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_lowater(void)
676 return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
677 dbuf_cache_lowater_bytes());
681 * Evict the oldest eligible dbuf from the dbuf cache.
686 int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache);
687 multilist_sublist_t *mls = multilist_sublist_lock(
688 dbuf_caches[DB_DBUF_CACHE].cache, idx);
690 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
692 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
693 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
694 db = multilist_sublist_prev(mls, db);
697 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
698 multilist_sublist_t *, mls);
701 multilist_sublist_remove(mls, db);
702 multilist_sublist_unlock(mls);
703 (void) zfs_refcount_remove_many(
704 &dbuf_caches[DB_DBUF_CACHE].size,
706 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
707 DBUF_STAT_BUMPDOWN(cache_count);
708 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
710 ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE);
711 db->db_caching_status = DB_NO_CACHE;
713 DBUF_STAT_BUMP(cache_total_evicts);
715 multilist_sublist_unlock(mls);
720 * The dbuf evict thread is responsible for aging out dbufs from the
721 * cache. Once the cache has reached it's maximum size, dbufs are removed
722 * and destroyed. The eviction thread will continue running until the size
723 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
724 * out of the cache it is destroyed and becomes eligible for arc eviction.
728 dbuf_evict_thread(void *unused __unused)
732 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
734 mutex_enter(&dbuf_evict_lock);
735 while (!dbuf_evict_thread_exit) {
736 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
737 CALLB_CPR_SAFE_BEGIN(&cpr);
738 (void) cv_timedwait_hires(&dbuf_evict_cv,
739 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
740 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
742 if (dbuf_ksp != NULL)
743 dbuf_ksp->ks_update(dbuf_ksp, KSTAT_READ);
746 mutex_exit(&dbuf_evict_lock);
749 * Keep evicting as long as we're above the low water mark
750 * for the cache. We do this without holding the locks to
751 * minimize lock contention.
753 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
757 mutex_enter(&dbuf_evict_lock);
760 dbuf_evict_thread_exit = B_FALSE;
761 cv_broadcast(&dbuf_evict_cv);
762 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
767 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
768 * If the dbuf cache is at its high water mark, then evict a dbuf from the
769 * dbuf cache using the callers context.
772 dbuf_evict_notify(uint64_t size)
775 * We check if we should evict without holding the dbuf_evict_lock,
776 * because it's OK to occasionally make the wrong decision here,
777 * and grabbing the lock results in massive lock contention.
779 if (size > dbuf_cache_max_bytes) {
780 if (size > dbuf_cache_hiwater_bytes())
782 cv_signal(&dbuf_evict_cv);
787 dbuf_kstat_update(kstat_t *ksp, int rw)
789 dbuf_stats_t *ds = ksp->ks_data;
791 if (rw == KSTAT_WRITE) {
792 return (SET_ERROR(EACCES));
794 ds->metadata_cache_size_bytes.value.ui64 =
795 zfs_refcount_count(&dbuf_caches[DB_DBUF_METADATA_CACHE].size);
796 ds->cache_size_bytes.value.ui64 =
797 zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size);
798 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
799 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
800 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
801 ds->hash_elements.value.ui64 = dbuf_hash_count;
810 uint64_t hsize = 1ULL << 16;
811 dbuf_hash_table_t *h = &dbuf_hash_table;
815 * The hash table is big enough to fill all of physical memory
816 * with an average 4K block size. The table will take up
817 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
819 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
823 h->hash_table_mask = hsize - 1;
824 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
825 if (h->hash_table == NULL) {
826 /* XXX - we should really return an error instead of assert */
827 ASSERT(hsize > (1ULL << 10));
832 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
833 sizeof (dmu_buf_impl_t),
834 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
836 for (i = 0; i < DBUF_MUTEXES; i++)
837 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
841 * Setup the parameters for the dbuf caches. We set the sizes of the
842 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
843 * of the size of the ARC, respectively. If the values are set in
844 * /etc/system and they're not greater than the size of the ARC, then
845 * we honor that value.
847 if (dbuf_cache_max_bytes == 0 ||
848 dbuf_cache_max_bytes >= arc_max_bytes()) {
849 dbuf_cache_max_bytes = arc_max_bytes() >> dbuf_cache_shift;
851 if (dbuf_metadata_cache_max_bytes == 0 ||
852 dbuf_metadata_cache_max_bytes >= arc_max_bytes()) {
853 dbuf_metadata_cache_max_bytes =
854 arc_max_bytes() >> dbuf_metadata_cache_shift;
858 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
859 * configuration is not required.
861 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
863 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
864 dbuf_caches[dcs].cache =
865 multilist_create(sizeof (dmu_buf_impl_t),
866 offsetof(dmu_buf_impl_t, db_cache_link),
867 dbuf_cache_multilist_index_func);
868 zfs_refcount_create(&dbuf_caches[dcs].size);
871 dbuf_evict_thread_exit = B_FALSE;
872 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
873 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
874 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
875 NULL, 0, &p0, TS_RUN, minclsyspri);
877 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
878 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
880 if (dbuf_ksp != NULL) {
881 for (i = 0; i < DN_MAX_LEVELS; i++) {
882 snprintf(dbuf_stats.cache_levels[i].name,
883 KSTAT_STRLEN, "cache_level_%d", i);
884 dbuf_stats.cache_levels[i].data_type =
886 snprintf(dbuf_stats.cache_levels_bytes[i].name,
887 KSTAT_STRLEN, "cache_level_%d_bytes", i);
888 dbuf_stats.cache_levels_bytes[i].data_type =
891 dbuf_ksp->ks_data = &dbuf_stats;
892 dbuf_ksp->ks_update = dbuf_kstat_update;
893 kstat_install(dbuf_ksp);
900 dbuf_hash_table_t *h = &dbuf_hash_table;
903 dbuf_stats_destroy();
905 for (i = 0; i < DBUF_MUTEXES; i++)
906 mutex_destroy(&h->hash_mutexes[i]);
907 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
908 kmem_cache_destroy(dbuf_kmem_cache);
909 taskq_destroy(dbu_evict_taskq);
911 mutex_enter(&dbuf_evict_lock);
912 dbuf_evict_thread_exit = B_TRUE;
913 while (dbuf_evict_thread_exit) {
914 cv_signal(&dbuf_evict_cv);
915 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
917 mutex_exit(&dbuf_evict_lock);
919 mutex_destroy(&dbuf_evict_lock);
920 cv_destroy(&dbuf_evict_cv);
922 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
923 zfs_refcount_destroy(&dbuf_caches[dcs].size);
924 multilist_destroy(dbuf_caches[dcs].cache);
927 if (dbuf_ksp != NULL) {
928 kstat_delete(dbuf_ksp);
939 dbuf_verify(dmu_buf_impl_t *db)
942 dbuf_dirty_record_t *dr;
944 ASSERT(MUTEX_HELD(&db->db_mtx));
946 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
949 ASSERT(db->db_objset != NULL);
953 ASSERT(db->db_parent == NULL);
954 ASSERT(db->db_blkptr == NULL);
956 ASSERT3U(db->db.db_object, ==, dn->dn_object);
957 ASSERT3P(db->db_objset, ==, dn->dn_objset);
958 ASSERT3U(db->db_level, <, dn->dn_nlevels);
959 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
960 db->db_blkid == DMU_SPILL_BLKID ||
961 !avl_is_empty(&dn->dn_dbufs));
963 if (db->db_blkid == DMU_BONUS_BLKID) {
965 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
966 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
967 } else if (db->db_blkid == DMU_SPILL_BLKID) {
969 ASSERT0(db->db.db_offset);
971 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
974 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
975 ASSERT(dr->dr_dbuf == db);
977 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
978 ASSERT(dr->dr_dbuf == db);
981 * We can't assert that db_size matches dn_datablksz because it
982 * can be momentarily different when another thread is doing
985 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
986 dr = db->db_data_pending;
988 * It should only be modified in syncing context, so
989 * make sure we only have one copy of the data.
991 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
994 /* verify db->db_blkptr */
996 if (db->db_parent == dn->dn_dbuf) {
997 /* db is pointed to by the dnode */
998 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
999 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
1000 ASSERT(db->db_parent == NULL);
1002 ASSERT(db->db_parent != NULL);
1003 if (db->db_blkid != DMU_SPILL_BLKID)
1004 ASSERT3P(db->db_blkptr, ==,
1005 &dn->dn_phys->dn_blkptr[db->db_blkid]);
1007 /* db is pointed to by an indirect block */
1008 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
1009 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
1010 ASSERT3U(db->db_parent->db.db_object, ==,
1013 * dnode_grow_indblksz() can make this fail if we don't
1014 * have the struct_rwlock. XXX indblksz no longer
1015 * grows. safe to do this now?
1017 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1018 ASSERT3P(db->db_blkptr, ==,
1019 ((blkptr_t *)db->db_parent->db.db_data +
1020 db->db_blkid % epb));
1024 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
1025 (db->db_buf == NULL || db->db_buf->b_data) &&
1026 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
1027 db->db_state != DB_FILL && !dn->dn_free_txg) {
1029 * If the blkptr isn't set but they have nonzero data,
1030 * it had better be dirty, otherwise we'll lose that
1031 * data when we evict this buffer.
1033 * There is an exception to this rule for indirect blocks; in
1034 * this case, if the indirect block is a hole, we fill in a few
1035 * fields on each of the child blocks (importantly, birth time)
1036 * to prevent hole birth times from being lost when you
1037 * partially fill in a hole.
1039 if (db->db_dirtycnt == 0) {
1040 if (db->db_level == 0) {
1041 uint64_t *buf = db->db.db_data;
1044 for (i = 0; i < db->db.db_size >> 3; i++) {
1045 ASSERT(buf[i] == 0);
1048 blkptr_t *bps = db->db.db_data;
1049 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
1052 * We want to verify that all the blkptrs in the
1053 * indirect block are holes, but we may have
1054 * automatically set up a few fields for them.
1055 * We iterate through each blkptr and verify
1056 * they only have those fields set.
1059 i < db->db.db_size / sizeof (blkptr_t);
1061 blkptr_t *bp = &bps[i];
1062 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1065 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
1066 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
1067 DVA_IS_EMPTY(&bp->blk_dva[2]));
1068 ASSERT0(bp->blk_fill);
1069 ASSERT0(bp->blk_pad[0]);
1070 ASSERT0(bp->blk_pad[1]);
1071 ASSERT(!BP_IS_EMBEDDED(bp));
1072 ASSERT(BP_IS_HOLE(bp));
1073 ASSERT0(bp->blk_phys_birth);
1083 dbuf_clear_data(dmu_buf_impl_t *db)
1085 ASSERT(MUTEX_HELD(&db->db_mtx));
1086 dbuf_evict_user(db);
1087 ASSERT3P(db->db_buf, ==, NULL);
1088 db->db.db_data = NULL;
1089 if (db->db_state != DB_NOFILL)
1090 db->db_state = DB_UNCACHED;
1094 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1096 ASSERT(MUTEX_HELD(&db->db_mtx));
1097 ASSERT(buf != NULL);
1100 ASSERT(buf->b_data != NULL);
1101 db->db.db_data = buf->b_data;
1105 * Loan out an arc_buf for read. Return the loaned arc_buf.
1108 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1112 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1113 mutex_enter(&db->db_mtx);
1114 if (arc_released(db->db_buf) || zfs_refcount_count(&db->db_holds) > 1) {
1115 int blksz = db->db.db_size;
1116 spa_t *spa = db->db_objset->os_spa;
1118 mutex_exit(&db->db_mtx);
1119 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1120 bcopy(db->db.db_data, abuf->b_data, blksz);
1123 arc_loan_inuse_buf(abuf, db);
1125 dbuf_clear_data(db);
1126 mutex_exit(&db->db_mtx);
1132 * Calculate which level n block references the data at the level 0 offset
1136 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
1138 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1140 * The level n blkid is equal to the level 0 blkid divided by
1141 * the number of level 0s in a level n block.
1143 * The level 0 blkid is offset >> datablkshift =
1144 * offset / 2^datablkshift.
1146 * The number of level 0s in a level n is the number of block
1147 * pointers in an indirect block, raised to the power of level.
1148 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1149 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1151 * Thus, the level n blkid is: offset /
1152 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
1153 * = offset / 2^(datablkshift + level *
1154 * (indblkshift - SPA_BLKPTRSHIFT))
1155 * = offset >> (datablkshift + level *
1156 * (indblkshift - SPA_BLKPTRSHIFT))
1158 return (offset >> (dn->dn_datablkshift + level *
1159 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
1161 ASSERT3U(offset, <, dn->dn_datablksz);
1167 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1168 arc_buf_t *buf, void *vdb)
1170 dmu_buf_impl_t *db = vdb;
1172 mutex_enter(&db->db_mtx);
1173 ASSERT3U(db->db_state, ==, DB_READ);
1175 * All reads are synchronous, so we must have a hold on the dbuf
1177 ASSERT(zfs_refcount_count(&db->db_holds) > 0);
1178 ASSERT(db->db_buf == NULL);
1179 ASSERT(db->db.db_data == NULL);
1182 ASSERT(zio == NULL || zio->io_error != 0);
1183 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1184 ASSERT3P(db->db_buf, ==, NULL);
1185 db->db_state = DB_UNCACHED;
1186 } else if (db->db_level == 0 && db->db_freed_in_flight) {
1187 /* freed in flight */
1188 ASSERT(zio == NULL || zio->io_error == 0);
1190 buf = arc_alloc_buf(db->db_objset->os_spa,
1191 db, DBUF_GET_BUFC_TYPE(db), db->db.db_size);
1193 arc_release(buf, db);
1194 bzero(buf->b_data, db->db.db_size);
1195 arc_buf_freeze(buf);
1196 db->db_freed_in_flight = FALSE;
1197 dbuf_set_data(db, buf);
1198 db->db_state = DB_CACHED;
1201 ASSERT(zio == NULL || zio->io_error == 0);
1202 dbuf_set_data(db, buf);
1203 db->db_state = DB_CACHED;
1205 cv_broadcast(&db->db_changed);
1206 dbuf_rele_and_unlock(db, NULL, B_FALSE);
1210 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1213 zbookmark_phys_t zb;
1214 arc_flags_t aflags = ARC_FLAG_NOWAIT;
1218 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1219 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1220 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1221 ASSERT(MUTEX_HELD(&db->db_mtx));
1222 ASSERT(db->db_state == DB_UNCACHED);
1223 ASSERT(db->db_buf == NULL);
1225 if (db->db_blkid == DMU_BONUS_BLKID) {
1227 * The bonus length stored in the dnode may be less than
1228 * the maximum available space in the bonus buffer.
1230 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1231 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1233 ASSERT3U(bonuslen, <=, db->db.db_size);
1234 db->db.db_data = zio_buf_alloc(max_bonuslen);
1235 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1236 if (bonuslen < max_bonuslen)
1237 bzero(db->db.db_data, max_bonuslen);
1239 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1241 db->db_state = DB_CACHED;
1242 mutex_exit(&db->db_mtx);
1247 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1248 * processes the delete record and clears the bp while we are waiting
1249 * for the dn_mtx (resulting in a "no" from block_freed).
1251 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1252 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1253 BP_IS_HOLE(db->db_blkptr)))) {
1254 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1256 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1258 bzero(db->db.db_data, db->db.db_size);
1260 if (db->db_blkptr != NULL && db->db_level > 0 &&
1261 BP_IS_HOLE(db->db_blkptr) &&
1262 db->db_blkptr->blk_birth != 0) {
1263 blkptr_t *bps = db->db.db_data;
1264 for (int i = 0; i < ((1 <<
1265 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1267 blkptr_t *bp = &bps[i];
1268 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1269 1 << dn->dn_indblkshift);
1271 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1273 BP_GET_LSIZE(db->db_blkptr));
1274 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1276 BP_GET_LEVEL(db->db_blkptr) - 1);
1277 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1281 db->db_state = DB_CACHED;
1282 mutex_exit(&db->db_mtx);
1288 db->db_state = DB_READ;
1289 mutex_exit(&db->db_mtx);
1291 if (DBUF_IS_L2CACHEABLE(db))
1292 aflags |= ARC_FLAG_L2CACHE;
1294 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1295 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1296 db->db.db_object, db->db_level, db->db_blkid);
1298 dbuf_add_ref(db, NULL);
1300 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1301 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1302 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1307 * This is our just-in-time copy function. It makes a copy of buffers that
1308 * have been modified in a previous transaction group before we access them in
1309 * the current active group.
1311 * This function is used in three places: when we are dirtying a buffer for the
1312 * first time in a txg, when we are freeing a range in a dnode that includes
1313 * this buffer, and when we are accessing a buffer which was received compressed
1314 * and later referenced in a WRITE_BYREF record.
1316 * Note that when we are called from dbuf_free_range() we do not put a hold on
1317 * the buffer, we just traverse the active dbuf list for the dnode.
1320 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1322 dbuf_dirty_record_t *dr = db->db_last_dirty;
1324 ASSERT(MUTEX_HELD(&db->db_mtx));
1325 ASSERT(db->db.db_data != NULL);
1326 ASSERT(db->db_level == 0);
1327 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1330 (dr->dt.dl.dr_data !=
1331 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1335 * If the last dirty record for this dbuf has not yet synced
1336 * and its referencing the dbuf data, either:
1337 * reset the reference to point to a new copy,
1338 * or (if there a no active holders)
1339 * just null out the current db_data pointer.
1341 ASSERT(dr->dr_txg >= txg - 2);
1342 if (db->db_blkid == DMU_BONUS_BLKID) {
1343 /* Note that the data bufs here are zio_bufs */
1344 dnode_t *dn = DB_DNODE(db);
1345 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1346 dr->dt.dl.dr_data = zio_buf_alloc(bonuslen);
1347 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1348 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1349 } else if (zfs_refcount_count(&db->db_holds) > db->db_dirtycnt) {
1350 int size = arc_buf_size(db->db_buf);
1351 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1352 spa_t *spa = db->db_objset->os_spa;
1353 enum zio_compress compress_type =
1354 arc_get_compression(db->db_buf);
1356 if (compress_type == ZIO_COMPRESS_OFF) {
1357 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1359 ASSERT3U(type, ==, ARC_BUFC_DATA);
1360 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1361 size, arc_buf_lsize(db->db_buf), compress_type);
1363 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1366 dbuf_clear_data(db);
1371 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1378 * We don't have to hold the mutex to check db_state because it
1379 * can't be freed while we have a hold on the buffer.
1381 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1383 if (db->db_state == DB_NOFILL)
1384 return (SET_ERROR(EIO));
1388 if ((flags & DB_RF_HAVESTRUCT) == 0)
1389 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1391 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1392 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1393 DBUF_IS_CACHEABLE(db);
1395 mutex_enter(&db->db_mtx);
1396 if (db->db_state == DB_CACHED) {
1398 * If the arc buf is compressed, we need to decompress it to
1399 * read the data. This could happen during the "zfs receive" of
1400 * a stream which is compressed and deduplicated.
1402 if (db->db_buf != NULL &&
1403 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1404 dbuf_fix_old_data(db,
1405 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1406 err = arc_decompress(db->db_buf);
1407 dbuf_set_data(db, db->db_buf);
1409 mutex_exit(&db->db_mtx);
1411 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1412 if ((flags & DB_RF_HAVESTRUCT) == 0)
1413 rw_exit(&dn->dn_struct_rwlock);
1415 DBUF_STAT_BUMP(hash_hits);
1416 } else if (db->db_state == DB_UNCACHED) {
1417 spa_t *spa = dn->dn_objset->os_spa;
1418 boolean_t need_wait = B_FALSE;
1421 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1422 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1425 dbuf_read_impl(db, zio, flags);
1427 /* dbuf_read_impl has dropped db_mtx for us */
1430 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1432 if ((flags & DB_RF_HAVESTRUCT) == 0)
1433 rw_exit(&dn->dn_struct_rwlock);
1435 DBUF_STAT_BUMP(hash_misses);
1438 err = zio_wait(zio);
1441 * Another reader came in while the dbuf was in flight
1442 * between UNCACHED and CACHED. Either a writer will finish
1443 * writing the buffer (sending the dbuf to CACHED) or the
1444 * first reader's request will reach the read_done callback
1445 * and send the dbuf to CACHED. Otherwise, a failure
1446 * occurred and the dbuf went to UNCACHED.
1448 mutex_exit(&db->db_mtx);
1450 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1451 if ((flags & DB_RF_HAVESTRUCT) == 0)
1452 rw_exit(&dn->dn_struct_rwlock);
1454 DBUF_STAT_BUMP(hash_misses);
1456 /* Skip the wait per the caller's request. */
1457 mutex_enter(&db->db_mtx);
1458 if ((flags & DB_RF_NEVERWAIT) == 0) {
1459 while (db->db_state == DB_READ ||
1460 db->db_state == DB_FILL) {
1461 ASSERT(db->db_state == DB_READ ||
1462 (flags & DB_RF_HAVESTRUCT) == 0);
1463 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1465 cv_wait(&db->db_changed, &db->db_mtx);
1467 if (db->db_state == DB_UNCACHED)
1468 err = SET_ERROR(EIO);
1470 mutex_exit(&db->db_mtx);
1477 dbuf_noread(dmu_buf_impl_t *db)
1479 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1480 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1481 mutex_enter(&db->db_mtx);
1482 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1483 cv_wait(&db->db_changed, &db->db_mtx);
1484 if (db->db_state == DB_UNCACHED) {
1485 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1486 spa_t *spa = db->db_objset->os_spa;
1488 ASSERT(db->db_buf == NULL);
1489 ASSERT(db->db.db_data == NULL);
1490 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1491 db->db_state = DB_FILL;
1492 } else if (db->db_state == DB_NOFILL) {
1493 dbuf_clear_data(db);
1495 ASSERT3U(db->db_state, ==, DB_CACHED);
1497 mutex_exit(&db->db_mtx);
1501 dbuf_unoverride(dbuf_dirty_record_t *dr)
1503 dmu_buf_impl_t *db = dr->dr_dbuf;
1504 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1505 uint64_t txg = dr->dr_txg;
1507 ASSERT(MUTEX_HELD(&db->db_mtx));
1509 * This assert is valid because dmu_sync() expects to be called by
1510 * a zilog's get_data while holding a range lock. This call only
1511 * comes from dbuf_dirty() callers who must also hold a range lock.
1513 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1514 ASSERT(db->db_level == 0);
1516 if (db->db_blkid == DMU_BONUS_BLKID ||
1517 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1520 ASSERT(db->db_data_pending != dr);
1522 /* free this block */
1523 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1524 zio_free(db->db_objset->os_spa, txg, bp);
1526 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1527 dr->dt.dl.dr_nopwrite = B_FALSE;
1530 * Release the already-written buffer, so we leave it in
1531 * a consistent dirty state. Note that all callers are
1532 * modifying the buffer, so they will immediately do
1533 * another (redundant) arc_release(). Therefore, leave
1534 * the buf thawed to save the effort of freezing &
1535 * immediately re-thawing it.
1537 arc_release(dr->dt.dl.dr_data, db);
1541 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1542 * data blocks in the free range, so that any future readers will find
1546 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1549 dmu_buf_impl_t db_search;
1550 dmu_buf_impl_t *db, *db_next;
1551 uint64_t txg = tx->tx_txg;
1554 if (end_blkid > dn->dn_maxblkid &&
1555 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1556 end_blkid = dn->dn_maxblkid;
1557 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1559 db_search.db_level = 0;
1560 db_search.db_blkid = start_blkid;
1561 db_search.db_state = DB_SEARCH;
1563 mutex_enter(&dn->dn_dbufs_mtx);
1564 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1565 ASSERT3P(db, ==, NULL);
1567 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1569 for (; db != NULL; db = db_next) {
1570 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1571 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1573 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1576 ASSERT3U(db->db_blkid, >=, start_blkid);
1578 /* found a level 0 buffer in the range */
1579 mutex_enter(&db->db_mtx);
1580 if (dbuf_undirty(db, tx)) {
1581 /* mutex has been dropped and dbuf destroyed */
1585 if (db->db_state == DB_UNCACHED ||
1586 db->db_state == DB_NOFILL ||
1587 db->db_state == DB_EVICTING) {
1588 ASSERT(db->db.db_data == NULL);
1589 mutex_exit(&db->db_mtx);
1592 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1593 /* will be handled in dbuf_read_done or dbuf_rele */
1594 db->db_freed_in_flight = TRUE;
1595 mutex_exit(&db->db_mtx);
1598 if (zfs_refcount_count(&db->db_holds) == 0) {
1603 /* The dbuf is referenced */
1605 if (db->db_last_dirty != NULL) {
1606 dbuf_dirty_record_t *dr = db->db_last_dirty;
1608 if (dr->dr_txg == txg) {
1610 * This buffer is "in-use", re-adjust the file
1611 * size to reflect that this buffer may
1612 * contain new data when we sync.
1614 if (db->db_blkid != DMU_SPILL_BLKID &&
1615 db->db_blkid > dn->dn_maxblkid)
1616 dn->dn_maxblkid = db->db_blkid;
1617 dbuf_unoverride(dr);
1620 * This dbuf is not dirty in the open context.
1621 * Either uncache it (if its not referenced in
1622 * the open context) or reset its contents to
1625 dbuf_fix_old_data(db, txg);
1628 /* clear the contents if its cached */
1629 if (db->db_state == DB_CACHED) {
1630 ASSERT(db->db.db_data != NULL);
1631 arc_release(db->db_buf, db);
1632 bzero(db->db.db_data, db->db.db_size);
1633 arc_buf_freeze(db->db_buf);
1636 mutex_exit(&db->db_mtx);
1638 mutex_exit(&dn->dn_dbufs_mtx);
1642 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1644 arc_buf_t *buf, *obuf;
1645 int osize = db->db.db_size;
1646 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1649 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1654 /* XXX does *this* func really need the lock? */
1655 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1658 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1659 * is OK, because there can be no other references to the db
1660 * when we are changing its size, so no concurrent DB_FILL can
1664 * XXX we should be doing a dbuf_read, checking the return
1665 * value and returning that up to our callers
1667 dmu_buf_will_dirty(&db->db, tx);
1669 /* create the data buffer for the new block */
1670 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1672 /* copy old block data to the new block */
1674 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1675 /* zero the remainder */
1677 bzero((uint8_t *)buf->b_data + osize, size - osize);
1679 mutex_enter(&db->db_mtx);
1680 dbuf_set_data(db, buf);
1681 arc_buf_destroy(obuf, db);
1682 db->db.db_size = size;
1684 if (db->db_level == 0) {
1685 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1686 db->db_last_dirty->dt.dl.dr_data = buf;
1688 mutex_exit(&db->db_mtx);
1690 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1695 dbuf_release_bp(dmu_buf_impl_t *db)
1697 objset_t *os = db->db_objset;
1699 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1700 ASSERT(arc_released(os->os_phys_buf) ||
1701 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1702 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1704 (void) arc_release(db->db_buf, db);
1708 * We already have a dirty record for this TXG, and we are being
1712 dbuf_redirty(dbuf_dirty_record_t *dr)
1714 dmu_buf_impl_t *db = dr->dr_dbuf;
1716 ASSERT(MUTEX_HELD(&db->db_mtx));
1718 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1720 * If this buffer has already been written out,
1721 * we now need to reset its state.
1723 dbuf_unoverride(dr);
1724 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1725 db->db_state != DB_NOFILL) {
1726 /* Already released on initial dirty, so just thaw. */
1727 ASSERT(arc_released(db->db_buf));
1728 arc_buf_thaw(db->db_buf);
1733 dbuf_dirty_record_t *
1734 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1738 dbuf_dirty_record_t **drp, *dr;
1739 int drop_struct_lock = FALSE;
1740 int txgoff = tx->tx_txg & TXG_MASK;
1742 ASSERT(tx->tx_txg != 0);
1743 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1744 DMU_TX_DIRTY_BUF(tx, db);
1749 * Shouldn't dirty a regular buffer in syncing context. Private
1750 * objects may be dirtied in syncing context, but only if they
1751 * were already pre-dirtied in open context.
1754 if (dn->dn_objset->os_dsl_dataset != NULL) {
1755 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1758 ASSERT(!dmu_tx_is_syncing(tx) ||
1759 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1760 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1761 dn->dn_objset->os_dsl_dataset == NULL);
1762 if (dn->dn_objset->os_dsl_dataset != NULL)
1763 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1766 * We make this assert for private objects as well, but after we
1767 * check if we're already dirty. They are allowed to re-dirty
1768 * in syncing context.
1770 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1771 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1772 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1774 mutex_enter(&db->db_mtx);
1776 * XXX make this true for indirects too? The problem is that
1777 * transactions created with dmu_tx_create_assigned() from
1778 * syncing context don't bother holding ahead.
1780 ASSERT(db->db_level != 0 ||
1781 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1782 db->db_state == DB_NOFILL);
1784 mutex_enter(&dn->dn_mtx);
1786 * Don't set dirtyctx to SYNC if we're just modifying this as we
1787 * initialize the objset.
1789 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1790 if (dn->dn_objset->os_dsl_dataset != NULL) {
1791 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1794 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1795 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1796 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1797 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1798 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1800 if (dn->dn_objset->os_dsl_dataset != NULL) {
1801 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1806 if (tx->tx_txg > dn->dn_dirty_txg)
1807 dn->dn_dirty_txg = tx->tx_txg;
1808 mutex_exit(&dn->dn_mtx);
1810 if (db->db_blkid == DMU_SPILL_BLKID)
1811 dn->dn_have_spill = B_TRUE;
1814 * If this buffer is already dirty, we're done.
1816 drp = &db->db_last_dirty;
1817 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1818 db->db.db_object == DMU_META_DNODE_OBJECT);
1819 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1821 if (dr && dr->dr_txg == tx->tx_txg) {
1825 mutex_exit(&db->db_mtx);
1830 * Only valid if not already dirty.
1832 ASSERT(dn->dn_object == 0 ||
1833 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1834 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1836 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1839 * We should only be dirtying in syncing context if it's the
1840 * mos or we're initializing the os or it's a special object.
1841 * However, we are allowed to dirty in syncing context provided
1842 * we already dirtied it in open context. Hence we must make
1843 * this assertion only if we're not already dirty.
1846 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1848 if (dn->dn_objset->os_dsl_dataset != NULL)
1849 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1850 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1851 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1852 if (dn->dn_objset->os_dsl_dataset != NULL)
1853 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1855 ASSERT(db->db.db_size != 0);
1857 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1859 if (db->db_blkid != DMU_BONUS_BLKID) {
1860 dmu_objset_willuse_space(os, db->db.db_size, tx);
1864 * If this buffer is dirty in an old transaction group we need
1865 * to make a copy of it so that the changes we make in this
1866 * transaction group won't leak out when we sync the older txg.
1868 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1869 list_link_init(&dr->dr_dirty_node);
1870 if (db->db_level == 0) {
1871 void *data_old = db->db_buf;
1873 if (db->db_state != DB_NOFILL) {
1874 if (db->db_blkid == DMU_BONUS_BLKID) {
1875 dbuf_fix_old_data(db, tx->tx_txg);
1876 data_old = db->db.db_data;
1877 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1879 * Release the data buffer from the cache so
1880 * that we can modify it without impacting
1881 * possible other users of this cached data
1882 * block. Note that indirect blocks and
1883 * private objects are not released until the
1884 * syncing state (since they are only modified
1887 arc_release(db->db_buf, db);
1888 dbuf_fix_old_data(db, tx->tx_txg);
1889 data_old = db->db_buf;
1891 ASSERT(data_old != NULL);
1893 dr->dt.dl.dr_data = data_old;
1895 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1896 list_create(&dr->dt.di.dr_children,
1897 sizeof (dbuf_dirty_record_t),
1898 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1900 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1901 dr->dr_accounted = db->db.db_size;
1903 dr->dr_txg = tx->tx_txg;
1908 * We could have been freed_in_flight between the dbuf_noread
1909 * and dbuf_dirty. We win, as though the dbuf_noread() had
1910 * happened after the free.
1912 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1913 db->db_blkid != DMU_SPILL_BLKID) {
1914 mutex_enter(&dn->dn_mtx);
1915 if (dn->dn_free_ranges[txgoff] != NULL) {
1916 range_tree_clear(dn->dn_free_ranges[txgoff],
1919 mutex_exit(&dn->dn_mtx);
1920 db->db_freed_in_flight = FALSE;
1924 * This buffer is now part of this txg
1926 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1927 db->db_dirtycnt += 1;
1928 ASSERT3U(db->db_dirtycnt, <=, 3);
1930 mutex_exit(&db->db_mtx);
1932 if (db->db_blkid == DMU_BONUS_BLKID ||
1933 db->db_blkid == DMU_SPILL_BLKID) {
1934 mutex_enter(&dn->dn_mtx);
1935 ASSERT(!list_link_active(&dr->dr_dirty_node));
1936 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1937 mutex_exit(&dn->dn_mtx);
1938 dnode_setdirty(dn, tx);
1944 * The dn_struct_rwlock prevents db_blkptr from changing
1945 * due to a write from syncing context completing
1946 * while we are running, so we want to acquire it before
1947 * looking at db_blkptr.
1949 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1950 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1951 drop_struct_lock = TRUE;
1955 * We need to hold the dn_struct_rwlock to make this assertion,
1956 * because it protects dn_phys / dn_next_nlevels from changing.
1958 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1959 dn->dn_phys->dn_nlevels > db->db_level ||
1960 dn->dn_next_nlevels[txgoff] > db->db_level ||
1961 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1962 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1965 * If we are overwriting a dedup BP, then unless it is snapshotted,
1966 * when we get to syncing context we will need to decrement its
1967 * refcount in the DDT. Prefetch the relevant DDT block so that
1968 * syncing context won't have to wait for the i/o.
1970 ddt_prefetch(os->os_spa, db->db_blkptr);
1972 if (db->db_level == 0) {
1973 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1974 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1977 if (db->db_level+1 < dn->dn_nlevels) {
1978 dmu_buf_impl_t *parent = db->db_parent;
1979 dbuf_dirty_record_t *di;
1980 int parent_held = FALSE;
1982 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1983 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1985 parent = dbuf_hold_level(dn, db->db_level+1,
1986 db->db_blkid >> epbs, FTAG);
1987 ASSERT(parent != NULL);
1990 if (drop_struct_lock)
1991 rw_exit(&dn->dn_struct_rwlock);
1992 ASSERT3U(db->db_level+1, ==, parent->db_level);
1993 di = dbuf_dirty(parent, tx);
1995 dbuf_rele(parent, FTAG);
1997 mutex_enter(&db->db_mtx);
1999 * Since we've dropped the mutex, it's possible that
2000 * dbuf_undirty() might have changed this out from under us.
2002 if (db->db_last_dirty == dr ||
2003 dn->dn_object == DMU_META_DNODE_OBJECT) {
2004 mutex_enter(&di->dt.di.dr_mtx);
2005 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
2006 ASSERT(!list_link_active(&dr->dr_dirty_node));
2007 list_insert_tail(&di->dt.di.dr_children, dr);
2008 mutex_exit(&di->dt.di.dr_mtx);
2011 mutex_exit(&db->db_mtx);
2013 ASSERT(db->db_level+1 == dn->dn_nlevels);
2014 ASSERT(db->db_blkid < dn->dn_nblkptr);
2015 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
2016 mutex_enter(&dn->dn_mtx);
2017 ASSERT(!list_link_active(&dr->dr_dirty_node));
2018 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2019 mutex_exit(&dn->dn_mtx);
2020 if (drop_struct_lock)
2021 rw_exit(&dn->dn_struct_rwlock);
2024 dnode_setdirty(dn, tx);
2030 * Undirty a buffer in the transaction group referenced by the given
2031 * transaction. Return whether this evicted the dbuf.
2034 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2037 uint64_t txg = tx->tx_txg;
2038 dbuf_dirty_record_t *dr, **drp;
2043 * Due to our use of dn_nlevels below, this can only be called
2044 * in open context, unless we are operating on the MOS.
2045 * From syncing context, dn_nlevels may be different from the
2046 * dn_nlevels used when dbuf was dirtied.
2048 ASSERT(db->db_objset ==
2049 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2050 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2051 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2052 ASSERT0(db->db_level);
2053 ASSERT(MUTEX_HELD(&db->db_mtx));
2056 * If this buffer is not dirty, we're done.
2058 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
2059 if (dr->dr_txg <= txg)
2061 if (dr == NULL || dr->dr_txg < txg)
2063 ASSERT(dr->dr_txg == txg);
2064 ASSERT(dr->dr_dbuf == db);
2069 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2071 ASSERT(db->db.db_size != 0);
2073 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2074 dr->dr_accounted, txg);
2079 * Note that there are three places in dbuf_dirty()
2080 * where this dirty record may be put on a list.
2081 * Make sure to do a list_remove corresponding to
2082 * every one of those list_insert calls.
2084 if (dr->dr_parent) {
2085 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2086 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2087 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2088 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2089 db->db_level + 1 == dn->dn_nlevels) {
2090 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2091 mutex_enter(&dn->dn_mtx);
2092 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2093 mutex_exit(&dn->dn_mtx);
2097 if (db->db_state != DB_NOFILL) {
2098 dbuf_unoverride(dr);
2100 ASSERT(db->db_buf != NULL);
2101 ASSERT(dr->dt.dl.dr_data != NULL);
2102 if (dr->dt.dl.dr_data != db->db_buf)
2103 arc_buf_destroy(dr->dt.dl.dr_data, db);
2106 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2108 ASSERT(db->db_dirtycnt > 0);
2109 db->db_dirtycnt -= 1;
2111 if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2112 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2121 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2123 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2124 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
2126 ASSERT(tx->tx_txg != 0);
2127 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2130 * Quick check for dirtyness. For already dirty blocks, this
2131 * reduces runtime of this function by >90%, and overall performance
2132 * by 50% for some workloads (e.g. file deletion with indirect blocks
2135 mutex_enter(&db->db_mtx);
2136 dbuf_dirty_record_t *dr;
2137 for (dr = db->db_last_dirty;
2138 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
2140 * It's possible that it is already dirty but not cached,
2141 * because there are some calls to dbuf_dirty() that don't
2142 * go through dmu_buf_will_dirty().
2144 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
2145 /* This dbuf is already dirty and cached. */
2147 mutex_exit(&db->db_mtx);
2151 mutex_exit(&db->db_mtx);
2154 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2155 rf |= DB_RF_HAVESTRUCT;
2157 (void) dbuf_read(db, NULL, rf);
2158 (void) dbuf_dirty(db, tx);
2162 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2164 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2166 db->db_state = DB_NOFILL;
2168 dmu_buf_will_fill(db_fake, tx);
2172 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2174 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2176 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2177 ASSERT(tx->tx_txg != 0);
2178 ASSERT(db->db_level == 0);
2179 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2181 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2182 dmu_tx_private_ok(tx));
2185 (void) dbuf_dirty(db, tx);
2188 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2191 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2193 mutex_enter(&db->db_mtx);
2196 if (db->db_state == DB_FILL) {
2197 if (db->db_level == 0 && db->db_freed_in_flight) {
2198 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2199 /* we were freed while filling */
2200 /* XXX dbuf_undirty? */
2201 bzero(db->db.db_data, db->db.db_size);
2202 db->db_freed_in_flight = FALSE;
2204 db->db_state = DB_CACHED;
2205 cv_broadcast(&db->db_changed);
2207 mutex_exit(&db->db_mtx);
2211 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2212 bp_embedded_type_t etype, enum zio_compress comp,
2213 int uncompressed_size, int compressed_size, int byteorder,
2216 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2217 struct dirty_leaf *dl;
2218 dmu_object_type_t type;
2220 if (etype == BP_EMBEDDED_TYPE_DATA) {
2221 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2222 SPA_FEATURE_EMBEDDED_DATA));
2226 type = DB_DNODE(db)->dn_type;
2229 ASSERT0(db->db_level);
2230 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2232 dmu_buf_will_not_fill(dbuf, tx);
2234 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2235 dl = &db->db_last_dirty->dt.dl;
2236 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2237 data, comp, uncompressed_size, compressed_size);
2238 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2239 BP_SET_TYPE(&dl->dr_overridden_by, type);
2240 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2241 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2243 dl->dr_override_state = DR_OVERRIDDEN;
2244 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2248 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2249 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2252 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2254 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2255 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2256 ASSERT(db->db_level == 0);
2257 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2258 ASSERT(buf != NULL);
2259 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2260 ASSERT(tx->tx_txg != 0);
2262 arc_return_buf(buf, db);
2263 ASSERT(arc_released(buf));
2265 mutex_enter(&db->db_mtx);
2267 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2268 cv_wait(&db->db_changed, &db->db_mtx);
2270 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2272 if (db->db_state == DB_CACHED &&
2273 zfs_refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2274 mutex_exit(&db->db_mtx);
2275 (void) dbuf_dirty(db, tx);
2276 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2277 arc_buf_destroy(buf, db);
2278 xuio_stat_wbuf_copied();
2282 xuio_stat_wbuf_nocopy();
2283 if (db->db_state == DB_CACHED) {
2284 dbuf_dirty_record_t *dr = db->db_last_dirty;
2286 ASSERT(db->db_buf != NULL);
2287 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2288 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2289 if (!arc_released(db->db_buf)) {
2290 ASSERT(dr->dt.dl.dr_override_state ==
2292 arc_release(db->db_buf, db);
2294 dr->dt.dl.dr_data = buf;
2295 arc_buf_destroy(db->db_buf, db);
2296 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2297 arc_release(db->db_buf, db);
2298 arc_buf_destroy(db->db_buf, db);
2302 ASSERT(db->db_buf == NULL);
2303 dbuf_set_data(db, buf);
2304 db->db_state = DB_FILL;
2305 mutex_exit(&db->db_mtx);
2306 (void) dbuf_dirty(db, tx);
2307 dmu_buf_fill_done(&db->db, tx);
2311 dbuf_destroy(dmu_buf_impl_t *db)
2314 dmu_buf_impl_t *parent = db->db_parent;
2315 dmu_buf_impl_t *dndb;
2317 ASSERT(MUTEX_HELD(&db->db_mtx));
2318 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2320 if (db->db_buf != NULL) {
2321 arc_buf_destroy(db->db_buf, db);
2325 if (db->db_blkid == DMU_BONUS_BLKID) {
2326 int slots = DB_DNODE(db)->dn_num_slots;
2327 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2328 if (db->db.db_data != NULL) {
2329 zio_buf_free(db->db.db_data, bonuslen);
2330 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2331 db->db_state = DB_UNCACHED;
2335 dbuf_clear_data(db);
2337 if (multilist_link_active(&db->db_cache_link)) {
2338 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
2339 db->db_caching_status == DB_DBUF_METADATA_CACHE);
2341 multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
2342 (void) zfs_refcount_remove_many(
2343 &dbuf_caches[db->db_caching_status].size,
2344 db->db.db_size, db);
2346 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2347 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2349 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2350 DBUF_STAT_BUMPDOWN(cache_count);
2351 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2354 db->db_caching_status = DB_NO_CACHE;
2357 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2358 ASSERT(db->db_data_pending == NULL);
2360 db->db_state = DB_EVICTING;
2361 db->db_blkptr = NULL;
2364 * Now that db_state is DB_EVICTING, nobody else can find this via
2365 * the hash table. We can now drop db_mtx, which allows us to
2366 * acquire the dn_dbufs_mtx.
2368 mutex_exit(&db->db_mtx);
2373 if (db->db_blkid != DMU_BONUS_BLKID) {
2374 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2376 mutex_enter(&dn->dn_dbufs_mtx);
2377 avl_remove(&dn->dn_dbufs, db);
2381 mutex_exit(&dn->dn_dbufs_mtx);
2383 * Decrementing the dbuf count means that the hold corresponding
2384 * to the removed dbuf is no longer discounted in dnode_move(),
2385 * so the dnode cannot be moved until after we release the hold.
2386 * The membar_producer() ensures visibility of the decremented
2387 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2390 mutex_enter(&dn->dn_mtx);
2391 dnode_rele_and_unlock(dn, db, B_TRUE);
2392 db->db_dnode_handle = NULL;
2394 dbuf_hash_remove(db);
2399 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2401 db->db_parent = NULL;
2403 ASSERT(db->db_buf == NULL);
2404 ASSERT(db->db.db_data == NULL);
2405 ASSERT(db->db_hash_next == NULL);
2406 ASSERT(db->db_blkptr == NULL);
2407 ASSERT(db->db_data_pending == NULL);
2408 ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE);
2409 ASSERT(!multilist_link_active(&db->db_cache_link));
2411 kmem_cache_free(dbuf_kmem_cache, db);
2412 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2415 * If this dbuf is referenced from an indirect dbuf,
2416 * decrement the ref count on the indirect dbuf.
2418 if (parent && parent != dndb) {
2419 mutex_enter(&parent->db_mtx);
2420 dbuf_rele_and_unlock(parent, db, B_TRUE);
2425 * Note: While bpp will always be updated if the function returns success,
2426 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2427 * this happens when the dnode is the meta-dnode, or a userused or groupused
2430 __attribute__((always_inline))
2432 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2433 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2438 ASSERT(blkid != DMU_BONUS_BLKID);
2440 if (blkid == DMU_SPILL_BLKID) {
2441 mutex_enter(&dn->dn_mtx);
2442 if (dn->dn_have_spill &&
2443 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2444 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2447 dbuf_add_ref(dn->dn_dbuf, NULL);
2448 *parentp = dn->dn_dbuf;
2449 mutex_exit(&dn->dn_mtx);
2454 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2455 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2457 ASSERT3U(level * epbs, <, 64);
2458 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2460 * This assertion shouldn't trip as long as the max indirect block size
2461 * is less than 1M. The reason for this is that up to that point,
2462 * the number of levels required to address an entire object with blocks
2463 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2464 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2465 * (i.e. we can address the entire object), objects will all use at most
2466 * N-1 levels and the assertion won't overflow. However, once epbs is
2467 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2468 * enough to address an entire object, so objects will have 5 levels,
2469 * but then this assertion will overflow.
2471 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2472 * need to redo this logic to handle overflows.
2474 ASSERT(level >= nlevels ||
2475 ((nlevels - level - 1) * epbs) +
2476 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2477 if (level >= nlevels ||
2478 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2479 ((nlevels - level - 1) * epbs)) ||
2481 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2482 /* the buffer has no parent yet */
2483 return (SET_ERROR(ENOENT));
2484 } else if (level < nlevels-1) {
2485 /* this block is referenced from an indirect block */
2488 err = dbuf_hold_impl(dn, level+1,
2489 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2491 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2492 blkid >> epbs, fail_sparse, FALSE, NULL,
2493 parentp, dh->dh_depth + 1);
2494 err = __dbuf_hold_impl(dh + 1);
2498 err = dbuf_read(*parentp, NULL,
2499 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2501 dbuf_rele(*parentp, NULL);
2505 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2506 (blkid & ((1ULL << epbs) - 1));
2507 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2508 ASSERT(BP_IS_HOLE(*bpp));
2511 /* the block is referenced from the dnode */
2512 ASSERT3U(level, ==, nlevels-1);
2513 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2514 blkid < dn->dn_phys->dn_nblkptr);
2516 dbuf_add_ref(dn->dn_dbuf, NULL);
2517 *parentp = dn->dn_dbuf;
2519 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2524 static dmu_buf_impl_t *
2525 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2526 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2528 objset_t *os = dn->dn_objset;
2529 dmu_buf_impl_t *db, *odb;
2531 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2532 ASSERT(dn->dn_type != DMU_OT_NONE);
2534 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2537 db->db.db_object = dn->dn_object;
2538 db->db_level = level;
2539 db->db_blkid = blkid;
2540 db->db_last_dirty = NULL;
2541 db->db_dirtycnt = 0;
2542 db->db_dnode_handle = dn->dn_handle;
2543 db->db_parent = parent;
2544 db->db_blkptr = blkptr;
2547 db->db_user_immediate_evict = FALSE;
2548 db->db_freed_in_flight = FALSE;
2549 db->db_pending_evict = FALSE;
2551 if (blkid == DMU_BONUS_BLKID) {
2552 ASSERT3P(parent, ==, dn->dn_dbuf);
2553 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2554 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2555 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2556 db->db.db_offset = DMU_BONUS_BLKID;
2557 db->db_state = DB_UNCACHED;
2558 db->db_caching_status = DB_NO_CACHE;
2559 /* the bonus dbuf is not placed in the hash table */
2560 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2562 } else if (blkid == DMU_SPILL_BLKID) {
2563 db->db.db_size = (blkptr != NULL) ?
2564 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2565 db->db.db_offset = 0;
2568 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2569 db->db.db_size = blocksize;
2570 db->db.db_offset = db->db_blkid * blocksize;
2574 * Hold the dn_dbufs_mtx while we get the new dbuf
2575 * in the hash table *and* added to the dbufs list.
2576 * This prevents a possible deadlock with someone
2577 * trying to look up this dbuf before its added to the
2580 mutex_enter(&dn->dn_dbufs_mtx);
2581 db->db_state = DB_EVICTING;
2582 if ((odb = dbuf_hash_insert(db)) != NULL) {
2583 /* someone else inserted it first */
2584 kmem_cache_free(dbuf_kmem_cache, db);
2585 mutex_exit(&dn->dn_dbufs_mtx);
2586 DBUF_STAT_BUMP(hash_insert_race);
2589 avl_add(&dn->dn_dbufs, db);
2591 db->db_state = DB_UNCACHED;
2592 db->db_caching_status = DB_NO_CACHE;
2593 mutex_exit(&dn->dn_dbufs_mtx);
2594 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2596 if (parent && parent != dn->dn_dbuf)
2597 dbuf_add_ref(parent, db);
2599 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2600 zfs_refcount_count(&dn->dn_holds) > 0);
2601 (void) zfs_refcount_add(&dn->dn_holds, db);
2603 dprintf_dbuf(db, "db=%p\n", db);
2608 typedef struct dbuf_prefetch_arg {
2609 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2610 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2611 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2612 int dpa_curlevel; /* The current level that we're reading */
2613 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2614 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2615 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2616 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2617 } dbuf_prefetch_arg_t;
2620 * Actually issue the prefetch read for the block given.
2623 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2625 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2628 arc_flags_t aflags =
2629 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2631 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2632 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2633 ASSERT(dpa->dpa_zio != NULL);
2634 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2635 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2636 &aflags, &dpa->dpa_zb);
2640 * Called when an indirect block above our prefetch target is read in. This
2641 * will either read in the next indirect block down the tree or issue the actual
2642 * prefetch if the next block down is our target.
2645 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
2646 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
2648 dbuf_prefetch_arg_t *dpa = private;
2650 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2651 ASSERT3S(dpa->dpa_curlevel, >, 0);
2654 ASSERT(zio == NULL || zio->io_error != 0);
2655 kmem_free(dpa, sizeof (*dpa));
2658 ASSERT(zio == NULL || zio->io_error == 0);
2661 * The dpa_dnode is only valid if we are called with a NULL
2662 * zio. This indicates that the arc_read() returned without
2663 * first calling zio_read() to issue a physical read. Once
2664 * a physical read is made the dpa_dnode must be invalidated
2665 * as the locks guarding it may have been dropped. If the
2666 * dpa_dnode is still valid, then we want to add it to the dbuf
2667 * cache. To do so, we must hold the dbuf associated with the block
2668 * we just prefetched, read its contents so that we associate it
2669 * with an arc_buf_t, and then release it.
2672 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2673 if (zio->io_flags & ZIO_FLAG_RAW) {
2674 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2676 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2678 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2680 dpa->dpa_dnode = NULL;
2681 } else if (dpa->dpa_dnode != NULL) {
2682 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2683 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2684 dpa->dpa_zb.zb_level));
2685 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2686 dpa->dpa_curlevel, curblkid, FTAG);
2687 (void) dbuf_read(db, NULL,
2688 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2689 dbuf_rele(db, FTAG);
2693 kmem_free(dpa, sizeof(*dpa));
2697 dpa->dpa_curlevel--;
2699 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2700 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2701 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2702 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2703 if (BP_IS_HOLE(bp)) {
2704 kmem_free(dpa, sizeof (*dpa));
2705 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2706 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2707 dbuf_issue_final_prefetch(dpa, bp);
2708 kmem_free(dpa, sizeof (*dpa));
2710 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2711 zbookmark_phys_t zb;
2713 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2714 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2715 iter_aflags |= ARC_FLAG_L2CACHE;
2717 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2719 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2720 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2722 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2723 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2724 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2728 arc_buf_destroy(abuf, private);
2732 * Issue prefetch reads for the given block on the given level. If the indirect
2733 * blocks above that block are not in memory, we will read them in
2734 * asynchronously. As a result, this call never blocks waiting for a read to
2738 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2742 int epbs, nlevels, curlevel;
2745 ASSERT(blkid != DMU_BONUS_BLKID);
2746 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2748 if (blkid > dn->dn_maxblkid)
2751 if (dnode_block_freed(dn, blkid))
2755 * This dnode hasn't been written to disk yet, so there's nothing to
2758 nlevels = dn->dn_phys->dn_nlevels;
2759 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2762 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2763 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2766 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2769 mutex_exit(&db->db_mtx);
2771 * This dbuf already exists. It is either CACHED, or
2772 * (we assume) about to be read or filled.
2778 * Find the closest ancestor (indirect block) of the target block
2779 * that is present in the cache. In this indirect block, we will
2780 * find the bp that is at curlevel, curblkid.
2784 while (curlevel < nlevels - 1) {
2785 int parent_level = curlevel + 1;
2786 uint64_t parent_blkid = curblkid >> epbs;
2789 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2790 FALSE, TRUE, FTAG, &db) == 0) {
2791 blkptr_t *bpp = db->db_buf->b_data;
2792 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2793 dbuf_rele(db, FTAG);
2797 curlevel = parent_level;
2798 curblkid = parent_blkid;
2801 if (curlevel == nlevels - 1) {
2802 /* No cached indirect blocks found. */
2803 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2804 bp = dn->dn_phys->dn_blkptr[curblkid];
2806 if (BP_IS_HOLE(&bp))
2809 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2811 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2814 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2815 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2816 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2817 dn->dn_object, level, blkid);
2818 dpa->dpa_curlevel = curlevel;
2819 dpa->dpa_prio = prio;
2820 dpa->dpa_aflags = aflags;
2821 dpa->dpa_spa = dn->dn_objset->os_spa;
2822 dpa->dpa_dnode = dn;
2823 dpa->dpa_epbs = epbs;
2826 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2827 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2828 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2831 * If we have the indirect just above us, no need to do the asynchronous
2832 * prefetch chain; we'll just run the last step ourselves. If we're at
2833 * a higher level, though, we want to issue the prefetches for all the
2834 * indirect blocks asynchronously, so we can go on with whatever we were
2837 if (curlevel == level) {
2838 ASSERT3U(curblkid, ==, blkid);
2839 dbuf_issue_final_prefetch(dpa, &bp);
2840 kmem_free(dpa, sizeof (*dpa));
2842 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2843 zbookmark_phys_t zb;
2845 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2846 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2847 iter_aflags |= ARC_FLAG_L2CACHE;
2849 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2850 dn->dn_object, curlevel, curblkid);
2851 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2852 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2853 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2857 * We use pio here instead of dpa_zio since it's possible that
2858 * dpa may have already been freed.
2863 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2866 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2867 * the case of encrypted, compressed and uncompressed buffers by
2868 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2869 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2871 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2873 noinline static void
2874 dbuf_hold_copy(struct dbuf_hold_impl_data *dh)
2876 dnode_t *dn = dh->dh_dn;
2877 dmu_buf_impl_t *db = dh->dh_db;
2878 dbuf_dirty_record_t *dr = dh->dh_dr;
2879 arc_buf_t *data = dr->dt.dl.dr_data;
2881 enum zio_compress compress_type = arc_get_compression(data);
2883 if (compress_type != ZIO_COMPRESS_OFF) {
2884 dbuf_set_data(db, arc_alloc_compressed_buf(
2885 dn->dn_objset->os_spa, db, arc_buf_size(data),
2886 arc_buf_lsize(data), compress_type));
2888 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db,
2889 DBUF_GET_BUFC_TYPE(db), db->db.db_size));
2892 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
2896 * Returns with db_holds incremented, and db_mtx not held.
2897 * Note: dn_struct_rwlock must be held.
2900 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2902 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2903 dh->dh_parent = NULL;
2905 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2906 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2907 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2909 *(dh->dh_dbp) = NULL;
2911 /* dbuf_find() returns with db_mtx held */
2912 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2913 dh->dh_level, dh->dh_blkid);
2915 if (dh->dh_db == NULL) {
2918 if (dh->dh_fail_uncached)
2919 return (SET_ERROR(ENOENT));
2921 ASSERT3P(dh->dh_parent, ==, NULL);
2922 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2923 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2924 if (dh->dh_fail_sparse) {
2925 if (dh->dh_err == 0 &&
2926 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2927 dh->dh_err = SET_ERROR(ENOENT);
2930 dbuf_rele(dh->dh_parent, NULL);
2931 return (dh->dh_err);
2934 if (dh->dh_err && dh->dh_err != ENOENT)
2935 return (dh->dh_err);
2936 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2937 dh->dh_parent, dh->dh_bp);
2940 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2941 mutex_exit(&dh->dh_db->db_mtx);
2942 return (SET_ERROR(ENOENT));
2945 if (dh->dh_db->db_buf != NULL) {
2946 arc_buf_access(dh->dh_db->db_buf);
2947 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2950 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2953 * If this buffer is currently syncing out, and we are are
2954 * still referencing it from db_data, we need to make a copy
2955 * of it in case we decide we want to dirty it again in this txg.
2957 if (dh->dh_db->db_level == 0 &&
2958 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2959 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2960 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2961 dh->dh_dr = dh->dh_db->db_data_pending;
2962 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf)
2966 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2967 ASSERT(zfs_refcount_is_zero(&dh->dh_db->db_holds));
2968 ASSERT(dh->dh_db->db_caching_status == DB_DBUF_CACHE ||
2969 dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE);
2972 dbuf_caches[dh->dh_db->db_caching_status].cache,
2974 (void) zfs_refcount_remove_many(
2975 &dbuf_caches[dh->dh_db->db_caching_status].size,
2976 dh->dh_db->db.db_size, dh->dh_db);
2978 if (dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2979 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2981 DBUF_STAT_BUMPDOWN(cache_levels[dh->dh_db->db_level]);
2982 DBUF_STAT_BUMPDOWN(cache_count);
2983 DBUF_STAT_DECR(cache_levels_bytes[dh->dh_db->db_level],
2984 dh->dh_db->db.db_size);
2986 dh->dh_db->db_caching_status = DB_NO_CACHE;
2988 (void) zfs_refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
2989 DBUF_VERIFY(dh->dh_db);
2990 mutex_exit(&dh->dh_db->db_mtx);
2992 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2994 dbuf_rele(dh->dh_parent, NULL);
2996 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
2997 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
2998 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
2999 *(dh->dh_dbp) = dh->dh_db;
3005 * The following code preserves the recursive function dbuf_hold_impl()
3006 * but moves the local variables AND function arguments to the heap to
3007 * minimize the stack frame size. Enough space is initially allocated
3008 * on the stack for 20 levels of recursion.
3011 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3012 boolean_t fail_sparse, boolean_t fail_uncached,
3013 void *tag, dmu_buf_impl_t **dbp)
3015 struct dbuf_hold_impl_data *dh;
3018 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
3019 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
3020 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
3021 fail_uncached, tag, dbp, 0);
3023 error = __dbuf_hold_impl(dh);
3025 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
3026 DBUF_HOLD_IMPL_MAX_DEPTH);
3032 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
3033 dnode_t *dn, uint8_t level, uint64_t blkid,
3034 boolean_t fail_sparse, boolean_t fail_uncached,
3035 void *tag, dmu_buf_impl_t **dbp, int depth)
3038 dh->dh_level = level;
3039 dh->dh_blkid = blkid;
3041 dh->dh_fail_sparse = fail_sparse;
3042 dh->dh_fail_uncached = fail_uncached;
3048 dh->dh_parent = NULL;
3053 dh->dh_depth = depth;
3057 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3059 return (dbuf_hold_level(dn, 0, blkid, tag));
3063 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3066 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3067 return (err ? NULL : db);
3071 dbuf_create_bonus(dnode_t *dn)
3073 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3075 ASSERT(dn->dn_bonus == NULL);
3076 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3080 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3082 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3085 if (db->db_blkid != DMU_SPILL_BLKID)
3086 return (SET_ERROR(ENOTSUP));
3088 blksz = SPA_MINBLOCKSIZE;
3089 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3090 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3094 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3095 dbuf_new_size(db, blksz, tx);
3096 rw_exit(&dn->dn_struct_rwlock);
3103 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3105 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3108 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3110 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3112 int64_t holds = zfs_refcount_add(&db->db_holds, tag);
3113 ASSERT3S(holds, >, 1);
3116 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3118 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3121 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3122 dmu_buf_impl_t *found_db;
3123 boolean_t result = B_FALSE;
3125 if (db->db_blkid == DMU_BONUS_BLKID)
3126 found_db = dbuf_find_bonus(os, obj);
3128 found_db = dbuf_find(os, obj, 0, blkid);
3130 if (found_db != NULL) {
3131 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3132 (void) zfs_refcount_add(&db->db_holds, tag);
3135 mutex_exit(&db->db_mtx);
3141 * If you call dbuf_rele() you had better not be referencing the dnode handle
3142 * unless you have some other direct or indirect hold on the dnode. (An indirect
3143 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3144 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3145 * dnode's parent dbuf evicting its dnode handles.
3148 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3150 mutex_enter(&db->db_mtx);
3151 dbuf_rele_and_unlock(db, tag, B_FALSE);
3155 dmu_buf_rele(dmu_buf_t *db, void *tag)
3157 dbuf_rele((dmu_buf_impl_t *)db, tag);
3161 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3162 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3163 * argument should be set if we are already in the dbuf-evicting code
3164 * path, in which case we don't want to recursively evict. This allows us to
3165 * avoid deeply nested stacks that would have a call flow similar to this:
3167 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3170 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3174 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
3179 ASSERT(MUTEX_HELD(&db->db_mtx));
3183 * Remove the reference to the dbuf before removing its hold on the
3184 * dnode so we can guarantee in dnode_move() that a referenced bonus
3185 * buffer has a corresponding dnode hold.
3187 holds = zfs_refcount_remove(&db->db_holds, tag);
3191 * We can't freeze indirects if there is a possibility that they
3192 * may be modified in the current syncing context.
3194 if (db->db_buf != NULL &&
3195 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3196 arc_buf_freeze(db->db_buf);
3199 if (holds == db->db_dirtycnt &&
3200 db->db_level == 0 && db->db_user_immediate_evict)
3201 dbuf_evict_user(db);
3204 if (db->db_blkid == DMU_BONUS_BLKID) {
3206 boolean_t evict_dbuf = db->db_pending_evict;
3209 * If the dnode moves here, we cannot cross this
3210 * barrier until the move completes.
3215 atomic_dec_32(&dn->dn_dbufs_count);
3218 * Decrementing the dbuf count means that the bonus
3219 * buffer's dnode hold is no longer discounted in
3220 * dnode_move(). The dnode cannot move until after
3221 * the dnode_rele() below.
3226 * Do not reference db after its lock is dropped.
3227 * Another thread may evict it.
3229 mutex_exit(&db->db_mtx);
3232 dnode_evict_bonus(dn);
3235 } else if (db->db_buf == NULL) {
3237 * This is a special case: we never associated this
3238 * dbuf with any data allocated from the ARC.
3240 ASSERT(db->db_state == DB_UNCACHED ||
3241 db->db_state == DB_NOFILL);
3243 } else if (arc_released(db->db_buf)) {
3245 * This dbuf has anonymous data associated with it.
3249 boolean_t do_arc_evict = B_FALSE;
3251 spa_t *spa = dmu_objset_spa(db->db_objset);
3253 if (!DBUF_IS_CACHEABLE(db) &&
3254 db->db_blkptr != NULL &&
3255 !BP_IS_HOLE(db->db_blkptr) &&
3256 !BP_IS_EMBEDDED(db->db_blkptr)) {
3257 do_arc_evict = B_TRUE;
3258 bp = *db->db_blkptr;
3261 if (!DBUF_IS_CACHEABLE(db) ||
3262 db->db_pending_evict) {
3264 } else if (!multilist_link_active(&db->db_cache_link)) {
3265 ASSERT3U(db->db_caching_status, ==,
3268 dbuf_cached_state_t dcs =
3269 dbuf_include_in_metadata_cache(db) ?
3270 DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
3271 db->db_caching_status = dcs;
3273 multilist_insert(dbuf_caches[dcs].cache, db);
3274 size = zfs_refcount_add_many(
3275 &dbuf_caches[dcs].size, db->db.db_size, db);
3277 if (dcs == DB_DBUF_METADATA_CACHE) {
3278 DBUF_STAT_BUMP(metadata_cache_count);
3280 metadata_cache_size_bytes_max,
3284 cache_levels[db->db_level]);
3285 DBUF_STAT_BUMP(cache_count);
3287 cache_levels_bytes[db->db_level],
3289 DBUF_STAT_MAX(cache_size_bytes_max,
3292 mutex_exit(&db->db_mtx);
3294 if (dcs == DB_DBUF_CACHE && !evicting)
3295 dbuf_evict_notify(size);
3299 arc_freed(spa, &bp);
3302 mutex_exit(&db->db_mtx);
3307 #pragma weak dmu_buf_refcount = dbuf_refcount
3309 dbuf_refcount(dmu_buf_impl_t *db)
3311 return (zfs_refcount_count(&db->db_holds));
3315 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3316 dmu_buf_user_t *new_user)
3318 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3320 mutex_enter(&db->db_mtx);
3321 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3322 if (db->db_user == old_user)
3323 db->db_user = new_user;
3325 old_user = db->db_user;
3326 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3327 mutex_exit(&db->db_mtx);
3333 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3335 return (dmu_buf_replace_user(db_fake, NULL, user));
3339 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3341 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3343 db->db_user_immediate_evict = TRUE;
3344 return (dmu_buf_set_user(db_fake, user));
3348 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3350 return (dmu_buf_replace_user(db_fake, user, NULL));
3354 dmu_buf_get_user(dmu_buf_t *db_fake)
3356 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3358 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3359 return (db->db_user);
3363 dmu_buf_user_evict_wait()
3365 taskq_wait(dbu_evict_taskq);
3369 dmu_buf_get_blkptr(dmu_buf_t *db)
3371 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3372 return (dbi->db_blkptr);
3376 dmu_buf_get_objset(dmu_buf_t *db)
3378 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3379 return (dbi->db_objset);
3383 dmu_buf_dnode_enter(dmu_buf_t *db)
3385 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3386 DB_DNODE_ENTER(dbi);
3387 return (DB_DNODE(dbi));
3391 dmu_buf_dnode_exit(dmu_buf_t *db)
3393 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3398 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3400 /* ASSERT(dmu_tx_is_syncing(tx) */
3401 ASSERT(MUTEX_HELD(&db->db_mtx));
3403 if (db->db_blkptr != NULL)
3406 if (db->db_blkid == DMU_SPILL_BLKID) {
3407 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3408 BP_ZERO(db->db_blkptr);
3411 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3413 * This buffer was allocated at a time when there was
3414 * no available blkptrs from the dnode, or it was
3415 * inappropriate to hook it in (i.e., nlevels mis-match).
3417 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3418 ASSERT(db->db_parent == NULL);
3419 db->db_parent = dn->dn_dbuf;
3420 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3423 dmu_buf_impl_t *parent = db->db_parent;
3424 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3426 ASSERT(dn->dn_phys->dn_nlevels > 1);
3427 if (parent == NULL) {
3428 mutex_exit(&db->db_mtx);
3429 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3430 parent = dbuf_hold_level(dn, db->db_level + 1,
3431 db->db_blkid >> epbs, db);
3432 rw_exit(&dn->dn_struct_rwlock);
3433 mutex_enter(&db->db_mtx);
3434 db->db_parent = parent;
3436 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3437 (db->db_blkid & ((1ULL << epbs) - 1));
3443 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3444 * is critical the we not allow the compiler to inline this function in to
3445 * dbuf_sync_list() thereby drastically bloating the stack usage.
3447 noinline static void
3448 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3450 dmu_buf_impl_t *db = dr->dr_dbuf;
3454 ASSERT(dmu_tx_is_syncing(tx));
3456 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3458 mutex_enter(&db->db_mtx);
3460 ASSERT(db->db_level > 0);
3463 /* Read the block if it hasn't been read yet. */
3464 if (db->db_buf == NULL) {
3465 mutex_exit(&db->db_mtx);
3466 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3467 mutex_enter(&db->db_mtx);
3469 ASSERT3U(db->db_state, ==, DB_CACHED);
3470 ASSERT(db->db_buf != NULL);
3474 /* Indirect block size must match what the dnode thinks it is. */
3475 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3476 dbuf_check_blkptr(dn, db);
3479 /* Provide the pending dirty record to child dbufs */
3480 db->db_data_pending = dr;
3482 mutex_exit(&db->db_mtx);
3484 dbuf_write(dr, db->db_buf, tx);
3487 mutex_enter(&dr->dt.di.dr_mtx);
3488 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3489 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3490 mutex_exit(&dr->dt.di.dr_mtx);
3495 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3496 * critical the we not allow the compiler to inline this function in to
3497 * dbuf_sync_list() thereby drastically bloating the stack usage.
3499 noinline static void
3500 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3502 arc_buf_t **datap = &dr->dt.dl.dr_data;
3503 dmu_buf_impl_t *db = dr->dr_dbuf;
3506 uint64_t txg = tx->tx_txg;
3508 ASSERT(dmu_tx_is_syncing(tx));
3510 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3512 mutex_enter(&db->db_mtx);
3514 * To be synced, we must be dirtied. But we
3515 * might have been freed after the dirty.
3517 if (db->db_state == DB_UNCACHED) {
3518 /* This buffer has been freed since it was dirtied */
3519 ASSERT(db->db.db_data == NULL);
3520 } else if (db->db_state == DB_FILL) {
3521 /* This buffer was freed and is now being re-filled */
3522 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3524 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3531 if (db->db_blkid == DMU_SPILL_BLKID) {
3532 mutex_enter(&dn->dn_mtx);
3533 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3535 * In the previous transaction group, the bonus buffer
3536 * was entirely used to store the attributes for the
3537 * dnode which overrode the dn_spill field. However,
3538 * when adding more attributes to the file a spill
3539 * block was required to hold the extra attributes.
3541 * Make sure to clear the garbage left in the dn_spill
3542 * field from the previous attributes in the bonus
3543 * buffer. Otherwise, after writing out the spill
3544 * block to the new allocated dva, it will free
3545 * the old block pointed to by the invalid dn_spill.
3547 db->db_blkptr = NULL;
3549 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3550 mutex_exit(&dn->dn_mtx);
3554 * If this is a bonus buffer, simply copy the bonus data into the
3555 * dnode. It will be written out when the dnode is synced (and it
3556 * will be synced, since it must have been dirty for dbuf_sync to
3559 if (db->db_blkid == DMU_BONUS_BLKID) {
3560 dbuf_dirty_record_t **drp;
3562 ASSERT(*datap != NULL);
3563 ASSERT0(db->db_level);
3564 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3565 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3566 bcopy(*datap, DN_BONUS(dn->dn_phys),
3567 DN_MAX_BONUS_LEN(dn->dn_phys));
3570 if (*datap != db->db.db_data) {
3571 int slots = DB_DNODE(db)->dn_num_slots;
3572 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3573 zio_buf_free(*datap, bonuslen);
3574 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3576 db->db_data_pending = NULL;
3577 drp = &db->db_last_dirty;
3579 drp = &(*drp)->dr_next;
3580 ASSERT(dr->dr_next == NULL);
3581 ASSERT(dr->dr_dbuf == db);
3583 if (dr->dr_dbuf->db_level != 0) {
3584 mutex_destroy(&dr->dt.di.dr_mtx);
3585 list_destroy(&dr->dt.di.dr_children);
3587 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3588 ASSERT(db->db_dirtycnt > 0);
3589 db->db_dirtycnt -= 1;
3590 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
3597 * This function may have dropped the db_mtx lock allowing a dmu_sync
3598 * operation to sneak in. As a result, we need to ensure that we
3599 * don't check the dr_override_state until we have returned from
3600 * dbuf_check_blkptr.
3602 dbuf_check_blkptr(dn, db);
3605 * If this buffer is in the middle of an immediate write,
3606 * wait for the synchronous IO to complete.
3608 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3609 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3610 cv_wait(&db->db_changed, &db->db_mtx);
3611 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3614 if (db->db_state != DB_NOFILL &&
3615 dn->dn_object != DMU_META_DNODE_OBJECT &&
3616 zfs_refcount_count(&db->db_holds) > 1 &&
3617 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3618 *datap == db->db_buf) {
3620 * If this buffer is currently "in use" (i.e., there
3621 * are active holds and db_data still references it),
3622 * then make a copy before we start the write so that
3623 * any modifications from the open txg will not leak
3626 * NOTE: this copy does not need to be made for
3627 * objects only modified in the syncing context (e.g.
3628 * DNONE_DNODE blocks).
3630 int psize = arc_buf_size(*datap);
3631 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3632 enum zio_compress compress_type = arc_get_compression(*datap);
3634 if (compress_type == ZIO_COMPRESS_OFF) {
3635 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3637 ASSERT3U(type, ==, ARC_BUFC_DATA);
3638 int lsize = arc_buf_lsize(*datap);
3639 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3640 psize, lsize, compress_type);
3642 bcopy(db->db.db_data, (*datap)->b_data, psize);
3644 db->db_data_pending = dr;
3646 mutex_exit(&db->db_mtx);
3648 dbuf_write(dr, *datap, tx);
3650 ASSERT(!list_link_active(&dr->dr_dirty_node));
3651 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3652 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3656 * Although zio_nowait() does not "wait for an IO", it does
3657 * initiate the IO. If this is an empty write it seems plausible
3658 * that the IO could actually be completed before the nowait
3659 * returns. We need to DB_DNODE_EXIT() first in case
3660 * zio_nowait() invalidates the dbuf.
3663 zio_nowait(dr->dr_zio);
3668 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3670 dbuf_dirty_record_t *dr;
3672 while (dr = list_head(list)) {
3673 if (dr->dr_zio != NULL) {
3675 * If we find an already initialized zio then we
3676 * are processing the meta-dnode, and we have finished.
3677 * The dbufs for all dnodes are put back on the list
3678 * during processing, so that we can zio_wait()
3679 * these IOs after initiating all child IOs.
3681 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3682 DMU_META_DNODE_OBJECT);
3685 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3686 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3687 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3689 list_remove(list, dr);
3690 if (dr->dr_dbuf->db_level > 0)
3691 dbuf_sync_indirect(dr, tx);
3693 dbuf_sync_leaf(dr, tx);
3699 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3701 dmu_buf_impl_t *db = vdb;
3703 blkptr_t *bp = zio->io_bp;
3704 blkptr_t *bp_orig = &zio->io_bp_orig;
3705 spa_t *spa = zio->io_spa;
3710 ASSERT3P(db->db_blkptr, !=, NULL);
3711 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3715 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3716 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3717 zio->io_prev_space_delta = delta;
3719 if (bp->blk_birth != 0) {
3720 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3721 BP_GET_TYPE(bp) == dn->dn_type) ||
3722 (db->db_blkid == DMU_SPILL_BLKID &&
3723 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3724 BP_IS_EMBEDDED(bp));
3725 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3728 mutex_enter(&db->db_mtx);
3731 if (db->db_blkid == DMU_SPILL_BLKID) {
3732 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3733 ASSERT(!(BP_IS_HOLE(bp)) &&
3734 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3738 if (db->db_level == 0) {
3739 mutex_enter(&dn->dn_mtx);
3740 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3741 db->db_blkid != DMU_SPILL_BLKID)
3742 dn->dn_phys->dn_maxblkid = db->db_blkid;
3743 mutex_exit(&dn->dn_mtx);
3745 if (dn->dn_type == DMU_OT_DNODE) {
3747 while (i < db->db.db_size) {
3749 (void *)(((char *)db->db.db_data) + i);
3751 i += DNODE_MIN_SIZE;
3752 if (dnp->dn_type != DMU_OT_NONE) {
3754 i += dnp->dn_extra_slots *
3759 if (BP_IS_HOLE(bp)) {
3766 blkptr_t *ibp = db->db.db_data;
3767 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3768 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3769 if (BP_IS_HOLE(ibp))
3771 fill += BP_GET_FILL(ibp);
3776 if (!BP_IS_EMBEDDED(bp))
3777 bp->blk_fill = fill;
3779 mutex_exit(&db->db_mtx);
3781 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3782 *db->db_blkptr = *bp;
3783 rw_exit(&dn->dn_struct_rwlock);
3788 * This function gets called just prior to running through the compression
3789 * stage of the zio pipeline. If we're an indirect block comprised of only
3790 * holes, then we want this indirect to be compressed away to a hole. In
3791 * order to do that we must zero out any information about the holes that
3792 * this indirect points to prior to before we try to compress it.
3795 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3797 dmu_buf_impl_t *db = vdb;
3800 unsigned int epbs, i;
3802 ASSERT3U(db->db_level, >, 0);
3805 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3806 ASSERT3U(epbs, <, 31);
3808 /* Determine if all our children are holes */
3809 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3810 if (!BP_IS_HOLE(bp))
3815 * If all the children are holes, then zero them all out so that
3816 * we may get compressed away.
3818 if (i == 1 << epbs) {
3820 * We only found holes. Grab the rwlock to prevent
3821 * anybody from reading the blocks we're about to
3824 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3825 bzero(db->db.db_data, db->db.db_size);
3826 rw_exit(&dn->dn_struct_rwlock);
3832 * The SPA will call this callback several times for each zio - once
3833 * for every physical child i/o (zio->io_phys_children times). This
3834 * allows the DMU to monitor the progress of each logical i/o. For example,
3835 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3836 * block. There may be a long delay before all copies/fragments are completed,
3837 * so this callback allows us to retire dirty space gradually, as the physical
3842 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3844 dmu_buf_impl_t *db = arg;
3845 objset_t *os = db->db_objset;
3846 dsl_pool_t *dp = dmu_objset_pool(os);
3847 dbuf_dirty_record_t *dr;
3850 dr = db->db_data_pending;
3851 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3854 * The callback will be called io_phys_children times. Retire one
3855 * portion of our dirty space each time we are called. Any rounding
3856 * error will be cleaned up by dsl_pool_sync()'s call to
3857 * dsl_pool_undirty_space().
3859 delta = dr->dr_accounted / zio->io_phys_children;
3860 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3865 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3867 dmu_buf_impl_t *db = vdb;
3868 blkptr_t *bp_orig = &zio->io_bp_orig;
3869 blkptr_t *bp = db->db_blkptr;
3870 objset_t *os = db->db_objset;
3871 dmu_tx_t *tx = os->os_synctx;
3872 dbuf_dirty_record_t **drp, *dr;
3874 ASSERT0(zio->io_error);
3875 ASSERT(db->db_blkptr == bp);
3878 * For nopwrites and rewrites we ensure that the bp matches our
3879 * original and bypass all the accounting.
3881 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3882 ASSERT(BP_EQUAL(bp, bp_orig));
3884 dsl_dataset_t *ds = os->os_dsl_dataset;
3885 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3886 dsl_dataset_block_born(ds, bp, tx);
3889 mutex_enter(&db->db_mtx);
3893 drp = &db->db_last_dirty;
3894 while ((dr = *drp) != db->db_data_pending)
3896 ASSERT(!list_link_active(&dr->dr_dirty_node));
3897 ASSERT(dr->dr_dbuf == db);
3898 ASSERT(dr->dr_next == NULL);
3902 if (db->db_blkid == DMU_SPILL_BLKID) {
3907 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3908 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3909 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3914 if (db->db_level == 0) {
3915 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3916 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3917 if (db->db_state != DB_NOFILL) {
3918 if (dr->dt.dl.dr_data != db->db_buf)
3919 arc_buf_destroy(dr->dt.dl.dr_data, db);
3926 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3927 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3928 if (!BP_IS_HOLE(db->db_blkptr)) {
3930 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3931 ASSERT3U(db->db_blkid, <=,
3932 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3933 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3937 mutex_destroy(&dr->dt.di.dr_mtx);
3938 list_destroy(&dr->dt.di.dr_children);
3940 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3942 cv_broadcast(&db->db_changed);
3943 ASSERT(db->db_dirtycnt > 0);
3944 db->db_dirtycnt -= 1;
3945 db->db_data_pending = NULL;
3946 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
3950 dbuf_write_nofill_ready(zio_t *zio)
3952 dbuf_write_ready(zio, NULL, zio->io_private);
3956 dbuf_write_nofill_done(zio_t *zio)
3958 dbuf_write_done(zio, NULL, zio->io_private);
3962 dbuf_write_override_ready(zio_t *zio)
3964 dbuf_dirty_record_t *dr = zio->io_private;
3965 dmu_buf_impl_t *db = dr->dr_dbuf;
3967 dbuf_write_ready(zio, NULL, db);
3971 dbuf_write_override_done(zio_t *zio)
3973 dbuf_dirty_record_t *dr = zio->io_private;
3974 dmu_buf_impl_t *db = dr->dr_dbuf;
3975 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3977 mutex_enter(&db->db_mtx);
3978 if (!BP_EQUAL(zio->io_bp, obp)) {
3979 if (!BP_IS_HOLE(obp))
3980 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3981 arc_release(dr->dt.dl.dr_data, db);
3983 mutex_exit(&db->db_mtx);
3984 dbuf_write_done(zio, NULL, db);
3986 if (zio->io_abd != NULL)
3987 abd_put(zio->io_abd);
3990 typedef struct dbuf_remap_impl_callback_arg {
3992 uint64_t drica_blk_birth;
3994 } dbuf_remap_impl_callback_arg_t;
3997 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
4000 dbuf_remap_impl_callback_arg_t *drica = arg;
4001 objset_t *os = drica->drica_os;
4002 spa_t *spa = dmu_objset_spa(os);
4003 dmu_tx_t *tx = drica->drica_tx;
4005 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4007 if (os == spa_meta_objset(spa)) {
4008 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
4010 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
4011 size, drica->drica_blk_birth, tx);
4016 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
4018 blkptr_t bp_copy = *bp;
4019 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4020 dbuf_remap_impl_callback_arg_t drica;
4022 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4024 drica.drica_os = dn->dn_objset;
4025 drica.drica_blk_birth = bp->blk_birth;
4026 drica.drica_tx = tx;
4027 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
4030 * The struct_rwlock prevents dbuf_read_impl() from
4031 * dereferencing the BP while we are changing it. To
4032 * avoid lock contention, only grab it when we are actually
4035 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
4037 rw_exit(&dn->dn_struct_rwlock);
4042 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
4043 * to remap a copy of every bp in the dbuf.
4046 dbuf_can_remap(const dmu_buf_impl_t *db)
4048 spa_t *spa = dmu_objset_spa(db->db_objset);
4049 blkptr_t *bp = db->db.db_data;
4050 boolean_t ret = B_FALSE;
4052 ASSERT3U(db->db_level, >, 0);
4053 ASSERT3S(db->db_state, ==, DB_CACHED);
4055 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4057 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4058 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4059 blkptr_t bp_copy = bp[i];
4060 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4065 spa_config_exit(spa, SCL_VDEV, FTAG);
4071 dnode_needs_remap(const dnode_t *dn)
4073 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4074 boolean_t ret = B_FALSE;
4076 if (dn->dn_phys->dn_nlevels == 0) {
4080 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4082 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4083 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
4084 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
4085 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4090 spa_config_exit(spa, SCL_VDEV, FTAG);
4096 * Remap any existing BP's to concrete vdevs, if possible.
4099 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
4101 spa_t *spa = dmu_objset_spa(db->db_objset);
4102 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4104 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
4107 if (db->db_level > 0) {
4108 blkptr_t *bp = db->db.db_data;
4109 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4110 dbuf_remap_impl(dn, &bp[i], tx);
4112 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
4113 dnode_phys_t *dnp = db->db.db_data;
4114 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
4116 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT;
4117 i += dnp[i].dn_extra_slots + 1) {
4118 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
4119 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
4126 /* Issue I/O to commit a dirty buffer to disk. */
4128 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4130 dmu_buf_impl_t *db = dr->dr_dbuf;
4133 dmu_buf_impl_t *parent = db->db_parent;
4134 uint64_t txg = tx->tx_txg;
4135 zbookmark_phys_t zb;
4140 ASSERT(dmu_tx_is_syncing(tx));
4146 if (db->db_state != DB_NOFILL) {
4147 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4149 * Private object buffers are released here rather
4150 * than in dbuf_dirty() since they are only modified
4151 * in the syncing context and we don't want the
4152 * overhead of making multiple copies of the data.
4154 if (BP_IS_HOLE(db->db_blkptr)) {
4157 dbuf_release_bp(db);
4159 dbuf_remap(dn, db, tx);
4163 if (parent != dn->dn_dbuf) {
4164 /* Our parent is an indirect block. */
4165 /* We have a dirty parent that has been scheduled for write. */
4166 ASSERT(parent && parent->db_data_pending);
4167 /* Our parent's buffer is one level closer to the dnode. */
4168 ASSERT(db->db_level == parent->db_level-1);
4170 * We're about to modify our parent's db_data by modifying
4171 * our block pointer, so the parent must be released.
4173 ASSERT(arc_released(parent->db_buf));
4174 zio = parent->db_data_pending->dr_zio;
4176 /* Our parent is the dnode itself. */
4177 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4178 db->db_blkid != DMU_SPILL_BLKID) ||
4179 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4180 if (db->db_blkid != DMU_SPILL_BLKID)
4181 ASSERT3P(db->db_blkptr, ==,
4182 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4186 ASSERT(db->db_level == 0 || data == db->db_buf);
4187 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4190 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4191 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4192 db->db.db_object, db->db_level, db->db_blkid);
4194 if (db->db_blkid == DMU_SPILL_BLKID)
4196 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
4198 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
4202 * We copy the blkptr now (rather than when we instantiate the dirty
4203 * record), because its value can change between open context and
4204 * syncing context. We do not need to hold dn_struct_rwlock to read
4205 * db_blkptr because we are in syncing context.
4207 dr->dr_bp_copy = *db->db_blkptr;
4209 if (db->db_level == 0 &&
4210 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
4212 * The BP for this block has been provided by open context
4213 * (by dmu_sync() or dmu_buf_write_embedded()).
4215 abd_t *contents = (data != NULL) ?
4216 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
4218 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
4219 contents, db->db.db_size, db->db.db_size, &zp,
4220 dbuf_write_override_ready, NULL, NULL,
4221 dbuf_write_override_done,
4222 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
4223 mutex_enter(&db->db_mtx);
4224 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
4225 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
4226 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
4227 mutex_exit(&db->db_mtx);
4228 } else if (db->db_state == DB_NOFILL) {
4229 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
4230 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
4231 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4232 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
4233 dbuf_write_nofill_ready, NULL, NULL,
4234 dbuf_write_nofill_done, db,
4235 ZIO_PRIORITY_ASYNC_WRITE,
4236 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
4238 ASSERT(arc_released(data));
4241 * For indirect blocks, we want to setup the children
4242 * ready callback so that we can properly handle an indirect
4243 * block that only contains holes.
4245 arc_write_done_func_t *children_ready_cb = NULL;
4246 if (db->db_level != 0)
4247 children_ready_cb = dbuf_write_children_ready;
4249 dr->dr_zio = arc_write(zio, os->os_spa, txg,
4250 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
4251 &zp, dbuf_write_ready, children_ready_cb,
4252 dbuf_write_physdone, dbuf_write_done, db,
4253 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);