);
1865 arc_buf_destroy(obuf
, db
);
1866 db
->db
.db_size
= size
;
1868 if (db
->db_level
== 0) {
1869 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1871 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1872 ASSERT3U(db
->db_last_dirty
->dr_accounted
, ==, osize
);
1873 db
->db_last_dirty
->dr_accounted
= size
;
1874 mutex_exit(&db
->db_mtx
);
1876 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1881 dbuf_release_bp(dmu_buf_impl_t
*db
)
1883 objset_t
*os __maybe_unused
= db
->db_objset
;
1885 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1886 ASSERT(arc_released(os
->os_phys_buf
) ||
1887 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1888 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1890 (void) arc_release(db
->db_buf
, db
);
1894 * We already have a dirty record for this TXG, and we are being
1898 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1900 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1902 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1904 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1906 * If this buffer has already been written out,
1907 * we now need to reset its state.
1909 dbuf_unoverride(dr
);
1910 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1911 db
->db_state
!= DB_NOFILL
) {
1912 /* Already released on initial dirty, so just thaw. */
1913 ASSERT(arc_released(db
->db_buf
));
1914 arc_buf_thaw(db
->db_buf
);
1919 dbuf_dirty_record_t
*
1920 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1924 dbuf_dirty_record_t
**drp
, *dr
;
1925 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1926 boolean_t drop_struct_rwlock
= B_FALSE
;
1928 ASSERT(tx
->tx_txg
!= 0);
1929 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1930 DMU_TX_DIRTY_BUF(tx
, db
);
1935 * Shouldn't dirty a regular buffer in syncing context. Private
1936 * objects may be dirtied in syncing context, but only if they
1937 * were already pre-dirtied in open context.
1940 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1941 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1944 ASSERT(!dmu_tx_is_syncing(tx
) ||
1945 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1946 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1947 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1948 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1949 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1952 * We make this assert for private objects as well, but after we
1953 * check if we're already dirty. They are allowed to re-dirty
1954 * in syncing context.
1956 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1957 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1958 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1960 mutex_enter(&db
->db_mtx
);
1962 * XXX make this true for indirects too? The problem is that
1963 * transactions created with dmu_tx_create_assigned() from
1964 * syncing context don't bother holding ahead.
1966 ASSERT(db
->db_level
!= 0 ||
1967 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1968 db
->db_state
== DB_NOFILL
);
1970 mutex_enter(&dn
->dn_mtx
);
1972 * Don't set dirtyctx to SYNC if we're just modifying this as we
1973 * initialize the objset.
1975 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1976 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1977 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1980 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1981 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1982 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1983 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1984 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1986 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1987 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1992 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
1993 dn
->dn_dirty_txg
= tx
->tx_txg
;
1994 mutex_exit(&dn
->dn_mtx
);
1996 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1997 dn
->dn_have_spill
= B_TRUE
;
2000 * If this buffer is already dirty, we're done.
2002 drp
= &db
->db_last_dirty
;
2003 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
2004 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
2005 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
2007 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
2011 mutex_exit(&db
->db_mtx
);
2016 * Only valid if not already dirty.
2018 ASSERT(dn
->dn_object
== 0 ||
2019 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2020 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2022 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
2025 * We should only be dirtying in syncing context if it's the
2026 * mos or we're initializing the os or it's a special object.
2027 * However, we are allowed to dirty in syncing context provided
2028 * we already dirtied it in open context. Hence we must make
2029 * this assertion only if we're not already dirty.
2032 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
2034 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2035 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
2036 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2037 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
2038 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2039 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2041 ASSERT(db
->db
.db_size
!= 0);
2043 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2045 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2046 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
2050 * If this buffer is dirty in an old transaction group we need
2051 * to make a copy of it so that the changes we make in this
2052 * transaction group won't leak out when we sync the older txg.
2054 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
2055 list_link_init(&dr
->dr_dirty_node
);
2056 if (db
->db_level
== 0) {
2057 void *data_old
= db
->db_buf
;
2059 if (db
->db_state
!= DB_NOFILL
) {
2060 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2061 dbuf_fix_old_data(db
, tx
->tx_txg
);
2062 data_old
= db
->db
.db_data
;
2063 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
2065 * Release the data buffer from the cache so
2066 * that we can modify it without impacting
2067 * possible other users of this cached data
2068 * block. Note that indirect blocks and
2069 * private objects are not released until the
2070 * syncing state (since they are only modified
2073 arc_release(db
->db_buf
, db
);
2074 dbuf_fix_old_data(db
, tx
->tx_txg
);
2075 data_old
= db
->db_buf
;
2077 ASSERT(data_old
!= NULL
);
2079 dr
->dt
.dl
.dr_data
= data_old
;
2081 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
2082 list_create(&dr
->dt
.di
.dr_children
,
2083 sizeof (dbuf_dirty_record_t
),
2084 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
2086 if (db
->db_blkid
!= DMU_BONUS_BLKID
)
2087 dr
->dr_accounted
= db
->db
.db_size
;
2089 dr
->dr_txg
= tx
->tx_txg
;
2094 * We could have been freed_in_flight between the dbuf_noread
2095 * and dbuf_dirty. We win, as though the dbuf_noread() had
2096 * happened after the free.
2098 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2099 db
->db_blkid
!= DMU_SPILL_BLKID
) {
2100 mutex_enter(&dn
->dn_mtx
);
2101 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2102 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
2105 mutex_exit(&dn
->dn_mtx
);
2106 db
->db_freed_in_flight
= FALSE
;
2110 * This buffer is now part of this txg
2112 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
2113 db
->db_dirtycnt
+= 1;
2114 ASSERT3U(db
->db_dirtycnt
, <=, 3);
2116 mutex_exit(&db
->db_mtx
);
2118 if (db
->db_blkid
== DMU_BONUS_BLKID
||
2119 db
->db_blkid
== DMU_SPILL_BLKID
) {
2120 mutex_enter(&dn
->dn_mtx
);
2121 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2122 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2123 mutex_exit(&dn
->dn_mtx
);
2124 dnode_setdirty(dn
, tx
);
2129 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
2130 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2131 drop_struct_rwlock
= B_TRUE
;
2135 * If we are overwriting a dedup BP, then unless it is snapshotted,
2136 * when we get to syncing context we will need to decrement its
2137 * refcount in the DDT. Prefetch the relevant DDT block so that
2138 * syncing context won't have to wait for the i/o.
2140 if (db
->db_blkptr
!= NULL
) {
2141 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
2142 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
2143 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
2147 * We need to hold the dn_struct_rwlock to make this assertion,
2148 * because it protects dn_phys / dn_next_nlevels from changing.
2150 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
2151 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
2152 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
2153 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
2154 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
2157 if (db
->db_level
== 0) {
2158 ASSERT(!db
->db_objset
->os_raw_receive
||
2159 dn
->dn_maxblkid
>= db
->db_blkid
);
2160 dnode_new_blkid(dn
, db
->db_blkid
, tx
,
2161 drop_struct_rwlock
, B_FALSE
);
2162 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
2165 if (db
->db_level
+1 < dn
->dn_nlevels
) {
2166 dmu_buf_impl_t
*parent
= db
->db_parent
;
2167 dbuf_dirty_record_t
*di
;
2168 int parent_held
= FALSE
;
2170 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
2171 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2172 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2173 db
->db_blkid
>> epbs
, FTAG
);
2174 ASSERT(parent
!= NULL
);
2177 if (drop_struct_rwlock
)
2178 rw_exit(&dn
->dn_struct_rwlock
);
2179 ASSERT3U(db
->db_level
+ 1, ==, parent
->db_level
);
2180 di
= dbuf_dirty(parent
, tx
);
2182 dbuf_rele(parent
, FTAG
);
2184 mutex_enter(&db
->db_mtx
);
2186 * Since we've dropped the mutex, it's possible that
2187 * dbuf_undirty() might have changed this out from under us.
2189 if (db
->db_last_dirty
== dr
||
2190 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2191 mutex_enter(&di
->dt
.di
.dr_mtx
);
2192 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2193 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2194 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2195 mutex_exit(&di
->dt
.di
.dr_mtx
);
2198 mutex_exit(&db
->db_mtx
);
2200 ASSERT(db
->db_level
+ 1 == dn
->dn_nlevels
);
2201 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2202 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2203 mutex_enter(&dn
->dn_mtx
);
2204 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2205 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2206 mutex_exit(&dn
->dn_mtx
);
2207 if (drop_struct_rwlock
)
2208 rw_exit(&dn
->dn_struct_rwlock
);
2211 dnode_setdirty(dn
, tx
);
2217 * Undirty a buffer in the transaction group referenced by the given
2218 * transaction. Return whether this evicted the dbuf.
2221 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2224 uint64_t txg
= tx
->tx_txg
;
2225 dbuf_dirty_record_t
*dr
, **drp
;
2230 * Due to our use of dn_nlevels below, this can only be called
2231 * in open context, unless we are operating on the MOS.
2232 * From syncing context, dn_nlevels may be different from the
2233 * dn_nlevels used when dbuf was dirtied.
2235 ASSERT(db
->db_objset
==
2236 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2237 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2238 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2239 ASSERT0(db
->db_level
);
2240 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2243 * If this buffer is not dirty, we're done.
2245 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
2246 if (dr
->dr_txg
<= txg
)
2248 if (dr
== NULL
|| dr
->dr_txg
< txg
)
2250 ASSERT(dr
->dr_txg
== txg
);
2251 ASSERT(dr
->dr_dbuf
== db
);
2256 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2258 ASSERT(db
->db
.db_size
!= 0);
2260 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2261 dr
->dr_accounted
, txg
);
2266 * Note that there are three places in dbuf_dirty()
2267 * where this dirty record may be put on a list.
2268 * Make sure to do a list_remove corresponding to
2269 * every one of those list_insert calls.
2271 if (dr
->dr_parent
) {
2272 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2273 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2274 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2275 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2276 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2277 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2278 mutex_enter(&dn
->dn_mtx
);
2279 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2280 mutex_exit(&dn
->dn_mtx
);
2284 if (db
->db_state
!= DB_NOFILL
) {
2285 dbuf_unoverride(dr
);
2287 ASSERT(db
->db_buf
!= NULL
);
2288 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2289 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2290 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2293 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2295 ASSERT(db
->db_dirtycnt
> 0);
2296 db
->db_dirtycnt
-= 1;
2298 if (zfs_refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2299 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
2308 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2310 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2312 ASSERT(tx
->tx_txg
!= 0);
2313 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2316 * Quick check for dirtiness. For already dirty blocks, this
2317 * reduces runtime of this function by >90%, and overall performance
2318 * by 50% for some workloads (e.g. file deletion with indirect blocks
2321 mutex_enter(&db
->db_mtx
);
2323 dbuf_dirty_record_t
*dr
;
2324 for (dr
= db
->db_last_dirty
;
2325 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
2327 * It's possible that it is already dirty but not cached,
2328 * because there are some calls to dbuf_dirty() that don't
2329 * go through dmu_buf_will_dirty().
2331 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
2332 /* This dbuf is already dirty and cached. */
2334 mutex_exit(&db
->db_mtx
);
2338 mutex_exit(&db
->db_mtx
);
2341 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2342 flags
|= DB_RF_HAVESTRUCT
;
2344 (void) dbuf_read(db
, NULL
, flags
);
2345 (void) dbuf_dirty(db
, tx
);
2349 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2351 dmu_buf_will_dirty_impl(db_fake
,
2352 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2356 dmu_buf_is_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2358 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2360 mutex_enter(&db
->db_mtx
);
2361 for (dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2362 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
2363 if (dr
->dr_txg
== tx
->tx_txg
) {
2364 mutex_exit(&db
->db_mtx
);
2368 mutex_exit(&db
->db_mtx
);
2373 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2375 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2377 db
->db_state
= DB_NOFILL
;
2379 dmu_buf_will_fill(db_fake
, tx
);
2383 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2385 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2387 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2388 ASSERT(tx
->tx_txg
!= 0);
2389 ASSERT(db
->db_level
== 0);
2390 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2392 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2393 dmu_tx_private_ok(tx
));
2396 (void) dbuf_dirty(db
, tx
);
2400 * This function is effectively the same as dmu_buf_will_dirty(), but
2401 * indicates the caller expects raw encrypted data in the db, and provides
2402 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2403 * blkptr_t when this dbuf is written. This is only used for blocks of
2404 * dnodes, during raw receive.
2407 dmu_buf_set_crypt_params(dmu_buf_t
*db_fake
, boolean_t byteorder
,
2408 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
, dmu_tx_t
*tx
)
2410 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2411 dbuf_dirty_record_t
*dr
;
2414 * dr_has_raw_params is only processed for blocks of dnodes
2415 * (see dbuf_sync_dnode_leaf_crypt()).
2417 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
2418 ASSERT3U(db
->db_level
, ==, 0);
2419 ASSERT(db
->db_objset
->os_raw_receive
);
2421 dmu_buf_will_dirty_impl(db_fake
,
2422 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2424 dr
= db
->db_last_dirty
;
2425 while (dr
!= NULL
&& dr
->dr_txg
> tx
->tx_txg
)
2428 ASSERT3P(dr
, !=, NULL
);
2429 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2431 dr
->dt
.dl
.dr_has_raw_params
= B_TRUE
;
2432 dr
->dt
.dl
.dr_byteorder
= byteorder
;
2433 bcopy(salt
, dr
->dt
.dl
.dr_salt
, ZIO_DATA_SALT_LEN
);
2434 bcopy(iv
, dr
->dt
.dl
.dr_iv
, ZIO_DATA_IV_LEN
);
2435 bcopy(mac
, dr
->dt
.dl
.dr_mac
, ZIO_DATA_MAC_LEN
);
2439 dbuf_override_impl(dmu_buf_impl_t
*db
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
2441 struct dirty_leaf
*dl
;
2443 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2444 dl
= &db
->db_last_dirty
->dt
.dl
;
2445 dl
->dr_overridden_by
= *bp
;
2446 dl
->dr_override_state
= DR_OVERRIDDEN
;
2447 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2452 dmu_buf_fill_done(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2454 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2455 mutex_enter(&db
->db_mtx
);
2458 if (db
->db_state
== DB_FILL
) {
2459 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2460 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2461 /* we were freed while filling */
2462 /* XXX dbuf_undirty? */
2463 bzero(db
->db
.db_data
, db
->db
.db_size
);
2464 db
->db_freed_in_flight
= FALSE
;
2466 db
->db_state
= DB_CACHED
;
2467 cv_broadcast(&db
->db_changed
);
2469 mutex_exit(&db
->db_mtx
);
2473 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2474 bp_embedded_type_t etype
, enum zio_compress comp
,
2475 int uncompressed_size
, int compressed_size
, int byteorder
,
2478 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2479 struct dirty_leaf
*dl
;
2480 dmu_object_type_t type
;
2482 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2483 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2484 SPA_FEATURE_EMBEDDED_DATA
));
2488 type
= DB_DNODE(db
)->dn_type
;
2491 ASSERT0(db
->db_level
);
2492 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2494 dmu_buf_will_not_fill(dbuf
, tx
);
2496 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2497 dl
= &db
->db_last_dirty
->dt
.dl
;
2498 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2499 data
, comp
, uncompressed_size
, compressed_size
);
2500 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2501 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2502 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2503 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2505 dl
->dr_override_state
= DR_OVERRIDDEN
;
2506 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2510 dmu_buf_redact(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2512 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2513 dmu_object_type_t type
;
2514 ASSERT(dsl_dataset_feature_is_active(db
->db_objset
->os_dsl_dataset
,
2515 SPA_FEATURE_REDACTED_DATASETS
));
2518 type
= DB_DNODE(db
)->dn_type
;
2521 ASSERT0(db
->db_level
);
2522 dmu_buf_will_not_fill(dbuf
, tx
);
2524 blkptr_t bp
= { { { {0} } } };
2525 BP_SET_TYPE(&bp
, type
);
2526 BP_SET_LEVEL(&bp
, 0);
2527 BP_SET_BIRTH(&bp
, tx
->tx_txg
, 0);
2528 BP_SET_REDACTED(&bp
);
2529 BPE_SET_LSIZE(&bp
, dbuf
->db_size
);
2531 dbuf_override_impl(db
, &bp
, tx
);
2535 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2536 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2539 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2541 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2542 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2543 ASSERT(db
->db_level
== 0);
2544 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2545 ASSERT(buf
!= NULL
);
2546 ASSERT3U(arc_buf_lsize(buf
), ==, db
->db
.db_size
);
2547 ASSERT(tx
->tx_txg
!= 0);
2549 arc_return_buf(buf
, db
);
2550 ASSERT(arc_released(buf
));
2552 mutex_enter(&db
->db_mtx
);
2554 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2555 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2557 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2559 if (db
->db_state
== DB_CACHED
&&
2560 zfs_refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2562 * In practice, we will never have a case where we have an
2563 * encrypted arc buffer while additional holds exist on the
2564 * dbuf. We don't handle this here so we simply assert that
2567 ASSERT(!arc_is_encrypted(buf
));
2568 mutex_exit(&db
->db_mtx
);
2569 (void) dbuf_dirty(db
, tx
);
2570 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2571 arc_buf_destroy(buf
, db
);
2572 xuio_stat_wbuf_copied();
2576 xuio_stat_wbuf_nocopy();
2577 if (db
->db_state
== DB_CACHED
) {
2578 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2580 ASSERT(db
->db_buf
!= NULL
);
2581 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2582 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2584 if (!arc_released(db
->db_buf
)) {
2585 ASSERT(dr
->dt
.dl
.dr_override_state
==
2587 arc_release(db
->db_buf
, db
);
2589 dr
->dt
.dl
.dr_data
= buf
;
2590 arc_buf_destroy(db
->db_buf
, db
);
2591 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2592 arc_release(db
->db_buf
, db
);
2593 arc_buf_destroy(db
->db_buf
, db
);
2597 ASSERT(db
->db_buf
== NULL
);
2598 dbuf_set_data(db
, buf
);
2599 db
->db_state
= DB_FILL
;
2600 mutex_exit(&db
->db_mtx
);
2601 (void) dbuf_dirty(db
, tx
);
2602 dmu_buf_fill_done(&db
->db
, tx
);
2606 dbuf_destroy(dmu_buf_impl_t
*db
)
2609 dmu_buf_impl_t
*parent
= db
->db_parent
;
2610 dmu_buf_impl_t
*dndb
;
2612 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2613 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
2615 if (db
->db_buf
!= NULL
) {
2616 arc_buf_destroy(db
->db_buf
, db
);
2620 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2621 int slots
= DB_DNODE(db
)->dn_num_slots
;
2622 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2623 if (db
->db
.db_data
!= NULL
) {
2624 kmem_free(db
->db
.db_data
, bonuslen
);
2625 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2626 db
->db_state
= DB_UNCACHED
;
2630 dbuf_clear_data(db
);
2632 if (multilist_link_active(&db
->db_cache_link
)) {
2633 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
2634 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
2636 multilist_remove(dbuf_caches
[db
->db_caching_status
].cache
, db
);
2637 (void) zfs_refcount_remove_many(
2638 &dbuf_caches
[db
->db_caching_status
].size
,
2639 db
->db
.db_size
, db
);
2641 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
2642 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
2644 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
2645 DBUF_STAT_BUMPDOWN(cache_count
);
2646 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
2649 db
->db_caching_status
= DB_NO_CACHE
;
2652 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2653 ASSERT(db
->db_data_pending
== NULL
);
2655 db
->db_state
= DB_EVICTING
;
2656 db
->db_blkptr
= NULL
;
2659 * Now that db_state is DB_EVICTING, nobody else can find this via
2660 * the hash table. We can now drop db_mtx, which allows us to
2661 * acquire the dn_dbufs_mtx.
2663 mutex_exit(&db
->db_mtx
);
2668 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2669 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2671 mutex_enter_nested(&dn
->dn_dbufs_mtx
,
2673 avl_remove(&dn
->dn_dbufs
, db
);
2674 atomic_dec_32(&dn
->dn_dbufs_count
);
2678 mutex_exit(&dn
->dn_dbufs_mtx
);
2680 * Decrementing the dbuf count means that the hold corresponding
2681 * to the removed dbuf is no longer discounted in dnode_move(),
2682 * so the dnode cannot be moved until after we release the hold.
2683 * The membar_producer() ensures visibility of the decremented
2684 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2687 mutex_enter(&dn
->dn_mtx
);
2688 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
2689 db
->db_dnode_handle
= NULL
;
2691 dbuf_hash_remove(db
);
2696 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
2698 db
->db_parent
= NULL
;
2700 ASSERT(db
->db_buf
== NULL
);
2701 ASSERT(db
->db
.db_data
== NULL
);
2702 ASSERT(db
->db_hash_next
== NULL
);
2703 ASSERT(db
->db_blkptr
== NULL
);
2704 ASSERT(db
->db_data_pending
== NULL
);
2705 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
2706 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2708 kmem_cache_free(dbuf_kmem_cache
, db
);
2709 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2712 * If this dbuf is referenced from an indirect dbuf,
2713 * decrement the ref count on the indirect dbuf.
2715 if (parent
&& parent
!= dndb
) {
2716 mutex_enter(&parent
->db_mtx
);
2717 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
2722 * Note: While bpp will always be updated if the function returns success,
2723 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2724 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2727 __attribute__((always_inline
))
2729 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2730 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
2735 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2737 if (blkid
== DMU_SPILL_BLKID
) {
2738 mutex_enter(&dn
->dn_mtx
);
2739 if (dn
->dn_have_spill
&&
2740 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2741 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2744 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2745 *parentp
= dn
->dn_dbuf
;
2746 mutex_exit(&dn
->dn_mtx
);
2751 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2752 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2754 ASSERT3U(level
* epbs
, <, 64);
2755 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2757 * This assertion shouldn't trip as long as the max indirect block size
2758 * is less than 1M. The reason for this is that up to that point,
2759 * the number of levels required to address an entire object with blocks
2760 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2761 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2762 * (i.e. we can address the entire object), objects will all use at most
2763 * N-1 levels and the assertion won't overflow. However, once epbs is
2764 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2765 * enough to address an entire object, so objects will have 5 levels,
2766 * but then this assertion will overflow.
2768 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2769 * need to redo this logic to handle overflows.
2771 ASSERT(level
>= nlevels
||
2772 ((nlevels
- level
- 1) * epbs
) +
2773 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2774 if (level
>= nlevels
||
2775 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2776 ((nlevels
- level
- 1) * epbs
)) ||
2778 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2779 /* the buffer has no parent yet */
2780 return (SET_ERROR(ENOENT
));
2781 } else if (level
< nlevels
-1) {
2782 /* this block is referenced from an indirect block */
2785 err
= dbuf_hold_impl(dn
, level
+ 1,
2786 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2790 err
= dbuf_read(*parentp
, NULL
,
2791 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2793 dbuf_rele(*parentp
, NULL
);
2797 rw_enter(&(*parentp
)->db_rwlock
, RW_READER
);
2798 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2799 (blkid
& ((1ULL << epbs
) - 1));
2800 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2801 ASSERT(BP_IS_HOLE(*bpp
));
2802 rw_exit(&(*parentp
)->db_rwlock
);
2805 /* the block is referenced from the dnode */
2806 ASSERT3U(level
, ==, nlevels
-1);
2807 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2808 blkid
< dn
->dn_phys
->dn_nblkptr
);
2810 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2811 *parentp
= dn
->dn_dbuf
;
2813 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2818 static dmu_buf_impl_t
*
2819 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2820 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2822 objset_t
*os
= dn
->dn_objset
;
2823 dmu_buf_impl_t
*db
, *odb
;
2825 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2826 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2828 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2831 db
->db
.db_object
= dn
->dn_object
;
2832 db
->db_level
= level
;
2833 db
->db_blkid
= blkid
;
2834 db
->db_last_dirty
= NULL
;
2835 db
->db_dirtycnt
= 0;
2836 db
->db_dnode_handle
= dn
->dn_handle
;
2837 db
->db_parent
= parent
;
2838 db
->db_blkptr
= blkptr
;
2841 db
->db_user_immediate_evict
= FALSE
;
2842 db
->db_freed_in_flight
= FALSE
;
2843 db
->db_pending_evict
= FALSE
;
2845 if (blkid
== DMU_BONUS_BLKID
) {
2846 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2847 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2848 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2849 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2850 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2851 db
->db_state
= DB_UNCACHED
;
2852 db
->db_caching_status
= DB_NO_CACHE
;
2853 /* the bonus dbuf is not placed in the hash table */
2854 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2856 } else if (blkid
== DMU_SPILL_BLKID
) {
2857 db
->db
.db_size
= (blkptr
!= NULL
) ?
2858 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2859 db
->db
.db_offset
= 0;
2862 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2863 db
->db
.db_size
= blocksize
;
2864 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2868 * Hold the dn_dbufs_mtx while we get the new dbuf
2869 * in the hash table *and* added to the dbufs list.
2870 * This prevents a possible deadlock with someone
2871 * trying to look up this dbuf before it's added to the
2874 mutex_enter(&dn
->dn_dbufs_mtx
);
2875 db
->db_state
= DB_EVICTING
;
2876 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2877 /* someone else inserted it first */
2878 kmem_cache_free(dbuf_kmem_cache
, db
);
2879 mutex_exit(&dn
->dn_dbufs_mtx
);
2880 DBUF_STAT_BUMP(hash_insert_race
);
2883 avl_add(&dn
->dn_dbufs
, db
);
2885 db
->db_state
= DB_UNCACHED
;
2886 db
->db_caching_status
= DB_NO_CACHE
;
2887 mutex_exit(&dn
->dn_dbufs_mtx
);
2888 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2890 if (parent
&& parent
!= dn
->dn_dbuf
)
2891 dbuf_add_ref(parent
, db
);
2893 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2894 zfs_refcount_count(&dn
->dn_holds
) > 0);
2895 (void) zfs_refcount_add(&dn
->dn_holds
, db
);
2896 atomic_inc_32(&dn
->dn_dbufs_count
);
2898 dprintf_dbuf(db
, "db=%p\n", db
);
2904 * This function returns a block pointer and information about the object,
2905 * given a dnode and a block. This is a publicly accessible version of
2906 * dbuf_findbp that only returns some information, rather than the
2907 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
2908 * should be locked as (at least) a reader.
2911 dbuf_dnode_findbp(dnode_t
*dn
, uint64_t level
, uint64_t blkid
,
2912 blkptr_t
*bp
, uint16_t *datablkszsec
, uint8_t *indblkshift
)
2914 dmu_buf_impl_t
*dbp
= NULL
;
2917 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2919 err
= dbuf_findbp(dn
, level
, blkid
, B_FALSE
, &dbp
, &bp2
);
2923 dbuf_rele(dbp
, NULL
);
2924 if (datablkszsec
!= NULL
)
2925 *datablkszsec
= dn
->dn_phys
->dn_datablkszsec
;
2926 if (indblkshift
!= NULL
)
2927 *indblkshift
= dn
->dn_phys
->dn_indblkshift
;
2933 typedef struct dbuf_prefetch_arg
{
2934 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2935 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2936 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2937 int dpa_curlevel
; /* The current level that we're reading */
2938 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2939 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2940 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2941 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2942 } dbuf_prefetch_arg_t
;
2945 * Actually issue the prefetch read for the block given.
2948 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2950 ASSERT(!BP_IS_REDACTED(bp
) ||
2951 dsl_dataset_feature_is_active(
2952 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
2953 SPA_FEATURE_REDACTED_DATASETS
));
2955 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
) || BP_IS_REDACTED(bp
))
2958 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2959 arc_flags_t aflags
=
2960 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2962 /* dnodes are always read as raw and then converted later */
2963 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
2964 dpa
->dpa_curlevel
== 0)
2965 zio_flags
|= ZIO_FLAG_RAW
;
2967 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2968 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2969 ASSERT(dpa
->dpa_zio
!= NULL
);
2970 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2971 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
2975 * Called when an indirect block above our prefetch target is read in. This
2976 * will either read in the next indirect block down the tree or issue the actual
2977 * prefetch if the next block down is our target.
2980 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
2981 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
2983 dbuf_prefetch_arg_t
*dpa
= private;
2985 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2986 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2989 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
2990 kmem_free(dpa
, sizeof (*dpa
));
2993 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
2996 * The dpa_dnode is only valid if we are called with a NULL
2997 * zio. This indicates that the arc_read() returned without
2998 * first calling zio_read() to issue a physical read. Once
2999 * a physical read is made the dpa_dnode must be invalidated
3000 * as the locks guarding it may have been dropped. If the
3001 * dpa_dnode is still valid, then we want to add it to the dbuf
3002 * cache. To do so, we must hold the dbuf associated with the block
3003 * we just prefetched, read its contents so that we associate it
3004 * with an arc_buf_t, and then release it.
3007 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
3008 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
3009 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
3011 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
3013 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
3015 dpa
->dpa_dnode
= NULL
;
3016 } else if (dpa
->dpa_dnode
!= NULL
) {
3017 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
3018 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
3019 dpa
->dpa_zb
.zb_level
));
3020 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
3021 dpa
->dpa_curlevel
, curblkid
, FTAG
);
3023 kmem_free(dpa
, sizeof (*dpa
));
3024 arc_buf_destroy(abuf
, private);
3028 (void) dbuf_read(db
, NULL
,
3029 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
3030 dbuf_rele(db
, FTAG
);
3033 dpa
->dpa_curlevel
--;
3034 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
3035 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
3036 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
3037 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
3039 ASSERT(!BP_IS_REDACTED(bp
) ||
3040 dsl_dataset_feature_is_active(
3041 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3042 SPA_FEATURE_REDACTED_DATASETS
));
3043 if (BP_IS_HOLE(bp
) || BP_IS_REDACTED(bp
)) {
3044 kmem_free(dpa
, sizeof (*dpa
));
3045 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
3046 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
3047 dbuf_issue_final_prefetch(dpa
, bp
);
3048 kmem_free(dpa
, sizeof (*dpa
));
3050 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3051 zbookmark_phys_t zb
;
3053 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3054 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
3055 iter_aflags
|= ARC_FLAG_L2CACHE
;
3057 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3059 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
3060 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
3062 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3063 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
3064 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3068 arc_buf_destroy(abuf
, private);
3072 * Issue prefetch reads for the given block on the given level. If the indirect
3073 * blocks above that block are not in memory, we will read them in
3074 * asynchronously. As a result, this call never blocks waiting for a read to
3075 * complete. Note that the prefetch might fail if the dataset is encrypted and
3076 * the encryption key is unmapped before the IO completes.
3079 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
3083 int epbs
, nlevels
, curlevel
;
3086 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3087 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3089 if (blkid
> dn
->dn_maxblkid
)
3092 if (level
== 0 && dnode_block_freed(dn
, blkid
))
3096 * This dnode hasn't been written to disk yet, so there's nothing to
3099 nlevels
= dn
->dn_phys
->dn_nlevels
;
3100 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
3103 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3104 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
3107 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
3110 mutex_exit(&db
->db_mtx
);
3112 * This dbuf already exists. It is either CACHED, or
3113 * (we assume) about to be read or filled.
3119 * Find the closest ancestor (indirect block) of the target block
3120 * that is present in the cache. In this indirect block, we will
3121 * find the bp that is at curlevel, curblkid.
3125 while (curlevel
< nlevels
- 1) {
3126 int parent_level
= curlevel
+ 1;
3127 uint64_t parent_blkid
= curblkid
>> epbs
;
3130 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
3131 FALSE
, TRUE
, FTAG
, &db
) == 0) {
3132 blkptr_t
*bpp
= db
->db_buf
->b_data
;
3133 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
3134 dbuf_rele(db
, FTAG
);
3138 curlevel
= parent_level
;
3139 curblkid
= parent_blkid
;
3142 if (curlevel
== nlevels
- 1) {
3143 /* No cached indirect blocks found. */
3144 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
3145 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
3147 ASSERT(!BP_IS_REDACTED(&bp
) ||
3148 dsl_dataset_feature_is_active(dn
->dn_objset
->os_dsl_dataset
,
3149 SPA_FEATURE_REDACTED_DATASETS
));
3150 if (BP_IS_HOLE(&bp
) || BP_IS_REDACTED(&bp
))
3153 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
3155 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
3158 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
3159 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
3160 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3161 dn
->dn_object
, level
, blkid
);
3162 dpa
->dpa_curlevel
= curlevel
;
3163 dpa
->dpa_prio
= prio
;
3164 dpa
->dpa_aflags
= aflags
;
3165 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
3166 dpa
->dpa_dnode
= dn
;
3167 dpa
->dpa_epbs
= epbs
;
3170 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3171 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
3172 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
3175 * If we have the indirect just above us, no need to do the asynchronous
3176 * prefetch chain; we'll just run the last step ourselves. If we're at
3177 * a higher level, though, we want to issue the prefetches for all the
3178 * indirect blocks asynchronously, so we can go on with whatever we were
3181 if (curlevel
== level
) {
3182 ASSERT3U(curblkid
, ==, blkid
);
3183 dbuf_issue_final_prefetch(dpa
, &bp
);
3184 kmem_free(dpa
, sizeof (*dpa
));
3186 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3187 zbookmark_phys_t zb
;
3189 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3190 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
3191 iter_aflags
|= ARC_FLAG_L2CACHE
;
3193 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3194 dn
->dn_object
, curlevel
, curblkid
);
3195 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3196 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
3197 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3201 * We use pio here instead of dpa_zio since it's possible that
3202 * dpa may have already been freed.
3208 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3209 * the case of encrypted, compressed and uncompressed buffers by
3210 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3211 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3213 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3215 noinline
static void
3216 dbuf_hold_copy(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3218 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3219 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
3220 enum zio_compress compress_type
= arc_get_compression(data
);
3222 if (arc_is_encrypted(data
)) {
3223 boolean_t byteorder
;
3224 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3225 uint8_t iv
[ZIO_DATA_IV_LEN
];
3226 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3228 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
3229 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
3230 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
3231 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
3233 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3234 dbuf_set_data(db
, arc_alloc_compressed_buf(
3235 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
3236 arc_buf_lsize(data
), compress_type
));
3238 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
3239 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
3242 rw_enter(&db
->db_rwlock
, RW_WRITER
);
3243 bcopy(data
->b_data
, db
->db
.db_data
, arc_buf_size(data
));
3244 rw_exit(&db
->db_rwlock
);
3248 * Returns with db_holds incremented, and db_mtx not held.
3249 * Note: dn_struct_rwlock must be held.
3252 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3253 boolean_t fail_sparse
, boolean_t fail_uncached
,
3254 void *tag
, dmu_buf_impl_t
**dbp
)
3256 dmu_buf_impl_t
*db
, *parent
= NULL
;
3258 /* If the pool has been created, verify the tx_sync_lock is not held */
3259 spa_t
*spa
= dn
->dn_objset
->os_spa
;
3260 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
3262 ASSERT(!MUTEX_HELD(&dp
->dp_tx
.tx_sync_lock
));
3265 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3266 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3267 ASSERT3U(dn
->dn_nlevels
, >, level
);
3271 /* dbuf_find() returns with db_mtx held */
3272 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
);
3275 blkptr_t
*bp
= NULL
;
3279 return (SET_ERROR(ENOENT
));
3281 ASSERT3P(parent
, ==, NULL
);
3282 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
3284 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
3285 err
= SET_ERROR(ENOENT
);
3288 dbuf_rele(parent
, NULL
);
3292 if (err
&& err
!= ENOENT
)
3294 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
);
3297 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
3298 mutex_exit(&db
->db_mtx
);
3299 return (SET_ERROR(ENOENT
));
3302 if (db
->db_buf
!= NULL
) {
3303 arc_buf_access(db
->db_buf
);
3304 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
3307 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
3310 * If this buffer is currently syncing out, and we are
3311 * still referencing it from db_data, we need to make a copy
3312 * of it in case we decide we want to dirty it again in this txg.
3314 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
3315 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3316 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
3317 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3318 if (dr
->dt
.dl
.dr_data
== db
->db_buf
)
3319 dbuf_hold_copy(dn
, db
);
3322 if (multilist_link_active(&db
->db_cache_link
)) {
3323 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3324 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3325 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3327 multilist_remove(dbuf_caches
[db
->db_caching_status
].cache
, db
);
3328 (void) zfs_refcount_remove_many(
3329 &dbuf_caches
[db
->db_caching_status
].size
,
3330 db
->db
.db_size
, db
);
3332 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3333 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3335 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3336 DBUF_STAT_BUMPDOWN(cache_count
);
3337 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3340 db
->db_caching_status
= DB_NO_CACHE
;
3342 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3344 mutex_exit(&db
->db_mtx
);
3346 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3348 dbuf_rele(parent
, NULL
);
3350 ASSERT3P(DB_DNODE(db
), ==, dn
);
3351 ASSERT3U(db
->db_blkid
, ==, blkid
);
3352 ASSERT3U(db
->db_level
, ==, level
);
3359 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
3361 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3365 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
3368 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3369 return (err
? NULL
: db
);
3373 dbuf_create_bonus(dnode_t
*dn
)
3375 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3377 ASSERT(dn
->dn_bonus
== NULL
);
3378 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
3382 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3384 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3386 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3387 return (SET_ERROR(ENOTSUP
));
3389 blksz
= SPA_MINBLOCKSIZE
;
3390 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3391 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3393 dbuf_new_size(db
, blksz
, tx
);
3399 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3401 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3404 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3406 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
3408 int64_t holds
= zfs_refcount_add(&db
->db_holds
, tag
);
3409 VERIFY3S(holds
, >, 1);
3412 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3414 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3417 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3418 dmu_buf_impl_t
*found_db
;
3419 boolean_t result
= B_FALSE
;
3421 if (blkid
== DMU_BONUS_BLKID
)
3422 found_db
= dbuf_find_bonus(os
, obj
);
3424 found_db
= dbuf_find(os
, obj
, 0, blkid
);
3426 if (found_db
!= NULL
) {
3427 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3428 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3431 mutex_exit(&found_db
->db_mtx
);
3437 * If you call dbuf_rele() you had better not be referencing the dnode handle
3438 * unless you have some other direct or indirect hold on the dnode. (An indirect
3439 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3440 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3441 * dnode's parent dbuf evicting its dnode handles.
3444 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
3446 mutex_enter(&db
->db_mtx
);
3447 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
3451 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
3453 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3457 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3458 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3459 * argument should be set if we are already in the dbuf-evicting code
3460 * path, in which case we don't want to recursively evict. This allows us to
3461 * avoid deeply nested stacks that would have a call flow similar to this:
3463 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3466 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3470 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
, boolean_t evicting
)
3474 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3478 * Remove the reference to the dbuf before removing its hold on the
3479 * dnode so we can guarantee in dnode_move() that a referenced bonus
3480 * buffer has a corresponding dnode hold.
3482 holds
= zfs_refcount_remove(&db
->db_holds
, tag
);
3486 * We can't freeze indirects if there is a possibility that they
3487 * may be modified in the current syncing context.
3489 if (db
->db_buf
!= NULL
&&
3490 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3491 arc_buf_freeze(db
->db_buf
);
3494 if (holds
== db
->db_dirtycnt
&&
3495 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3496 dbuf_evict_user(db
);
3499 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3501 boolean_t evict_dbuf
= db
->db_pending_evict
;
3504 * If the dnode moves here, we cannot cross this
3505 * barrier until the move completes.
3510 atomic_dec_32(&dn
->dn_dbufs_count
);
3513 * Decrementing the dbuf count means that the bonus
3514 * buffer's dnode hold is no longer discounted in
3515 * dnode_move(). The dnode cannot move until after
3516 * the dnode_rele() below.
3521 * Do not reference db after its lock is dropped.
3522 * Another thread may evict it.
3524 mutex_exit(&db
->db_mtx
);
3527 dnode_evict_bonus(dn
);
3530 } else if (db
->db_buf
== NULL
) {
3532 * This is a special case: we never associated this
3533 * dbuf with any data allocated from the ARC.
3535 ASSERT(db
->db_state
== DB_UNCACHED
||
3536 db
->db_state
== DB_NOFILL
);
3538 } else if (arc_released(db
->db_buf
)) {
3540 * This dbuf has anonymous data associated with it.
3544 boolean_t do_arc_evict
= B_FALSE
;
3546 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3548 if (!DBUF_IS_CACHEABLE(db
) &&
3549 db
->db_blkptr
!= NULL
&&
3550 !BP_IS_HOLE(db
->db_blkptr
) &&
3551 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
3552 do_arc_evict
= B_TRUE
;
3553 bp
= *db
->db_blkptr
;
3556 if (!DBUF_IS_CACHEABLE(db
) ||
3557 db
->db_pending_evict
) {
3559 } else if (!multilist_link_active(&db
->db_cache_link
)) {
3560 ASSERT3U(db
->db_caching_status
, ==,
3563 dbuf_cached_state_t dcs
=
3564 dbuf_include_in_metadata_cache(db
) ?
3565 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
3566 db
->db_caching_status
= dcs
;
3568 multilist_insert(dbuf_caches
[dcs
].cache
, db
);
3569 (void) zfs_refcount_add_many(
3570 &dbuf_caches
[dcs
].size
,
3571 db
->db
.db_size
, db
);
3573 if (dcs
== DB_DBUF_METADATA_CACHE
) {
3574 DBUF_STAT_BUMP(metadata_cache_count
);
3576 metadata_cache_size_bytes_max
,
3578 &dbuf_caches
[dcs
].size
));
3581 cache_levels
[db
->db_level
]);
3582 DBUF_STAT_BUMP(cache_count
);
3584 cache_levels_bytes
[db
->db_level
],
3586 DBUF_STAT_MAX(cache_size_bytes_max
,
3588 &dbuf_caches
[dcs
].size
));
3590 mutex_exit(&db
->db_mtx
);
3592 if (db
->db_caching_status
== DB_DBUF_CACHE
&&
3594 dbuf_evict_notify();
3599 arc_freed(spa
, &bp
);
3602 mutex_exit(&db
->db_mtx
);
3607 #pragma weak dmu_buf_refcount = dbuf_refcount
3609 dbuf_refcount(dmu_buf_impl_t
*db
)
3611 return (zfs_refcount_count(&db
->db_holds
));
3615 dmu_buf_user_refcount(dmu_buf_t
*db_fake
)
3618 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3620 mutex_enter(&db
->db_mtx
);
3621 ASSERT3U(zfs_refcount_count(&db
->db_holds
), >=, db
->db_dirtycnt
);
3622 holds
= zfs_refcount_count(&db
->db_holds
) - db
->db_dirtycnt
;
3623 mutex_exit(&db
->db_mtx
);
3629 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3630 dmu_buf_user_t
*new_user
)
3632 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3634 mutex_enter(&db
->db_mtx
);
3635 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3636 if (db
->db_user
== old_user
)
3637 db
->db_user
= new_user
;
3639 old_user
= db
->db_user
;
3640 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3641 mutex_exit(&db
->db_mtx
);
3647 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3649 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3653 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3655 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3657 db
->db_user_immediate_evict
= TRUE
;
3658 return (dmu_buf_set_user(db_fake
, user
));
3662 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3664 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3668 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3670 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3672 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3673 return (db
->db_user
);
3677 dmu_buf_user_evict_wait()
3679 taskq_wait(dbu_evict_taskq
);
3683 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3685 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3686 return (dbi
->db_blkptr
);
3690 dmu_buf_get_objset(dmu_buf_t
*db
)
3692 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3693 return (dbi
->db_objset
);
3697 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3699 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3700 DB_DNODE_ENTER(dbi
);
3701 return (DB_DNODE(dbi
));
3705 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3707 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3712 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3714 /* ASSERT(dmu_tx_is_syncing(tx) */
3715 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3717 if (db
->db_blkptr
!= NULL
)
3720 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3721 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3722 BP_ZERO(db
->db_blkptr
);
3725 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3727 * This buffer was allocated at a time when there was
3728 * no available blkptrs from the dnode, or it was
3729 * inappropriate to hook it in (i.e., nlevels mismatch).
3731 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3732 ASSERT(db
->db_parent
== NULL
);
3733 db
->db_parent
= dn
->dn_dbuf
;
3734 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3737 dmu_buf_impl_t
*parent
= db
->db_parent
;
3738 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3740 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3741 if (parent
== NULL
) {
3742 mutex_exit(&db
->db_mtx
);
3743 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3744 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3745 db
->db_blkid
>> epbs
, db
);
3746 rw_exit(&dn
->dn_struct_rwlock
);
3747 mutex_enter(&db
->db_mtx
);
3748 db
->db_parent
= parent
;
3750 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3751 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3757 * When syncing out a blocks of dnodes, adjust the block to deal with
3758 * encryption. Normally, we make sure the block is decrypted before writing
3759 * it. If we have crypt params, then we are writing a raw (encrypted) block,
3760 * from a raw receive. In this case, set the ARC buf's crypt params so
3761 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
3764 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t
*dr
)
3767 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3769 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3770 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
3771 ASSERT3U(db
->db_level
, ==, 0);
3773 if (!db
->db_objset
->os_raw_receive
&& arc_is_encrypted(db
->db_buf
)) {
3774 zbookmark_phys_t zb
;
3777 * Unfortunately, there is currently no mechanism for
3778 * syncing context to handle decryption errors. An error
3779 * here is only possible if an attacker maliciously
3780 * changed a dnode block and updated the associated
3781 * checksums going up the block tree.
3783 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
3784 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3785 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
3788 panic("Invalid dnode block MAC");
3789 } else if (dr
->dt
.dl
.dr_has_raw_params
) {
3790 (void) arc_release(dr
->dt
.dl
.dr_data
, db
);
3791 arc_convert_to_raw(dr
->dt
.dl
.dr_data
,
3792 dmu_objset_id(db
->db_objset
),
3793 dr
->dt
.dl
.dr_byteorder
, DMU_OT_DNODE
,
3794 dr
->dt
.dl
.dr_salt
, dr
->dt
.dl
.dr_iv
, dr
->dt
.dl
.dr_mac
);
3799 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3800 * is critical the we not allow the compiler to inline this function in to
3801 * dbuf_sync_list() thereby drastically bloating the stack usage.
3803 noinline
static void
3804 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3806 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3810 ASSERT(dmu_tx_is_syncing(tx
));
3812 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3814 mutex_enter(&db
->db_mtx
);
3816 ASSERT(db
->db_level
> 0);
3819 /* Read the block if it hasn't been read yet. */
3820 if (db
->db_buf
== NULL
) {
3821 mutex_exit(&db
->db_mtx
);
3822 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3823 mutex_enter(&db
->db_mtx
);
3825 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3826 ASSERT(db
->db_buf
!= NULL
);
3830 /* Indirect block size must match what the dnode thinks it is. */
3831 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3832 dbuf_check_blkptr(dn
, db
);
3835 /* Provide the pending dirty record to child dbufs */
3836 db
->db_data_pending
= dr
;
3838 mutex_exit(&db
->db_mtx
);
3840 dbuf_write(dr
, db
->db_buf
, tx
);
3843 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3844 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3845 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3846 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3852 * Verify that the size of the data in our bonus buffer does not exceed
3853 * its recorded size.
3855 * The purpose of this verification is to catch any cases in development
3856 * where the size of a phys structure (i.e space_map_phys_t) grows and,
3857 * due to incorrect feature management, older pools expect to read more
3858 * data even though they didn't actually write it to begin with.
3860 * For a example, this would catch an error in the feature logic where we
3861 * open an older pool and we expect to write the space map histogram of
3862 * a space map with size SPACE_MAP_SIZE_V0.
3865 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
)
3867 dnode_t
*dn
= DB_DNODE(dr
->dr_dbuf
);
3870 * Encrypted bonus buffers can have data past their bonuslen.
3871 * Skip the verification of these blocks.
3873 if (DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))
3876 uint16_t bonuslen
= dn
->dn_phys
->dn_bonuslen
;
3877 uint16_t maxbonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
3878 ASSERT3U(bonuslen
, <=, maxbonuslen
);
3880 arc_buf_t
*datap
= dr
->dt
.dl
.dr_data
;
3881 char *datap_end
= ((char *)datap
) + bonuslen
;
3882 char *datap_max
= ((char *)datap
) + maxbonuslen
;
3884 /* ensure that everything is zero after our data */
3885 for (; datap_end
< datap_max
; datap_end
++)
3886 ASSERT(*datap_end
== 0);
3891 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3892 * critical the we not allow the compiler to inline this function in to
3893 * dbuf_sync_list() thereby drastically bloating the stack usage.
3895 noinline
static void
3896 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3898 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3899 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3902 uint64_t txg
= tx
->tx_txg
;
3904 ASSERT(dmu_tx_is_syncing(tx
));
3906 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3908 mutex_enter(&db
->db_mtx
);
3910 * To be synced, we must be dirtied. But we
3911 * might have been freed after the dirty.
3913 if (db
->db_state
== DB_UNCACHED
) {
3914 /* This buffer has been freed since it was dirtied */
3915 ASSERT(db
->db
.db_data
== NULL
);
3916 } else if (db
->db_state
== DB_FILL
) {
3917 /* This buffer was freed and is now being re-filled */
3918 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3920 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3927 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3928 mutex_enter(&dn
->dn_mtx
);
3929 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3931 * In the previous transaction group, the bonus buffer
3932 * was entirely used to store the attributes for the
3933 * dnode which overrode the dn_spill field. However,
3934 * when adding more attributes to the file a spill
3935 * block was required to hold the extra attributes.
3937 * Make sure to clear the garbage left in the dn_spill
3938 * field from the previous attributes in the bonus
3939 * buffer. Otherwise, after writing out the spill
3940 * block to the new allocated dva, it will free
3941 * the old block pointed to by the invalid dn_spill.
3943 db
->db_blkptr
= NULL
;
3945 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3946 mutex_exit(&dn
->dn_mtx
);
3950 * If this is a bonus buffer, simply copy the bonus data into the
3951 * dnode. It will be written out when the dnode is synced (and it
3952 * will be synced, since it must have been dirty for dbuf_sync to
3955 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3956 dbuf_dirty_record_t
**drp
;
3958 ASSERT(*datap
!= NULL
);
3959 ASSERT0(db
->db_level
);
3960 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
3961 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3962 bcopy(*datap
, DN_BONUS(dn
->dn_phys
),
3963 DN_MAX_BONUS_LEN(dn
->dn_phys
));
3967 dbuf_sync_leaf_verify_bonus_dnode(dr
);
3970 if (*datap
!= db
->db
.db_data
) {
3971 int slots
= DB_DNODE(db
)->dn_num_slots
;
3972 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3973 kmem_free(*datap
, bonuslen
);
3974 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3976 db
->db_data_pending
= NULL
;
3977 drp
= &db
->db_last_dirty
;
3979 drp
= &(*drp
)->dr_next
;
3980 ASSERT(dr
->dr_next
== NULL
);
3981 ASSERT(dr
->dr_dbuf
== db
);
3983 if (dr
->dr_dbuf
->db_level
!= 0) {
3984 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3985 list_destroy(&dr
->dt
.di
.dr_children
);
3987 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3988 ASSERT(db
->db_dirtycnt
> 0);
3989 db
->db_dirtycnt
-= 1;
3990 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
, B_FALSE
);
3997 * This function may have dropped the db_mtx lock allowing a dmu_sync
3998 * operation to sneak in. As a result, we need to ensure that we
3999 * don't check the dr_override_state until we have returned from
4000 * dbuf_check_blkptr.
4002 dbuf_check_blkptr(dn
, db
);
4005 * If this buffer is in the middle of an immediate write,
4006 * wait for the synchronous IO to complete.
4008 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
4009 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
4010 cv_wait(&db
->db_changed
, &db
->db_mtx
);
4011 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
4015 * If this is a dnode block, ensure it is appropriately encrypted
4016 * or decrypted, depending on what we are writing to it this txg.
4018 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
4019 dbuf_prepare_encrypted_dnode_leaf(dr
);
4021 if (db
->db_state
!= DB_NOFILL
&&
4022 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
4023 zfs_refcount_count(&db
->db_holds
) > 1 &&
4024 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
4025 *datap
== db
->db_buf
) {
4027 * If this buffer is currently "in use" (i.e., there
4028 * are active holds and db_data still references it),
4029 * then make a copy before we start the write so that
4030 * any modifications from the open txg will not leak
4033 * NOTE: this copy does not need to be made for
4034 * objects only modified in the syncing context (e.g.
4035 * DNONE_DNODE blocks).
4037 int psize
= arc_buf_size(*datap
);
4038 int lsize
= arc_buf_lsize(*datap
);
4039 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
4040 enum zio_compress compress_type
= arc_get_compression(*datap
);
4042 if (arc_is_encrypted(*datap
)) {
4043 boolean_t byteorder
;
4044 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4045 uint8_t iv
[ZIO_DATA_IV_LEN
];
4046 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4048 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
4049 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
4050 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
4051 dn
->dn_type
, psize
, lsize
, compress_type
);
4052 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
4053 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
4054 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
4055 psize
, lsize
, compress_type
);
4057 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
4059 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
4061 db
->db_data_pending
= dr
;
4063 mutex_exit(&db
->db_mtx
);
4065 dbuf_write(dr
, *datap
, tx
);
4067 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4068 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
4069 list_insert_tail(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
4073 * Although zio_nowait() does not "wait for an IO", it does
4074 * initiate the IO. If this is an empty write it seems plausible
4075 * that the IO could actually be completed before the nowait
4076 * returns. We need to DB_DNODE_EXIT() first in case
4077 * zio_nowait() invalidates the dbuf.
4080 zio_nowait(dr
->dr_zio
);
4085 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
4087 dbuf_dirty_record_t
*dr
;
4089 while ((dr
= list_head(list
))) {
4090 if (dr
->dr_zio
!= NULL
) {
4092 * If we find an already initialized zio then we
4093 * are processing the meta-dnode, and we have finished.
4094 * The dbufs for all dnodes are put back on the list
4095 * during processing, so that we can zio_wait()
4096 * these IOs after initiating all child IOs.
4098 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
4099 DMU_META_DNODE_OBJECT
);
4102 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
4103 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
4104 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
4106 list_remove(list
, dr
);
4107 if (dr
->dr_dbuf
->db_level
> 0)
4108 dbuf_sync_indirect(dr
, tx
);
4110 dbuf_sync_leaf(dr
, tx
);
4116 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4118 dmu_buf_impl_t
*db
= vdb
;
4120 blkptr_t
*bp
= zio
->io_bp
;
4121 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4122 spa_t
*spa
= zio
->io_spa
;
4127 ASSERT3P(db
->db_blkptr
, !=, NULL
);
4128 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
4132 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
4133 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
4134 zio
->io_prev_space_delta
= delta
;
4136 if (bp
->blk_birth
!= 0) {
4137 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
4138 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
4139 (db
->db_blkid
== DMU_SPILL_BLKID
&&
4140 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
4141 BP_IS_EMBEDDED(bp
));
4142 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
4145 mutex_enter(&db
->db_mtx
);
4148 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4149 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4150 ASSERT(!(BP_IS_HOLE(bp
)) &&
4151 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4155 if (db
->db_level
== 0) {
4156 mutex_enter(&dn
->dn_mtx
);
4157 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
4158 db
->db_blkid
!= DMU_SPILL_BLKID
) {
4159 ASSERT0(db
->db_objset
->os_raw_receive
);
4160 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
4162 mutex_exit(&dn
->dn_mtx
);
4164 if (dn
->dn_type
== DMU_OT_DNODE
) {
4166 while (i
< db
->db
.db_size
) {
4168 (void *)(((char *)db
->db
.db_data
) + i
);
4170 i
+= DNODE_MIN_SIZE
;
4171 if (dnp
->dn_type
!= DMU_OT_NONE
) {
4173 i
+= dnp
->dn_extra_slots
*
4178 if (BP_IS_HOLE(bp
)) {
4185 blkptr_t
*ibp
= db
->db
.db_data
;
4186 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4187 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
4188 if (BP_IS_HOLE(ibp
))
4190 fill
+= BP_GET_FILL(ibp
);
4195 if (!BP_IS_EMBEDDED(bp
))
4196 BP_SET_FILL(bp
, fill
);
4198 mutex_exit(&db
->db_mtx
);
4200 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_WRITER
, FTAG
);
4201 *db
->db_blkptr
= *bp
;
4202 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
4207 * This function gets called just prior to running through the compression
4208 * stage of the zio pipeline. If we're an indirect block comprised of only
4209 * holes, then we want this indirect to be compressed away to a hole. In
4210 * order to do that we must zero out any information about the holes that
4211 * this indirect points to prior to before we try to compress it.
4214 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4216 dmu_buf_impl_t
*db
= vdb
;
4219 unsigned int epbs
, i
;
4221 ASSERT3U(db
->db_level
, >, 0);
4224 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4225 ASSERT3U(epbs
, <, 31);
4227 /* Determine if all our children are holes */
4228 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
4229 if (!BP_IS_HOLE(bp
))
4234 * If all the children are holes, then zero them all out so that
4235 * we may get compressed away.
4237 if (i
== 1ULL << epbs
) {
4239 * We only found holes. Grab the rwlock to prevent
4240 * anybody from reading the blocks we're about to
4243 rw_enter(&db
->db_rwlock
, RW_WRITER
);
4244 bzero(db
->db
.db_data
, db
->db
.db_size
);
4245 rw_exit(&db
->db_rwlock
);
4251 * The SPA will call this callback several times for each zio - once
4252 * for every physical child i/o (zio->io_phys_children times). This
4253 * allows the DMU to monitor the progress of each logical i/o. For example,
4254 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
4255 * block. There may be a long delay before all copies/fragments are completed,
4256 * so this callback allows us to retire dirty space gradually, as the physical
4261 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4263 dmu_buf_impl_t
*db
= arg
;
4264 objset_t
*os
= db
->db_objset
;
4265 dsl_pool_t
*dp
= dmu_objset_pool(os
);
4266 dbuf_dirty_record_t
*dr
;
4269 dr
= db
->db_data_pending
;
4270 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
4273 * The callback will be called io_phys_children times. Retire one
4274 * portion of our dirty space each time we are called. Any rounding
4275 * error will be cleaned up by dbuf_write_done().
4277 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
4278 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
4283 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4285 dmu_buf_impl_t
*db
= vdb
;
4286 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4287 blkptr_t
*bp
= db
->db_blkptr
;
4288 objset_t
*os
= db
->db_objset
;
4289 dmu_tx_t
*tx
= os
->os_synctx
;
4290 dbuf_dirty_record_t
**drp
, *dr
;
4292 ASSERT0(zio
->io_error
);
4293 ASSERT(db
->db_blkptr
== bp
);
4296 * For nopwrites and rewrites we ensure that the bp matches our
4297 * original and bypass all the accounting.
4299 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4300 ASSERT(BP_EQUAL(bp
, bp_orig
));
4302 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4303 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
4304 dsl_dataset_block_born(ds
, bp
, tx
);
4307 mutex_enter(&db
->db_mtx
);
4311 drp
= &db
->db_last_dirty
;
4312 while ((dr
= *drp
) != db
->db_data_pending
)
4314 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4315 ASSERT(dr
->dr_dbuf
== db
);
4316 ASSERT(dr
->dr_next
== NULL
);
4320 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4325 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4326 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
4327 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4332 if (db
->db_level
== 0) {
4333 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
4334 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
4335 if (db
->db_state
!= DB_NOFILL
) {
4336 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
4337 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
4344 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4345 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
4346 if (!BP_IS_HOLE(db
->db_blkptr
)) {
4347 int epbs __maybe_unused
= dn
->dn_phys
->dn_indblkshift
-
4349 ASSERT3U(db
->db_blkid
, <=,
4350 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
4351 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4355 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4356 list_destroy(&dr
->dt
.di
.dr_children
);
4359 cv_broadcast(&db
->db_changed
);
4360 ASSERT(db
->db_dirtycnt
> 0);
4361 db
->db_dirtycnt
-= 1;
4362 db
->db_data_pending
= NULL
;
4363 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4366 * If we didn't do a physical write in this ZIO and we
4367 * still ended up here, it means that the space of the
4368 * dbuf that we just released (and undirtied) above hasn't
4369 * been marked as undirtied in the pool's accounting.
4371 * Thus, we undirty that space in the pool's view of the
4372 * world here. For physical writes this type of update
4373 * happens in dbuf_write_physdone().
4375 * If we did a physical write, cleanup any rounding errors
4376 * that came up due to writing multiple copies of a block
4377 * on disk [see dbuf_write_physdone()].
4379 if (zio
->io_phys_children
== 0) {
4380 dsl_pool_undirty_space(dmu_objset_pool(os
),
4381 dr
->dr_accounted
, zio
->io_txg
);
4383 dsl_pool_undirty_space(dmu_objset_pool(os
),
4384 dr
->dr_accounted
% zio
->io_phys_children
, zio
->io_txg
);
4387 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4391 dbuf_write_nofill_ready(zio_t
*zio
)
4393 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4397 dbuf_write_nofill_done(zio_t
*zio
)
4399 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4403 dbuf_write_override_ready(zio_t
*zio
)
4405 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4406 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4408 dbuf_write_ready(zio
, NULL
, db
);
4412 dbuf_write_override_done(zio_t
*zio
)
4414 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4415 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4416 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4418 mutex_enter(&db
->db_mtx
);
4419 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4420 if (!BP_IS_HOLE(obp
))
4421 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4422 arc_release(dr
->dt
.dl
.dr_data
, db
);
4424 mutex_exit(&db
->db_mtx
);
4426 dbuf_write_done(zio
, NULL
, db
);
4428 if (zio
->io_abd
!= NULL
)
4429 abd_put(zio
->io_abd
);
4432 typedef struct dbuf_remap_impl_callback_arg
{
4434 uint64_t drica_blk_birth
;
4436 } dbuf_remap_impl_callback_arg_t
;
4439 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
4442 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
4443 objset_t
*os
= drica
->drica_os
;
4444 spa_t
*spa
= dmu_objset_spa(os
);
4445 dmu_tx_t
*tx
= drica
->drica_tx
;
4447 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4449 if (os
== spa_meta_objset(spa
)) {
4450 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
4452 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
4453 size
, drica
->drica_blk_birth
, tx
);
4458 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, krwlock_t
*rw
, dmu_tx_t
*tx
)
4460 blkptr_t bp_copy
= *bp
;
4461 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4462 dbuf_remap_impl_callback_arg_t drica
;
4464 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4466 drica
.drica_os
= dn
->dn_objset
;
4467 drica
.drica_blk_birth
= bp
->blk_birth
;
4468 drica
.drica_tx
= tx
;
4469 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
4472 * If the blkptr being remapped is tracked by a livelist,
4473 * then we need to make sure the livelist reflects the update.
4474 * First, cancel out the old blkptr by appending a 'FREE'
4475 * entry. Next, add an 'ALLOC' to track the new version. This
4476 * way we avoid trying to free an inaccurate blkptr at delete.
4477 * Note that embedded blkptrs are not tracked in livelists.
4479 if (dn
->dn_objset
!= spa_meta_objset(spa
)) {
4480 dsl_dataset_t
*ds
= dmu_objset_ds(dn
->dn_objset
);
4481 if (dsl_deadlist_is_open(&ds
->ds_dir
->dd_livelist
) &&
4482 bp
->blk_birth
> ds
->ds_dir
->dd_origin_txg
) {
4483 ASSERT(!BP_IS_EMBEDDED(bp
));
4484 ASSERT(dsl_dir_is_clone(ds
->ds_dir
));
4485 ASSERT(spa_feature_is_enabled(spa
,
4486 SPA_FEATURE_LIVELIST
));
4487 bplist_append(&ds
->ds_dir
->dd_pending_frees
,
4489 bplist_append(&ds
->ds_dir
->dd_pending_allocs
,
4495 * The db_rwlock prevents dbuf_read_impl() from
4496 * dereferencing the BP while we are changing it. To
4497 * avoid lock contention, only grab it when we are actually
4501 rw_enter(rw
, RW_WRITER
);
4509 * Remap any existing BP's to concrete vdevs, if possible.
4512 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
4514 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
4515 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4517 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
4520 if (db
->db_level
> 0) {
4521 blkptr_t
*bp
= db
->db
.db_data
;
4522 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
4523 dbuf_remap_impl(dn
, &bp
[i
], &db
->db_rwlock
, tx
);
4525 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
4526 dnode_phys_t
*dnp
= db
->db
.db_data
;
4527 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
4529 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
;
4530 i
+= dnp
[i
].dn_extra_slots
+ 1) {
4531 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
4532 krwlock_t
*lock
= (dn
->dn_dbuf
== NULL
? NULL
:
4533 &dn
->dn_dbuf
->db_rwlock
);
4534 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], lock
,
4542 /* Issue I/O to commit a dirty buffer to disk. */
4544 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
4546 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4549 dmu_buf_impl_t
*parent
= db
->db_parent
;
4550 uint64_t txg
= tx
->tx_txg
;
4551 zbookmark_phys_t zb
;
4556 ASSERT(dmu_tx_is_syncing(tx
));
4562 if (db
->db_state
!= DB_NOFILL
) {
4563 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
4565 * Private object buffers are released here rather
4566 * than in dbuf_dirty() since they are only modified
4567 * in the syncing context and we don't want the
4568 * overhead of making multiple copies of the data.
4570 if (BP_IS_HOLE(db
->db_blkptr
)) {
4573 dbuf_release_bp(db
);
4575 dbuf_remap(dn
, db
, tx
);
4579 if (parent
!= dn
->dn_dbuf
) {
4580 /* Our parent is an indirect block. */
4581 /* We have a dirty parent that has been scheduled for write. */
4582 ASSERT(parent
&& parent
->db_data_pending
);
4583 /* Our parent's buffer is one level closer to the dnode. */
4584 ASSERT(db
->db_level
== parent
->db_level
-1);
4586 * We're about to modify our parent's db_data by modifying
4587 * our block pointer, so the parent must be released.
4589 ASSERT(arc_released(parent
->db_buf
));
4590 zio
= parent
->db_data_pending
->dr_zio
;
4592 /* Our parent is the dnode itself. */
4593 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
4594 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
4595 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
4596 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
4597 ASSERT3P(db
->db_blkptr
, ==,
4598 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
4602 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
4603 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
4606 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
4607 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
4608 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4610 if (db
->db_blkid
== DMU_SPILL_BLKID
)
4612 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
4614 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
4618 * We copy the blkptr now (rather than when we instantiate the dirty
4619 * record), because its value can change between open context and
4620 * syncing context. We do not need to hold dn_struct_rwlock to read
4621 * db_blkptr because we are in syncing context.
4623 dr
->dr_bp_copy
= *db
->db_blkptr
;
4625 if (db
->db_level
== 0 &&
4626 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
4628 * The BP for this block has been provided by open context
4629 * (by dmu_sync() or dmu_buf_write_embedded()).
4631 abd_t
*contents
= (data
!= NULL
) ?
4632 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
4634 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4635 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
4636 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
4637 dbuf_write_override_done
,
4638 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
4639 mutex_enter(&db
->db_mtx
);
4640 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
4641 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
4642 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
4643 mutex_exit(&db
->db_mtx
);
4644 } else if (db
->db_state
== DB_NOFILL
) {
4645 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
4646 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
4647 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4648 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
4649 dbuf_write_nofill_ready
, NULL
, NULL
,
4650 dbuf_write_nofill_done
, db
,
4651 ZIO_PRIORITY_ASYNC_WRITE
,
4652 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
4654 ASSERT(arc_released(data
));
4657 * For indirect blocks, we want to setup the children
4658 * ready callback so that we can properly handle an indirect
4659 * block that only contains holes.
4661 arc_write_done_func_t
*children_ready_cb
= NULL
;
4662 if (db
->db_level
!= 0)
4663 children_ready_cb
= dbuf_write_children_ready
;
4665 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
4666 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
4667 &zp
, dbuf_write_ready
,
4668 children_ready_cb
, dbuf_write_physdone
,
4669 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
4670 ZIO_FLAG_MUSTSUCCEED
, &zb
);
4674 EXPORT_SYMBOL(dbuf_find
);
4675 EXPORT_SYMBOL(dbuf_is_metadata
);
4676 EXPORT_SYMBOL(dbuf_destroy
);
4677 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
4678 EXPORT_SYMBOL(dbuf_whichblock
);
4679 EXPORT_SYMBOL(dbuf_read
);
4680 EXPORT_SYMBOL(dbuf_unoverride
);
4681 EXPORT_SYMBOL(dbuf_free_range
);
4682 EXPORT_SYMBOL(dbuf_new_size
);
4683 EXPORT_SYMBOL(dbuf_release_bp
);
4684 EXPORT_SYMBOL(dbuf_dirty
);
4685 EXPORT_SYMBOL(dmu_buf_set_crypt_params
);
4686 EXPORT_SYMBOL(dmu_buf_will_dirty
);
4687 EXPORT_SYMBOL(dmu_buf_is_dirty
);
4688 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
4689 EXPORT_SYMBOL(dmu_buf_will_fill
);
4690 EXPORT_SYMBOL(dmu_buf_fill_done
);
4691 EXPORT_SYMBOL(dmu_buf_rele
);
4692 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
4693 EXPORT_SYMBOL(dbuf_prefetch
);
4694 EXPORT_SYMBOL(dbuf_hold_impl
);
4695 EXPORT_SYMBOL(dbuf_hold
);
4696 EXPORT_SYMBOL(dbuf_hold_level
);
4697 EXPORT_SYMBOL(dbuf_create_bonus
);
4698 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
4699 EXPORT_SYMBOL(dbuf_rm_spill
);
4700 EXPORT_SYMBOL(dbuf_add_ref
);
4701 EXPORT_SYMBOL(dbuf_rele
);
4702 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
4703 EXPORT_SYMBOL(dbuf_refcount
);
4704 EXPORT_SYMBOL(dbuf_sync_list
);
4705 EXPORT_SYMBOL(dmu_buf_set_user
);
4706 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
4707 EXPORT_SYMBOL(dmu_buf_get_user
);
4708 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
4711 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, max_bytes
, ULONG
, ZMOD_RW
,
4712 "Maximum size in bytes of the dbuf cache.");
4714 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, hiwater_pct
, UINT
, ZMOD_RW
,
4715 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4718 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, lowater_pct
, UINT
, ZMOD_RW
,
4719 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4722 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_max_bytes
, ULONG
, ZMOD_RW
,
4723 "Maximum size in bytes of the dbuf metadata cache.");
4725 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, cache_shift
, INT
, ZMOD_RW
,
4726 "Set the size of the dbuf cache to a log2 fraction of arc size.");
4728 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_shift
, INT
, ZMOD_RW
,
4729 "Set the size of the dbuf metadata cache to a log2 fraction of arc "