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_QUAD(_vfs_zfs, OID_AUTO, dbuf_metadata_cache_max_bytes, CTLFLAG_RWTUN,
296 &dbuf_metadata_cache_max_bytes, 0, "dbuf metadata cache size in bytes");
297 SYSCTL_INT(_vfs_zfs, OID_AUTO, dbuf_cache_shift, CTLFLAG_RDTUN,
298 &dbuf_cache_shift, 0, "dbuf cache size as log2 fraction of ARC");
299 SYSCTL_INT(_vfs_zfs, OID_AUTO, dbuf_metadata_cache_shift, CTLFLAG_RDTUN,
300 &dbuf_metadata_cache_shift, 0,
301 "dbuf metadata cache size as log2 fraction of ARC");
302 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, dbuf_metadata_cache_overflow, CTLFLAG_RD,
303 &dbuf_metadata_cache_overflow, 0, "dbuf metadata cache overflow");
304 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_hiwater_pct, CTLFLAG_RWTUN,
305 &dbuf_cache_hiwater_pct, 0, "max percents above the dbuf cache size");
306 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_lowater_pct, CTLFLAG_RWTUN,
307 &dbuf_cache_lowater_pct, 0, "max percents below the dbuf cache size");
311 dbuf_cons(void *vdb, void *unused, int kmflag)
313 dmu_buf_impl_t *db = vdb;
314 bzero(db, sizeof (dmu_buf_impl_t));
316 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
317 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
318 multilist_link_init(&db->db_cache_link);
319 zfs_refcount_create(&db->db_holds);
326 dbuf_dest(void *vdb, void *unused)
328 dmu_buf_impl_t *db = vdb;
329 mutex_destroy(&db->db_mtx);
330 cv_destroy(&db->db_changed);
331 ASSERT(!multilist_link_active(&db->db_cache_link));
332 zfs_refcount_destroy(&db->db_holds);
336 * dbuf hash table routines
338 static dbuf_hash_table_t dbuf_hash_table;
340 static uint64_t dbuf_hash_count;
343 * We use Cityhash for this. It's fast, and has good hash properties without
344 * requiring any large static buffers.
347 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
349 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid));
352 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
353 ((dbuf)->db.db_object == (obj) && \
354 (dbuf)->db_objset == (os) && \
355 (dbuf)->db_level == (level) && \
356 (dbuf)->db_blkid == (blkid))
359 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
361 dbuf_hash_table_t *h = &dbuf_hash_table;
362 uint64_t hv = dbuf_hash(os, obj, level, blkid);
363 uint64_t idx = hv & h->hash_table_mask;
366 mutex_enter(DBUF_HASH_MUTEX(h, idx));
367 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
368 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
369 mutex_enter(&db->db_mtx);
370 if (db->db_state != DB_EVICTING) {
371 mutex_exit(DBUF_HASH_MUTEX(h, idx));
374 mutex_exit(&db->db_mtx);
377 mutex_exit(DBUF_HASH_MUTEX(h, idx));
381 static dmu_buf_impl_t *
382 dbuf_find_bonus(objset_t *os, uint64_t object)
385 dmu_buf_impl_t *db = NULL;
387 if (dnode_hold(os, object, FTAG, &dn) == 0) {
388 rw_enter(&dn->dn_struct_rwlock, RW_READER);
389 if (dn->dn_bonus != NULL) {
391 mutex_enter(&db->db_mtx);
393 rw_exit(&dn->dn_struct_rwlock);
394 dnode_rele(dn, FTAG);
400 * Insert an entry into the hash table. If there is already an element
401 * equal to elem in the hash table, then the already existing element
402 * will be returned and the new element will not be inserted.
403 * Otherwise returns NULL.
405 static dmu_buf_impl_t *
406 dbuf_hash_insert(dmu_buf_impl_t *db)
408 dbuf_hash_table_t *h = &dbuf_hash_table;
409 objset_t *os = db->db_objset;
410 uint64_t obj = db->db.db_object;
411 int level = db->db_level;
412 uint64_t blkid, hv, idx;
416 blkid = db->db_blkid;
417 hv = dbuf_hash(os, obj, level, blkid);
418 idx = hv & h->hash_table_mask;
420 mutex_enter(DBUF_HASH_MUTEX(h, idx));
421 for (dbf = h->hash_table[idx], i = 0; dbf != NULL;
422 dbf = dbf->db_hash_next, i++) {
423 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
424 mutex_enter(&dbf->db_mtx);
425 if (dbf->db_state != DB_EVICTING) {
426 mutex_exit(DBUF_HASH_MUTEX(h, idx));
429 mutex_exit(&dbf->db_mtx);
434 DBUF_STAT_BUMP(hash_collisions);
436 DBUF_STAT_BUMP(hash_chains);
438 DBUF_STAT_MAX(hash_chain_max, i);
441 mutex_enter(&db->db_mtx);
442 db->db_hash_next = h->hash_table[idx];
443 h->hash_table[idx] = db;
444 mutex_exit(DBUF_HASH_MUTEX(h, idx));
445 atomic_inc_64(&dbuf_hash_count);
446 DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count);
452 * Remove an entry from the hash table. It must be in the EVICTING state.
455 dbuf_hash_remove(dmu_buf_impl_t *db)
457 dbuf_hash_table_t *h = &dbuf_hash_table;
459 dmu_buf_impl_t *dbf, **dbp;
461 hv = dbuf_hash(db->db_objset, db->db.db_object,
462 db->db_level, db->db_blkid);
463 idx = hv & h->hash_table_mask;
466 * We mustn't hold db_mtx to maintain lock ordering:
467 * DBUF_HASH_MUTEX > db_mtx.
469 ASSERT(zfs_refcount_is_zero(&db->db_holds));
470 ASSERT(db->db_state == DB_EVICTING);
471 ASSERT(!MUTEX_HELD(&db->db_mtx));
473 mutex_enter(DBUF_HASH_MUTEX(h, idx));
474 dbp = &h->hash_table[idx];
475 while ((dbf = *dbp) != db) {
476 dbp = &dbf->db_hash_next;
479 *dbp = db->db_hash_next;
480 db->db_hash_next = NULL;
481 if (h->hash_table[idx] &&
482 h->hash_table[idx]->db_hash_next == NULL)
483 DBUF_STAT_BUMPDOWN(hash_chains);
484 mutex_exit(DBUF_HASH_MUTEX(h, idx));
485 atomic_dec_64(&dbuf_hash_count);
491 } dbvu_verify_type_t;
494 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
499 if (db->db_user == NULL)
502 /* Only data blocks support the attachment of user data. */
503 ASSERT(db->db_level == 0);
505 /* Clients must resolve a dbuf before attaching user data. */
506 ASSERT(db->db.db_data != NULL);
507 ASSERT3U(db->db_state, ==, DB_CACHED);
509 holds = zfs_refcount_count(&db->db_holds);
510 if (verify_type == DBVU_EVICTING) {
512 * Immediate eviction occurs when holds == dirtycnt.
513 * For normal eviction buffers, holds is zero on
514 * eviction, except when dbuf_fix_old_data() calls
515 * dbuf_clear_data(). However, the hold count can grow
516 * during eviction even though db_mtx is held (see
517 * dmu_bonus_hold() for an example), so we can only
518 * test the generic invariant that holds >= dirtycnt.
520 ASSERT3U(holds, >=, db->db_dirtycnt);
522 if (db->db_user_immediate_evict == TRUE)
523 ASSERT3U(holds, >=, db->db_dirtycnt);
525 ASSERT3U(holds, >, 0);
531 dbuf_evict_user(dmu_buf_impl_t *db)
533 dmu_buf_user_t *dbu = db->db_user;
535 ASSERT(MUTEX_HELD(&db->db_mtx));
540 dbuf_verify_user(db, DBVU_EVICTING);
544 if (dbu->dbu_clear_on_evict_dbufp != NULL)
545 *dbu->dbu_clear_on_evict_dbufp = NULL;
549 * There are two eviction callbacks - one that we call synchronously
550 * and one that we invoke via a taskq. The async one is useful for
551 * avoiding lock order reversals and limiting stack depth.
553 * Note that if we have a sync callback but no async callback,
554 * it's likely that the sync callback will free the structure
555 * containing the dbu. In that case we need to take care to not
556 * dereference dbu after calling the sync evict func.
558 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
560 if (dbu->dbu_evict_func_sync != NULL)
561 dbu->dbu_evict_func_sync(dbu);
564 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
565 dbu, 0, &dbu->dbu_tqent);
570 dbuf_is_metadata(dmu_buf_impl_t *db)
572 if (db->db_level > 0) {
575 boolean_t is_metadata;
578 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
581 return (is_metadata);
586 * This returns whether this dbuf should be stored in the metadata cache, which
587 * is based on whether it's from one of the dnode types that store data related
588 * to traversing dataset hierarchies.
591 dbuf_include_in_metadata_cache(dmu_buf_impl_t *db)
594 dmu_object_type_t type = DB_DNODE(db)->dn_type;
597 /* Check if this dbuf is one of the types we care about */
598 if (DMU_OT_IS_METADATA_CACHED(type)) {
599 /* If we hit this, then we set something up wrong in dmu_ot */
600 ASSERT(DMU_OT_IS_METADATA(type));
603 * Sanity check for small-memory systems: don't allocate too
604 * much memory for this purpose.
606 if (zfs_refcount_count(
607 &dbuf_caches[DB_DBUF_METADATA_CACHE].size) >
608 dbuf_metadata_cache_max_bytes) {
609 dbuf_metadata_cache_overflow++;
610 DTRACE_PROBE1(dbuf__metadata__cache__overflow,
611 dmu_buf_impl_t *, db);
622 * This function *must* return indices evenly distributed between all
623 * sublists of the multilist. This is needed due to how the dbuf eviction
624 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
625 * distributed between all sublists and uses this assumption when
626 * deciding which sublist to evict from and how much to evict from it.
629 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
631 dmu_buf_impl_t *db = obj;
634 * The assumption here, is the hash value for a given
635 * dmu_buf_impl_t will remain constant throughout it's lifetime
636 * (i.e. it's objset, object, level and blkid fields don't change).
637 * Thus, we don't need to store the dbuf's sublist index
638 * on insertion, as this index can be recalculated on removal.
640 * Also, the low order bits of the hash value are thought to be
641 * distributed evenly. Otherwise, in the case that the multilist
642 * has a power of two number of sublists, each sublists' usage
643 * would not be evenly distributed.
645 return (dbuf_hash(db->db_objset, db->db.db_object,
646 db->db_level, db->db_blkid) %
647 multilist_get_num_sublists(ml));
650 static inline unsigned long
651 dbuf_cache_target_bytes(void)
653 return MIN(dbuf_cache_max_bytes,
654 arc_max_bytes() >> dbuf_cache_shift);
657 static inline uint64_t
658 dbuf_cache_hiwater_bytes(void)
660 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
661 return (dbuf_cache_target +
662 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100);
665 static inline uint64_t
666 dbuf_cache_lowater_bytes(void)
668 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
669 return (dbuf_cache_target -
670 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100);
673 static inline boolean_t
674 dbuf_cache_above_hiwater(void)
676 return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
677 dbuf_cache_hiwater_bytes());
680 static inline boolean_t
681 dbuf_cache_above_lowater(void)
683 return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
684 dbuf_cache_lowater_bytes());
688 * Evict the oldest eligible dbuf from the dbuf cache.
693 int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache);
694 multilist_sublist_t *mls = multilist_sublist_lock(
695 dbuf_caches[DB_DBUF_CACHE].cache, idx);
697 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
699 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
700 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
701 db = multilist_sublist_prev(mls, db);
704 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
705 multilist_sublist_t *, mls);
708 multilist_sublist_remove(mls, db);
709 multilist_sublist_unlock(mls);
710 (void) zfs_refcount_remove_many(
711 &dbuf_caches[DB_DBUF_CACHE].size,
713 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
714 DBUF_STAT_BUMPDOWN(cache_count);
715 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
717 ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE);
718 db->db_caching_status = DB_NO_CACHE;
720 DBUF_STAT_MAX(cache_size_bytes_max,
721 zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size));
722 DBUF_STAT_BUMP(cache_total_evicts);
724 multilist_sublist_unlock(mls);
729 * The dbuf evict thread is responsible for aging out dbufs from the
730 * cache. Once the cache has reached it's maximum size, dbufs are removed
731 * and destroyed. The eviction thread will continue running until the size
732 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
733 * out of the cache it is destroyed and becomes eligible for arc eviction.
737 dbuf_evict_thread(void *unused __unused)
741 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
743 mutex_enter(&dbuf_evict_lock);
744 while (!dbuf_evict_thread_exit) {
745 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
746 CALLB_CPR_SAFE_BEGIN(&cpr);
747 (void) cv_timedwait_hires(&dbuf_evict_cv,
748 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
749 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
751 mutex_exit(&dbuf_evict_lock);
754 * Keep evicting as long as we're above the low water mark
755 * for the cache. We do this without holding the locks to
756 * minimize lock contention.
758 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
762 mutex_enter(&dbuf_evict_lock);
765 dbuf_evict_thread_exit = B_FALSE;
766 cv_broadcast(&dbuf_evict_cv);
767 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
772 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
773 * If the dbuf cache is at its high water mark, then evict a dbuf from the
774 * dbuf cache using the callers context.
777 dbuf_evict_notify(void)
780 * We check if we should evict without holding the dbuf_evict_lock,
781 * because it's OK to occasionally make the wrong decision here,
782 * and grabbing the lock results in massive lock contention.
784 if (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
785 dbuf_cache_max_bytes) {
786 if (dbuf_cache_above_hiwater())
788 cv_signal(&dbuf_evict_cv);
793 dbuf_kstat_update(kstat_t *ksp, int rw)
795 dbuf_stats_t *ds = ksp->ks_data;
797 if (rw == KSTAT_WRITE) {
798 return (SET_ERROR(EACCES));
800 ds->metadata_cache_size_bytes.value.ui64 =
801 zfs_refcount_count(&dbuf_caches[DB_DBUF_METADATA_CACHE].size);
802 ds->cache_size_bytes.value.ui64 =
803 zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size);
804 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
805 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
806 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
807 ds->hash_elements.value.ui64 = dbuf_hash_count;
816 uint64_t hsize = 1ULL << 16;
817 dbuf_hash_table_t *h = &dbuf_hash_table;
821 * The hash table is big enough to fill all of physical memory
822 * with an average 4K block size. The table will take up
823 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
825 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
829 h->hash_table_mask = hsize - 1;
830 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
831 if (h->hash_table == NULL) {
832 /* XXX - we should really return an error instead of assert */
833 ASSERT(hsize > (1ULL << 10));
838 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
839 sizeof (dmu_buf_impl_t),
840 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
842 for (i = 0; i < DBUF_MUTEXES; i++)
843 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
847 * Setup the parameters for the dbuf caches. We set the sizes of the
848 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
849 * of the size of the ARC, respectively. If the values are set in
850 * /etc/system and they're not greater than the size of the ARC, then
851 * we honor that value.
853 if (dbuf_cache_max_bytes == 0 ||
854 dbuf_cache_max_bytes >= arc_max_bytes()) {
855 dbuf_cache_max_bytes = arc_max_bytes() >> dbuf_cache_shift;
857 if (dbuf_metadata_cache_max_bytes == 0 ||
858 dbuf_metadata_cache_max_bytes >= arc_max_bytes()) {
859 dbuf_metadata_cache_max_bytes =
860 arc_max_bytes() >> dbuf_metadata_cache_shift;
864 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
865 * configuration is not required.
867 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
869 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
870 dbuf_caches[dcs].cache =
871 multilist_create(sizeof (dmu_buf_impl_t),
872 offsetof(dmu_buf_impl_t, db_cache_link),
873 dbuf_cache_multilist_index_func);
874 zfs_refcount_create(&dbuf_caches[dcs].size);
877 dbuf_evict_thread_exit = B_FALSE;
878 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
879 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
880 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
881 NULL, 0, &p0, TS_RUN, minclsyspri);
885 * XXX FreeBSD's SPL lacks KSTAT_TYPE_NAMED support - TODO
887 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
888 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
890 if (dbuf_ksp != NULL) {
891 dbuf_ksp->ks_data = &dbuf_stats;
892 dbuf_ksp->ks_update = dbuf_kstat_update;
893 kstat_install(dbuf_ksp);
895 for (i = 0; i < DN_MAX_LEVELS; i++) {
896 snprintf(dbuf_stats.cache_levels[i].name,
897 KSTAT_STRLEN, "cache_level_%d", i);
898 dbuf_stats.cache_levels[i].data_type =
900 snprintf(dbuf_stats.cache_levels_bytes[i].name,
901 KSTAT_STRLEN, "cache_level_%d_bytes", i);
902 dbuf_stats.cache_levels_bytes[i].data_type =
912 dbuf_hash_table_t *h = &dbuf_hash_table;
915 dbuf_stats_destroy();
917 for (i = 0; i < DBUF_MUTEXES; i++)
918 mutex_destroy(&h->hash_mutexes[i]);
919 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
920 kmem_cache_destroy(dbuf_kmem_cache);
921 taskq_destroy(dbu_evict_taskq);
923 mutex_enter(&dbuf_evict_lock);
924 dbuf_evict_thread_exit = B_TRUE;
925 while (dbuf_evict_thread_exit) {
926 cv_signal(&dbuf_evict_cv);
927 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
929 mutex_exit(&dbuf_evict_lock);
931 mutex_destroy(&dbuf_evict_lock);
932 cv_destroy(&dbuf_evict_cv);
934 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
935 zfs_refcount_destroy(&dbuf_caches[dcs].size);
936 multilist_destroy(dbuf_caches[dcs].cache);
939 if (dbuf_ksp != NULL) {
940 kstat_delete(dbuf_ksp);
951 dbuf_verify(dmu_buf_impl_t *db)
954 dbuf_dirty_record_t *dr;
956 ASSERT(MUTEX_HELD(&db->db_mtx));
958 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
961 ASSERT(db->db_objset != NULL);
965 ASSERT(db->db_parent == NULL);
966 ASSERT(db->db_blkptr == NULL);
968 ASSERT3U(db->db.db_object, ==, dn->dn_object);
969 ASSERT3P(db->db_objset, ==, dn->dn_objset);
970 ASSERT3U(db->db_level, <, dn->dn_nlevels);
971 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
972 db->db_blkid == DMU_SPILL_BLKID ||
973 !avl_is_empty(&dn->dn_dbufs));
975 if (db->db_blkid == DMU_BONUS_BLKID) {
977 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
978 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
979 } else if (db->db_blkid == DMU_SPILL_BLKID) {
981 ASSERT0(db->db.db_offset);
983 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
986 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
987 ASSERT(dr->dr_dbuf == db);
989 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
990 ASSERT(dr->dr_dbuf == db);
993 * We can't assert that db_size matches dn_datablksz because it
994 * can be momentarily different when another thread is doing
997 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
998 dr = db->db_data_pending;
1000 * It should only be modified in syncing context, so
1001 * make sure we only have one copy of the data.
1003 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
1006 /* verify db->db_blkptr */
1007 if (db->db_blkptr) {
1008 if (db->db_parent == dn->dn_dbuf) {
1009 /* db is pointed to by the dnode */
1010 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1011 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
1012 ASSERT(db->db_parent == NULL);
1014 ASSERT(db->db_parent != NULL);
1015 if (db->db_blkid != DMU_SPILL_BLKID)
1016 ASSERT3P(db->db_blkptr, ==,
1017 &dn->dn_phys->dn_blkptr[db->db_blkid]);
1019 /* db is pointed to by an indirect block */
1020 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
1021 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
1022 ASSERT3U(db->db_parent->db.db_object, ==,
1025 * dnode_grow_indblksz() can make this fail if we don't
1026 * have the struct_rwlock. XXX indblksz no longer
1027 * grows. safe to do this now?
1029 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1030 ASSERT3P(db->db_blkptr, ==,
1031 ((blkptr_t *)db->db_parent->db.db_data +
1032 db->db_blkid % epb));
1036 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
1037 (db->db_buf == NULL || db->db_buf->b_data) &&
1038 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
1039 db->db_state != DB_FILL && !dn->dn_free_txg) {
1041 * If the blkptr isn't set but they have nonzero data,
1042 * it had better be dirty, otherwise we'll lose that
1043 * data when we evict this buffer.
1045 * There is an exception to this rule for indirect blocks; in
1046 * this case, if the indirect block is a hole, we fill in a few
1047 * fields on each of the child blocks (importantly, birth time)
1048 * to prevent hole birth times from being lost when you
1049 * partially fill in a hole.
1051 if (db->db_dirtycnt == 0) {
1052 if (db->db_level == 0) {
1053 uint64_t *buf = db->db.db_data;
1056 for (i = 0; i < db->db.db_size >> 3; i++) {
1057 ASSERT(buf[i] == 0);
1060 blkptr_t *bps = db->db.db_data;
1061 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
1064 * We want to verify that all the blkptrs in the
1065 * indirect block are holes, but we may have
1066 * automatically set up a few fields for them.
1067 * We iterate through each blkptr and verify
1068 * they only have those fields set.
1071 i < db->db.db_size / sizeof (blkptr_t);
1073 blkptr_t *bp = &bps[i];
1074 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1077 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
1078 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
1079 DVA_IS_EMPTY(&bp->blk_dva[2]));
1080 ASSERT0(bp->blk_fill);
1081 ASSERT0(bp->blk_pad[0]);
1082 ASSERT0(bp->blk_pad[1]);
1083 ASSERT(!BP_IS_EMBEDDED(bp));
1084 ASSERT(BP_IS_HOLE(bp));
1085 ASSERT0(bp->blk_phys_birth);
1095 dbuf_clear_data(dmu_buf_impl_t *db)
1097 ASSERT(MUTEX_HELD(&db->db_mtx));
1098 dbuf_evict_user(db);
1099 ASSERT3P(db->db_buf, ==, NULL);
1100 db->db.db_data = NULL;
1101 if (db->db_state != DB_NOFILL)
1102 db->db_state = DB_UNCACHED;
1106 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1108 ASSERT(MUTEX_HELD(&db->db_mtx));
1109 ASSERT(buf != NULL);
1112 ASSERT(buf->b_data != NULL);
1113 db->db.db_data = buf->b_data;
1117 * Loan out an arc_buf for read. Return the loaned arc_buf.
1120 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1124 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1125 mutex_enter(&db->db_mtx);
1126 if (arc_released(db->db_buf) || zfs_refcount_count(&db->db_holds) > 1) {
1127 int blksz = db->db.db_size;
1128 spa_t *spa = db->db_objset->os_spa;
1130 mutex_exit(&db->db_mtx);
1131 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1132 bcopy(db->db.db_data, abuf->b_data, blksz);
1135 arc_loan_inuse_buf(abuf, db);
1137 dbuf_clear_data(db);
1138 mutex_exit(&db->db_mtx);
1144 * Calculate which level n block references the data at the level 0 offset
1148 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
1150 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1152 * The level n blkid is equal to the level 0 blkid divided by
1153 * the number of level 0s in a level n block.
1155 * The level 0 blkid is offset >> datablkshift =
1156 * offset / 2^datablkshift.
1158 * The number of level 0s in a level n is the number of block
1159 * pointers in an indirect block, raised to the power of level.
1160 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1161 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1163 * Thus, the level n blkid is: offset /
1164 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
1165 * = offset / 2^(datablkshift + level *
1166 * (indblkshift - SPA_BLKPTRSHIFT))
1167 * = offset >> (datablkshift + level *
1168 * (indblkshift - SPA_BLKPTRSHIFT))
1170 return (offset >> (dn->dn_datablkshift + level *
1171 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
1173 ASSERT3U(offset, <, dn->dn_datablksz);
1179 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1180 arc_buf_t *buf, void *vdb)
1182 dmu_buf_impl_t *db = vdb;
1184 mutex_enter(&db->db_mtx);
1185 ASSERT3U(db->db_state, ==, DB_READ);
1187 * All reads are synchronous, so we must have a hold on the dbuf
1189 ASSERT(zfs_refcount_count(&db->db_holds) > 0);
1190 ASSERT(db->db_buf == NULL);
1191 ASSERT(db->db.db_data == NULL);
1194 ASSERT(zio == NULL || zio->io_error != 0);
1195 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1196 ASSERT3P(db->db_buf, ==, NULL);
1197 db->db_state = DB_UNCACHED;
1198 } else if (db->db_level == 0 && db->db_freed_in_flight) {
1199 /* freed in flight */
1200 ASSERT(zio == NULL || zio->io_error == 0);
1202 buf = arc_alloc_buf(db->db_objset->os_spa,
1203 db, DBUF_GET_BUFC_TYPE(db), db->db.db_size);
1205 arc_release(buf, db);
1206 bzero(buf->b_data, db->db.db_size);
1207 arc_buf_freeze(buf);
1208 db->db_freed_in_flight = FALSE;
1209 dbuf_set_data(db, buf);
1210 db->db_state = DB_CACHED;
1213 ASSERT(zio == NULL || zio->io_error == 0);
1214 dbuf_set_data(db, buf);
1215 db->db_state = DB_CACHED;
1217 cv_broadcast(&db->db_changed);
1218 dbuf_rele_and_unlock(db, NULL, B_FALSE);
1222 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1225 zbookmark_phys_t zb;
1226 arc_flags_t aflags = ARC_FLAG_NOWAIT;
1230 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1231 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1232 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1233 ASSERT(MUTEX_HELD(&db->db_mtx));
1234 ASSERT(db->db_state == DB_UNCACHED);
1235 ASSERT(db->db_buf == NULL);
1237 if (db->db_blkid == DMU_BONUS_BLKID) {
1239 * The bonus length stored in the dnode may be less than
1240 * the maximum available space in the bonus buffer.
1242 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1243 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1245 ASSERT3U(bonuslen, <=, db->db.db_size);
1246 db->db.db_data = zio_buf_alloc(max_bonuslen);
1247 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1248 if (bonuslen < max_bonuslen)
1249 bzero(db->db.db_data, max_bonuslen);
1251 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1253 db->db_state = DB_CACHED;
1254 mutex_exit(&db->db_mtx);
1259 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1260 * processes the delete record and clears the bp while we are waiting
1261 * for the dn_mtx (resulting in a "no" from block_freed).
1263 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1264 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1265 BP_IS_HOLE(db->db_blkptr)))) {
1266 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1268 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1270 bzero(db->db.db_data, db->db.db_size);
1272 if (db->db_blkptr != NULL && db->db_level > 0 &&
1273 BP_IS_HOLE(db->db_blkptr) &&
1274 db->db_blkptr->blk_birth != 0) {
1275 blkptr_t *bps = db->db.db_data;
1276 for (int i = 0; i < ((1 <<
1277 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1279 blkptr_t *bp = &bps[i];
1280 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1281 1 << dn->dn_indblkshift);
1283 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1285 BP_GET_LSIZE(db->db_blkptr));
1286 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1288 BP_GET_LEVEL(db->db_blkptr) - 1);
1289 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1293 db->db_state = DB_CACHED;
1294 mutex_exit(&db->db_mtx);
1300 db->db_state = DB_READ;
1301 mutex_exit(&db->db_mtx);
1303 if (DBUF_IS_L2CACHEABLE(db))
1304 aflags |= ARC_FLAG_L2CACHE;
1306 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1307 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1308 db->db.db_object, db->db_level, db->db_blkid);
1310 dbuf_add_ref(db, NULL);
1312 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1313 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1314 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1319 * This is our just-in-time copy function. It makes a copy of buffers that
1320 * have been modified in a previous transaction group before we access them in
1321 * the current active group.
1323 * This function is used in three places: when we are dirtying a buffer for the
1324 * first time in a txg, when we are freeing a range in a dnode that includes
1325 * this buffer, and when we are accessing a buffer which was received compressed
1326 * and later referenced in a WRITE_BYREF record.
1328 * Note that when we are called from dbuf_free_range() we do not put a hold on
1329 * the buffer, we just traverse the active dbuf list for the dnode.
1332 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1334 dbuf_dirty_record_t *dr = db->db_last_dirty;
1336 ASSERT(MUTEX_HELD(&db->db_mtx));
1337 ASSERT(db->db.db_data != NULL);
1338 ASSERT(db->db_level == 0);
1339 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1342 (dr->dt.dl.dr_data !=
1343 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1347 * If the last dirty record for this dbuf has not yet synced
1348 * and its referencing the dbuf data, either:
1349 * reset the reference to point to a new copy,
1350 * or (if there a no active holders)
1351 * just null out the current db_data pointer.
1353 ASSERT(dr->dr_txg >= txg - 2);
1354 if (db->db_blkid == DMU_BONUS_BLKID) {
1355 /* Note that the data bufs here are zio_bufs */
1356 dnode_t *dn = DB_DNODE(db);
1357 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1358 dr->dt.dl.dr_data = zio_buf_alloc(bonuslen);
1359 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1360 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1361 } else if (zfs_refcount_count(&db->db_holds) > db->db_dirtycnt) {
1362 int size = arc_buf_size(db->db_buf);
1363 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1364 spa_t *spa = db->db_objset->os_spa;
1365 enum zio_compress compress_type =
1366 arc_get_compression(db->db_buf);
1368 if (compress_type == ZIO_COMPRESS_OFF) {
1369 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1371 ASSERT3U(type, ==, ARC_BUFC_DATA);
1372 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1373 size, arc_buf_lsize(db->db_buf), compress_type);
1375 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1378 dbuf_clear_data(db);
1383 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1390 * We don't have to hold the mutex to check db_state because it
1391 * can't be freed while we have a hold on the buffer.
1393 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1395 if (db->db_state == DB_NOFILL)
1396 return (SET_ERROR(EIO));
1400 if ((flags & DB_RF_HAVESTRUCT) == 0)
1401 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1403 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1404 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1405 DBUF_IS_CACHEABLE(db);
1407 mutex_enter(&db->db_mtx);
1408 if (db->db_state == DB_CACHED) {
1410 * If the arc buf is compressed, we need to decompress it to
1411 * read the data. This could happen during the "zfs receive" of
1412 * a stream which is compressed and deduplicated.
1414 if (db->db_buf != NULL &&
1415 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1416 dbuf_fix_old_data(db,
1417 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1418 err = arc_decompress(db->db_buf);
1419 dbuf_set_data(db, db->db_buf);
1421 mutex_exit(&db->db_mtx);
1423 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1424 if ((flags & DB_RF_HAVESTRUCT) == 0)
1425 rw_exit(&dn->dn_struct_rwlock);
1427 DBUF_STAT_BUMP(hash_hits);
1428 } else if (db->db_state == DB_UNCACHED) {
1429 spa_t *spa = dn->dn_objset->os_spa;
1430 boolean_t need_wait = B_FALSE;
1433 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1434 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1437 dbuf_read_impl(db, zio, flags);
1439 /* dbuf_read_impl has dropped db_mtx for us */
1442 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1444 if ((flags & DB_RF_HAVESTRUCT) == 0)
1445 rw_exit(&dn->dn_struct_rwlock);
1447 DBUF_STAT_BUMP(hash_misses);
1450 err = zio_wait(zio);
1453 * Another reader came in while the dbuf was in flight
1454 * between UNCACHED and CACHED. Either a writer will finish
1455 * writing the buffer (sending the dbuf to CACHED) or the
1456 * first reader's request will reach the read_done callback
1457 * and send the dbuf to CACHED. Otherwise, a failure
1458 * occurred and the dbuf went to UNCACHED.
1460 mutex_exit(&db->db_mtx);
1462 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1463 if ((flags & DB_RF_HAVESTRUCT) == 0)
1464 rw_exit(&dn->dn_struct_rwlock);
1466 DBUF_STAT_BUMP(hash_misses);
1468 /* Skip the wait per the caller's request. */
1469 mutex_enter(&db->db_mtx);
1470 if ((flags & DB_RF_NEVERWAIT) == 0) {
1471 while (db->db_state == DB_READ ||
1472 db->db_state == DB_FILL) {
1473 ASSERT(db->db_state == DB_READ ||
1474 (flags & DB_RF_HAVESTRUCT) == 0);
1475 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1477 cv_wait(&db->db_changed, &db->db_mtx);
1479 if (db->db_state == DB_UNCACHED)
1480 err = SET_ERROR(EIO);
1482 mutex_exit(&db->db_mtx);
1489 dbuf_noread(dmu_buf_impl_t *db)
1491 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1492 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1493 mutex_enter(&db->db_mtx);
1494 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1495 cv_wait(&db->db_changed, &db->db_mtx);
1496 if (db->db_state == DB_UNCACHED) {
1497 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1498 spa_t *spa = db->db_objset->os_spa;
1500 ASSERT(db->db_buf == NULL);
1501 ASSERT(db->db.db_data == NULL);
1502 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1503 db->db_state = DB_FILL;
1504 } else if (db->db_state == DB_NOFILL) {
1505 dbuf_clear_data(db);
1507 ASSERT3U(db->db_state, ==, DB_CACHED);
1509 mutex_exit(&db->db_mtx);
1513 dbuf_unoverride(dbuf_dirty_record_t *dr)
1515 dmu_buf_impl_t *db = dr->dr_dbuf;
1516 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1517 uint64_t txg = dr->dr_txg;
1519 ASSERT(MUTEX_HELD(&db->db_mtx));
1521 * This assert is valid because dmu_sync() expects to be called by
1522 * a zilog's get_data while holding a range lock. This call only
1523 * comes from dbuf_dirty() callers who must also hold a range lock.
1525 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1526 ASSERT(db->db_level == 0);
1528 if (db->db_blkid == DMU_BONUS_BLKID ||
1529 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1532 ASSERT(db->db_data_pending != dr);
1534 /* free this block */
1535 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1536 zio_free(db->db_objset->os_spa, txg, bp);
1538 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1539 dr->dt.dl.dr_nopwrite = B_FALSE;
1542 * Release the already-written buffer, so we leave it in
1543 * a consistent dirty state. Note that all callers are
1544 * modifying the buffer, so they will immediately do
1545 * another (redundant) arc_release(). Therefore, leave
1546 * the buf thawed to save the effort of freezing &
1547 * immediately re-thawing it.
1549 arc_release(dr->dt.dl.dr_data, db);
1553 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1554 * data blocks in the free range, so that any future readers will find
1558 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1561 dmu_buf_impl_t db_search;
1562 dmu_buf_impl_t *db, *db_next;
1563 uint64_t txg = tx->tx_txg;
1566 if (end_blkid > dn->dn_maxblkid &&
1567 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1568 end_blkid = dn->dn_maxblkid;
1569 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1571 db_search.db_level = 0;
1572 db_search.db_blkid = start_blkid;
1573 db_search.db_state = DB_SEARCH;
1575 mutex_enter(&dn->dn_dbufs_mtx);
1576 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1577 ASSERT3P(db, ==, NULL);
1579 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1581 for (; db != NULL; db = db_next) {
1582 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1583 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1585 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1588 ASSERT3U(db->db_blkid, >=, start_blkid);
1590 /* found a level 0 buffer in the range */
1591 mutex_enter(&db->db_mtx);
1592 if (dbuf_undirty(db, tx)) {
1593 /* mutex has been dropped and dbuf destroyed */
1597 if (db->db_state == DB_UNCACHED ||
1598 db->db_state == DB_NOFILL ||
1599 db->db_state == DB_EVICTING) {
1600 ASSERT(db->db.db_data == NULL);
1601 mutex_exit(&db->db_mtx);
1604 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1605 /* will be handled in dbuf_read_done or dbuf_rele */
1606 db->db_freed_in_flight = TRUE;
1607 mutex_exit(&db->db_mtx);
1610 if (zfs_refcount_count(&db->db_holds) == 0) {
1615 /* The dbuf is referenced */
1617 if (db->db_last_dirty != NULL) {
1618 dbuf_dirty_record_t *dr = db->db_last_dirty;
1620 if (dr->dr_txg == txg) {
1622 * This buffer is "in-use", re-adjust the file
1623 * size to reflect that this buffer may
1624 * contain new data when we sync.
1626 if (db->db_blkid != DMU_SPILL_BLKID &&
1627 db->db_blkid > dn->dn_maxblkid)
1628 dn->dn_maxblkid = db->db_blkid;
1629 dbuf_unoverride(dr);
1632 * This dbuf is not dirty in the open context.
1633 * Either uncache it (if its not referenced in
1634 * the open context) or reset its contents to
1637 dbuf_fix_old_data(db, txg);
1640 /* clear the contents if its cached */
1641 if (db->db_state == DB_CACHED) {
1642 ASSERT(db->db.db_data != NULL);
1643 arc_release(db->db_buf, db);
1644 bzero(db->db.db_data, db->db.db_size);
1645 arc_buf_freeze(db->db_buf);
1648 mutex_exit(&db->db_mtx);
1650 mutex_exit(&dn->dn_dbufs_mtx);
1654 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1656 arc_buf_t *buf, *obuf;
1657 int osize = db->db.db_size;
1658 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1661 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1666 /* XXX does *this* func really need the lock? */
1667 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1670 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1671 * is OK, because there can be no other references to the db
1672 * when we are changing its size, so no concurrent DB_FILL can
1676 * XXX we should be doing a dbuf_read, checking the return
1677 * value and returning that up to our callers
1679 dmu_buf_will_dirty(&db->db, tx);
1681 /* create the data buffer for the new block */
1682 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1684 /* copy old block data to the new block */
1686 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1687 /* zero the remainder */
1689 bzero((uint8_t *)buf->b_data + osize, size - osize);
1691 mutex_enter(&db->db_mtx);
1692 dbuf_set_data(db, buf);
1693 arc_buf_destroy(obuf, db);
1694 db->db.db_size = size;
1696 if (db->db_level == 0) {
1697 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1698 db->db_last_dirty->dt.dl.dr_data = buf;
1700 mutex_exit(&db->db_mtx);
1702 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1707 dbuf_release_bp(dmu_buf_impl_t *db)
1709 objset_t *os = db->db_objset;
1711 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1712 ASSERT(arc_released(os->os_phys_buf) ||
1713 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1714 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1716 (void) arc_release(db->db_buf, db);
1720 * We already have a dirty record for this TXG, and we are being
1724 dbuf_redirty(dbuf_dirty_record_t *dr)
1726 dmu_buf_impl_t *db = dr->dr_dbuf;
1728 ASSERT(MUTEX_HELD(&db->db_mtx));
1730 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1732 * If this buffer has already been written out,
1733 * we now need to reset its state.
1735 dbuf_unoverride(dr);
1736 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1737 db->db_state != DB_NOFILL) {
1738 /* Already released on initial dirty, so just thaw. */
1739 ASSERT(arc_released(db->db_buf));
1740 arc_buf_thaw(db->db_buf);
1745 dbuf_dirty_record_t *
1746 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1750 dbuf_dirty_record_t **drp, *dr;
1751 int drop_struct_lock = FALSE;
1752 int txgoff = tx->tx_txg & TXG_MASK;
1754 ASSERT(tx->tx_txg != 0);
1755 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1756 DMU_TX_DIRTY_BUF(tx, db);
1761 * Shouldn't dirty a regular buffer in syncing context. Private
1762 * objects may be dirtied in syncing context, but only if they
1763 * were already pre-dirtied in open context.
1766 if (dn->dn_objset->os_dsl_dataset != NULL) {
1767 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1770 ASSERT(!dmu_tx_is_syncing(tx) ||
1771 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1772 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1773 dn->dn_objset->os_dsl_dataset == NULL);
1774 if (dn->dn_objset->os_dsl_dataset != NULL)
1775 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1778 * We make this assert for private objects as well, but after we
1779 * check if we're already dirty. They are allowed to re-dirty
1780 * in syncing context.
1782 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1783 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1784 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1786 mutex_enter(&db->db_mtx);
1788 * XXX make this true for indirects too? The problem is that
1789 * transactions created with dmu_tx_create_assigned() from
1790 * syncing context don't bother holding ahead.
1792 ASSERT(db->db_level != 0 ||
1793 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1794 db->db_state == DB_NOFILL);
1796 mutex_enter(&dn->dn_mtx);
1798 * Don't set dirtyctx to SYNC if we're just modifying this as we
1799 * initialize the objset.
1801 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1802 if (dn->dn_objset->os_dsl_dataset != NULL) {
1803 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1806 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1807 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1808 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1809 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1810 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1812 if (dn->dn_objset->os_dsl_dataset != NULL) {
1813 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1818 if (tx->tx_txg > dn->dn_dirty_txg)
1819 dn->dn_dirty_txg = tx->tx_txg;
1820 mutex_exit(&dn->dn_mtx);
1822 if (db->db_blkid == DMU_SPILL_BLKID)
1823 dn->dn_have_spill = B_TRUE;
1826 * If this buffer is already dirty, we're done.
1828 drp = &db->db_last_dirty;
1829 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1830 db->db.db_object == DMU_META_DNODE_OBJECT);
1831 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1833 if (dr && dr->dr_txg == tx->tx_txg) {
1837 mutex_exit(&db->db_mtx);
1842 * Only valid if not already dirty.
1844 ASSERT(dn->dn_object == 0 ||
1845 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1846 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1848 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1851 * We should only be dirtying in syncing context if it's the
1852 * mos or we're initializing the os or it's a special object.
1853 * However, we are allowed to dirty in syncing context provided
1854 * we already dirtied it in open context. Hence we must make
1855 * this assertion only if we're not already dirty.
1858 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1860 if (dn->dn_objset->os_dsl_dataset != NULL)
1861 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1862 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1863 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1864 if (dn->dn_objset->os_dsl_dataset != NULL)
1865 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1867 ASSERT(db->db.db_size != 0);
1869 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1871 if (db->db_blkid != DMU_BONUS_BLKID) {
1872 dmu_objset_willuse_space(os, db->db.db_size, tx);
1876 * If this buffer is dirty in an old transaction group we need
1877 * to make a copy of it so that the changes we make in this
1878 * transaction group won't leak out when we sync the older txg.
1880 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1881 list_link_init(&dr->dr_dirty_node);
1882 if (db->db_level == 0) {
1883 void *data_old = db->db_buf;
1885 if (db->db_state != DB_NOFILL) {
1886 if (db->db_blkid == DMU_BONUS_BLKID) {
1887 dbuf_fix_old_data(db, tx->tx_txg);
1888 data_old = db->db.db_data;
1889 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1891 * Release the data buffer from the cache so
1892 * that we can modify it without impacting
1893 * possible other users of this cached data
1894 * block. Note that indirect blocks and
1895 * private objects are not released until the
1896 * syncing state (since they are only modified
1899 arc_release(db->db_buf, db);
1900 dbuf_fix_old_data(db, tx->tx_txg);
1901 data_old = db->db_buf;
1903 ASSERT(data_old != NULL);
1905 dr->dt.dl.dr_data = data_old;
1907 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1908 list_create(&dr->dt.di.dr_children,
1909 sizeof (dbuf_dirty_record_t),
1910 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1912 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1913 dr->dr_accounted = db->db.db_size;
1915 dr->dr_txg = tx->tx_txg;
1920 * We could have been freed_in_flight between the dbuf_noread
1921 * and dbuf_dirty. We win, as though the dbuf_noread() had
1922 * happened after the free.
1924 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1925 db->db_blkid != DMU_SPILL_BLKID) {
1926 mutex_enter(&dn->dn_mtx);
1927 if (dn->dn_free_ranges[txgoff] != NULL) {
1928 range_tree_clear(dn->dn_free_ranges[txgoff],
1931 mutex_exit(&dn->dn_mtx);
1932 db->db_freed_in_flight = FALSE;
1936 * This buffer is now part of this txg
1938 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1939 db->db_dirtycnt += 1;
1940 ASSERT3U(db->db_dirtycnt, <=, 3);
1942 mutex_exit(&db->db_mtx);
1944 if (db->db_blkid == DMU_BONUS_BLKID ||
1945 db->db_blkid == DMU_SPILL_BLKID) {
1946 mutex_enter(&dn->dn_mtx);
1947 ASSERT(!list_link_active(&dr->dr_dirty_node));
1948 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1949 mutex_exit(&dn->dn_mtx);
1950 dnode_setdirty(dn, tx);
1956 * The dn_struct_rwlock prevents db_blkptr from changing
1957 * due to a write from syncing context completing
1958 * while we are running, so we want to acquire it before
1959 * looking at db_blkptr.
1961 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1962 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1963 drop_struct_lock = TRUE;
1967 * We need to hold the dn_struct_rwlock to make this assertion,
1968 * because it protects dn_phys / dn_next_nlevels from changing.
1970 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1971 dn->dn_phys->dn_nlevels > db->db_level ||
1972 dn->dn_next_nlevels[txgoff] > db->db_level ||
1973 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1974 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1977 * If we are overwriting a dedup BP, then unless it is snapshotted,
1978 * when we get to syncing context we will need to decrement its
1979 * refcount in the DDT. Prefetch the relevant DDT block so that
1980 * syncing context won't have to wait for the i/o.
1982 ddt_prefetch(os->os_spa, db->db_blkptr);
1984 if (db->db_level == 0) {
1985 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1986 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1989 if (db->db_level+1 < dn->dn_nlevels) {
1990 dmu_buf_impl_t *parent = db->db_parent;
1991 dbuf_dirty_record_t *di;
1992 int parent_held = FALSE;
1994 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1995 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1997 parent = dbuf_hold_level(dn, db->db_level+1,
1998 db->db_blkid >> epbs, FTAG);
1999 ASSERT(parent != NULL);
2002 if (drop_struct_lock)
2003 rw_exit(&dn->dn_struct_rwlock);
2004 ASSERT3U(db->db_level+1, ==, parent->db_level);
2005 di = dbuf_dirty(parent, tx);
2007 dbuf_rele(parent, FTAG);
2009 mutex_enter(&db->db_mtx);
2011 * Since we've dropped the mutex, it's possible that
2012 * dbuf_undirty() might have changed this out from under us.
2014 if (db->db_last_dirty == dr ||
2015 dn->dn_object == DMU_META_DNODE_OBJECT) {
2016 mutex_enter(&di->dt.di.dr_mtx);
2017 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
2018 ASSERT(!list_link_active(&dr->dr_dirty_node));
2019 list_insert_tail(&di->dt.di.dr_children, dr);
2020 mutex_exit(&di->dt.di.dr_mtx);
2023 mutex_exit(&db->db_mtx);
2025 ASSERT(db->db_level+1 == dn->dn_nlevels);
2026 ASSERT(db->db_blkid < dn->dn_nblkptr);
2027 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
2028 mutex_enter(&dn->dn_mtx);
2029 ASSERT(!list_link_active(&dr->dr_dirty_node));
2030 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2031 mutex_exit(&dn->dn_mtx);
2032 if (drop_struct_lock)
2033 rw_exit(&dn->dn_struct_rwlock);
2036 dnode_setdirty(dn, tx);
2042 * Undirty a buffer in the transaction group referenced by the given
2043 * transaction. Return whether this evicted the dbuf.
2046 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2049 uint64_t txg = tx->tx_txg;
2050 dbuf_dirty_record_t *dr, **drp;
2055 * Due to our use of dn_nlevels below, this can only be called
2056 * in open context, unless we are operating on the MOS.
2057 * From syncing context, dn_nlevels may be different from the
2058 * dn_nlevels used when dbuf was dirtied.
2060 ASSERT(db->db_objset ==
2061 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2062 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2063 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2064 ASSERT0(db->db_level);
2065 ASSERT(MUTEX_HELD(&db->db_mtx));
2068 * If this buffer is not dirty, we're done.
2070 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
2071 if (dr->dr_txg <= txg)
2073 if (dr == NULL || dr->dr_txg < txg)
2075 ASSERT(dr->dr_txg == txg);
2076 ASSERT(dr->dr_dbuf == db);
2081 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2083 ASSERT(db->db.db_size != 0);
2085 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2086 dr->dr_accounted, txg);
2091 * Note that there are three places in dbuf_dirty()
2092 * where this dirty record may be put on a list.
2093 * Make sure to do a list_remove corresponding to
2094 * every one of those list_insert calls.
2096 if (dr->dr_parent) {
2097 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2098 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2099 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2100 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2101 db->db_level + 1 == dn->dn_nlevels) {
2102 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2103 mutex_enter(&dn->dn_mtx);
2104 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2105 mutex_exit(&dn->dn_mtx);
2109 if (db->db_state != DB_NOFILL) {
2110 dbuf_unoverride(dr);
2112 ASSERT(db->db_buf != NULL);
2113 ASSERT(dr->dt.dl.dr_data != NULL);
2114 if (dr->dt.dl.dr_data != db->db_buf)
2115 arc_buf_destroy(dr->dt.dl.dr_data, db);
2118 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2120 ASSERT(db->db_dirtycnt > 0);
2121 db->db_dirtycnt -= 1;
2123 if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2124 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2133 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2135 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2136 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
2138 ASSERT(tx->tx_txg != 0);
2139 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2142 * Quick check for dirtyness. For already dirty blocks, this
2143 * reduces runtime of this function by >90%, and overall performance
2144 * by 50% for some workloads (e.g. file deletion with indirect blocks
2147 mutex_enter(&db->db_mtx);
2148 dbuf_dirty_record_t *dr;
2149 for (dr = db->db_last_dirty;
2150 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
2152 * It's possible that it is already dirty but not cached,
2153 * because there are some calls to dbuf_dirty() that don't
2154 * go through dmu_buf_will_dirty().
2156 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
2157 /* This dbuf is already dirty and cached. */
2159 mutex_exit(&db->db_mtx);
2163 mutex_exit(&db->db_mtx);
2166 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2167 rf |= DB_RF_HAVESTRUCT;
2169 (void) dbuf_read(db, NULL, rf);
2170 (void) dbuf_dirty(db, tx);
2174 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2176 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2178 db->db_state = DB_NOFILL;
2180 dmu_buf_will_fill(db_fake, tx);
2184 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2186 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2188 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2189 ASSERT(tx->tx_txg != 0);
2190 ASSERT(db->db_level == 0);
2191 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2193 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2194 dmu_tx_private_ok(tx));
2197 (void) dbuf_dirty(db, tx);
2200 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2203 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2205 mutex_enter(&db->db_mtx);
2208 if (db->db_state == DB_FILL) {
2209 if (db->db_level == 0 && db->db_freed_in_flight) {
2210 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2211 /* we were freed while filling */
2212 /* XXX dbuf_undirty? */
2213 bzero(db->db.db_data, db->db.db_size);
2214 db->db_freed_in_flight = FALSE;
2216 db->db_state = DB_CACHED;
2217 cv_broadcast(&db->db_changed);
2219 mutex_exit(&db->db_mtx);
2223 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2224 bp_embedded_type_t etype, enum zio_compress comp,
2225 int uncompressed_size, int compressed_size, int byteorder,
2228 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2229 struct dirty_leaf *dl;
2230 dmu_object_type_t type;
2232 if (etype == BP_EMBEDDED_TYPE_DATA) {
2233 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2234 SPA_FEATURE_EMBEDDED_DATA));
2238 type = DB_DNODE(db)->dn_type;
2241 ASSERT0(db->db_level);
2242 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2244 dmu_buf_will_not_fill(dbuf, tx);
2246 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2247 dl = &db->db_last_dirty->dt.dl;
2248 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2249 data, comp, uncompressed_size, compressed_size);
2250 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2251 BP_SET_TYPE(&dl->dr_overridden_by, type);
2252 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2253 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2255 dl->dr_override_state = DR_OVERRIDDEN;
2256 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2260 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2261 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2264 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2266 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2267 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2268 ASSERT(db->db_level == 0);
2269 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2270 ASSERT(buf != NULL);
2271 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2272 ASSERT(tx->tx_txg != 0);
2274 arc_return_buf(buf, db);
2275 ASSERT(arc_released(buf));
2277 mutex_enter(&db->db_mtx);
2279 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2280 cv_wait(&db->db_changed, &db->db_mtx);
2282 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2284 if (db->db_state == DB_CACHED &&
2285 zfs_refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2286 mutex_exit(&db->db_mtx);
2287 (void) dbuf_dirty(db, tx);
2288 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2289 arc_buf_destroy(buf, db);
2290 xuio_stat_wbuf_copied();
2294 xuio_stat_wbuf_nocopy();
2295 if (db->db_state == DB_CACHED) {
2296 dbuf_dirty_record_t *dr = db->db_last_dirty;
2298 ASSERT(db->db_buf != NULL);
2299 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2300 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2301 if (!arc_released(db->db_buf)) {
2302 ASSERT(dr->dt.dl.dr_override_state ==
2304 arc_release(db->db_buf, db);
2306 dr->dt.dl.dr_data = buf;
2307 arc_buf_destroy(db->db_buf, db);
2308 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2309 arc_release(db->db_buf, db);
2310 arc_buf_destroy(db->db_buf, db);
2314 ASSERT(db->db_buf == NULL);
2315 dbuf_set_data(db, buf);
2316 db->db_state = DB_FILL;
2317 mutex_exit(&db->db_mtx);
2318 (void) dbuf_dirty(db, tx);
2319 dmu_buf_fill_done(&db->db, tx);
2323 dbuf_destroy(dmu_buf_impl_t *db)
2326 dmu_buf_impl_t *parent = db->db_parent;
2327 dmu_buf_impl_t *dndb;
2329 ASSERT(MUTEX_HELD(&db->db_mtx));
2330 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2332 if (db->db_buf != NULL) {
2333 arc_buf_destroy(db->db_buf, db);
2337 if (db->db_blkid == DMU_BONUS_BLKID) {
2338 int slots = DB_DNODE(db)->dn_num_slots;
2339 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2340 if (db->db.db_data != NULL) {
2341 zio_buf_free(db->db.db_data, bonuslen);
2342 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2343 db->db_state = DB_UNCACHED;
2347 dbuf_clear_data(db);
2349 if (multilist_link_active(&db->db_cache_link)) {
2350 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
2351 db->db_caching_status == DB_DBUF_METADATA_CACHE);
2353 multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
2354 (void) zfs_refcount_remove_many(
2355 &dbuf_caches[db->db_caching_status].size,
2356 db->db.db_size, db);
2358 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2359 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2361 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2362 DBUF_STAT_BUMPDOWN(cache_count);
2363 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2366 db->db_caching_status = DB_NO_CACHE;
2369 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2370 ASSERT(db->db_data_pending == NULL);
2372 db->db_state = DB_EVICTING;
2373 db->db_blkptr = NULL;
2376 * Now that db_state is DB_EVICTING, nobody else can find this via
2377 * the hash table. We can now drop db_mtx, which allows us to
2378 * acquire the dn_dbufs_mtx.
2380 mutex_exit(&db->db_mtx);
2385 if (db->db_blkid != DMU_BONUS_BLKID) {
2386 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2388 mutex_enter(&dn->dn_dbufs_mtx);
2389 avl_remove(&dn->dn_dbufs, db);
2390 atomic_dec_32(&dn->dn_dbufs_count);
2394 mutex_exit(&dn->dn_dbufs_mtx);
2396 * Decrementing the dbuf count means that the hold corresponding
2397 * to the removed dbuf is no longer discounted in dnode_move(),
2398 * so the dnode cannot be moved until after we release the hold.
2399 * The membar_producer() ensures visibility of the decremented
2400 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2403 mutex_enter(&dn->dn_mtx);
2404 dnode_rele_and_unlock(dn, db, B_TRUE);
2405 db->db_dnode_handle = NULL;
2407 dbuf_hash_remove(db);
2412 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2414 db->db_parent = NULL;
2416 ASSERT(db->db_buf == NULL);
2417 ASSERT(db->db.db_data == NULL);
2418 ASSERT(db->db_hash_next == NULL);
2419 ASSERT(db->db_blkptr == NULL);
2420 ASSERT(db->db_data_pending == NULL);
2421 ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE);
2422 ASSERT(!multilist_link_active(&db->db_cache_link));
2424 kmem_cache_free(dbuf_kmem_cache, db);
2425 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2428 * If this dbuf is referenced from an indirect dbuf,
2429 * decrement the ref count on the indirect dbuf.
2431 if (parent && parent != dndb) {
2432 mutex_enter(&parent->db_mtx);
2433 dbuf_rele_and_unlock(parent, db, B_TRUE);
2438 * Note: While bpp will always be updated if the function returns success,
2439 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2440 * this happens when the dnode is the meta-dnode, or a userused or groupused
2443 __attribute__((always_inline))
2445 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2446 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2451 ASSERT(blkid != DMU_BONUS_BLKID);
2453 if (blkid == DMU_SPILL_BLKID) {
2454 mutex_enter(&dn->dn_mtx);
2455 if (dn->dn_have_spill &&
2456 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2457 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2460 dbuf_add_ref(dn->dn_dbuf, NULL);
2461 *parentp = dn->dn_dbuf;
2462 mutex_exit(&dn->dn_mtx);
2467 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2468 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2470 ASSERT3U(level * epbs, <, 64);
2471 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2473 * This assertion shouldn't trip as long as the max indirect block size
2474 * is less than 1M. The reason for this is that up to that point,
2475 * the number of levels required to address an entire object with blocks
2476 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2477 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2478 * (i.e. we can address the entire object), objects will all use at most
2479 * N-1 levels and the assertion won't overflow. However, once epbs is
2480 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2481 * enough to address an entire object, so objects will have 5 levels,
2482 * but then this assertion will overflow.
2484 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2485 * need to redo this logic to handle overflows.
2487 ASSERT(level >= nlevels ||
2488 ((nlevels - level - 1) * epbs) +
2489 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2490 if (level >= nlevels ||
2491 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2492 ((nlevels - level - 1) * epbs)) ||
2494 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2495 /* the buffer has no parent yet */
2496 return (SET_ERROR(ENOENT));
2497 } else if (level < nlevels-1) {
2498 /* this block is referenced from an indirect block */
2501 err = dbuf_hold_impl(dn, level+1,
2502 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2504 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2505 blkid >> epbs, fail_sparse, FALSE, NULL,
2506 parentp, dh->dh_depth + 1);
2507 err = __dbuf_hold_impl(dh + 1);
2511 err = dbuf_read(*parentp, NULL,
2512 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2514 dbuf_rele(*parentp, NULL);
2518 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2519 (blkid & ((1ULL << epbs) - 1));
2520 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2521 ASSERT(BP_IS_HOLE(*bpp));
2524 /* the block is referenced from the dnode */
2525 ASSERT3U(level, ==, nlevels-1);
2526 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2527 blkid < dn->dn_phys->dn_nblkptr);
2529 dbuf_add_ref(dn->dn_dbuf, NULL);
2530 *parentp = dn->dn_dbuf;
2532 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2537 static dmu_buf_impl_t *
2538 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2539 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2541 objset_t *os = dn->dn_objset;
2542 dmu_buf_impl_t *db, *odb;
2544 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2545 ASSERT(dn->dn_type != DMU_OT_NONE);
2547 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2550 db->db.db_object = dn->dn_object;
2551 db->db_level = level;
2552 db->db_blkid = blkid;
2553 db->db_last_dirty = NULL;
2554 db->db_dirtycnt = 0;
2555 db->db_dnode_handle = dn->dn_handle;
2556 db->db_parent = parent;
2557 db->db_blkptr = blkptr;
2560 db->db_user_immediate_evict = FALSE;
2561 db->db_freed_in_flight = FALSE;
2562 db->db_pending_evict = FALSE;
2564 if (blkid == DMU_BONUS_BLKID) {
2565 ASSERT3P(parent, ==, dn->dn_dbuf);
2566 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2567 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2568 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2569 db->db.db_offset = DMU_BONUS_BLKID;
2570 db->db_state = DB_UNCACHED;
2571 db->db_caching_status = DB_NO_CACHE;
2572 /* the bonus dbuf is not placed in the hash table */
2573 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2575 } else if (blkid == DMU_SPILL_BLKID) {
2576 db->db.db_size = (blkptr != NULL) ?
2577 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2578 db->db.db_offset = 0;
2581 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2582 db->db.db_size = blocksize;
2583 db->db.db_offset = db->db_blkid * blocksize;
2587 * Hold the dn_dbufs_mtx while we get the new dbuf
2588 * in the hash table *and* added to the dbufs list.
2589 * This prevents a possible deadlock with someone
2590 * trying to look up this dbuf before its added to the
2593 mutex_enter(&dn->dn_dbufs_mtx);
2594 db->db_state = DB_EVICTING;
2595 if ((odb = dbuf_hash_insert(db)) != NULL) {
2596 /* someone else inserted it first */
2597 kmem_cache_free(dbuf_kmem_cache, db);
2598 mutex_exit(&dn->dn_dbufs_mtx);
2599 DBUF_STAT_BUMP(hash_insert_race);
2602 avl_add(&dn->dn_dbufs, db);
2604 db->db_state = DB_UNCACHED;
2605 db->db_caching_status = DB_NO_CACHE;
2606 mutex_exit(&dn->dn_dbufs_mtx);
2607 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2609 if (parent && parent != dn->dn_dbuf)
2610 dbuf_add_ref(parent, db);
2612 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2613 zfs_refcount_count(&dn->dn_holds) > 0);
2614 (void) zfs_refcount_add(&dn->dn_holds, db);
2615 atomic_inc_32(&dn->dn_dbufs_count);
2617 dprintf_dbuf(db, "db=%p\n", db);
2622 typedef struct dbuf_prefetch_arg {
2623 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2624 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2625 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2626 int dpa_curlevel; /* The current level that we're reading */
2627 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2628 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2629 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2630 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2631 } dbuf_prefetch_arg_t;
2634 * Actually issue the prefetch read for the block given.
2637 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2639 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2642 arc_flags_t aflags =
2643 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2645 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2646 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2647 ASSERT(dpa->dpa_zio != NULL);
2648 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2649 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2650 &aflags, &dpa->dpa_zb);
2654 * Called when an indirect block above our prefetch target is read in. This
2655 * will either read in the next indirect block down the tree or issue the actual
2656 * prefetch if the next block down is our target.
2659 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
2660 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
2662 dbuf_prefetch_arg_t *dpa = private;
2664 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2665 ASSERT3S(dpa->dpa_curlevel, >, 0);
2668 ASSERT(zio == NULL || zio->io_error != 0);
2669 kmem_free(dpa, sizeof (*dpa));
2672 ASSERT(zio == NULL || zio->io_error == 0);
2675 * The dpa_dnode is only valid if we are called with a NULL
2676 * zio. This indicates that the arc_read() returned without
2677 * first calling zio_read() to issue a physical read. Once
2678 * a physical read is made the dpa_dnode must be invalidated
2679 * as the locks guarding it may have been dropped. If the
2680 * dpa_dnode is still valid, then we want to add it to the dbuf
2681 * cache. To do so, we must hold the dbuf associated with the block
2682 * we just prefetched, read its contents so that we associate it
2683 * with an arc_buf_t, and then release it.
2686 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2687 if (zio->io_flags & ZIO_FLAG_RAW) {
2688 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2690 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2692 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2694 dpa->dpa_dnode = NULL;
2695 } else if (dpa->dpa_dnode != NULL) {
2696 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2697 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2698 dpa->dpa_zb.zb_level));
2699 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2700 dpa->dpa_curlevel, curblkid, FTAG);
2701 (void) dbuf_read(db, NULL,
2702 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2703 dbuf_rele(db, FTAG);
2707 kmem_free(dpa, sizeof(*dpa));
2711 dpa->dpa_curlevel--;
2713 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2714 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2715 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2716 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2717 if (BP_IS_HOLE(bp)) {
2718 kmem_free(dpa, sizeof (*dpa));
2719 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2720 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2721 dbuf_issue_final_prefetch(dpa, bp);
2722 kmem_free(dpa, sizeof (*dpa));
2724 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2725 zbookmark_phys_t zb;
2727 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2728 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2729 iter_aflags |= ARC_FLAG_L2CACHE;
2731 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2733 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2734 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2736 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2737 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2738 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2742 arc_buf_destroy(abuf, private);
2746 * Issue prefetch reads for the given block on the given level. If the indirect
2747 * blocks above that block are not in memory, we will read them in
2748 * asynchronously. As a result, this call never blocks waiting for a read to
2752 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2756 int epbs, nlevels, curlevel;
2759 ASSERT(blkid != DMU_BONUS_BLKID);
2760 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2762 if (blkid > dn->dn_maxblkid)
2765 if (dnode_block_freed(dn, blkid))
2769 * This dnode hasn't been written to disk yet, so there's nothing to
2772 nlevels = dn->dn_phys->dn_nlevels;
2773 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2776 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2777 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2780 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2783 mutex_exit(&db->db_mtx);
2785 * This dbuf already exists. It is either CACHED, or
2786 * (we assume) about to be read or filled.
2792 * Find the closest ancestor (indirect block) of the target block
2793 * that is present in the cache. In this indirect block, we will
2794 * find the bp that is at curlevel, curblkid.
2798 while (curlevel < nlevels - 1) {
2799 int parent_level = curlevel + 1;
2800 uint64_t parent_blkid = curblkid >> epbs;
2803 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2804 FALSE, TRUE, FTAG, &db) == 0) {
2805 blkptr_t *bpp = db->db_buf->b_data;
2806 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2807 dbuf_rele(db, FTAG);
2811 curlevel = parent_level;
2812 curblkid = parent_blkid;
2815 if (curlevel == nlevels - 1) {
2816 /* No cached indirect blocks found. */
2817 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2818 bp = dn->dn_phys->dn_blkptr[curblkid];
2820 if (BP_IS_HOLE(&bp))
2823 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2825 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2828 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2829 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2830 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2831 dn->dn_object, level, blkid);
2832 dpa->dpa_curlevel = curlevel;
2833 dpa->dpa_prio = prio;
2834 dpa->dpa_aflags = aflags;
2835 dpa->dpa_spa = dn->dn_objset->os_spa;
2836 dpa->dpa_dnode = dn;
2837 dpa->dpa_epbs = epbs;
2840 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2841 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2842 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2845 * If we have the indirect just above us, no need to do the asynchronous
2846 * prefetch chain; we'll just run the last step ourselves. If we're at
2847 * a higher level, though, we want to issue the prefetches for all the
2848 * indirect blocks asynchronously, so we can go on with whatever we were
2851 if (curlevel == level) {
2852 ASSERT3U(curblkid, ==, blkid);
2853 dbuf_issue_final_prefetch(dpa, &bp);
2854 kmem_free(dpa, sizeof (*dpa));
2856 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2857 zbookmark_phys_t zb;
2859 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2860 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2861 iter_aflags |= ARC_FLAG_L2CACHE;
2863 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2864 dn->dn_object, curlevel, curblkid);
2865 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2866 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2867 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2871 * We use pio here instead of dpa_zio since it's possible that
2872 * dpa may have already been freed.
2877 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2880 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2881 * the case of encrypted, compressed and uncompressed buffers by
2882 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2883 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2885 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2887 noinline static void
2888 dbuf_hold_copy(struct dbuf_hold_impl_data *dh)
2890 dnode_t *dn = dh->dh_dn;
2891 dmu_buf_impl_t *db = dh->dh_db;
2892 dbuf_dirty_record_t *dr = dh->dh_dr;
2893 arc_buf_t *data = dr->dt.dl.dr_data;
2895 enum zio_compress compress_type = arc_get_compression(data);
2897 if (compress_type != ZIO_COMPRESS_OFF) {
2898 dbuf_set_data(db, arc_alloc_compressed_buf(
2899 dn->dn_objset->os_spa, db, arc_buf_size(data),
2900 arc_buf_lsize(data), compress_type));
2902 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db,
2903 DBUF_GET_BUFC_TYPE(db), db->db.db_size));
2906 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
2910 * Returns with db_holds incremented, and db_mtx not held.
2911 * Note: dn_struct_rwlock must be held.
2914 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2916 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2917 dh->dh_parent = NULL;
2919 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2920 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2921 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2923 *(dh->dh_dbp) = NULL;
2925 /* dbuf_find() returns with db_mtx held */
2926 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2927 dh->dh_level, dh->dh_blkid);
2929 if (dh->dh_db == NULL) {
2932 if (dh->dh_fail_uncached)
2933 return (SET_ERROR(ENOENT));
2935 ASSERT3P(dh->dh_parent, ==, NULL);
2936 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2937 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2938 if (dh->dh_fail_sparse) {
2939 if (dh->dh_err == 0 &&
2940 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2941 dh->dh_err = SET_ERROR(ENOENT);
2944 dbuf_rele(dh->dh_parent, NULL);
2945 return (dh->dh_err);
2948 if (dh->dh_err && dh->dh_err != ENOENT)
2949 return (dh->dh_err);
2950 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2951 dh->dh_parent, dh->dh_bp);
2954 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2955 mutex_exit(&dh->dh_db->db_mtx);
2956 return (SET_ERROR(ENOENT));
2959 if (dh->dh_db->db_buf != NULL) {
2960 arc_buf_access(dh->dh_db->db_buf);
2961 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2964 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2967 * If this buffer is currently syncing out, and we are are
2968 * still referencing it from db_data, we need to make a copy
2969 * of it in case we decide we want to dirty it again in this txg.
2971 if (dh->dh_db->db_level == 0 &&
2972 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2973 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2974 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2975 dh->dh_dr = dh->dh_db->db_data_pending;
2976 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf)
2980 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2981 ASSERT(zfs_refcount_is_zero(&dh->dh_db->db_holds));
2982 ASSERT(dh->dh_db->db_caching_status == DB_DBUF_CACHE ||
2983 dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE);
2986 dbuf_caches[dh->dh_db->db_caching_status].cache,
2988 (void) zfs_refcount_remove_many(
2989 &dbuf_caches[dh->dh_db->db_caching_status].size,
2990 dh->dh_db->db.db_size, dh->dh_db);
2992 if (dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2993 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2995 DBUF_STAT_BUMPDOWN(cache_levels[dh->dh_db->db_level]);
2996 DBUF_STAT_BUMPDOWN(cache_count);
2997 DBUF_STAT_DECR(cache_levels_bytes[dh->dh_db->db_level],
2998 dh->dh_db->db.db_size);
3000 dh->dh_db->db_caching_status = DB_NO_CACHE;
3002 (void) zfs_refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
3003 DBUF_VERIFY(dh->dh_db);
3004 mutex_exit(&dh->dh_db->db_mtx);
3006 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3008 dbuf_rele(dh->dh_parent, NULL);
3010 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
3011 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
3012 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
3013 *(dh->dh_dbp) = dh->dh_db;
3019 * The following code preserves the recursive function dbuf_hold_impl()
3020 * but moves the local variables AND function arguments to the heap to
3021 * minimize the stack frame size. Enough space is initially allocated
3022 * on the stack for 20 levels of recursion.
3025 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3026 boolean_t fail_sparse, boolean_t fail_uncached,
3027 void *tag, dmu_buf_impl_t **dbp)
3029 struct dbuf_hold_impl_data *dh;
3032 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
3033 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
3034 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
3035 fail_uncached, tag, dbp, 0);
3037 error = __dbuf_hold_impl(dh);
3039 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
3040 DBUF_HOLD_IMPL_MAX_DEPTH);
3046 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
3047 dnode_t *dn, uint8_t level, uint64_t blkid,
3048 boolean_t fail_sparse, boolean_t fail_uncached,
3049 void *tag, dmu_buf_impl_t **dbp, int depth)
3052 dh->dh_level = level;
3053 dh->dh_blkid = blkid;
3055 dh->dh_fail_sparse = fail_sparse;
3056 dh->dh_fail_uncached = fail_uncached;
3062 dh->dh_parent = NULL;
3067 dh->dh_depth = depth;
3071 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3073 return (dbuf_hold_level(dn, 0, blkid, tag));
3077 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3080 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3081 return (err ? NULL : db);
3085 dbuf_create_bonus(dnode_t *dn)
3087 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3089 ASSERT(dn->dn_bonus == NULL);
3090 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3094 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3096 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3099 if (db->db_blkid != DMU_SPILL_BLKID)
3100 return (SET_ERROR(ENOTSUP));
3102 blksz = SPA_MINBLOCKSIZE;
3103 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3104 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3108 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3109 dbuf_new_size(db, blksz, tx);
3110 rw_exit(&dn->dn_struct_rwlock);
3117 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3119 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3122 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3124 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3126 int64_t holds = zfs_refcount_add(&db->db_holds, tag);
3127 ASSERT3S(holds, >, 1);
3130 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3132 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3135 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3136 dmu_buf_impl_t *found_db;
3137 boolean_t result = B_FALSE;
3139 if (db->db_blkid == DMU_BONUS_BLKID)
3140 found_db = dbuf_find_bonus(os, obj);
3142 found_db = dbuf_find(os, obj, 0, blkid);
3144 if (found_db != NULL) {
3145 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3146 (void) zfs_refcount_add(&db->db_holds, tag);
3149 mutex_exit(&db->db_mtx);
3155 * If you call dbuf_rele() you had better not be referencing the dnode handle
3156 * unless you have some other direct or indirect hold on the dnode. (An indirect
3157 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3158 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3159 * dnode's parent dbuf evicting its dnode handles.
3162 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3164 mutex_enter(&db->db_mtx);
3165 dbuf_rele_and_unlock(db, tag, B_FALSE);
3169 dmu_buf_rele(dmu_buf_t *db, void *tag)
3171 dbuf_rele((dmu_buf_impl_t *)db, tag);
3175 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3176 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3177 * argument should be set if we are already in the dbuf-evicting code
3178 * path, in which case we don't want to recursively evict. This allows us to
3179 * avoid deeply nested stacks that would have a call flow similar to this:
3181 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3184 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3188 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
3192 ASSERT(MUTEX_HELD(&db->db_mtx));
3196 * Remove the reference to the dbuf before removing its hold on the
3197 * dnode so we can guarantee in dnode_move() that a referenced bonus
3198 * buffer has a corresponding dnode hold.
3200 holds = zfs_refcount_remove(&db->db_holds, tag);
3204 * We can't freeze indirects if there is a possibility that they
3205 * may be modified in the current syncing context.
3207 if (db->db_buf != NULL &&
3208 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3209 arc_buf_freeze(db->db_buf);
3212 if (holds == db->db_dirtycnt &&
3213 db->db_level == 0 && db->db_user_immediate_evict)
3214 dbuf_evict_user(db);
3217 if (db->db_blkid == DMU_BONUS_BLKID) {
3219 boolean_t evict_dbuf = db->db_pending_evict;
3222 * If the dnode moves here, we cannot cross this
3223 * barrier until the move completes.
3228 atomic_dec_32(&dn->dn_dbufs_count);
3231 * Decrementing the dbuf count means that the bonus
3232 * buffer's dnode hold is no longer discounted in
3233 * dnode_move(). The dnode cannot move until after
3234 * the dnode_rele() below.
3239 * Do not reference db after its lock is dropped.
3240 * Another thread may evict it.
3242 mutex_exit(&db->db_mtx);
3245 dnode_evict_bonus(dn);
3248 } else if (db->db_buf == NULL) {
3250 * This is a special case: we never associated this
3251 * dbuf with any data allocated from the ARC.
3253 ASSERT(db->db_state == DB_UNCACHED ||
3254 db->db_state == DB_NOFILL);
3256 } else if (arc_released(db->db_buf)) {
3258 * This dbuf has anonymous data associated with it.
3262 boolean_t do_arc_evict = B_FALSE;
3264 spa_t *spa = dmu_objset_spa(db->db_objset);
3266 if (!DBUF_IS_CACHEABLE(db) &&
3267 db->db_blkptr != NULL &&
3268 !BP_IS_HOLE(db->db_blkptr) &&
3269 !BP_IS_EMBEDDED(db->db_blkptr)) {
3270 do_arc_evict = B_TRUE;
3271 bp = *db->db_blkptr;
3274 if (!DBUF_IS_CACHEABLE(db) ||
3275 db->db_pending_evict) {
3277 } else if (!multilist_link_active(&db->db_cache_link)) {
3278 ASSERT3U(db->db_caching_status, ==,
3281 dbuf_cached_state_t dcs =
3282 dbuf_include_in_metadata_cache(db) ?
3283 DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
3284 db->db_caching_status = dcs;
3286 multilist_insert(dbuf_caches[dcs].cache, db);
3287 (void) zfs_refcount_add_many(
3288 &dbuf_caches[dcs].size, db->db.db_size, db);
3290 if (dcs == DB_DBUF_METADATA_CACHE) {
3291 DBUF_STAT_BUMP(metadata_cache_count);
3293 metadata_cache_size_bytes_max,
3295 &dbuf_caches[dcs].size));
3298 cache_levels[db->db_level]);
3299 DBUF_STAT_BUMP(cache_count);
3301 cache_levels_bytes[db->db_level],
3303 DBUF_STAT_MAX(cache_size_bytes_max,
3305 &dbuf_caches[dcs].size));
3307 mutex_exit(&db->db_mtx);
3309 if (db->db_caching_status == DB_DBUF_CACHE &&
3311 dbuf_evict_notify();
3316 arc_freed(spa, &bp);
3319 mutex_exit(&db->db_mtx);
3324 #pragma weak dmu_buf_refcount = dbuf_refcount
3326 dbuf_refcount(dmu_buf_impl_t *db)
3328 return (zfs_refcount_count(&db->db_holds));
3332 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3333 dmu_buf_user_t *new_user)
3335 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3337 mutex_enter(&db->db_mtx);
3338 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3339 if (db->db_user == old_user)
3340 db->db_user = new_user;
3342 old_user = db->db_user;
3343 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3344 mutex_exit(&db->db_mtx);
3350 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3352 return (dmu_buf_replace_user(db_fake, NULL, user));
3356 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3358 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3360 db->db_user_immediate_evict = TRUE;
3361 return (dmu_buf_set_user(db_fake, user));
3365 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3367 return (dmu_buf_replace_user(db_fake, user, NULL));
3371 dmu_buf_get_user(dmu_buf_t *db_fake)
3373 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3375 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3376 return (db->db_user);
3380 dmu_buf_user_evict_wait()
3382 taskq_wait(dbu_evict_taskq);
3386 dmu_buf_get_blkptr(dmu_buf_t *db)
3388 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3389 return (dbi->db_blkptr);
3393 dmu_buf_get_objset(dmu_buf_t *db)
3395 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3396 return (dbi->db_objset);
3400 dmu_buf_dnode_enter(dmu_buf_t *db)
3402 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3403 DB_DNODE_ENTER(dbi);
3404 return (DB_DNODE(dbi));
3408 dmu_buf_dnode_exit(dmu_buf_t *db)
3410 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3415 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3417 /* ASSERT(dmu_tx_is_syncing(tx) */
3418 ASSERT(MUTEX_HELD(&db->db_mtx));
3420 if (db->db_blkptr != NULL)
3423 if (db->db_blkid == DMU_SPILL_BLKID) {
3424 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3425 BP_ZERO(db->db_blkptr);
3428 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3430 * This buffer was allocated at a time when there was
3431 * no available blkptrs from the dnode, or it was
3432 * inappropriate to hook it in (i.e., nlevels mis-match).
3434 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3435 ASSERT(db->db_parent == NULL);
3436 db->db_parent = dn->dn_dbuf;
3437 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3440 dmu_buf_impl_t *parent = db->db_parent;
3441 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3443 ASSERT(dn->dn_phys->dn_nlevels > 1);
3444 if (parent == NULL) {
3445 mutex_exit(&db->db_mtx);
3446 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3447 parent = dbuf_hold_level(dn, db->db_level + 1,
3448 db->db_blkid >> epbs, db);
3449 rw_exit(&dn->dn_struct_rwlock);
3450 mutex_enter(&db->db_mtx);
3451 db->db_parent = parent;
3453 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3454 (db->db_blkid & ((1ULL << epbs) - 1));
3460 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3461 * is critical the we not allow the compiler to inline this function in to
3462 * dbuf_sync_list() thereby drastically bloating the stack usage.
3464 noinline static void
3465 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3467 dmu_buf_impl_t *db = dr->dr_dbuf;
3471 ASSERT(dmu_tx_is_syncing(tx));
3473 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3475 mutex_enter(&db->db_mtx);
3477 ASSERT(db->db_level > 0);
3480 /* Read the block if it hasn't been read yet. */
3481 if (db->db_buf == NULL) {
3482 mutex_exit(&db->db_mtx);
3483 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3484 mutex_enter(&db->db_mtx);
3486 ASSERT3U(db->db_state, ==, DB_CACHED);
3487 ASSERT(db->db_buf != NULL);
3491 /* Indirect block size must match what the dnode thinks it is. */
3492 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3493 dbuf_check_blkptr(dn, db);
3496 /* Provide the pending dirty record to child dbufs */
3497 db->db_data_pending = dr;
3499 mutex_exit(&db->db_mtx);
3501 dbuf_write(dr, db->db_buf, tx);
3504 mutex_enter(&dr->dt.di.dr_mtx);
3505 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3506 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3507 mutex_exit(&dr->dt.di.dr_mtx);
3512 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3513 * critical the we not allow the compiler to inline this function in to
3514 * dbuf_sync_list() thereby drastically bloating the stack usage.
3516 noinline static void
3517 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3519 arc_buf_t **datap = &dr->dt.dl.dr_data;
3520 dmu_buf_impl_t *db = dr->dr_dbuf;
3523 uint64_t txg = tx->tx_txg;
3525 ASSERT(dmu_tx_is_syncing(tx));
3527 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3529 mutex_enter(&db->db_mtx);
3531 * To be synced, we must be dirtied. But we
3532 * might have been freed after the dirty.
3534 if (db->db_state == DB_UNCACHED) {
3535 /* This buffer has been freed since it was dirtied */
3536 ASSERT(db->db.db_data == NULL);
3537 } else if (db->db_state == DB_FILL) {
3538 /* This buffer was freed and is now being re-filled */
3539 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3541 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3548 if (db->db_blkid == DMU_SPILL_BLKID) {
3549 mutex_enter(&dn->dn_mtx);
3550 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3552 * In the previous transaction group, the bonus buffer
3553 * was entirely used to store the attributes for the
3554 * dnode which overrode the dn_spill field. However,
3555 * when adding more attributes to the file a spill
3556 * block was required to hold the extra attributes.
3558 * Make sure to clear the garbage left in the dn_spill
3559 * field from the previous attributes in the bonus
3560 * buffer. Otherwise, after writing out the spill
3561 * block to the new allocated dva, it will free
3562 * the old block pointed to by the invalid dn_spill.
3564 db->db_blkptr = NULL;
3566 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3567 mutex_exit(&dn->dn_mtx);
3571 * If this is a bonus buffer, simply copy the bonus data into the
3572 * dnode. It will be written out when the dnode is synced (and it
3573 * will be synced, since it must have been dirty for dbuf_sync to
3576 if (db->db_blkid == DMU_BONUS_BLKID) {
3577 dbuf_dirty_record_t **drp;
3579 ASSERT(*datap != NULL);
3580 ASSERT0(db->db_level);
3581 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3582 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3583 bcopy(*datap, DN_BONUS(dn->dn_phys),
3584 DN_MAX_BONUS_LEN(dn->dn_phys));
3587 if (*datap != db->db.db_data) {
3588 int slots = DB_DNODE(db)->dn_num_slots;
3589 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3590 zio_buf_free(*datap, bonuslen);
3591 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3593 db->db_data_pending = NULL;
3594 drp = &db->db_last_dirty;
3596 drp = &(*drp)->dr_next;
3597 ASSERT(dr->dr_next == NULL);
3598 ASSERT(dr->dr_dbuf == db);
3600 if (dr->dr_dbuf->db_level != 0) {
3601 mutex_destroy(&dr->dt.di.dr_mtx);
3602 list_destroy(&dr->dt.di.dr_children);
3604 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3605 ASSERT(db->db_dirtycnt > 0);
3606 db->db_dirtycnt -= 1;
3607 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
3614 * This function may have dropped the db_mtx lock allowing a dmu_sync
3615 * operation to sneak in. As a result, we need to ensure that we
3616 * don't check the dr_override_state until we have returned from
3617 * dbuf_check_blkptr.
3619 dbuf_check_blkptr(dn, db);
3622 * If this buffer is in the middle of an immediate write,
3623 * wait for the synchronous IO to complete.
3625 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3626 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3627 cv_wait(&db->db_changed, &db->db_mtx);
3628 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3631 if (db->db_state != DB_NOFILL &&
3632 dn->dn_object != DMU_META_DNODE_OBJECT &&
3633 zfs_refcount_count(&db->db_holds) > 1 &&
3634 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3635 *datap == db->db_buf) {
3637 * If this buffer is currently "in use" (i.e., there
3638 * are active holds and db_data still references it),
3639 * then make a copy before we start the write so that
3640 * any modifications from the open txg will not leak
3643 * NOTE: this copy does not need to be made for
3644 * objects only modified in the syncing context (e.g.
3645 * DNONE_DNODE blocks).
3647 int psize = arc_buf_size(*datap);
3648 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3649 enum zio_compress compress_type = arc_get_compression(*datap);
3651 if (compress_type == ZIO_COMPRESS_OFF) {
3652 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3654 ASSERT3U(type, ==, ARC_BUFC_DATA);
3655 int lsize = arc_buf_lsize(*datap);
3656 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3657 psize, lsize, compress_type);
3659 bcopy(db->db.db_data, (*datap)->b_data, psize);
3661 db->db_data_pending = dr;
3663 mutex_exit(&db->db_mtx);
3665 dbuf_write(dr, *datap, tx);
3667 ASSERT(!list_link_active(&dr->dr_dirty_node));
3668 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3669 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3673 * Although zio_nowait() does not "wait for an IO", it does
3674 * initiate the IO. If this is an empty write it seems plausible
3675 * that the IO could actually be completed before the nowait
3676 * returns. We need to DB_DNODE_EXIT() first in case
3677 * zio_nowait() invalidates the dbuf.
3680 zio_nowait(dr->dr_zio);
3685 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3687 dbuf_dirty_record_t *dr;
3689 while (dr = list_head(list)) {
3690 if (dr->dr_zio != NULL) {
3692 * If we find an already initialized zio then we
3693 * are processing the meta-dnode, and we have finished.
3694 * The dbufs for all dnodes are put back on the list
3695 * during processing, so that we can zio_wait()
3696 * these IOs after initiating all child IOs.
3698 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3699 DMU_META_DNODE_OBJECT);
3702 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3703 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3704 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3706 list_remove(list, dr);
3707 if (dr->dr_dbuf->db_level > 0)
3708 dbuf_sync_indirect(dr, tx);
3710 dbuf_sync_leaf(dr, tx);
3716 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3718 dmu_buf_impl_t *db = vdb;
3720 blkptr_t *bp = zio->io_bp;
3721 blkptr_t *bp_orig = &zio->io_bp_orig;
3722 spa_t *spa = zio->io_spa;
3727 ASSERT3P(db->db_blkptr, !=, NULL);
3728 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3732 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3733 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3734 zio->io_prev_space_delta = delta;
3736 if (bp->blk_birth != 0) {
3737 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3738 BP_GET_TYPE(bp) == dn->dn_type) ||
3739 (db->db_blkid == DMU_SPILL_BLKID &&
3740 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3741 BP_IS_EMBEDDED(bp));
3742 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3745 mutex_enter(&db->db_mtx);
3748 if (db->db_blkid == DMU_SPILL_BLKID) {
3749 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3750 ASSERT(!(BP_IS_HOLE(bp)) &&
3751 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3755 if (db->db_level == 0) {
3756 mutex_enter(&dn->dn_mtx);
3757 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3758 db->db_blkid != DMU_SPILL_BLKID)
3759 dn->dn_phys->dn_maxblkid = db->db_blkid;
3760 mutex_exit(&dn->dn_mtx);
3762 if (dn->dn_type == DMU_OT_DNODE) {
3764 while (i < db->db.db_size) {
3766 (void *)(((char *)db->db.db_data) + i);
3768 i += DNODE_MIN_SIZE;
3769 if (dnp->dn_type != DMU_OT_NONE) {
3771 i += dnp->dn_extra_slots *
3776 if (BP_IS_HOLE(bp)) {
3783 blkptr_t *ibp = db->db.db_data;
3784 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3785 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3786 if (BP_IS_HOLE(ibp))
3788 fill += BP_GET_FILL(ibp);
3793 if (!BP_IS_EMBEDDED(bp))
3794 bp->blk_fill = fill;
3796 mutex_exit(&db->db_mtx);
3798 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3799 *db->db_blkptr = *bp;
3800 rw_exit(&dn->dn_struct_rwlock);
3805 * This function gets called just prior to running through the compression
3806 * stage of the zio pipeline. If we're an indirect block comprised of only
3807 * holes, then we want this indirect to be compressed away to a hole. In
3808 * order to do that we must zero out any information about the holes that
3809 * this indirect points to prior to before we try to compress it.
3812 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3814 dmu_buf_impl_t *db = vdb;
3817 unsigned int epbs, i;
3819 ASSERT3U(db->db_level, >, 0);
3822 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3823 ASSERT3U(epbs, <, 31);
3825 /* Determine if all our children are holes */
3826 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3827 if (!BP_IS_HOLE(bp))
3832 * If all the children are holes, then zero them all out so that
3833 * we may get compressed away.
3835 if (i == 1 << epbs) {
3837 * We only found holes. Grab the rwlock to prevent
3838 * anybody from reading the blocks we're about to
3841 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3842 bzero(db->db.db_data, db->db.db_size);
3843 rw_exit(&dn->dn_struct_rwlock);
3849 * The SPA will call this callback several times for each zio - once
3850 * for every physical child i/o (zio->io_phys_children times). This
3851 * allows the DMU to monitor the progress of each logical i/o. For example,
3852 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3853 * block. There may be a long delay before all copies/fragments are completed,
3854 * so this callback allows us to retire dirty space gradually, as the physical
3859 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3861 dmu_buf_impl_t *db = arg;
3862 objset_t *os = db->db_objset;
3863 dsl_pool_t *dp = dmu_objset_pool(os);
3864 dbuf_dirty_record_t *dr;
3867 dr = db->db_data_pending;
3868 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3871 * The callback will be called io_phys_children times. Retire one
3872 * portion of our dirty space each time we are called. Any rounding
3873 * error will be cleaned up by dsl_pool_sync()'s call to
3874 * dsl_pool_undirty_space().
3876 delta = dr->dr_accounted / zio->io_phys_children;
3877 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3882 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3884 dmu_buf_impl_t *db = vdb;
3885 blkptr_t *bp_orig = &zio->io_bp_orig;
3886 blkptr_t *bp = db->db_blkptr;
3887 objset_t *os = db->db_objset;
3888 dmu_tx_t *tx = os->os_synctx;
3889 dbuf_dirty_record_t **drp, *dr;
3891 ASSERT0(zio->io_error);
3892 ASSERT(db->db_blkptr == bp);
3895 * For nopwrites and rewrites we ensure that the bp matches our
3896 * original and bypass all the accounting.
3898 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3899 ASSERT(BP_EQUAL(bp, bp_orig));
3901 dsl_dataset_t *ds = os->os_dsl_dataset;
3902 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3903 dsl_dataset_block_born(ds, bp, tx);
3906 mutex_enter(&db->db_mtx);
3910 drp = &db->db_last_dirty;
3911 while ((dr = *drp) != db->db_data_pending)
3913 ASSERT(!list_link_active(&dr->dr_dirty_node));
3914 ASSERT(dr->dr_dbuf == db);
3915 ASSERT(dr->dr_next == NULL);
3919 if (db->db_blkid == DMU_SPILL_BLKID) {
3924 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3925 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3926 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3931 if (db->db_level == 0) {
3932 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3933 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3934 if (db->db_state != DB_NOFILL) {
3935 if (dr->dt.dl.dr_data != db->db_buf)
3936 arc_buf_destroy(dr->dt.dl.dr_data, db);
3943 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3944 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3945 if (!BP_IS_HOLE(db->db_blkptr)) {
3947 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3948 ASSERT3U(db->db_blkid, <=,
3949 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3950 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3954 mutex_destroy(&dr->dt.di.dr_mtx);
3955 list_destroy(&dr->dt.di.dr_children);
3957 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3959 cv_broadcast(&db->db_changed);
3960 ASSERT(db->db_dirtycnt > 0);
3961 db->db_dirtycnt -= 1;
3962 db->db_data_pending = NULL;
3963 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
3967 dbuf_write_nofill_ready(zio_t *zio)
3969 dbuf_write_ready(zio, NULL, zio->io_private);
3973 dbuf_write_nofill_done(zio_t *zio)
3975 dbuf_write_done(zio, NULL, zio->io_private);
3979 dbuf_write_override_ready(zio_t *zio)
3981 dbuf_dirty_record_t *dr = zio->io_private;
3982 dmu_buf_impl_t *db = dr->dr_dbuf;
3984 dbuf_write_ready(zio, NULL, db);
3988 dbuf_write_override_done(zio_t *zio)
3990 dbuf_dirty_record_t *dr = zio->io_private;
3991 dmu_buf_impl_t *db = dr->dr_dbuf;
3992 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3994 mutex_enter(&db->db_mtx);
3995 if (!BP_EQUAL(zio->io_bp, obp)) {
3996 if (!BP_IS_HOLE(obp))
3997 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3998 arc_release(dr->dt.dl.dr_data, db);
4000 mutex_exit(&db->db_mtx);
4001 dbuf_write_done(zio, NULL, db);
4003 if (zio->io_abd != NULL)
4004 abd_put(zio->io_abd);
4007 typedef struct dbuf_remap_impl_callback_arg {
4009 uint64_t drica_blk_birth;
4011 } dbuf_remap_impl_callback_arg_t;
4014 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
4017 dbuf_remap_impl_callback_arg_t *drica = arg;
4018 objset_t *os = drica->drica_os;
4019 spa_t *spa = dmu_objset_spa(os);
4020 dmu_tx_t *tx = drica->drica_tx;
4022 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4024 if (os == spa_meta_objset(spa)) {
4025 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
4027 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
4028 size, drica->drica_blk_birth, tx);
4033 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
4035 blkptr_t bp_copy = *bp;
4036 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4037 dbuf_remap_impl_callback_arg_t drica;
4039 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4041 drica.drica_os = dn->dn_objset;
4042 drica.drica_blk_birth = bp->blk_birth;
4043 drica.drica_tx = tx;
4044 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
4047 * The struct_rwlock prevents dbuf_read_impl() from
4048 * dereferencing the BP while we are changing it. To
4049 * avoid lock contention, only grab it when we are actually
4052 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
4054 rw_exit(&dn->dn_struct_rwlock);
4059 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
4060 * to remap a copy of every bp in the dbuf.
4063 dbuf_can_remap(const dmu_buf_impl_t *db)
4065 spa_t *spa = dmu_objset_spa(db->db_objset);
4066 blkptr_t *bp = db->db.db_data;
4067 boolean_t ret = B_FALSE;
4069 ASSERT3U(db->db_level, >, 0);
4070 ASSERT3S(db->db_state, ==, DB_CACHED);
4072 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4074 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4075 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4076 blkptr_t bp_copy = bp[i];
4077 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4082 spa_config_exit(spa, SCL_VDEV, FTAG);
4088 dnode_needs_remap(const dnode_t *dn)
4090 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4091 boolean_t ret = B_FALSE;
4093 if (dn->dn_phys->dn_nlevels == 0) {
4097 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4099 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4100 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
4101 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
4102 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4107 spa_config_exit(spa, SCL_VDEV, FTAG);
4113 * Remap any existing BP's to concrete vdevs, if possible.
4116 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
4118 spa_t *spa = dmu_objset_spa(db->db_objset);
4119 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4121 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
4124 if (db->db_level > 0) {
4125 blkptr_t *bp = db->db.db_data;
4126 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4127 dbuf_remap_impl(dn, &bp[i], tx);
4129 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
4130 dnode_phys_t *dnp = db->db.db_data;
4131 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
4133 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT;
4134 i += dnp[i].dn_extra_slots + 1) {
4135 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
4136 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
4143 /* Issue I/O to commit a dirty buffer to disk. */
4145 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4147 dmu_buf_impl_t *db = dr->dr_dbuf;
4150 dmu_buf_impl_t *parent = db->db_parent;
4151 uint64_t txg = tx->tx_txg;
4152 zbookmark_phys_t zb;
4157 ASSERT(dmu_tx_is_syncing(tx));
4163 if (db->db_state != DB_NOFILL) {
4164 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4166 * Private object buffers are released here rather
4167 * than in dbuf_dirty() since they are only modified
4168 * in the syncing context and we don't want the
4169 * overhead of making multiple copies of the data.
4171 if (BP_IS_HOLE(db->db_blkptr)) {
4174 dbuf_release_bp(db);
4176 dbuf_remap(dn, db, tx);
4180 if (parent != dn->dn_dbuf) {
4181 /* Our parent is an indirect block. */
4182 /* We have a dirty parent that has been scheduled for write. */
4183 ASSERT(parent && parent->db_data_pending);
4184 /* Our parent's buffer is one level closer to the dnode. */
4185 ASSERT(db->db_level == parent->db_level-1);
4187 * We're about to modify our parent's db_data by modifying
4188 * our block pointer, so the parent must be released.
4190 ASSERT(arc_released(parent->db_buf));
4191 zio = parent->db_data_pending->dr_zio;
4193 /* Our parent is the dnode itself. */
4194 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4195 db->db_blkid != DMU_SPILL_BLKID) ||
4196 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4197 if (db->db_blkid != DMU_SPILL_BLKID)
4198 ASSERT3P(db->db_blkptr, ==,
4199 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4203 ASSERT(db->db_level == 0 || data == db->db_buf);
4204 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4207 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4208 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4209 db->db.db_object, db->db_level, db->db_blkid);
4211 if (db->db_blkid == DMU_SPILL_BLKID)
4213 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
4215 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
4219 * We copy the blkptr now (rather than when we instantiate the dirty
4220 * record), because its value can change between open context and
4221 * syncing context. We do not need to hold dn_struct_rwlock to read
4222 * db_blkptr because we are in syncing context.
4224 dr->dr_bp_copy = *db->db_blkptr;
4226 if (db->db_level == 0 &&
4227 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
4229 * The BP for this block has been provided by open context
4230 * (by dmu_sync() or dmu_buf_write_embedded()).
4232 abd_t *contents = (data != NULL) ?
4233 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
4235 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
4236 contents, db->db.db_size, db->db.db_size, &zp,
4237 dbuf_write_override_ready, NULL, NULL,
4238 dbuf_write_override_done,
4239 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
4240 mutex_enter(&db->db_mtx);
4241 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
4242 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
4243 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
4244 mutex_exit(&db->db_mtx);
4245 } else if (db->db_state == DB_NOFILL) {
4246 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
4247 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
4248 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4249 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
4250 dbuf_write_nofill_ready, NULL, NULL,
4251 dbuf_write_nofill_done, db,
4252 ZIO_PRIORITY_ASYNC_WRITE,
4253 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
4255 ASSERT(arc_released(data));
4258 * For indirect blocks, we want to setup the children
4259 * ready callback so that we can properly handle an indirect
4260 * block that only contains holes.
4262 arc_write_done_func_t *children_ready_cb = NULL;
4263 if (db->db_level != 0)
4264 children_ready_cb = dbuf_write_children_ready;
4266 dr->dr_zio = arc_write(zio, os->os_spa, txg,
4267 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
4268 &zp, dbuf_write_ready, children_ready_cb,
4269 dbuf_write_physdone, dbuf_write_done, db,
4270 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);