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_lowater(void)
633 return (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
634 dbuf_cache_lowater_bytes());
638 * Evict the oldest eligible dbuf from the dbuf cache.
643 int idx
= multilist_get_random_index(dbuf_caches
[DB_DBUF_CACHE
].cache
);
644 multilist_sublist_t
*mls
= multilist_sublist_lock(
645 dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
647 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
649 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
650 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
651 db
= multilist_sublist_prev(mls
, db
);
654 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
655 multilist_sublist_t
*, mls
);
658 multilist_sublist_remove(mls
, db
);
659 multilist_sublist_unlock(mls
);
660 (void) zfs_refcount_remove_many(
661 &dbuf_caches
[DB_DBUF_CACHE
].size
, db
->db
.db_size
, db
);
662 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
663 DBUF_STAT_BUMPDOWN(cache_count
);
664 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
666 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
667 db
->db_caching_status
= DB_NO_CACHE
;
669 DBUF_STAT_BUMP(cache_total_evicts
);
671 multilist_sublist_unlock(mls
);
676 * The dbuf evict thread is responsible for aging out dbufs from the
677 * cache. Once the cache has reached it's maximum size, dbufs are removed
678 * and destroyed. The eviction thread will continue running until the size
679 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
680 * out of the cache it is destroyed and becomes eligible for arc eviction.
684 dbuf_evict_thread(void *unused
)
688 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
690 mutex_enter(&dbuf_evict_lock
);
691 while (!dbuf_evict_thread_exit
) {
692 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
693 CALLB_CPR_SAFE_BEGIN(&cpr
);
694 (void) cv_timedwait_sig_hires(&dbuf_evict_cv
,
695 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
696 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
698 mutex_exit(&dbuf_evict_lock
);
701 * Keep evicting as long as we're above the low water mark
702 * for the cache. We do this without holding the locks to
703 * minimize lock contention.
705 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
709 mutex_enter(&dbuf_evict_lock
);
712 dbuf_evict_thread_exit
= B_FALSE
;
713 cv_broadcast(&dbuf_evict_cv
);
714 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
719 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
720 * If the dbuf cache is at its high water mark, then evict a dbuf from the
721 * dbuf cache using the callers context.
724 dbuf_evict_notify(uint64_t size
)
727 * We check if we should evict without holding the dbuf_evict_lock,
728 * because it's OK to occasionally make the wrong decision here,
729 * and grabbing the lock results in massive lock contention.
731 if (size
> dbuf_cache_target_bytes()) {
732 if (size
> dbuf_cache_hiwater_bytes())
734 cv_signal(&dbuf_evict_cv
);
739 dbuf_kstat_update(kstat_t
*ksp
, int rw
)
741 dbuf_stats_t
*ds
= ksp
->ks_data
;
743 if (rw
== KSTAT_WRITE
) {
744 return (SET_ERROR(EACCES
));
746 ds
->metadata_cache_size_bytes
.value
.ui64
= zfs_refcount_count(
747 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
);
748 ds
->cache_size_bytes
.value
.ui64
=
749 zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
);
750 ds
->cache_target_bytes
.value
.ui64
= dbuf_cache_target_bytes();
751 ds
->cache_hiwater_bytes
.value
.ui64
= dbuf_cache_hiwater_bytes();
752 ds
->cache_lowater_bytes
.value
.ui64
= dbuf_cache_lowater_bytes();
753 ds
->hash_elements
.value
.ui64
= dbuf_hash_count
;
762 uint64_t hsize
= 1ULL << 16;
763 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
767 * The hash table is big enough to fill all of physical memory
768 * with an average block size of zfs_arc_average_blocksize (default 8K).
769 * By default, the table will take up
770 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
772 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
776 h
->hash_table_mask
= hsize
- 1;
779 * Large allocations which do not require contiguous pages
780 * should be using vmem_alloc() in the linux kernel
782 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
784 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
786 if (h
->hash_table
== NULL
) {
787 /* XXX - we should really return an error instead of assert */
788 ASSERT(hsize
> (1ULL << 10));
793 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
794 sizeof (dmu_buf_impl_t
),
795 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
797 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
798 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
803 * Setup the parameters for the dbuf caches. We set the sizes of the
804 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
805 * of the target size of the ARC. If the values has been specified as
806 * a module option and they're not greater than the target size of the
807 * ARC, then we honor that value.
809 if (dbuf_cache_max_bytes
== 0 ||
810 dbuf_cache_max_bytes
>= arc_target_bytes()) {
811 dbuf_cache_max_bytes
= arc_target_bytes() >> dbuf_cache_shift
;
813 if (dbuf_metadata_cache_max_bytes
== 0 ||
814 dbuf_metadata_cache_max_bytes
>= arc_target_bytes()) {
815 dbuf_metadata_cache_max_bytes
=
816 arc_target_bytes() >> dbuf_metadata_cache_shift
;
820 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
821 * configuration is not required.
823 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
825 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
826 dbuf_caches
[dcs
].cache
=
827 multilist_create(sizeof (dmu_buf_impl_t
),
828 offsetof(dmu_buf_impl_t
, db_cache_link
),
829 dbuf_cache_multilist_index_func
);
830 zfs_refcount_create(&dbuf_caches
[dcs
].size
);
833 dbuf_evict_thread_exit
= B_FALSE
;
834 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
835 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
836 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
837 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
839 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
840 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
842 if (dbuf_ksp
!= NULL
) {
843 for (i
= 0; i
< DN_MAX_LEVELS
; i
++) {
844 snprintf(dbuf_stats
.cache_levels
[i
].name
,
845 KSTAT_STRLEN
, "cache_level_%d", i
);
846 dbuf_stats
.cache_levels
[i
].data_type
=
848 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
849 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
850 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
853 dbuf_ksp
->ks_data
= &dbuf_stats
;
854 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
855 kstat_install(dbuf_ksp
);
862 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
865 dbuf_stats_destroy();
867 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
868 mutex_destroy(&h
->hash_mutexes
[i
]);
871 * Large allocations which do not require contiguous pages
872 * should be using vmem_free() in the linux kernel
874 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
876 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
878 kmem_cache_destroy(dbuf_kmem_cache
);
879 taskq_destroy(dbu_evict_taskq
);
881 mutex_enter(&dbuf_evict_lock
);
882 dbuf_evict_thread_exit
= B_TRUE
;
883 while (dbuf_evict_thread_exit
) {
884 cv_signal(&dbuf_evict_cv
);
885 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
887 mutex_exit(&dbuf_evict_lock
);
889 mutex_destroy(&dbuf_evict_lock
);
890 cv_destroy(&dbuf_evict_cv
);
892 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
893 zfs_refcount_destroy(&dbuf_caches
[dcs
].size
);
894 multilist_destroy(dbuf_caches
[dcs
].cache
);
897 if (dbuf_ksp
!= NULL
) {
898 kstat_delete(dbuf_ksp
);
909 dbuf_verify(dmu_buf_impl_t
*db
)
912 dbuf_dirty_record_t
*dr
;
915 ASSERT(MUTEX_HELD(&db
->db_mtx
));
917 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
920 ASSERT(db
->db_objset
!= NULL
);
924 ASSERT(db
->db_parent
== NULL
);
925 ASSERT(db
->db_blkptr
== NULL
);
927 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
928 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
929 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
930 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
931 db
->db_blkid
== DMU_SPILL_BLKID
||
932 !avl_is_empty(&dn
->dn_dbufs
));
934 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
936 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
937 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
938 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
940 ASSERT0(db
->db
.db_offset
);
942 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
945 if ((dr
= list_head(&db
->db_dirty_records
)) != NULL
) {
946 ASSERT(dr
->dr_dbuf
== db
);
947 txg_prev
= dr
->dr_txg
;
948 for (dr
= list_next(&db
->db_dirty_records
, dr
); dr
!= NULL
;
949 dr
= list_next(&db
->db_dirty_records
, dr
)) {
950 ASSERT(dr
->dr_dbuf
== db
);
951 ASSERT(txg_prev
> dr
->dr_txg
);
952 txg_prev
= dr
->dr_txg
;
957 * We can't assert that db_size matches dn_datablksz because it
958 * can be momentarily different when another thread is doing
961 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
962 dr
= db
->db_data_pending
;
964 * It should only be modified in syncing context, so
965 * make sure we only have one copy of the data.
967 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
970 /* verify db->db_blkptr */
972 if (db
->db_parent
== dn
->dn_dbuf
) {
973 /* db is pointed to by the dnode */
974 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
975 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
976 ASSERT(db
->db_parent
== NULL
);
978 ASSERT(db
->db_parent
!= NULL
);
979 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
980 ASSERT3P(db
->db_blkptr
, ==,
981 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
983 /* db is pointed to by an indirect block */
984 int epb __maybe_unused
= db
->db_parent
->db
.db_size
>>
986 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
987 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
990 * dnode_grow_indblksz() can make this fail if we don't
991 * have the parent's rwlock. XXX indblksz no longer
992 * grows. safe to do this now?
994 if (RW_LOCK_HELD(&db
->db_parent
->db_rwlock
)) {
995 ASSERT3P(db
->db_blkptr
, ==,
996 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
997 db
->db_blkid
% epb
));
1001 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
1002 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
1003 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
1004 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
1006 * If the blkptr isn't set but they have nonzero data,
1007 * it had better be dirty, otherwise we'll lose that
1008 * data when we evict this buffer.
1010 * There is an exception to this rule for indirect blocks; in
1011 * this case, if the indirect block is a hole, we fill in a few
1012 * fields on each of the child blocks (importantly, birth time)
1013 * to prevent hole birth times from being lost when you
1014 * partially fill in a hole.
1016 if (db
->db_dirtycnt
== 0) {
1017 if (db
->db_level
== 0) {
1018 uint64_t *buf
= db
->db
.db_data
;
1021 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
1022 ASSERT(buf
[i
] == 0);
1025 blkptr_t
*bps
= db
->db
.db_data
;
1026 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1029 * We want to verify that all the blkptrs in the
1030 * indirect block are holes, but we may have
1031 * automatically set up a few fields for them.
1032 * We iterate through each blkptr and verify
1033 * they only have those fields set.
1036 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1038 blkptr_t
*bp
= &bps
[i
];
1039 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1042 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1043 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1044 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1045 ASSERT0(bp
->blk_fill
);
1046 ASSERT0(bp
->blk_pad
[0]);
1047 ASSERT0(bp
->blk_pad
[1]);
1048 ASSERT(!BP_IS_EMBEDDED(bp
));
1049 ASSERT(BP_IS_HOLE(bp
));
1050 ASSERT0(bp
->blk_phys_birth
);
1060 dbuf_clear_data(dmu_buf_impl_t
*db
)
1062 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1063 dbuf_evict_user(db
);
1064 ASSERT3P(db
->db_buf
, ==, NULL
);
1065 db
->db
.db_data
= NULL
;
1066 if (db
->db_state
!= DB_NOFILL
)
1067 db
->db_state
= DB_UNCACHED
;
1071 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1073 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1074 ASSERT(buf
!= NULL
);
1077 ASSERT(buf
->b_data
!= NULL
);
1078 db
->db
.db_data
= buf
->b_data
;
1082 * Loan out an arc_buf for read. Return the loaned arc_buf.
1085 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1089 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1090 mutex_enter(&db
->db_mtx
);
1091 if (arc_released(db
->db_buf
) || zfs_refcount_count(&db
->db_holds
) > 1) {
1092 int blksz
= db
->db
.db_size
;
1093 spa_t
*spa
= db
->db_objset
->os_spa
;
1095 mutex_exit(&db
->db_mtx
);
1096 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1097 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
1100 arc_loan_inuse_buf(abuf
, db
);
1102 dbuf_clear_data(db
);
1103 mutex_exit(&db
->db_mtx
);
1109 * Calculate which level n block references the data at the level 0 offset
1113 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1115 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1117 * The level n blkid is equal to the level 0 blkid divided by
1118 * the number of level 0s in a level n block.
1120 * The level 0 blkid is offset >> datablkshift =
1121 * offset / 2^datablkshift.
1123 * The number of level 0s in a level n is the number of block
1124 * pointers in an indirect block, raised to the power of level.
1125 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1126 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1128 * Thus, the level n blkid is: offset /
1129 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1130 * = offset / 2^(datablkshift + level *
1131 * (indblkshift - SPA_BLKPTRSHIFT))
1132 * = offset >> (datablkshift + level *
1133 * (indblkshift - SPA_BLKPTRSHIFT))
1136 const unsigned exp
= dn
->dn_datablkshift
+
1137 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1139 if (exp
>= 8 * sizeof (offset
)) {
1140 /* This only happens on the highest indirection level */
1141 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1145 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1147 return (offset
>> exp
);
1149 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1155 * This function is used to lock the parent of the provided dbuf. This should be
1156 * used when modifying or reading db_blkptr.
1159 dmu_buf_lock_parent(dmu_buf_impl_t
*db
, krw_t rw
, void *tag
)
1161 enum db_lock_type ret
= DLT_NONE
;
1162 if (db
->db_parent
!= NULL
) {
1163 rw_enter(&db
->db_parent
->db_rwlock
, rw
);
1165 } else if (dmu_objset_ds(db
->db_objset
) != NULL
) {
1166 rrw_enter(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, rw
,
1171 * We only return a DLT_NONE lock when it's the top-most indirect block
1172 * of the meta-dnode of the MOS.
1178 * We need to pass the lock type in because it's possible that the block will
1179 * move from being the topmost indirect block in a dnode (and thus, have no
1180 * parent) to not the top-most via an indirection increase. This would cause a
1181 * panic if we didn't pass the lock type in.
1184 dmu_buf_unlock_parent(dmu_buf_impl_t
*db
, db_lock_type_t type
, void *tag
)
1186 if (type
== DLT_PARENT
)
1187 rw_exit(&db
->db_parent
->db_rwlock
);
1188 else if (type
== DLT_OBJSET
)
1189 rrw_exit(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, tag
);
1193 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1194 arc_buf_t
*buf
, void *vdb
)
1196 dmu_buf_impl_t
*db
= vdb
;
1198 mutex_enter(&db
->db_mtx
);
1199 ASSERT3U(db
->db_state
, ==, DB_READ
);
1201 * All reads are synchronous, so we must have a hold on the dbuf
1203 ASSERT(zfs_refcount_count(&db
->db_holds
) > 0);
1204 ASSERT(db
->db_buf
== NULL
);
1205 ASSERT(db
->db
.db_data
== NULL
);
1208 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
1209 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1210 ASSERT3P(db
->db_buf
, ==, NULL
);
1211 db
->db_state
= DB_UNCACHED
;
1212 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1213 /* freed in flight */
1214 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1215 arc_release(buf
, db
);
1216 bzero(buf
->b_data
, db
->db
.db_size
);
1217 arc_buf_freeze(buf
);
1218 db
->db_freed_in_flight
= FALSE
;
1219 dbuf_set_data(db
, buf
);
1220 db
->db_state
= DB_CACHED
;
1223 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1224 dbuf_set_data(db
, buf
);
1225 db
->db_state
= DB_CACHED
;
1227 cv_broadcast(&db
->db_changed
);
1228 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
1233 * This function ensures that, when doing a decrypting read of a block,
1234 * we make sure we have decrypted the dnode associated with it. We must do
1235 * this so that we ensure we are fully authenticating the checksum-of-MACs
1236 * tree from the root of the objset down to this block. Indirect blocks are
1237 * always verified against their secure checksum-of-MACs assuming that the
1238 * dnode containing them is correct. Now that we are doing a decrypting read,
1239 * we can be sure that the key is loaded and verify that assumption. This is
1240 * especially important considering that we always read encrypted dnode
1241 * blocks as raw data (without verifying their MACs) to start, and
1242 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1245 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
)
1248 objset_t
*os
= db
->db_objset
;
1249 arc_buf_t
*dnode_abuf
;
1251 zbookmark_phys_t zb
;
1253 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1255 if (!os
->os_encrypted
|| os
->os_raw_receive
||
1256 (flags
& DB_RF_NO_DECRYPT
) != 0)
1261 dnode_abuf
= (dn
->dn_dbuf
!= NULL
) ? dn
->dn_dbuf
->db_buf
: NULL
;
1263 if (dnode_abuf
== NULL
|| !arc_is_encrypted(dnode_abuf
)) {
1268 SET_BOOKMARK(&zb
, dmu_objset_id(os
),
1269 DMU_META_DNODE_OBJECT
, 0, dn
->dn_dbuf
->db_blkid
);
1270 err
= arc_untransform(dnode_abuf
, os
->os_spa
, &zb
, B_TRUE
);
1273 * An error code of EACCES tells us that the key is still not
1274 * available. This is ok if we are only reading authenticated
1275 * (and therefore non-encrypted) blocks.
1277 if (err
== EACCES
&& ((db
->db_blkid
!= DMU_BONUS_BLKID
&&
1278 !DMU_OT_IS_ENCRYPTED(dn
->dn_type
)) ||
1279 (db
->db_blkid
== DMU_BONUS_BLKID
&&
1280 !DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))))
1289 * Drops db_mtx and the parent lock specified by dblt and tag before
1293 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
,
1294 db_lock_type_t dblt
, void *tag
)
1297 zbookmark_phys_t zb
;
1298 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1299 int err
, zio_flags
= 0;
1303 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1304 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1305 ASSERT(db
->db_state
== DB_UNCACHED
);
1306 ASSERT(db
->db_buf
== NULL
);
1307 ASSERT(db
->db_parent
== NULL
||
1308 RW_LOCK_HELD(&db
->db_parent
->db_rwlock
));
1310 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1312 * The bonus length stored in the dnode may be less than
1313 * the maximum available space in the bonus buffer.
1315 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1316 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1318 /* if the underlying dnode block is encrypted, decrypt it */
1319 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1322 mutex_exit(&db
->db_mtx
);
1326 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1327 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1328 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1329 if (bonuslen
< max_bonuslen
)
1330 bzero(db
->db
.db_data
, max_bonuslen
);
1332 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1334 db
->db_state
= DB_CACHED
;
1335 mutex_exit(&db
->db_mtx
);
1336 dmu_buf_unlock_parent(db
, dblt
, tag
);
1341 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1342 * processes the delete record and clears the bp while we are waiting
1343 * for the dn_mtx (resulting in a "no" from block_freed).
1345 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1346 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1347 BP_IS_HOLE(db
->db_blkptr
)))) {
1348 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1350 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1352 bzero(db
->db
.db_data
, db
->db
.db_size
);
1354 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1355 BP_IS_HOLE(db
->db_blkptr
) &&
1356 db
->db_blkptr
->blk_birth
!= 0) {
1357 blkptr_t
*bps
= db
->db
.db_data
;
1358 for (int i
= 0; i
< ((1 <<
1359 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1361 blkptr_t
*bp
= &bps
[i
];
1362 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1363 1 << dn
->dn_indblkshift
);
1365 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1367 BP_GET_LSIZE(db
->db_blkptr
));
1368 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1370 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1371 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1375 db
->db_state
= DB_CACHED
;
1376 mutex_exit(&db
->db_mtx
);
1377 dmu_buf_unlock_parent(db
, dblt
, tag
);
1382 * Any attempt to read a redacted block should result in an error. This
1383 * will never happen under normal conditions, but can be useful for
1384 * debugging purposes.
1386 if (BP_IS_REDACTED(db
->db_blkptr
)) {
1387 ASSERT(dsl_dataset_feature_is_active(
1388 db
->db_objset
->os_dsl_dataset
,
1389 SPA_FEATURE_REDACTED_DATASETS
));
1391 mutex_exit(&db
->db_mtx
);
1392 return (SET_ERROR(EIO
));
1396 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1397 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1400 * All bps of an encrypted os should have the encryption bit set.
1401 * If this is not true it indicates tampering and we report an error.
1403 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(db
->db_blkptr
)) {
1404 spa_log_error(db
->db_objset
->os_spa
, &zb
);
1405 zfs_panic_recover("unencrypted block in encrypted "
1406 "object set %llu", dmu_objset_id(db
->db_objset
));
1408 mutex_exit(&db
->db_mtx
);
1409 dmu_buf_unlock_parent(db
, dblt
, tag
);
1410 return (SET_ERROR(EIO
));
1413 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1416 dmu_buf_unlock_parent(db
, dblt
, tag
);
1417 mutex_exit(&db
->db_mtx
);
1423 db
->db_state
= DB_READ
;
1424 mutex_exit(&db
->db_mtx
);
1426 if (DBUF_IS_L2CACHEABLE(db
))
1427 aflags
|= ARC_FLAG_L2CACHE
;
1429 dbuf_add_ref(db
, NULL
);
1431 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1432 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1434 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1435 zio_flags
|= ZIO_FLAG_RAW
;
1437 * The zio layer will copy the provided blkptr later, but we need to
1438 * do this now so that we can release the parent's rwlock. We have to
1439 * do that now so that if dbuf_read_done is called synchronously (on
1440 * an l1 cache hit) we don't acquire the db_mtx while holding the
1441 * parent's rwlock, which would be a lock ordering violation.
1443 blkptr_t bp
= *db
->db_blkptr
;
1444 dmu_buf_unlock_parent(db
, dblt
, tag
);
1445 (void) arc_read(zio
, db
->db_objset
->os_spa
, &bp
,
1446 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1452 * This is our just-in-time copy function. It makes a copy of buffers that
1453 * have been modified in a previous transaction group before we access them in
1454 * the current active group.
1456 * This function is used in three places: when we are dirtying a buffer for the
1457 * first time in a txg, when we are freeing a range in a dnode that includes
1458 * this buffer, and when we are accessing a buffer which was received compressed
1459 * and later referenced in a WRITE_BYREF record.
1461 * Note that when we are called from dbuf_free_range() we do not put a hold on
1462 * the buffer, we just traverse the active dbuf list for the dnode.
1465 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1467 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
1469 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1470 ASSERT(db
->db
.db_data
!= NULL
);
1471 ASSERT(db
->db_level
== 0);
1472 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1475 (dr
->dt
.dl
.dr_data
!=
1476 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1480 * If the last dirty record for this dbuf has not yet synced
1481 * and its referencing the dbuf data, either:
1482 * reset the reference to point to a new copy,
1483 * or (if there a no active holders)
1484 * just null out the current db_data pointer.
1486 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1487 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1488 dnode_t
*dn
= DB_DNODE(db
);
1489 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1490 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1491 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1492 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1493 } else if (zfs_refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1494 dnode_t
*dn
= DB_DNODE(db
);
1495 int size
= arc_buf_size(db
->db_buf
);
1496 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1497 spa_t
*spa
= db
->db_objset
->os_spa
;
1498 enum zio_compress compress_type
=
1499 arc_get_compression(db
->db_buf
);
1501 if (arc_is_encrypted(db
->db_buf
)) {
1502 boolean_t byteorder
;
1503 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1504 uint8_t iv
[ZIO_DATA_IV_LEN
];
1505 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1507 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1509 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1510 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1511 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1513 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1514 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1515 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1516 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1518 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1520 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1523 dbuf_clear_data(db
);
1528 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1535 * We don't have to hold the mutex to check db_state because it
1536 * can't be freed while we have a hold on the buffer.
1538 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1540 if (db
->db_state
== DB_NOFILL
)
1541 return (SET_ERROR(EIO
));
1546 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1547 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1548 DBUF_IS_CACHEABLE(db
);
1550 mutex_enter(&db
->db_mtx
);
1551 if (db
->db_state
== DB_CACHED
) {
1552 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1555 * Ensure that this block's dnode has been decrypted if
1556 * the caller has requested decrypted data.
1558 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1561 * If the arc buf is compressed or encrypted and the caller
1562 * requested uncompressed data, we need to untransform it
1563 * before returning. We also call arc_untransform() on any
1564 * unauthenticated blocks, which will verify their MAC if
1565 * the key is now available.
1567 if (err
== 0 && db
->db_buf
!= NULL
&&
1568 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1569 (arc_is_encrypted(db
->db_buf
) ||
1570 arc_is_unauthenticated(db
->db_buf
) ||
1571 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1572 zbookmark_phys_t zb
;
1574 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1575 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1576 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1577 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1578 dbuf_set_data(db
, db
->db_buf
);
1580 mutex_exit(&db
->db_mtx
);
1581 if (err
== 0 && prefetch
) {
1582 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1583 flags
& DB_RF_HAVESTRUCT
);
1586 DBUF_STAT_BUMP(hash_hits
);
1587 } else if (db
->db_state
== DB_UNCACHED
) {
1588 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1589 boolean_t need_wait
= B_FALSE
;
1591 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
1594 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1595 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1598 err
= dbuf_read_impl(db
, zio
, flags
, dblt
, FTAG
);
1600 * dbuf_read_impl has dropped db_mtx and our parent's rwlock
1603 if (!err
&& prefetch
) {
1604 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1605 flags
& DB_RF_HAVESTRUCT
);
1609 DBUF_STAT_BUMP(hash_misses
);
1612 * If we created a zio_root we must execute it to avoid
1613 * leaking it, even if it isn't attached to any work due
1614 * to an error in dbuf_read_impl().
1618 err
= zio_wait(zio
);
1620 VERIFY0(zio_wait(zio
));
1624 * Another reader came in while the dbuf was in flight
1625 * between UNCACHED and CACHED. Either a writer will finish
1626 * writing the buffer (sending the dbuf to CACHED) or the
1627 * first reader's request will reach the read_done callback
1628 * and send the dbuf to CACHED. Otherwise, a failure
1629 * occurred and the dbuf went to UNCACHED.
1631 mutex_exit(&db
->db_mtx
);
1633 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1634 flags
& DB_RF_HAVESTRUCT
);
1637 DBUF_STAT_BUMP(hash_misses
);
1639 /* Skip the wait per the caller's request. */
1640 mutex_enter(&db
->db_mtx
);
1641 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1642 while (db
->db_state
== DB_READ
||
1643 db
->db_state
== DB_FILL
) {
1644 ASSERT(db
->db_state
== DB_READ
||
1645 (flags
& DB_RF_HAVESTRUCT
) == 0);
1646 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1648 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1650 if (db
->db_state
== DB_UNCACHED
)
1651 err
= SET_ERROR(EIO
);
1653 mutex_exit(&db
->db_mtx
);
1660 dbuf_noread(dmu_buf_impl_t
*db
)
1662 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1663 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1664 mutex_enter(&db
->db_mtx
);
1665 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1666 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1667 if (db
->db_state
== DB_UNCACHED
) {
1668 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1669 spa_t
*spa
= db
->db_objset
->os_spa
;
1671 ASSERT(db
->db_buf
== NULL
);
1672 ASSERT(db
->db
.db_data
== NULL
);
1673 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1674 db
->db_state
= DB_FILL
;
1675 } else if (db
->db_state
== DB_NOFILL
) {
1676 dbuf_clear_data(db
);
1678 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1680 mutex_exit(&db
->db_mtx
);
1684 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1686 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1687 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1688 uint64_t txg
= dr
->dr_txg
;
1690 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1692 * This assert is valid because dmu_sync() expects to be called by
1693 * a zilog's get_data while holding a range lock. This call only
1694 * comes from dbuf_dirty() callers who must also hold a range lock.
1696 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1697 ASSERT(db
->db_level
== 0);
1699 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1700 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1703 ASSERT(db
->db_data_pending
!= dr
);
1705 /* free this block */
1706 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1707 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1709 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1710 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1711 dr
->dt
.dl
.dr_has_raw_params
= B_FALSE
;
1714 * Release the already-written buffer, so we leave it in
1715 * a consistent dirty state. Note that all callers are
1716 * modifying the buffer, so they will immediately do
1717 * another (redundant) arc_release(). Therefore, leave
1718 * the buf thawed to save the effort of freezing &
1719 * immediately re-thawing it.
1721 arc_release(dr
->dt
.dl
.dr_data
, db
);
1725 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1726 * data blocks in the free range, so that any future readers will find
1730 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1733 dmu_buf_impl_t
*db_search
;
1734 dmu_buf_impl_t
*db
, *db_next
;
1735 uint64_t txg
= tx
->tx_txg
;
1738 if (end_blkid
> dn
->dn_maxblkid
&&
1739 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1740 end_blkid
= dn
->dn_maxblkid
;
1741 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1743 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1744 db_search
->db_level
= 0;
1745 db_search
->db_blkid
= start_blkid
;
1746 db_search
->db_state
= DB_SEARCH
;
1748 mutex_enter(&dn
->dn_dbufs_mtx
);
1749 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1750 ASSERT3P(db
, ==, NULL
);
1752 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1754 for (; db
!= NULL
; db
= db_next
) {
1755 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1756 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1758 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1761 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1763 /* found a level 0 buffer in the range */
1764 mutex_enter(&db
->db_mtx
);
1765 if (dbuf_undirty(db
, tx
)) {
1766 /* mutex has been dropped and dbuf destroyed */
1770 if (db
->db_state
== DB_UNCACHED
||
1771 db
->db_state
== DB_NOFILL
||
1772 db
->db_state
== DB_EVICTING
) {
1773 ASSERT(db
->db
.db_data
== NULL
);
1774 mutex_exit(&db
->db_mtx
);
1777 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1778 /* will be handled in dbuf_read_done or dbuf_rele */
1779 db
->db_freed_in_flight
= TRUE
;
1780 mutex_exit(&db
->db_mtx
);
1783 if (zfs_refcount_count(&db
->db_holds
) == 0) {
1788 /* The dbuf is referenced */
1790 if (!list_is_empty(&db
->db_dirty_records
)) {
1791 dbuf_dirty_record_t
*dr
;
1793 dr
= list_head(&db
->db_dirty_records
);
1794 if (dr
->dr_txg
== txg
) {
1796 * This buffer is "in-use", re-adjust the file
1797 * size to reflect that this buffer may
1798 * contain new data when we sync.
1800 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1801 db
->db_blkid
> dn
->dn_maxblkid
)
1802 dn
->dn_maxblkid
= db
->db_blkid
;
1803 dbuf_unoverride(dr
);
1806 * This dbuf is not dirty in the open context.
1807 * Either uncache it (if its not referenced in
1808 * the open context) or reset its contents to
1811 dbuf_fix_old_data(db
, txg
);
1814 /* clear the contents if its cached */
1815 if (db
->db_state
== DB_CACHED
) {
1816 ASSERT(db
->db
.db_data
!= NULL
);
1817 arc_release(db
->db_buf
, db
);
1818 rw_enter(&db
->db_rwlock
, RW_WRITER
);
1819 bzero(db
->db
.db_data
, db
->db
.db_size
);
1820 rw_exit(&db
->db_rwlock
);
1821 arc_buf_freeze(db
->db_buf
);
1824 mutex_exit(&db
->db_mtx
);
1827 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1828 mutex_exit(&dn
->dn_dbufs_mtx
);
1832 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1834 arc_buf_t
*buf
, *obuf
;
1835 dbuf_dirty_record_t
*dr
;
1836 int osize
= db
->db
.db_size
;
1837 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1840 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1846 * XXX we should be doing a dbuf_read, checking the return
1847 * value and returning that up to our callers
1849 dmu_buf_will_dirty(&db
->db
, tx
);
1851 /* create the data buffer for the new block */
1852 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1854 /* copy old block data to the new block */
1856 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1857 /* zero the remainder */
1859 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1861 mutex_enter(&db
->db_mtx
);
1862 dbuf_set_data(db
, buf
);
1863 arc_buf_destroy(obuf
, db
);
1864 db
->db
.db_size
= size
;
1866 dr
= list_head(&db
->db_dirty_records
);
1867 if (db
->db_level
== 0)
1868 dr
->dt
.dl
.dr_data
= buf
;
1869 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
1870 ASSERT3U(dr
->dr_accounted
, ==, osize
);
1871 dr
->dr_accounted
= size
;
1872 mutex_exit(&db
->db_mtx
);
1874 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1879 dbuf_release_bp(dmu_buf_impl_t
*db
)
1881 objset_t
*os __maybe_unused
= db
->db_objset
;
1883 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1884 ASSERT(arc_released(os
->os_phys_buf
) ||
1885 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1886 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1888 (void) arc_release(db
->db_buf
, db
);
1892 * We already have a dirty record for this TXG, and we are being
1896 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1898 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1900 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1902 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1904 * If this buffer has already been written out,
1905 * we now need to reset its state.
1907 dbuf_unoverride(dr
);
1908 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1909 db
->db_state
!= DB_NOFILL
) {
1910 /* Already released on initial dirty, so just thaw. */
1911 ASSERT(arc_released(db
->db_buf
));
1912 arc_buf_thaw(db
->db_buf
);
1917 dbuf_dirty_record_t
*
1918 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1922 dbuf_dirty_record_t
*dr
, *dr_next
, *dr_head
;
1923 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1924 boolean_t drop_struct_rwlock
= B_FALSE
;
1926 ASSERT(tx
->tx_txg
!= 0);
1927 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1928 DMU_TX_DIRTY_BUF(tx
, db
);
1933 * Shouldn't dirty a regular buffer in syncing context. Private
1934 * objects may be dirtied in syncing context, but only if they
1935 * were already pre-dirtied in open context.
1938 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1939 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1942 ASSERT(!dmu_tx_is_syncing(tx
) ||
1943 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1944 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1945 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1946 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1947 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1950 * We make this assert for private objects as well, but after we
1951 * check if we're already dirty. They are allowed to re-dirty
1952 * in syncing context.
1954 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1955 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1956 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1958 mutex_enter(&db
->db_mtx
);
1960 * XXX make this true for indirects too? The problem is that
1961 * transactions created with dmu_tx_create_assigned() from
1962 * syncing context don't bother holding ahead.
1964 ASSERT(db
->db_level
!= 0 ||
1965 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1966 db
->db_state
== DB_NOFILL
);
1968 mutex_enter(&dn
->dn_mtx
);
1970 * Don't set dirtyctx to SYNC if we're just modifying this as we
1971 * initialize the objset.
1973 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1974 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1975 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1978 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1979 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1980 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1981 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1982 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1984 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1985 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1990 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
1991 dn
->dn_dirty_txg
= tx
->tx_txg
;
1992 mutex_exit(&dn
->dn_mtx
);
1994 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1995 dn
->dn_have_spill
= B_TRUE
;
1998 * If this buffer is already dirty, we're done.
2000 dr_head
= list_head(&db
->db_dirty_records
);
2001 ASSERT(dr_head
== NULL
|| dr_head
->dr_txg
<= tx
->tx_txg
||
2002 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
2003 dr_next
= dbuf_find_dirty_lte(db
, tx
->tx_txg
);
2004 if (dr_next
&& dr_next
->dr_txg
== tx
->tx_txg
) {
2007 dbuf_redirty(dr_next
);
2008 mutex_exit(&db
->db_mtx
);
2013 * Only valid if not already dirty.
2015 ASSERT(dn
->dn_object
== 0 ||
2016 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2017 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2019 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
2022 * We should only be dirtying in syncing context if it's the
2023 * mos or we're initializing the os or it's a special object.
2024 * However, we are allowed to dirty in syncing context provided
2025 * we already dirtied it in open context. Hence we must make
2026 * this assertion only if we're not already dirty.
2029 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
2031 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2032 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
2033 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2034 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
2035 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2036 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2038 ASSERT(db
->db
.db_size
!= 0);
2040 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2042 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2043 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
2047 * If this buffer is dirty in an old transaction group we need
2048 * to make a copy of it so that the changes we make in this
2049 * transaction group won't leak out when we sync the older txg.
2051 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
2052 list_link_init(&dr
->dr_dirty_node
);
2053 list_link_init(&dr
->dr_dbuf_node
);
2054 if (db
->db_level
== 0) {
2055 void *data_old
= db
->db_buf
;
2057 if (db
->db_state
!= DB_NOFILL
) {
2058 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2059 dbuf_fix_old_data(db
, tx
->tx_txg
);
2060 data_old
= db
->db
.db_data
;
2061 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
2063 * Release the data buffer from the cache so
2064 * that we can modify it without impacting
2065 * possible other users of this cached data
2066 * block. Note that indirect blocks and
2067 * private objects are not released until the
2068 * syncing state (since they are only modified
2071 arc_release(db
->db_buf
, db
);
2072 dbuf_fix_old_data(db
, tx
->tx_txg
);
2073 data_old
= db
->db_buf
;
2075 ASSERT(data_old
!= NULL
);
2077 dr
->dt
.dl
.dr_data
= data_old
;
2079 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
2080 list_create(&dr
->dt
.di
.dr_children
,
2081 sizeof (dbuf_dirty_record_t
),
2082 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
2084 if (db
->db_blkid
!= DMU_BONUS_BLKID
)
2085 dr
->dr_accounted
= db
->db
.db_size
;
2087 dr
->dr_txg
= tx
->tx_txg
;
2088 list_insert_before(&db
->db_dirty_records
, dr_next
, dr
);
2091 * We could have been freed_in_flight between the dbuf_noread
2092 * and dbuf_dirty. We win, as though the dbuf_noread() had
2093 * happened after the free.
2095 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2096 db
->db_blkid
!= DMU_SPILL_BLKID
) {
2097 mutex_enter(&dn
->dn_mtx
);
2098 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2099 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
2102 mutex_exit(&dn
->dn_mtx
);
2103 db
->db_freed_in_flight
= FALSE
;
2107 * This buffer is now part of this txg
2109 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
2110 db
->db_dirtycnt
+= 1;
2111 ASSERT3U(db
->db_dirtycnt
, <=, 3);
2113 mutex_exit(&db
->db_mtx
);
2115 if (db
->db_blkid
== DMU_BONUS_BLKID
||
2116 db
->db_blkid
== DMU_SPILL_BLKID
) {
2117 mutex_enter(&dn
->dn_mtx
);
2118 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2119 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2120 mutex_exit(&dn
->dn_mtx
);
2121 dnode_setdirty(dn
, tx
);
2126 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
2127 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2128 drop_struct_rwlock
= B_TRUE
;
2132 * If we are overwriting a dedup BP, then unless it is snapshotted,
2133 * when we get to syncing context we will need to decrement its
2134 * refcount in the DDT. Prefetch the relevant DDT block so that
2135 * syncing context won't have to wait for the i/o.
2137 if (db
->db_blkptr
!= NULL
) {
2138 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
2139 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
2140 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
2144 * We need to hold the dn_struct_rwlock to make this assertion,
2145 * because it protects dn_phys / dn_next_nlevels from changing.
2147 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
2148 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
2149 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
2150 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
2151 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
2154 if (db
->db_level
== 0) {
2155 ASSERT(!db
->db_objset
->os_raw_receive
||
2156 dn
->dn_maxblkid
>= db
->db_blkid
);
2157 dnode_new_blkid(dn
, db
->db_blkid
, tx
,
2158 drop_struct_rwlock
, B_FALSE
);
2159 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
2162 if (db
->db_level
+1 < dn
->dn_nlevels
) {
2163 dmu_buf_impl_t
*parent
= db
->db_parent
;
2164 dbuf_dirty_record_t
*di
;
2165 int parent_held
= FALSE
;
2167 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
2168 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2169 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2170 db
->db_blkid
>> epbs
, FTAG
);
2171 ASSERT(parent
!= NULL
);
2174 if (drop_struct_rwlock
)
2175 rw_exit(&dn
->dn_struct_rwlock
);
2176 ASSERT3U(db
->db_level
+ 1, ==, parent
->db_level
);
2177 di
= dbuf_dirty(parent
, tx
);
2179 dbuf_rele(parent
, FTAG
);
2181 mutex_enter(&db
->db_mtx
);
2183 * Since we've dropped the mutex, it's possible that
2184 * dbuf_undirty() might have changed this out from under us.
2186 if (list_head(&db
->db_dirty_records
) == dr
||
2187 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2188 mutex_enter(&di
->dt
.di
.dr_mtx
);
2189 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2190 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2191 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2192 mutex_exit(&di
->dt
.di
.dr_mtx
);
2195 mutex_exit(&db
->db_mtx
);
2197 ASSERT(db
->db_level
+ 1 == dn
->dn_nlevels
);
2198 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2199 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2200 mutex_enter(&dn
->dn_mtx
);
2201 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2202 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2203 mutex_exit(&dn
->dn_mtx
);
2204 if (drop_struct_rwlock
)
2205 rw_exit(&dn
->dn_struct_rwlock
);
2208 dnode_setdirty(dn
, tx
);
2214 * Undirty a buffer in the transaction group referenced by the given
2215 * transaction. Return whether this evicted the dbuf.
2218 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2221 uint64_t txg
= tx
->tx_txg
;
2222 dbuf_dirty_record_t
*dr
;
2227 * Due to our use of dn_nlevels below, this can only be called
2228 * in open context, unless we are operating on the MOS.
2229 * From syncing context, dn_nlevels may be different from the
2230 * dn_nlevels used when dbuf was dirtied.
2232 ASSERT(db
->db_objset
==
2233 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2234 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2235 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2236 ASSERT0(db
->db_level
);
2237 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2240 * If this buffer is not dirty, we're done.
2242 dr
= dbuf_find_dirty_eq(db
, txg
);
2245 ASSERT(dr
->dr_dbuf
== db
);
2250 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2252 ASSERT(db
->db
.db_size
!= 0);
2254 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2255 dr
->dr_accounted
, txg
);
2257 list_remove(&db
->db_dirty_records
, dr
);
2260 * Note that there are three places in dbuf_dirty()
2261 * where this dirty record may be put on a list.
2262 * Make sure to do a list_remove corresponding to
2263 * every one of those list_insert calls.
2265 if (dr
->dr_parent
) {
2266 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2267 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2268 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2269 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2270 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2271 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2272 mutex_enter(&dn
->dn_mtx
);
2273 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2274 mutex_exit(&dn
->dn_mtx
);
2278 if (db
->db_state
!= DB_NOFILL
) {
2279 dbuf_unoverride(dr
);
2281 ASSERT(db
->db_buf
!= NULL
);
2282 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2283 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2284 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2287 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2289 ASSERT(db
->db_dirtycnt
> 0);
2290 db
->db_dirtycnt
-= 1;
2292 if (zfs_refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2293 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
2302 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2304 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2306 ASSERT(tx
->tx_txg
!= 0);
2307 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2310 * Quick check for dirtiness. For already dirty blocks, this
2311 * reduces runtime of this function by >90%, and overall performance
2312 * by 50% for some workloads (e.g. file deletion with indirect blocks
2315 mutex_enter(&db
->db_mtx
);
2317 if (db
->db_state
== DB_CACHED
) {
2318 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2320 * It's possible that it is already dirty but not cached,
2321 * because there are some calls to dbuf_dirty() that don't
2322 * go through dmu_buf_will_dirty().
2325 /* This dbuf is already dirty and cached. */
2327 mutex_exit(&db
->db_mtx
);
2331 mutex_exit(&db
->db_mtx
);
2334 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2335 flags
|= DB_RF_HAVESTRUCT
;
2337 (void) dbuf_read(db
, NULL
, flags
);
2338 (void) dbuf_dirty(db
, tx
);
2342 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2344 dmu_buf_will_dirty_impl(db_fake
,
2345 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2349 dmu_buf_is_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2351 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2352 dbuf_dirty_record_t
*dr
;
2354 mutex_enter(&db
->db_mtx
);
2355 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2356 mutex_exit(&db
->db_mtx
);
2357 return (dr
!= NULL
);
2361 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2363 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2365 db
->db_state
= DB_NOFILL
;
2367 dmu_buf_will_fill(db_fake
, tx
);
2371 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2373 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2375 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2376 ASSERT(tx
->tx_txg
!= 0);
2377 ASSERT(db
->db_level
== 0);
2378 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2380 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2381 dmu_tx_private_ok(tx
));
2384 (void) dbuf_dirty(db
, tx
);
2388 * This function is effectively the same as dmu_buf_will_dirty(), but
2389 * indicates the caller expects raw encrypted data in the db, and provides
2390 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2391 * blkptr_t when this dbuf is written. This is only used for blocks of
2392 * dnodes, during raw receive.
2395 dmu_buf_set_crypt_params(dmu_buf_t
*db_fake
, boolean_t byteorder
,
2396 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
, dmu_tx_t
*tx
)
2398 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2399 dbuf_dirty_record_t
*dr
;
2402 * dr_has_raw_params is only processed for blocks of dnodes
2403 * (see dbuf_sync_dnode_leaf_crypt()).
2405 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
2406 ASSERT3U(db
->db_level
, ==, 0);
2407 ASSERT(db
->db_objset
->os_raw_receive
);
2409 dmu_buf_will_dirty_impl(db_fake
,
2410 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2412 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2414 ASSERT3P(dr
, !=, NULL
);
2416 dr
->dt
.dl
.dr_has_raw_params
= B_TRUE
;
2417 dr
->dt
.dl
.dr_byteorder
= byteorder
;
2418 bcopy(salt
, dr
->dt
.dl
.dr_salt
, ZIO_DATA_SALT_LEN
);
2419 bcopy(iv
, dr
->dt
.dl
.dr_iv
, ZIO_DATA_IV_LEN
);
2420 bcopy(mac
, dr
->dt
.dl
.dr_mac
, ZIO_DATA_MAC_LEN
);
2424 dbuf_override_impl(dmu_buf_impl_t
*db
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
2426 struct dirty_leaf
*dl
;
2427 dbuf_dirty_record_t
*dr
;
2429 dr
= list_head(&db
->db_dirty_records
);
2430 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2432 dl
->dr_overridden_by
= *bp
;
2433 dl
->dr_override_state
= DR_OVERRIDDEN
;
2434 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2439 dmu_buf_fill_done(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2441 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2442 mutex_enter(&db
->db_mtx
);
2445 if (db
->db_state
== DB_FILL
) {
2446 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2447 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2448 /* we were freed while filling */
2449 /* XXX dbuf_undirty? */
2450 bzero(db
->db
.db_data
, db
->db
.db_size
);
2451 db
->db_freed_in_flight
= FALSE
;
2453 db
->db_state
= DB_CACHED
;
2454 cv_broadcast(&db
->db_changed
);
2456 mutex_exit(&db
->db_mtx
);
2460 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2461 bp_embedded_type_t etype
, enum zio_compress comp
,
2462 int uncompressed_size
, int compressed_size
, int byteorder
,
2465 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2466 struct dirty_leaf
*dl
;
2467 dmu_object_type_t type
;
2468 dbuf_dirty_record_t
*dr
;
2470 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2471 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2472 SPA_FEATURE_EMBEDDED_DATA
));
2476 type
= DB_DNODE(db
)->dn_type
;
2479 ASSERT0(db
->db_level
);
2480 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2482 dmu_buf_will_not_fill(dbuf
, tx
);
2484 dr
= list_head(&db
->db_dirty_records
);
2485 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2487 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2488 data
, comp
, uncompressed_size
, compressed_size
);
2489 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2490 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2491 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2492 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2494 dl
->dr_override_state
= DR_OVERRIDDEN
;
2495 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2499 dmu_buf_redact(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2501 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2502 dmu_object_type_t type
;
2503 ASSERT(dsl_dataset_feature_is_active(db
->db_objset
->os_dsl_dataset
,
2504 SPA_FEATURE_REDACTED_DATASETS
));
2507 type
= DB_DNODE(db
)->dn_type
;
2510 ASSERT0(db
->db_level
);
2511 dmu_buf_will_not_fill(dbuf
, tx
);
2513 blkptr_t bp
= { { { {0} } } };
2514 BP_SET_TYPE(&bp
, type
);
2515 BP_SET_LEVEL(&bp
, 0);
2516 BP_SET_BIRTH(&bp
, tx
->tx_txg
, 0);
2517 BP_SET_REDACTED(&bp
);
2518 BPE_SET_LSIZE(&bp
, dbuf
->db_size
);
2520 dbuf_override_impl(db
, &bp
, tx
);
2524 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2525 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2528 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2530 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2531 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2532 ASSERT(db
->db_level
== 0);
2533 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2534 ASSERT(buf
!= NULL
);
2535 ASSERT3U(arc_buf_lsize(buf
), ==, db
->db
.db_size
);
2536 ASSERT(tx
->tx_txg
!= 0);
2538 arc_return_buf(buf
, db
);
2539 ASSERT(arc_released(buf
));
2541 mutex_enter(&db
->db_mtx
);
2543 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2544 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2546 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2548 if (db
->db_state
== DB_CACHED
&&
2549 zfs_refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2551 * In practice, we will never have a case where we have an
2552 * encrypted arc buffer while additional holds exist on the
2553 * dbuf. We don't handle this here so we simply assert that
2556 ASSERT(!arc_is_encrypted(buf
));
2557 mutex_exit(&db
->db_mtx
);
2558 (void) dbuf_dirty(db
, tx
);
2559 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2560 arc_buf_destroy(buf
, db
);
2561 xuio_stat_wbuf_copied();
2565 xuio_stat_wbuf_nocopy();
2566 if (db
->db_state
== DB_CACHED
) {
2567 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
2569 ASSERT(db
->db_buf
!= NULL
);
2570 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2571 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2573 if (!arc_released(db
->db_buf
)) {
2574 ASSERT(dr
->dt
.dl
.dr_override_state
==
2576 arc_release(db
->db_buf
, db
);
2578 dr
->dt
.dl
.dr_data
= buf
;
2579 arc_buf_destroy(db
->db_buf
, db
);
2580 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2581 arc_release(db
->db_buf
, db
);
2582 arc_buf_destroy(db
->db_buf
, db
);
2586 ASSERT(db
->db_buf
== NULL
);
2587 dbuf_set_data(db
, buf
);
2588 db
->db_state
= DB_FILL
;
2589 mutex_exit(&db
->db_mtx
);
2590 (void) dbuf_dirty(db
, tx
);
2591 dmu_buf_fill_done(&db
->db
, tx
);
2595 dbuf_destroy(dmu_buf_impl_t
*db
)
2598 dmu_buf_impl_t
*parent
= db
->db_parent
;
2599 dmu_buf_impl_t
*dndb
;
2601 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2602 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
2604 if (db
->db_buf
!= NULL
) {
2605 arc_buf_destroy(db
->db_buf
, db
);
2609 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2610 int slots
= DB_DNODE(db
)->dn_num_slots
;
2611 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2612 if (db
->db
.db_data
!= NULL
) {
2613 kmem_free(db
->db
.db_data
, bonuslen
);
2614 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2615 db
->db_state
= DB_UNCACHED
;
2619 dbuf_clear_data(db
);
2621 if (multilist_link_active(&db
->db_cache_link
)) {
2622 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
2623 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
2625 multilist_remove(dbuf_caches
[db
->db_caching_status
].cache
, db
);
2626 (void) zfs_refcount_remove_many(
2627 &dbuf_caches
[db
->db_caching_status
].size
,
2628 db
->db
.db_size
, db
);
2630 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
2631 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
2633 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
2634 DBUF_STAT_BUMPDOWN(cache_count
);
2635 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
2638 db
->db_caching_status
= DB_NO_CACHE
;
2641 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2642 ASSERT(db
->db_data_pending
== NULL
);
2644 db
->db_state
= DB_EVICTING
;
2645 db
->db_blkptr
= NULL
;
2648 * Now that db_state is DB_EVICTING, nobody else can find this via
2649 * the hash table. We can now drop db_mtx, which allows us to
2650 * acquire the dn_dbufs_mtx.
2652 mutex_exit(&db
->db_mtx
);
2657 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2658 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2660 mutex_enter_nested(&dn
->dn_dbufs_mtx
,
2662 avl_remove(&dn
->dn_dbufs
, db
);
2663 atomic_dec_32(&dn
->dn_dbufs_count
);
2667 mutex_exit(&dn
->dn_dbufs_mtx
);
2669 * Decrementing the dbuf count means that the hold corresponding
2670 * to the removed dbuf is no longer discounted in dnode_move(),
2671 * so the dnode cannot be moved until after we release the hold.
2672 * The membar_producer() ensures visibility of the decremented
2673 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2676 mutex_enter(&dn
->dn_mtx
);
2677 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
2678 db
->db_dnode_handle
= NULL
;
2680 dbuf_hash_remove(db
);
2685 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
2687 db
->db_parent
= NULL
;
2689 ASSERT(db
->db_buf
== NULL
);
2690 ASSERT(db
->db
.db_data
== NULL
);
2691 ASSERT(db
->db_hash_next
== NULL
);
2692 ASSERT(db
->db_blkptr
== NULL
);
2693 ASSERT(db
->db_data_pending
== NULL
);
2694 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
2695 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2697 kmem_cache_free(dbuf_kmem_cache
, db
);
2698 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2701 * If this dbuf is referenced from an indirect dbuf,
2702 * decrement the ref count on the indirect dbuf.
2704 if (parent
&& parent
!= dndb
) {
2705 mutex_enter(&parent
->db_mtx
);
2706 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
2711 * Note: While bpp will always be updated if the function returns success,
2712 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2713 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2716 __attribute__((always_inline
))
2718 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2719 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
2724 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2726 if (blkid
== DMU_SPILL_BLKID
) {
2727 mutex_enter(&dn
->dn_mtx
);
2728 if (dn
->dn_have_spill
&&
2729 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2730 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2733 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2734 *parentp
= dn
->dn_dbuf
;
2735 mutex_exit(&dn
->dn_mtx
);
2740 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2741 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2743 ASSERT3U(level
* epbs
, <, 64);
2744 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2746 * This assertion shouldn't trip as long as the max indirect block size
2747 * is less than 1M. The reason for this is that up to that point,
2748 * the number of levels required to address an entire object with blocks
2749 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2750 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2751 * (i.e. we can address the entire object), objects will all use at most
2752 * N-1 levels and the assertion won't overflow. However, once epbs is
2753 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2754 * enough to address an entire object, so objects will have 5 levels,
2755 * but then this assertion will overflow.
2757 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2758 * need to redo this logic to handle overflows.
2760 ASSERT(level
>= nlevels
||
2761 ((nlevels
- level
- 1) * epbs
) +
2762 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2763 if (level
>= nlevels
||
2764 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2765 ((nlevels
- level
- 1) * epbs
)) ||
2767 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2768 /* the buffer has no parent yet */
2769 return (SET_ERROR(ENOENT
));
2770 } else if (level
< nlevels
-1) {
2771 /* this block is referenced from an indirect block */
2774 err
= dbuf_hold_impl(dn
, level
+ 1,
2775 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2779 err
= dbuf_read(*parentp
, NULL
,
2780 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2782 dbuf_rele(*parentp
, NULL
);
2786 rw_enter(&(*parentp
)->db_rwlock
, RW_READER
);
2787 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2788 (blkid
& ((1ULL << epbs
) - 1));
2789 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2790 ASSERT(BP_IS_HOLE(*bpp
));
2791 rw_exit(&(*parentp
)->db_rwlock
);
2794 /* the block is referenced from the dnode */
2795 ASSERT3U(level
, ==, nlevels
-1);
2796 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2797 blkid
< dn
->dn_phys
->dn_nblkptr
);
2799 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2800 *parentp
= dn
->dn_dbuf
;
2802 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2807 static dmu_buf_impl_t
*
2808 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2809 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2811 objset_t
*os
= dn
->dn_objset
;
2812 dmu_buf_impl_t
*db
, *odb
;
2814 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2815 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2817 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2819 list_create(&db
->db_dirty_records
, sizeof (dbuf_dirty_record_t
),
2820 offsetof(dbuf_dirty_record_t
, dr_dbuf_node
));
2823 db
->db
.db_object
= dn
->dn_object
;
2824 db
->db_level
= level
;
2825 db
->db_blkid
= blkid
;
2826 db
->db_dirtycnt
= 0;
2827 db
->db_dnode_handle
= dn
->dn_handle
;
2828 db
->db_parent
= parent
;
2829 db
->db_blkptr
= blkptr
;
2832 db
->db_user_immediate_evict
= FALSE
;
2833 db
->db_freed_in_flight
= FALSE
;
2834 db
->db_pending_evict
= FALSE
;
2836 if (blkid
== DMU_BONUS_BLKID
) {
2837 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2838 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2839 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2840 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2841 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2842 db
->db_state
= DB_UNCACHED
;
2843 db
->db_caching_status
= DB_NO_CACHE
;
2844 /* the bonus dbuf is not placed in the hash table */
2845 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2847 } else if (blkid
== DMU_SPILL_BLKID
) {
2848 db
->db
.db_size
= (blkptr
!= NULL
) ?
2849 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2850 db
->db
.db_offset
= 0;
2853 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2854 db
->db
.db_size
= blocksize
;
2855 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2859 * Hold the dn_dbufs_mtx while we get the new dbuf
2860 * in the hash table *and* added to the dbufs list.
2861 * This prevents a possible deadlock with someone
2862 * trying to look up this dbuf before it's added to the
2865 mutex_enter(&dn
->dn_dbufs_mtx
);
2866 db
->db_state
= DB_EVICTING
;
2867 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2868 /* someone else inserted it first */
2869 kmem_cache_free(dbuf_kmem_cache
, db
);
2870 mutex_exit(&dn
->dn_dbufs_mtx
);
2871 DBUF_STAT_BUMP(hash_insert_race
);
2874 avl_add(&dn
->dn_dbufs
, db
);
2876 db
->db_state
= DB_UNCACHED
;
2877 db
->db_caching_status
= DB_NO_CACHE
;
2878 mutex_exit(&dn
->dn_dbufs_mtx
);
2879 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2881 if (parent
&& parent
!= dn
->dn_dbuf
)
2882 dbuf_add_ref(parent
, db
);
2884 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2885 zfs_refcount_count(&dn
->dn_holds
) > 0);
2886 (void) zfs_refcount_add(&dn
->dn_holds
, db
);
2887 atomic_inc_32(&dn
->dn_dbufs_count
);
2889 dprintf_dbuf(db
, "db=%p\n", db
);
2895 * This function returns a block pointer and information about the object,
2896 * given a dnode and a block. This is a publicly accessible version of
2897 * dbuf_findbp that only returns some information, rather than the
2898 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
2899 * should be locked as (at least) a reader.
2902 dbuf_dnode_findbp(dnode_t
*dn
, uint64_t level
, uint64_t blkid
,
2903 blkptr_t
*bp
, uint16_t *datablkszsec
, uint8_t *indblkshift
)
2905 dmu_buf_impl_t
*dbp
= NULL
;
2908 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2910 err
= dbuf_findbp(dn
, level
, blkid
, B_FALSE
, &dbp
, &bp2
);
2914 dbuf_rele(dbp
, NULL
);
2915 if (datablkszsec
!= NULL
)
2916 *datablkszsec
= dn
->dn_phys
->dn_datablkszsec
;
2917 if (indblkshift
!= NULL
)
2918 *indblkshift
= dn
->dn_phys
->dn_indblkshift
;
2924 typedef struct dbuf_prefetch_arg
{
2925 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2926 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2927 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2928 int dpa_curlevel
; /* The current level that we're reading */
2929 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2930 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2931 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2932 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2933 } dbuf_prefetch_arg_t
;
2936 * Actually issue the prefetch read for the block given.
2939 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2941 ASSERT(!BP_IS_REDACTED(bp
) ||
2942 dsl_dataset_feature_is_active(
2943 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
2944 SPA_FEATURE_REDACTED_DATASETS
));
2946 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
) || BP_IS_REDACTED(bp
))
2949 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2950 arc_flags_t aflags
=
2951 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2953 /* dnodes are always read as raw and then converted later */
2954 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
2955 dpa
->dpa_curlevel
== 0)
2956 zio_flags
|= ZIO_FLAG_RAW
;
2958 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2959 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2960 ASSERT(dpa
->dpa_zio
!= NULL
);
2961 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2962 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
2966 * Called when an indirect block above our prefetch target is read in. This
2967 * will either read in the next indirect block down the tree or issue the actual
2968 * prefetch if the next block down is our target.
2971 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
2972 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
2974 dbuf_prefetch_arg_t
*dpa
= private;
2976 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2977 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2980 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
2981 kmem_free(dpa
, sizeof (*dpa
));
2984 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
2987 * The dpa_dnode is only valid if we are called with a NULL
2988 * zio. This indicates that the arc_read() returned without
2989 * first calling zio_read() to issue a physical read. Once
2990 * a physical read is made the dpa_dnode must be invalidated
2991 * as the locks guarding it may have been dropped. If the
2992 * dpa_dnode is still valid, then we want to add it to the dbuf
2993 * cache. To do so, we must hold the dbuf associated with the block
2994 * we just prefetched, read its contents so that we associate it
2995 * with an arc_buf_t, and then release it.
2998 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2999 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
3000 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
3002 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
3004 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
3006 dpa
->dpa_dnode
= NULL
;
3007 } else if (dpa
->dpa_dnode
!= NULL
) {
3008 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
3009 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
3010 dpa
->dpa_zb
.zb_level
));
3011 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
3012 dpa
->dpa_curlevel
, curblkid
, FTAG
);
3014 kmem_free(dpa
, sizeof (*dpa
));
3015 arc_buf_destroy(abuf
, private);
3019 (void) dbuf_read(db
, NULL
,
3020 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
3021 dbuf_rele(db
, FTAG
);
3024 dpa
->dpa_curlevel
--;
3025 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
3026 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
3027 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
3028 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
3030 ASSERT(!BP_IS_REDACTED(bp
) ||
3031 dsl_dataset_feature_is_active(
3032 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3033 SPA_FEATURE_REDACTED_DATASETS
));
3034 if (BP_IS_HOLE(bp
) || BP_IS_REDACTED(bp
)) {
3035 kmem_free(dpa
, sizeof (*dpa
));
3036 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
3037 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
3038 dbuf_issue_final_prefetch(dpa
, bp
);
3039 kmem_free(dpa
, sizeof (*dpa
));
3041 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3042 zbookmark_phys_t zb
;
3044 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3045 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
3046 iter_aflags
|= ARC_FLAG_L2CACHE
;
3048 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3050 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
3051 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
3053 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3054 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
3055 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3059 arc_buf_destroy(abuf
, private);
3063 * Issue prefetch reads for the given block on the given level. If the indirect
3064 * blocks above that block are not in memory, we will read them in
3065 * asynchronously. As a result, this call never blocks waiting for a read to
3066 * complete. Note that the prefetch might fail if the dataset is encrypted and
3067 * the encryption key is unmapped before the IO completes.
3070 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
3074 int epbs
, nlevels
, curlevel
;
3077 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3078 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3080 if (blkid
> dn
->dn_maxblkid
)
3083 if (level
== 0 && dnode_block_freed(dn
, blkid
))
3087 * This dnode hasn't been written to disk yet, so there's nothing to
3090 nlevels
= dn
->dn_phys
->dn_nlevels
;
3091 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
3094 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3095 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
3098 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
3101 mutex_exit(&db
->db_mtx
);
3103 * This dbuf already exists. It is either CACHED, or
3104 * (we assume) about to be read or filled.
3110 * Find the closest ancestor (indirect block) of the target block
3111 * that is present in the cache. In this indirect block, we will
3112 * find the bp that is at curlevel, curblkid.
3116 while (curlevel
< nlevels
- 1) {
3117 int parent_level
= curlevel
+ 1;
3118 uint64_t parent_blkid
= curblkid
>> epbs
;
3121 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
3122 FALSE
, TRUE
, FTAG
, &db
) == 0) {
3123 blkptr_t
*bpp
= db
->db_buf
->b_data
;
3124 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
3125 dbuf_rele(db
, FTAG
);
3129 curlevel
= parent_level
;
3130 curblkid
= parent_blkid
;
3133 if (curlevel
== nlevels
- 1) {
3134 /* No cached indirect blocks found. */
3135 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
3136 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
3138 ASSERT(!BP_IS_REDACTED(&bp
) ||
3139 dsl_dataset_feature_is_active(dn
->dn_objset
->os_dsl_dataset
,
3140 SPA_FEATURE_REDACTED_DATASETS
));
3141 if (BP_IS_HOLE(&bp
) || BP_IS_REDACTED(&bp
))
3144 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
3146 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
3149 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
3150 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
3151 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3152 dn
->dn_object
, level
, blkid
);
3153 dpa
->dpa_curlevel
= curlevel
;
3154 dpa
->dpa_prio
= prio
;
3155 dpa
->dpa_aflags
= aflags
;
3156 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
3157 dpa
->dpa_dnode
= dn
;
3158 dpa
->dpa_epbs
= epbs
;
3161 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3162 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
3163 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
3166 * If we have the indirect just above us, no need to do the asynchronous
3167 * prefetch chain; we'll just run the last step ourselves. If we're at
3168 * a higher level, though, we want to issue the prefetches for all the
3169 * indirect blocks asynchronously, so we can go on with whatever we were
3172 if (curlevel
== level
) {
3173 ASSERT3U(curblkid
, ==, blkid
);
3174 dbuf_issue_final_prefetch(dpa
, &bp
);
3175 kmem_free(dpa
, sizeof (*dpa
));
3177 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3178 zbookmark_phys_t zb
;
3180 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3181 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
3182 iter_aflags
|= ARC_FLAG_L2CACHE
;
3184 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3185 dn
->dn_object
, curlevel
, curblkid
);
3186 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3187 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
3188 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3192 * We use pio here instead of dpa_zio since it's possible that
3193 * dpa may have already been freed.
3199 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3200 * the case of encrypted, compressed and uncompressed buffers by
3201 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3202 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3204 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3206 noinline
static void
3207 dbuf_hold_copy(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3209 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3210 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
3211 enum zio_compress compress_type
= arc_get_compression(data
);
3213 if (arc_is_encrypted(data
)) {
3214 boolean_t byteorder
;
3215 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3216 uint8_t iv
[ZIO_DATA_IV_LEN
];
3217 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3219 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
3220 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
3221 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
3222 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
3224 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3225 dbuf_set_data(db
, arc_alloc_compressed_buf(
3226 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
3227 arc_buf_lsize(data
), compress_type
));
3229 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
3230 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
3233 rw_enter(&db
->db_rwlock
, RW_WRITER
);
3234 bcopy(data
->b_data
, db
->db
.db_data
, arc_buf_size(data
));
3235 rw_exit(&db
->db_rwlock
);
3239 * Returns with db_holds incremented, and db_mtx not held.
3240 * Note: dn_struct_rwlock must be held.
3243 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3244 boolean_t fail_sparse
, boolean_t fail_uncached
,
3245 void *tag
, dmu_buf_impl_t
**dbp
)
3247 dmu_buf_impl_t
*db
, *parent
= NULL
;
3249 /* If the pool has been created, verify the tx_sync_lock is not held */
3250 spa_t
*spa
= dn
->dn_objset
->os_spa
;
3251 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
3253 ASSERT(!MUTEX_HELD(&dp
->dp_tx
.tx_sync_lock
));
3256 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3257 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3258 ASSERT3U(dn
->dn_nlevels
, >, level
);
3262 /* dbuf_find() returns with db_mtx held */
3263 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
);
3266 blkptr_t
*bp
= NULL
;
3270 return (SET_ERROR(ENOENT
));
3272 ASSERT3P(parent
, ==, NULL
);
3273 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
3275 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
3276 err
= SET_ERROR(ENOENT
);
3279 dbuf_rele(parent
, NULL
);
3283 if (err
&& err
!= ENOENT
)
3285 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
);
3288 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
3289 mutex_exit(&db
->db_mtx
);
3290 return (SET_ERROR(ENOENT
));
3293 if (db
->db_buf
!= NULL
) {
3294 arc_buf_access(db
->db_buf
);
3295 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
3298 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
3301 * If this buffer is currently syncing out, and we are
3302 * still referencing it from db_data, we need to make a copy
3303 * of it in case we decide we want to dirty it again in this txg.
3305 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
3306 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3307 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
3308 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3309 if (dr
->dt
.dl
.dr_data
== db
->db_buf
)
3310 dbuf_hold_copy(dn
, db
);
3313 if (multilist_link_active(&db
->db_cache_link
)) {
3314 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3315 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3316 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3318 multilist_remove(dbuf_caches
[db
->db_caching_status
].cache
, db
);
3319 (void) zfs_refcount_remove_many(
3320 &dbuf_caches
[db
->db_caching_status
].size
,
3321 db
->db
.db_size
, db
);
3323 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3324 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3326 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3327 DBUF_STAT_BUMPDOWN(cache_count
);
3328 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3331 db
->db_caching_status
= DB_NO_CACHE
;
3333 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3335 mutex_exit(&db
->db_mtx
);
3337 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3339 dbuf_rele(parent
, NULL
);
3341 ASSERT3P(DB_DNODE(db
), ==, dn
);
3342 ASSERT3U(db
->db_blkid
, ==, blkid
);
3343 ASSERT3U(db
->db_level
, ==, level
);
3350 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
3352 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3356 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
3359 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3360 return (err
? NULL
: db
);
3364 dbuf_create_bonus(dnode_t
*dn
)
3366 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3368 ASSERT(dn
->dn_bonus
== NULL
);
3369 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
3373 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3375 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3377 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3378 return (SET_ERROR(ENOTSUP
));
3380 blksz
= SPA_MINBLOCKSIZE
;
3381 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3382 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3384 dbuf_new_size(db
, blksz
, tx
);
3390 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3392 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3395 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3397 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
3399 int64_t holds
= zfs_refcount_add(&db
->db_holds
, tag
);
3400 VERIFY3S(holds
, >, 1);
3403 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3405 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3408 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3409 dmu_buf_impl_t
*found_db
;
3410 boolean_t result
= B_FALSE
;
3412 if (blkid
== DMU_BONUS_BLKID
)
3413 found_db
= dbuf_find_bonus(os
, obj
);
3415 found_db
= dbuf_find(os
, obj
, 0, blkid
);
3417 if (found_db
!= NULL
) {
3418 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3419 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3422 mutex_exit(&found_db
->db_mtx
);
3428 * If you call dbuf_rele() you had better not be referencing the dnode handle
3429 * unless you have some other direct or indirect hold on the dnode. (An indirect
3430 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3431 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3432 * dnode's parent dbuf evicting its dnode handles.
3435 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
3437 mutex_enter(&db
->db_mtx
);
3438 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
3442 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
3444 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3448 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3449 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3450 * argument should be set if we are already in the dbuf-evicting code
3451 * path, in which case we don't want to recursively evict. This allows us to
3452 * avoid deeply nested stacks that would have a call flow similar to this:
3454 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3457 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3461 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
, boolean_t evicting
)
3466 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3470 * Remove the reference to the dbuf before removing its hold on the
3471 * dnode so we can guarantee in dnode_move() that a referenced bonus
3472 * buffer has a corresponding dnode hold.
3474 holds
= zfs_refcount_remove(&db
->db_holds
, tag
);
3478 * We can't freeze indirects if there is a possibility that they
3479 * may be modified in the current syncing context.
3481 if (db
->db_buf
!= NULL
&&
3482 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3483 arc_buf_freeze(db
->db_buf
);
3486 if (holds
== db
->db_dirtycnt
&&
3487 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3488 dbuf_evict_user(db
);
3491 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3493 boolean_t evict_dbuf
= db
->db_pending_evict
;
3496 * If the dnode moves here, we cannot cross this
3497 * barrier until the move completes.
3502 atomic_dec_32(&dn
->dn_dbufs_count
);
3505 * Decrementing the dbuf count means that the bonus
3506 * buffer's dnode hold is no longer discounted in
3507 * dnode_move(). The dnode cannot move until after
3508 * the dnode_rele() below.
3513 * Do not reference db after its lock is dropped.
3514 * Another thread may evict it.
3516 mutex_exit(&db
->db_mtx
);
3519 dnode_evict_bonus(dn
);
3522 } else if (db
->db_buf
== NULL
) {
3524 * This is a special case: we never associated this
3525 * dbuf with any data allocated from the ARC.
3527 ASSERT(db
->db_state
== DB_UNCACHED
||
3528 db
->db_state
== DB_NOFILL
);
3530 } else if (arc_released(db
->db_buf
)) {
3532 * This dbuf has anonymous data associated with it.
3536 boolean_t do_arc_evict
= B_FALSE
;
3538 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3540 if (!DBUF_IS_CACHEABLE(db
) &&
3541 db
->db_blkptr
!= NULL
&&
3542 !BP_IS_HOLE(db
->db_blkptr
) &&
3543 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
3544 do_arc_evict
= B_TRUE
;
3545 bp
= *db
->db_blkptr
;
3548 if (!DBUF_IS_CACHEABLE(db
) ||
3549 db
->db_pending_evict
) {
3551 } else if (!multilist_link_active(&db
->db_cache_link
)) {
3552 ASSERT3U(db
->db_caching_status
, ==,
3555 dbuf_cached_state_t dcs
=
3556 dbuf_include_in_metadata_cache(db
) ?
3557 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
3558 db
->db_caching_status
= dcs
;
3560 multilist_insert(dbuf_caches
[dcs
].cache
, db
);
3561 size
= zfs_refcount_add_many(
3562 &dbuf_caches
[dcs
].size
,
3563 db
->db
.db_size
, db
);
3565 if (dcs
== DB_DBUF_METADATA_CACHE
) {
3566 DBUF_STAT_BUMP(metadata_cache_count
);
3568 metadata_cache_size_bytes_max
,
3572 cache_levels
[db
->db_level
]);
3573 DBUF_STAT_BUMP(cache_count
);
3575 cache_levels_bytes
[db
->db_level
],
3577 DBUF_STAT_MAX(cache_size_bytes_max
,
3580 mutex_exit(&db
->db_mtx
);
3582 if (dcs
== DB_DBUF_CACHE
&& !evicting
)
3583 dbuf_evict_notify(size
);
3587 arc_freed(spa
, &bp
);
3590 mutex_exit(&db
->db_mtx
);
3595 #pragma weak dmu_buf_refcount = dbuf_refcount
3597 dbuf_refcount(dmu_buf_impl_t
*db
)
3599 return (zfs_refcount_count(&db
->db_holds
));
3603 dmu_buf_user_refcount(dmu_buf_t
*db_fake
)
3606 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3608 mutex_enter(&db
->db_mtx
);
3609 ASSERT3U(zfs_refcount_count(&db
->db_holds
), >=, db
->db_dirtycnt
);
3610 holds
= zfs_refcount_count(&db
->db_holds
) - db
->db_dirtycnt
;
3611 mutex_exit(&db
->db_mtx
);
3617 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3618 dmu_buf_user_t
*new_user
)
3620 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3622 mutex_enter(&db
->db_mtx
);
3623 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3624 if (db
->db_user
== old_user
)
3625 db
->db_user
= new_user
;
3627 old_user
= db
->db_user
;
3628 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3629 mutex_exit(&db
->db_mtx
);
3635 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3637 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3641 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3643 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3645 db
->db_user_immediate_evict
= TRUE
;
3646 return (dmu_buf_set_user(db_fake
, user
));
3650 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3652 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3656 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3658 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3660 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3661 return (db
->db_user
);
3665 dmu_buf_user_evict_wait()
3667 taskq_wait(dbu_evict_taskq
);
3671 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3673 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3674 return (dbi
->db_blkptr
);
3678 dmu_buf_get_objset(dmu_buf_t
*db
)
3680 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3681 return (dbi
->db_objset
);
3685 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3687 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3688 DB_DNODE_ENTER(dbi
);
3689 return (DB_DNODE(dbi
));
3693 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3695 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3700 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3702 /* ASSERT(dmu_tx_is_syncing(tx) */
3703 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3705 if (db
->db_blkptr
!= NULL
)
3708 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3709 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3710 BP_ZERO(db
->db_blkptr
);
3713 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3715 * This buffer was allocated at a time when there was
3716 * no available blkptrs from the dnode, or it was
3717 * inappropriate to hook it in (i.e., nlevels mismatch).
3719 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3720 ASSERT(db
->db_parent
== NULL
);
3721 db
->db_parent
= dn
->dn_dbuf
;
3722 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3725 dmu_buf_impl_t
*parent
= db
->db_parent
;
3726 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3728 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3729 if (parent
== NULL
) {
3730 mutex_exit(&db
->db_mtx
);
3731 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3732 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3733 db
->db_blkid
>> epbs
, db
);
3734 rw_exit(&dn
->dn_struct_rwlock
);
3735 mutex_enter(&db
->db_mtx
);
3736 db
->db_parent
= parent
;
3738 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3739 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3745 * When syncing out a blocks of dnodes, adjust the block to deal with
3746 * encryption. Normally, we make sure the block is decrypted before writing
3747 * it. If we have crypt params, then we are writing a raw (encrypted) block,
3748 * from a raw receive. In this case, set the ARC buf's crypt params so
3749 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
3752 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t
*dr
)
3755 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3757 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3758 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
3759 ASSERT3U(db
->db_level
, ==, 0);
3761 if (!db
->db_objset
->os_raw_receive
&& arc_is_encrypted(db
->db_buf
)) {
3762 zbookmark_phys_t zb
;
3765 * Unfortunately, there is currently no mechanism for
3766 * syncing context to handle decryption errors. An error
3767 * here is only possible if an attacker maliciously
3768 * changed a dnode block and updated the associated
3769 * checksums going up the block tree.
3771 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
3772 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3773 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
3776 panic("Invalid dnode block MAC");
3777 } else if (dr
->dt
.dl
.dr_has_raw_params
) {
3778 (void) arc_release(dr
->dt
.dl
.dr_data
, db
);
3779 arc_convert_to_raw(dr
->dt
.dl
.dr_data
,
3780 dmu_objset_id(db
->db_objset
),
3781 dr
->dt
.dl
.dr_byteorder
, DMU_OT_DNODE
,
3782 dr
->dt
.dl
.dr_salt
, dr
->dt
.dl
.dr_iv
, dr
->dt
.dl
.dr_mac
);
3787 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3788 * is critical the we not allow the compiler to inline this function in to
3789 * dbuf_sync_list() thereby drastically bloating the stack usage.
3791 noinline
static void
3792 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3794 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3798 ASSERT(dmu_tx_is_syncing(tx
));
3800 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3802 mutex_enter(&db
->db_mtx
);
3804 ASSERT(db
->db_level
> 0);
3807 /* Read the block if it hasn't been read yet. */
3808 if (db
->db_buf
== NULL
) {
3809 mutex_exit(&db
->db_mtx
);
3810 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3811 mutex_enter(&db
->db_mtx
);
3813 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3814 ASSERT(db
->db_buf
!= NULL
);
3818 /* Indirect block size must match what the dnode thinks it is. */
3819 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3820 dbuf_check_blkptr(dn
, db
);
3823 /* Provide the pending dirty record to child dbufs */
3824 db
->db_data_pending
= dr
;
3826 mutex_exit(&db
->db_mtx
);
3828 dbuf_write(dr
, db
->db_buf
, tx
);
3831 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3832 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3833 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3834 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3840 * Verify that the size of the data in our bonus buffer does not exceed
3841 * its recorded size.
3843 * The purpose of this verification is to catch any cases in development
3844 * where the size of a phys structure (i.e space_map_phys_t) grows and,
3845 * due to incorrect feature management, older pools expect to read more
3846 * data even though they didn't actually write it to begin with.
3848 * For a example, this would catch an error in the feature logic where we
3849 * open an older pool and we expect to write the space map histogram of
3850 * a space map with size SPACE_MAP_SIZE_V0.
3853 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
)
3855 dnode_t
*dn
= DB_DNODE(dr
->dr_dbuf
);
3858 * Encrypted bonus buffers can have data past their bonuslen.
3859 * Skip the verification of these blocks.
3861 if (DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))
3864 uint16_t bonuslen
= dn
->dn_phys
->dn_bonuslen
;
3865 uint16_t maxbonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
3866 ASSERT3U(bonuslen
, <=, maxbonuslen
);
3868 arc_buf_t
*datap
= dr
->dt
.dl
.dr_data
;
3869 char *datap_end
= ((char *)datap
) + bonuslen
;
3870 char *datap_max
= ((char *)datap
) + maxbonuslen
;
3872 /* ensure that everything is zero after our data */
3873 for (; datap_end
< datap_max
; datap_end
++)
3874 ASSERT(*datap_end
== 0);
3879 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3880 * critical the we not allow the compiler to inline this function in to
3881 * dbuf_sync_list() thereby drastically bloating the stack usage.
3883 noinline
static void
3884 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3886 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3887 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3890 uint64_t txg
= tx
->tx_txg
;
3892 ASSERT(dmu_tx_is_syncing(tx
));
3894 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3896 mutex_enter(&db
->db_mtx
);
3898 * To be synced, we must be dirtied. But we
3899 * might have been freed after the dirty.
3901 if (db
->db_state
== DB_UNCACHED
) {
3902 /* This buffer has been freed since it was dirtied */
3903 ASSERT(db
->db
.db_data
== NULL
);
3904 } else if (db
->db_state
== DB_FILL
) {
3905 /* This buffer was freed and is now being re-filled */
3906 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3908 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3915 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3916 mutex_enter(&dn
->dn_mtx
);
3917 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3919 * In the previous transaction group, the bonus buffer
3920 * was entirely used to store the attributes for the
3921 * dnode which overrode the dn_spill field. However,
3922 * when adding more attributes to the file a spill
3923 * block was required to hold the extra attributes.
3925 * Make sure to clear the garbage left in the dn_spill
3926 * field from the previous attributes in the bonus
3927 * buffer. Otherwise, after writing out the spill
3928 * block to the new allocated dva, it will free
3929 * the old block pointed to by the invalid dn_spill.
3931 db
->db_blkptr
= NULL
;
3933 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3934 mutex_exit(&dn
->dn_mtx
);
3938 * If this is a bonus buffer, simply copy the bonus data into the
3939 * dnode. It will be written out when the dnode is synced (and it
3940 * will be synced, since it must have been dirty for dbuf_sync to
3943 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3944 ASSERT(*datap
!= NULL
);
3945 ASSERT0(db
->db_level
);
3946 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
3947 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3948 bcopy(*datap
, DN_BONUS(dn
->dn_phys
),
3949 DN_MAX_BONUS_LEN(dn
->dn_phys
));
3953 dbuf_sync_leaf_verify_bonus_dnode(dr
);
3956 if (*datap
!= db
->db
.db_data
) {
3957 int slots
= DB_DNODE(db
)->dn_num_slots
;
3958 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3959 kmem_free(*datap
, bonuslen
);
3960 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3962 db
->db_data_pending
= NULL
;
3963 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
3964 ASSERT(dr
->dr_dbuf
== db
);
3965 list_remove(&db
->db_dirty_records
, dr
);
3966 if (dr
->dr_dbuf
->db_level
!= 0) {
3967 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3968 list_destroy(&dr
->dt
.di
.dr_children
);
3970 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3971 ASSERT(db
->db_dirtycnt
> 0);
3972 db
->db_dirtycnt
-= 1;
3973 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
, B_FALSE
);
3980 * This function may have dropped the db_mtx lock allowing a dmu_sync
3981 * operation to sneak in. As a result, we need to ensure that we
3982 * don't check the dr_override_state until we have returned from
3983 * dbuf_check_blkptr.
3985 dbuf_check_blkptr(dn
, db
);
3988 * If this buffer is in the middle of an immediate write,
3989 * wait for the synchronous IO to complete.
3991 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3992 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3993 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3994 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3998 * If this is a dnode block, ensure it is appropriately encrypted
3999 * or decrypted, depending on what we are writing to it this txg.
4001 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
4002 dbuf_prepare_encrypted_dnode_leaf(dr
);
4004 if (db
->db_state
!= DB_NOFILL
&&
4005 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
4006 zfs_refcount_count(&db
->db_holds
) > 1 &&
4007 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
4008 *datap
== db
->db_buf
) {
4010 * If this buffer is currently "in use" (i.e., there
4011 * are active holds and db_data still references it),
4012 * then make a copy before we start the write so that
4013 * any modifications from the open txg will not leak
4016 * NOTE: this copy does not need to be made for
4017 * objects only modified in the syncing context (e.g.
4018 * DNONE_DNODE blocks).
4020 int psize
= arc_buf_size(*datap
);
4021 int lsize
= arc_buf_lsize(*datap
);
4022 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
4023 enum zio_compress compress_type
= arc_get_compression(*datap
);
4025 if (arc_is_encrypted(*datap
)) {
4026 boolean_t byteorder
;
4027 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4028 uint8_t iv
[ZIO_DATA_IV_LEN
];
4029 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4031 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
4032 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
4033 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
4034 dn
->dn_type
, psize
, lsize
, compress_type
);
4035 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
4036 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
4037 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
4038 psize
, lsize
, compress_type
);
4040 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
4042 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
4044 db
->db_data_pending
= dr
;
4046 mutex_exit(&db
->db_mtx
);
4048 dbuf_write(dr
, *datap
, tx
);
4050 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4051 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
4052 list_insert_tail(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
4056 * Although zio_nowait() does not "wait for an IO", it does
4057 * initiate the IO. If this is an empty write it seems plausible
4058 * that the IO could actually be completed before the nowait
4059 * returns. We need to DB_DNODE_EXIT() first in case
4060 * zio_nowait() invalidates the dbuf.
4063 zio_nowait(dr
->dr_zio
);
4068 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
4070 dbuf_dirty_record_t
*dr
;
4072 while ((dr
= list_head(list
))) {
4073 if (dr
->dr_zio
!= NULL
) {
4075 * If we find an already initialized zio then we
4076 * are processing the meta-dnode, and we have finished.
4077 * The dbufs for all dnodes are put back on the list
4078 * during processing, so that we can zio_wait()
4079 * these IOs after initiating all child IOs.
4081 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
4082 DMU_META_DNODE_OBJECT
);
4085 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
4086 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
4087 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
4089 list_remove(list
, dr
);
4090 if (dr
->dr_dbuf
->db_level
> 0)
4091 dbuf_sync_indirect(dr
, tx
);
4093 dbuf_sync_leaf(dr
, tx
);
4099 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4101 dmu_buf_impl_t
*db
= vdb
;
4103 blkptr_t
*bp
= zio
->io_bp
;
4104 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4105 spa_t
*spa
= zio
->io_spa
;
4110 ASSERT3P(db
->db_blkptr
, !=, NULL
);
4111 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
4115 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
4116 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
4117 zio
->io_prev_space_delta
= delta
;
4119 if (bp
->blk_birth
!= 0) {
4120 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
4121 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
4122 (db
->db_blkid
== DMU_SPILL_BLKID
&&
4123 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
4124 BP_IS_EMBEDDED(bp
));
4125 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
4128 mutex_enter(&db
->db_mtx
);
4131 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4132 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4133 ASSERT(!(BP_IS_HOLE(bp
)) &&
4134 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4138 if (db
->db_level
== 0) {
4139 mutex_enter(&dn
->dn_mtx
);
4140 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
4141 db
->db_blkid
!= DMU_SPILL_BLKID
) {
4142 ASSERT0(db
->db_objset
->os_raw_receive
);
4143 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
4145 mutex_exit(&dn
->dn_mtx
);
4147 if (dn
->dn_type
== DMU_OT_DNODE
) {
4149 while (i
< db
->db
.db_size
) {
4151 (void *)(((char *)db
->db
.db_data
) + i
);
4153 i
+= DNODE_MIN_SIZE
;
4154 if (dnp
->dn_type
!= DMU_OT_NONE
) {
4156 i
+= dnp
->dn_extra_slots
*
4161 if (BP_IS_HOLE(bp
)) {
4168 blkptr_t
*ibp
= db
->db
.db_data
;
4169 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4170 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
4171 if (BP_IS_HOLE(ibp
))
4173 fill
+= BP_GET_FILL(ibp
);
4178 if (!BP_IS_EMBEDDED(bp
))
4179 BP_SET_FILL(bp
, fill
);
4181 mutex_exit(&db
->db_mtx
);
4183 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_WRITER
, FTAG
);
4184 *db
->db_blkptr
= *bp
;
4185 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
4190 * This function gets called just prior to running through the compression
4191 * stage of the zio pipeline. If we're an indirect block comprised of only
4192 * holes, then we want this indirect to be compressed away to a hole. In
4193 * order to do that we must zero out any information about the holes that
4194 * this indirect points to prior to before we try to compress it.
4197 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4199 dmu_buf_impl_t
*db
= vdb
;
4202 unsigned int epbs
, i
;
4204 ASSERT3U(db
->db_level
, >, 0);
4207 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4208 ASSERT3U(epbs
, <, 31);
4210 /* Determine if all our children are holes */
4211 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
4212 if (!BP_IS_HOLE(bp
))
4217 * If all the children are holes, then zero them all out so that
4218 * we may get compressed away.
4220 if (i
== 1ULL << epbs
) {
4222 * We only found holes. Grab the rwlock to prevent
4223 * anybody from reading the blocks we're about to
4226 rw_enter(&db
->db_rwlock
, RW_WRITER
);
4227 bzero(db
->db
.db_data
, db
->db
.db_size
);
4228 rw_exit(&db
->db_rwlock
);
4234 * The SPA will call this callback several times for each zio - once
4235 * for every physical child i/o (zio->io_phys_children times). This
4236 * allows the DMU to monitor the progress of each logical i/o. For example,
4237 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
4238 * block. There may be a long delay before all copies/fragments are completed,
4239 * so this callback allows us to retire dirty space gradually, as the physical
4244 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4246 dmu_buf_impl_t
*db
= arg
;
4247 objset_t
*os
= db
->db_objset
;
4248 dsl_pool_t
*dp
= dmu_objset_pool(os
);
4249 dbuf_dirty_record_t
*dr
;
4252 dr
= db
->db_data_pending
;
4253 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
4256 * The callback will be called io_phys_children times. Retire one
4257 * portion of our dirty space each time we are called. Any rounding
4258 * error will be cleaned up by dbuf_write_done().
4260 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
4261 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
4266 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4268 dmu_buf_impl_t
*db
= vdb
;
4269 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4270 blkptr_t
*bp
= db
->db_blkptr
;
4271 objset_t
*os
= db
->db_objset
;
4272 dmu_tx_t
*tx
= os
->os_synctx
;
4273 dbuf_dirty_record_t
*dr
;
4275 ASSERT0(zio
->io_error
);
4276 ASSERT(db
->db_blkptr
== bp
);
4279 * For nopwrites and rewrites we ensure that the bp matches our
4280 * original and bypass all the accounting.
4282 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4283 ASSERT(BP_EQUAL(bp
, bp_orig
));
4285 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4286 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
4287 dsl_dataset_block_born(ds
, bp
, tx
);
4290 mutex_enter(&db
->db_mtx
);
4294 dr
= db
->db_data_pending
;
4295 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4296 ASSERT(dr
->dr_dbuf
== db
);
4297 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
4298 list_remove(&db
->db_dirty_records
, dr
);
4301 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4306 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4307 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
4308 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4313 if (db
->db_level
== 0) {
4314 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
4315 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
4316 if (db
->db_state
!= DB_NOFILL
) {
4317 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
4318 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
4325 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4326 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
4327 if (!BP_IS_HOLE(db
->db_blkptr
)) {
4328 int epbs __maybe_unused
= dn
->dn_phys
->dn_indblkshift
-
4330 ASSERT3U(db
->db_blkid
, <=,
4331 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
4332 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4336 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4337 list_destroy(&dr
->dt
.di
.dr_children
);
4340 cv_broadcast(&db
->db_changed
);
4341 ASSERT(db
->db_dirtycnt
> 0);
4342 db
->db_dirtycnt
-= 1;
4343 db
->db_data_pending
= NULL
;
4344 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4347 * If we didn't do a physical write in this ZIO and we
4348 * still ended up here, it means that the space of the
4349 * dbuf that we just released (and undirtied) above hasn't
4350 * been marked as undirtied in the pool's accounting.
4352 * Thus, we undirty that space in the pool's view of the
4353 * world here. For physical writes this type of update
4354 * happens in dbuf_write_physdone().
4356 * If we did a physical write, cleanup any rounding errors
4357 * that came up due to writing multiple copies of a block
4358 * on disk [see dbuf_write_physdone()].
4360 if (zio
->io_phys_children
== 0) {
4361 dsl_pool_undirty_space(dmu_objset_pool(os
),
4362 dr
->dr_accounted
, zio
->io_txg
);
4364 dsl_pool_undirty_space(dmu_objset_pool(os
),
4365 dr
->dr_accounted
% zio
->io_phys_children
, zio
->io_txg
);
4368 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4372 dbuf_write_nofill_ready(zio_t
*zio
)
4374 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4378 dbuf_write_nofill_done(zio_t
*zio
)
4380 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4384 dbuf_write_override_ready(zio_t
*zio
)
4386 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4387 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4389 dbuf_write_ready(zio
, NULL
, db
);
4393 dbuf_write_override_done(zio_t
*zio
)
4395 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4396 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4397 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4399 mutex_enter(&db
->db_mtx
);
4400 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4401 if (!BP_IS_HOLE(obp
))
4402 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4403 arc_release(dr
->dt
.dl
.dr_data
, db
);
4405 mutex_exit(&db
->db_mtx
);
4407 dbuf_write_done(zio
, NULL
, db
);
4409 if (zio
->io_abd
!= NULL
)
4410 abd_put(zio
->io_abd
);
4413 typedef struct dbuf_remap_impl_callback_arg
{
4415 uint64_t drica_blk_birth
;
4417 } dbuf_remap_impl_callback_arg_t
;
4420 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
4423 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
4424 objset_t
*os
= drica
->drica_os
;
4425 spa_t
*spa
= dmu_objset_spa(os
);
4426 dmu_tx_t
*tx
= drica
->drica_tx
;
4428 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4430 if (os
== spa_meta_objset(spa
)) {
4431 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
4433 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
4434 size
, drica
->drica_blk_birth
, tx
);
4439 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, krwlock_t
*rw
, dmu_tx_t
*tx
)
4441 blkptr_t bp_copy
= *bp
;
4442 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4443 dbuf_remap_impl_callback_arg_t drica
;
4445 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4447 drica
.drica_os
= dn
->dn_objset
;
4448 drica
.drica_blk_birth
= bp
->blk_birth
;
4449 drica
.drica_tx
= tx
;
4450 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
4453 * If the blkptr being remapped is tracked by a livelist,
4454 * then we need to make sure the livelist reflects the update.
4455 * First, cancel out the old blkptr by appending a 'FREE'
4456 * entry. Next, add an 'ALLOC' to track the new version. This
4457 * way we avoid trying to free an inaccurate blkptr at delete.
4458 * Note that embedded blkptrs are not tracked in livelists.
4460 if (dn
->dn_objset
!= spa_meta_objset(spa
)) {
4461 dsl_dataset_t
*ds
= dmu_objset_ds(dn
->dn_objset
);
4462 if (dsl_deadlist_is_open(&ds
->ds_dir
->dd_livelist
) &&
4463 bp
->blk_birth
> ds
->ds_dir
->dd_origin_txg
) {
4464 ASSERT(!BP_IS_EMBEDDED(bp
));
4465 ASSERT(dsl_dir_is_clone(ds
->ds_dir
));
4466 ASSERT(spa_feature_is_enabled(spa
,
4467 SPA_FEATURE_LIVELIST
));
4468 bplist_append(&ds
->ds_dir
->dd_pending_frees
,
4470 bplist_append(&ds
->ds_dir
->dd_pending_allocs
,
4476 * The db_rwlock prevents dbuf_read_impl() from
4477 * dereferencing the BP while we are changing it. To
4478 * avoid lock contention, only grab it when we are actually
4482 rw_enter(rw
, RW_WRITER
);
4490 * Remap any existing BP's to concrete vdevs, if possible.
4493 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
4495 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
4496 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4498 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
4501 if (db
->db_level
> 0) {
4502 blkptr_t
*bp
= db
->db
.db_data
;
4503 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
4504 dbuf_remap_impl(dn
, &bp
[i
], &db
->db_rwlock
, tx
);
4506 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
4507 dnode_phys_t
*dnp
= db
->db
.db_data
;
4508 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
4510 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
;
4511 i
+= dnp
[i
].dn_extra_slots
+ 1) {
4512 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
4513 krwlock_t
*lock
= (dn
->dn_dbuf
== NULL
? NULL
:
4514 &dn
->dn_dbuf
->db_rwlock
);
4515 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], lock
,
4523 /* Issue I/O to commit a dirty buffer to disk. */
4525 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
4527 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4530 dmu_buf_impl_t
*parent
= db
->db_parent
;
4531 uint64_t txg
= tx
->tx_txg
;
4532 zbookmark_phys_t zb
;
4537 ASSERT(dmu_tx_is_syncing(tx
));
4543 if (db
->db_state
!= DB_NOFILL
) {
4544 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
4546 * Private object buffers are released here rather
4547 * than in dbuf_dirty() since they are only modified
4548 * in the syncing context and we don't want the
4549 * overhead of making multiple copies of the data.
4551 if (BP_IS_HOLE(db
->db_blkptr
)) {
4554 dbuf_release_bp(db
);
4556 dbuf_remap(dn
, db
, tx
);
4560 if (parent
!= dn
->dn_dbuf
) {
4561 /* Our parent is an indirect block. */
4562 /* We have a dirty parent that has been scheduled for write. */
4563 ASSERT(parent
&& parent
->db_data_pending
);
4564 /* Our parent's buffer is one level closer to the dnode. */
4565 ASSERT(db
->db_level
== parent
->db_level
-1);
4567 * We're about to modify our parent's db_data by modifying
4568 * our block pointer, so the parent must be released.
4570 ASSERT(arc_released(parent
->db_buf
));
4571 zio
= parent
->db_data_pending
->dr_zio
;
4573 /* Our parent is the dnode itself. */
4574 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
4575 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
4576 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
4577 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
4578 ASSERT3P(db
->db_blkptr
, ==,
4579 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
4583 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
4584 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
4587 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
4588 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
4589 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4591 if (db
->db_blkid
== DMU_SPILL_BLKID
)
4593 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
4595 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
4599 * We copy the blkptr now (rather than when we instantiate the dirty
4600 * record), because its value can change between open context and
4601 * syncing context. We do not need to hold dn_struct_rwlock to read
4602 * db_blkptr because we are in syncing context.
4604 dr
->dr_bp_copy
= *db
->db_blkptr
;
4606 if (db
->db_level
== 0 &&
4607 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
4609 * The BP for this block has been provided by open context
4610 * (by dmu_sync() or dmu_buf_write_embedded()).
4612 abd_t
*contents
= (data
!= NULL
) ?
4613 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
4615 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4616 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
4617 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
4618 dbuf_write_override_done
,
4619 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
4620 mutex_enter(&db
->db_mtx
);
4621 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
4622 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
4623 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
4624 mutex_exit(&db
->db_mtx
);
4625 } else if (db
->db_state
== DB_NOFILL
) {
4626 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
4627 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
4628 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4629 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
4630 dbuf_write_nofill_ready
, NULL
, NULL
,
4631 dbuf_write_nofill_done
, db
,
4632 ZIO_PRIORITY_ASYNC_WRITE
,
4633 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
4635 ASSERT(arc_released(data
));
4638 * For indirect blocks, we want to setup the children
4639 * ready callback so that we can properly handle an indirect
4640 * block that only contains holes.
4642 arc_write_done_func_t
*children_ready_cb
= NULL
;
4643 if (db
->db_level
!= 0)
4644 children_ready_cb
= dbuf_write_children_ready
;
4646 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
4647 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
4648 &zp
, dbuf_write_ready
,
4649 children_ready_cb
, dbuf_write_physdone
,
4650 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
4651 ZIO_FLAG_MUSTSUCCEED
, &zb
);
4655 EXPORT_SYMBOL(dbuf_find
);
4656 EXPORT_SYMBOL(dbuf_is_metadata
);
4657 EXPORT_SYMBOL(dbuf_destroy
);
4658 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
4659 EXPORT_SYMBOL(dbuf_whichblock
);
4660 EXPORT_SYMBOL(dbuf_read
);
4661 EXPORT_SYMBOL(dbuf_unoverride
);
4662 EXPORT_SYMBOL(dbuf_free_range
);
4663 EXPORT_SYMBOL(dbuf_new_size
);
4664 EXPORT_SYMBOL(dbuf_release_bp
);
4665 EXPORT_SYMBOL(dbuf_dirty
);
4666 EXPORT_SYMBOL(dmu_buf_set_crypt_params
);
4667 EXPORT_SYMBOL(dmu_buf_will_dirty
);
4668 EXPORT_SYMBOL(dmu_buf_is_dirty
);
4669 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
4670 EXPORT_SYMBOL(dmu_buf_will_fill
);
4671 EXPORT_SYMBOL(dmu_buf_fill_done
);
4672 EXPORT_SYMBOL(dmu_buf_rele
);
4673 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
4674 EXPORT_SYMBOL(dbuf_prefetch
);
4675 EXPORT_SYMBOL(dbuf_hold_impl
);
4676 EXPORT_SYMBOL(dbuf_hold
);
4677 EXPORT_SYMBOL(dbuf_hold_level
);
4678 EXPORT_SYMBOL(dbuf_create_bonus
);
4679 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
4680 EXPORT_SYMBOL(dbuf_rm_spill
);
4681 EXPORT_SYMBOL(dbuf_add_ref
);
4682 EXPORT_SYMBOL(dbuf_rele
);
4683 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
4684 EXPORT_SYMBOL(dbuf_refcount
);
4685 EXPORT_SYMBOL(dbuf_sync_list
);
4686 EXPORT_SYMBOL(dmu_buf_set_user
);
4687 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
4688 EXPORT_SYMBOL(dmu_buf_get_user
);
4689 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
4692 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, max_bytes
, ULONG
, ZMOD_RW
,
4693 "Maximum size in bytes of the dbuf cache.");
4695 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, hiwater_pct
, UINT
, ZMOD_RW
,
4696 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4699 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, lowater_pct
, UINT
, ZMOD_RW
,
4700 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4703 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_max_bytes
, ULONG
, ZMOD_RW
,
4704 "Maximum size in bytes of the dbuf metadata cache.");
4706 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, cache_shift
, INT
, ZMOD_RW
,
4707 "Set the size of the dbuf cache to a log2 fraction of arc size.");
4709 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_shift
, INT
, ZMOD_RW
,
4710 "Set the size of the dbuf metadata cache to a log2 fraction of arc "