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, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
42 #include <sys/dmu_zfetch.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
51 #include <sys/cityhash.h>
52 #include <sys/spa_impl.h>
56 typedef struct dbuf_stats {
58 * Various statistics about the size of the dbuf cache.
60 kstat_named_t cache_count;
61 kstat_named_t cache_size_bytes;
62 kstat_named_t cache_size_bytes_max;
64 * Statistics regarding the bounds on the dbuf cache size.
66 kstat_named_t cache_target_bytes;
67 kstat_named_t cache_lowater_bytes;
68 kstat_named_t cache_hiwater_bytes;
70 * Total number of dbuf cache evictions that have occurred.
72 kstat_named_t cache_total_evicts;
74 * The distribution of dbuf levels in the dbuf cache and
75 * the total size of all dbufs at each level.
77 kstat_named_t cache_levels[DN_MAX_LEVELS];
78 kstat_named_t cache_levels_bytes[DN_MAX_LEVELS];
80 * Statistics about the dbuf hash table.
82 kstat_named_t hash_hits;
83 kstat_named_t hash_misses;
84 kstat_named_t hash_collisions;
85 kstat_named_t hash_elements;
86 kstat_named_t hash_elements_max;
88 * Number of sublists containing more than one dbuf in the dbuf
89 * hash table. Keep track of the longest hash chain.
91 kstat_named_t hash_chains;
92 kstat_named_t hash_chain_max;
94 * Number of times a dbuf_create() discovers that a dbuf was
95 * already created and in the dbuf hash table.
97 kstat_named_t hash_insert_race;
99 * Statistics about the size of the metadata dbuf cache.
101 kstat_named_t metadata_cache_count;
102 kstat_named_t metadata_cache_size_bytes;
103 kstat_named_t metadata_cache_size_bytes_max;
105 * For diagnostic purposes, this is incremented whenever we can't add
106 * something to the metadata cache because it's full, and instead put
107 * the data in the regular dbuf cache.
109 kstat_named_t metadata_cache_overflow;
112 dbuf_stats_t dbuf_stats = {
113 { "cache_count", KSTAT_DATA_UINT64 },
114 { "cache_size_bytes", KSTAT_DATA_UINT64 },
115 { "cache_size_bytes_max", KSTAT_DATA_UINT64 },
116 { "cache_target_bytes", KSTAT_DATA_UINT64 },
117 { "cache_lowater_bytes", KSTAT_DATA_UINT64 },
118 { "cache_hiwater_bytes", KSTAT_DATA_UINT64 },
119 { "cache_total_evicts", KSTAT_DATA_UINT64 },
120 { { "cache_levels_N", KSTAT_DATA_UINT64 } },
121 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } },
122 { "hash_hits", KSTAT_DATA_UINT64 },
123 { "hash_misses", KSTAT_DATA_UINT64 },
124 { "hash_collisions", KSTAT_DATA_UINT64 },
125 { "hash_elements", KSTAT_DATA_UINT64 },
126 { "hash_elements_max", KSTAT_DATA_UINT64 },
127 { "hash_chains", KSTAT_DATA_UINT64 },
128 { "hash_chain_max", KSTAT_DATA_UINT64 },
129 { "hash_insert_race", KSTAT_DATA_UINT64 },
130 { "metadata_cache_count", KSTAT_DATA_UINT64 },
131 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64 },
132 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64 },
133 { "metadata_cache_overflow", KSTAT_DATA_UINT64 }
136 #define DBUF_STAT_INCR(stat, val) \
137 atomic_add_64(&dbuf_stats.stat.value.ui64, (val));
138 #define DBUF_STAT_DECR(stat, val) \
139 DBUF_STAT_INCR(stat, -(val));
140 #define DBUF_STAT_BUMP(stat) \
141 DBUF_STAT_INCR(stat, 1);
142 #define DBUF_STAT_BUMPDOWN(stat) \
143 DBUF_STAT_INCR(stat, -1);
144 #define DBUF_STAT_MAX(stat, v) { \
146 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
147 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
151 struct dbuf_hold_impl_data {
152 /* Function arguments */
156 boolean_t dh_fail_sparse;
157 boolean_t dh_fail_uncached;
159 dmu_buf_impl_t **dh_dbp;
160 /* Local variables */
161 dmu_buf_impl_t *dh_db;
162 dmu_buf_impl_t *dh_parent;
165 dbuf_dirty_record_t *dh_dr;
169 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
170 dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse,
171 boolean_t fail_uncached,
172 void *tag, dmu_buf_impl_t **dbp, int depth);
173 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
175 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
176 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
179 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
180 dmu_buf_evict_func_t *evict_func_sync,
181 dmu_buf_evict_func_t *evict_func_async,
182 dmu_buf_t **clear_on_evict_dbufp);
183 #endif /* ! __lint */
186 * Global data structures and functions for the dbuf cache.
188 static kmem_cache_t *dbuf_kmem_cache;
189 static taskq_t *dbu_evict_taskq;
191 static kthread_t *dbuf_cache_evict_thread;
192 static kmutex_t dbuf_evict_lock;
193 static kcondvar_t dbuf_evict_cv;
194 static boolean_t dbuf_evict_thread_exit;
197 * There are two dbuf caches; each dbuf can only be in one of them at a time.
199 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
200 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
201 * that represent the metadata that describes filesystems/snapshots/
202 * bookmarks/properties/etc. We only evict from this cache when we export a
203 * pool, to short-circuit as much I/O as possible for all administrative
204 * commands that need the metadata. There is no eviction policy for this
205 * cache, because we try to only include types in it which would occupy a
206 * very small amount of space per object but create a large impact on the
207 * performance of these commands. Instead, after it reaches a maximum size
208 * (which should only happen on very small memory systems with a very large
209 * number of filesystem objects), we stop taking new dbufs into the
210 * metadata cache, instead putting them in the normal dbuf cache.
212 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
213 * are not currently held but have been recently released. These dbufs
214 * are not eligible for arc eviction until they are aged out of the cache.
215 * Dbufs that are aged out of the cache will be immediately destroyed and
216 * become eligible for arc eviction.
218 * Dbufs are added to these caches once the last hold is released. If a dbuf is
219 * later accessed and still exists in the dbuf cache, then it will be removed
220 * from the cache and later re-added to the head of the cache.
222 * If a given dbuf meets the requirements for the metadata cache, it will go
223 * there, otherwise it will be considered for the generic LRU dbuf cache. The
224 * caches and the refcounts tracking their sizes are stored in an array indexed
225 * by those caches' matching enum values (from dbuf_cached_state_t).
227 typedef struct dbuf_cache {
231 dbuf_cache_t dbuf_caches[DB_CACHE_MAX];
233 /* Size limits for the caches */
234 uint64_t dbuf_cache_max_bytes = 0;
235 uint64_t dbuf_metadata_cache_max_bytes = 0;
236 /* Set the default sizes of the caches to log2 fraction of arc size */
237 int dbuf_cache_shift = 5;
238 int dbuf_metadata_cache_shift = 6;
241 * For diagnostic purposes, this is incremented whenever we can't add
242 * something to the metadata cache because it's full, and instead put
243 * the data in the regular dbuf cache.
245 uint64_t dbuf_metadata_cache_overflow;
248 * The LRU dbuf cache uses a three-stage eviction policy:
249 * - A low water marker designates when the dbuf eviction thread
250 * should stop evicting from the dbuf cache.
251 * - When we reach the maximum size (aka mid water mark), we
252 * signal the eviction thread to run.
253 * - The high water mark indicates when the eviction thread
254 * is unable to keep up with the incoming load and eviction must
255 * happen in the context of the calling thread.
259 * low water mid water hi water
260 * +----------------------------------------+----------+----------+
265 * +----------------------------------------+----------+----------+
267 * evicting eviction directly
270 * The high and low water marks indicate the operating range for the eviction
271 * thread. The low water mark is, by default, 90% of the total size of the
272 * cache and the high water mark is at 110% (both of these percentages can be
273 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
274 * respectively). The eviction thread will try to ensure that the cache remains
275 * within this range by waking up every second and checking if the cache is
276 * above the low water mark. The thread can also be woken up by callers adding
277 * elements into the cache if the cache is larger than the mid water (i.e max
278 * cache size). Once the eviction thread is woken up and eviction is required,
279 * it will continue evicting buffers until it's able to reduce the cache size
280 * to the low water mark. If the cache size continues to grow and hits the high
281 * water mark, then callers adding elments to the cache will begin to evict
282 * directly from the cache until the cache is no longer above the high water
287 * The percentage above and below the maximum cache size.
289 uint_t dbuf_cache_hiwater_pct = 10;
290 uint_t dbuf_cache_lowater_pct = 10;
292 SYSCTL_DECL(_vfs_zfs);
293 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, dbuf_cache_max_bytes, CTLFLAG_RWTUN,
294 &dbuf_cache_max_bytes, 0, "dbuf cache size in bytes");
295 SYSCTL_INT(_vfs_zfs, OID_AUTO, dbuf_cache_shift, CTLFLAG_RDTUN,
296 &dbuf_cache_shift, 0, "dbuf cache size as log2 fraction of ARC");
297 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_hiwater_pct, CTLFLAG_RWTUN,
298 &dbuf_cache_hiwater_pct, 0, "max percents above the dbuf cache size");
299 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_lowater_pct, CTLFLAG_RWTUN,
300 &dbuf_cache_lowater_pct, 0, "max percents below the dbuf cache size");
304 dbuf_cons(void *vdb, void *unused, int kmflag)
306 dmu_buf_impl_t *db = vdb;
307 bzero(db, sizeof (dmu_buf_impl_t));
309 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
310 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
311 multilist_link_init(&db->db_cache_link);
312 refcount_create(&db->db_holds);
319 dbuf_dest(void *vdb, void *unused)
321 dmu_buf_impl_t *db = vdb;
322 mutex_destroy(&db->db_mtx);
323 cv_destroy(&db->db_changed);
324 ASSERT(!multilist_link_active(&db->db_cache_link));
325 refcount_destroy(&db->db_holds);
329 * dbuf hash table routines
331 static dbuf_hash_table_t dbuf_hash_table;
333 static uint64_t dbuf_hash_count;
336 * We use Cityhash for this. It's fast, and has good hash properties without
337 * requiring any large static buffers.
340 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
342 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid));
345 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
346 ((dbuf)->db.db_object == (obj) && \
347 (dbuf)->db_objset == (os) && \
348 (dbuf)->db_level == (level) && \
349 (dbuf)->db_blkid == (blkid))
352 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
354 dbuf_hash_table_t *h = &dbuf_hash_table;
355 uint64_t hv = dbuf_hash(os, obj, level, blkid);
356 uint64_t idx = hv & h->hash_table_mask;
359 mutex_enter(DBUF_HASH_MUTEX(h, idx));
360 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
361 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
362 mutex_enter(&db->db_mtx);
363 if (db->db_state != DB_EVICTING) {
364 mutex_exit(DBUF_HASH_MUTEX(h, idx));
367 mutex_exit(&db->db_mtx);
370 mutex_exit(DBUF_HASH_MUTEX(h, idx));
374 static dmu_buf_impl_t *
375 dbuf_find_bonus(objset_t *os, uint64_t object)
378 dmu_buf_impl_t *db = NULL;
380 if (dnode_hold(os, object, FTAG, &dn) == 0) {
381 rw_enter(&dn->dn_struct_rwlock, RW_READER);
382 if (dn->dn_bonus != NULL) {
384 mutex_enter(&db->db_mtx);
386 rw_exit(&dn->dn_struct_rwlock);
387 dnode_rele(dn, FTAG);
393 * Insert an entry into the hash table. If there is already an element
394 * equal to elem in the hash table, then the already existing element
395 * will be returned and the new element will not be inserted.
396 * Otherwise returns NULL.
398 static dmu_buf_impl_t *
399 dbuf_hash_insert(dmu_buf_impl_t *db)
401 dbuf_hash_table_t *h = &dbuf_hash_table;
402 objset_t *os = db->db_objset;
403 uint64_t obj = db->db.db_object;
404 int level = db->db_level;
405 uint64_t blkid, hv, idx;
409 blkid = db->db_blkid;
410 hv = dbuf_hash(os, obj, level, blkid);
411 idx = hv & h->hash_table_mask;
413 mutex_enter(DBUF_HASH_MUTEX(h, idx));
414 for (dbf = h->hash_table[idx], i = 0; dbf != NULL;
415 dbf = dbf->db_hash_next, i++) {
416 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
417 mutex_enter(&dbf->db_mtx);
418 if (dbf->db_state != DB_EVICTING) {
419 mutex_exit(DBUF_HASH_MUTEX(h, idx));
422 mutex_exit(&dbf->db_mtx);
427 DBUF_STAT_BUMP(hash_collisions);
429 DBUF_STAT_BUMP(hash_chains);
431 DBUF_STAT_MAX(hash_chain_max, i);
434 mutex_enter(&db->db_mtx);
435 db->db_hash_next = h->hash_table[idx];
436 h->hash_table[idx] = db;
437 mutex_exit(DBUF_HASH_MUTEX(h, idx));
438 atomic_inc_64(&dbuf_hash_count);
439 DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count);
445 * Remove an entry from the hash table. It must be in the EVICTING state.
448 dbuf_hash_remove(dmu_buf_impl_t *db)
450 dbuf_hash_table_t *h = &dbuf_hash_table;
452 dmu_buf_impl_t *dbf, **dbp;
454 hv = dbuf_hash(db->db_objset, db->db.db_object,
455 db->db_level, db->db_blkid);
456 idx = hv & h->hash_table_mask;
459 * We mustn't hold db_mtx to maintain lock ordering:
460 * DBUF_HASH_MUTEX > db_mtx.
462 ASSERT(refcount_is_zero(&db->db_holds));
463 ASSERT(db->db_state == DB_EVICTING);
464 ASSERT(!MUTEX_HELD(&db->db_mtx));
466 mutex_enter(DBUF_HASH_MUTEX(h, idx));
467 dbp = &h->hash_table[idx];
468 while ((dbf = *dbp) != db) {
469 dbp = &dbf->db_hash_next;
472 *dbp = db->db_hash_next;
473 db->db_hash_next = NULL;
474 if (h->hash_table[idx] &&
475 h->hash_table[idx]->db_hash_next == NULL)
476 DBUF_STAT_BUMPDOWN(hash_chains);
477 mutex_exit(DBUF_HASH_MUTEX(h, idx));
478 atomic_dec_64(&dbuf_hash_count);
484 } dbvu_verify_type_t;
487 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
492 if (db->db_user == NULL)
495 /* Only data blocks support the attachment of user data. */
496 ASSERT(db->db_level == 0);
498 /* Clients must resolve a dbuf before attaching user data. */
499 ASSERT(db->db.db_data != NULL);
500 ASSERT3U(db->db_state, ==, DB_CACHED);
502 holds = refcount_count(&db->db_holds);
503 if (verify_type == DBVU_EVICTING) {
505 * Immediate eviction occurs when holds == dirtycnt.
506 * For normal eviction buffers, holds is zero on
507 * eviction, except when dbuf_fix_old_data() calls
508 * dbuf_clear_data(). However, the hold count can grow
509 * during eviction even though db_mtx is held (see
510 * dmu_bonus_hold() for an example), so we can only
511 * test the generic invariant that holds >= dirtycnt.
513 ASSERT3U(holds, >=, db->db_dirtycnt);
515 if (db->db_user_immediate_evict == TRUE)
516 ASSERT3U(holds, >=, db->db_dirtycnt);
518 ASSERT3U(holds, >, 0);
524 dbuf_evict_user(dmu_buf_impl_t *db)
526 dmu_buf_user_t *dbu = db->db_user;
528 ASSERT(MUTEX_HELD(&db->db_mtx));
533 dbuf_verify_user(db, DBVU_EVICTING);
537 if (dbu->dbu_clear_on_evict_dbufp != NULL)
538 *dbu->dbu_clear_on_evict_dbufp = NULL;
542 * There are two eviction callbacks - one that we call synchronously
543 * and one that we invoke via a taskq. The async one is useful for
544 * avoiding lock order reversals and limiting stack depth.
546 * Note that if we have a sync callback but no async callback,
547 * it's likely that the sync callback will free the structure
548 * containing the dbu. In that case we need to take care to not
549 * dereference dbu after calling the sync evict func.
551 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
553 if (dbu->dbu_evict_func_sync != NULL)
554 dbu->dbu_evict_func_sync(dbu);
557 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
558 dbu, 0, &dbu->dbu_tqent);
563 dbuf_is_metadata(dmu_buf_impl_t *db)
565 if (db->db_level > 0) {
568 boolean_t is_metadata;
571 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
574 return (is_metadata);
579 * This returns whether this dbuf should be stored in the metadata cache, which
580 * is based on whether it's from one of the dnode types that store data related
581 * to traversing dataset hierarchies.
584 dbuf_include_in_metadata_cache(dmu_buf_impl_t *db)
587 dmu_object_type_t type = DB_DNODE(db)->dn_type;
590 /* Check if this dbuf is one of the types we care about */
591 if (DMU_OT_IS_METADATA_CACHED(type)) {
592 /* If we hit this, then we set something up wrong in dmu_ot */
593 ASSERT(DMU_OT_IS_METADATA(type));
596 * Sanity check for small-memory systems: don't allocate too
597 * much memory for this purpose.
599 if (refcount_count(&dbuf_caches[DB_DBUF_METADATA_CACHE].size) >
600 dbuf_metadata_cache_max_bytes) {
601 dbuf_metadata_cache_overflow++;
602 DTRACE_PROBE1(dbuf__metadata__cache__overflow,
603 dmu_buf_impl_t *, db);
614 * This function *must* return indices evenly distributed between all
615 * sublists of the multilist. This is needed due to how the dbuf eviction
616 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
617 * distributed between all sublists and uses this assumption when
618 * deciding which sublist to evict from and how much to evict from it.
621 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
623 dmu_buf_impl_t *db = obj;
626 * The assumption here, is the hash value for a given
627 * dmu_buf_impl_t will remain constant throughout it's lifetime
628 * (i.e. it's objset, object, level and blkid fields don't change).
629 * Thus, we don't need to store the dbuf's sublist index
630 * on insertion, as this index can be recalculated on removal.
632 * Also, the low order bits of the hash value are thought to be
633 * distributed evenly. Otherwise, in the case that the multilist
634 * has a power of two number of sublists, each sublists' usage
635 * would not be evenly distributed.
637 return (dbuf_hash(db->db_objset, db->db.db_object,
638 db->db_level, db->db_blkid) %
639 multilist_get_num_sublists(ml));
642 static inline unsigned long
643 dbuf_cache_target_bytes(void)
645 return MIN(dbuf_cache_max_bytes,
646 arc_max_bytes() >> dbuf_cache_shift);
649 static inline uint64_t
650 dbuf_cache_hiwater_bytes(void)
652 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
653 return (dbuf_cache_target +
654 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100);
657 static inline uint64_t
658 dbuf_cache_lowater_bytes(void)
660 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
661 return (dbuf_cache_target -
662 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100);
665 static inline boolean_t
666 dbuf_cache_above_hiwater(void)
668 return (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
669 dbuf_cache_hiwater_bytes());
672 static inline boolean_t
673 dbuf_cache_above_lowater(void)
675 return (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
676 dbuf_cache_lowater_bytes());
680 * Evict the oldest eligible dbuf from the dbuf cache.
685 int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache);
686 multilist_sublist_t *mls = multilist_sublist_lock(
687 dbuf_caches[DB_DBUF_CACHE].cache, idx);
689 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
691 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
692 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
693 db = multilist_sublist_prev(mls, db);
696 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
697 multilist_sublist_t *, mls);
700 multilist_sublist_remove(mls, db);
701 multilist_sublist_unlock(mls);
702 (void) refcount_remove_many(&dbuf_caches[DB_DBUF_CACHE].size,
704 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
705 DBUF_STAT_BUMPDOWN(cache_count);
706 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
708 ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE);
709 db->db_caching_status = DB_NO_CACHE;
711 DBUF_STAT_MAX(cache_size_bytes_max,
712 refcount_count(&dbuf_caches[DB_DBUF_CACHE].size));
713 DBUF_STAT_BUMP(cache_total_evicts);
715 multilist_sublist_unlock(mls);
720 * The dbuf evict thread is responsible for aging out dbufs from the
721 * cache. Once the cache has reached it's maximum size, dbufs are removed
722 * and destroyed. The eviction thread will continue running until the size
723 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
724 * out of the cache it is destroyed and becomes eligible for arc eviction.
728 dbuf_evict_thread(void *unused __unused)
732 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
734 mutex_enter(&dbuf_evict_lock);
735 while (!dbuf_evict_thread_exit) {
736 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
737 CALLB_CPR_SAFE_BEGIN(&cpr);
738 (void) cv_timedwait_hires(&dbuf_evict_cv,
739 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
740 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
742 mutex_exit(&dbuf_evict_lock);
745 * Keep evicting as long as we're above the low water mark
746 * for the cache. We do this without holding the locks to
747 * minimize lock contention.
749 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
753 mutex_enter(&dbuf_evict_lock);
756 dbuf_evict_thread_exit = B_FALSE;
757 cv_broadcast(&dbuf_evict_cv);
758 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
763 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
764 * If the dbuf cache is at its high water mark, then evict a dbuf from the
765 * dbuf cache using the callers context.
768 dbuf_evict_notify(void)
771 * We check if we should evict without holding the dbuf_evict_lock,
772 * because it's OK to occasionally make the wrong decision here,
773 * and grabbing the lock results in massive lock contention.
775 if (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
776 dbuf_cache_max_bytes) {
777 if (dbuf_cache_above_hiwater())
779 cv_signal(&dbuf_evict_cv);
784 dbuf_kstat_update(kstat_t *ksp, int rw)
786 dbuf_stats_t *ds = ksp->ks_data;
788 if (rw == KSTAT_WRITE) {
789 return (SET_ERROR(EACCES));
791 ds->metadata_cache_size_bytes.value.ui64 =
792 refcount_count(&dbuf_caches[DB_DBUF_METADATA_CACHE].size);
793 ds->cache_size_bytes.value.ui64 =
794 refcount_count(&dbuf_caches[DB_DBUF_CACHE].size);
795 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
796 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
797 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
798 ds->hash_elements.value.ui64 = dbuf_hash_count;
807 uint64_t hsize = 1ULL << 16;
808 dbuf_hash_table_t *h = &dbuf_hash_table;
812 * The hash table is big enough to fill all of physical memory
813 * with an average 4K block size. The table will take up
814 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
816 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
820 h->hash_table_mask = hsize - 1;
821 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
822 if (h->hash_table == NULL) {
823 /* XXX - we should really return an error instead of assert */
824 ASSERT(hsize > (1ULL << 10));
829 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
830 sizeof (dmu_buf_impl_t),
831 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
833 for (i = 0; i < DBUF_MUTEXES; i++)
834 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
838 * Setup the parameters for the dbuf caches. We set the sizes of the
839 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
840 * of the size of the ARC, respectively. If the values are set in
841 * /etc/system and they're not greater than the size of the ARC, then
842 * we honor that value.
844 if (dbuf_cache_max_bytes == 0 ||
845 dbuf_cache_max_bytes >= arc_max_bytes()) {
846 dbuf_cache_max_bytes = arc_max_bytes() >> dbuf_cache_shift;
848 if (dbuf_metadata_cache_max_bytes == 0 ||
849 dbuf_metadata_cache_max_bytes >= arc_max_bytes()) {
850 dbuf_metadata_cache_max_bytes =
851 arc_max_bytes() >> dbuf_metadata_cache_shift;
855 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
856 * configuration is not required.
858 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
860 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
861 dbuf_caches[dcs].cache =
862 multilist_create(sizeof (dmu_buf_impl_t),
863 offsetof(dmu_buf_impl_t, db_cache_link),
864 dbuf_cache_multilist_index_func);
865 refcount_create(&dbuf_caches[dcs].size);
868 dbuf_evict_thread_exit = B_FALSE;
869 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
870 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
871 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
872 NULL, 0, &p0, TS_RUN, minclsyspri);
876 * XXX FreeBSD's SPL lacks KSTAT_TYPE_NAMED support - TODO
878 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
879 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
881 if (dbuf_ksp != NULL) {
882 dbuf_ksp->ks_data = &dbuf_stats;
883 dbuf_ksp->ks_update = dbuf_kstat_update;
884 kstat_install(dbuf_ksp);
886 for (i = 0; i < DN_MAX_LEVELS; i++) {
887 snprintf(dbuf_stats.cache_levels[i].name,
888 KSTAT_STRLEN, "cache_level_%d", i);
889 dbuf_stats.cache_levels[i].data_type =
891 snprintf(dbuf_stats.cache_levels_bytes[i].name,
892 KSTAT_STRLEN, "cache_level_%d_bytes", i);
893 dbuf_stats.cache_levels_bytes[i].data_type =
903 dbuf_hash_table_t *h = &dbuf_hash_table;
906 dbuf_stats_destroy();
908 for (i = 0; i < DBUF_MUTEXES; i++)
909 mutex_destroy(&h->hash_mutexes[i]);
910 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
911 kmem_cache_destroy(dbuf_kmem_cache);
912 taskq_destroy(dbu_evict_taskq);
914 mutex_enter(&dbuf_evict_lock);
915 dbuf_evict_thread_exit = B_TRUE;
916 while (dbuf_evict_thread_exit) {
917 cv_signal(&dbuf_evict_cv);
918 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
920 mutex_exit(&dbuf_evict_lock);
922 mutex_destroy(&dbuf_evict_lock);
923 cv_destroy(&dbuf_evict_cv);
925 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
926 refcount_destroy(&dbuf_caches[dcs].size);
927 multilist_destroy(dbuf_caches[dcs].cache);
930 if (dbuf_ksp != NULL) {
931 kstat_delete(dbuf_ksp);
942 dbuf_verify(dmu_buf_impl_t *db)
945 dbuf_dirty_record_t *dr;
947 ASSERT(MUTEX_HELD(&db->db_mtx));
949 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
952 ASSERT(db->db_objset != NULL);
956 ASSERT(db->db_parent == NULL);
957 ASSERT(db->db_blkptr == NULL);
959 ASSERT3U(db->db.db_object, ==, dn->dn_object);
960 ASSERT3P(db->db_objset, ==, dn->dn_objset);
961 ASSERT3U(db->db_level, <, dn->dn_nlevels);
962 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
963 db->db_blkid == DMU_SPILL_BLKID ||
964 !avl_is_empty(&dn->dn_dbufs));
966 if (db->db_blkid == DMU_BONUS_BLKID) {
968 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
969 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
970 } else if (db->db_blkid == DMU_SPILL_BLKID) {
972 ASSERT0(db->db.db_offset);
974 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
977 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
978 ASSERT(dr->dr_dbuf == db);
980 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
981 ASSERT(dr->dr_dbuf == db);
984 * We can't assert that db_size matches dn_datablksz because it
985 * can be momentarily different when another thread is doing
988 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
989 dr = db->db_data_pending;
991 * It should only be modified in syncing context, so
992 * make sure we only have one copy of the data.
994 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
997 /* verify db->db_blkptr */
999 if (db->db_parent == dn->dn_dbuf) {
1000 /* db is pointed to by the dnode */
1001 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1002 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
1003 ASSERT(db->db_parent == NULL);
1005 ASSERT(db->db_parent != NULL);
1006 if (db->db_blkid != DMU_SPILL_BLKID)
1007 ASSERT3P(db->db_blkptr, ==,
1008 &dn->dn_phys->dn_blkptr[db->db_blkid]);
1010 /* db is pointed to by an indirect block */
1011 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
1012 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
1013 ASSERT3U(db->db_parent->db.db_object, ==,
1016 * dnode_grow_indblksz() can make this fail if we don't
1017 * have the struct_rwlock. XXX indblksz no longer
1018 * grows. safe to do this now?
1020 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1021 ASSERT3P(db->db_blkptr, ==,
1022 ((blkptr_t *)db->db_parent->db.db_data +
1023 db->db_blkid % epb));
1027 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
1028 (db->db_buf == NULL || db->db_buf->b_data) &&
1029 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
1030 db->db_state != DB_FILL && !dn->dn_free_txg) {
1032 * If the blkptr isn't set but they have nonzero data,
1033 * it had better be dirty, otherwise we'll lose that
1034 * data when we evict this buffer.
1036 * There is an exception to this rule for indirect blocks; in
1037 * this case, if the indirect block is a hole, we fill in a few
1038 * fields on each of the child blocks (importantly, birth time)
1039 * to prevent hole birth times from being lost when you
1040 * partially fill in a hole.
1042 if (db->db_dirtycnt == 0) {
1043 if (db->db_level == 0) {
1044 uint64_t *buf = db->db.db_data;
1047 for (i = 0; i < db->db.db_size >> 3; i++) {
1048 ASSERT(buf[i] == 0);
1051 blkptr_t *bps = db->db.db_data;
1052 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
1055 * We want to verify that all the blkptrs in the
1056 * indirect block are holes, but we may have
1057 * automatically set up a few fields for them.
1058 * We iterate through each blkptr and verify
1059 * they only have those fields set.
1062 i < db->db.db_size / sizeof (blkptr_t);
1064 blkptr_t *bp = &bps[i];
1065 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1068 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
1069 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
1070 DVA_IS_EMPTY(&bp->blk_dva[2]));
1071 ASSERT0(bp->blk_fill);
1072 ASSERT0(bp->blk_pad[0]);
1073 ASSERT0(bp->blk_pad[1]);
1074 ASSERT(!BP_IS_EMBEDDED(bp));
1075 ASSERT(BP_IS_HOLE(bp));
1076 ASSERT0(bp->blk_phys_birth);
1086 dbuf_clear_data(dmu_buf_impl_t *db)
1088 ASSERT(MUTEX_HELD(&db->db_mtx));
1089 dbuf_evict_user(db);
1090 ASSERT3P(db->db_buf, ==, NULL);
1091 db->db.db_data = NULL;
1092 if (db->db_state != DB_NOFILL)
1093 db->db_state = DB_UNCACHED;
1097 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1099 ASSERT(MUTEX_HELD(&db->db_mtx));
1100 ASSERT(buf != NULL);
1103 ASSERT(buf->b_data != NULL);
1104 db->db.db_data = buf->b_data;
1108 * Loan out an arc_buf for read. Return the loaned arc_buf.
1111 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1115 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1116 mutex_enter(&db->db_mtx);
1117 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
1118 int blksz = db->db.db_size;
1119 spa_t *spa = db->db_objset->os_spa;
1121 mutex_exit(&db->db_mtx);
1122 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1123 bcopy(db->db.db_data, abuf->b_data, blksz);
1126 arc_loan_inuse_buf(abuf, db);
1128 dbuf_clear_data(db);
1129 mutex_exit(&db->db_mtx);
1135 * Calculate which level n block references the data at the level 0 offset
1139 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
1141 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1143 * The level n blkid is equal to the level 0 blkid divided by
1144 * the number of level 0s in a level n block.
1146 * The level 0 blkid is offset >> datablkshift =
1147 * offset / 2^datablkshift.
1149 * The number of level 0s in a level n is the number of block
1150 * pointers in an indirect block, raised to the power of level.
1151 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1152 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1154 * Thus, the level n blkid is: offset /
1155 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
1156 * = offset / 2^(datablkshift + level *
1157 * (indblkshift - SPA_BLKPTRSHIFT))
1158 * = offset >> (datablkshift + level *
1159 * (indblkshift - SPA_BLKPTRSHIFT))
1161 return (offset >> (dn->dn_datablkshift + level *
1162 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
1164 ASSERT3U(offset, <, dn->dn_datablksz);
1170 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1171 arc_buf_t *buf, void *vdb)
1173 dmu_buf_impl_t *db = vdb;
1175 mutex_enter(&db->db_mtx);
1176 ASSERT3U(db->db_state, ==, DB_READ);
1178 * All reads are synchronous, so we must have a hold on the dbuf
1180 ASSERT(refcount_count(&db->db_holds) > 0);
1181 ASSERT(db->db_buf == NULL);
1182 ASSERT(db->db.db_data == NULL);
1185 ASSERT(zio == NULL || zio->io_error != 0);
1186 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1187 ASSERT3P(db->db_buf, ==, NULL);
1188 db->db_state = DB_UNCACHED;
1189 } else if (db->db_level == 0 && db->db_freed_in_flight) {
1190 /* freed in flight */
1191 ASSERT(zio == NULL || zio->io_error == 0);
1193 buf = arc_alloc_buf(db->db_objset->os_spa,
1194 db, DBUF_GET_BUFC_TYPE(db), db->db.db_size);
1196 arc_release(buf, db);
1197 bzero(buf->b_data, db->db.db_size);
1198 arc_buf_freeze(buf);
1199 db->db_freed_in_flight = FALSE;
1200 dbuf_set_data(db, buf);
1201 db->db_state = DB_CACHED;
1204 ASSERT(zio == NULL || zio->io_error == 0);
1205 dbuf_set_data(db, buf);
1206 db->db_state = DB_CACHED;
1208 cv_broadcast(&db->db_changed);
1209 dbuf_rele_and_unlock(db, NULL, B_FALSE);
1213 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1216 zbookmark_phys_t zb;
1217 arc_flags_t aflags = ARC_FLAG_NOWAIT;
1221 ASSERT(!refcount_is_zero(&db->db_holds));
1222 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1223 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1224 ASSERT(MUTEX_HELD(&db->db_mtx));
1225 ASSERT(db->db_state == DB_UNCACHED);
1226 ASSERT(db->db_buf == NULL);
1228 if (db->db_blkid == DMU_BONUS_BLKID) {
1230 * The bonus length stored in the dnode may be less than
1231 * the maximum available space in the bonus buffer.
1233 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1234 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1236 ASSERT3U(bonuslen, <=, db->db.db_size);
1237 db->db.db_data = zio_buf_alloc(max_bonuslen);
1238 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1239 if (bonuslen < max_bonuslen)
1240 bzero(db->db.db_data, max_bonuslen);
1242 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1244 db->db_state = DB_CACHED;
1245 mutex_exit(&db->db_mtx);
1250 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1251 * processes the delete record and clears the bp while we are waiting
1252 * for the dn_mtx (resulting in a "no" from block_freed).
1254 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1255 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1256 BP_IS_HOLE(db->db_blkptr)))) {
1257 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1259 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1261 bzero(db->db.db_data, db->db.db_size);
1263 if (db->db_blkptr != NULL && db->db_level > 0 &&
1264 BP_IS_HOLE(db->db_blkptr) &&
1265 db->db_blkptr->blk_birth != 0) {
1266 blkptr_t *bps = db->db.db_data;
1267 for (int i = 0; i < ((1 <<
1268 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1270 blkptr_t *bp = &bps[i];
1271 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1272 1 << dn->dn_indblkshift);
1274 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1276 BP_GET_LSIZE(db->db_blkptr));
1277 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1279 BP_GET_LEVEL(db->db_blkptr) - 1);
1280 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1284 db->db_state = DB_CACHED;
1285 mutex_exit(&db->db_mtx);
1291 db->db_state = DB_READ;
1292 mutex_exit(&db->db_mtx);
1294 if (DBUF_IS_L2CACHEABLE(db))
1295 aflags |= ARC_FLAG_L2CACHE;
1297 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1298 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1299 db->db.db_object, db->db_level, db->db_blkid);
1301 dbuf_add_ref(db, NULL);
1303 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1304 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1305 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1310 * This is our just-in-time copy function. It makes a copy of buffers that
1311 * have been modified in a previous transaction group before we access them in
1312 * the current active group.
1314 * This function is used in three places: when we are dirtying a buffer for the
1315 * first time in a txg, when we are freeing a range in a dnode that includes
1316 * this buffer, and when we are accessing a buffer which was received compressed
1317 * and later referenced in a WRITE_BYREF record.
1319 * Note that when we are called from dbuf_free_range() we do not put a hold on
1320 * the buffer, we just traverse the active dbuf list for the dnode.
1323 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1325 dbuf_dirty_record_t *dr = db->db_last_dirty;
1327 ASSERT(MUTEX_HELD(&db->db_mtx));
1328 ASSERT(db->db.db_data != NULL);
1329 ASSERT(db->db_level == 0);
1330 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1333 (dr->dt.dl.dr_data !=
1334 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1338 * If the last dirty record for this dbuf has not yet synced
1339 * and its referencing the dbuf data, either:
1340 * reset the reference to point to a new copy,
1341 * or (if there a no active holders)
1342 * just null out the current db_data pointer.
1344 ASSERT(dr->dr_txg >= txg - 2);
1345 if (db->db_blkid == DMU_BONUS_BLKID) {
1346 /* Note that the data bufs here are zio_bufs */
1347 dnode_t *dn = DB_DNODE(db);
1348 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1349 dr->dt.dl.dr_data = zio_buf_alloc(bonuslen);
1350 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1351 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1352 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1353 int size = arc_buf_size(db->db_buf);
1354 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1355 spa_t *spa = db->db_objset->os_spa;
1356 enum zio_compress compress_type =
1357 arc_get_compression(db->db_buf);
1359 if (compress_type == ZIO_COMPRESS_OFF) {
1360 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1362 ASSERT3U(type, ==, ARC_BUFC_DATA);
1363 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1364 size, arc_buf_lsize(db->db_buf), compress_type);
1366 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1369 dbuf_clear_data(db);
1374 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1381 * We don't have to hold the mutex to check db_state because it
1382 * can't be freed while we have a hold on the buffer.
1384 ASSERT(!refcount_is_zero(&db->db_holds));
1386 if (db->db_state == DB_NOFILL)
1387 return (SET_ERROR(EIO));
1391 if ((flags & DB_RF_HAVESTRUCT) == 0)
1392 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1394 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1395 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1396 DBUF_IS_CACHEABLE(db);
1398 mutex_enter(&db->db_mtx);
1399 if (db->db_state == DB_CACHED) {
1401 * If the arc buf is compressed, we need to decompress it to
1402 * read the data. This could happen during the "zfs receive" of
1403 * a stream which is compressed and deduplicated.
1405 if (db->db_buf != NULL &&
1406 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1407 dbuf_fix_old_data(db,
1408 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1409 err = arc_decompress(db->db_buf);
1410 dbuf_set_data(db, db->db_buf);
1412 mutex_exit(&db->db_mtx);
1414 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1415 if ((flags & DB_RF_HAVESTRUCT) == 0)
1416 rw_exit(&dn->dn_struct_rwlock);
1418 DBUF_STAT_BUMP(hash_hits);
1419 } else if (db->db_state == DB_UNCACHED) {
1420 spa_t *spa = dn->dn_objset->os_spa;
1421 boolean_t need_wait = B_FALSE;
1424 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1425 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1428 dbuf_read_impl(db, zio, flags);
1430 /* dbuf_read_impl has dropped db_mtx for us */
1433 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1435 if ((flags & DB_RF_HAVESTRUCT) == 0)
1436 rw_exit(&dn->dn_struct_rwlock);
1438 DBUF_STAT_BUMP(hash_misses);
1441 err = zio_wait(zio);
1444 * Another reader came in while the dbuf was in flight
1445 * between UNCACHED and CACHED. Either a writer will finish
1446 * writing the buffer (sending the dbuf to CACHED) or the
1447 * first reader's request will reach the read_done callback
1448 * and send the dbuf to CACHED. Otherwise, a failure
1449 * occurred and the dbuf went to UNCACHED.
1451 mutex_exit(&db->db_mtx);
1453 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1454 if ((flags & DB_RF_HAVESTRUCT) == 0)
1455 rw_exit(&dn->dn_struct_rwlock);
1457 DBUF_STAT_BUMP(hash_misses);
1459 /* Skip the wait per the caller's request. */
1460 mutex_enter(&db->db_mtx);
1461 if ((flags & DB_RF_NEVERWAIT) == 0) {
1462 while (db->db_state == DB_READ ||
1463 db->db_state == DB_FILL) {
1464 ASSERT(db->db_state == DB_READ ||
1465 (flags & DB_RF_HAVESTRUCT) == 0);
1466 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1468 cv_wait(&db->db_changed, &db->db_mtx);
1470 if (db->db_state == DB_UNCACHED)
1471 err = SET_ERROR(EIO);
1473 mutex_exit(&db->db_mtx);
1480 dbuf_noread(dmu_buf_impl_t *db)
1482 ASSERT(!refcount_is_zero(&db->db_holds));
1483 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1484 mutex_enter(&db->db_mtx);
1485 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1486 cv_wait(&db->db_changed, &db->db_mtx);
1487 if (db->db_state == DB_UNCACHED) {
1488 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1489 spa_t *spa = db->db_objset->os_spa;
1491 ASSERT(db->db_buf == NULL);
1492 ASSERT(db->db.db_data == NULL);
1493 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1494 db->db_state = DB_FILL;
1495 } else if (db->db_state == DB_NOFILL) {
1496 dbuf_clear_data(db);
1498 ASSERT3U(db->db_state, ==, DB_CACHED);
1500 mutex_exit(&db->db_mtx);
1504 dbuf_unoverride(dbuf_dirty_record_t *dr)
1506 dmu_buf_impl_t *db = dr->dr_dbuf;
1507 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1508 uint64_t txg = dr->dr_txg;
1510 ASSERT(MUTEX_HELD(&db->db_mtx));
1512 * This assert is valid because dmu_sync() expects to be called by
1513 * a zilog's get_data while holding a range lock. This call only
1514 * comes from dbuf_dirty() callers who must also hold a range lock.
1516 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1517 ASSERT(db->db_level == 0);
1519 if (db->db_blkid == DMU_BONUS_BLKID ||
1520 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1523 ASSERT(db->db_data_pending != dr);
1525 /* free this block */
1526 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1527 zio_free(db->db_objset->os_spa, txg, bp);
1529 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1530 dr->dt.dl.dr_nopwrite = B_FALSE;
1533 * Release the already-written buffer, so we leave it in
1534 * a consistent dirty state. Note that all callers are
1535 * modifying the buffer, so they will immediately do
1536 * another (redundant) arc_release(). Therefore, leave
1537 * the buf thawed to save the effort of freezing &
1538 * immediately re-thawing it.
1540 arc_release(dr->dt.dl.dr_data, db);
1544 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1545 * data blocks in the free range, so that any future readers will find
1549 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1552 dmu_buf_impl_t db_search;
1553 dmu_buf_impl_t *db, *db_next;
1554 uint64_t txg = tx->tx_txg;
1557 if (end_blkid > dn->dn_maxblkid &&
1558 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1559 end_blkid = dn->dn_maxblkid;
1560 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1562 db_search.db_level = 0;
1563 db_search.db_blkid = start_blkid;
1564 db_search.db_state = DB_SEARCH;
1566 mutex_enter(&dn->dn_dbufs_mtx);
1567 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1568 ASSERT3P(db, ==, NULL);
1570 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1572 for (; db != NULL; db = db_next) {
1573 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1574 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1576 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1579 ASSERT3U(db->db_blkid, >=, start_blkid);
1581 /* found a level 0 buffer in the range */
1582 mutex_enter(&db->db_mtx);
1583 if (dbuf_undirty(db, tx)) {
1584 /* mutex has been dropped and dbuf destroyed */
1588 if (db->db_state == DB_UNCACHED ||
1589 db->db_state == DB_NOFILL ||
1590 db->db_state == DB_EVICTING) {
1591 ASSERT(db->db.db_data == NULL);
1592 mutex_exit(&db->db_mtx);
1595 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1596 /* will be handled in dbuf_read_done or dbuf_rele */
1597 db->db_freed_in_flight = TRUE;
1598 mutex_exit(&db->db_mtx);
1601 if (refcount_count(&db->db_holds) == 0) {
1606 /* The dbuf is referenced */
1608 if (db->db_last_dirty != NULL) {
1609 dbuf_dirty_record_t *dr = db->db_last_dirty;
1611 if (dr->dr_txg == txg) {
1613 * This buffer is "in-use", re-adjust the file
1614 * size to reflect that this buffer may
1615 * contain new data when we sync.
1617 if (db->db_blkid != DMU_SPILL_BLKID &&
1618 db->db_blkid > dn->dn_maxblkid)
1619 dn->dn_maxblkid = db->db_blkid;
1620 dbuf_unoverride(dr);
1623 * This dbuf is not dirty in the open context.
1624 * Either uncache it (if its not referenced in
1625 * the open context) or reset its contents to
1628 dbuf_fix_old_data(db, txg);
1631 /* clear the contents if its cached */
1632 if (db->db_state == DB_CACHED) {
1633 ASSERT(db->db.db_data != NULL);
1634 arc_release(db->db_buf, db);
1635 bzero(db->db.db_data, db->db.db_size);
1636 arc_buf_freeze(db->db_buf);
1639 mutex_exit(&db->db_mtx);
1641 mutex_exit(&dn->dn_dbufs_mtx);
1645 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1647 arc_buf_t *buf, *obuf;
1648 int osize = db->db.db_size;
1649 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1652 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1657 /* XXX does *this* func really need the lock? */
1658 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1661 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1662 * is OK, because there can be no other references to the db
1663 * when we are changing its size, so no concurrent DB_FILL can
1667 * XXX we should be doing a dbuf_read, checking the return
1668 * value and returning that up to our callers
1670 dmu_buf_will_dirty(&db->db, tx);
1672 /* create the data buffer for the new block */
1673 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1675 /* copy old block data to the new block */
1677 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1678 /* zero the remainder */
1680 bzero((uint8_t *)buf->b_data + osize, size - osize);
1682 mutex_enter(&db->db_mtx);
1683 dbuf_set_data(db, buf);
1684 arc_buf_destroy(obuf, db);
1685 db->db.db_size = size;
1687 if (db->db_level == 0) {
1688 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1689 db->db_last_dirty->dt.dl.dr_data = buf;
1691 mutex_exit(&db->db_mtx);
1693 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1698 dbuf_release_bp(dmu_buf_impl_t *db)
1700 objset_t *os = db->db_objset;
1702 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1703 ASSERT(arc_released(os->os_phys_buf) ||
1704 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1705 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1707 (void) arc_release(db->db_buf, db);
1711 * We already have a dirty record for this TXG, and we are being
1715 dbuf_redirty(dbuf_dirty_record_t *dr)
1717 dmu_buf_impl_t *db = dr->dr_dbuf;
1719 ASSERT(MUTEX_HELD(&db->db_mtx));
1721 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1723 * If this buffer has already been written out,
1724 * we now need to reset its state.
1726 dbuf_unoverride(dr);
1727 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1728 db->db_state != DB_NOFILL) {
1729 /* Already released on initial dirty, so just thaw. */
1730 ASSERT(arc_released(db->db_buf));
1731 arc_buf_thaw(db->db_buf);
1736 dbuf_dirty_record_t *
1737 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1741 dbuf_dirty_record_t **drp, *dr;
1742 int drop_struct_lock = FALSE;
1743 int txgoff = tx->tx_txg & TXG_MASK;
1745 ASSERT(tx->tx_txg != 0);
1746 ASSERT(!refcount_is_zero(&db->db_holds));
1747 DMU_TX_DIRTY_BUF(tx, db);
1752 * Shouldn't dirty a regular buffer in syncing context. Private
1753 * objects may be dirtied in syncing context, but only if they
1754 * were already pre-dirtied in open context.
1757 if (dn->dn_objset->os_dsl_dataset != NULL) {
1758 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1761 ASSERT(!dmu_tx_is_syncing(tx) ||
1762 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1763 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1764 dn->dn_objset->os_dsl_dataset == NULL);
1765 if (dn->dn_objset->os_dsl_dataset != NULL)
1766 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1769 * We make this assert for private objects as well, but after we
1770 * check if we're already dirty. They are allowed to re-dirty
1771 * in syncing context.
1773 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1774 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1775 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1777 mutex_enter(&db->db_mtx);
1779 * XXX make this true for indirects too? The problem is that
1780 * transactions created with dmu_tx_create_assigned() from
1781 * syncing context don't bother holding ahead.
1783 ASSERT(db->db_level != 0 ||
1784 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1785 db->db_state == DB_NOFILL);
1787 mutex_enter(&dn->dn_mtx);
1789 * Don't set dirtyctx to SYNC if we're just modifying this as we
1790 * initialize the objset.
1792 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1793 if (dn->dn_objset->os_dsl_dataset != NULL) {
1794 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1797 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1798 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1799 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1800 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1801 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1803 if (dn->dn_objset->os_dsl_dataset != NULL) {
1804 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1808 mutex_exit(&dn->dn_mtx);
1810 if (db->db_blkid == DMU_SPILL_BLKID)
1811 dn->dn_have_spill = B_TRUE;
1814 * If this buffer is already dirty, we're done.
1816 drp = &db->db_last_dirty;
1817 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1818 db->db.db_object == DMU_META_DNODE_OBJECT);
1819 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1821 if (dr && dr->dr_txg == tx->tx_txg) {
1825 mutex_exit(&db->db_mtx);
1830 * Only valid if not already dirty.
1832 ASSERT(dn->dn_object == 0 ||
1833 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1834 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1836 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1839 * We should only be dirtying in syncing context if it's the
1840 * mos or we're initializing the os or it's a special object.
1841 * However, we are allowed to dirty in syncing context provided
1842 * we already dirtied it in open context. Hence we must make
1843 * this assertion only if we're not already dirty.
1846 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1848 if (dn->dn_objset->os_dsl_dataset != NULL)
1849 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1850 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1851 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1852 if (dn->dn_objset->os_dsl_dataset != NULL)
1853 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1855 ASSERT(db->db.db_size != 0);
1857 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1859 if (db->db_blkid != DMU_BONUS_BLKID) {
1860 dmu_objset_willuse_space(os, db->db.db_size, tx);
1864 * If this buffer is dirty in an old transaction group we need
1865 * to make a copy of it so that the changes we make in this
1866 * transaction group won't leak out when we sync the older txg.
1868 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1869 list_link_init(&dr->dr_dirty_node);
1870 if (db->db_level == 0) {
1871 void *data_old = db->db_buf;
1873 if (db->db_state != DB_NOFILL) {
1874 if (db->db_blkid == DMU_BONUS_BLKID) {
1875 dbuf_fix_old_data(db, tx->tx_txg);
1876 data_old = db->db.db_data;
1877 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1879 * Release the data buffer from the cache so
1880 * that we can modify it without impacting
1881 * possible other users of this cached data
1882 * block. Note that indirect blocks and
1883 * private objects are not released until the
1884 * syncing state (since they are only modified
1887 arc_release(db->db_buf, db);
1888 dbuf_fix_old_data(db, tx->tx_txg);
1889 data_old = db->db_buf;
1891 ASSERT(data_old != NULL);
1893 dr->dt.dl.dr_data = data_old;
1895 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1896 list_create(&dr->dt.di.dr_children,
1897 sizeof (dbuf_dirty_record_t),
1898 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1900 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1901 dr->dr_accounted = db->db.db_size;
1903 dr->dr_txg = tx->tx_txg;
1908 * We could have been freed_in_flight between the dbuf_noread
1909 * and dbuf_dirty. We win, as though the dbuf_noread() had
1910 * happened after the free.
1912 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1913 db->db_blkid != DMU_SPILL_BLKID) {
1914 mutex_enter(&dn->dn_mtx);
1915 if (dn->dn_free_ranges[txgoff] != NULL) {
1916 range_tree_clear(dn->dn_free_ranges[txgoff],
1919 mutex_exit(&dn->dn_mtx);
1920 db->db_freed_in_flight = FALSE;
1924 * This buffer is now part of this txg
1926 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1927 db->db_dirtycnt += 1;
1928 ASSERT3U(db->db_dirtycnt, <=, 3);
1930 mutex_exit(&db->db_mtx);
1932 if (db->db_blkid == DMU_BONUS_BLKID ||
1933 db->db_blkid == DMU_SPILL_BLKID) {
1934 mutex_enter(&dn->dn_mtx);
1935 ASSERT(!list_link_active(&dr->dr_dirty_node));
1936 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1937 mutex_exit(&dn->dn_mtx);
1938 dnode_setdirty(dn, tx);
1944 * The dn_struct_rwlock prevents db_blkptr from changing
1945 * due to a write from syncing context completing
1946 * while we are running, so we want to acquire it before
1947 * looking at db_blkptr.
1949 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1950 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1951 drop_struct_lock = TRUE;
1955 * We need to hold the dn_struct_rwlock to make this assertion,
1956 * because it protects dn_phys / dn_next_nlevels from changing.
1958 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1959 dn->dn_phys->dn_nlevels > db->db_level ||
1960 dn->dn_next_nlevels[txgoff] > db->db_level ||
1961 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1962 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1965 * If we are overwriting a dedup BP, then unless it is snapshotted,
1966 * when we get to syncing context we will need to decrement its
1967 * refcount in the DDT. Prefetch the relevant DDT block so that
1968 * syncing context won't have to wait for the i/o.
1970 ddt_prefetch(os->os_spa, db->db_blkptr);
1972 if (db->db_level == 0) {
1973 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1974 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1977 if (db->db_level+1 < dn->dn_nlevels) {
1978 dmu_buf_impl_t *parent = db->db_parent;
1979 dbuf_dirty_record_t *di;
1980 int parent_held = FALSE;
1982 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1983 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1985 parent = dbuf_hold_level(dn, db->db_level+1,
1986 db->db_blkid >> epbs, FTAG);
1987 ASSERT(parent != NULL);
1990 if (drop_struct_lock)
1991 rw_exit(&dn->dn_struct_rwlock);
1992 ASSERT3U(db->db_level+1, ==, parent->db_level);
1993 di = dbuf_dirty(parent, tx);
1995 dbuf_rele(parent, FTAG);
1997 mutex_enter(&db->db_mtx);
1999 * Since we've dropped the mutex, it's possible that
2000 * dbuf_undirty() might have changed this out from under us.
2002 if (db->db_last_dirty == dr ||
2003 dn->dn_object == DMU_META_DNODE_OBJECT) {
2004 mutex_enter(&di->dt.di.dr_mtx);
2005 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
2006 ASSERT(!list_link_active(&dr->dr_dirty_node));
2007 list_insert_tail(&di->dt.di.dr_children, dr);
2008 mutex_exit(&di->dt.di.dr_mtx);
2011 mutex_exit(&db->db_mtx);
2013 ASSERT(db->db_level+1 == dn->dn_nlevels);
2014 ASSERT(db->db_blkid < dn->dn_nblkptr);
2015 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
2016 mutex_enter(&dn->dn_mtx);
2017 ASSERT(!list_link_active(&dr->dr_dirty_node));
2018 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2019 mutex_exit(&dn->dn_mtx);
2020 if (drop_struct_lock)
2021 rw_exit(&dn->dn_struct_rwlock);
2024 dnode_setdirty(dn, tx);
2030 * Undirty a buffer in the transaction group referenced by the given
2031 * transaction. Return whether this evicted the dbuf.
2034 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2037 uint64_t txg = tx->tx_txg;
2038 dbuf_dirty_record_t *dr, **drp;
2043 * Due to our use of dn_nlevels below, this can only be called
2044 * in open context, unless we are operating on the MOS.
2045 * From syncing context, dn_nlevels may be different from the
2046 * dn_nlevels used when dbuf was dirtied.
2048 ASSERT(db->db_objset ==
2049 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2050 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2051 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2052 ASSERT0(db->db_level);
2053 ASSERT(MUTEX_HELD(&db->db_mtx));
2056 * If this buffer is not dirty, we're done.
2058 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
2059 if (dr->dr_txg <= txg)
2061 if (dr == NULL || dr->dr_txg < txg)
2063 ASSERT(dr->dr_txg == txg);
2064 ASSERT(dr->dr_dbuf == db);
2069 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2071 ASSERT(db->db.db_size != 0);
2073 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2074 dr->dr_accounted, txg);
2079 * Note that there are three places in dbuf_dirty()
2080 * where this dirty record may be put on a list.
2081 * Make sure to do a list_remove corresponding to
2082 * every one of those list_insert calls.
2084 if (dr->dr_parent) {
2085 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2086 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2087 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2088 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2089 db->db_level + 1 == dn->dn_nlevels) {
2090 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2091 mutex_enter(&dn->dn_mtx);
2092 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2093 mutex_exit(&dn->dn_mtx);
2097 if (db->db_state != DB_NOFILL) {
2098 dbuf_unoverride(dr);
2100 ASSERT(db->db_buf != NULL);
2101 ASSERT(dr->dt.dl.dr_data != NULL);
2102 if (dr->dt.dl.dr_data != db->db_buf)
2103 arc_buf_destroy(dr->dt.dl.dr_data, db);
2106 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2108 ASSERT(db->db_dirtycnt > 0);
2109 db->db_dirtycnt -= 1;
2111 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2112 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2121 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2123 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2124 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
2126 ASSERT(tx->tx_txg != 0);
2127 ASSERT(!refcount_is_zero(&db->db_holds));
2130 * Quick check for dirtyness. For already dirty blocks, this
2131 * reduces runtime of this function by >90%, and overall performance
2132 * by 50% for some workloads (e.g. file deletion with indirect blocks
2135 mutex_enter(&db->db_mtx);
2136 dbuf_dirty_record_t *dr;
2137 for (dr = db->db_last_dirty;
2138 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
2140 * It's possible that it is already dirty but not cached,
2141 * because there are some calls to dbuf_dirty() that don't
2142 * go through dmu_buf_will_dirty().
2144 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
2145 /* This dbuf is already dirty and cached. */
2147 mutex_exit(&db->db_mtx);
2151 mutex_exit(&db->db_mtx);
2154 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2155 rf |= DB_RF_HAVESTRUCT;
2157 (void) dbuf_read(db, NULL, rf);
2158 (void) dbuf_dirty(db, tx);
2162 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2164 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2166 db->db_state = DB_NOFILL;
2168 dmu_buf_will_fill(db_fake, tx);
2172 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2174 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2176 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2177 ASSERT(tx->tx_txg != 0);
2178 ASSERT(db->db_level == 0);
2179 ASSERT(!refcount_is_zero(&db->db_holds));
2181 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2182 dmu_tx_private_ok(tx));
2185 (void) dbuf_dirty(db, tx);
2188 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2191 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2193 mutex_enter(&db->db_mtx);
2196 if (db->db_state == DB_FILL) {
2197 if (db->db_level == 0 && db->db_freed_in_flight) {
2198 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2199 /* we were freed while filling */
2200 /* XXX dbuf_undirty? */
2201 bzero(db->db.db_data, db->db.db_size);
2202 db->db_freed_in_flight = FALSE;
2204 db->db_state = DB_CACHED;
2205 cv_broadcast(&db->db_changed);
2207 mutex_exit(&db->db_mtx);
2211 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2212 bp_embedded_type_t etype, enum zio_compress comp,
2213 int uncompressed_size, int compressed_size, int byteorder,
2216 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2217 struct dirty_leaf *dl;
2218 dmu_object_type_t type;
2220 if (etype == BP_EMBEDDED_TYPE_DATA) {
2221 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2222 SPA_FEATURE_EMBEDDED_DATA));
2226 type = DB_DNODE(db)->dn_type;
2229 ASSERT0(db->db_level);
2230 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2232 dmu_buf_will_not_fill(dbuf, tx);
2234 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2235 dl = &db->db_last_dirty->dt.dl;
2236 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2237 data, comp, uncompressed_size, compressed_size);
2238 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2239 BP_SET_TYPE(&dl->dr_overridden_by, type);
2240 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2241 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2243 dl->dr_override_state = DR_OVERRIDDEN;
2244 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2248 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2249 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2252 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2254 ASSERT(!refcount_is_zero(&db->db_holds));
2255 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2256 ASSERT(db->db_level == 0);
2257 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2258 ASSERT(buf != NULL);
2259 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2260 ASSERT(tx->tx_txg != 0);
2262 arc_return_buf(buf, db);
2263 ASSERT(arc_released(buf));
2265 mutex_enter(&db->db_mtx);
2267 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2268 cv_wait(&db->db_changed, &db->db_mtx);
2270 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2272 if (db->db_state == DB_CACHED &&
2273 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2274 mutex_exit(&db->db_mtx);
2275 (void) dbuf_dirty(db, tx);
2276 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2277 arc_buf_destroy(buf, db);
2278 xuio_stat_wbuf_copied();
2282 xuio_stat_wbuf_nocopy();
2283 if (db->db_state == DB_CACHED) {
2284 dbuf_dirty_record_t *dr = db->db_last_dirty;
2286 ASSERT(db->db_buf != NULL);
2287 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2288 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2289 if (!arc_released(db->db_buf)) {
2290 ASSERT(dr->dt.dl.dr_override_state ==
2292 arc_release(db->db_buf, db);
2294 dr->dt.dl.dr_data = buf;
2295 arc_buf_destroy(db->db_buf, db);
2296 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2297 arc_release(db->db_buf, db);
2298 arc_buf_destroy(db->db_buf, db);
2302 ASSERT(db->db_buf == NULL);
2303 dbuf_set_data(db, buf);
2304 db->db_state = DB_FILL;
2305 mutex_exit(&db->db_mtx);
2306 (void) dbuf_dirty(db, tx);
2307 dmu_buf_fill_done(&db->db, tx);
2311 dbuf_destroy(dmu_buf_impl_t *db)
2314 dmu_buf_impl_t *parent = db->db_parent;
2315 dmu_buf_impl_t *dndb;
2317 ASSERT(MUTEX_HELD(&db->db_mtx));
2318 ASSERT(refcount_is_zero(&db->db_holds));
2320 if (db->db_buf != NULL) {
2321 arc_buf_destroy(db->db_buf, db);
2325 if (db->db_blkid == DMU_BONUS_BLKID) {
2326 int slots = DB_DNODE(db)->dn_num_slots;
2327 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2328 if (db->db.db_data != NULL) {
2329 zio_buf_free(db->db.db_data, bonuslen);
2330 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2331 db->db_state = DB_UNCACHED;
2335 dbuf_clear_data(db);
2337 if (multilist_link_active(&db->db_cache_link)) {
2338 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
2339 db->db_caching_status == DB_DBUF_METADATA_CACHE);
2341 multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
2342 (void) refcount_remove_many(
2343 &dbuf_caches[db->db_caching_status].size,
2344 db->db.db_size, db);
2346 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2347 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2349 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2350 DBUF_STAT_BUMPDOWN(cache_count);
2351 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2354 db->db_caching_status = DB_NO_CACHE;
2357 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2358 ASSERT(db->db_data_pending == NULL);
2360 db->db_state = DB_EVICTING;
2361 db->db_blkptr = NULL;
2364 * Now that db_state is DB_EVICTING, nobody else can find this via
2365 * the hash table. We can now drop db_mtx, which allows us to
2366 * acquire the dn_dbufs_mtx.
2368 mutex_exit(&db->db_mtx);
2373 if (db->db_blkid != DMU_BONUS_BLKID) {
2374 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2376 mutex_enter(&dn->dn_dbufs_mtx);
2377 avl_remove(&dn->dn_dbufs, db);
2378 atomic_dec_32(&dn->dn_dbufs_count);
2382 mutex_exit(&dn->dn_dbufs_mtx);
2384 * Decrementing the dbuf count means that the hold corresponding
2385 * to the removed dbuf is no longer discounted in dnode_move(),
2386 * so the dnode cannot be moved until after we release the hold.
2387 * The membar_producer() ensures visibility of the decremented
2388 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2391 mutex_enter(&dn->dn_mtx);
2392 dnode_rele_and_unlock(dn, db, B_TRUE);
2393 db->db_dnode_handle = NULL;
2395 dbuf_hash_remove(db);
2400 ASSERT(refcount_is_zero(&db->db_holds));
2402 db->db_parent = NULL;
2404 ASSERT(db->db_buf == NULL);
2405 ASSERT(db->db.db_data == NULL);
2406 ASSERT(db->db_hash_next == NULL);
2407 ASSERT(db->db_blkptr == NULL);
2408 ASSERT(db->db_data_pending == NULL);
2409 ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE);
2410 ASSERT(!multilist_link_active(&db->db_cache_link));
2412 kmem_cache_free(dbuf_kmem_cache, db);
2413 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2416 * If this dbuf is referenced from an indirect dbuf,
2417 * decrement the ref count on the indirect dbuf.
2419 if (parent && parent != dndb) {
2420 mutex_enter(&parent->db_mtx);
2421 dbuf_rele_and_unlock(parent, db, B_TRUE);
2426 * Note: While bpp will always be updated if the function returns success,
2427 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2428 * this happens when the dnode is the meta-dnode, or a userused or groupused
2431 __attribute__((always_inline))
2433 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2434 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2439 ASSERT(blkid != DMU_BONUS_BLKID);
2441 if (blkid == DMU_SPILL_BLKID) {
2442 mutex_enter(&dn->dn_mtx);
2443 if (dn->dn_have_spill &&
2444 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2445 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2448 dbuf_add_ref(dn->dn_dbuf, NULL);
2449 *parentp = dn->dn_dbuf;
2450 mutex_exit(&dn->dn_mtx);
2455 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2456 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2458 ASSERT3U(level * epbs, <, 64);
2459 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2461 * This assertion shouldn't trip as long as the max indirect block size
2462 * is less than 1M. The reason for this is that up to that point,
2463 * the number of levels required to address an entire object with blocks
2464 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2465 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2466 * (i.e. we can address the entire object), objects will all use at most
2467 * N-1 levels and the assertion won't overflow. However, once epbs is
2468 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2469 * enough to address an entire object, so objects will have 5 levels,
2470 * but then this assertion will overflow.
2472 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2473 * need to redo this logic to handle overflows.
2475 ASSERT(level >= nlevels ||
2476 ((nlevels - level - 1) * epbs) +
2477 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2478 if (level >= nlevels ||
2479 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2480 ((nlevels - level - 1) * epbs)) ||
2482 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2483 /* the buffer has no parent yet */
2484 return (SET_ERROR(ENOENT));
2485 } else if (level < nlevels-1) {
2486 /* this block is referenced from an indirect block */
2489 err = dbuf_hold_impl(dn, level+1,
2490 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2492 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2493 blkid >> epbs, fail_sparse, FALSE, NULL,
2494 parentp, dh->dh_depth + 1);
2495 err = __dbuf_hold_impl(dh + 1);
2499 err = dbuf_read(*parentp, NULL,
2500 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2502 dbuf_rele(*parentp, NULL);
2506 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2507 (blkid & ((1ULL << epbs) - 1));
2508 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2509 ASSERT(BP_IS_HOLE(*bpp));
2512 /* the block is referenced from the dnode */
2513 ASSERT3U(level, ==, nlevels-1);
2514 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2515 blkid < dn->dn_phys->dn_nblkptr);
2517 dbuf_add_ref(dn->dn_dbuf, NULL);
2518 *parentp = dn->dn_dbuf;
2520 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2525 static dmu_buf_impl_t *
2526 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2527 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2529 objset_t *os = dn->dn_objset;
2530 dmu_buf_impl_t *db, *odb;
2532 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2533 ASSERT(dn->dn_type != DMU_OT_NONE);
2535 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2538 db->db.db_object = dn->dn_object;
2539 db->db_level = level;
2540 db->db_blkid = blkid;
2541 db->db_last_dirty = NULL;
2542 db->db_dirtycnt = 0;
2543 db->db_dnode_handle = dn->dn_handle;
2544 db->db_parent = parent;
2545 db->db_blkptr = blkptr;
2548 db->db_user_immediate_evict = FALSE;
2549 db->db_freed_in_flight = FALSE;
2550 db->db_pending_evict = FALSE;
2552 if (blkid == DMU_BONUS_BLKID) {
2553 ASSERT3P(parent, ==, dn->dn_dbuf);
2554 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2555 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2556 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2557 db->db.db_offset = DMU_BONUS_BLKID;
2558 db->db_state = DB_UNCACHED;
2559 db->db_caching_status = DB_NO_CACHE;
2560 /* the bonus dbuf is not placed in the hash table */
2561 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2563 } else if (blkid == DMU_SPILL_BLKID) {
2564 db->db.db_size = (blkptr != NULL) ?
2565 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2566 db->db.db_offset = 0;
2569 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2570 db->db.db_size = blocksize;
2571 db->db.db_offset = db->db_blkid * blocksize;
2575 * Hold the dn_dbufs_mtx while we get the new dbuf
2576 * in the hash table *and* added to the dbufs list.
2577 * This prevents a possible deadlock with someone
2578 * trying to look up this dbuf before its added to the
2581 mutex_enter(&dn->dn_dbufs_mtx);
2582 db->db_state = DB_EVICTING;
2583 if ((odb = dbuf_hash_insert(db)) != NULL) {
2584 /* someone else inserted it first */
2585 kmem_cache_free(dbuf_kmem_cache, db);
2586 mutex_exit(&dn->dn_dbufs_mtx);
2587 DBUF_STAT_BUMP(hash_insert_race);
2590 avl_add(&dn->dn_dbufs, db);
2592 db->db_state = DB_UNCACHED;
2593 db->db_caching_status = DB_NO_CACHE;
2594 mutex_exit(&dn->dn_dbufs_mtx);
2595 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2597 if (parent && parent != dn->dn_dbuf)
2598 dbuf_add_ref(parent, db);
2600 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2601 refcount_count(&dn->dn_holds) > 0);
2602 (void) refcount_add(&dn->dn_holds, db);
2603 atomic_inc_32(&dn->dn_dbufs_count);
2605 dprintf_dbuf(db, "db=%p\n", db);
2610 typedef struct dbuf_prefetch_arg {
2611 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2612 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2613 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2614 int dpa_curlevel; /* The current level that we're reading */
2615 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2616 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2617 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2618 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2619 } dbuf_prefetch_arg_t;
2622 * Actually issue the prefetch read for the block given.
2625 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2627 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2630 arc_flags_t aflags =
2631 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2633 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2634 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2635 ASSERT(dpa->dpa_zio != NULL);
2636 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2637 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2638 &aflags, &dpa->dpa_zb);
2642 * Called when an indirect block above our prefetch target is read in. This
2643 * will either read in the next indirect block down the tree or issue the actual
2644 * prefetch if the next block down is our target.
2647 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
2648 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
2650 dbuf_prefetch_arg_t *dpa = private;
2652 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2653 ASSERT3S(dpa->dpa_curlevel, >, 0);
2656 ASSERT(zio == NULL || zio->io_error != 0);
2657 kmem_free(dpa, sizeof (*dpa));
2660 ASSERT(zio == NULL || zio->io_error == 0);
2663 * The dpa_dnode is only valid if we are called with a NULL
2664 * zio. This indicates that the arc_read() returned without
2665 * first calling zio_read() to issue a physical read. Once
2666 * a physical read is made the dpa_dnode must be invalidated
2667 * as the locks guarding it may have been dropped. If the
2668 * dpa_dnode is still valid, then we want to add it to the dbuf
2669 * cache. To do so, we must hold the dbuf associated with the block
2670 * we just prefetched, read its contents so that we associate it
2671 * with an arc_buf_t, and then release it.
2674 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2675 if (zio->io_flags & ZIO_FLAG_RAW) {
2676 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2678 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2680 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2682 dpa->dpa_dnode = NULL;
2683 } else if (dpa->dpa_dnode != NULL) {
2684 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2685 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2686 dpa->dpa_zb.zb_level));
2687 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2688 dpa->dpa_curlevel, curblkid, FTAG);
2689 (void) dbuf_read(db, NULL,
2690 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2691 dbuf_rele(db, FTAG);
2695 kmem_free(dpa, sizeof(*dpa));
2699 dpa->dpa_curlevel--;
2701 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2702 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2703 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2704 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2705 if (BP_IS_HOLE(bp)) {
2706 kmem_free(dpa, sizeof (*dpa));
2707 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2708 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2709 dbuf_issue_final_prefetch(dpa, bp);
2710 kmem_free(dpa, sizeof (*dpa));
2712 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2713 zbookmark_phys_t zb;
2715 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2716 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2717 iter_aflags |= ARC_FLAG_L2CACHE;
2719 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2721 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2722 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2724 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2725 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2726 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2730 arc_buf_destroy(abuf, private);
2734 * Issue prefetch reads for the given block on the given level. If the indirect
2735 * blocks above that block are not in memory, we will read them in
2736 * asynchronously. As a result, this call never blocks waiting for a read to
2740 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2744 int epbs, nlevels, curlevel;
2747 ASSERT(blkid != DMU_BONUS_BLKID);
2748 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2750 if (blkid > dn->dn_maxblkid)
2753 if (dnode_block_freed(dn, blkid))
2757 * This dnode hasn't been written to disk yet, so there's nothing to
2760 nlevels = dn->dn_phys->dn_nlevels;
2761 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2764 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2765 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2768 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2771 mutex_exit(&db->db_mtx);
2773 * This dbuf already exists. It is either CACHED, or
2774 * (we assume) about to be read or filled.
2780 * Find the closest ancestor (indirect block) of the target block
2781 * that is present in the cache. In this indirect block, we will
2782 * find the bp that is at curlevel, curblkid.
2786 while (curlevel < nlevels - 1) {
2787 int parent_level = curlevel + 1;
2788 uint64_t parent_blkid = curblkid >> epbs;
2791 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2792 FALSE, TRUE, FTAG, &db) == 0) {
2793 blkptr_t *bpp = db->db_buf->b_data;
2794 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2795 dbuf_rele(db, FTAG);
2799 curlevel = parent_level;
2800 curblkid = parent_blkid;
2803 if (curlevel == nlevels - 1) {
2804 /* No cached indirect blocks found. */
2805 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2806 bp = dn->dn_phys->dn_blkptr[curblkid];
2808 if (BP_IS_HOLE(&bp))
2811 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2813 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2816 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2817 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2818 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2819 dn->dn_object, level, blkid);
2820 dpa->dpa_curlevel = curlevel;
2821 dpa->dpa_prio = prio;
2822 dpa->dpa_aflags = aflags;
2823 dpa->dpa_spa = dn->dn_objset->os_spa;
2824 dpa->dpa_dnode = dn;
2825 dpa->dpa_epbs = epbs;
2828 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2829 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2830 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2833 * If we have the indirect just above us, no need to do the asynchronous
2834 * prefetch chain; we'll just run the last step ourselves. If we're at
2835 * a higher level, though, we want to issue the prefetches for all the
2836 * indirect blocks asynchronously, so we can go on with whatever we were
2839 if (curlevel == level) {
2840 ASSERT3U(curblkid, ==, blkid);
2841 dbuf_issue_final_prefetch(dpa, &bp);
2842 kmem_free(dpa, sizeof (*dpa));
2844 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2845 zbookmark_phys_t zb;
2847 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2848 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2849 iter_aflags |= ARC_FLAG_L2CACHE;
2851 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2852 dn->dn_object, curlevel, curblkid);
2853 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2854 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2855 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2859 * We use pio here instead of dpa_zio since it's possible that
2860 * dpa may have already been freed.
2865 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2868 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2869 * the case of encrypted, compressed and uncompressed buffers by
2870 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2871 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2873 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2875 noinline static void
2876 dbuf_hold_copy(struct dbuf_hold_impl_data *dh)
2878 dnode_t *dn = dh->dh_dn;
2879 dmu_buf_impl_t *db = dh->dh_db;
2880 dbuf_dirty_record_t *dr = dh->dh_dr;
2881 arc_buf_t *data = dr->dt.dl.dr_data;
2883 enum zio_compress compress_type = arc_get_compression(data);
2885 if (compress_type != ZIO_COMPRESS_OFF) {
2886 dbuf_set_data(db, arc_alloc_compressed_buf(
2887 dn->dn_objset->os_spa, db, arc_buf_size(data),
2888 arc_buf_lsize(data), compress_type));
2890 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db,
2891 DBUF_GET_BUFC_TYPE(db), db->db.db_size));
2894 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
2898 * Returns with db_holds incremented, and db_mtx not held.
2899 * Note: dn_struct_rwlock must be held.
2902 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2904 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2905 dh->dh_parent = NULL;
2907 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2908 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2909 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2911 *(dh->dh_dbp) = NULL;
2913 /* dbuf_find() returns with db_mtx held */
2914 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2915 dh->dh_level, dh->dh_blkid);
2917 if (dh->dh_db == NULL) {
2920 if (dh->dh_fail_uncached)
2921 return (SET_ERROR(ENOENT));
2923 ASSERT3P(dh->dh_parent, ==, NULL);
2924 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2925 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2926 if (dh->dh_fail_sparse) {
2927 if (dh->dh_err == 0 &&
2928 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2929 dh->dh_err = SET_ERROR(ENOENT);
2932 dbuf_rele(dh->dh_parent, NULL);
2933 return (dh->dh_err);
2936 if (dh->dh_err && dh->dh_err != ENOENT)
2937 return (dh->dh_err);
2938 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2939 dh->dh_parent, dh->dh_bp);
2942 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2943 mutex_exit(&dh->dh_db->db_mtx);
2944 return (SET_ERROR(ENOENT));
2947 if (dh->dh_db->db_buf != NULL) {
2948 arc_buf_access(dh->dh_db->db_buf);
2949 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2952 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2955 * If this buffer is currently syncing out, and we are are
2956 * still referencing it from db_data, we need to make a copy
2957 * of it in case we decide we want to dirty it again in this txg.
2959 if (dh->dh_db->db_level == 0 &&
2960 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2961 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2962 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2963 dh->dh_dr = dh->dh_db->db_data_pending;
2964 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf)
2968 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2969 ASSERT(refcount_is_zero(&dh->dh_db->db_holds));
2970 ASSERT(dh->dh_db->db_caching_status == DB_DBUF_CACHE ||
2971 dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE);
2974 dbuf_caches[dh->dh_db->db_caching_status].cache,
2976 (void) refcount_remove_many(
2977 &dbuf_caches[dh->dh_db->db_caching_status].size,
2978 dh->dh_db->db.db_size, dh->dh_db);
2980 if (dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2981 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2983 DBUF_STAT_BUMPDOWN(cache_levels[dh->dh_db->db_level]);
2984 DBUF_STAT_BUMPDOWN(cache_count);
2985 DBUF_STAT_DECR(cache_levels_bytes[dh->dh_db->db_level],
2986 dh->dh_db->db.db_size);
2988 dh->dh_db->db_caching_status = DB_NO_CACHE;
2990 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
2991 DBUF_VERIFY(dh->dh_db);
2992 mutex_exit(&dh->dh_db->db_mtx);
2994 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2996 dbuf_rele(dh->dh_parent, NULL);
2998 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
2999 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
3000 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
3001 *(dh->dh_dbp) = dh->dh_db;
3007 * The following code preserves the recursive function dbuf_hold_impl()
3008 * but moves the local variables AND function arguments to the heap to
3009 * minimize the stack frame size. Enough space is initially allocated
3010 * on the stack for 20 levels of recursion.
3013 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3014 boolean_t fail_sparse, boolean_t fail_uncached,
3015 void *tag, dmu_buf_impl_t **dbp)
3017 struct dbuf_hold_impl_data *dh;
3020 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
3021 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
3022 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
3023 fail_uncached, tag, dbp, 0);
3025 error = __dbuf_hold_impl(dh);
3027 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
3028 DBUF_HOLD_IMPL_MAX_DEPTH);
3034 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
3035 dnode_t *dn, uint8_t level, uint64_t blkid,
3036 boolean_t fail_sparse, boolean_t fail_uncached,
3037 void *tag, dmu_buf_impl_t **dbp, int depth)
3040 dh->dh_level = level;
3041 dh->dh_blkid = blkid;
3043 dh->dh_fail_sparse = fail_sparse;
3044 dh->dh_fail_uncached = fail_uncached;
3050 dh->dh_parent = NULL;
3055 dh->dh_depth = depth;
3059 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3061 return (dbuf_hold_level(dn, 0, blkid, tag));
3065 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3068 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3069 return (err ? NULL : db);
3073 dbuf_create_bonus(dnode_t *dn)
3075 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3077 ASSERT(dn->dn_bonus == NULL);
3078 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3082 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3084 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3087 if (db->db_blkid != DMU_SPILL_BLKID)
3088 return (SET_ERROR(ENOTSUP));
3090 blksz = SPA_MINBLOCKSIZE;
3091 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3092 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3096 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3097 dbuf_new_size(db, blksz, tx);
3098 rw_exit(&dn->dn_struct_rwlock);
3105 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3107 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3110 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3112 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3114 int64_t holds = refcount_add(&db->db_holds, tag);
3115 ASSERT3S(holds, >, 1);
3118 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3120 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3123 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3124 dmu_buf_impl_t *found_db;
3125 boolean_t result = B_FALSE;
3127 if (db->db_blkid == DMU_BONUS_BLKID)
3128 found_db = dbuf_find_bonus(os, obj);
3130 found_db = dbuf_find(os, obj, 0, blkid);
3132 if (found_db != NULL) {
3133 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3134 (void) refcount_add(&db->db_holds, tag);
3137 mutex_exit(&db->db_mtx);
3143 * If you call dbuf_rele() you had better not be referencing the dnode handle
3144 * unless you have some other direct or indirect hold on the dnode. (An indirect
3145 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3146 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3147 * dnode's parent dbuf evicting its dnode handles.
3150 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3152 mutex_enter(&db->db_mtx);
3153 dbuf_rele_and_unlock(db, tag, B_FALSE);
3157 dmu_buf_rele(dmu_buf_t *db, void *tag)
3159 dbuf_rele((dmu_buf_impl_t *)db, tag);
3163 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3164 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3165 * argument should be set if we are already in the dbuf-evicting code
3166 * path, in which case we don't want to recursively evict. This allows us to
3167 * avoid deeply nested stacks that would have a call flow similar to this:
3169 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3172 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3176 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
3180 ASSERT(MUTEX_HELD(&db->db_mtx));
3184 * Remove the reference to the dbuf before removing its hold on the
3185 * dnode so we can guarantee in dnode_move() that a referenced bonus
3186 * buffer has a corresponding dnode hold.
3188 holds = refcount_remove(&db->db_holds, tag);
3192 * We can't freeze indirects if there is a possibility that they
3193 * may be modified in the current syncing context.
3195 if (db->db_buf != NULL &&
3196 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3197 arc_buf_freeze(db->db_buf);
3200 if (holds == db->db_dirtycnt &&
3201 db->db_level == 0 && db->db_user_immediate_evict)
3202 dbuf_evict_user(db);
3205 if (db->db_blkid == DMU_BONUS_BLKID) {
3207 boolean_t evict_dbuf = db->db_pending_evict;
3210 * If the dnode moves here, we cannot cross this
3211 * barrier until the move completes.
3216 atomic_dec_32(&dn->dn_dbufs_count);
3219 * Decrementing the dbuf count means that the bonus
3220 * buffer's dnode hold is no longer discounted in
3221 * dnode_move(). The dnode cannot move until after
3222 * the dnode_rele() below.
3227 * Do not reference db after its lock is dropped.
3228 * Another thread may evict it.
3230 mutex_exit(&db->db_mtx);
3233 dnode_evict_bonus(dn);
3236 } else if (db->db_buf == NULL) {
3238 * This is a special case: we never associated this
3239 * dbuf with any data allocated from the ARC.
3241 ASSERT(db->db_state == DB_UNCACHED ||
3242 db->db_state == DB_NOFILL);
3244 } else if (arc_released(db->db_buf)) {
3246 * This dbuf has anonymous data associated with it.
3250 boolean_t do_arc_evict = B_FALSE;
3252 spa_t *spa = dmu_objset_spa(db->db_objset);
3254 if (!DBUF_IS_CACHEABLE(db) &&
3255 db->db_blkptr != NULL &&
3256 !BP_IS_HOLE(db->db_blkptr) &&
3257 !BP_IS_EMBEDDED(db->db_blkptr)) {
3258 do_arc_evict = B_TRUE;
3259 bp = *db->db_blkptr;
3262 if (!DBUF_IS_CACHEABLE(db) ||
3263 db->db_pending_evict) {
3265 } else if (!multilist_link_active(&db->db_cache_link)) {
3266 ASSERT3U(db->db_caching_status, ==,
3269 dbuf_cached_state_t dcs =
3270 dbuf_include_in_metadata_cache(db) ?
3271 DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
3272 db->db_caching_status = dcs;
3274 multilist_insert(dbuf_caches[dcs].cache, db);
3275 (void) refcount_add_many(&dbuf_caches[dcs].size,
3276 db->db.db_size, db);
3278 if (dcs == DB_DBUF_METADATA_CACHE) {
3279 DBUF_STAT_BUMP(metadata_cache_count);
3281 metadata_cache_size_bytes_max,
3283 &dbuf_caches[dcs].size));
3286 cache_levels[db->db_level]);
3287 DBUF_STAT_BUMP(cache_count);
3289 cache_levels_bytes[db->db_level],
3291 DBUF_STAT_MAX(cache_size_bytes_max,
3293 &dbuf_caches[dcs].size));
3295 mutex_exit(&db->db_mtx);
3297 if (db->db_caching_status == DB_DBUF_CACHE &&
3299 dbuf_evict_notify();
3304 arc_freed(spa, &bp);
3307 mutex_exit(&db->db_mtx);
3312 #pragma weak dmu_buf_refcount = dbuf_refcount
3314 dbuf_refcount(dmu_buf_impl_t *db)
3316 return (refcount_count(&db->db_holds));
3320 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3321 dmu_buf_user_t *new_user)
3323 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3325 mutex_enter(&db->db_mtx);
3326 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3327 if (db->db_user == old_user)
3328 db->db_user = new_user;
3330 old_user = db->db_user;
3331 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3332 mutex_exit(&db->db_mtx);
3338 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3340 return (dmu_buf_replace_user(db_fake, NULL, user));
3344 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3346 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3348 db->db_user_immediate_evict = TRUE;
3349 return (dmu_buf_set_user(db_fake, user));
3353 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3355 return (dmu_buf_replace_user(db_fake, user, NULL));
3359 dmu_buf_get_user(dmu_buf_t *db_fake)
3361 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3363 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3364 return (db->db_user);
3368 dmu_buf_user_evict_wait()
3370 taskq_wait(dbu_evict_taskq);
3374 dmu_buf_get_blkptr(dmu_buf_t *db)
3376 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3377 return (dbi->db_blkptr);
3381 dmu_buf_get_objset(dmu_buf_t *db)
3383 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3384 return (dbi->db_objset);
3388 dmu_buf_dnode_enter(dmu_buf_t *db)
3390 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3391 DB_DNODE_ENTER(dbi);
3392 return (DB_DNODE(dbi));
3396 dmu_buf_dnode_exit(dmu_buf_t *db)
3398 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3403 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3405 /* ASSERT(dmu_tx_is_syncing(tx) */
3406 ASSERT(MUTEX_HELD(&db->db_mtx));
3408 if (db->db_blkptr != NULL)
3411 if (db->db_blkid == DMU_SPILL_BLKID) {
3412 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3413 BP_ZERO(db->db_blkptr);
3416 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3418 * This buffer was allocated at a time when there was
3419 * no available blkptrs from the dnode, or it was
3420 * inappropriate to hook it in (i.e., nlevels mis-match).
3422 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3423 ASSERT(db->db_parent == NULL);
3424 db->db_parent = dn->dn_dbuf;
3425 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3428 dmu_buf_impl_t *parent = db->db_parent;
3429 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3431 ASSERT(dn->dn_phys->dn_nlevels > 1);
3432 if (parent == NULL) {
3433 mutex_exit(&db->db_mtx);
3434 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3435 parent = dbuf_hold_level(dn, db->db_level + 1,
3436 db->db_blkid >> epbs, db);
3437 rw_exit(&dn->dn_struct_rwlock);
3438 mutex_enter(&db->db_mtx);
3439 db->db_parent = parent;
3441 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3442 (db->db_blkid & ((1ULL << epbs) - 1));
3448 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3449 * is critical the we not allow the compiler to inline this function in to
3450 * dbuf_sync_list() thereby drastically bloating the stack usage.
3452 noinline static void
3453 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3455 dmu_buf_impl_t *db = dr->dr_dbuf;
3459 ASSERT(dmu_tx_is_syncing(tx));
3461 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3463 mutex_enter(&db->db_mtx);
3465 ASSERT(db->db_level > 0);
3468 /* Read the block if it hasn't been read yet. */
3469 if (db->db_buf == NULL) {
3470 mutex_exit(&db->db_mtx);
3471 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3472 mutex_enter(&db->db_mtx);
3474 ASSERT3U(db->db_state, ==, DB_CACHED);
3475 ASSERT(db->db_buf != NULL);
3479 /* Indirect block size must match what the dnode thinks it is. */
3480 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3481 dbuf_check_blkptr(dn, db);
3484 /* Provide the pending dirty record to child dbufs */
3485 db->db_data_pending = dr;
3487 mutex_exit(&db->db_mtx);
3489 dbuf_write(dr, db->db_buf, tx);
3492 mutex_enter(&dr->dt.di.dr_mtx);
3493 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3494 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3495 mutex_exit(&dr->dt.di.dr_mtx);
3500 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3501 * critical the we not allow the compiler to inline this function in to
3502 * dbuf_sync_list() thereby drastically bloating the stack usage.
3504 noinline static void
3505 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3507 arc_buf_t **datap = &dr->dt.dl.dr_data;
3508 dmu_buf_impl_t *db = dr->dr_dbuf;
3511 uint64_t txg = tx->tx_txg;
3513 ASSERT(dmu_tx_is_syncing(tx));
3515 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3517 mutex_enter(&db->db_mtx);
3519 * To be synced, we must be dirtied. But we
3520 * might have been freed after the dirty.
3522 if (db->db_state == DB_UNCACHED) {
3523 /* This buffer has been freed since it was dirtied */
3524 ASSERT(db->db.db_data == NULL);
3525 } else if (db->db_state == DB_FILL) {
3526 /* This buffer was freed and is now being re-filled */
3527 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3529 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3536 if (db->db_blkid == DMU_SPILL_BLKID) {
3537 mutex_enter(&dn->dn_mtx);
3538 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3540 * In the previous transaction group, the bonus buffer
3541 * was entirely used to store the attributes for the
3542 * dnode which overrode the dn_spill field. However,
3543 * when adding more attributes to the file a spill
3544 * block was required to hold the extra attributes.
3546 * Make sure to clear the garbage left in the dn_spill
3547 * field from the previous attributes in the bonus
3548 * buffer. Otherwise, after writing out the spill
3549 * block to the new allocated dva, it will free
3550 * the old block pointed to by the invalid dn_spill.
3552 db->db_blkptr = NULL;
3554 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3555 mutex_exit(&dn->dn_mtx);
3559 * If this is a bonus buffer, simply copy the bonus data into the
3560 * dnode. It will be written out when the dnode is synced (and it
3561 * will be synced, since it must have been dirty for dbuf_sync to
3564 if (db->db_blkid == DMU_BONUS_BLKID) {
3565 dbuf_dirty_record_t **drp;
3567 ASSERT(*datap != NULL);
3568 ASSERT0(db->db_level);
3569 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3570 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3571 bcopy(*datap, DN_BONUS(dn->dn_phys),
3572 DN_MAX_BONUS_LEN(dn->dn_phys));
3575 if (*datap != db->db.db_data) {
3576 int slots = DB_DNODE(db)->dn_num_slots;
3577 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3578 zio_buf_free(*datap, bonuslen);
3579 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3581 db->db_data_pending = NULL;
3582 drp = &db->db_last_dirty;
3584 drp = &(*drp)->dr_next;
3585 ASSERT(dr->dr_next == NULL);
3586 ASSERT(dr->dr_dbuf == db);
3588 if (dr->dr_dbuf->db_level != 0) {
3589 mutex_destroy(&dr->dt.di.dr_mtx);
3590 list_destroy(&dr->dt.di.dr_children);
3592 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3593 ASSERT(db->db_dirtycnt > 0);
3594 db->db_dirtycnt -= 1;
3595 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
3602 * This function may have dropped the db_mtx lock allowing a dmu_sync
3603 * operation to sneak in. As a result, we need to ensure that we
3604 * don't check the dr_override_state until we have returned from
3605 * dbuf_check_blkptr.
3607 dbuf_check_blkptr(dn, db);
3610 * If this buffer is in the middle of an immediate write,
3611 * wait for the synchronous IO to complete.
3613 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3614 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3615 cv_wait(&db->db_changed, &db->db_mtx);
3616 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3619 if (db->db_state != DB_NOFILL &&
3620 dn->dn_object != DMU_META_DNODE_OBJECT &&
3621 refcount_count(&db->db_holds) > 1 &&
3622 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3623 *datap == db->db_buf) {
3625 * If this buffer is currently "in use" (i.e., there
3626 * are active holds and db_data still references it),
3627 * then make a copy before we start the write so that
3628 * any modifications from the open txg will not leak
3631 * NOTE: this copy does not need to be made for
3632 * objects only modified in the syncing context (e.g.
3633 * DNONE_DNODE blocks).
3635 int psize = arc_buf_size(*datap);
3636 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3637 enum zio_compress compress_type = arc_get_compression(*datap);
3639 if (compress_type == ZIO_COMPRESS_OFF) {
3640 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3642 ASSERT3U(type, ==, ARC_BUFC_DATA);
3643 int lsize = arc_buf_lsize(*datap);
3644 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3645 psize, lsize, compress_type);
3647 bcopy(db->db.db_data, (*datap)->b_data, psize);
3649 db->db_data_pending = dr;
3651 mutex_exit(&db->db_mtx);
3653 dbuf_write(dr, *datap, tx);
3655 ASSERT(!list_link_active(&dr->dr_dirty_node));
3656 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3657 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3661 * Although zio_nowait() does not "wait for an IO", it does
3662 * initiate the IO. If this is an empty write it seems plausible
3663 * that the IO could actually be completed before the nowait
3664 * returns. We need to DB_DNODE_EXIT() first in case
3665 * zio_nowait() invalidates the dbuf.
3668 zio_nowait(dr->dr_zio);
3673 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3675 dbuf_dirty_record_t *dr;
3677 while (dr = list_head(list)) {
3678 if (dr->dr_zio != NULL) {
3680 * If we find an already initialized zio then we
3681 * are processing the meta-dnode, and we have finished.
3682 * The dbufs for all dnodes are put back on the list
3683 * during processing, so that we can zio_wait()
3684 * these IOs after initiating all child IOs.
3686 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3687 DMU_META_DNODE_OBJECT);
3690 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3691 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3692 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3694 list_remove(list, dr);
3695 if (dr->dr_dbuf->db_level > 0)
3696 dbuf_sync_indirect(dr, tx);
3698 dbuf_sync_leaf(dr, tx);
3704 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3706 dmu_buf_impl_t *db = vdb;
3708 blkptr_t *bp = zio->io_bp;
3709 blkptr_t *bp_orig = &zio->io_bp_orig;
3710 spa_t *spa = zio->io_spa;
3715 ASSERT3P(db->db_blkptr, !=, NULL);
3716 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3720 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3721 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3722 zio->io_prev_space_delta = delta;
3724 if (bp->blk_birth != 0) {
3725 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3726 BP_GET_TYPE(bp) == dn->dn_type) ||
3727 (db->db_blkid == DMU_SPILL_BLKID &&
3728 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3729 BP_IS_EMBEDDED(bp));
3730 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3733 mutex_enter(&db->db_mtx);
3736 if (db->db_blkid == DMU_SPILL_BLKID) {
3737 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3738 ASSERT(!(BP_IS_HOLE(bp)) &&
3739 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3743 if (db->db_level == 0) {
3744 mutex_enter(&dn->dn_mtx);
3745 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3746 db->db_blkid != DMU_SPILL_BLKID)
3747 dn->dn_phys->dn_maxblkid = db->db_blkid;
3748 mutex_exit(&dn->dn_mtx);
3750 if (dn->dn_type == DMU_OT_DNODE) {
3752 while (i < db->db.db_size) {
3753 dnode_phys_t *dnp = db->db.db_data + i;
3755 i += DNODE_MIN_SIZE;
3756 if (dnp->dn_type != DMU_OT_NONE) {
3758 i += dnp->dn_extra_slots *
3763 if (BP_IS_HOLE(bp)) {
3770 blkptr_t *ibp = db->db.db_data;
3771 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3772 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3773 if (BP_IS_HOLE(ibp))
3775 fill += BP_GET_FILL(ibp);
3780 if (!BP_IS_EMBEDDED(bp))
3781 bp->blk_fill = fill;
3783 mutex_exit(&db->db_mtx);
3785 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3786 *db->db_blkptr = *bp;
3787 rw_exit(&dn->dn_struct_rwlock);
3792 * This function gets called just prior to running through the compression
3793 * stage of the zio pipeline. If we're an indirect block comprised of only
3794 * holes, then we want this indirect to be compressed away to a hole. In
3795 * order to do that we must zero out any information about the holes that
3796 * this indirect points to prior to before we try to compress it.
3799 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3801 dmu_buf_impl_t *db = vdb;
3804 unsigned int epbs, i;
3806 ASSERT3U(db->db_level, >, 0);
3809 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3810 ASSERT3U(epbs, <, 31);
3812 /* Determine if all our children are holes */
3813 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3814 if (!BP_IS_HOLE(bp))
3819 * If all the children are holes, then zero them all out so that
3820 * we may get compressed away.
3822 if (i == 1 << epbs) {
3824 * We only found holes. Grab the rwlock to prevent
3825 * anybody from reading the blocks we're about to
3828 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3829 bzero(db->db.db_data, db->db.db_size);
3830 rw_exit(&dn->dn_struct_rwlock);
3836 * The SPA will call this callback several times for each zio - once
3837 * for every physical child i/o (zio->io_phys_children times). This
3838 * allows the DMU to monitor the progress of each logical i/o. For example,
3839 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3840 * block. There may be a long delay before all copies/fragments are completed,
3841 * so this callback allows us to retire dirty space gradually, as the physical
3846 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3848 dmu_buf_impl_t *db = arg;
3849 objset_t *os = db->db_objset;
3850 dsl_pool_t *dp = dmu_objset_pool(os);
3851 dbuf_dirty_record_t *dr;
3854 dr = db->db_data_pending;
3855 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3858 * The callback will be called io_phys_children times. Retire one
3859 * portion of our dirty space each time we are called. Any rounding
3860 * error will be cleaned up by dsl_pool_sync()'s call to
3861 * dsl_pool_undirty_space().
3863 delta = dr->dr_accounted / zio->io_phys_children;
3864 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3869 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3871 dmu_buf_impl_t *db = vdb;
3872 blkptr_t *bp_orig = &zio->io_bp_orig;
3873 blkptr_t *bp = db->db_blkptr;
3874 objset_t *os = db->db_objset;
3875 dmu_tx_t *tx = os->os_synctx;
3876 dbuf_dirty_record_t **drp, *dr;
3878 ASSERT0(zio->io_error);
3879 ASSERT(db->db_blkptr == bp);
3882 * For nopwrites and rewrites we ensure that the bp matches our
3883 * original and bypass all the accounting.
3885 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3886 ASSERT(BP_EQUAL(bp, bp_orig));
3888 dsl_dataset_t *ds = os->os_dsl_dataset;
3889 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3890 dsl_dataset_block_born(ds, bp, tx);
3893 mutex_enter(&db->db_mtx);
3897 drp = &db->db_last_dirty;
3898 while ((dr = *drp) != db->db_data_pending)
3900 ASSERT(!list_link_active(&dr->dr_dirty_node));
3901 ASSERT(dr->dr_dbuf == db);
3902 ASSERT(dr->dr_next == NULL);
3906 if (db->db_blkid == DMU_SPILL_BLKID) {
3911 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3912 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3913 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3918 if (db->db_level == 0) {
3919 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3920 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3921 if (db->db_state != DB_NOFILL) {
3922 if (dr->dt.dl.dr_data != db->db_buf)
3923 arc_buf_destroy(dr->dt.dl.dr_data, db);
3930 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3931 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3932 if (!BP_IS_HOLE(db->db_blkptr)) {
3934 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3935 ASSERT3U(db->db_blkid, <=,
3936 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3937 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3941 mutex_destroy(&dr->dt.di.dr_mtx);
3942 list_destroy(&dr->dt.di.dr_children);
3944 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3946 cv_broadcast(&db->db_changed);
3947 ASSERT(db->db_dirtycnt > 0);
3948 db->db_dirtycnt -= 1;
3949 db->db_data_pending = NULL;
3950 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
3954 dbuf_write_nofill_ready(zio_t *zio)
3956 dbuf_write_ready(zio, NULL, zio->io_private);
3960 dbuf_write_nofill_done(zio_t *zio)
3962 dbuf_write_done(zio, NULL, zio->io_private);
3966 dbuf_write_override_ready(zio_t *zio)
3968 dbuf_dirty_record_t *dr = zio->io_private;
3969 dmu_buf_impl_t *db = dr->dr_dbuf;
3971 dbuf_write_ready(zio, NULL, db);
3975 dbuf_write_override_done(zio_t *zio)
3977 dbuf_dirty_record_t *dr = zio->io_private;
3978 dmu_buf_impl_t *db = dr->dr_dbuf;
3979 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3981 mutex_enter(&db->db_mtx);
3982 if (!BP_EQUAL(zio->io_bp, obp)) {
3983 if (!BP_IS_HOLE(obp))
3984 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3985 arc_release(dr->dt.dl.dr_data, db);
3987 mutex_exit(&db->db_mtx);
3988 dbuf_write_done(zio, NULL, db);
3990 if (zio->io_abd != NULL)
3991 abd_put(zio->io_abd);
3994 typedef struct dbuf_remap_impl_callback_arg {
3996 uint64_t drica_blk_birth;
3998 } dbuf_remap_impl_callback_arg_t;
4001 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
4004 dbuf_remap_impl_callback_arg_t *drica = arg;
4005 objset_t *os = drica->drica_os;
4006 spa_t *spa = dmu_objset_spa(os);
4007 dmu_tx_t *tx = drica->drica_tx;
4009 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4011 if (os == spa_meta_objset(spa)) {
4012 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
4014 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
4015 size, drica->drica_blk_birth, tx);
4020 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
4022 blkptr_t bp_copy = *bp;
4023 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4024 dbuf_remap_impl_callback_arg_t drica;
4026 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4028 drica.drica_os = dn->dn_objset;
4029 drica.drica_blk_birth = bp->blk_birth;
4030 drica.drica_tx = tx;
4031 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
4034 * The struct_rwlock prevents dbuf_read_impl() from
4035 * dereferencing the BP while we are changing it. To
4036 * avoid lock contention, only grab it when we are actually
4039 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
4041 rw_exit(&dn->dn_struct_rwlock);
4046 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
4047 * to remap a copy of every bp in the dbuf.
4050 dbuf_can_remap(const dmu_buf_impl_t *db)
4052 spa_t *spa = dmu_objset_spa(db->db_objset);
4053 blkptr_t *bp = db->db.db_data;
4054 boolean_t ret = B_FALSE;
4056 ASSERT3U(db->db_level, >, 0);
4057 ASSERT3S(db->db_state, ==, DB_CACHED);
4059 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4061 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4062 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4063 blkptr_t bp_copy = bp[i];
4064 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4069 spa_config_exit(spa, SCL_VDEV, FTAG);
4075 dnode_needs_remap(const dnode_t *dn)
4077 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4078 boolean_t ret = B_FALSE;
4080 if (dn->dn_phys->dn_nlevels == 0) {
4084 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4086 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4087 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
4088 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
4089 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4094 spa_config_exit(spa, SCL_VDEV, FTAG);
4100 * Remap any existing BP's to concrete vdevs, if possible.
4103 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
4105 spa_t *spa = dmu_objset_spa(db->db_objset);
4106 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4108 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
4111 if (db->db_level > 0) {
4112 blkptr_t *bp = db->db.db_data;
4113 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4114 dbuf_remap_impl(dn, &bp[i], tx);
4116 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
4117 dnode_phys_t *dnp = db->db.db_data;
4118 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
4120 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT;
4121 i += dnp[i].dn_extra_slots + 1) {
4122 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
4123 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
4130 /* Issue I/O to commit a dirty buffer to disk. */
4132 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4134 dmu_buf_impl_t *db = dr->dr_dbuf;
4137 dmu_buf_impl_t *parent = db->db_parent;
4138 uint64_t txg = tx->tx_txg;
4139 zbookmark_phys_t zb;
4144 ASSERT(dmu_tx_is_syncing(tx));
4150 if (db->db_state != DB_NOFILL) {
4151 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4153 * Private object buffers are released here rather
4154 * than in dbuf_dirty() since they are only modified
4155 * in the syncing context and we don't want the
4156 * overhead of making multiple copies of the data.
4158 if (BP_IS_HOLE(db->db_blkptr)) {
4161 dbuf_release_bp(db);
4163 dbuf_remap(dn, db, tx);
4167 if (parent != dn->dn_dbuf) {
4168 /* Our parent is an indirect block. */
4169 /* We have a dirty parent that has been scheduled for write. */
4170 ASSERT(parent && parent->db_data_pending);
4171 /* Our parent's buffer is one level closer to the dnode. */
4172 ASSERT(db->db_level == parent->db_level-1);
4174 * We're about to modify our parent's db_data by modifying
4175 * our block pointer, so the parent must be released.
4177 ASSERT(arc_released(parent->db_buf));
4178 zio = parent->db_data_pending->dr_zio;
4180 /* Our parent is the dnode itself. */
4181 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4182 db->db_blkid != DMU_SPILL_BLKID) ||
4183 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4184 if (db->db_blkid != DMU_SPILL_BLKID)
4185 ASSERT3P(db->db_blkptr, ==,
4186 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4190 ASSERT(db->db_level == 0 || data == db->db_buf);
4191 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4194 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4195 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4196 db->db.db_object, db->db_level, db->db_blkid);
4198 if (db->db_blkid == DMU_SPILL_BLKID)
4200 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
4202 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
4206 * We copy the blkptr now (rather than when we instantiate the dirty
4207 * record), because its value can change between open context and
4208 * syncing context. We do not need to hold dn_struct_rwlock to read
4209 * db_blkptr because we are in syncing context.
4211 dr->dr_bp_copy = *db->db_blkptr;
4213 if (db->db_level == 0 &&
4214 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
4216 * The BP for this block has been provided by open context
4217 * (by dmu_sync() or dmu_buf_write_embedded()).
4219 abd_t *contents = (data != NULL) ?
4220 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
4222 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
4223 contents, db->db.db_size, db->db.db_size, &zp,
4224 dbuf_write_override_ready, NULL, NULL,
4225 dbuf_write_override_done,
4226 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
4227 mutex_enter(&db->db_mtx);
4228 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
4229 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
4230 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
4231 mutex_exit(&db->db_mtx);
4232 } else if (db->db_state == DB_NOFILL) {
4233 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
4234 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
4235 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4236 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
4237 dbuf_write_nofill_ready, NULL, NULL,
4238 dbuf_write_nofill_done, db,
4239 ZIO_PRIORITY_ASYNC_WRITE,
4240 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
4242 ASSERT(arc_released(data));
4245 * For indirect blocks, we want to setup the children
4246 * ready callback so that we can properly handle an indirect
4247 * block that only contains holes.
4249 arc_write_done_func_t *children_ready_cb = NULL;
4250 if (db->db_level != 0)
4251 children_ready_cb = dbuf_write_children_ready;
4253 dr->dr_zio = arc_write(zio, os->os_spa, txg,
4254 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
4255 &zp, dbuf_write_ready, children_ready_cb,
4256 dbuf_write_physdone, dbuf_write_done, db,
4257 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);