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>
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);
153 static void dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t *dr);
154 static int dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags);
156 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
157 dmu_buf_evict_func_t *evict_func_sync,
158 dmu_buf_evict_func_t *evict_func_async,
159 dmu_buf_t **clear_on_evict_dbufp);
162 * Global data structures and functions for the dbuf cache.
164 static kmem_cache_t *dbuf_kmem_cache;
165 static taskq_t *dbu_evict_taskq;
167 static kthread_t *dbuf_cache_evict_thread;
168 static kmutex_t dbuf_evict_lock;
169 static kcondvar_t dbuf_evict_cv;
170 static boolean_t dbuf_evict_thread_exit;
173 * There are two dbuf caches; each dbuf can only be in one of them at a time.
175 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
176 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
177 * that represent the metadata that describes filesystems/snapshots/
178 * bookmarks/properties/etc. We only evict from this cache when we export a
179 * pool, to short-circuit as much I/O as possible for all administrative
180 * commands that need the metadata. There is no eviction policy for this
181 * cache, because we try to only include types in it which would occupy a
182 * very small amount of space per object but create a large impact on the
183 * performance of these commands. Instead, after it reaches a maximum size
184 * (which should only happen on very small memory systems with a very large
185 * number of filesystem objects), we stop taking new dbufs into the
186 * metadata cache, instead putting them in the normal dbuf cache.
188 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
189 * are not currently held but have been recently released. These dbufs
190 * are not eligible for arc eviction until they are aged out of the cache.
191 * Dbufs that are aged out of the cache will be immediately destroyed and
192 * become eligible for arc eviction.
194 * Dbufs are added to these caches once the last hold is released. If a dbuf is
195 * later accessed and still exists in the dbuf cache, then it will be removed
196 * from the cache and later re-added to the head of the cache.
198 * If a given dbuf meets the requirements for the metadata cache, it will go
199 * there, otherwise it will be considered for the generic LRU dbuf cache. The
200 * caches and the refcounts tracking their sizes are stored in an array indexed
201 * by those caches' matching enum values (from dbuf_cached_state_t).
203 typedef struct dbuf_cache {
207 dbuf_cache_t dbuf_caches[DB_CACHE_MAX];
209 /* Size limits for the caches */
210 unsigned long dbuf_cache_max_bytes = ULONG_MAX;
211 unsigned long dbuf_metadata_cache_max_bytes = ULONG_MAX;
213 /* Set the default sizes of the caches to log2 fraction of arc size */
214 int dbuf_cache_shift = 5;
215 int dbuf_metadata_cache_shift = 6;
217 static unsigned long dbuf_cache_target_bytes(void);
218 static unsigned long dbuf_metadata_cache_target_bytes(void);
221 * The LRU dbuf cache uses a three-stage eviction policy:
222 * - A low water marker designates when the dbuf eviction thread
223 * should stop evicting from the dbuf cache.
224 * - When we reach the maximum size (aka mid water mark), we
225 * signal the eviction thread to run.
226 * - The high water mark indicates when the eviction thread
227 * is unable to keep up with the incoming load and eviction must
228 * happen in the context of the calling thread.
232 * low water mid water hi water
233 * +----------------------------------------+----------+----------+
238 * +----------------------------------------+----------+----------+
240 * evicting eviction directly
243 * The high and low water marks indicate the operating range for the eviction
244 * thread. The low water mark is, by default, 90% of the total size of the
245 * cache and the high water mark is at 110% (both of these percentages can be
246 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
247 * respectively). The eviction thread will try to ensure that the cache remains
248 * within this range by waking up every second and checking if the cache is
249 * above the low water mark. The thread can also be woken up by callers adding
250 * elements into the cache if the cache is larger than the mid water (i.e max
251 * cache size). Once the eviction thread is woken up and eviction is required,
252 * it will continue evicting buffers until it's able to reduce the cache size
253 * to the low water mark. If the cache size continues to grow and hits the high
254 * water mark, then callers adding elements to the cache will begin to evict
255 * directly from the cache until the cache is no longer above the high water
260 * The percentage above and below the maximum cache size.
262 uint_t dbuf_cache_hiwater_pct = 10;
263 uint_t dbuf_cache_lowater_pct = 10;
267 dbuf_cons(void *vdb, void *unused, int kmflag)
269 dmu_buf_impl_t *db = vdb;
270 bzero(db, sizeof (dmu_buf_impl_t));
272 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
273 rw_init(&db->db_rwlock, NULL, RW_DEFAULT, NULL);
274 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
275 multilist_link_init(&db->db_cache_link);
276 zfs_refcount_create(&db->db_holds);
283 dbuf_dest(void *vdb, void *unused)
285 dmu_buf_impl_t *db = vdb;
286 mutex_destroy(&db->db_mtx);
287 rw_destroy(&db->db_rwlock);
288 cv_destroy(&db->db_changed);
289 ASSERT(!multilist_link_active(&db->db_cache_link));
290 zfs_refcount_destroy(&db->db_holds);
294 * dbuf hash table routines
296 static dbuf_hash_table_t dbuf_hash_table;
298 static uint64_t dbuf_hash_count;
301 * We use Cityhash for this. It's fast, and has good hash properties without
302 * requiring any large static buffers.
305 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
307 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid));
310 #define DTRACE_SET_STATE(db, why) \
311 DTRACE_PROBE2(dbuf__state_change, dmu_buf_impl_t *, db, \
314 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
315 ((dbuf)->db.db_object == (obj) && \
316 (dbuf)->db_objset == (os) && \
317 (dbuf)->db_level == (level) && \
318 (dbuf)->db_blkid == (blkid))
321 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
323 dbuf_hash_table_t *h = &dbuf_hash_table;
328 hv = dbuf_hash(os, obj, level, blkid);
329 idx = hv & h->hash_table_mask;
331 mutex_enter(DBUF_HASH_MUTEX(h, idx));
332 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
333 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
334 mutex_enter(&db->db_mtx);
335 if (db->db_state != DB_EVICTING) {
336 mutex_exit(DBUF_HASH_MUTEX(h, idx));
339 mutex_exit(&db->db_mtx);
342 mutex_exit(DBUF_HASH_MUTEX(h, idx));
346 static dmu_buf_impl_t *
347 dbuf_find_bonus(objset_t *os, uint64_t object)
350 dmu_buf_impl_t *db = NULL;
352 if (dnode_hold(os, object, FTAG, &dn) == 0) {
353 rw_enter(&dn->dn_struct_rwlock, RW_READER);
354 if (dn->dn_bonus != NULL) {
356 mutex_enter(&db->db_mtx);
358 rw_exit(&dn->dn_struct_rwlock);
359 dnode_rele(dn, FTAG);
365 * Insert an entry into the hash table. If there is already an element
366 * equal to elem in the hash table, then the already existing element
367 * will be returned and the new element will not be inserted.
368 * Otherwise returns NULL.
370 static dmu_buf_impl_t *
371 dbuf_hash_insert(dmu_buf_impl_t *db)
373 dbuf_hash_table_t *h = &dbuf_hash_table;
374 objset_t *os = db->db_objset;
375 uint64_t obj = db->db.db_object;
376 int level = db->db_level;
377 uint64_t blkid, hv, idx;
381 blkid = db->db_blkid;
382 hv = dbuf_hash(os, obj, level, blkid);
383 idx = hv & h->hash_table_mask;
385 mutex_enter(DBUF_HASH_MUTEX(h, idx));
386 for (dbf = h->hash_table[idx], i = 0; dbf != NULL;
387 dbf = dbf->db_hash_next, i++) {
388 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
389 mutex_enter(&dbf->db_mtx);
390 if (dbf->db_state != DB_EVICTING) {
391 mutex_exit(DBUF_HASH_MUTEX(h, idx));
394 mutex_exit(&dbf->db_mtx);
399 DBUF_STAT_BUMP(hash_collisions);
401 DBUF_STAT_BUMP(hash_chains);
403 DBUF_STAT_MAX(hash_chain_max, i);
406 mutex_enter(&db->db_mtx);
407 db->db_hash_next = h->hash_table[idx];
408 h->hash_table[idx] = db;
409 mutex_exit(DBUF_HASH_MUTEX(h, idx));
410 atomic_inc_64(&dbuf_hash_count);
411 DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count);
417 * This returns whether this dbuf should be stored in the metadata cache, which
418 * is based on whether it's from one of the dnode types that store data related
419 * to traversing dataset hierarchies.
422 dbuf_include_in_metadata_cache(dmu_buf_impl_t *db)
425 dmu_object_type_t type = DB_DNODE(db)->dn_type;
428 /* Check if this dbuf is one of the types we care about */
429 if (DMU_OT_IS_METADATA_CACHED(type)) {
430 /* If we hit this, then we set something up wrong in dmu_ot */
431 ASSERT(DMU_OT_IS_METADATA(type));
434 * Sanity check for small-memory systems: don't allocate too
435 * much memory for this purpose.
437 if (zfs_refcount_count(
438 &dbuf_caches[DB_DBUF_METADATA_CACHE].size) >
439 dbuf_metadata_cache_target_bytes()) {
440 DBUF_STAT_BUMP(metadata_cache_overflow);
451 * Remove an entry from the hash table. It must be in the EVICTING state.
454 dbuf_hash_remove(dmu_buf_impl_t *db)
456 dbuf_hash_table_t *h = &dbuf_hash_table;
458 dmu_buf_impl_t *dbf, **dbp;
460 hv = dbuf_hash(db->db_objset, db->db.db_object,
461 db->db_level, db->db_blkid);
462 idx = hv & h->hash_table_mask;
465 * We mustn't hold db_mtx to maintain lock ordering:
466 * DBUF_HASH_MUTEX > db_mtx.
468 ASSERT(zfs_refcount_is_zero(&db->db_holds));
469 ASSERT(db->db_state == DB_EVICTING);
470 ASSERT(!MUTEX_HELD(&db->db_mtx));
472 mutex_enter(DBUF_HASH_MUTEX(h, idx));
473 dbp = &h->hash_table[idx];
474 while ((dbf = *dbp) != db) {
475 dbp = &dbf->db_hash_next;
478 *dbp = db->db_hash_next;
479 db->db_hash_next = NULL;
480 if (h->hash_table[idx] &&
481 h->hash_table[idx]->db_hash_next == NULL)
482 DBUF_STAT_BUMPDOWN(hash_chains);
483 mutex_exit(DBUF_HASH_MUTEX(h, idx));
484 atomic_dec_64(&dbuf_hash_count);
490 } dbvu_verify_type_t;
493 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
498 if (db->db_user == NULL)
501 /* Only data blocks support the attachment of user data. */
502 ASSERT(db->db_level == 0);
504 /* Clients must resolve a dbuf before attaching user data. */
505 ASSERT(db->db.db_data != NULL);
506 ASSERT3U(db->db_state, ==, DB_CACHED);
508 holds = zfs_refcount_count(&db->db_holds);
509 if (verify_type == DBVU_EVICTING) {
511 * Immediate eviction occurs when holds == dirtycnt.
512 * For normal eviction buffers, holds is zero on
513 * eviction, except when dbuf_fix_old_data() calls
514 * dbuf_clear_data(). However, the hold count can grow
515 * during eviction even though db_mtx is held (see
516 * dmu_bonus_hold() for an example), so we can only
517 * test the generic invariant that holds >= dirtycnt.
519 ASSERT3U(holds, >=, db->db_dirtycnt);
521 if (db->db_user_immediate_evict == TRUE)
522 ASSERT3U(holds, >=, db->db_dirtycnt);
524 ASSERT3U(holds, >, 0);
530 dbuf_evict_user(dmu_buf_impl_t *db)
532 dmu_buf_user_t *dbu = db->db_user;
534 ASSERT(MUTEX_HELD(&db->db_mtx));
539 dbuf_verify_user(db, DBVU_EVICTING);
543 if (dbu->dbu_clear_on_evict_dbufp != NULL)
544 *dbu->dbu_clear_on_evict_dbufp = NULL;
548 * There are two eviction callbacks - one that we call synchronously
549 * and one that we invoke via a taskq. The async one is useful for
550 * avoiding lock order reversals and limiting stack depth.
552 * Note that if we have a sync callback but no async callback,
553 * it's likely that the sync callback will free the structure
554 * containing the dbu. In that case we need to take care to not
555 * dereference dbu after calling the sync evict func.
557 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
559 if (dbu->dbu_evict_func_sync != NULL)
560 dbu->dbu_evict_func_sync(dbu);
563 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
564 dbu, 0, &dbu->dbu_tqent);
569 dbuf_is_metadata(dmu_buf_impl_t *db)
572 * Consider indirect blocks and spill blocks to be meta data.
574 if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) {
577 boolean_t is_metadata;
580 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
583 return (is_metadata);
589 * This function *must* return indices evenly distributed between all
590 * sublists of the multilist. This is needed due to how the dbuf eviction
591 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
592 * distributed between all sublists and uses this assumption when
593 * deciding which sublist to evict from and how much to evict from it.
596 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
598 dmu_buf_impl_t *db = obj;
601 * The assumption here, is the hash value for a given
602 * dmu_buf_impl_t will remain constant throughout it's lifetime
603 * (i.e. it's objset, object, level and blkid fields don't change).
604 * Thus, we don't need to store the dbuf's sublist index
605 * on insertion, as this index can be recalculated on removal.
607 * Also, the low order bits of the hash value are thought to be
608 * distributed evenly. Otherwise, in the case that the multilist
609 * has a power of two number of sublists, each sublists' usage
610 * would not be evenly distributed.
612 return (dbuf_hash(db->db_objset, db->db.db_object,
613 db->db_level, db->db_blkid) %
614 multilist_get_num_sublists(ml));
618 * The target size of the dbuf cache can grow with the ARC target,
619 * unless limited by the tunable dbuf_cache_max_bytes.
621 static inline unsigned long
622 dbuf_cache_target_bytes(void)
624 return (MIN(dbuf_cache_max_bytes,
625 arc_target_bytes() >> dbuf_cache_shift));
629 * The target size of the dbuf metadata cache can grow with the ARC target,
630 * unless limited by the tunable dbuf_metadata_cache_max_bytes.
632 static inline unsigned long
633 dbuf_metadata_cache_target_bytes(void)
635 return (MIN(dbuf_metadata_cache_max_bytes,
636 arc_target_bytes() >> dbuf_metadata_cache_shift));
639 static inline uint64_t
640 dbuf_cache_hiwater_bytes(void)
642 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
643 return (dbuf_cache_target +
644 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100);
647 static inline uint64_t
648 dbuf_cache_lowater_bytes(void)
650 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
651 return (dbuf_cache_target -
652 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100);
655 static inline boolean_t
656 dbuf_cache_above_lowater(void)
658 return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
659 dbuf_cache_lowater_bytes());
663 * Evict the oldest eligible dbuf from the dbuf cache.
668 int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache);
669 multilist_sublist_t *mls = multilist_sublist_lock(
670 dbuf_caches[DB_DBUF_CACHE].cache, idx);
672 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
674 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
675 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
676 db = multilist_sublist_prev(mls, db);
679 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
680 multilist_sublist_t *, mls);
683 multilist_sublist_remove(mls, db);
684 multilist_sublist_unlock(mls);
685 (void) zfs_refcount_remove_many(
686 &dbuf_caches[DB_DBUF_CACHE].size, db->db.db_size, db);
687 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
688 DBUF_STAT_BUMPDOWN(cache_count);
689 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
691 ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE);
692 db->db_caching_status = DB_NO_CACHE;
694 DBUF_STAT_BUMP(cache_total_evicts);
696 multilist_sublist_unlock(mls);
701 * The dbuf evict thread is responsible for aging out dbufs from the
702 * cache. Once the cache has reached it's maximum size, dbufs are removed
703 * and destroyed. The eviction thread will continue running until the size
704 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
705 * out of the cache it is destroyed and becomes eligible for arc eviction.
709 dbuf_evict_thread(void *unused)
713 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
715 mutex_enter(&dbuf_evict_lock);
716 while (!dbuf_evict_thread_exit) {
717 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
718 CALLB_CPR_SAFE_BEGIN(&cpr);
719 (void) cv_timedwait_sig_hires(&dbuf_evict_cv,
720 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
721 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
723 mutex_exit(&dbuf_evict_lock);
726 * Keep evicting as long as we're above the low water mark
727 * for the cache. We do this without holding the locks to
728 * minimize lock contention.
730 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
734 mutex_enter(&dbuf_evict_lock);
737 dbuf_evict_thread_exit = B_FALSE;
738 cv_broadcast(&dbuf_evict_cv);
739 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
744 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
745 * If the dbuf cache is at its high water mark, then evict a dbuf from the
746 * dbuf cache using the callers context.
749 dbuf_evict_notify(uint64_t size)
752 * We check if we should evict without holding the dbuf_evict_lock,
753 * because it's OK to occasionally make the wrong decision here,
754 * and grabbing the lock results in massive lock contention.
756 if (size > dbuf_cache_target_bytes()) {
757 if (size > dbuf_cache_hiwater_bytes())
759 cv_signal(&dbuf_evict_cv);
764 dbuf_kstat_update(kstat_t *ksp, int rw)
766 dbuf_stats_t *ds = ksp->ks_data;
768 if (rw == KSTAT_WRITE) {
769 return (SET_ERROR(EACCES));
771 ds->metadata_cache_size_bytes.value.ui64 = zfs_refcount_count(
772 &dbuf_caches[DB_DBUF_METADATA_CACHE].size);
773 ds->cache_size_bytes.value.ui64 =
774 zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size);
775 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
776 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
777 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
778 ds->hash_elements.value.ui64 = dbuf_hash_count;
787 uint64_t hsize = 1ULL << 16;
788 dbuf_hash_table_t *h = &dbuf_hash_table;
792 * The hash table is big enough to fill all of physical memory
793 * with an average block size of zfs_arc_average_blocksize (default 8K).
794 * By default, the table will take up
795 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
797 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
801 h->hash_table_mask = hsize - 1;
804 * Large allocations which do not require contiguous pages
805 * should be using vmem_alloc() in the linux kernel
807 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP);
809 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
811 if (h->hash_table == NULL) {
812 /* XXX - we should really return an error instead of assert */
813 ASSERT(hsize > (1ULL << 10));
818 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
819 sizeof (dmu_buf_impl_t),
820 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
822 for (i = 0; i < DBUF_MUTEXES; i++)
823 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
828 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
829 * configuration is not required.
831 dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
833 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
834 dbuf_caches[dcs].cache =
835 multilist_create(sizeof (dmu_buf_impl_t),
836 offsetof(dmu_buf_impl_t, db_cache_link),
837 dbuf_cache_multilist_index_func);
838 zfs_refcount_create(&dbuf_caches[dcs].size);
841 dbuf_evict_thread_exit = B_FALSE;
842 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
843 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
844 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
845 NULL, 0, &p0, TS_RUN, minclsyspri);
847 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
848 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
850 if (dbuf_ksp != NULL) {
851 for (i = 0; i < DN_MAX_LEVELS; i++) {
852 snprintf(dbuf_stats.cache_levels[i].name,
853 KSTAT_STRLEN, "cache_level_%d", i);
854 dbuf_stats.cache_levels[i].data_type =
856 snprintf(dbuf_stats.cache_levels_bytes[i].name,
857 KSTAT_STRLEN, "cache_level_%d_bytes", i);
858 dbuf_stats.cache_levels_bytes[i].data_type =
861 dbuf_ksp->ks_data = &dbuf_stats;
862 dbuf_ksp->ks_update = dbuf_kstat_update;
863 kstat_install(dbuf_ksp);
870 dbuf_hash_table_t *h = &dbuf_hash_table;
873 dbuf_stats_destroy();
875 for (i = 0; i < DBUF_MUTEXES; i++)
876 mutex_destroy(&h->hash_mutexes[i]);
879 * Large allocations which do not require contiguous pages
880 * should be using vmem_free() in the linux kernel
882 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
884 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
886 kmem_cache_destroy(dbuf_kmem_cache);
887 taskq_destroy(dbu_evict_taskq);
889 mutex_enter(&dbuf_evict_lock);
890 dbuf_evict_thread_exit = B_TRUE;
891 while (dbuf_evict_thread_exit) {
892 cv_signal(&dbuf_evict_cv);
893 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
895 mutex_exit(&dbuf_evict_lock);
897 mutex_destroy(&dbuf_evict_lock);
898 cv_destroy(&dbuf_evict_cv);
900 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
901 zfs_refcount_destroy(&dbuf_caches[dcs].size);
902 multilist_destroy(dbuf_caches[dcs].cache);
905 if (dbuf_ksp != NULL) {
906 kstat_delete(dbuf_ksp);
917 dbuf_verify(dmu_buf_impl_t *db)
920 dbuf_dirty_record_t *dr;
923 ASSERT(MUTEX_HELD(&db->db_mtx));
925 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
928 ASSERT(db->db_objset != NULL);
932 ASSERT(db->db_parent == NULL);
933 ASSERT(db->db_blkptr == NULL);
935 ASSERT3U(db->db.db_object, ==, dn->dn_object);
936 ASSERT3P(db->db_objset, ==, dn->dn_objset);
937 ASSERT3U(db->db_level, <, dn->dn_nlevels);
938 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
939 db->db_blkid == DMU_SPILL_BLKID ||
940 !avl_is_empty(&dn->dn_dbufs));
942 if (db->db_blkid == DMU_BONUS_BLKID) {
944 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
945 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
946 } else if (db->db_blkid == DMU_SPILL_BLKID) {
948 ASSERT0(db->db.db_offset);
950 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
953 if ((dr = list_head(&db->db_dirty_records)) != NULL) {
954 ASSERT(dr->dr_dbuf == db);
955 txg_prev = dr->dr_txg;
956 for (dr = list_next(&db->db_dirty_records, dr); dr != NULL;
957 dr = list_next(&db->db_dirty_records, dr)) {
958 ASSERT(dr->dr_dbuf == db);
959 ASSERT(txg_prev > dr->dr_txg);
960 txg_prev = dr->dr_txg;
965 * We can't assert that db_size matches dn_datablksz because it
966 * can be momentarily different when another thread is doing
969 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
970 dr = db->db_data_pending;
972 * It should only be modified in syncing context, so
973 * make sure we only have one copy of the data.
975 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
978 /* verify db->db_blkptr */
980 if (db->db_parent == dn->dn_dbuf) {
981 /* db is pointed to by the dnode */
982 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
983 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
984 ASSERT(db->db_parent == NULL);
986 ASSERT(db->db_parent != NULL);
987 if (db->db_blkid != DMU_SPILL_BLKID)
988 ASSERT3P(db->db_blkptr, ==,
989 &dn->dn_phys->dn_blkptr[db->db_blkid]);
991 /* db is pointed to by an indirect block */
992 int epb __maybe_unused = db->db_parent->db.db_size >>
994 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
995 ASSERT3U(db->db_parent->db.db_object, ==,
998 * dnode_grow_indblksz() can make this fail if we don't
999 * have the parent's rwlock. XXX indblksz no longer
1000 * grows. safe to do this now?
1002 if (RW_LOCK_HELD(&db->db_parent->db_rwlock)) {
1003 ASSERT3P(db->db_blkptr, ==,
1004 ((blkptr_t *)db->db_parent->db.db_data +
1005 db->db_blkid % epb));
1009 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
1010 (db->db_buf == NULL || db->db_buf->b_data) &&
1011 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
1012 db->db_state != DB_FILL && !dn->dn_free_txg) {
1014 * If the blkptr isn't set but they have nonzero data,
1015 * it had better be dirty, otherwise we'll lose that
1016 * data when we evict this buffer.
1018 * There is an exception to this rule for indirect blocks; in
1019 * this case, if the indirect block is a hole, we fill in a few
1020 * fields on each of the child blocks (importantly, birth time)
1021 * to prevent hole birth times from being lost when you
1022 * partially fill in a hole.
1024 if (db->db_dirtycnt == 0) {
1025 if (db->db_level == 0) {
1026 uint64_t *buf = db->db.db_data;
1029 for (i = 0; i < db->db.db_size >> 3; i++) {
1030 ASSERT(buf[i] == 0);
1033 blkptr_t *bps = db->db.db_data;
1034 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
1037 * We want to verify that all the blkptrs in the
1038 * indirect block are holes, but we may have
1039 * automatically set up a few fields for them.
1040 * We iterate through each blkptr and verify
1041 * they only have those fields set.
1044 i < db->db.db_size / sizeof (blkptr_t);
1046 blkptr_t *bp = &bps[i];
1047 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1050 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
1051 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
1052 DVA_IS_EMPTY(&bp->blk_dva[2]));
1053 ASSERT0(bp->blk_fill);
1054 ASSERT0(bp->blk_pad[0]);
1055 ASSERT0(bp->blk_pad[1]);
1056 ASSERT(!BP_IS_EMBEDDED(bp));
1057 ASSERT(BP_IS_HOLE(bp));
1058 ASSERT0(bp->blk_phys_birth);
1068 dbuf_clear_data(dmu_buf_impl_t *db)
1070 ASSERT(MUTEX_HELD(&db->db_mtx));
1071 dbuf_evict_user(db);
1072 ASSERT3P(db->db_buf, ==, NULL);
1073 db->db.db_data = NULL;
1074 if (db->db_state != DB_NOFILL) {
1075 db->db_state = DB_UNCACHED;
1076 DTRACE_SET_STATE(db, "clear data");
1081 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1083 ASSERT(MUTEX_HELD(&db->db_mtx));
1084 ASSERT(buf != NULL);
1087 ASSERT(buf->b_data != NULL);
1088 db->db.db_data = buf->b_data;
1092 dbuf_alloc_arcbuf_from_arcbuf(dmu_buf_impl_t *db, arc_buf_t *data)
1094 objset_t *os = db->db_objset;
1095 spa_t *spa = os->os_spa;
1096 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1097 enum zio_compress compress_type;
1100 psize = arc_buf_size(data);
1101 lsize = arc_buf_lsize(data);
1102 compress_type = arc_get_compression(data);
1104 if (arc_is_encrypted(data)) {
1105 boolean_t byteorder;
1106 uint8_t salt[ZIO_DATA_SALT_LEN];
1107 uint8_t iv[ZIO_DATA_IV_LEN];
1108 uint8_t mac[ZIO_DATA_MAC_LEN];
1109 dnode_t *dn = DB_DNODE(db);
1111 arc_get_raw_params(data, &byteorder, salt, iv, mac);
1112 data = arc_alloc_raw_buf(spa, db, dmu_objset_id(os),
1113 byteorder, salt, iv, mac, dn->dn_type, psize, lsize,
1115 } else if (compress_type != ZIO_COMPRESS_OFF) {
1116 ASSERT3U(type, ==, ARC_BUFC_DATA);
1117 data = arc_alloc_compressed_buf(spa, db,
1118 psize, lsize, compress_type);
1120 data = arc_alloc_buf(spa, db, type, psize);
1126 dbuf_alloc_arcbuf(dmu_buf_impl_t *db)
1128 spa_t *spa = db->db_objset->os_spa;
1130 return (arc_alloc_buf(spa, db, DBUF_GET_BUFC_TYPE(db), db->db.db_size));
1134 * Loan out an arc_buf for read. Return the loaned arc_buf.
1137 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1141 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1142 mutex_enter(&db->db_mtx);
1143 if (arc_released(db->db_buf) || zfs_refcount_count(&db->db_holds) > 1) {
1144 int blksz = db->db.db_size;
1145 spa_t *spa = db->db_objset->os_spa;
1147 mutex_exit(&db->db_mtx);
1148 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1149 bcopy(db->db.db_data, abuf->b_data, blksz);
1152 arc_loan_inuse_buf(abuf, db);
1154 dbuf_clear_data(db);
1155 mutex_exit(&db->db_mtx);
1161 * Calculate which level n block references the data at the level 0 offset
1165 dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset)
1167 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1169 * The level n blkid is equal to the level 0 blkid divided by
1170 * the number of level 0s in a level n block.
1172 * The level 0 blkid is offset >> datablkshift =
1173 * offset / 2^datablkshift.
1175 * The number of level 0s in a level n is the number of block
1176 * pointers in an indirect block, raised to the power of level.
1177 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1178 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1180 * Thus, the level n blkid is: offset /
1181 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1182 * = offset / 2^(datablkshift + level *
1183 * (indblkshift - SPA_BLKPTRSHIFT))
1184 * = offset >> (datablkshift + level *
1185 * (indblkshift - SPA_BLKPTRSHIFT))
1188 const unsigned exp = dn->dn_datablkshift +
1189 level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
1191 if (exp >= 8 * sizeof (offset)) {
1192 /* This only happens on the highest indirection level */
1193 ASSERT3U(level, ==, dn->dn_nlevels - 1);
1197 ASSERT3U(exp, <, 8 * sizeof (offset));
1199 return (offset >> exp);
1201 ASSERT3U(offset, <, dn->dn_datablksz);
1207 * This function is used to lock the parent of the provided dbuf. This should be
1208 * used when modifying or reading db_blkptr.
1211 dmu_buf_lock_parent(dmu_buf_impl_t *db, krw_t rw, void *tag)
1213 enum db_lock_type ret = DLT_NONE;
1214 if (db->db_parent != NULL) {
1215 rw_enter(&db->db_parent->db_rwlock, rw);
1217 } else if (dmu_objset_ds(db->db_objset) != NULL) {
1218 rrw_enter(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, rw,
1223 * We only return a DLT_NONE lock when it's the top-most indirect block
1224 * of the meta-dnode of the MOS.
1230 * We need to pass the lock type in because it's possible that the block will
1231 * move from being the topmost indirect block in a dnode (and thus, have no
1232 * parent) to not the top-most via an indirection increase. This would cause a
1233 * panic if we didn't pass the lock type in.
1236 dmu_buf_unlock_parent(dmu_buf_impl_t *db, db_lock_type_t type, void *tag)
1238 if (type == DLT_PARENT)
1239 rw_exit(&db->db_parent->db_rwlock);
1240 else if (type == DLT_OBJSET)
1241 rrw_exit(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, tag);
1245 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1246 arc_buf_t *buf, void *vdb)
1248 dmu_buf_impl_t *db = vdb;
1250 mutex_enter(&db->db_mtx);
1251 ASSERT3U(db->db_state, ==, DB_READ);
1253 * All reads are synchronous, so we must have a hold on the dbuf
1255 ASSERT(zfs_refcount_count(&db->db_holds) > 0);
1256 ASSERT(db->db_buf == NULL);
1257 ASSERT(db->db.db_data == NULL);
1260 ASSERT(zio == NULL || zio->io_error != 0);
1261 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1262 ASSERT3P(db->db_buf, ==, NULL);
1263 db->db_state = DB_UNCACHED;
1264 DTRACE_SET_STATE(db, "i/o error");
1265 } else if (db->db_level == 0 && db->db_freed_in_flight) {
1266 /* freed in flight */
1267 ASSERT(zio == NULL || zio->io_error == 0);
1268 arc_release(buf, db);
1269 bzero(buf->b_data, db->db.db_size);
1270 arc_buf_freeze(buf);
1271 db->db_freed_in_flight = FALSE;
1272 dbuf_set_data(db, buf);
1273 db->db_state = DB_CACHED;
1274 DTRACE_SET_STATE(db, "freed in flight");
1277 ASSERT(zio == NULL || zio->io_error == 0);
1278 dbuf_set_data(db, buf);
1279 db->db_state = DB_CACHED;
1280 DTRACE_SET_STATE(db, "successful read");
1282 cv_broadcast(&db->db_changed);
1283 dbuf_rele_and_unlock(db, NULL, B_FALSE);
1287 * Shortcut for performing reads on bonus dbufs. Returns
1288 * an error if we fail to verify the dnode associated with
1289 * a decrypted block. Otherwise success.
1292 dbuf_read_bonus(dmu_buf_impl_t *db, dnode_t *dn, uint32_t flags)
1294 int bonuslen, max_bonuslen, err;
1296 err = dbuf_read_verify_dnode_crypt(db, flags);
1300 bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1301 max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1302 ASSERT(MUTEX_HELD(&db->db_mtx));
1303 ASSERT(DB_DNODE_HELD(db));
1304 ASSERT3U(bonuslen, <=, db->db.db_size);
1305 db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP);
1306 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1307 if (bonuslen < max_bonuslen)
1308 bzero(db->db.db_data, max_bonuslen);
1310 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1311 db->db_state = DB_CACHED;
1312 DTRACE_SET_STATE(db, "bonus buffer filled");
1317 dbuf_handle_indirect_hole(dmu_buf_impl_t *db, dnode_t *dn)
1319 blkptr_t *bps = db->db.db_data;
1320 uint32_t indbs = 1ULL << dn->dn_indblkshift;
1321 int n_bps = indbs >> SPA_BLKPTRSHIFT;
1323 for (int i = 0; i < n_bps; i++) {
1324 blkptr_t *bp = &bps[i];
1326 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, indbs);
1327 BP_SET_LSIZE(bp, BP_GET_LEVEL(db->db_blkptr) == 1 ?
1328 dn->dn_datablksz : BP_GET_LSIZE(db->db_blkptr));
1329 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1330 BP_SET_LEVEL(bp, BP_GET_LEVEL(db->db_blkptr) - 1);
1331 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1336 * Handle reads on dbufs that are holes, if necessary. This function
1337 * requires that the dbuf's mutex is held. Returns success (0) if action
1338 * was taken, ENOENT if no action was taken.
1341 dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn, uint32_t flags)
1343 ASSERT(MUTEX_HELD(&db->db_mtx));
1345 int is_hole = db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr);
1347 * For level 0 blocks only, if the above check fails:
1348 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1349 * processes the delete record and clears the bp while we are waiting
1350 * for the dn_mtx (resulting in a "no" from block_freed).
1352 if (!is_hole && db->db_level == 0) {
1353 is_hole = dnode_block_freed(dn, db->db_blkid) ||
1354 BP_IS_HOLE(db->db_blkptr);
1358 dbuf_set_data(db, dbuf_alloc_arcbuf(db));
1359 bzero(db->db.db_data, db->db.db_size);
1361 if (db->db_blkptr != NULL && db->db_level > 0 &&
1362 BP_IS_HOLE(db->db_blkptr) &&
1363 db->db_blkptr->blk_birth != 0) {
1364 dbuf_handle_indirect_hole(db, dn);
1366 db->db_state = DB_CACHED;
1367 DTRACE_SET_STATE(db, "hole read satisfied");
1374 * This function ensures that, when doing a decrypting read of a block,
1375 * we make sure we have decrypted the dnode associated with it. We must do
1376 * this so that we ensure we are fully authenticating the checksum-of-MACs
1377 * tree from the root of the objset down to this block. Indirect blocks are
1378 * always verified against their secure checksum-of-MACs assuming that the
1379 * dnode containing them is correct. Now that we are doing a decrypting read,
1380 * we can be sure that the key is loaded and verify that assumption. This is
1381 * especially important considering that we always read encrypted dnode
1382 * blocks as raw data (without verifying their MACs) to start, and
1383 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1386 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags)
1389 objset_t *os = db->db_objset;
1390 arc_buf_t *dnode_abuf;
1392 zbookmark_phys_t zb;
1394 ASSERT(MUTEX_HELD(&db->db_mtx));
1396 if (!os->os_encrypted || os->os_raw_receive ||
1397 (flags & DB_RF_NO_DECRYPT) != 0)
1402 dnode_abuf = (dn->dn_dbuf != NULL) ? dn->dn_dbuf->db_buf : NULL;
1404 if (dnode_abuf == NULL || !arc_is_encrypted(dnode_abuf)) {
1409 SET_BOOKMARK(&zb, dmu_objset_id(os),
1410 DMU_META_DNODE_OBJECT, 0, dn->dn_dbuf->db_blkid);
1411 err = arc_untransform(dnode_abuf, os->os_spa, &zb, B_TRUE);
1414 * An error code of EACCES tells us that the key is still not
1415 * available. This is ok if we are only reading authenticated
1416 * (and therefore non-encrypted) blocks.
1418 if (err == EACCES && ((db->db_blkid != DMU_BONUS_BLKID &&
1419 !DMU_OT_IS_ENCRYPTED(dn->dn_type)) ||
1420 (db->db_blkid == DMU_BONUS_BLKID &&
1421 !DMU_OT_IS_ENCRYPTED(dn->dn_bonustype))))
1430 * Drops db_mtx and the parent lock specified by dblt and tag before
1434 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags,
1435 db_lock_type_t dblt, void *tag)
1438 zbookmark_phys_t zb;
1439 uint32_t aflags = ARC_FLAG_NOWAIT;
1441 boolean_t bonus_read;
1443 err = zio_flags = 0;
1444 bonus_read = B_FALSE;
1447 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1448 ASSERT(MUTEX_HELD(&db->db_mtx));
1449 ASSERT(db->db_state == DB_UNCACHED);
1450 ASSERT(db->db_buf == NULL);
1451 ASSERT(db->db_parent == NULL ||
1452 RW_LOCK_HELD(&db->db_parent->db_rwlock));
1454 if (db->db_blkid == DMU_BONUS_BLKID) {
1455 err = dbuf_read_bonus(db, dn, flags);
1459 err = dbuf_read_hole(db, dn, flags);
1464 * Any attempt to read a redacted block should result in an error. This
1465 * will never happen under normal conditions, but can be useful for
1466 * debugging purposes.
1468 if (BP_IS_REDACTED(db->db_blkptr)) {
1469 ASSERT(dsl_dataset_feature_is_active(
1470 db->db_objset->os_dsl_dataset,
1471 SPA_FEATURE_REDACTED_DATASETS));
1472 err = SET_ERROR(EIO);
1476 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
1477 db->db.db_object, db->db_level, db->db_blkid);
1480 * All bps of an encrypted os should have the encryption bit set.
1481 * If this is not true it indicates tampering and we report an error.
1483 if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) {
1484 spa_log_error(db->db_objset->os_spa, &zb);
1485 zfs_panic_recover("unencrypted block in encrypted "
1486 "object set %llu", dmu_objset_id(db->db_objset));
1487 err = SET_ERROR(EIO);
1491 err = dbuf_read_verify_dnode_crypt(db, flags);
1497 db->db_state = DB_READ;
1498 DTRACE_SET_STATE(db, "read issued");
1499 mutex_exit(&db->db_mtx);
1501 if (DBUF_IS_L2CACHEABLE(db))
1502 aflags |= ARC_FLAG_L2CACHE;
1504 dbuf_add_ref(db, NULL);
1506 zio_flags = (flags & DB_RF_CANFAIL) ?
1507 ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED;
1509 if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr))
1510 zio_flags |= ZIO_FLAG_RAW;
1512 * The zio layer will copy the provided blkptr later, but we need to
1513 * do this now so that we can release the parent's rwlock. We have to
1514 * do that now so that if dbuf_read_done is called synchronously (on
1515 * an l1 cache hit) we don't acquire the db_mtx while holding the
1516 * parent's rwlock, which would be a lock ordering violation.
1518 blkptr_t bp = *db->db_blkptr;
1519 dmu_buf_unlock_parent(db, dblt, tag);
1520 (void) arc_read(zio, db->db_objset->os_spa, &bp,
1521 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags,
1526 mutex_exit(&db->db_mtx);
1527 dmu_buf_unlock_parent(db, dblt, tag);
1532 * This is our just-in-time copy function. It makes a copy of buffers that
1533 * have been modified in a previous transaction group before we access them in
1534 * the current active group.
1536 * This function is used in three places: when we are dirtying a buffer for the
1537 * first time in a txg, when we are freeing a range in a dnode that includes
1538 * this buffer, and when we are accessing a buffer which was received compressed
1539 * and later referenced in a WRITE_BYREF record.
1541 * Note that when we are called from dbuf_free_range() we do not put a hold on
1542 * the buffer, we just traverse the active dbuf list for the dnode.
1545 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1547 dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records);
1549 ASSERT(MUTEX_HELD(&db->db_mtx));
1550 ASSERT(db->db.db_data != NULL);
1551 ASSERT(db->db_level == 0);
1552 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1555 (dr->dt.dl.dr_data !=
1556 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1560 * If the last dirty record for this dbuf has not yet synced
1561 * and its referencing the dbuf data, either:
1562 * reset the reference to point to a new copy,
1563 * or (if there a no active holders)
1564 * just null out the current db_data pointer.
1566 ASSERT3U(dr->dr_txg, >=, txg - 2);
1567 if (db->db_blkid == DMU_BONUS_BLKID) {
1568 dnode_t *dn = DB_DNODE(db);
1569 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1570 dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP);
1571 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1572 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1573 } else if (zfs_refcount_count(&db->db_holds) > db->db_dirtycnt) {
1574 arc_buf_t *buf = dbuf_alloc_arcbuf_from_arcbuf(db, db->db_buf);
1575 dr->dt.dl.dr_data = buf;
1576 bcopy(db->db.db_data, buf->b_data, arc_buf_size(buf));
1579 dbuf_clear_data(db);
1584 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1591 * We don't have to hold the mutex to check db_state because it
1592 * can't be freed while we have a hold on the buffer.
1594 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1596 if (db->db_state == DB_NOFILL)
1597 return (SET_ERROR(EIO));
1602 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1603 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1604 DBUF_IS_CACHEABLE(db);
1606 mutex_enter(&db->db_mtx);
1607 if (db->db_state == DB_CACHED) {
1608 spa_t *spa = dn->dn_objset->os_spa;
1611 * Ensure that this block's dnode has been decrypted if
1612 * the caller has requested decrypted data.
1614 err = dbuf_read_verify_dnode_crypt(db, flags);
1617 * If the arc buf is compressed or encrypted and the caller
1618 * requested uncompressed data, we need to untransform it
1619 * before returning. We also call arc_untransform() on any
1620 * unauthenticated blocks, which will verify their MAC if
1621 * the key is now available.
1623 if (err == 0 && db->db_buf != NULL &&
1624 (flags & DB_RF_NO_DECRYPT) == 0 &&
1625 (arc_is_encrypted(db->db_buf) ||
1626 arc_is_unauthenticated(db->db_buf) ||
1627 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) {
1628 zbookmark_phys_t zb;
1630 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
1631 db->db.db_object, db->db_level, db->db_blkid);
1632 dbuf_fix_old_data(db, spa_syncing_txg(spa));
1633 err = arc_untransform(db->db_buf, spa, &zb, B_FALSE);
1634 dbuf_set_data(db, db->db_buf);
1636 mutex_exit(&db->db_mtx);
1637 if (err == 0 && prefetch) {
1638 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE,
1639 flags & DB_RF_HAVESTRUCT);
1642 DBUF_STAT_BUMP(hash_hits);
1643 } else if (db->db_state == DB_UNCACHED) {
1644 spa_t *spa = dn->dn_objset->os_spa;
1645 boolean_t need_wait = B_FALSE;
1647 db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
1650 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1651 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1654 err = dbuf_read_impl(db, zio, flags, dblt, FTAG);
1656 * dbuf_read_impl has dropped db_mtx and our parent's rwlock
1659 if (!err && prefetch) {
1660 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE,
1661 flags & DB_RF_HAVESTRUCT);
1665 DBUF_STAT_BUMP(hash_misses);
1668 * If we created a zio_root we must execute it to avoid
1669 * leaking it, even if it isn't attached to any work due
1670 * to an error in dbuf_read_impl().
1674 err = zio_wait(zio);
1676 VERIFY0(zio_wait(zio));
1680 * Another reader came in while the dbuf was in flight
1681 * between UNCACHED and CACHED. Either a writer will finish
1682 * writing the buffer (sending the dbuf to CACHED) or the
1683 * first reader's request will reach the read_done callback
1684 * and send the dbuf to CACHED. Otherwise, a failure
1685 * occurred and the dbuf went to UNCACHED.
1687 mutex_exit(&db->db_mtx);
1689 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE,
1690 flags & DB_RF_HAVESTRUCT);
1693 DBUF_STAT_BUMP(hash_misses);
1695 /* Skip the wait per the caller's request. */
1696 if ((flags & DB_RF_NEVERWAIT) == 0) {
1697 mutex_enter(&db->db_mtx);
1698 while (db->db_state == DB_READ ||
1699 db->db_state == DB_FILL) {
1700 ASSERT(db->db_state == DB_READ ||
1701 (flags & DB_RF_HAVESTRUCT) == 0);
1702 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1704 cv_wait(&db->db_changed, &db->db_mtx);
1706 if (db->db_state == DB_UNCACHED)
1707 err = SET_ERROR(EIO);
1708 mutex_exit(&db->db_mtx);
1716 dbuf_noread(dmu_buf_impl_t *db)
1718 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1719 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1720 mutex_enter(&db->db_mtx);
1721 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1722 cv_wait(&db->db_changed, &db->db_mtx);
1723 if (db->db_state == DB_UNCACHED) {
1724 ASSERT(db->db_buf == NULL);
1725 ASSERT(db->db.db_data == NULL);
1726 dbuf_set_data(db, dbuf_alloc_arcbuf(db));
1727 db->db_state = DB_FILL;
1728 DTRACE_SET_STATE(db, "assigning filled buffer");
1729 } else if (db->db_state == DB_NOFILL) {
1730 dbuf_clear_data(db);
1732 ASSERT3U(db->db_state, ==, DB_CACHED);
1734 mutex_exit(&db->db_mtx);
1738 dbuf_unoverride(dbuf_dirty_record_t *dr)
1740 dmu_buf_impl_t *db = dr->dr_dbuf;
1741 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1742 uint64_t txg = dr->dr_txg;
1744 ASSERT(MUTEX_HELD(&db->db_mtx));
1746 * This assert is valid because dmu_sync() expects to be called by
1747 * a zilog's get_data while holding a range lock. This call only
1748 * comes from dbuf_dirty() callers who must also hold a range lock.
1750 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1751 ASSERT(db->db_level == 0);
1753 if (db->db_blkid == DMU_BONUS_BLKID ||
1754 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1757 ASSERT(db->db_data_pending != dr);
1759 /* free this block */
1760 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1761 zio_free(db->db_objset->os_spa, txg, bp);
1763 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1764 dr->dt.dl.dr_nopwrite = B_FALSE;
1765 dr->dt.dl.dr_has_raw_params = B_FALSE;
1768 * Release the already-written buffer, so we leave it in
1769 * a consistent dirty state. Note that all callers are
1770 * modifying the buffer, so they will immediately do
1771 * another (redundant) arc_release(). Therefore, leave
1772 * the buf thawed to save the effort of freezing &
1773 * immediately re-thawing it.
1775 arc_release(dr->dt.dl.dr_data, db);
1779 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1780 * data blocks in the free range, so that any future readers will find
1784 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1787 dmu_buf_impl_t *db_search;
1788 dmu_buf_impl_t *db, *db_next;
1789 uint64_t txg = tx->tx_txg;
1791 dbuf_dirty_record_t *dr;
1793 if (end_blkid > dn->dn_maxblkid &&
1794 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1795 end_blkid = dn->dn_maxblkid;
1796 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1798 db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1799 db_search->db_level = 0;
1800 db_search->db_blkid = start_blkid;
1801 db_search->db_state = DB_SEARCH;
1803 mutex_enter(&dn->dn_dbufs_mtx);
1804 db = avl_find(&dn->dn_dbufs, db_search, &where);
1805 ASSERT3P(db, ==, NULL);
1807 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1809 for (; db != NULL; db = db_next) {
1810 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1811 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1813 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1816 ASSERT3U(db->db_blkid, >=, start_blkid);
1818 /* found a level 0 buffer in the range */
1819 mutex_enter(&db->db_mtx);
1820 if (dbuf_undirty(db, tx)) {
1821 /* mutex has been dropped and dbuf destroyed */
1825 if (db->db_state == DB_UNCACHED ||
1826 db->db_state == DB_NOFILL ||
1827 db->db_state == DB_EVICTING) {
1828 ASSERT(db->db.db_data == NULL);
1829 mutex_exit(&db->db_mtx);
1832 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1833 /* will be handled in dbuf_read_done or dbuf_rele */
1834 db->db_freed_in_flight = TRUE;
1835 mutex_exit(&db->db_mtx);
1838 if (zfs_refcount_count(&db->db_holds) == 0) {
1843 /* The dbuf is referenced */
1845 dr = list_head(&db->db_dirty_records);
1847 if (dr->dr_txg == txg) {
1849 * This buffer is "in-use", re-adjust the file
1850 * size to reflect that this buffer may
1851 * contain new data when we sync.
1853 if (db->db_blkid != DMU_SPILL_BLKID &&
1854 db->db_blkid > dn->dn_maxblkid)
1855 dn->dn_maxblkid = db->db_blkid;
1856 dbuf_unoverride(dr);
1859 * This dbuf is not dirty in the open context.
1860 * Either uncache it (if its not referenced in
1861 * the open context) or reset its contents to
1864 dbuf_fix_old_data(db, txg);
1867 /* clear the contents if its cached */
1868 if (db->db_state == DB_CACHED) {
1869 ASSERT(db->db.db_data != NULL);
1870 arc_release(db->db_buf, db);
1871 rw_enter(&db->db_rwlock, RW_WRITER);
1872 bzero(db->db.db_data, db->db.db_size);
1873 rw_exit(&db->db_rwlock);
1874 arc_buf_freeze(db->db_buf);
1877 mutex_exit(&db->db_mtx);
1880 kmem_free(db_search, sizeof (dmu_buf_impl_t));
1881 mutex_exit(&dn->dn_dbufs_mtx);
1885 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1887 arc_buf_t *buf, *old_buf;
1888 dbuf_dirty_record_t *dr;
1889 int osize = db->db.db_size;
1890 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1893 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1899 * XXX we should be doing a dbuf_read, checking the return
1900 * value and returning that up to our callers
1902 dmu_buf_will_dirty(&db->db, tx);
1904 /* create the data buffer for the new block */
1905 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1907 /* copy old block data to the new block */
1908 old_buf = db->db_buf;
1909 bcopy(old_buf->b_data, buf->b_data, MIN(osize, size));
1910 /* zero the remainder */
1912 bzero((uint8_t *)buf->b_data + osize, size - osize);
1914 mutex_enter(&db->db_mtx);
1915 dbuf_set_data(db, buf);
1916 arc_buf_destroy(old_buf, db);
1917 db->db.db_size = size;
1919 dr = list_head(&db->db_dirty_records);
1920 /* dirty record added by dmu_buf_will_dirty() */
1922 if (db->db_level == 0)
1923 dr->dt.dl.dr_data = buf;
1924 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
1925 ASSERT3U(dr->dr_accounted, ==, osize);
1926 dr->dr_accounted = size;
1927 mutex_exit(&db->db_mtx);
1929 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1934 dbuf_release_bp(dmu_buf_impl_t *db)
1936 objset_t *os __maybe_unused = db->db_objset;
1938 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1939 ASSERT(arc_released(os->os_phys_buf) ||
1940 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1941 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1943 (void) arc_release(db->db_buf, db);
1947 * We already have a dirty record for this TXG, and we are being
1951 dbuf_redirty(dbuf_dirty_record_t *dr)
1953 dmu_buf_impl_t *db = dr->dr_dbuf;
1955 ASSERT(MUTEX_HELD(&db->db_mtx));
1957 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1959 * If this buffer has already been written out,
1960 * we now need to reset its state.
1962 dbuf_unoverride(dr);
1963 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1964 db->db_state != DB_NOFILL) {
1965 /* Already released on initial dirty, so just thaw. */
1966 ASSERT(arc_released(db->db_buf));
1967 arc_buf_thaw(db->db_buf);
1972 dbuf_dirty_record_t *
1973 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1977 dbuf_dirty_record_t *dr, *dr_next, *dr_head;
1978 int txgoff = tx->tx_txg & TXG_MASK;
1979 boolean_t drop_struct_rwlock = B_FALSE;
1981 ASSERT(tx->tx_txg != 0);
1982 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1983 DMU_TX_DIRTY_BUF(tx, db);
1988 * Shouldn't dirty a regular buffer in syncing context. Private
1989 * objects may be dirtied in syncing context, but only if they
1990 * were already pre-dirtied in open context.
1993 if (dn->dn_objset->os_dsl_dataset != NULL) {
1994 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1997 ASSERT(!dmu_tx_is_syncing(tx) ||
1998 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1999 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
2000 dn->dn_objset->os_dsl_dataset == NULL);
2001 if (dn->dn_objset->os_dsl_dataset != NULL)
2002 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
2005 * We make this assert for private objects as well, but after we
2006 * check if we're already dirty. They are allowed to re-dirty
2007 * in syncing context.
2009 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2010 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
2011 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
2013 mutex_enter(&db->db_mtx);
2015 * XXX make this true for indirects too? The problem is that
2016 * transactions created with dmu_tx_create_assigned() from
2017 * syncing context don't bother holding ahead.
2019 ASSERT(db->db_level != 0 ||
2020 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
2021 db->db_state == DB_NOFILL);
2023 mutex_enter(&dn->dn_mtx);
2024 dnode_set_dirtyctx(dn, tx, db);
2025 if (tx->tx_txg > dn->dn_dirty_txg)
2026 dn->dn_dirty_txg = tx->tx_txg;
2027 mutex_exit(&dn->dn_mtx);
2029 if (db->db_blkid == DMU_SPILL_BLKID)
2030 dn->dn_have_spill = B_TRUE;
2033 * If this buffer is already dirty, we're done.
2035 dr_head = list_head(&db->db_dirty_records);
2036 ASSERT(dr_head == NULL || dr_head->dr_txg <= tx->tx_txg ||
2037 db->db.db_object == DMU_META_DNODE_OBJECT);
2038 dr_next = dbuf_find_dirty_lte(db, tx->tx_txg);
2039 if (dr_next && dr_next->dr_txg == tx->tx_txg) {
2042 dbuf_redirty(dr_next);
2043 mutex_exit(&db->db_mtx);
2048 * Only valid if not already dirty.
2050 ASSERT(dn->dn_object == 0 ||
2051 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
2052 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
2054 ASSERT3U(dn->dn_nlevels, >, db->db_level);
2057 * We should only be dirtying in syncing context if it's the
2058 * mos or we're initializing the os or it's a special object.
2059 * However, we are allowed to dirty in syncing context provided
2060 * we already dirtied it in open context. Hence we must make
2061 * this assertion only if we're not already dirty.
2064 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
2066 if (dn->dn_objset->os_dsl_dataset != NULL)
2067 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
2068 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
2069 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
2070 if (dn->dn_objset->os_dsl_dataset != NULL)
2071 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
2073 ASSERT(db->db.db_size != 0);
2075 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2077 if (db->db_blkid != DMU_BONUS_BLKID) {
2078 dmu_objset_willuse_space(os, db->db.db_size, tx);
2082 * If this buffer is dirty in an old transaction group we need
2083 * to make a copy of it so that the changes we make in this
2084 * transaction group won't leak out when we sync the older txg.
2086 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
2087 list_link_init(&dr->dr_dirty_node);
2088 list_link_init(&dr->dr_dbuf_node);
2089 if (db->db_level == 0) {
2090 void *data_old = db->db_buf;
2092 if (db->db_state != DB_NOFILL) {
2093 if (db->db_blkid == DMU_BONUS_BLKID) {
2094 dbuf_fix_old_data(db, tx->tx_txg);
2095 data_old = db->db.db_data;
2096 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
2098 * Release the data buffer from the cache so
2099 * that we can modify it without impacting
2100 * possible other users of this cached data
2101 * block. Note that indirect blocks and
2102 * private objects are not released until the
2103 * syncing state (since they are only modified
2106 arc_release(db->db_buf, db);
2107 dbuf_fix_old_data(db, tx->tx_txg);
2108 data_old = db->db_buf;
2110 ASSERT(data_old != NULL);
2112 dr->dt.dl.dr_data = data_old;
2114 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL);
2115 list_create(&dr->dt.di.dr_children,
2116 sizeof (dbuf_dirty_record_t),
2117 offsetof(dbuf_dirty_record_t, dr_dirty_node));
2119 if (db->db_blkid != DMU_BONUS_BLKID)
2120 dr->dr_accounted = db->db.db_size;
2122 dr->dr_txg = tx->tx_txg;
2123 list_insert_before(&db->db_dirty_records, dr_next, dr);
2126 * We could have been freed_in_flight between the dbuf_noread
2127 * and dbuf_dirty. We win, as though the dbuf_noread() had
2128 * happened after the free.
2130 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2131 db->db_blkid != DMU_SPILL_BLKID) {
2132 mutex_enter(&dn->dn_mtx);
2133 if (dn->dn_free_ranges[txgoff] != NULL) {
2134 range_tree_clear(dn->dn_free_ranges[txgoff],
2137 mutex_exit(&dn->dn_mtx);
2138 db->db_freed_in_flight = FALSE;
2142 * This buffer is now part of this txg
2144 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
2145 db->db_dirtycnt += 1;
2146 ASSERT3U(db->db_dirtycnt, <=, 3);
2148 mutex_exit(&db->db_mtx);
2150 if (db->db_blkid == DMU_BONUS_BLKID ||
2151 db->db_blkid == DMU_SPILL_BLKID) {
2152 mutex_enter(&dn->dn_mtx);
2153 ASSERT(!list_link_active(&dr->dr_dirty_node));
2154 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2155 mutex_exit(&dn->dn_mtx);
2156 dnode_setdirty(dn, tx);
2161 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
2162 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2163 drop_struct_rwlock = B_TRUE;
2167 * If we are overwriting a dedup BP, then unless it is snapshotted,
2168 * when we get to syncing context we will need to decrement its
2169 * refcount in the DDT. Prefetch the relevant DDT block so that
2170 * syncing context won't have to wait for the i/o.
2172 if (db->db_blkptr != NULL) {
2173 db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
2174 ddt_prefetch(os->os_spa, db->db_blkptr);
2175 dmu_buf_unlock_parent(db, dblt, FTAG);
2179 * We need to hold the dn_struct_rwlock to make this assertion,
2180 * because it protects dn_phys / dn_next_nlevels from changing.
2182 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
2183 dn->dn_phys->dn_nlevels > db->db_level ||
2184 dn->dn_next_nlevels[txgoff] > db->db_level ||
2185 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
2186 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
2189 if (db->db_level == 0) {
2190 ASSERT(!db->db_objset->os_raw_receive ||
2191 dn->dn_maxblkid >= db->db_blkid);
2192 dnode_new_blkid(dn, db->db_blkid, tx,
2193 drop_struct_rwlock, B_FALSE);
2194 ASSERT(dn->dn_maxblkid >= db->db_blkid);
2197 if (db->db_level+1 < dn->dn_nlevels) {
2198 dmu_buf_impl_t *parent = db->db_parent;
2199 dbuf_dirty_record_t *di;
2200 int parent_held = FALSE;
2202 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
2203 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2204 parent = dbuf_hold_level(dn, db->db_level + 1,
2205 db->db_blkid >> epbs, FTAG);
2206 ASSERT(parent != NULL);
2209 if (drop_struct_rwlock)
2210 rw_exit(&dn->dn_struct_rwlock);
2211 ASSERT3U(db->db_level + 1, ==, parent->db_level);
2212 di = dbuf_dirty(parent, tx);
2214 dbuf_rele(parent, FTAG);
2216 mutex_enter(&db->db_mtx);
2218 * Since we've dropped the mutex, it's possible that
2219 * dbuf_undirty() might have changed this out from under us.
2221 if (list_head(&db->db_dirty_records) == dr ||
2222 dn->dn_object == DMU_META_DNODE_OBJECT) {
2223 mutex_enter(&di->dt.di.dr_mtx);
2224 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
2225 ASSERT(!list_link_active(&dr->dr_dirty_node));
2226 list_insert_tail(&di->dt.di.dr_children, dr);
2227 mutex_exit(&di->dt.di.dr_mtx);
2230 mutex_exit(&db->db_mtx);
2232 ASSERT(db->db_level + 1 == dn->dn_nlevels);
2233 ASSERT(db->db_blkid < dn->dn_nblkptr);
2234 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
2235 mutex_enter(&dn->dn_mtx);
2236 ASSERT(!list_link_active(&dr->dr_dirty_node));
2237 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2238 mutex_exit(&dn->dn_mtx);
2239 if (drop_struct_rwlock)
2240 rw_exit(&dn->dn_struct_rwlock);
2243 dnode_setdirty(dn, tx);
2249 dbuf_undirty_bonus(dbuf_dirty_record_t *dr)
2251 dmu_buf_impl_t *db = dr->dr_dbuf;
2253 if (dr->dt.dl.dr_data != db->db.db_data) {
2254 struct dnode *dn = DB_DNODE(db);
2255 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
2257 kmem_free(dr->dt.dl.dr_data, max_bonuslen);
2258 arc_space_return(max_bonuslen, ARC_SPACE_BONUS);
2260 db->db_data_pending = NULL;
2261 ASSERT(list_next(&db->db_dirty_records, dr) == NULL);
2262 list_remove(&db->db_dirty_records, dr);
2263 if (dr->dr_dbuf->db_level != 0) {
2264 mutex_destroy(&dr->dt.di.dr_mtx);
2265 list_destroy(&dr->dt.di.dr_children);
2267 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2268 ASSERT3U(db->db_dirtycnt, >, 0);
2269 db->db_dirtycnt -= 1;
2273 * Undirty a buffer in the transaction group referenced by the given
2274 * transaction. Return whether this evicted the dbuf.
2277 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2280 uint64_t txg = tx->tx_txg;
2281 dbuf_dirty_record_t *dr;
2286 * Due to our use of dn_nlevels below, this can only be called
2287 * in open context, unless we are operating on the MOS.
2288 * From syncing context, dn_nlevels may be different from the
2289 * dn_nlevels used when dbuf was dirtied.
2291 ASSERT(db->db_objset ==
2292 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2293 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2294 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2295 ASSERT0(db->db_level);
2296 ASSERT(MUTEX_HELD(&db->db_mtx));
2299 * If this buffer is not dirty, we're done.
2301 dr = dbuf_find_dirty_eq(db, txg);
2304 ASSERT(dr->dr_dbuf == db);
2309 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2311 ASSERT(db->db.db_size != 0);
2313 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2314 dr->dr_accounted, txg);
2316 list_remove(&db->db_dirty_records, dr);
2319 * Note that there are three places in dbuf_dirty()
2320 * where this dirty record may be put on a list.
2321 * Make sure to do a list_remove corresponding to
2322 * every one of those list_insert calls.
2324 if (dr->dr_parent) {
2325 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2326 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2327 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2328 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2329 db->db_level + 1 == dn->dn_nlevels) {
2330 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2331 mutex_enter(&dn->dn_mtx);
2332 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2333 mutex_exit(&dn->dn_mtx);
2337 if (db->db_state != DB_NOFILL) {
2338 dbuf_unoverride(dr);
2340 ASSERT(db->db_buf != NULL);
2341 ASSERT(dr->dt.dl.dr_data != NULL);
2342 if (dr->dt.dl.dr_data != db->db_buf)
2343 arc_buf_destroy(dr->dt.dl.dr_data, db);
2346 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2348 ASSERT(db->db_dirtycnt > 0);
2349 db->db_dirtycnt -= 1;
2351 if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2352 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2361 dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx)
2363 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2365 ASSERT(tx->tx_txg != 0);
2366 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2369 * Quick check for dirtiness. For already dirty blocks, this
2370 * reduces runtime of this function by >90%, and overall performance
2371 * by 50% for some workloads (e.g. file deletion with indirect blocks
2374 mutex_enter(&db->db_mtx);
2376 if (db->db_state == DB_CACHED) {
2377 dbuf_dirty_record_t *dr = dbuf_find_dirty_eq(db, tx->tx_txg);
2379 * It's possible that it is already dirty but not cached,
2380 * because there are some calls to dbuf_dirty() that don't
2381 * go through dmu_buf_will_dirty().
2384 /* This dbuf is already dirty and cached. */
2386 mutex_exit(&db->db_mtx);
2390 mutex_exit(&db->db_mtx);
2393 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2394 flags |= DB_RF_HAVESTRUCT;
2396 (void) dbuf_read(db, NULL, flags);
2397 (void) dbuf_dirty(db, tx);
2401 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2403 dmu_buf_will_dirty_impl(db_fake,
2404 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx);
2408 dmu_buf_is_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2410 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2411 dbuf_dirty_record_t *dr;
2413 mutex_enter(&db->db_mtx);
2414 dr = dbuf_find_dirty_eq(db, tx->tx_txg);
2415 mutex_exit(&db->db_mtx);
2416 return (dr != NULL);
2420 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2422 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2424 db->db_state = DB_NOFILL;
2425 DTRACE_SET_STATE(db, "allocating NOFILL buffer");
2426 dmu_buf_will_fill(db_fake, tx);
2430 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2432 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2434 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2435 ASSERT(tx->tx_txg != 0);
2436 ASSERT(db->db_level == 0);
2437 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2439 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2440 dmu_tx_private_ok(tx));
2443 (void) dbuf_dirty(db, tx);
2447 * This function is effectively the same as dmu_buf_will_dirty(), but
2448 * indicates the caller expects raw encrypted data in the db, and provides
2449 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2450 * blkptr_t when this dbuf is written. This is only used for blocks of
2451 * dnodes, during raw receive.
2454 dmu_buf_set_crypt_params(dmu_buf_t *db_fake, boolean_t byteorder,
2455 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx)
2457 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2458 dbuf_dirty_record_t *dr;
2461 * dr_has_raw_params is only processed for blocks of dnodes
2462 * (see dbuf_sync_dnode_leaf_crypt()).
2464 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT);
2465 ASSERT3U(db->db_level, ==, 0);
2466 ASSERT(db->db_objset->os_raw_receive);
2468 dmu_buf_will_dirty_impl(db_fake,
2469 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx);
2471 dr = dbuf_find_dirty_eq(db, tx->tx_txg);
2473 ASSERT3P(dr, !=, NULL);
2475 dr->dt.dl.dr_has_raw_params = B_TRUE;
2476 dr->dt.dl.dr_byteorder = byteorder;
2477 bcopy(salt, dr->dt.dl.dr_salt, ZIO_DATA_SALT_LEN);
2478 bcopy(iv, dr->dt.dl.dr_iv, ZIO_DATA_IV_LEN);
2479 bcopy(mac, dr->dt.dl.dr_mac, ZIO_DATA_MAC_LEN);
2483 dbuf_override_impl(dmu_buf_impl_t *db, const blkptr_t *bp, dmu_tx_t *tx)
2485 struct dirty_leaf *dl;
2486 dbuf_dirty_record_t *dr;
2488 dr = list_head(&db->db_dirty_records);
2489 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2491 dl->dr_overridden_by = *bp;
2492 dl->dr_override_state = DR_OVERRIDDEN;
2493 dl->dr_overridden_by.blk_birth = dr->dr_txg;
2498 dmu_buf_fill_done(dmu_buf_t *dbuf, dmu_tx_t *tx)
2500 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2501 dbuf_states_t old_state;
2502 mutex_enter(&db->db_mtx);
2505 old_state = db->db_state;
2506 db->db_state = DB_CACHED;
2507 if (old_state == DB_FILL) {
2508 if (db->db_level == 0 && db->db_freed_in_flight) {
2509 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2510 /* we were freed while filling */
2511 /* XXX dbuf_undirty? */
2512 bzero(db->db.db_data, db->db.db_size);
2513 db->db_freed_in_flight = FALSE;
2514 DTRACE_SET_STATE(db,
2515 "fill done handling freed in flight");
2517 DTRACE_SET_STATE(db, "fill done");
2519 cv_broadcast(&db->db_changed);
2521 mutex_exit(&db->db_mtx);
2525 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2526 bp_embedded_type_t etype, enum zio_compress comp,
2527 int uncompressed_size, int compressed_size, int byteorder,
2530 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2531 struct dirty_leaf *dl;
2532 dmu_object_type_t type;
2533 dbuf_dirty_record_t *dr;
2535 if (etype == BP_EMBEDDED_TYPE_DATA) {
2536 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2537 SPA_FEATURE_EMBEDDED_DATA));
2541 type = DB_DNODE(db)->dn_type;
2544 ASSERT0(db->db_level);
2545 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2547 dmu_buf_will_not_fill(dbuf, tx);
2549 dr = list_head(&db->db_dirty_records);
2550 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2552 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2553 data, comp, uncompressed_size, compressed_size);
2554 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2555 BP_SET_TYPE(&dl->dr_overridden_by, type);
2556 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2557 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2559 dl->dr_override_state = DR_OVERRIDDEN;
2560 dl->dr_overridden_by.blk_birth = dr->dr_txg;
2564 dmu_buf_redact(dmu_buf_t *dbuf, dmu_tx_t *tx)
2566 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2567 dmu_object_type_t type;
2568 ASSERT(dsl_dataset_feature_is_active(db->db_objset->os_dsl_dataset,
2569 SPA_FEATURE_REDACTED_DATASETS));
2572 type = DB_DNODE(db)->dn_type;
2575 ASSERT0(db->db_level);
2576 dmu_buf_will_not_fill(dbuf, tx);
2578 blkptr_t bp = { { { {0} } } };
2579 BP_SET_TYPE(&bp, type);
2580 BP_SET_LEVEL(&bp, 0);
2581 BP_SET_BIRTH(&bp, tx->tx_txg, 0);
2582 BP_SET_REDACTED(&bp);
2583 BPE_SET_LSIZE(&bp, dbuf->db_size);
2585 dbuf_override_impl(db, &bp, tx);
2589 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2590 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2593 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2595 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2596 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2597 ASSERT(db->db_level == 0);
2598 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2599 ASSERT(buf != NULL);
2600 ASSERT3U(arc_buf_lsize(buf), ==, db->db.db_size);
2601 ASSERT(tx->tx_txg != 0);
2603 arc_return_buf(buf, db);
2604 ASSERT(arc_released(buf));
2606 mutex_enter(&db->db_mtx);
2608 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2609 cv_wait(&db->db_changed, &db->db_mtx);
2611 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2613 if (db->db_state == DB_CACHED &&
2614 zfs_refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2616 * In practice, we will never have a case where we have an
2617 * encrypted arc buffer while additional holds exist on the
2618 * dbuf. We don't handle this here so we simply assert that
2621 ASSERT(!arc_is_encrypted(buf));
2622 mutex_exit(&db->db_mtx);
2623 (void) dbuf_dirty(db, tx);
2624 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2625 arc_buf_destroy(buf, db);
2626 xuio_stat_wbuf_copied();
2630 xuio_stat_wbuf_nocopy();
2631 if (db->db_state == DB_CACHED) {
2632 dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records);
2634 ASSERT(db->db_buf != NULL);
2635 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2636 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2638 if (!arc_released(db->db_buf)) {
2639 ASSERT(dr->dt.dl.dr_override_state ==
2641 arc_release(db->db_buf, db);
2643 dr->dt.dl.dr_data = buf;
2644 arc_buf_destroy(db->db_buf, db);
2645 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2646 arc_release(db->db_buf, db);
2647 arc_buf_destroy(db->db_buf, db);
2651 ASSERT(db->db_buf == NULL);
2652 dbuf_set_data(db, buf);
2653 db->db_state = DB_FILL;
2654 DTRACE_SET_STATE(db, "filling assigned arcbuf");
2655 mutex_exit(&db->db_mtx);
2656 (void) dbuf_dirty(db, tx);
2657 dmu_buf_fill_done(&db->db, tx);
2661 dbuf_destroy(dmu_buf_impl_t *db)
2664 dmu_buf_impl_t *parent = db->db_parent;
2665 dmu_buf_impl_t *dndb;
2667 ASSERT(MUTEX_HELD(&db->db_mtx));
2668 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2670 if (db->db_buf != NULL) {
2671 arc_buf_destroy(db->db_buf, db);
2675 if (db->db_blkid == DMU_BONUS_BLKID) {
2676 int slots = DB_DNODE(db)->dn_num_slots;
2677 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2678 if (db->db.db_data != NULL) {
2679 kmem_free(db->db.db_data, bonuslen);
2680 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2681 db->db_state = DB_UNCACHED;
2682 DTRACE_SET_STATE(db, "buffer cleared");
2686 dbuf_clear_data(db);
2688 if (multilist_link_active(&db->db_cache_link)) {
2689 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
2690 db->db_caching_status == DB_DBUF_METADATA_CACHE);
2692 multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
2693 (void) zfs_refcount_remove_many(
2694 &dbuf_caches[db->db_caching_status].size,
2695 db->db.db_size, db);
2697 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2698 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2700 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2701 DBUF_STAT_BUMPDOWN(cache_count);
2702 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2705 db->db_caching_status = DB_NO_CACHE;
2708 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2709 ASSERT(db->db_data_pending == NULL);
2710 ASSERT(list_is_empty(&db->db_dirty_records));
2712 db->db_state = DB_EVICTING;
2713 DTRACE_SET_STATE(db, "buffer eviction started");
2714 db->db_blkptr = NULL;
2717 * Now that db_state is DB_EVICTING, nobody else can find this via
2718 * the hash table. We can now drop db_mtx, which allows us to
2719 * acquire the dn_dbufs_mtx.
2721 mutex_exit(&db->db_mtx);
2726 if (db->db_blkid != DMU_BONUS_BLKID) {
2727 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2729 mutex_enter_nested(&dn->dn_dbufs_mtx,
2731 avl_remove(&dn->dn_dbufs, db);
2735 mutex_exit(&dn->dn_dbufs_mtx);
2737 * Decrementing the dbuf count means that the hold corresponding
2738 * to the removed dbuf is no longer discounted in dnode_move(),
2739 * so the dnode cannot be moved until after we release the hold.
2740 * The membar_producer() ensures visibility of the decremented
2741 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2744 mutex_enter(&dn->dn_mtx);
2745 dnode_rele_and_unlock(dn, db, B_TRUE);
2746 db->db_dnode_handle = NULL;
2748 dbuf_hash_remove(db);
2753 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2755 db->db_parent = NULL;
2757 ASSERT(db->db_buf == NULL);
2758 ASSERT(db->db.db_data == NULL);
2759 ASSERT(db->db_hash_next == NULL);
2760 ASSERT(db->db_blkptr == NULL);
2761 ASSERT(db->db_data_pending == NULL);
2762 ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE);
2763 ASSERT(!multilist_link_active(&db->db_cache_link));
2765 kmem_cache_free(dbuf_kmem_cache, db);
2766 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2769 * If this dbuf is referenced from an indirect dbuf,
2770 * decrement the ref count on the indirect dbuf.
2772 if (parent && parent != dndb) {
2773 mutex_enter(&parent->db_mtx);
2774 dbuf_rele_and_unlock(parent, db, B_TRUE);
2779 * Note: While bpp will always be updated if the function returns success,
2780 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2781 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2784 __attribute__((always_inline))
2786 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2787 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2792 ASSERT(blkid != DMU_BONUS_BLKID);
2794 if (blkid == DMU_SPILL_BLKID) {
2795 mutex_enter(&dn->dn_mtx);
2796 if (dn->dn_have_spill &&
2797 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2798 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2801 dbuf_add_ref(dn->dn_dbuf, NULL);
2802 *parentp = dn->dn_dbuf;
2803 mutex_exit(&dn->dn_mtx);
2808 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2809 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2811 ASSERT3U(level * epbs, <, 64);
2812 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2814 * This assertion shouldn't trip as long as the max indirect block size
2815 * is less than 1M. The reason for this is that up to that point,
2816 * the number of levels required to address an entire object with blocks
2817 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2818 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2819 * (i.e. we can address the entire object), objects will all use at most
2820 * N-1 levels and the assertion won't overflow. However, once epbs is
2821 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2822 * enough to address an entire object, so objects will have 5 levels,
2823 * but then this assertion will overflow.
2825 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2826 * need to redo this logic to handle overflows.
2828 ASSERT(level >= nlevels ||
2829 ((nlevels - level - 1) * epbs) +
2830 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2831 if (level >= nlevels ||
2832 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2833 ((nlevels - level - 1) * epbs)) ||
2835 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2836 /* the buffer has no parent yet */
2837 return (SET_ERROR(ENOENT));
2838 } else if (level < nlevels-1) {
2839 /* this block is referenced from an indirect block */
2842 err = dbuf_hold_impl(dn, level + 1,
2843 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2847 err = dbuf_read(*parentp, NULL,
2848 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2850 dbuf_rele(*parentp, NULL);
2854 rw_enter(&(*parentp)->db_rwlock, RW_READER);
2855 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2856 (blkid & ((1ULL << epbs) - 1));
2857 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2858 ASSERT(BP_IS_HOLE(*bpp));
2859 rw_exit(&(*parentp)->db_rwlock);
2862 /* the block is referenced from the dnode */
2863 ASSERT3U(level, ==, nlevels-1);
2864 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2865 blkid < dn->dn_phys->dn_nblkptr);
2867 dbuf_add_ref(dn->dn_dbuf, NULL);
2868 *parentp = dn->dn_dbuf;
2870 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2875 static dmu_buf_impl_t *
2876 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2877 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2879 objset_t *os = dn->dn_objset;
2880 dmu_buf_impl_t *db, *odb;
2882 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2883 ASSERT(dn->dn_type != DMU_OT_NONE);
2885 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2887 list_create(&db->db_dirty_records, sizeof (dbuf_dirty_record_t),
2888 offsetof(dbuf_dirty_record_t, dr_dbuf_node));
2891 db->db.db_object = dn->dn_object;
2892 db->db_level = level;
2893 db->db_blkid = blkid;
2894 db->db_dirtycnt = 0;
2895 db->db_dnode_handle = dn->dn_handle;
2896 db->db_parent = parent;
2897 db->db_blkptr = blkptr;
2900 db->db_user_immediate_evict = FALSE;
2901 db->db_freed_in_flight = FALSE;
2902 db->db_pending_evict = FALSE;
2904 if (blkid == DMU_BONUS_BLKID) {
2905 ASSERT3P(parent, ==, dn->dn_dbuf);
2906 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2907 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2908 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2909 db->db.db_offset = DMU_BONUS_BLKID;
2910 db->db_state = DB_UNCACHED;
2911 DTRACE_SET_STATE(db, "bonus buffer created");
2912 db->db_caching_status = DB_NO_CACHE;
2913 /* the bonus dbuf is not placed in the hash table */
2914 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2916 } else if (blkid == DMU_SPILL_BLKID) {
2917 db->db.db_size = (blkptr != NULL) ?
2918 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2919 db->db.db_offset = 0;
2922 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2923 db->db.db_size = blocksize;
2924 db->db.db_offset = db->db_blkid * blocksize;
2928 * Hold the dn_dbufs_mtx while we get the new dbuf
2929 * in the hash table *and* added to the dbufs list.
2930 * This prevents a possible deadlock with someone
2931 * trying to look up this dbuf before it's added to the
2934 mutex_enter(&dn->dn_dbufs_mtx);
2935 db->db_state = DB_EVICTING; /* not worth logging this state change */
2936 if ((odb = dbuf_hash_insert(db)) != NULL) {
2937 /* someone else inserted it first */
2938 kmem_cache_free(dbuf_kmem_cache, db);
2939 mutex_exit(&dn->dn_dbufs_mtx);
2940 DBUF_STAT_BUMP(hash_insert_race);
2943 avl_add(&dn->dn_dbufs, db);
2945 db->db_state = DB_UNCACHED;
2946 DTRACE_SET_STATE(db, "regular buffer created");
2947 db->db_caching_status = DB_NO_CACHE;
2948 mutex_exit(&dn->dn_dbufs_mtx);
2949 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2951 if (parent && parent != dn->dn_dbuf)
2952 dbuf_add_ref(parent, db);
2954 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2955 zfs_refcount_count(&dn->dn_holds) > 0);
2956 (void) zfs_refcount_add(&dn->dn_holds, db);
2958 dprintf_dbuf(db, "db=%p\n", db);
2964 * This function returns a block pointer and information about the object,
2965 * given a dnode and a block. This is a publicly accessible version of
2966 * dbuf_findbp that only returns some information, rather than the
2967 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
2968 * should be locked as (at least) a reader.
2971 dbuf_dnode_findbp(dnode_t *dn, uint64_t level, uint64_t blkid,
2972 blkptr_t *bp, uint16_t *datablkszsec, uint8_t *indblkshift)
2974 dmu_buf_impl_t *dbp = NULL;
2977 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2979 err = dbuf_findbp(dn, level, blkid, B_FALSE, &dbp, &bp2);
2983 dbuf_rele(dbp, NULL);
2984 if (datablkszsec != NULL)
2985 *datablkszsec = dn->dn_phys->dn_datablkszsec;
2986 if (indblkshift != NULL)
2987 *indblkshift = dn->dn_phys->dn_indblkshift;
2993 typedef struct dbuf_prefetch_arg {
2994 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2995 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2996 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2997 int dpa_curlevel; /* The current level that we're reading */
2998 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2999 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
3000 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
3001 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
3002 } dbuf_prefetch_arg_t;
3005 * Actually issue the prefetch read for the block given.
3008 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
3010 ASSERT(!BP_IS_REDACTED(bp) ||
3011 dsl_dataset_feature_is_active(
3012 dpa->dpa_dnode->dn_objset->os_dsl_dataset,
3013 SPA_FEATURE_REDACTED_DATASETS));
3015 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp))
3018 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
3019 arc_flags_t aflags =
3020 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
3022 /* dnodes are always read as raw and then converted later */
3023 if (BP_GET_TYPE(bp) == DMU_OT_DNODE && BP_IS_PROTECTED(bp) &&
3024 dpa->dpa_curlevel == 0)
3025 zio_flags |= ZIO_FLAG_RAW;
3027 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
3028 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
3029 ASSERT(dpa->dpa_zio != NULL);
3030 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
3031 dpa->dpa_prio, zio_flags, &aflags, &dpa->dpa_zb);
3035 * Called when an indirect block above our prefetch target is read in. This
3036 * will either read in the next indirect block down the tree or issue the actual
3037 * prefetch if the next block down is our target.
3040 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
3041 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
3043 dbuf_prefetch_arg_t *dpa = private;
3045 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
3046 ASSERT3S(dpa->dpa_curlevel, >, 0);
3049 ASSERT(zio == NULL || zio->io_error != 0);
3050 kmem_free(dpa, sizeof (*dpa));
3053 ASSERT(zio == NULL || zio->io_error == 0);
3056 * The dpa_dnode is only valid if we are called with a NULL
3057 * zio. This indicates that the arc_read() returned without
3058 * first calling zio_read() to issue a physical read. Once
3059 * a physical read is made the dpa_dnode must be invalidated
3060 * as the locks guarding it may have been dropped. If the
3061 * dpa_dnode is still valid, then we want to add it to the dbuf
3062 * cache. To do so, we must hold the dbuf associated with the block
3063 * we just prefetched, read its contents so that we associate it
3064 * with an arc_buf_t, and then release it.
3067 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
3068 if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
3069 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
3071 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
3073 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
3075 dpa->dpa_dnode = NULL;
3076 } else if (dpa->dpa_dnode != NULL) {
3077 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
3078 (dpa->dpa_epbs * (dpa->dpa_curlevel -
3079 dpa->dpa_zb.zb_level));
3080 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
3081 dpa->dpa_curlevel, curblkid, FTAG);
3083 kmem_free(dpa, sizeof (*dpa));
3084 arc_buf_destroy(abuf, private);
3088 (void) dbuf_read(db, NULL,
3089 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
3090 dbuf_rele(db, FTAG);
3093 dpa->dpa_curlevel--;
3094 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
3095 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
3096 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
3097 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
3099 ASSERT(!BP_IS_REDACTED(bp) ||
3100 dsl_dataset_feature_is_active(
3101 dpa->dpa_dnode->dn_objset->os_dsl_dataset,
3102 SPA_FEATURE_REDACTED_DATASETS));
3103 if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
3104 kmem_free(dpa, sizeof (*dpa));
3105 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
3106 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
3107 dbuf_issue_final_prefetch(dpa, bp);
3108 kmem_free(dpa, sizeof (*dpa));
3110 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
3111 zbookmark_phys_t zb;
3113 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3114 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
3115 iter_aflags |= ARC_FLAG_L2CACHE;
3117 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
3119 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
3120 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
3122 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
3123 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
3124 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3128 arc_buf_destroy(abuf, private);
3132 * Issue prefetch reads for the given block on the given level. If the indirect
3133 * blocks above that block are not in memory, we will read them in
3134 * asynchronously. As a result, this call never blocks waiting for a read to
3135 * complete. Note that the prefetch might fail if the dataset is encrypted and
3136 * the encryption key is unmapped before the IO completes.
3139 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
3143 int epbs, nlevels, curlevel;
3146 ASSERT(blkid != DMU_BONUS_BLKID);
3147 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
3149 if (blkid > dn->dn_maxblkid)
3152 if (level == 0 && dnode_block_freed(dn, blkid))
3156 * This dnode hasn't been written to disk yet, so there's nothing to
3159 nlevels = dn->dn_phys->dn_nlevels;
3160 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
3163 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3164 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
3167 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
3170 mutex_exit(&db->db_mtx);
3172 * This dbuf already exists. It is either CACHED, or
3173 * (we assume) about to be read or filled.
3179 * Find the closest ancestor (indirect block) of the target block
3180 * that is present in the cache. In this indirect block, we will
3181 * find the bp that is at curlevel, curblkid.
3185 while (curlevel < nlevels - 1) {
3186 int parent_level = curlevel + 1;
3187 uint64_t parent_blkid = curblkid >> epbs;
3190 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
3191 FALSE, TRUE, FTAG, &db) == 0) {
3192 blkptr_t *bpp = db->db_buf->b_data;
3193 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
3194 dbuf_rele(db, FTAG);
3198 curlevel = parent_level;
3199 curblkid = parent_blkid;
3202 if (curlevel == nlevels - 1) {
3203 /* No cached indirect blocks found. */
3204 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
3205 bp = dn->dn_phys->dn_blkptr[curblkid];
3207 ASSERT(!BP_IS_REDACTED(&bp) ||
3208 dsl_dataset_feature_is_active(dn->dn_objset->os_dsl_dataset,
3209 SPA_FEATURE_REDACTED_DATASETS));
3210 if (BP_IS_HOLE(&bp) || BP_IS_REDACTED(&bp))
3213 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
3215 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
3218 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
3219 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
3220 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
3221 dn->dn_object, level, blkid);
3222 dpa->dpa_curlevel = curlevel;
3223 dpa->dpa_prio = prio;
3224 dpa->dpa_aflags = aflags;
3225 dpa->dpa_spa = dn->dn_objset->os_spa;
3226 dpa->dpa_dnode = dn;
3227 dpa->dpa_epbs = epbs;
3230 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3231 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
3232 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
3235 * If we have the indirect just above us, no need to do the asynchronous
3236 * prefetch chain; we'll just run the last step ourselves. If we're at
3237 * a higher level, though, we want to issue the prefetches for all the
3238 * indirect blocks asynchronously, so we can go on with whatever we were
3241 if (curlevel == level) {
3242 ASSERT3U(curblkid, ==, blkid);
3243 dbuf_issue_final_prefetch(dpa, &bp);
3244 kmem_free(dpa, sizeof (*dpa));
3246 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
3247 zbookmark_phys_t zb;
3249 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3250 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
3251 iter_aflags |= ARC_FLAG_L2CACHE;
3253 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
3254 dn->dn_object, curlevel, curblkid);
3255 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
3256 &bp, dbuf_prefetch_indirect_done, dpa, prio,
3257 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3261 * We use pio here instead of dpa_zio since it's possible that
3262 * dpa may have already been freed.
3268 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3269 * the case of encrypted, compressed and uncompressed buffers by
3270 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3271 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3273 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3275 noinline static void
3276 dbuf_hold_copy(dnode_t *dn, dmu_buf_impl_t *db)
3278 dbuf_dirty_record_t *dr = db->db_data_pending;
3279 arc_buf_t *newdata, *data = dr->dt.dl.dr_data;
3281 newdata = dbuf_alloc_arcbuf_from_arcbuf(db, data);
3282 dbuf_set_data(db, newdata);
3283 rw_enter(&db->db_rwlock, RW_WRITER);
3284 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
3285 rw_exit(&db->db_rwlock);
3289 * Returns with db_holds incremented, and db_mtx not held.
3290 * Note: dn_struct_rwlock must be held.
3293 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3294 boolean_t fail_sparse, boolean_t fail_uncached,
3295 void *tag, dmu_buf_impl_t **dbp)
3297 dmu_buf_impl_t *db, *parent = NULL;
3299 /* If the pool has been created, verify the tx_sync_lock is not held */
3300 spa_t *spa = dn->dn_objset->os_spa;
3301 dsl_pool_t *dp = spa->spa_dsl_pool;
3303 ASSERT(!MUTEX_HELD(&dp->dp_tx.tx_sync_lock));
3306 ASSERT(blkid != DMU_BONUS_BLKID);
3307 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
3308 ASSERT3U(dn->dn_nlevels, >, level);
3312 /* dbuf_find() returns with db_mtx held */
3313 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
3316 blkptr_t *bp = NULL;
3320 return (SET_ERROR(ENOENT));
3322 ASSERT3P(parent, ==, NULL);
3323 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
3325 if (err == 0 && bp && BP_IS_HOLE(bp))
3326 err = SET_ERROR(ENOENT);
3329 dbuf_rele(parent, NULL);
3333 if (err && err != ENOENT)
3335 db = dbuf_create(dn, level, blkid, parent, bp);
3338 if (fail_uncached && db->db_state != DB_CACHED) {
3339 mutex_exit(&db->db_mtx);
3340 return (SET_ERROR(ENOENT));
3343 if (db->db_buf != NULL) {
3344 arc_buf_access(db->db_buf);
3345 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
3348 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
3351 * If this buffer is currently syncing out, and we are
3352 * still referencing it from db_data, we need to make a copy
3353 * of it in case we decide we want to dirty it again in this txg.
3355 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
3356 dn->dn_object != DMU_META_DNODE_OBJECT &&
3357 db->db_state == DB_CACHED && db->db_data_pending) {
3358 dbuf_dirty_record_t *dr = db->db_data_pending;
3359 if (dr->dt.dl.dr_data == db->db_buf)
3360 dbuf_hold_copy(dn, db);
3363 if (multilist_link_active(&db->db_cache_link)) {
3364 ASSERT(zfs_refcount_is_zero(&db->db_holds));
3365 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
3366 db->db_caching_status == DB_DBUF_METADATA_CACHE);
3368 multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
3369 (void) zfs_refcount_remove_many(
3370 &dbuf_caches[db->db_caching_status].size,
3371 db->db.db_size, db);
3373 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
3374 DBUF_STAT_BUMPDOWN(metadata_cache_count);
3376 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
3377 DBUF_STAT_BUMPDOWN(cache_count);
3378 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
3381 db->db_caching_status = DB_NO_CACHE;
3383 (void) zfs_refcount_add(&db->db_holds, tag);
3385 mutex_exit(&db->db_mtx);
3387 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3389 dbuf_rele(parent, NULL);
3391 ASSERT3P(DB_DNODE(db), ==, dn);
3392 ASSERT3U(db->db_blkid, ==, blkid);
3393 ASSERT3U(db->db_level, ==, level);
3400 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3402 return (dbuf_hold_level(dn, 0, blkid, tag));
3406 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3409 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3410 return (err ? NULL : db);
3414 dbuf_create_bonus(dnode_t *dn)
3416 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3418 ASSERT(dn->dn_bonus == NULL);
3419 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3423 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3425 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3427 if (db->db_blkid != DMU_SPILL_BLKID)
3428 return (SET_ERROR(ENOTSUP));
3430 blksz = SPA_MINBLOCKSIZE;
3431 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3432 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3434 dbuf_new_size(db, blksz, tx);
3440 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3442 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3445 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3447 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3449 int64_t holds = zfs_refcount_add(&db->db_holds, tag);
3450 VERIFY3S(holds, >, 1);
3453 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3455 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3458 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3459 dmu_buf_impl_t *found_db;
3460 boolean_t result = B_FALSE;
3462 if (blkid == DMU_BONUS_BLKID)
3463 found_db = dbuf_find_bonus(os, obj);
3465 found_db = dbuf_find(os, obj, 0, blkid);
3467 if (found_db != NULL) {
3468 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3469 (void) zfs_refcount_add(&db->db_holds, tag);
3472 mutex_exit(&found_db->db_mtx);
3478 * If you call dbuf_rele() you had better not be referencing the dnode handle
3479 * unless you have some other direct or indirect hold on the dnode. (An indirect
3480 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3481 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3482 * dnode's parent dbuf evicting its dnode handles.
3485 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3487 mutex_enter(&db->db_mtx);
3488 dbuf_rele_and_unlock(db, tag, B_FALSE);
3492 dmu_buf_rele(dmu_buf_t *db, void *tag)
3494 dbuf_rele((dmu_buf_impl_t *)db, tag);
3498 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3499 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3500 * argument should be set if we are already in the dbuf-evicting code
3501 * path, in which case we don't want to recursively evict. This allows us to
3502 * avoid deeply nested stacks that would have a call flow similar to this:
3504 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3507 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3511 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
3516 ASSERT(MUTEX_HELD(&db->db_mtx));
3520 * Remove the reference to the dbuf before removing its hold on the
3521 * dnode so we can guarantee in dnode_move() that a referenced bonus
3522 * buffer has a corresponding dnode hold.
3524 holds = zfs_refcount_remove(&db->db_holds, tag);
3528 * We can't freeze indirects if there is a possibility that they
3529 * may be modified in the current syncing context.
3531 if (db->db_buf != NULL &&
3532 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3533 arc_buf_freeze(db->db_buf);
3536 if (holds == db->db_dirtycnt &&
3537 db->db_level == 0 && db->db_user_immediate_evict)
3538 dbuf_evict_user(db);
3541 if (db->db_blkid == DMU_BONUS_BLKID) {
3543 boolean_t evict_dbuf = db->db_pending_evict;
3546 * If the dnode moves here, we cannot cross this
3547 * barrier until the move completes.
3552 atomic_dec_32(&dn->dn_dbufs_count);
3555 * Decrementing the dbuf count means that the bonus
3556 * buffer's dnode hold is no longer discounted in
3557 * dnode_move(). The dnode cannot move until after
3558 * the dnode_rele() below.
3563 * Do not reference db after its lock is dropped.
3564 * Another thread may evict it.
3566 mutex_exit(&db->db_mtx);
3569 dnode_evict_bonus(dn);
3572 } else if (db->db_buf == NULL) {
3574 * This is a special case: we never associated this
3575 * dbuf with any data allocated from the ARC.
3577 ASSERT(db->db_state == DB_UNCACHED ||
3578 db->db_state == DB_NOFILL);
3580 } else if (arc_released(db->db_buf)) {
3582 * This dbuf has anonymous data associated with it.
3586 boolean_t do_arc_evict = B_FALSE;
3588 spa_t *spa = dmu_objset_spa(db->db_objset);
3590 if (!DBUF_IS_CACHEABLE(db) &&
3591 db->db_blkptr != NULL &&
3592 !BP_IS_HOLE(db->db_blkptr) &&
3593 !BP_IS_EMBEDDED(db->db_blkptr)) {
3594 do_arc_evict = B_TRUE;
3595 bp = *db->db_blkptr;
3598 if (!DBUF_IS_CACHEABLE(db) ||
3599 db->db_pending_evict) {
3601 } else if (!multilist_link_active(&db->db_cache_link)) {
3602 ASSERT3U(db->db_caching_status, ==,
3605 dbuf_cached_state_t dcs =
3606 dbuf_include_in_metadata_cache(db) ?
3607 DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
3608 db->db_caching_status = dcs;
3610 multilist_insert(dbuf_caches[dcs].cache, db);
3611 size = zfs_refcount_add_many(
3612 &dbuf_caches[dcs].size,
3613 db->db.db_size, db);
3615 if (dcs == DB_DBUF_METADATA_CACHE) {
3616 DBUF_STAT_BUMP(metadata_cache_count);
3618 metadata_cache_size_bytes_max,
3622 cache_levels[db->db_level]);
3623 DBUF_STAT_BUMP(cache_count);
3625 cache_levels_bytes[db->db_level],
3627 DBUF_STAT_MAX(cache_size_bytes_max,
3630 mutex_exit(&db->db_mtx);
3632 if (dcs == DB_DBUF_CACHE && !evicting)
3633 dbuf_evict_notify(size);
3637 arc_freed(spa, &bp);
3640 mutex_exit(&db->db_mtx);
3645 #pragma weak dmu_buf_refcount = dbuf_refcount
3647 dbuf_refcount(dmu_buf_impl_t *db)
3649 return (zfs_refcount_count(&db->db_holds));
3653 dmu_buf_user_refcount(dmu_buf_t *db_fake)
3656 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3658 mutex_enter(&db->db_mtx);
3659 ASSERT3U(zfs_refcount_count(&db->db_holds), >=, db->db_dirtycnt);
3660 holds = zfs_refcount_count(&db->db_holds) - db->db_dirtycnt;
3661 mutex_exit(&db->db_mtx);
3667 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3668 dmu_buf_user_t *new_user)
3670 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3672 mutex_enter(&db->db_mtx);
3673 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3674 if (db->db_user == old_user)
3675 db->db_user = new_user;
3677 old_user = db->db_user;
3678 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3679 mutex_exit(&db->db_mtx);
3685 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3687 return (dmu_buf_replace_user(db_fake, NULL, user));
3691 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3693 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3695 db->db_user_immediate_evict = TRUE;
3696 return (dmu_buf_set_user(db_fake, user));
3700 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3702 return (dmu_buf_replace_user(db_fake, user, NULL));
3706 dmu_buf_get_user(dmu_buf_t *db_fake)
3708 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3710 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3711 return (db->db_user);
3715 dmu_buf_user_evict_wait()
3717 taskq_wait(dbu_evict_taskq);
3721 dmu_buf_get_blkptr(dmu_buf_t *db)
3723 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3724 return (dbi->db_blkptr);
3728 dmu_buf_get_objset(dmu_buf_t *db)
3730 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3731 return (dbi->db_objset);
3735 dmu_buf_dnode_enter(dmu_buf_t *db)
3737 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3738 DB_DNODE_ENTER(dbi);
3739 return (DB_DNODE(dbi));
3743 dmu_buf_dnode_exit(dmu_buf_t *db)
3745 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3750 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3752 /* ASSERT(dmu_tx_is_syncing(tx) */
3753 ASSERT(MUTEX_HELD(&db->db_mtx));
3755 if (db->db_blkptr != NULL)
3758 if (db->db_blkid == DMU_SPILL_BLKID) {
3759 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3760 BP_ZERO(db->db_blkptr);
3763 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3765 * This buffer was allocated at a time when there was
3766 * no available blkptrs from the dnode, or it was
3767 * inappropriate to hook it in (i.e., nlevels mismatch).
3769 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3770 ASSERT(db->db_parent == NULL);
3771 db->db_parent = dn->dn_dbuf;
3772 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3775 dmu_buf_impl_t *parent = db->db_parent;
3776 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3778 ASSERT(dn->dn_phys->dn_nlevels > 1);
3779 if (parent == NULL) {
3780 mutex_exit(&db->db_mtx);
3781 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3782 parent = dbuf_hold_level(dn, db->db_level + 1,
3783 db->db_blkid >> epbs, db);
3784 rw_exit(&dn->dn_struct_rwlock);
3785 mutex_enter(&db->db_mtx);
3786 db->db_parent = parent;
3788 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3789 (db->db_blkid & ((1ULL << epbs) - 1));
3795 dbuf_sync_bonus(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3797 dmu_buf_impl_t *db = dr->dr_dbuf;
3798 void *data = dr->dt.dl.dr_data;
3800 ASSERT0(db->db_level);
3801 ASSERT(MUTEX_HELD(&db->db_mtx));
3802 ASSERT(DB_DNODE_HELD(db));
3803 ASSERT(db->db_blkid == DMU_BONUS_BLKID);
3804 ASSERT(data != NULL);
3806 dnode_t *dn = DB_DNODE(db);
3807 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3808 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3809 bcopy(data, DN_BONUS(dn->dn_phys), DN_MAX_BONUS_LEN(dn->dn_phys));
3812 dbuf_sync_leaf_verify_bonus_dnode(dr);
3814 dbuf_undirty_bonus(dr);
3815 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
3819 * When syncing out a blocks of dnodes, adjust the block to deal with
3820 * encryption. Normally, we make sure the block is decrypted before writing
3821 * it. If we have crypt params, then we are writing a raw (encrypted) block,
3822 * from a raw receive. In this case, set the ARC buf's crypt params so
3823 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
3826 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t *dr)
3829 dmu_buf_impl_t *db = dr->dr_dbuf;
3831 ASSERT(MUTEX_HELD(&db->db_mtx));
3832 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT);
3833 ASSERT3U(db->db_level, ==, 0);
3835 if (!db->db_objset->os_raw_receive && arc_is_encrypted(db->db_buf)) {
3836 zbookmark_phys_t zb;
3839 * Unfortunately, there is currently no mechanism for
3840 * syncing context to handle decryption errors. An error
3841 * here is only possible if an attacker maliciously
3842 * changed a dnode block and updated the associated
3843 * checksums going up the block tree.
3845 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
3846 db->db.db_object, db->db_level, db->db_blkid);
3847 err = arc_untransform(db->db_buf, db->db_objset->os_spa,
3850 panic("Invalid dnode block MAC");
3851 } else if (dr->dt.dl.dr_has_raw_params) {
3852 (void) arc_release(dr->dt.dl.dr_data, db);
3853 arc_convert_to_raw(dr->dt.dl.dr_data,
3854 dmu_objset_id(db->db_objset),
3855 dr->dt.dl.dr_byteorder, DMU_OT_DNODE,
3856 dr->dt.dl.dr_salt, dr->dt.dl.dr_iv, dr->dt.dl.dr_mac);
3861 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3862 * is critical the we not allow the compiler to inline this function in to
3863 * dbuf_sync_list() thereby drastically bloating the stack usage.
3865 noinline static void
3866 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3868 dmu_buf_impl_t *db = dr->dr_dbuf;
3872 ASSERT(dmu_tx_is_syncing(tx));
3874 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3876 mutex_enter(&db->db_mtx);
3878 ASSERT(db->db_level > 0);
3881 /* Read the block if it hasn't been read yet. */
3882 if (db->db_buf == NULL) {
3883 mutex_exit(&db->db_mtx);
3884 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3885 mutex_enter(&db->db_mtx);
3887 ASSERT3U(db->db_state, ==, DB_CACHED);
3888 ASSERT(db->db_buf != NULL);
3892 /* Indirect block size must match what the dnode thinks it is. */
3893 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3894 dbuf_check_blkptr(dn, db);
3897 /* Provide the pending dirty record to child dbufs */
3898 db->db_data_pending = dr;
3900 mutex_exit(&db->db_mtx);
3902 dbuf_write(dr, db->db_buf, tx);
3905 mutex_enter(&dr->dt.di.dr_mtx);
3906 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3907 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3908 mutex_exit(&dr->dt.di.dr_mtx);
3913 * Verify that the size of the data in our bonus buffer does not exceed
3914 * its recorded size.
3916 * The purpose of this verification is to catch any cases in development
3917 * where the size of a phys structure (i.e space_map_phys_t) grows and,
3918 * due to incorrect feature management, older pools expect to read more
3919 * data even though they didn't actually write it to begin with.
3921 * For a example, this would catch an error in the feature logic where we
3922 * open an older pool and we expect to write the space map histogram of
3923 * a space map with size SPACE_MAP_SIZE_V0.
3926 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t *dr)
3929 dnode_t *dn = DB_DNODE(dr->dr_dbuf);
3932 * Encrypted bonus buffers can have data past their bonuslen.
3933 * Skip the verification of these blocks.
3935 if (DMU_OT_IS_ENCRYPTED(dn->dn_bonustype))
3938 uint16_t bonuslen = dn->dn_phys->dn_bonuslen;
3939 uint16_t maxbonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
3940 ASSERT3U(bonuslen, <=, maxbonuslen);
3942 arc_buf_t *datap = dr->dt.dl.dr_data;
3943 char *datap_end = ((char *)datap) + bonuslen;
3944 char *datap_max = ((char *)datap) + maxbonuslen;
3946 /* ensure that everything is zero after our data */
3947 for (; datap_end < datap_max; datap_end++)
3948 ASSERT(*datap_end == 0);
3953 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3954 * critical the we not allow the compiler to inline this function in to
3955 * dbuf_sync_list() thereby drastically bloating the stack usage.
3957 noinline static void
3958 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3960 arc_buf_t **datap = &dr->dt.dl.dr_data;
3961 dmu_buf_impl_t *db = dr->dr_dbuf;
3964 uint64_t txg = tx->tx_txg;
3966 ASSERT(dmu_tx_is_syncing(tx));
3968 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3970 mutex_enter(&db->db_mtx);
3972 * To be synced, we must be dirtied. But we
3973 * might have been freed after the dirty.
3975 if (db->db_state == DB_UNCACHED) {
3976 /* This buffer has been freed since it was dirtied */
3977 ASSERT(db->db.db_data == NULL);
3978 } else if (db->db_state == DB_FILL) {
3979 /* This buffer was freed and is now being re-filled */
3980 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3982 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3989 if (db->db_blkid == DMU_SPILL_BLKID) {
3990 mutex_enter(&dn->dn_mtx);
3991 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3993 * In the previous transaction group, the bonus buffer
3994 * was entirely used to store the attributes for the
3995 * dnode which overrode the dn_spill field. However,
3996 * when adding more attributes to the file a spill
3997 * block was required to hold the extra attributes.
3999 * Make sure to clear the garbage left in the dn_spill
4000 * field from the previous attributes in the bonus
4001 * buffer. Otherwise, after writing out the spill
4002 * block to the new allocated dva, it will free
4003 * the old block pointed to by the invalid dn_spill.
4005 db->db_blkptr = NULL;
4007 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
4008 mutex_exit(&dn->dn_mtx);
4012 * If this is a bonus buffer, simply copy the bonus data into the
4013 * dnode. It will be written out when the dnode is synced (and it
4014 * will be synced, since it must have been dirty for dbuf_sync to
4017 if (db->db_blkid == DMU_BONUS_BLKID) {
4018 ASSERT(dr->dr_dbuf == db);
4019 dbuf_sync_bonus(dr, tx);
4026 * This function may have dropped the db_mtx lock allowing a dmu_sync
4027 * operation to sneak in. As a result, we need to ensure that we
4028 * don't check the dr_override_state until we have returned from
4029 * dbuf_check_blkptr.
4031 dbuf_check_blkptr(dn, db);
4034 * If this buffer is in the middle of an immediate write,
4035 * wait for the synchronous IO to complete.
4037 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
4038 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
4039 cv_wait(&db->db_changed, &db->db_mtx);
4040 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
4044 * If this is a dnode block, ensure it is appropriately encrypted
4045 * or decrypted, depending on what we are writing to it this txg.
4047 if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT)
4048 dbuf_prepare_encrypted_dnode_leaf(dr);
4050 if (db->db_state != DB_NOFILL &&
4051 dn->dn_object != DMU_META_DNODE_OBJECT &&
4052 zfs_refcount_count(&db->db_holds) > 1 &&
4053 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
4054 *datap == db->db_buf) {
4056 * If this buffer is currently "in use" (i.e., there
4057 * are active holds and db_data still references it),
4058 * then make a copy before we start the write so that
4059 * any modifications from the open txg will not leak
4062 * NOTE: this copy does not need to be made for
4063 * objects only modified in the syncing context (e.g.
4064 * DNONE_DNODE blocks).
4066 *datap = dbuf_alloc_arcbuf_from_arcbuf(db, db->db_buf);
4067 bcopy(db->db.db_data, (*datap)->b_data, arc_buf_size(*datap));
4069 db->db_data_pending = dr;
4071 mutex_exit(&db->db_mtx);
4073 dbuf_write(dr, *datap, tx);
4075 ASSERT(!list_link_active(&dr->dr_dirty_node));
4076 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
4077 list_insert_tail(&dn->dn_dirty_records[txg & TXG_MASK], dr);
4081 * Although zio_nowait() does not "wait for an IO", it does
4082 * initiate the IO. If this is an empty write it seems plausible
4083 * that the IO could actually be completed before the nowait
4084 * returns. We need to DB_DNODE_EXIT() first in case
4085 * zio_nowait() invalidates the dbuf.
4088 zio_nowait(dr->dr_zio);
4093 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
4095 dbuf_dirty_record_t *dr;
4097 while ((dr = list_head(list))) {
4098 if (dr->dr_zio != NULL) {
4100 * If we find an already initialized zio then we
4101 * are processing the meta-dnode, and we have finished.
4102 * The dbufs for all dnodes are put back on the list
4103 * during processing, so that we can zio_wait()
4104 * these IOs after initiating all child IOs.
4106 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
4107 DMU_META_DNODE_OBJECT);
4110 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
4111 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
4112 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
4114 list_remove(list, dr);
4115 if (dr->dr_dbuf->db_level > 0)
4116 dbuf_sync_indirect(dr, tx);
4118 dbuf_sync_leaf(dr, tx);
4124 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
4126 dmu_buf_impl_t *db = vdb;
4128 blkptr_t *bp = zio->io_bp;
4129 blkptr_t *bp_orig = &zio->io_bp_orig;
4130 spa_t *spa = zio->io_spa;
4135 ASSERT3P(db->db_blkptr, !=, NULL);
4136 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
4140 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
4141 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
4142 zio->io_prev_space_delta = delta;
4144 if (bp->blk_birth != 0) {
4145 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
4146 BP_GET_TYPE(bp) == dn->dn_type) ||
4147 (db->db_blkid == DMU_SPILL_BLKID &&
4148 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
4149 BP_IS_EMBEDDED(bp));
4150 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
4153 mutex_enter(&db->db_mtx);
4156 if (db->db_blkid == DMU_SPILL_BLKID) {
4157 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
4158 ASSERT(!(BP_IS_HOLE(bp)) &&
4159 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
4163 if (db->db_level == 0) {
4164 mutex_enter(&dn->dn_mtx);
4165 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
4166 db->db_blkid != DMU_SPILL_BLKID) {
4167 ASSERT0(db->db_objset->os_raw_receive);
4168 dn->dn_phys->dn_maxblkid = db->db_blkid;
4170 mutex_exit(&dn->dn_mtx);
4172 if (dn->dn_type == DMU_OT_DNODE) {
4174 while (i < db->db.db_size) {
4176 (void *)(((char *)db->db.db_data) + i);
4178 i += DNODE_MIN_SIZE;
4179 if (dnp->dn_type != DMU_OT_NONE) {
4181 i += dnp->dn_extra_slots *
4186 if (BP_IS_HOLE(bp)) {
4193 blkptr_t *ibp = db->db.db_data;
4194 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
4195 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
4196 if (BP_IS_HOLE(ibp))
4198 fill += BP_GET_FILL(ibp);
4203 if (!BP_IS_EMBEDDED(bp))
4204 BP_SET_FILL(bp, fill);
4206 mutex_exit(&db->db_mtx);
4208 db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_WRITER, FTAG);
4209 *db->db_blkptr = *bp;
4210 dmu_buf_unlock_parent(db, dblt, FTAG);
4215 * This function gets called just prior to running through the compression
4216 * stage of the zio pipeline. If we're an indirect block comprised of only
4217 * holes, then we want this indirect to be compressed away to a hole. In
4218 * order to do that we must zero out any information about the holes that
4219 * this indirect points to prior to before we try to compress it.
4222 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
4224 dmu_buf_impl_t *db = vdb;
4227 unsigned int epbs, i;
4229 ASSERT3U(db->db_level, >, 0);
4232 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
4233 ASSERT3U(epbs, <, 31);
4235 /* Determine if all our children are holes */
4236 for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) {
4237 if (!BP_IS_HOLE(bp))
4242 * If all the children are holes, then zero them all out so that
4243 * we may get compressed away.
4245 if (i == 1ULL << epbs) {
4247 * We only found holes. Grab the rwlock to prevent
4248 * anybody from reading the blocks we're about to
4251 rw_enter(&db->db_rwlock, RW_WRITER);
4252 bzero(db->db.db_data, db->db.db_size);
4253 rw_exit(&db->db_rwlock);
4259 * The SPA will call this callback several times for each zio - once
4260 * for every physical child i/o (zio->io_phys_children times). This
4261 * allows the DMU to monitor the progress of each logical i/o. For example,
4262 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
4263 * block. There may be a long delay before all copies/fragments are completed,
4264 * so this callback allows us to retire dirty space gradually, as the physical
4269 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
4271 dmu_buf_impl_t *db = arg;
4272 objset_t *os = db->db_objset;
4273 dsl_pool_t *dp = dmu_objset_pool(os);
4274 dbuf_dirty_record_t *dr;
4277 dr = db->db_data_pending;
4278 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
4281 * The callback will be called io_phys_children times. Retire one
4282 * portion of our dirty space each time we are called. Any rounding
4283 * error will be cleaned up by dbuf_write_done().
4285 delta = dr->dr_accounted / zio->io_phys_children;
4286 dsl_pool_undirty_space(dp, delta, zio->io_txg);
4291 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
4293 dmu_buf_impl_t *db = vdb;
4294 blkptr_t *bp_orig = &zio->io_bp_orig;
4295 blkptr_t *bp = db->db_blkptr;
4296 objset_t *os = db->db_objset;
4297 dmu_tx_t *tx = os->os_synctx;
4298 dbuf_dirty_record_t *dr;
4300 ASSERT0(zio->io_error);
4301 ASSERT(db->db_blkptr == bp);
4304 * For nopwrites and rewrites we ensure that the bp matches our
4305 * original and bypass all the accounting.
4307 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
4308 ASSERT(BP_EQUAL(bp, bp_orig));
4310 dsl_dataset_t *ds = os->os_dsl_dataset;
4311 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
4312 dsl_dataset_block_born(ds, bp, tx);
4315 mutex_enter(&db->db_mtx);
4319 dr = db->db_data_pending;
4320 ASSERT(!list_link_active(&dr->dr_dirty_node));
4321 ASSERT(dr->dr_dbuf == db);
4322 ASSERT(list_next(&db->db_dirty_records, dr) == NULL);
4323 list_remove(&db->db_dirty_records, dr);
4326 if (db->db_blkid == DMU_SPILL_BLKID) {
4331 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
4332 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
4333 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
4338 if (db->db_level == 0) {
4339 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
4340 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
4341 if (db->db_state != DB_NOFILL) {
4342 if (dr->dt.dl.dr_data != db->db_buf)
4343 arc_buf_destroy(dr->dt.dl.dr_data, db);
4350 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
4351 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
4352 if (!BP_IS_HOLE(db->db_blkptr)) {
4353 int epbs __maybe_unused = dn->dn_phys->dn_indblkshift -
4355 ASSERT3U(db->db_blkid, <=,
4356 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
4357 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
4361 mutex_destroy(&dr->dt.di.dr_mtx);
4362 list_destroy(&dr->dt.di.dr_children);
4365 cv_broadcast(&db->db_changed);
4366 ASSERT(db->db_dirtycnt > 0);
4367 db->db_dirtycnt -= 1;
4368 db->db_data_pending = NULL;
4369 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
4372 * If we didn't do a physical write in this ZIO and we
4373 * still ended up here, it means that the space of the
4374 * dbuf that we just released (and undirtied) above hasn't
4375 * been marked as undirtied in the pool's accounting.
4377 * Thus, we undirty that space in the pool's view of the
4378 * world here. For physical writes this type of update
4379 * happens in dbuf_write_physdone().
4381 * If we did a physical write, cleanup any rounding errors
4382 * that came up due to writing multiple copies of a block
4383 * on disk [see dbuf_write_physdone()].
4385 if (zio->io_phys_children == 0) {
4386 dsl_pool_undirty_space(dmu_objset_pool(os),
4387 dr->dr_accounted, zio->io_txg);
4389 dsl_pool_undirty_space(dmu_objset_pool(os),
4390 dr->dr_accounted % zio->io_phys_children, zio->io_txg);
4393 kmem_free(dr, sizeof (dbuf_dirty_record_t));
4397 dbuf_write_nofill_ready(zio_t *zio)
4399 dbuf_write_ready(zio, NULL, zio->io_private);
4403 dbuf_write_nofill_done(zio_t *zio)
4405 dbuf_write_done(zio, NULL, zio->io_private);
4409 dbuf_write_override_ready(zio_t *zio)
4411 dbuf_dirty_record_t *dr = zio->io_private;
4412 dmu_buf_impl_t *db = dr->dr_dbuf;
4414 dbuf_write_ready(zio, NULL, db);
4418 dbuf_write_override_done(zio_t *zio)
4420 dbuf_dirty_record_t *dr = zio->io_private;
4421 dmu_buf_impl_t *db = dr->dr_dbuf;
4422 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
4424 mutex_enter(&db->db_mtx);
4425 if (!BP_EQUAL(zio->io_bp, obp)) {
4426 if (!BP_IS_HOLE(obp))
4427 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
4428 arc_release(dr->dt.dl.dr_data, db);
4430 mutex_exit(&db->db_mtx);
4432 dbuf_write_done(zio, NULL, db);
4434 if (zio->io_abd != NULL)
4435 abd_put(zio->io_abd);
4438 typedef struct dbuf_remap_impl_callback_arg {
4440 uint64_t drica_blk_birth;
4442 } dbuf_remap_impl_callback_arg_t;
4445 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
4448 dbuf_remap_impl_callback_arg_t *drica = arg;
4449 objset_t *os = drica->drica_os;
4450 spa_t *spa = dmu_objset_spa(os);
4451 dmu_tx_t *tx = drica->drica_tx;
4453 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4455 if (os == spa_meta_objset(spa)) {
4456 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
4458 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
4459 size, drica->drica_blk_birth, tx);
4464 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, krwlock_t *rw, dmu_tx_t *tx)
4466 blkptr_t bp_copy = *bp;
4467 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4468 dbuf_remap_impl_callback_arg_t drica;
4470 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4472 drica.drica_os = dn->dn_objset;
4473 drica.drica_blk_birth = bp->blk_birth;
4474 drica.drica_tx = tx;
4475 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
4478 * If the blkptr being remapped is tracked by a livelist,
4479 * then we need to make sure the livelist reflects the update.
4480 * First, cancel out the old blkptr by appending a 'FREE'
4481 * entry. Next, add an 'ALLOC' to track the new version. This
4482 * way we avoid trying to free an inaccurate blkptr at delete.
4483 * Note that embedded blkptrs are not tracked in livelists.
4485 if (dn->dn_objset != spa_meta_objset(spa)) {
4486 dsl_dataset_t *ds = dmu_objset_ds(dn->dn_objset);
4487 if (dsl_deadlist_is_open(&ds->ds_dir->dd_livelist) &&
4488 bp->blk_birth > ds->ds_dir->dd_origin_txg) {
4489 ASSERT(!BP_IS_EMBEDDED(bp));
4490 ASSERT(dsl_dir_is_clone(ds->ds_dir));
4491 ASSERT(spa_feature_is_enabled(spa,
4492 SPA_FEATURE_LIVELIST));
4493 bplist_append(&ds->ds_dir->dd_pending_frees,
4495 bplist_append(&ds->ds_dir->dd_pending_allocs,
4501 * The db_rwlock prevents dbuf_read_impl() from
4502 * dereferencing the BP while we are changing it. To
4503 * avoid lock contention, only grab it when we are actually
4507 rw_enter(rw, RW_WRITER);
4515 * Remap any existing BP's to concrete vdevs, if possible.
4518 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
4520 spa_t *spa = dmu_objset_spa(db->db_objset);
4521 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4523 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
4526 if (db->db_level > 0) {
4527 blkptr_t *bp = db->db.db_data;
4528 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4529 dbuf_remap_impl(dn, &bp[i], &db->db_rwlock, tx);
4531 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
4532 dnode_phys_t *dnp = db->db.db_data;
4533 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
4535 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT;
4536 i += dnp[i].dn_extra_slots + 1) {
4537 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
4538 krwlock_t *lock = (dn->dn_dbuf == NULL ? NULL :
4539 &dn->dn_dbuf->db_rwlock);
4540 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], lock,
4548 /* Issue I/O to commit a dirty buffer to disk. */
4550 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4552 dmu_buf_impl_t *db = dr->dr_dbuf;
4555 dmu_buf_impl_t *parent = db->db_parent;
4556 uint64_t txg = tx->tx_txg;
4557 zbookmark_phys_t zb;
4559 zio_t *pio; /* parent I/O */
4562 ASSERT(dmu_tx_is_syncing(tx));
4568 if (db->db_state != DB_NOFILL) {
4569 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4571 * Private object buffers are released here rather
4572 * than in dbuf_dirty() since they are only modified
4573 * in the syncing context and we don't want the
4574 * overhead of making multiple copies of the data.
4576 if (BP_IS_HOLE(db->db_blkptr)) {
4579 dbuf_release_bp(db);
4581 dbuf_remap(dn, db, tx);
4585 if (parent != dn->dn_dbuf) {
4586 /* Our parent is an indirect block. */
4587 /* We have a dirty parent that has been scheduled for write. */
4588 ASSERT(parent && parent->db_data_pending);
4589 /* Our parent's buffer is one level closer to the dnode. */
4590 ASSERT(db->db_level == parent->db_level-1);
4592 * We're about to modify our parent's db_data by modifying
4593 * our block pointer, so the parent must be released.
4595 ASSERT(arc_released(parent->db_buf));
4596 pio = parent->db_data_pending->dr_zio;
4598 /* Our parent is the dnode itself. */
4599 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4600 db->db_blkid != DMU_SPILL_BLKID) ||
4601 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4602 if (db->db_blkid != DMU_SPILL_BLKID)
4603 ASSERT3P(db->db_blkptr, ==,
4604 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4608 ASSERT(db->db_level == 0 || data == db->db_buf);
4609 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4612 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4613 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4614 db->db.db_object, db->db_level, db->db_blkid);
4616 if (db->db_blkid == DMU_SPILL_BLKID)
4618 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
4620 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
4624 * We copy the blkptr now (rather than when we instantiate the dirty
4625 * record), because its value can change between open context and
4626 * syncing context. We do not need to hold dn_struct_rwlock to read
4627 * db_blkptr because we are in syncing context.
4629 dr->dr_bp_copy = *db->db_blkptr;
4631 if (db->db_level == 0 &&
4632 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
4634 * The BP for this block has been provided by open context
4635 * (by dmu_sync() or dmu_buf_write_embedded()).
4637 abd_t *contents = (data != NULL) ?
4638 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
4640 dr->dr_zio = zio_write(pio, os->os_spa, txg, &dr->dr_bp_copy,
4641 contents, db->db.db_size, db->db.db_size, &zp,
4642 dbuf_write_override_ready, NULL, NULL,
4643 dbuf_write_override_done,
4644 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
4645 mutex_enter(&db->db_mtx);
4646 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
4647 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
4648 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
4649 mutex_exit(&db->db_mtx);
4650 } else if (db->db_state == DB_NOFILL) {
4651 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
4652 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
4653 dr->dr_zio = zio_write(pio, os->os_spa, txg,
4654 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
4655 dbuf_write_nofill_ready, NULL, NULL,
4656 dbuf_write_nofill_done, db,
4657 ZIO_PRIORITY_ASYNC_WRITE,
4658 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
4660 ASSERT(arc_released(data));
4663 * For indirect blocks, we want to setup the children
4664 * ready callback so that we can properly handle an indirect
4665 * block that only contains holes.
4667 arc_write_done_func_t *children_ready_cb = NULL;
4668 if (db->db_level != 0)
4669 children_ready_cb = dbuf_write_children_ready;
4671 dr->dr_zio = arc_write(pio, os->os_spa, txg,
4672 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
4673 &zp, dbuf_write_ready,
4674 children_ready_cb, dbuf_write_physdone,
4675 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
4676 ZIO_FLAG_MUSTSUCCEED, &zb);
4680 EXPORT_SYMBOL(dbuf_find);
4681 EXPORT_SYMBOL(dbuf_is_metadata);
4682 EXPORT_SYMBOL(dbuf_destroy);
4683 EXPORT_SYMBOL(dbuf_loan_arcbuf);
4684 EXPORT_SYMBOL(dbuf_whichblock);
4685 EXPORT_SYMBOL(dbuf_read);
4686 EXPORT_SYMBOL(dbuf_unoverride);
4687 EXPORT_SYMBOL(dbuf_free_range);
4688 EXPORT_SYMBOL(dbuf_new_size);
4689 EXPORT_SYMBOL(dbuf_release_bp);
4690 EXPORT_SYMBOL(dbuf_dirty);
4691 EXPORT_SYMBOL(dmu_buf_set_crypt_params);
4692 EXPORT_SYMBOL(dmu_buf_will_dirty);
4693 EXPORT_SYMBOL(dmu_buf_is_dirty);
4694 EXPORT_SYMBOL(dmu_buf_will_not_fill);
4695 EXPORT_SYMBOL(dmu_buf_will_fill);
4696 EXPORT_SYMBOL(dmu_buf_fill_done);
4697 EXPORT_SYMBOL(dmu_buf_rele);
4698 EXPORT_SYMBOL(dbuf_assign_arcbuf);
4699 EXPORT_SYMBOL(dbuf_prefetch);
4700 EXPORT_SYMBOL(dbuf_hold_impl);
4701 EXPORT_SYMBOL(dbuf_hold);
4702 EXPORT_SYMBOL(dbuf_hold_level);
4703 EXPORT_SYMBOL(dbuf_create_bonus);
4704 EXPORT_SYMBOL(dbuf_spill_set_blksz);
4705 EXPORT_SYMBOL(dbuf_rm_spill);
4706 EXPORT_SYMBOL(dbuf_add_ref);
4707 EXPORT_SYMBOL(dbuf_rele);
4708 EXPORT_SYMBOL(dbuf_rele_and_unlock);
4709 EXPORT_SYMBOL(dbuf_refcount);
4710 EXPORT_SYMBOL(dbuf_sync_list);
4711 EXPORT_SYMBOL(dmu_buf_set_user);
4712 EXPORT_SYMBOL(dmu_buf_set_user_ie);
4713 EXPORT_SYMBOL(dmu_buf_get_user);
4714 EXPORT_SYMBOL(dmu_buf_get_blkptr);
4717 ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, max_bytes, ULONG, ZMOD_RW,
4718 "Maximum size in bytes of the dbuf cache.");
4720 ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, hiwater_pct, UINT, ZMOD_RW,
4721 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4724 ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, lowater_pct, UINT, ZMOD_RW,
4725 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4728 ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_max_bytes, ULONG, ZMOD_RW,
4729 "Maximum size in bytes of the dbuf metadata cache.");
4731 ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, cache_shift, INT, ZMOD_RW,
4732 "Set the size of the dbuf cache to a log2 fraction of arc size.");
4734 ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_shift, INT, ZMOD_RW,
4735 "Set the size of the dbuf metadata cache to a log2 fraction of arc "