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, 2020 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
27 * Copyright (c) 2019, Klara Inc.
28 * Copyright (c) 2019, Allan Jude
31 #include <sys/zfs_context.h>
34 #include <sys/dmu_send.h>
35 #include <sys/dmu_impl.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dmu_tx.h>
43 #include <sys/dmu_zfetch.h>
45 #include <sys/sa_impl.h>
46 #include <sys/zfeature.h>
47 #include <sys/blkptr.h>
48 #include <sys/range_tree.h>
49 #include <sys/trace_zfs.h>
50 #include <sys/callb.h>
54 #include <sys/spa_impl.h>
55 #include <sys/wmsum.h>
56 #include <sys/vdev_impl.h>
58 static kstat_t *dbuf_ksp;
60 typedef struct dbuf_stats {
62 * Various statistics about the size of the dbuf cache.
64 kstat_named_t cache_count;
65 kstat_named_t cache_size_bytes;
66 kstat_named_t cache_size_bytes_max;
68 * Statistics regarding the bounds on the dbuf cache size.
70 kstat_named_t cache_target_bytes;
71 kstat_named_t cache_lowater_bytes;
72 kstat_named_t cache_hiwater_bytes;
74 * Total number of dbuf cache evictions that have occurred.
76 kstat_named_t cache_total_evicts;
78 * The distribution of dbuf levels in the dbuf cache and
79 * the total size of all dbufs at each level.
81 kstat_named_t cache_levels[DN_MAX_LEVELS];
82 kstat_named_t cache_levels_bytes[DN_MAX_LEVELS];
84 * Statistics about the dbuf hash table.
86 kstat_named_t hash_hits;
87 kstat_named_t hash_misses;
88 kstat_named_t hash_collisions;
89 kstat_named_t hash_elements;
90 kstat_named_t hash_elements_max;
92 * Number of sublists containing more than one dbuf in the dbuf
93 * hash table. Keep track of the longest hash chain.
95 kstat_named_t hash_chains;
96 kstat_named_t hash_chain_max;
98 * Number of times a dbuf_create() discovers that a dbuf was
99 * already created and in the dbuf hash table.
101 kstat_named_t hash_insert_race;
103 * Statistics about the size of the metadata dbuf cache.
105 kstat_named_t metadata_cache_count;
106 kstat_named_t metadata_cache_size_bytes;
107 kstat_named_t metadata_cache_size_bytes_max;
109 * For diagnostic purposes, this is incremented whenever we can't add
110 * something to the metadata cache because it's full, and instead put
111 * the data in the regular dbuf cache.
113 kstat_named_t metadata_cache_overflow;
116 dbuf_stats_t dbuf_stats = {
117 { "cache_count", KSTAT_DATA_UINT64 },
118 { "cache_size_bytes", KSTAT_DATA_UINT64 },
119 { "cache_size_bytes_max", KSTAT_DATA_UINT64 },
120 { "cache_target_bytes", KSTAT_DATA_UINT64 },
121 { "cache_lowater_bytes", KSTAT_DATA_UINT64 },
122 { "cache_hiwater_bytes", KSTAT_DATA_UINT64 },
123 { "cache_total_evicts", KSTAT_DATA_UINT64 },
124 { { "cache_levels_N", KSTAT_DATA_UINT64 } },
125 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } },
126 { "hash_hits", KSTAT_DATA_UINT64 },
127 { "hash_misses", KSTAT_DATA_UINT64 },
128 { "hash_collisions", KSTAT_DATA_UINT64 },
129 { "hash_elements", KSTAT_DATA_UINT64 },
130 { "hash_elements_max", KSTAT_DATA_UINT64 },
131 { "hash_chains", KSTAT_DATA_UINT64 },
132 { "hash_chain_max", KSTAT_DATA_UINT64 },
133 { "hash_insert_race", KSTAT_DATA_UINT64 },
134 { "metadata_cache_count", KSTAT_DATA_UINT64 },
135 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64 },
136 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64 },
137 { "metadata_cache_overflow", KSTAT_DATA_UINT64 }
142 wmsum_t cache_total_evicts;
143 wmsum_t cache_levels[DN_MAX_LEVELS];
144 wmsum_t cache_levels_bytes[DN_MAX_LEVELS];
147 wmsum_t hash_collisions;
149 wmsum_t hash_insert_race;
150 wmsum_t metadata_cache_count;
151 wmsum_t metadata_cache_overflow;
154 #define DBUF_STAT_INCR(stat, val) \
155 wmsum_add(&dbuf_sums.stat, val);
156 #define DBUF_STAT_DECR(stat, val) \
157 DBUF_STAT_INCR(stat, -(val));
158 #define DBUF_STAT_BUMP(stat) \
159 DBUF_STAT_INCR(stat, 1);
160 #define DBUF_STAT_BUMPDOWN(stat) \
161 DBUF_STAT_INCR(stat, -1);
162 #define DBUF_STAT_MAX(stat, v) { \
164 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
165 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
169 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
170 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
171 static void dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t *dr);
172 static int dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags);
174 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
175 dmu_buf_evict_func_t *evict_func_sync,
176 dmu_buf_evict_func_t *evict_func_async,
177 dmu_buf_t **clear_on_evict_dbufp);
180 * Global data structures and functions for the dbuf cache.
182 static kmem_cache_t *dbuf_kmem_cache;
183 static taskq_t *dbu_evict_taskq;
185 static kthread_t *dbuf_cache_evict_thread;
186 static kmutex_t dbuf_evict_lock;
187 static kcondvar_t dbuf_evict_cv;
188 static boolean_t dbuf_evict_thread_exit;
191 * There are two dbuf caches; each dbuf can only be in one of them at a time.
193 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
194 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
195 * that represent the metadata that describes filesystems/snapshots/
196 * bookmarks/properties/etc. We only evict from this cache when we export a
197 * pool, to short-circuit as much I/O as possible for all administrative
198 * commands that need the metadata. There is no eviction policy for this
199 * cache, because we try to only include types in it which would occupy a
200 * very small amount of space per object but create a large impact on the
201 * performance of these commands. Instead, after it reaches a maximum size
202 * (which should only happen on very small memory systems with a very large
203 * number of filesystem objects), we stop taking new dbufs into the
204 * metadata cache, instead putting them in the normal dbuf cache.
206 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
207 * are not currently held but have been recently released. These dbufs
208 * are not eligible for arc eviction until they are aged out of the cache.
209 * Dbufs that are aged out of the cache will be immediately destroyed and
210 * become eligible for arc eviction.
212 * Dbufs are added to these caches once the last hold is released. If a dbuf is
213 * later accessed and still exists in the dbuf cache, then it will be removed
214 * from the cache and later re-added to the head of the cache.
216 * If a given dbuf meets the requirements for the metadata cache, it will go
217 * there, otherwise it will be considered for the generic LRU dbuf cache. The
218 * caches and the refcounts tracking their sizes are stored in an array indexed
219 * by those caches' matching enum values (from dbuf_cached_state_t).
221 typedef struct dbuf_cache {
223 zfs_refcount_t size ____cacheline_aligned;
225 dbuf_cache_t dbuf_caches[DB_CACHE_MAX];
227 /* Size limits for the caches */
228 static unsigned long dbuf_cache_max_bytes = ULONG_MAX;
229 static unsigned long dbuf_metadata_cache_max_bytes = ULONG_MAX;
231 /* Set the default sizes of the caches to log2 fraction of arc size */
232 static int dbuf_cache_shift = 5;
233 static int dbuf_metadata_cache_shift = 6;
235 static unsigned long dbuf_cache_target_bytes(void);
236 static unsigned long dbuf_metadata_cache_target_bytes(void);
239 * The LRU dbuf cache uses a three-stage eviction policy:
240 * - A low water marker designates when the dbuf eviction thread
241 * should stop evicting from the dbuf cache.
242 * - When we reach the maximum size (aka mid water mark), we
243 * signal the eviction thread to run.
244 * - The high water mark indicates when the eviction thread
245 * is unable to keep up with the incoming load and eviction must
246 * happen in the context of the calling thread.
250 * low water mid water hi water
251 * +----------------------------------------+----------+----------+
256 * +----------------------------------------+----------+----------+
258 * evicting eviction directly
261 * The high and low water marks indicate the operating range for the eviction
262 * thread. The low water mark is, by default, 90% of the total size of the
263 * cache and the high water mark is at 110% (both of these percentages can be
264 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
265 * respectively). The eviction thread will try to ensure that the cache remains
266 * within this range by waking up every second and checking if the cache is
267 * above the low water mark. The thread can also be woken up by callers adding
268 * elements into the cache if the cache is larger than the mid water (i.e max
269 * cache size). Once the eviction thread is woken up and eviction is required,
270 * it will continue evicting buffers until it's able to reduce the cache size
271 * to the low water mark. If the cache size continues to grow and hits the high
272 * water mark, then callers adding elements to the cache will begin to evict
273 * directly from the cache until the cache is no longer above the high water
278 * The percentage above and below the maximum cache size.
280 static uint_t dbuf_cache_hiwater_pct = 10;
281 static uint_t dbuf_cache_lowater_pct = 10;
284 dbuf_cons(void *vdb, void *unused, int kmflag)
286 (void) unused, (void) kmflag;
287 dmu_buf_impl_t *db = vdb;
288 bzero(db, sizeof (dmu_buf_impl_t));
290 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
291 rw_init(&db->db_rwlock, NULL, RW_DEFAULT, NULL);
292 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
293 multilist_link_init(&db->db_cache_link);
294 zfs_refcount_create(&db->db_holds);
300 dbuf_dest(void *vdb, void *unused)
303 dmu_buf_impl_t *db = vdb;
304 mutex_destroy(&db->db_mtx);
305 rw_destroy(&db->db_rwlock);
306 cv_destroy(&db->db_changed);
307 ASSERT(!multilist_link_active(&db->db_cache_link));
308 zfs_refcount_destroy(&db->db_holds);
312 * dbuf hash table routines
314 static dbuf_hash_table_t dbuf_hash_table;
317 * We use Cityhash for this. It's fast, and has good hash properties without
318 * requiring any large static buffers.
321 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
323 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid));
326 #define DTRACE_SET_STATE(db, why) \
327 DTRACE_PROBE2(dbuf__state_change, dmu_buf_impl_t *, db, \
330 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
331 ((dbuf)->db.db_object == (obj) && \
332 (dbuf)->db_objset == (os) && \
333 (dbuf)->db_level == (level) && \
334 (dbuf)->db_blkid == (blkid))
337 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
339 dbuf_hash_table_t *h = &dbuf_hash_table;
344 hv = dbuf_hash(os, obj, level, blkid);
345 idx = hv & h->hash_table_mask;
347 mutex_enter(DBUF_HASH_MUTEX(h, idx));
348 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
349 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
350 mutex_enter(&db->db_mtx);
351 if (db->db_state != DB_EVICTING) {
352 mutex_exit(DBUF_HASH_MUTEX(h, idx));
355 mutex_exit(&db->db_mtx);
358 mutex_exit(DBUF_HASH_MUTEX(h, idx));
362 static dmu_buf_impl_t *
363 dbuf_find_bonus(objset_t *os, uint64_t object)
366 dmu_buf_impl_t *db = NULL;
368 if (dnode_hold(os, object, FTAG, &dn) == 0) {
369 rw_enter(&dn->dn_struct_rwlock, RW_READER);
370 if (dn->dn_bonus != NULL) {
372 mutex_enter(&db->db_mtx);
374 rw_exit(&dn->dn_struct_rwlock);
375 dnode_rele(dn, FTAG);
381 * Insert an entry into the hash table. If there is already an element
382 * equal to elem in the hash table, then the already existing element
383 * will be returned and the new element will not be inserted.
384 * Otherwise returns NULL.
386 static dmu_buf_impl_t *
387 dbuf_hash_insert(dmu_buf_impl_t *db)
389 dbuf_hash_table_t *h = &dbuf_hash_table;
390 objset_t *os = db->db_objset;
391 uint64_t obj = db->db.db_object;
392 int level = db->db_level;
393 uint64_t blkid, hv, idx;
397 blkid = db->db_blkid;
398 hv = dbuf_hash(os, obj, level, blkid);
399 idx = hv & h->hash_table_mask;
401 mutex_enter(DBUF_HASH_MUTEX(h, idx));
402 for (dbf = h->hash_table[idx], i = 0; dbf != NULL;
403 dbf = dbf->db_hash_next, i++) {
404 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
405 mutex_enter(&dbf->db_mtx);
406 if (dbf->db_state != DB_EVICTING) {
407 mutex_exit(DBUF_HASH_MUTEX(h, idx));
410 mutex_exit(&dbf->db_mtx);
415 DBUF_STAT_BUMP(hash_collisions);
417 DBUF_STAT_BUMP(hash_chains);
419 DBUF_STAT_MAX(hash_chain_max, i);
422 mutex_enter(&db->db_mtx);
423 db->db_hash_next = h->hash_table[idx];
424 h->hash_table[idx] = db;
425 mutex_exit(DBUF_HASH_MUTEX(h, idx));
426 uint64_t he = atomic_inc_64_nv(&dbuf_stats.hash_elements.value.ui64);
427 DBUF_STAT_MAX(hash_elements_max, he);
433 * This returns whether this dbuf should be stored in the metadata cache, which
434 * is based on whether it's from one of the dnode types that store data related
435 * to traversing dataset hierarchies.
438 dbuf_include_in_metadata_cache(dmu_buf_impl_t *db)
441 dmu_object_type_t type = DB_DNODE(db)->dn_type;
444 /* Check if this dbuf is one of the types we care about */
445 if (DMU_OT_IS_METADATA_CACHED(type)) {
446 /* If we hit this, then we set something up wrong in dmu_ot */
447 ASSERT(DMU_OT_IS_METADATA(type));
450 * Sanity check for small-memory systems: don't allocate too
451 * much memory for this purpose.
453 if (zfs_refcount_count(
454 &dbuf_caches[DB_DBUF_METADATA_CACHE].size) >
455 dbuf_metadata_cache_target_bytes()) {
456 DBUF_STAT_BUMP(metadata_cache_overflow);
467 * Remove an entry from the hash table. It must be in the EVICTING state.
470 dbuf_hash_remove(dmu_buf_impl_t *db)
472 dbuf_hash_table_t *h = &dbuf_hash_table;
474 dmu_buf_impl_t *dbf, **dbp;
476 hv = dbuf_hash(db->db_objset, db->db.db_object,
477 db->db_level, db->db_blkid);
478 idx = hv & h->hash_table_mask;
481 * We mustn't hold db_mtx to maintain lock ordering:
482 * DBUF_HASH_MUTEX > db_mtx.
484 ASSERT(zfs_refcount_is_zero(&db->db_holds));
485 ASSERT(db->db_state == DB_EVICTING);
486 ASSERT(!MUTEX_HELD(&db->db_mtx));
488 mutex_enter(DBUF_HASH_MUTEX(h, idx));
489 dbp = &h->hash_table[idx];
490 while ((dbf = *dbp) != db) {
491 dbp = &dbf->db_hash_next;
494 *dbp = db->db_hash_next;
495 db->db_hash_next = NULL;
496 if (h->hash_table[idx] &&
497 h->hash_table[idx]->db_hash_next == NULL)
498 DBUF_STAT_BUMPDOWN(hash_chains);
499 mutex_exit(DBUF_HASH_MUTEX(h, idx));
500 atomic_dec_64(&dbuf_stats.hash_elements.value.ui64);
506 } dbvu_verify_type_t;
509 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
514 if (db->db_user == NULL)
517 /* Only data blocks support the attachment of user data. */
518 ASSERT(db->db_level == 0);
520 /* Clients must resolve a dbuf before attaching user data. */
521 ASSERT(db->db.db_data != NULL);
522 ASSERT3U(db->db_state, ==, DB_CACHED);
524 holds = zfs_refcount_count(&db->db_holds);
525 if (verify_type == DBVU_EVICTING) {
527 * Immediate eviction occurs when holds == dirtycnt.
528 * For normal eviction buffers, holds is zero on
529 * eviction, except when dbuf_fix_old_data() calls
530 * dbuf_clear_data(). However, the hold count can grow
531 * during eviction even though db_mtx is held (see
532 * dmu_bonus_hold() for an example), so we can only
533 * test the generic invariant that holds >= dirtycnt.
535 ASSERT3U(holds, >=, db->db_dirtycnt);
537 if (db->db_user_immediate_evict == TRUE)
538 ASSERT3U(holds, >=, db->db_dirtycnt);
540 ASSERT3U(holds, >, 0);
546 dbuf_evict_user(dmu_buf_impl_t *db)
548 dmu_buf_user_t *dbu = db->db_user;
550 ASSERT(MUTEX_HELD(&db->db_mtx));
555 dbuf_verify_user(db, DBVU_EVICTING);
559 if (dbu->dbu_clear_on_evict_dbufp != NULL)
560 *dbu->dbu_clear_on_evict_dbufp = NULL;
564 * There are two eviction callbacks - one that we call synchronously
565 * and one that we invoke via a taskq. The async one is useful for
566 * avoiding lock order reversals and limiting stack depth.
568 * Note that if we have a sync callback but no async callback,
569 * it's likely that the sync callback will free the structure
570 * containing the dbu. In that case we need to take care to not
571 * dereference dbu after calling the sync evict func.
573 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
575 if (dbu->dbu_evict_func_sync != NULL)
576 dbu->dbu_evict_func_sync(dbu);
579 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
580 dbu, 0, &dbu->dbu_tqent);
585 dbuf_is_metadata(dmu_buf_impl_t *db)
588 * Consider indirect blocks and spill blocks to be meta data.
590 if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) {
593 boolean_t is_metadata;
596 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
599 return (is_metadata);
604 * We want to exclude buffers that are on a special allocation class from
608 dbuf_is_l2cacheable(dmu_buf_impl_t *db)
611 zfs_cache_type_t cache = db->db_objset->os_secondary_cache;
612 blkptr_t *bp = db->db_blkptr;
614 if (bp != NULL && !BP_IS_HOLE(bp)) {
615 uint64_t vdev = DVA_GET_VDEV(bp->blk_dva);
616 vdev_t *rvd = db->db_objset->os_spa->spa_root_vdev;
618 if (vdev < rvd->vdev_children)
619 vd = rvd->vdev_child[vdev];
621 if (cache == ZFS_CACHE_ALL ||
622 (dbuf_is_metadata(db) && cache == ZFS_CACHE_METADATA)) {
626 if ((vd->vdev_alloc_bias != VDEV_BIAS_SPECIAL &&
627 vd->vdev_alloc_bias != VDEV_BIAS_DEDUP) ||
628 l2arc_exclude_special == 0)
636 static inline boolean_t
637 dnode_level_is_l2cacheable(blkptr_t *bp, dnode_t *dn, int64_t level)
640 zfs_cache_type_t cache = dn->dn_objset->os_secondary_cache;
642 if (bp != NULL && !BP_IS_HOLE(bp)) {
643 uint64_t vdev = DVA_GET_VDEV(bp->blk_dva);
644 vdev_t *rvd = dn->dn_objset->os_spa->spa_root_vdev;
646 if (vdev < rvd->vdev_children)
647 vd = rvd->vdev_child[vdev];
649 if (cache == ZFS_CACHE_ALL || ((level > 0 ||
650 DMU_OT_IS_METADATA(dn->dn_handle->dnh_dnode->dn_type)) &&
651 cache == ZFS_CACHE_METADATA)) {
655 if ((vd->vdev_alloc_bias != VDEV_BIAS_SPECIAL &&
656 vd->vdev_alloc_bias != VDEV_BIAS_DEDUP) ||
657 l2arc_exclude_special == 0)
667 * This function *must* return indices evenly distributed between all
668 * sublists of the multilist. This is needed due to how the dbuf eviction
669 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
670 * distributed between all sublists and uses this assumption when
671 * deciding which sublist to evict from and how much to evict from it.
674 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
676 dmu_buf_impl_t *db = obj;
679 * The assumption here, is the hash value for a given
680 * dmu_buf_impl_t will remain constant throughout it's lifetime
681 * (i.e. it's objset, object, level and blkid fields don't change).
682 * Thus, we don't need to store the dbuf's sublist index
683 * on insertion, as this index can be recalculated on removal.
685 * Also, the low order bits of the hash value are thought to be
686 * distributed evenly. Otherwise, in the case that the multilist
687 * has a power of two number of sublists, each sublists' usage
688 * would not be evenly distributed. In this context full 64bit
689 * division would be a waste of time, so limit it to 32 bits.
691 return ((unsigned int)dbuf_hash(db->db_objset, db->db.db_object,
692 db->db_level, db->db_blkid) %
693 multilist_get_num_sublists(ml));
697 * The target size of the dbuf cache can grow with the ARC target,
698 * unless limited by the tunable dbuf_cache_max_bytes.
700 static inline unsigned long
701 dbuf_cache_target_bytes(void)
703 return (MIN(dbuf_cache_max_bytes,
704 arc_target_bytes() >> dbuf_cache_shift));
708 * The target size of the dbuf metadata cache can grow with the ARC target,
709 * unless limited by the tunable dbuf_metadata_cache_max_bytes.
711 static inline unsigned long
712 dbuf_metadata_cache_target_bytes(void)
714 return (MIN(dbuf_metadata_cache_max_bytes,
715 arc_target_bytes() >> dbuf_metadata_cache_shift));
718 static inline uint64_t
719 dbuf_cache_hiwater_bytes(void)
721 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
722 return (dbuf_cache_target +
723 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100);
726 static inline uint64_t
727 dbuf_cache_lowater_bytes(void)
729 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
730 return (dbuf_cache_target -
731 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100);
734 static inline boolean_t
735 dbuf_cache_above_lowater(void)
737 return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
738 dbuf_cache_lowater_bytes());
742 * Evict the oldest eligible dbuf from the dbuf cache.
747 int idx = multilist_get_random_index(&dbuf_caches[DB_DBUF_CACHE].cache);
748 multilist_sublist_t *mls = multilist_sublist_lock(
749 &dbuf_caches[DB_DBUF_CACHE].cache, idx);
751 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
753 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
754 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
755 db = multilist_sublist_prev(mls, db);
758 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
759 multilist_sublist_t *, mls);
762 multilist_sublist_remove(mls, db);
763 multilist_sublist_unlock(mls);
764 (void) zfs_refcount_remove_many(
765 &dbuf_caches[DB_DBUF_CACHE].size, db->db.db_size, db);
766 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
767 DBUF_STAT_BUMPDOWN(cache_count);
768 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
770 ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE);
771 db->db_caching_status = DB_NO_CACHE;
773 DBUF_STAT_BUMP(cache_total_evicts);
775 multilist_sublist_unlock(mls);
780 * The dbuf evict thread is responsible for aging out dbufs from the
781 * cache. Once the cache has reached it's maximum size, dbufs are removed
782 * and destroyed. The eviction thread will continue running until the size
783 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
784 * out of the cache it is destroyed and becomes eligible for arc eviction.
787 dbuf_evict_thread(void *unused)
792 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
794 mutex_enter(&dbuf_evict_lock);
795 while (!dbuf_evict_thread_exit) {
796 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
797 CALLB_CPR_SAFE_BEGIN(&cpr);
798 (void) cv_timedwait_idle_hires(&dbuf_evict_cv,
799 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
800 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
802 mutex_exit(&dbuf_evict_lock);
805 * Keep evicting as long as we're above the low water mark
806 * for the cache. We do this without holding the locks to
807 * minimize lock contention.
809 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
813 mutex_enter(&dbuf_evict_lock);
816 dbuf_evict_thread_exit = B_FALSE;
817 cv_broadcast(&dbuf_evict_cv);
818 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
823 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
824 * If the dbuf cache is at its high water mark, then evict a dbuf from the
825 * dbuf cache using the callers context.
828 dbuf_evict_notify(uint64_t size)
831 * We check if we should evict without holding the dbuf_evict_lock,
832 * because it's OK to occasionally make the wrong decision here,
833 * and grabbing the lock results in massive lock contention.
835 if (size > dbuf_cache_target_bytes()) {
836 if (size > dbuf_cache_hiwater_bytes())
838 cv_signal(&dbuf_evict_cv);
843 dbuf_kstat_update(kstat_t *ksp, int rw)
845 dbuf_stats_t *ds = ksp->ks_data;
847 if (rw == KSTAT_WRITE)
848 return (SET_ERROR(EACCES));
850 ds->cache_count.value.ui64 =
851 wmsum_value(&dbuf_sums.cache_count);
852 ds->cache_size_bytes.value.ui64 =
853 zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size);
854 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
855 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
856 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
857 ds->cache_total_evicts.value.ui64 =
858 wmsum_value(&dbuf_sums.cache_total_evicts);
859 for (int i = 0; i < DN_MAX_LEVELS; i++) {
860 ds->cache_levels[i].value.ui64 =
861 wmsum_value(&dbuf_sums.cache_levels[i]);
862 ds->cache_levels_bytes[i].value.ui64 =
863 wmsum_value(&dbuf_sums.cache_levels_bytes[i]);
865 ds->hash_hits.value.ui64 =
866 wmsum_value(&dbuf_sums.hash_hits);
867 ds->hash_misses.value.ui64 =
868 wmsum_value(&dbuf_sums.hash_misses);
869 ds->hash_collisions.value.ui64 =
870 wmsum_value(&dbuf_sums.hash_collisions);
871 ds->hash_chains.value.ui64 =
872 wmsum_value(&dbuf_sums.hash_chains);
873 ds->hash_insert_race.value.ui64 =
874 wmsum_value(&dbuf_sums.hash_insert_race);
875 ds->metadata_cache_count.value.ui64 =
876 wmsum_value(&dbuf_sums.metadata_cache_count);
877 ds->metadata_cache_size_bytes.value.ui64 = zfs_refcount_count(
878 &dbuf_caches[DB_DBUF_METADATA_CACHE].size);
879 ds->metadata_cache_overflow.value.ui64 =
880 wmsum_value(&dbuf_sums.metadata_cache_overflow);
887 uint64_t hsize = 1ULL << 16;
888 dbuf_hash_table_t *h = &dbuf_hash_table;
892 * The hash table is big enough to fill one eighth of physical memory
893 * with an average block size of zfs_arc_average_blocksize (default 8K).
894 * By default, the table will take up
895 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
897 while (hsize * zfs_arc_average_blocksize < arc_all_memory() / 8)
901 h->hash_table_mask = hsize - 1;
904 * Large allocations which do not require contiguous pages
905 * should be using vmem_alloc() in the linux kernel
907 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP);
909 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
911 if (h->hash_table == NULL) {
912 /* XXX - we should really return an error instead of assert */
913 ASSERT(hsize > (1ULL << 10));
918 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
919 sizeof (dmu_buf_impl_t),
920 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
922 for (i = 0; i < DBUF_MUTEXES; i++)
923 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
928 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
929 * configuration is not required.
931 dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
933 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
934 multilist_create(&dbuf_caches[dcs].cache,
935 sizeof (dmu_buf_impl_t),
936 offsetof(dmu_buf_impl_t, db_cache_link),
937 dbuf_cache_multilist_index_func);
938 zfs_refcount_create(&dbuf_caches[dcs].size);
941 dbuf_evict_thread_exit = B_FALSE;
942 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
943 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
944 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
945 NULL, 0, &p0, TS_RUN, minclsyspri);
947 wmsum_init(&dbuf_sums.cache_count, 0);
948 wmsum_init(&dbuf_sums.cache_total_evicts, 0);
949 for (i = 0; i < DN_MAX_LEVELS; i++) {
950 wmsum_init(&dbuf_sums.cache_levels[i], 0);
951 wmsum_init(&dbuf_sums.cache_levels_bytes[i], 0);
953 wmsum_init(&dbuf_sums.hash_hits, 0);
954 wmsum_init(&dbuf_sums.hash_misses, 0);
955 wmsum_init(&dbuf_sums.hash_collisions, 0);
956 wmsum_init(&dbuf_sums.hash_chains, 0);
957 wmsum_init(&dbuf_sums.hash_insert_race, 0);
958 wmsum_init(&dbuf_sums.metadata_cache_count, 0);
959 wmsum_init(&dbuf_sums.metadata_cache_overflow, 0);
961 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
962 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
964 if (dbuf_ksp != NULL) {
965 for (i = 0; i < DN_MAX_LEVELS; i++) {
966 snprintf(dbuf_stats.cache_levels[i].name,
967 KSTAT_STRLEN, "cache_level_%d", i);
968 dbuf_stats.cache_levels[i].data_type =
970 snprintf(dbuf_stats.cache_levels_bytes[i].name,
971 KSTAT_STRLEN, "cache_level_%d_bytes", i);
972 dbuf_stats.cache_levels_bytes[i].data_type =
975 dbuf_ksp->ks_data = &dbuf_stats;
976 dbuf_ksp->ks_update = dbuf_kstat_update;
977 kstat_install(dbuf_ksp);
984 dbuf_hash_table_t *h = &dbuf_hash_table;
987 dbuf_stats_destroy();
989 for (i = 0; i < DBUF_MUTEXES; i++)
990 mutex_destroy(&h->hash_mutexes[i]);
993 * Large allocations which do not require contiguous pages
994 * should be using vmem_free() in the linux kernel
996 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
998 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
1000 kmem_cache_destroy(dbuf_kmem_cache);
1001 taskq_destroy(dbu_evict_taskq);
1003 mutex_enter(&dbuf_evict_lock);
1004 dbuf_evict_thread_exit = B_TRUE;
1005 while (dbuf_evict_thread_exit) {
1006 cv_signal(&dbuf_evict_cv);
1007 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
1009 mutex_exit(&dbuf_evict_lock);
1011 mutex_destroy(&dbuf_evict_lock);
1012 cv_destroy(&dbuf_evict_cv);
1014 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
1015 zfs_refcount_destroy(&dbuf_caches[dcs].size);
1016 multilist_destroy(&dbuf_caches[dcs].cache);
1019 if (dbuf_ksp != NULL) {
1020 kstat_delete(dbuf_ksp);
1024 wmsum_fini(&dbuf_sums.cache_count);
1025 wmsum_fini(&dbuf_sums.cache_total_evicts);
1026 for (i = 0; i < DN_MAX_LEVELS; i++) {
1027 wmsum_fini(&dbuf_sums.cache_levels[i]);
1028 wmsum_fini(&dbuf_sums.cache_levels_bytes[i]);
1030 wmsum_fini(&dbuf_sums.hash_hits);
1031 wmsum_fini(&dbuf_sums.hash_misses);
1032 wmsum_fini(&dbuf_sums.hash_collisions);
1033 wmsum_fini(&dbuf_sums.hash_chains);
1034 wmsum_fini(&dbuf_sums.hash_insert_race);
1035 wmsum_fini(&dbuf_sums.metadata_cache_count);
1036 wmsum_fini(&dbuf_sums.metadata_cache_overflow);
1045 dbuf_verify(dmu_buf_impl_t *db)
1048 dbuf_dirty_record_t *dr;
1051 ASSERT(MUTEX_HELD(&db->db_mtx));
1053 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
1056 ASSERT(db->db_objset != NULL);
1060 ASSERT(db->db_parent == NULL);
1061 ASSERT(db->db_blkptr == NULL);
1063 ASSERT3U(db->db.db_object, ==, dn->dn_object);
1064 ASSERT3P(db->db_objset, ==, dn->dn_objset);
1065 ASSERT3U(db->db_level, <, dn->dn_nlevels);
1066 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
1067 db->db_blkid == DMU_SPILL_BLKID ||
1068 !avl_is_empty(&dn->dn_dbufs));
1070 if (db->db_blkid == DMU_BONUS_BLKID) {
1072 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1073 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
1074 } else if (db->db_blkid == DMU_SPILL_BLKID) {
1076 ASSERT0(db->db.db_offset);
1078 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
1081 if ((dr = list_head(&db->db_dirty_records)) != NULL) {
1082 ASSERT(dr->dr_dbuf == db);
1083 txg_prev = dr->dr_txg;
1084 for (dr = list_next(&db->db_dirty_records, dr); dr != NULL;
1085 dr = list_next(&db->db_dirty_records, dr)) {
1086 ASSERT(dr->dr_dbuf == db);
1087 ASSERT(txg_prev > dr->dr_txg);
1088 txg_prev = dr->dr_txg;
1093 * We can't assert that db_size matches dn_datablksz because it
1094 * can be momentarily different when another thread is doing
1095 * dnode_set_blksz().
1097 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
1098 dr = db->db_data_pending;
1100 * It should only be modified in syncing context, so
1101 * make sure we only have one copy of the data.
1103 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
1106 /* verify db->db_blkptr */
1107 if (db->db_blkptr) {
1108 if (db->db_parent == dn->dn_dbuf) {
1109 /* db is pointed to by the dnode */
1110 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1111 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
1112 ASSERT(db->db_parent == NULL);
1114 ASSERT(db->db_parent != NULL);
1115 if (db->db_blkid != DMU_SPILL_BLKID)
1116 ASSERT3P(db->db_blkptr, ==,
1117 &dn->dn_phys->dn_blkptr[db->db_blkid]);
1119 /* db is pointed to by an indirect block */
1120 int epb __maybe_unused = db->db_parent->db.db_size >>
1122 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
1123 ASSERT3U(db->db_parent->db.db_object, ==,
1126 * dnode_grow_indblksz() can make this fail if we don't
1127 * have the parent's rwlock. XXX indblksz no longer
1128 * grows. safe to do this now?
1130 if (RW_LOCK_HELD(&db->db_parent->db_rwlock)) {
1131 ASSERT3P(db->db_blkptr, ==,
1132 ((blkptr_t *)db->db_parent->db.db_data +
1133 db->db_blkid % epb));
1137 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
1138 (db->db_buf == NULL || db->db_buf->b_data) &&
1139 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
1140 db->db_state != DB_FILL && !dn->dn_free_txg) {
1142 * If the blkptr isn't set but they have nonzero data,
1143 * it had better be dirty, otherwise we'll lose that
1144 * data when we evict this buffer.
1146 * There is an exception to this rule for indirect blocks; in
1147 * this case, if the indirect block is a hole, we fill in a few
1148 * fields on each of the child blocks (importantly, birth time)
1149 * to prevent hole birth times from being lost when you
1150 * partially fill in a hole.
1152 if (db->db_dirtycnt == 0) {
1153 if (db->db_level == 0) {
1154 uint64_t *buf = db->db.db_data;
1157 for (i = 0; i < db->db.db_size >> 3; i++) {
1158 ASSERT(buf[i] == 0);
1161 blkptr_t *bps = db->db.db_data;
1162 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
1165 * We want to verify that all the blkptrs in the
1166 * indirect block are holes, but we may have
1167 * automatically set up a few fields for them.
1168 * We iterate through each blkptr and verify
1169 * they only have those fields set.
1172 i < db->db.db_size / sizeof (blkptr_t);
1174 blkptr_t *bp = &bps[i];
1175 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1178 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
1179 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
1180 DVA_IS_EMPTY(&bp->blk_dva[2]));
1181 ASSERT0(bp->blk_fill);
1182 ASSERT0(bp->blk_pad[0]);
1183 ASSERT0(bp->blk_pad[1]);
1184 ASSERT(!BP_IS_EMBEDDED(bp));
1185 ASSERT(BP_IS_HOLE(bp));
1186 ASSERT0(bp->blk_phys_birth);
1196 dbuf_clear_data(dmu_buf_impl_t *db)
1198 ASSERT(MUTEX_HELD(&db->db_mtx));
1199 dbuf_evict_user(db);
1200 ASSERT3P(db->db_buf, ==, NULL);
1201 db->db.db_data = NULL;
1202 if (db->db_state != DB_NOFILL) {
1203 db->db_state = DB_UNCACHED;
1204 DTRACE_SET_STATE(db, "clear data");
1209 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1211 ASSERT(MUTEX_HELD(&db->db_mtx));
1212 ASSERT(buf != NULL);
1215 ASSERT(buf->b_data != NULL);
1216 db->db.db_data = buf->b_data;
1220 dbuf_alloc_arcbuf(dmu_buf_impl_t *db)
1222 spa_t *spa = db->db_objset->os_spa;
1224 return (arc_alloc_buf(spa, db, DBUF_GET_BUFC_TYPE(db), db->db.db_size));
1228 * Loan out an arc_buf for read. Return the loaned arc_buf.
1231 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1235 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1236 mutex_enter(&db->db_mtx);
1237 if (arc_released(db->db_buf) || zfs_refcount_count(&db->db_holds) > 1) {
1238 int blksz = db->db.db_size;
1239 spa_t *spa = db->db_objset->os_spa;
1241 mutex_exit(&db->db_mtx);
1242 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1243 bcopy(db->db.db_data, abuf->b_data, blksz);
1246 arc_loan_inuse_buf(abuf, db);
1248 dbuf_clear_data(db);
1249 mutex_exit(&db->db_mtx);
1255 * Calculate which level n block references the data at the level 0 offset
1259 dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset)
1261 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1263 * The level n blkid is equal to the level 0 blkid divided by
1264 * the number of level 0s in a level n block.
1266 * The level 0 blkid is offset >> datablkshift =
1267 * offset / 2^datablkshift.
1269 * The number of level 0s in a level n is the number of block
1270 * pointers in an indirect block, raised to the power of level.
1271 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1272 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1274 * Thus, the level n blkid is: offset /
1275 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1276 * = offset / 2^(datablkshift + level *
1277 * (indblkshift - SPA_BLKPTRSHIFT))
1278 * = offset >> (datablkshift + level *
1279 * (indblkshift - SPA_BLKPTRSHIFT))
1282 const unsigned exp = dn->dn_datablkshift +
1283 level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
1285 if (exp >= 8 * sizeof (offset)) {
1286 /* This only happens on the highest indirection level */
1287 ASSERT3U(level, ==, dn->dn_nlevels - 1);
1291 ASSERT3U(exp, <, 8 * sizeof (offset));
1293 return (offset >> exp);
1295 ASSERT3U(offset, <, dn->dn_datablksz);
1301 * This function is used to lock the parent of the provided dbuf. This should be
1302 * used when modifying or reading db_blkptr.
1305 dmu_buf_lock_parent(dmu_buf_impl_t *db, krw_t rw, void *tag)
1307 enum db_lock_type ret = DLT_NONE;
1308 if (db->db_parent != NULL) {
1309 rw_enter(&db->db_parent->db_rwlock, rw);
1311 } else if (dmu_objset_ds(db->db_objset) != NULL) {
1312 rrw_enter(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, rw,
1317 * We only return a DLT_NONE lock when it's the top-most indirect block
1318 * of the meta-dnode of the MOS.
1324 * We need to pass the lock type in because it's possible that the block will
1325 * move from being the topmost indirect block in a dnode (and thus, have no
1326 * parent) to not the top-most via an indirection increase. This would cause a
1327 * panic if we didn't pass the lock type in.
1330 dmu_buf_unlock_parent(dmu_buf_impl_t *db, db_lock_type_t type, void *tag)
1332 if (type == DLT_PARENT)
1333 rw_exit(&db->db_parent->db_rwlock);
1334 else if (type == DLT_OBJSET)
1335 rrw_exit(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, tag);
1339 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1340 arc_buf_t *buf, void *vdb)
1342 (void) zb, (void) bp;
1343 dmu_buf_impl_t *db = vdb;
1345 mutex_enter(&db->db_mtx);
1346 ASSERT3U(db->db_state, ==, DB_READ);
1348 * All reads are synchronous, so we must have a hold on the dbuf
1350 ASSERT(zfs_refcount_count(&db->db_holds) > 0);
1351 ASSERT(db->db_buf == NULL);
1352 ASSERT(db->db.db_data == NULL);
1355 ASSERT(zio == NULL || zio->io_error != 0);
1356 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1357 ASSERT3P(db->db_buf, ==, NULL);
1358 db->db_state = DB_UNCACHED;
1359 DTRACE_SET_STATE(db, "i/o error");
1360 } else if (db->db_level == 0 && db->db_freed_in_flight) {
1361 /* freed in flight */
1362 ASSERT(zio == NULL || zio->io_error == 0);
1363 arc_release(buf, db);
1364 bzero(buf->b_data, db->db.db_size);
1365 arc_buf_freeze(buf);
1366 db->db_freed_in_flight = FALSE;
1367 dbuf_set_data(db, buf);
1368 db->db_state = DB_CACHED;
1369 DTRACE_SET_STATE(db, "freed in flight");
1372 ASSERT(zio == NULL || zio->io_error == 0);
1373 dbuf_set_data(db, buf);
1374 db->db_state = DB_CACHED;
1375 DTRACE_SET_STATE(db, "successful read");
1377 cv_broadcast(&db->db_changed);
1378 dbuf_rele_and_unlock(db, NULL, B_FALSE);
1382 * Shortcut for performing reads on bonus dbufs. Returns
1383 * an error if we fail to verify the dnode associated with
1384 * a decrypted block. Otherwise success.
1387 dbuf_read_bonus(dmu_buf_impl_t *db, dnode_t *dn, uint32_t flags)
1389 int bonuslen, max_bonuslen, err;
1391 err = dbuf_read_verify_dnode_crypt(db, flags);
1395 bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1396 max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1397 ASSERT(MUTEX_HELD(&db->db_mtx));
1398 ASSERT(DB_DNODE_HELD(db));
1399 ASSERT3U(bonuslen, <=, db->db.db_size);
1400 db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP);
1401 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1402 if (bonuslen < max_bonuslen)
1403 bzero(db->db.db_data, max_bonuslen);
1405 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1406 db->db_state = DB_CACHED;
1407 DTRACE_SET_STATE(db, "bonus buffer filled");
1412 dbuf_handle_indirect_hole(dmu_buf_impl_t *db, dnode_t *dn)
1414 blkptr_t *bps = db->db.db_data;
1415 uint32_t indbs = 1ULL << dn->dn_indblkshift;
1416 int n_bps = indbs >> SPA_BLKPTRSHIFT;
1418 for (int i = 0; i < n_bps; i++) {
1419 blkptr_t *bp = &bps[i];
1421 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, indbs);
1422 BP_SET_LSIZE(bp, BP_GET_LEVEL(db->db_blkptr) == 1 ?
1423 dn->dn_datablksz : BP_GET_LSIZE(db->db_blkptr));
1424 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1425 BP_SET_LEVEL(bp, BP_GET_LEVEL(db->db_blkptr) - 1);
1426 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1431 * Handle reads on dbufs that are holes, if necessary. This function
1432 * requires that the dbuf's mutex is held. Returns success (0) if action
1433 * was taken, ENOENT if no action was taken.
1436 dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn)
1438 ASSERT(MUTEX_HELD(&db->db_mtx));
1440 int is_hole = db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr);
1442 * For level 0 blocks only, if the above check fails:
1443 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1444 * processes the delete record and clears the bp while we are waiting
1445 * for the dn_mtx (resulting in a "no" from block_freed).
1447 if (!is_hole && db->db_level == 0) {
1448 is_hole = dnode_block_freed(dn, db->db_blkid) ||
1449 BP_IS_HOLE(db->db_blkptr);
1453 dbuf_set_data(db, dbuf_alloc_arcbuf(db));
1454 bzero(db->db.db_data, db->db.db_size);
1456 if (db->db_blkptr != NULL && db->db_level > 0 &&
1457 BP_IS_HOLE(db->db_blkptr) &&
1458 db->db_blkptr->blk_birth != 0) {
1459 dbuf_handle_indirect_hole(db, dn);
1461 db->db_state = DB_CACHED;
1462 DTRACE_SET_STATE(db, "hole read satisfied");
1469 * This function ensures that, when doing a decrypting read of a block,
1470 * we make sure we have decrypted the dnode associated with it. We must do
1471 * this so that we ensure we are fully authenticating the checksum-of-MACs
1472 * tree from the root of the objset down to this block. Indirect blocks are
1473 * always verified against their secure checksum-of-MACs assuming that the
1474 * dnode containing them is correct. Now that we are doing a decrypting read,
1475 * we can be sure that the key is loaded and verify that assumption. This is
1476 * especially important considering that we always read encrypted dnode
1477 * blocks as raw data (without verifying their MACs) to start, and
1478 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1481 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags)
1484 objset_t *os = db->db_objset;
1485 arc_buf_t *dnode_abuf;
1487 zbookmark_phys_t zb;
1489 ASSERT(MUTEX_HELD(&db->db_mtx));
1491 if (!os->os_encrypted || os->os_raw_receive ||
1492 (flags & DB_RF_NO_DECRYPT) != 0)
1497 dnode_abuf = (dn->dn_dbuf != NULL) ? dn->dn_dbuf->db_buf : NULL;
1499 if (dnode_abuf == NULL || !arc_is_encrypted(dnode_abuf)) {
1504 SET_BOOKMARK(&zb, dmu_objset_id(os),
1505 DMU_META_DNODE_OBJECT, 0, dn->dn_dbuf->db_blkid);
1506 err = arc_untransform(dnode_abuf, os->os_spa, &zb, B_TRUE);
1509 * An error code of EACCES tells us that the key is still not
1510 * available. This is ok if we are only reading authenticated
1511 * (and therefore non-encrypted) blocks.
1513 if (err == EACCES && ((db->db_blkid != DMU_BONUS_BLKID &&
1514 !DMU_OT_IS_ENCRYPTED(dn->dn_type)) ||
1515 (db->db_blkid == DMU_BONUS_BLKID &&
1516 !DMU_OT_IS_ENCRYPTED(dn->dn_bonustype))))
1525 * Drops db_mtx and the parent lock specified by dblt and tag before
1529 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags,
1530 db_lock_type_t dblt, void *tag)
1533 zbookmark_phys_t zb;
1534 uint32_t aflags = ARC_FLAG_NOWAIT;
1537 err = zio_flags = 0;
1540 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1541 ASSERT(MUTEX_HELD(&db->db_mtx));
1542 ASSERT(db->db_state == DB_UNCACHED);
1543 ASSERT(db->db_buf == NULL);
1544 ASSERT(db->db_parent == NULL ||
1545 RW_LOCK_HELD(&db->db_parent->db_rwlock));
1547 if (db->db_blkid == DMU_BONUS_BLKID) {
1548 err = dbuf_read_bonus(db, dn, flags);
1552 err = dbuf_read_hole(db, dn);
1557 * Any attempt to read a redacted block should result in an error. This
1558 * will never happen under normal conditions, but can be useful for
1559 * debugging purposes.
1561 if (BP_IS_REDACTED(db->db_blkptr)) {
1562 ASSERT(dsl_dataset_feature_is_active(
1563 db->db_objset->os_dsl_dataset,
1564 SPA_FEATURE_REDACTED_DATASETS));
1565 err = SET_ERROR(EIO);
1569 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
1570 db->db.db_object, db->db_level, db->db_blkid);
1573 * All bps of an encrypted os should have the encryption bit set.
1574 * If this is not true it indicates tampering and we report an error.
1576 if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) {
1577 spa_log_error(db->db_objset->os_spa, &zb);
1578 zfs_panic_recover("unencrypted block in encrypted "
1579 "object set %llu", dmu_objset_id(db->db_objset));
1580 err = SET_ERROR(EIO);
1584 err = dbuf_read_verify_dnode_crypt(db, flags);
1590 db->db_state = DB_READ;
1591 DTRACE_SET_STATE(db, "read issued");
1592 mutex_exit(&db->db_mtx);
1594 if (dbuf_is_l2cacheable(db))
1595 aflags |= ARC_FLAG_L2CACHE;
1597 dbuf_add_ref(db, NULL);
1599 zio_flags = (flags & DB_RF_CANFAIL) ?
1600 ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED;
1602 if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr))
1603 zio_flags |= ZIO_FLAG_RAW;
1605 * The zio layer will copy the provided blkptr later, but we need to
1606 * do this now so that we can release the parent's rwlock. We have to
1607 * do that now so that if dbuf_read_done is called synchronously (on
1608 * an l1 cache hit) we don't acquire the db_mtx while holding the
1609 * parent's rwlock, which would be a lock ordering violation.
1611 blkptr_t bp = *db->db_blkptr;
1612 dmu_buf_unlock_parent(db, dblt, tag);
1613 (void) arc_read(zio, db->db_objset->os_spa, &bp,
1614 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags,
1619 mutex_exit(&db->db_mtx);
1620 dmu_buf_unlock_parent(db, dblt, tag);
1625 * This is our just-in-time copy function. It makes a copy of buffers that
1626 * have been modified in a previous transaction group before we access them in
1627 * the current active group.
1629 * This function is used in three places: when we are dirtying a buffer for the
1630 * first time in a txg, when we are freeing a range in a dnode that includes
1631 * this buffer, and when we are accessing a buffer which was received compressed
1632 * and later referenced in a WRITE_BYREF record.
1634 * Note that when we are called from dbuf_free_range() we do not put a hold on
1635 * the buffer, we just traverse the active dbuf list for the dnode.
1638 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1640 dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records);
1642 ASSERT(MUTEX_HELD(&db->db_mtx));
1643 ASSERT(db->db.db_data != NULL);
1644 ASSERT(db->db_level == 0);
1645 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1648 (dr->dt.dl.dr_data !=
1649 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1653 * If the last dirty record for this dbuf has not yet synced
1654 * and its referencing the dbuf data, either:
1655 * reset the reference to point to a new copy,
1656 * or (if there a no active holders)
1657 * just null out the current db_data pointer.
1659 ASSERT3U(dr->dr_txg, >=, txg - 2);
1660 if (db->db_blkid == DMU_BONUS_BLKID) {
1661 dnode_t *dn = DB_DNODE(db);
1662 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1663 dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP);
1664 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1665 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1666 } else if (zfs_refcount_count(&db->db_holds) > db->db_dirtycnt) {
1667 dnode_t *dn = DB_DNODE(db);
1668 int size = arc_buf_size(db->db_buf);
1669 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1670 spa_t *spa = db->db_objset->os_spa;
1671 enum zio_compress compress_type =
1672 arc_get_compression(db->db_buf);
1673 uint8_t complevel = arc_get_complevel(db->db_buf);
1675 if (arc_is_encrypted(db->db_buf)) {
1676 boolean_t byteorder;
1677 uint8_t salt[ZIO_DATA_SALT_LEN];
1678 uint8_t iv[ZIO_DATA_IV_LEN];
1679 uint8_t mac[ZIO_DATA_MAC_LEN];
1681 arc_get_raw_params(db->db_buf, &byteorder, salt,
1683 dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db,
1684 dmu_objset_id(dn->dn_objset), byteorder, salt, iv,
1685 mac, dn->dn_type, size, arc_buf_lsize(db->db_buf),
1686 compress_type, complevel);
1687 } else if (compress_type != ZIO_COMPRESS_OFF) {
1688 ASSERT3U(type, ==, ARC_BUFC_DATA);
1689 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1690 size, arc_buf_lsize(db->db_buf), compress_type,
1693 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1695 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1698 dbuf_clear_data(db);
1703 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1710 * We don't have to hold the mutex to check db_state because it
1711 * can't be freed while we have a hold on the buffer.
1713 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1715 if (db->db_state == DB_NOFILL)
1716 return (SET_ERROR(EIO));
1721 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1722 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1723 DBUF_IS_CACHEABLE(db);
1725 mutex_enter(&db->db_mtx);
1726 if (db->db_state == DB_CACHED) {
1727 spa_t *spa = dn->dn_objset->os_spa;
1730 * Ensure that this block's dnode has been decrypted if
1731 * the caller has requested decrypted data.
1733 err = dbuf_read_verify_dnode_crypt(db, flags);
1736 * If the arc buf is compressed or encrypted and the caller
1737 * requested uncompressed data, we need to untransform it
1738 * before returning. We also call arc_untransform() on any
1739 * unauthenticated blocks, which will verify their MAC if
1740 * the key is now available.
1742 if (err == 0 && db->db_buf != NULL &&
1743 (flags & DB_RF_NO_DECRYPT) == 0 &&
1744 (arc_is_encrypted(db->db_buf) ||
1745 arc_is_unauthenticated(db->db_buf) ||
1746 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) {
1747 zbookmark_phys_t zb;
1749 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
1750 db->db.db_object, db->db_level, db->db_blkid);
1751 dbuf_fix_old_data(db, spa_syncing_txg(spa));
1752 err = arc_untransform(db->db_buf, spa, &zb, B_FALSE);
1753 dbuf_set_data(db, db->db_buf);
1755 mutex_exit(&db->db_mtx);
1756 if (err == 0 && prefetch) {
1757 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE,
1758 B_FALSE, flags & DB_RF_HAVESTRUCT);
1761 DBUF_STAT_BUMP(hash_hits);
1762 } else if (db->db_state == DB_UNCACHED) {
1763 spa_t *spa = dn->dn_objset->os_spa;
1764 boolean_t need_wait = B_FALSE;
1766 db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
1769 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1770 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1773 err = dbuf_read_impl(db, zio, flags, dblt, FTAG);
1775 * dbuf_read_impl has dropped db_mtx and our parent's rwlock
1778 if (!err && prefetch) {
1779 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE,
1780 db->db_state != DB_CACHED,
1781 flags & DB_RF_HAVESTRUCT);
1785 DBUF_STAT_BUMP(hash_misses);
1788 * If we created a zio_root we must execute it to avoid
1789 * leaking it, even if it isn't attached to any work due
1790 * to an error in dbuf_read_impl().
1794 err = zio_wait(zio);
1796 VERIFY0(zio_wait(zio));
1800 * Another reader came in while the dbuf was in flight
1801 * between UNCACHED and CACHED. Either a writer will finish
1802 * writing the buffer (sending the dbuf to CACHED) or the
1803 * first reader's request will reach the read_done callback
1804 * and send the dbuf to CACHED. Otherwise, a failure
1805 * occurred and the dbuf went to UNCACHED.
1807 mutex_exit(&db->db_mtx);
1809 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE,
1810 B_TRUE, flags & DB_RF_HAVESTRUCT);
1813 DBUF_STAT_BUMP(hash_misses);
1815 /* Skip the wait per the caller's request. */
1816 if ((flags & DB_RF_NEVERWAIT) == 0) {
1817 mutex_enter(&db->db_mtx);
1818 while (db->db_state == DB_READ ||
1819 db->db_state == DB_FILL) {
1820 ASSERT(db->db_state == DB_READ ||
1821 (flags & DB_RF_HAVESTRUCT) == 0);
1822 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1824 cv_wait(&db->db_changed, &db->db_mtx);
1826 if (db->db_state == DB_UNCACHED)
1827 err = SET_ERROR(EIO);
1828 mutex_exit(&db->db_mtx);
1836 dbuf_noread(dmu_buf_impl_t *db)
1838 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
1839 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1840 mutex_enter(&db->db_mtx);
1841 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1842 cv_wait(&db->db_changed, &db->db_mtx);
1843 if (db->db_state == DB_UNCACHED) {
1844 ASSERT(db->db_buf == NULL);
1845 ASSERT(db->db.db_data == NULL);
1846 dbuf_set_data(db, dbuf_alloc_arcbuf(db));
1847 db->db_state = DB_FILL;
1848 DTRACE_SET_STATE(db, "assigning filled buffer");
1849 } else if (db->db_state == DB_NOFILL) {
1850 dbuf_clear_data(db);
1852 ASSERT3U(db->db_state, ==, DB_CACHED);
1854 mutex_exit(&db->db_mtx);
1858 dbuf_unoverride(dbuf_dirty_record_t *dr)
1860 dmu_buf_impl_t *db = dr->dr_dbuf;
1861 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1862 uint64_t txg = dr->dr_txg;
1864 ASSERT(MUTEX_HELD(&db->db_mtx));
1866 * This assert is valid because dmu_sync() expects to be called by
1867 * a zilog's get_data while holding a range lock. This call only
1868 * comes from dbuf_dirty() callers who must also hold a range lock.
1870 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1871 ASSERT(db->db_level == 0);
1873 if (db->db_blkid == DMU_BONUS_BLKID ||
1874 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1877 ASSERT(db->db_data_pending != dr);
1879 /* free this block */
1880 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1881 zio_free(db->db_objset->os_spa, txg, bp);
1883 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1884 dr->dt.dl.dr_nopwrite = B_FALSE;
1885 dr->dt.dl.dr_has_raw_params = B_FALSE;
1888 * Release the already-written buffer, so we leave it in
1889 * a consistent dirty state. Note that all callers are
1890 * modifying the buffer, so they will immediately do
1891 * another (redundant) arc_release(). Therefore, leave
1892 * the buf thawed to save the effort of freezing &
1893 * immediately re-thawing it.
1895 arc_release(dr->dt.dl.dr_data, db);
1899 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1900 * data blocks in the free range, so that any future readers will find
1904 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1907 dmu_buf_impl_t *db_search;
1908 dmu_buf_impl_t *db, *db_next;
1909 uint64_t txg = tx->tx_txg;
1911 dbuf_dirty_record_t *dr;
1913 if (end_blkid > dn->dn_maxblkid &&
1914 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1915 end_blkid = dn->dn_maxblkid;
1916 dprintf_dnode(dn, "start=%llu end=%llu\n", (u_longlong_t)start_blkid,
1917 (u_longlong_t)end_blkid);
1919 db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1920 db_search->db_level = 0;
1921 db_search->db_blkid = start_blkid;
1922 db_search->db_state = DB_SEARCH;
1924 mutex_enter(&dn->dn_dbufs_mtx);
1925 db = avl_find(&dn->dn_dbufs, db_search, &where);
1926 ASSERT3P(db, ==, NULL);
1928 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1930 for (; db != NULL; db = db_next) {
1931 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1932 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1934 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1937 ASSERT3U(db->db_blkid, >=, start_blkid);
1939 /* found a level 0 buffer in the range */
1940 mutex_enter(&db->db_mtx);
1941 if (dbuf_undirty(db, tx)) {
1942 /* mutex has been dropped and dbuf destroyed */
1946 if (db->db_state == DB_UNCACHED ||
1947 db->db_state == DB_NOFILL ||
1948 db->db_state == DB_EVICTING) {
1949 ASSERT(db->db.db_data == NULL);
1950 mutex_exit(&db->db_mtx);
1953 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1954 /* will be handled in dbuf_read_done or dbuf_rele */
1955 db->db_freed_in_flight = TRUE;
1956 mutex_exit(&db->db_mtx);
1959 if (zfs_refcount_count(&db->db_holds) == 0) {
1964 /* The dbuf is referenced */
1966 dr = list_head(&db->db_dirty_records);
1968 if (dr->dr_txg == txg) {
1970 * This buffer is "in-use", re-adjust the file
1971 * size to reflect that this buffer may
1972 * contain new data when we sync.
1974 if (db->db_blkid != DMU_SPILL_BLKID &&
1975 db->db_blkid > dn->dn_maxblkid)
1976 dn->dn_maxblkid = db->db_blkid;
1977 dbuf_unoverride(dr);
1980 * This dbuf is not dirty in the open context.
1981 * Either uncache it (if its not referenced in
1982 * the open context) or reset its contents to
1985 dbuf_fix_old_data(db, txg);
1988 /* clear the contents if its cached */
1989 if (db->db_state == DB_CACHED) {
1990 ASSERT(db->db.db_data != NULL);
1991 arc_release(db->db_buf, db);
1992 rw_enter(&db->db_rwlock, RW_WRITER);
1993 bzero(db->db.db_data, db->db.db_size);
1994 rw_exit(&db->db_rwlock);
1995 arc_buf_freeze(db->db_buf);
1998 mutex_exit(&db->db_mtx);
2001 mutex_exit(&dn->dn_dbufs_mtx);
2002 kmem_free(db_search, sizeof (dmu_buf_impl_t));
2006 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
2008 arc_buf_t *buf, *old_buf;
2009 dbuf_dirty_record_t *dr;
2010 int osize = db->db.db_size;
2011 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2014 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2020 * XXX we should be doing a dbuf_read, checking the return
2021 * value and returning that up to our callers
2023 dmu_buf_will_dirty(&db->db, tx);
2025 /* create the data buffer for the new block */
2026 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
2028 /* copy old block data to the new block */
2029 old_buf = db->db_buf;
2030 bcopy(old_buf->b_data, buf->b_data, MIN(osize, size));
2031 /* zero the remainder */
2033 bzero((uint8_t *)buf->b_data + osize, size - osize);
2035 mutex_enter(&db->db_mtx);
2036 dbuf_set_data(db, buf);
2037 arc_buf_destroy(old_buf, db);
2038 db->db.db_size = size;
2040 dr = list_head(&db->db_dirty_records);
2041 /* dirty record added by dmu_buf_will_dirty() */
2043 if (db->db_level == 0)
2044 dr->dt.dl.dr_data = buf;
2045 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2046 ASSERT3U(dr->dr_accounted, ==, osize);
2047 dr->dr_accounted = size;
2048 mutex_exit(&db->db_mtx);
2050 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
2055 dbuf_release_bp(dmu_buf_impl_t *db)
2057 objset_t *os __maybe_unused = db->db_objset;
2059 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
2060 ASSERT(arc_released(os->os_phys_buf) ||
2061 list_link_active(&os->os_dsl_dataset->ds_synced_link));
2062 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
2064 (void) arc_release(db->db_buf, db);
2068 * We already have a dirty record for this TXG, and we are being
2072 dbuf_redirty(dbuf_dirty_record_t *dr)
2074 dmu_buf_impl_t *db = dr->dr_dbuf;
2076 ASSERT(MUTEX_HELD(&db->db_mtx));
2078 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
2080 * If this buffer has already been written out,
2081 * we now need to reset its state.
2083 dbuf_unoverride(dr);
2084 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
2085 db->db_state != DB_NOFILL) {
2086 /* Already released on initial dirty, so just thaw. */
2087 ASSERT(arc_released(db->db_buf));
2088 arc_buf_thaw(db->db_buf);
2093 dbuf_dirty_record_t *
2094 dbuf_dirty_lightweight(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx)
2096 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2097 IMPLY(dn->dn_objset->os_raw_receive, dn->dn_maxblkid >= blkid);
2098 dnode_new_blkid(dn, blkid, tx, B_TRUE, B_FALSE);
2099 ASSERT(dn->dn_maxblkid >= blkid);
2101 dbuf_dirty_record_t *dr = kmem_zalloc(sizeof (*dr), KM_SLEEP);
2102 list_link_init(&dr->dr_dirty_node);
2103 list_link_init(&dr->dr_dbuf_node);
2105 dr->dr_txg = tx->tx_txg;
2106 dr->dt.dll.dr_blkid = blkid;
2107 dr->dr_accounted = dn->dn_datablksz;
2110 * There should not be any dbuf for the block that we're dirtying.
2111 * Otherwise the buffer contents could be inconsistent between the
2112 * dbuf and the lightweight dirty record.
2114 ASSERT3P(NULL, ==, dbuf_find(dn->dn_objset, dn->dn_object, 0, blkid));
2116 mutex_enter(&dn->dn_mtx);
2117 int txgoff = tx->tx_txg & TXG_MASK;
2118 if (dn->dn_free_ranges[txgoff] != NULL) {
2119 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, 1);
2122 if (dn->dn_nlevels == 1) {
2123 ASSERT3U(blkid, <, dn->dn_nblkptr);
2124 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2125 mutex_exit(&dn->dn_mtx);
2126 rw_exit(&dn->dn_struct_rwlock);
2127 dnode_setdirty(dn, tx);
2129 mutex_exit(&dn->dn_mtx);
2131 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2132 dmu_buf_impl_t *parent_db = dbuf_hold_level(dn,
2133 1, blkid >> epbs, FTAG);
2134 rw_exit(&dn->dn_struct_rwlock);
2135 if (parent_db == NULL) {
2136 kmem_free(dr, sizeof (*dr));
2139 int err = dbuf_read(parent_db, NULL,
2140 (DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2142 dbuf_rele(parent_db, FTAG);
2143 kmem_free(dr, sizeof (*dr));
2147 dbuf_dirty_record_t *parent_dr = dbuf_dirty(parent_db, tx);
2148 dbuf_rele(parent_db, FTAG);
2149 mutex_enter(&parent_dr->dt.di.dr_mtx);
2150 ASSERT3U(parent_dr->dr_txg, ==, tx->tx_txg);
2151 list_insert_tail(&parent_dr->dt.di.dr_children, dr);
2152 mutex_exit(&parent_dr->dt.di.dr_mtx);
2153 dr->dr_parent = parent_dr;
2156 dmu_objset_willuse_space(dn->dn_objset, dr->dr_accounted, tx);
2161 dbuf_dirty_record_t *
2162 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2166 dbuf_dirty_record_t *dr, *dr_next, *dr_head;
2167 int txgoff = tx->tx_txg & TXG_MASK;
2168 boolean_t drop_struct_rwlock = B_FALSE;
2170 ASSERT(tx->tx_txg != 0);
2171 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2172 DMU_TX_DIRTY_BUF(tx, db);
2177 * Shouldn't dirty a regular buffer in syncing context. Private
2178 * objects may be dirtied in syncing context, but only if they
2179 * were already pre-dirtied in open context.
2182 if (dn->dn_objset->os_dsl_dataset != NULL) {
2183 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
2186 ASSERT(!dmu_tx_is_syncing(tx) ||
2187 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
2188 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
2189 dn->dn_objset->os_dsl_dataset == NULL);
2190 if (dn->dn_objset->os_dsl_dataset != NULL)
2191 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
2194 * We make this assert for private objects as well, but after we
2195 * check if we're already dirty. They are allowed to re-dirty
2196 * in syncing context.
2198 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2199 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
2200 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
2202 mutex_enter(&db->db_mtx);
2204 * XXX make this true for indirects too? The problem is that
2205 * transactions created with dmu_tx_create_assigned() from
2206 * syncing context don't bother holding ahead.
2208 ASSERT(db->db_level != 0 ||
2209 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
2210 db->db_state == DB_NOFILL);
2212 mutex_enter(&dn->dn_mtx);
2213 dnode_set_dirtyctx(dn, tx, db);
2214 if (tx->tx_txg > dn->dn_dirty_txg)
2215 dn->dn_dirty_txg = tx->tx_txg;
2216 mutex_exit(&dn->dn_mtx);
2218 if (db->db_blkid == DMU_SPILL_BLKID)
2219 dn->dn_have_spill = B_TRUE;
2222 * If this buffer is already dirty, we're done.
2224 dr_head = list_head(&db->db_dirty_records);
2225 ASSERT(dr_head == NULL || dr_head->dr_txg <= tx->tx_txg ||
2226 db->db.db_object == DMU_META_DNODE_OBJECT);
2227 dr_next = dbuf_find_dirty_lte(db, tx->tx_txg);
2228 if (dr_next && dr_next->dr_txg == tx->tx_txg) {
2231 dbuf_redirty(dr_next);
2232 mutex_exit(&db->db_mtx);
2237 * Only valid if not already dirty.
2239 ASSERT(dn->dn_object == 0 ||
2240 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
2241 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
2243 ASSERT3U(dn->dn_nlevels, >, db->db_level);
2246 * We should only be dirtying in syncing context if it's the
2247 * mos or we're initializing the os or it's a special object.
2248 * However, we are allowed to dirty in syncing context provided
2249 * we already dirtied it in open context. Hence we must make
2250 * this assertion only if we're not already dirty.
2253 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
2255 if (dn->dn_objset->os_dsl_dataset != NULL)
2256 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
2257 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
2258 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
2259 if (dn->dn_objset->os_dsl_dataset != NULL)
2260 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
2262 ASSERT(db->db.db_size != 0);
2264 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2266 if (db->db_blkid != DMU_BONUS_BLKID) {
2267 dmu_objset_willuse_space(os, db->db.db_size, tx);
2271 * If this buffer is dirty in an old transaction group we need
2272 * to make a copy of it so that the changes we make in this
2273 * transaction group won't leak out when we sync the older txg.
2275 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
2276 list_link_init(&dr->dr_dirty_node);
2277 list_link_init(&dr->dr_dbuf_node);
2279 if (db->db_level == 0) {
2280 void *data_old = db->db_buf;
2282 if (db->db_state != DB_NOFILL) {
2283 if (db->db_blkid == DMU_BONUS_BLKID) {
2284 dbuf_fix_old_data(db, tx->tx_txg);
2285 data_old = db->db.db_data;
2286 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
2288 * Release the data buffer from the cache so
2289 * that we can modify it without impacting
2290 * possible other users of this cached data
2291 * block. Note that indirect blocks and
2292 * private objects are not released until the
2293 * syncing state (since they are only modified
2296 arc_release(db->db_buf, db);
2297 dbuf_fix_old_data(db, tx->tx_txg);
2298 data_old = db->db_buf;
2300 ASSERT(data_old != NULL);
2302 dr->dt.dl.dr_data = data_old;
2304 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL);
2305 list_create(&dr->dt.di.dr_children,
2306 sizeof (dbuf_dirty_record_t),
2307 offsetof(dbuf_dirty_record_t, dr_dirty_node));
2309 if (db->db_blkid != DMU_BONUS_BLKID)
2310 dr->dr_accounted = db->db.db_size;
2312 dr->dr_txg = tx->tx_txg;
2313 list_insert_before(&db->db_dirty_records, dr_next, dr);
2316 * We could have been freed_in_flight between the dbuf_noread
2317 * and dbuf_dirty. We win, as though the dbuf_noread() had
2318 * happened after the free.
2320 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2321 db->db_blkid != DMU_SPILL_BLKID) {
2322 mutex_enter(&dn->dn_mtx);
2323 if (dn->dn_free_ranges[txgoff] != NULL) {
2324 range_tree_clear(dn->dn_free_ranges[txgoff],
2327 mutex_exit(&dn->dn_mtx);
2328 db->db_freed_in_flight = FALSE;
2332 * This buffer is now part of this txg
2334 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
2335 db->db_dirtycnt += 1;
2336 ASSERT3U(db->db_dirtycnt, <=, 3);
2338 mutex_exit(&db->db_mtx);
2340 if (db->db_blkid == DMU_BONUS_BLKID ||
2341 db->db_blkid == DMU_SPILL_BLKID) {
2342 mutex_enter(&dn->dn_mtx);
2343 ASSERT(!list_link_active(&dr->dr_dirty_node));
2344 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2345 mutex_exit(&dn->dn_mtx);
2346 dnode_setdirty(dn, tx);
2351 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
2352 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2353 drop_struct_rwlock = B_TRUE;
2357 * If we are overwriting a dedup BP, then unless it is snapshotted,
2358 * when we get to syncing context we will need to decrement its
2359 * refcount in the DDT. Prefetch the relevant DDT block so that
2360 * syncing context won't have to wait for the i/o.
2362 if (db->db_blkptr != NULL) {
2363 db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
2364 ddt_prefetch(os->os_spa, db->db_blkptr);
2365 dmu_buf_unlock_parent(db, dblt, FTAG);
2369 * We need to hold the dn_struct_rwlock to make this assertion,
2370 * because it protects dn_phys / dn_next_nlevels from changing.
2372 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
2373 dn->dn_phys->dn_nlevels > db->db_level ||
2374 dn->dn_next_nlevels[txgoff] > db->db_level ||
2375 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
2376 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
2379 if (db->db_level == 0) {
2380 ASSERT(!db->db_objset->os_raw_receive ||
2381 dn->dn_maxblkid >= db->db_blkid);
2382 dnode_new_blkid(dn, db->db_blkid, tx,
2383 drop_struct_rwlock, B_FALSE);
2384 ASSERT(dn->dn_maxblkid >= db->db_blkid);
2387 if (db->db_level+1 < dn->dn_nlevels) {
2388 dmu_buf_impl_t *parent = db->db_parent;
2389 dbuf_dirty_record_t *di;
2390 int parent_held = FALSE;
2392 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
2393 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2394 parent = dbuf_hold_level(dn, db->db_level + 1,
2395 db->db_blkid >> epbs, FTAG);
2396 ASSERT(parent != NULL);
2399 if (drop_struct_rwlock)
2400 rw_exit(&dn->dn_struct_rwlock);
2401 ASSERT3U(db->db_level + 1, ==, parent->db_level);
2402 di = dbuf_dirty(parent, tx);
2404 dbuf_rele(parent, FTAG);
2406 mutex_enter(&db->db_mtx);
2408 * Since we've dropped the mutex, it's possible that
2409 * dbuf_undirty() might have changed this out from under us.
2411 if (list_head(&db->db_dirty_records) == dr ||
2412 dn->dn_object == DMU_META_DNODE_OBJECT) {
2413 mutex_enter(&di->dt.di.dr_mtx);
2414 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
2415 ASSERT(!list_link_active(&dr->dr_dirty_node));
2416 list_insert_tail(&di->dt.di.dr_children, dr);
2417 mutex_exit(&di->dt.di.dr_mtx);
2420 mutex_exit(&db->db_mtx);
2422 ASSERT(db->db_level + 1 == dn->dn_nlevels);
2423 ASSERT(db->db_blkid < dn->dn_nblkptr);
2424 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
2425 mutex_enter(&dn->dn_mtx);
2426 ASSERT(!list_link_active(&dr->dr_dirty_node));
2427 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2428 mutex_exit(&dn->dn_mtx);
2429 if (drop_struct_rwlock)
2430 rw_exit(&dn->dn_struct_rwlock);
2433 dnode_setdirty(dn, tx);
2439 dbuf_undirty_bonus(dbuf_dirty_record_t *dr)
2441 dmu_buf_impl_t *db = dr->dr_dbuf;
2443 if (dr->dt.dl.dr_data != db->db.db_data) {
2444 struct dnode *dn = dr->dr_dnode;
2445 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
2447 kmem_free(dr->dt.dl.dr_data, max_bonuslen);
2448 arc_space_return(max_bonuslen, ARC_SPACE_BONUS);
2450 db->db_data_pending = NULL;
2451 ASSERT(list_next(&db->db_dirty_records, dr) == NULL);
2452 list_remove(&db->db_dirty_records, dr);
2453 if (dr->dr_dbuf->db_level != 0) {
2454 mutex_destroy(&dr->dt.di.dr_mtx);
2455 list_destroy(&dr->dt.di.dr_children);
2457 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2458 ASSERT3U(db->db_dirtycnt, >, 0);
2459 db->db_dirtycnt -= 1;
2463 * Undirty a buffer in the transaction group referenced by the given
2464 * transaction. Return whether this evicted the dbuf.
2467 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2469 uint64_t txg = tx->tx_txg;
2474 * Due to our use of dn_nlevels below, this can only be called
2475 * in open context, unless we are operating on the MOS.
2476 * From syncing context, dn_nlevels may be different from the
2477 * dn_nlevels used when dbuf was dirtied.
2479 ASSERT(db->db_objset ==
2480 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2481 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2482 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2483 ASSERT0(db->db_level);
2484 ASSERT(MUTEX_HELD(&db->db_mtx));
2487 * If this buffer is not dirty, we're done.
2489 dbuf_dirty_record_t *dr = dbuf_find_dirty_eq(db, txg);
2492 ASSERT(dr->dr_dbuf == db);
2494 dnode_t *dn = dr->dr_dnode;
2496 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2498 ASSERT(db->db.db_size != 0);
2500 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2501 dr->dr_accounted, txg);
2503 list_remove(&db->db_dirty_records, dr);
2506 * Note that there are three places in dbuf_dirty()
2507 * where this dirty record may be put on a list.
2508 * Make sure to do a list_remove corresponding to
2509 * every one of those list_insert calls.
2511 if (dr->dr_parent) {
2512 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2513 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2514 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2515 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2516 db->db_level + 1 == dn->dn_nlevels) {
2517 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2518 mutex_enter(&dn->dn_mtx);
2519 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2520 mutex_exit(&dn->dn_mtx);
2523 if (db->db_state != DB_NOFILL) {
2524 dbuf_unoverride(dr);
2526 ASSERT(db->db_buf != NULL);
2527 ASSERT(dr->dt.dl.dr_data != NULL);
2528 if (dr->dt.dl.dr_data != db->db_buf)
2529 arc_buf_destroy(dr->dt.dl.dr_data, db);
2532 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2534 ASSERT(db->db_dirtycnt > 0);
2535 db->db_dirtycnt -= 1;
2537 if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2538 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2547 dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx)
2549 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2551 ASSERT(tx->tx_txg != 0);
2552 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2555 * Quick check for dirtiness. For already dirty blocks, this
2556 * reduces runtime of this function by >90%, and overall performance
2557 * by 50% for some workloads (e.g. file deletion with indirect blocks
2560 mutex_enter(&db->db_mtx);
2562 if (db->db_state == DB_CACHED) {
2563 dbuf_dirty_record_t *dr = dbuf_find_dirty_eq(db, tx->tx_txg);
2565 * It's possible that it is already dirty but not cached,
2566 * because there are some calls to dbuf_dirty() that don't
2567 * go through dmu_buf_will_dirty().
2570 /* This dbuf is already dirty and cached. */
2572 mutex_exit(&db->db_mtx);
2576 mutex_exit(&db->db_mtx);
2579 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2580 flags |= DB_RF_HAVESTRUCT;
2582 (void) dbuf_read(db, NULL, flags);
2583 (void) dbuf_dirty(db, tx);
2587 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2589 dmu_buf_will_dirty_impl(db_fake,
2590 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx);
2594 dmu_buf_is_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2596 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2597 dbuf_dirty_record_t *dr;
2599 mutex_enter(&db->db_mtx);
2600 dr = dbuf_find_dirty_eq(db, tx->tx_txg);
2601 mutex_exit(&db->db_mtx);
2602 return (dr != NULL);
2606 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2608 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2610 db->db_state = DB_NOFILL;
2611 DTRACE_SET_STATE(db, "allocating NOFILL buffer");
2612 dmu_buf_will_fill(db_fake, tx);
2616 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2618 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2620 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2621 ASSERT(tx->tx_txg != 0);
2622 ASSERT(db->db_level == 0);
2623 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2625 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2626 dmu_tx_private_ok(tx));
2629 (void) dbuf_dirty(db, tx);
2633 * This function is effectively the same as dmu_buf_will_dirty(), but
2634 * indicates the caller expects raw encrypted data in the db, and provides
2635 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2636 * blkptr_t when this dbuf is written. This is only used for blocks of
2637 * dnodes, during raw receive.
2640 dmu_buf_set_crypt_params(dmu_buf_t *db_fake, boolean_t byteorder,
2641 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx)
2643 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2644 dbuf_dirty_record_t *dr;
2647 * dr_has_raw_params is only processed for blocks of dnodes
2648 * (see dbuf_sync_dnode_leaf_crypt()).
2650 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT);
2651 ASSERT3U(db->db_level, ==, 0);
2652 ASSERT(db->db_objset->os_raw_receive);
2654 dmu_buf_will_dirty_impl(db_fake,
2655 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx);
2657 dr = dbuf_find_dirty_eq(db, tx->tx_txg);
2659 ASSERT3P(dr, !=, NULL);
2661 dr->dt.dl.dr_has_raw_params = B_TRUE;
2662 dr->dt.dl.dr_byteorder = byteorder;
2663 bcopy(salt, dr->dt.dl.dr_salt, ZIO_DATA_SALT_LEN);
2664 bcopy(iv, dr->dt.dl.dr_iv, ZIO_DATA_IV_LEN);
2665 bcopy(mac, dr->dt.dl.dr_mac, ZIO_DATA_MAC_LEN);
2669 dbuf_override_impl(dmu_buf_impl_t *db, const blkptr_t *bp, dmu_tx_t *tx)
2671 struct dirty_leaf *dl;
2672 dbuf_dirty_record_t *dr;
2674 dr = list_head(&db->db_dirty_records);
2675 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2677 dl->dr_overridden_by = *bp;
2678 dl->dr_override_state = DR_OVERRIDDEN;
2679 dl->dr_overridden_by.blk_birth = dr->dr_txg;
2683 dmu_buf_fill_done(dmu_buf_t *dbuf, dmu_tx_t *tx)
2686 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2687 dbuf_states_t old_state;
2688 mutex_enter(&db->db_mtx);
2691 old_state = db->db_state;
2692 db->db_state = DB_CACHED;
2693 if (old_state == DB_FILL) {
2694 if (db->db_level == 0 && db->db_freed_in_flight) {
2695 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2696 /* we were freed while filling */
2697 /* XXX dbuf_undirty? */
2698 bzero(db->db.db_data, db->db.db_size);
2699 db->db_freed_in_flight = FALSE;
2700 DTRACE_SET_STATE(db,
2701 "fill done handling freed in flight");
2703 DTRACE_SET_STATE(db, "fill done");
2705 cv_broadcast(&db->db_changed);
2707 mutex_exit(&db->db_mtx);
2711 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2712 bp_embedded_type_t etype, enum zio_compress comp,
2713 int uncompressed_size, int compressed_size, int byteorder,
2716 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2717 struct dirty_leaf *dl;
2718 dmu_object_type_t type;
2719 dbuf_dirty_record_t *dr;
2721 if (etype == BP_EMBEDDED_TYPE_DATA) {
2722 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2723 SPA_FEATURE_EMBEDDED_DATA));
2727 type = DB_DNODE(db)->dn_type;
2730 ASSERT0(db->db_level);
2731 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2733 dmu_buf_will_not_fill(dbuf, tx);
2735 dr = list_head(&db->db_dirty_records);
2736 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2738 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2739 data, comp, uncompressed_size, compressed_size);
2740 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2741 BP_SET_TYPE(&dl->dr_overridden_by, type);
2742 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2743 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2745 dl->dr_override_state = DR_OVERRIDDEN;
2746 dl->dr_overridden_by.blk_birth = dr->dr_txg;
2750 dmu_buf_redact(dmu_buf_t *dbuf, dmu_tx_t *tx)
2752 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2753 dmu_object_type_t type;
2754 ASSERT(dsl_dataset_feature_is_active(db->db_objset->os_dsl_dataset,
2755 SPA_FEATURE_REDACTED_DATASETS));
2758 type = DB_DNODE(db)->dn_type;
2761 ASSERT0(db->db_level);
2762 dmu_buf_will_not_fill(dbuf, tx);
2764 blkptr_t bp = { { { {0} } } };
2765 BP_SET_TYPE(&bp, type);
2766 BP_SET_LEVEL(&bp, 0);
2767 BP_SET_BIRTH(&bp, tx->tx_txg, 0);
2768 BP_SET_REDACTED(&bp);
2769 BPE_SET_LSIZE(&bp, dbuf->db_size);
2771 dbuf_override_impl(db, &bp, tx);
2775 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2776 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2779 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2781 ASSERT(!zfs_refcount_is_zero(&db->db_holds));
2782 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2783 ASSERT(db->db_level == 0);
2784 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2785 ASSERT(buf != NULL);
2786 ASSERT3U(arc_buf_lsize(buf), ==, db->db.db_size);
2787 ASSERT(tx->tx_txg != 0);
2789 arc_return_buf(buf, db);
2790 ASSERT(arc_released(buf));
2792 mutex_enter(&db->db_mtx);
2794 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2795 cv_wait(&db->db_changed, &db->db_mtx);
2797 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2799 if (db->db_state == DB_CACHED &&
2800 zfs_refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2802 * In practice, we will never have a case where we have an
2803 * encrypted arc buffer while additional holds exist on the
2804 * dbuf. We don't handle this here so we simply assert that
2807 ASSERT(!arc_is_encrypted(buf));
2808 mutex_exit(&db->db_mtx);
2809 (void) dbuf_dirty(db, tx);
2810 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2811 arc_buf_destroy(buf, db);
2815 if (db->db_state == DB_CACHED) {
2816 dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records);
2818 ASSERT(db->db_buf != NULL);
2819 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2820 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2822 if (!arc_released(db->db_buf)) {
2823 ASSERT(dr->dt.dl.dr_override_state ==
2825 arc_release(db->db_buf, db);
2827 dr->dt.dl.dr_data = buf;
2828 arc_buf_destroy(db->db_buf, db);
2829 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2830 arc_release(db->db_buf, db);
2831 arc_buf_destroy(db->db_buf, db);
2835 ASSERT(db->db_buf == NULL);
2836 dbuf_set_data(db, buf);
2837 db->db_state = DB_FILL;
2838 DTRACE_SET_STATE(db, "filling assigned arcbuf");
2839 mutex_exit(&db->db_mtx);
2840 (void) dbuf_dirty(db, tx);
2841 dmu_buf_fill_done(&db->db, tx);
2845 dbuf_destroy(dmu_buf_impl_t *db)
2848 dmu_buf_impl_t *parent = db->db_parent;
2849 dmu_buf_impl_t *dndb;
2851 ASSERT(MUTEX_HELD(&db->db_mtx));
2852 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2854 if (db->db_buf != NULL) {
2855 arc_buf_destroy(db->db_buf, db);
2859 if (db->db_blkid == DMU_BONUS_BLKID) {
2860 int slots = DB_DNODE(db)->dn_num_slots;
2861 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2862 if (db->db.db_data != NULL) {
2863 kmem_free(db->db.db_data, bonuslen);
2864 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2865 db->db_state = DB_UNCACHED;
2866 DTRACE_SET_STATE(db, "buffer cleared");
2870 dbuf_clear_data(db);
2872 if (multilist_link_active(&db->db_cache_link)) {
2873 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
2874 db->db_caching_status == DB_DBUF_METADATA_CACHE);
2876 multilist_remove(&dbuf_caches[db->db_caching_status].cache, db);
2877 (void) zfs_refcount_remove_many(
2878 &dbuf_caches[db->db_caching_status].size,
2879 db->db.db_size, db);
2881 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2882 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2884 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2885 DBUF_STAT_BUMPDOWN(cache_count);
2886 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2889 db->db_caching_status = DB_NO_CACHE;
2892 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2893 ASSERT(db->db_data_pending == NULL);
2894 ASSERT(list_is_empty(&db->db_dirty_records));
2896 db->db_state = DB_EVICTING;
2897 DTRACE_SET_STATE(db, "buffer eviction started");
2898 db->db_blkptr = NULL;
2901 * Now that db_state is DB_EVICTING, nobody else can find this via
2902 * the hash table. We can now drop db_mtx, which allows us to
2903 * acquire the dn_dbufs_mtx.
2905 mutex_exit(&db->db_mtx);
2910 if (db->db_blkid != DMU_BONUS_BLKID) {
2911 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2913 mutex_enter_nested(&dn->dn_dbufs_mtx,
2915 avl_remove(&dn->dn_dbufs, db);
2919 mutex_exit(&dn->dn_dbufs_mtx);
2921 * Decrementing the dbuf count means that the hold corresponding
2922 * to the removed dbuf is no longer discounted in dnode_move(),
2923 * so the dnode cannot be moved until after we release the hold.
2924 * The membar_producer() ensures visibility of the decremented
2925 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2928 mutex_enter(&dn->dn_mtx);
2929 dnode_rele_and_unlock(dn, db, B_TRUE);
2930 db->db_dnode_handle = NULL;
2932 dbuf_hash_remove(db);
2937 ASSERT(zfs_refcount_is_zero(&db->db_holds));
2939 db->db_parent = NULL;
2941 ASSERT(db->db_buf == NULL);
2942 ASSERT(db->db.db_data == NULL);
2943 ASSERT(db->db_hash_next == NULL);
2944 ASSERT(db->db_blkptr == NULL);
2945 ASSERT(db->db_data_pending == NULL);
2946 ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE);
2947 ASSERT(!multilist_link_active(&db->db_cache_link));
2949 kmem_cache_free(dbuf_kmem_cache, db);
2950 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2953 * If this dbuf is referenced from an indirect dbuf,
2954 * decrement the ref count on the indirect dbuf.
2956 if (parent && parent != dndb) {
2957 mutex_enter(&parent->db_mtx);
2958 dbuf_rele_and_unlock(parent, db, B_TRUE);
2963 * Note: While bpp will always be updated if the function returns success,
2964 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2965 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2968 __attribute__((always_inline))
2970 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2971 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2976 ASSERT(blkid != DMU_BONUS_BLKID);
2978 if (blkid == DMU_SPILL_BLKID) {
2979 mutex_enter(&dn->dn_mtx);
2980 if (dn->dn_have_spill &&
2981 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2982 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2985 dbuf_add_ref(dn->dn_dbuf, NULL);
2986 *parentp = dn->dn_dbuf;
2987 mutex_exit(&dn->dn_mtx);
2992 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2993 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2995 ASSERT3U(level * epbs, <, 64);
2996 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2998 * This assertion shouldn't trip as long as the max indirect block size
2999 * is less than 1M. The reason for this is that up to that point,
3000 * the number of levels required to address an entire object with blocks
3001 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
3002 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
3003 * (i.e. we can address the entire object), objects will all use at most
3004 * N-1 levels and the assertion won't overflow. However, once epbs is
3005 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
3006 * enough to address an entire object, so objects will have 5 levels,
3007 * but then this assertion will overflow.
3009 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
3010 * need to redo this logic to handle overflows.
3012 ASSERT(level >= nlevels ||
3013 ((nlevels - level - 1) * epbs) +
3014 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
3015 if (level >= nlevels ||
3016 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
3017 ((nlevels - level - 1) * epbs)) ||
3019 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
3020 /* the buffer has no parent yet */
3021 return (SET_ERROR(ENOENT));
3022 } else if (level < nlevels-1) {
3023 /* this block is referenced from an indirect block */
3026 err = dbuf_hold_impl(dn, level + 1,
3027 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
3031 err = dbuf_read(*parentp, NULL,
3032 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
3034 dbuf_rele(*parentp, NULL);
3038 rw_enter(&(*parentp)->db_rwlock, RW_READER);
3039 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
3040 (blkid & ((1ULL << epbs) - 1));
3041 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
3042 ASSERT(BP_IS_HOLE(*bpp));
3043 rw_exit(&(*parentp)->db_rwlock);
3046 /* the block is referenced from the dnode */
3047 ASSERT3U(level, ==, nlevels-1);
3048 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
3049 blkid < dn->dn_phys->dn_nblkptr);
3051 dbuf_add_ref(dn->dn_dbuf, NULL);
3052 *parentp = dn->dn_dbuf;
3054 *bpp = &dn->dn_phys->dn_blkptr[blkid];
3059 static dmu_buf_impl_t *
3060 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
3061 dmu_buf_impl_t *parent, blkptr_t *blkptr)
3063 objset_t *os = dn->dn_objset;
3064 dmu_buf_impl_t *db, *odb;
3066 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
3067 ASSERT(dn->dn_type != DMU_OT_NONE);
3069 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
3071 list_create(&db->db_dirty_records, sizeof (dbuf_dirty_record_t),
3072 offsetof(dbuf_dirty_record_t, dr_dbuf_node));
3075 db->db.db_object = dn->dn_object;
3076 db->db_level = level;
3077 db->db_blkid = blkid;
3078 db->db_dirtycnt = 0;
3079 db->db_dnode_handle = dn->dn_handle;
3080 db->db_parent = parent;
3081 db->db_blkptr = blkptr;
3084 db->db_user_immediate_evict = FALSE;
3085 db->db_freed_in_flight = FALSE;
3086 db->db_pending_evict = FALSE;
3088 if (blkid == DMU_BONUS_BLKID) {
3089 ASSERT3P(parent, ==, dn->dn_dbuf);
3090 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
3091 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
3092 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
3093 db->db.db_offset = DMU_BONUS_BLKID;
3094 db->db_state = DB_UNCACHED;
3095 DTRACE_SET_STATE(db, "bonus buffer created");
3096 db->db_caching_status = DB_NO_CACHE;
3097 /* the bonus dbuf is not placed in the hash table */
3098 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
3100 } else if (blkid == DMU_SPILL_BLKID) {
3101 db->db.db_size = (blkptr != NULL) ?
3102 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
3103 db->db.db_offset = 0;
3106 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
3107 db->db.db_size = blocksize;
3108 db->db.db_offset = db->db_blkid * blocksize;
3112 * Hold the dn_dbufs_mtx while we get the new dbuf
3113 * in the hash table *and* added to the dbufs list.
3114 * This prevents a possible deadlock with someone
3115 * trying to look up this dbuf before it's added to the
3118 mutex_enter(&dn->dn_dbufs_mtx);
3119 db->db_state = DB_EVICTING; /* not worth logging this state change */
3120 if ((odb = dbuf_hash_insert(db)) != NULL) {
3121 /* someone else inserted it first */
3122 mutex_exit(&dn->dn_dbufs_mtx);
3123 kmem_cache_free(dbuf_kmem_cache, db);
3124 DBUF_STAT_BUMP(hash_insert_race);
3127 avl_add(&dn->dn_dbufs, db);
3129 db->db_state = DB_UNCACHED;
3130 DTRACE_SET_STATE(db, "regular buffer created");
3131 db->db_caching_status = DB_NO_CACHE;
3132 mutex_exit(&dn->dn_dbufs_mtx);
3133 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
3135 if (parent && parent != dn->dn_dbuf)
3136 dbuf_add_ref(parent, db);
3138 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
3139 zfs_refcount_count(&dn->dn_holds) > 0);
3140 (void) zfs_refcount_add(&dn->dn_holds, db);
3142 dprintf_dbuf(db, "db=%p\n", db);
3148 * This function returns a block pointer and information about the object,
3149 * given a dnode and a block. This is a publicly accessible version of
3150 * dbuf_findbp that only returns some information, rather than the
3151 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
3152 * should be locked as (at least) a reader.
3155 dbuf_dnode_findbp(dnode_t *dn, uint64_t level, uint64_t blkid,
3156 blkptr_t *bp, uint16_t *datablkszsec, uint8_t *indblkshift)
3158 dmu_buf_impl_t *dbp = NULL;
3161 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
3163 err = dbuf_findbp(dn, level, blkid, B_FALSE, &dbp, &bp2);
3167 dbuf_rele(dbp, NULL);
3168 if (datablkszsec != NULL)
3169 *datablkszsec = dn->dn_phys->dn_datablkszsec;
3170 if (indblkshift != NULL)
3171 *indblkshift = dn->dn_phys->dn_indblkshift;
3177 typedef struct dbuf_prefetch_arg {
3178 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
3179 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
3180 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
3181 int dpa_curlevel; /* The current level that we're reading */
3182 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
3183 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
3184 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
3185 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
3186 dbuf_prefetch_fn dpa_cb; /* prefetch completion callback */
3187 void *dpa_arg; /* prefetch completion arg */
3188 } dbuf_prefetch_arg_t;
3191 dbuf_prefetch_fini(dbuf_prefetch_arg_t *dpa, boolean_t io_done)
3193 if (dpa->dpa_cb != NULL)
3194 dpa->dpa_cb(dpa->dpa_arg, io_done);
3195 kmem_free(dpa, sizeof (*dpa));
3199 dbuf_issue_final_prefetch_done(zio_t *zio, const zbookmark_phys_t *zb,
3200 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
3202 (void) zio, (void) zb, (void) iobp;
3203 dbuf_prefetch_arg_t *dpa = private;
3205 dbuf_prefetch_fini(dpa, B_TRUE);
3207 arc_buf_destroy(abuf, private);
3211 * Actually issue the prefetch read for the block given.
3214 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
3216 ASSERT(!BP_IS_REDACTED(bp) ||
3217 dsl_dataset_feature_is_active(
3218 dpa->dpa_dnode->dn_objset->os_dsl_dataset,
3219 SPA_FEATURE_REDACTED_DATASETS));
3221 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp))
3222 return (dbuf_prefetch_fini(dpa, B_FALSE));
3224 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
3225 arc_flags_t aflags =
3226 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
3229 /* dnodes are always read as raw and then converted later */
3230 if (BP_GET_TYPE(bp) == DMU_OT_DNODE && BP_IS_PROTECTED(bp) &&
3231 dpa->dpa_curlevel == 0)
3232 zio_flags |= ZIO_FLAG_RAW;
3234 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
3235 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
3236 ASSERT(dpa->dpa_zio != NULL);
3237 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp,
3238 dbuf_issue_final_prefetch_done, dpa,
3239 dpa->dpa_prio, zio_flags, &aflags, &dpa->dpa_zb);
3243 * Called when an indirect block above our prefetch target is read in. This
3244 * will either read in the next indirect block down the tree or issue the actual
3245 * prefetch if the next block down is our target.
3248 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
3249 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
3251 (void) zb, (void) iobp;
3252 dbuf_prefetch_arg_t *dpa = private;
3254 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
3255 ASSERT3S(dpa->dpa_curlevel, >, 0);
3258 ASSERT(zio == NULL || zio->io_error != 0);
3259 return (dbuf_prefetch_fini(dpa, B_TRUE));
3261 ASSERT(zio == NULL || zio->io_error == 0);
3264 * The dpa_dnode is only valid if we are called with a NULL
3265 * zio. This indicates that the arc_read() returned without
3266 * first calling zio_read() to issue a physical read. Once
3267 * a physical read is made the dpa_dnode must be invalidated
3268 * as the locks guarding it may have been dropped. If the
3269 * dpa_dnode is still valid, then we want to add it to the dbuf
3270 * cache. To do so, we must hold the dbuf associated with the block
3271 * we just prefetched, read its contents so that we associate it
3272 * with an arc_buf_t, and then release it.
3275 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
3276 if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
3277 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
3279 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
3281 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
3283 dpa->dpa_dnode = NULL;
3284 } else if (dpa->dpa_dnode != NULL) {
3285 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
3286 (dpa->dpa_epbs * (dpa->dpa_curlevel -
3287 dpa->dpa_zb.zb_level));
3288 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
3289 dpa->dpa_curlevel, curblkid, FTAG);
3291 arc_buf_destroy(abuf, private);
3292 return (dbuf_prefetch_fini(dpa, B_TRUE));
3294 (void) dbuf_read(db, NULL,
3295 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
3296 dbuf_rele(db, FTAG);
3299 dpa->dpa_curlevel--;
3300 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
3301 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
3302 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
3303 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
3305 ASSERT(!BP_IS_REDACTED(bp) ||
3306 dsl_dataset_feature_is_active(
3307 dpa->dpa_dnode->dn_objset->os_dsl_dataset,
3308 SPA_FEATURE_REDACTED_DATASETS));
3309 if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
3310 dbuf_prefetch_fini(dpa, B_TRUE);
3311 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
3312 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
3313 dbuf_issue_final_prefetch(dpa, bp);
3315 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
3316 zbookmark_phys_t zb;
3318 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3319 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
3320 iter_aflags |= ARC_FLAG_L2CACHE;
3322 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
3324 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
3325 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
3327 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
3328 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
3329 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3333 arc_buf_destroy(abuf, private);
3337 * Issue prefetch reads for the given block on the given level. If the indirect
3338 * blocks above that block are not in memory, we will read them in
3339 * asynchronously. As a result, this call never blocks waiting for a read to
3340 * complete. Note that the prefetch might fail if the dataset is encrypted and
3341 * the encryption key is unmapped before the IO completes.
3344 dbuf_prefetch_impl(dnode_t *dn, int64_t level, uint64_t blkid,
3345 zio_priority_t prio, arc_flags_t aflags, dbuf_prefetch_fn cb,
3349 int epbs, nlevels, curlevel;
3352 ASSERT(blkid != DMU_BONUS_BLKID);
3353 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
3355 if (blkid > dn->dn_maxblkid)
3358 if (level == 0 && dnode_block_freed(dn, blkid))
3362 * This dnode hasn't been written to disk yet, so there's nothing to
3365 nlevels = dn->dn_phys->dn_nlevels;
3366 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
3369 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3370 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
3373 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
3376 mutex_exit(&db->db_mtx);
3378 * This dbuf already exists. It is either CACHED, or
3379 * (we assume) about to be read or filled.
3385 * Find the closest ancestor (indirect block) of the target block
3386 * that is present in the cache. In this indirect block, we will
3387 * find the bp that is at curlevel, curblkid.
3391 while (curlevel < nlevels - 1) {
3392 int parent_level = curlevel + 1;
3393 uint64_t parent_blkid = curblkid >> epbs;
3396 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
3397 FALSE, TRUE, FTAG, &db) == 0) {
3398 blkptr_t *bpp = db->db_buf->b_data;
3399 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
3400 dbuf_rele(db, FTAG);
3404 curlevel = parent_level;
3405 curblkid = parent_blkid;
3408 if (curlevel == nlevels - 1) {
3409 /* No cached indirect blocks found. */
3410 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
3411 bp = dn->dn_phys->dn_blkptr[curblkid];
3413 ASSERT(!BP_IS_REDACTED(&bp) ||
3414 dsl_dataset_feature_is_active(dn->dn_objset->os_dsl_dataset,
3415 SPA_FEATURE_REDACTED_DATASETS));
3416 if (BP_IS_HOLE(&bp) || BP_IS_REDACTED(&bp))
3419 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
3421 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
3424 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
3425 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
3426 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
3427 dn->dn_object, level, blkid);
3428 dpa->dpa_curlevel = curlevel;
3429 dpa->dpa_prio = prio;
3430 dpa->dpa_aflags = aflags;
3431 dpa->dpa_spa = dn->dn_objset->os_spa;
3432 dpa->dpa_dnode = dn;
3433 dpa->dpa_epbs = epbs;
3438 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3439 if (dnode_level_is_l2cacheable(&bp, dn, level))
3440 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
3443 * If we have the indirect just above us, no need to do the asynchronous
3444 * prefetch chain; we'll just run the last step ourselves. If we're at
3445 * a higher level, though, we want to issue the prefetches for all the
3446 * indirect blocks asynchronously, so we can go on with whatever we were
3449 if (curlevel == level) {
3450 ASSERT3U(curblkid, ==, blkid);
3451 dbuf_issue_final_prefetch(dpa, &bp);
3453 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
3454 zbookmark_phys_t zb;
3456 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3457 if (dnode_level_is_l2cacheable(&bp, dn, level))
3458 iter_aflags |= ARC_FLAG_L2CACHE;
3460 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
3461 dn->dn_object, curlevel, curblkid);
3462 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
3463 &bp, dbuf_prefetch_indirect_done, dpa, prio,
3464 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3468 * We use pio here instead of dpa_zio since it's possible that
3469 * dpa may have already been freed.
3480 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
3484 return (dbuf_prefetch_impl(dn, level, blkid, prio, aflags, NULL, NULL));
3488 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3489 * the case of encrypted, compressed and uncompressed buffers by
3490 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3491 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3493 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3495 noinline static void
3496 dbuf_hold_copy(dnode_t *dn, dmu_buf_impl_t *db)
3498 dbuf_dirty_record_t *dr = db->db_data_pending;
3499 arc_buf_t *data = dr->dt.dl.dr_data;
3500 enum zio_compress compress_type = arc_get_compression(data);
3501 uint8_t complevel = arc_get_complevel(data);
3503 if (arc_is_encrypted(data)) {
3504 boolean_t byteorder;
3505 uint8_t salt[ZIO_DATA_SALT_LEN];
3506 uint8_t iv[ZIO_DATA_IV_LEN];
3507 uint8_t mac[ZIO_DATA_MAC_LEN];
3509 arc_get_raw_params(data, &byteorder, salt, iv, mac);
3510 dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db,
3511 dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac,
3512 dn->dn_type, arc_buf_size(data), arc_buf_lsize(data),
3513 compress_type, complevel));
3514 } else if (compress_type != ZIO_COMPRESS_OFF) {
3515 dbuf_set_data(db, arc_alloc_compressed_buf(
3516 dn->dn_objset->os_spa, db, arc_buf_size(data),
3517 arc_buf_lsize(data), compress_type, complevel));
3519 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db,
3520 DBUF_GET_BUFC_TYPE(db), db->db.db_size));
3523 rw_enter(&db->db_rwlock, RW_WRITER);
3524 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
3525 rw_exit(&db->db_rwlock);
3529 * Returns with db_holds incremented, and db_mtx not held.
3530 * Note: dn_struct_rwlock must be held.
3533 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3534 boolean_t fail_sparse, boolean_t fail_uncached,
3535 void *tag, dmu_buf_impl_t **dbp)
3537 dmu_buf_impl_t *db, *parent = NULL;
3539 /* If the pool has been created, verify the tx_sync_lock is not held */
3540 spa_t *spa = dn->dn_objset->os_spa;
3541 dsl_pool_t *dp = spa->spa_dsl_pool;
3543 ASSERT(!MUTEX_HELD(&dp->dp_tx.tx_sync_lock));
3546 ASSERT(blkid != DMU_BONUS_BLKID);
3547 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
3548 ASSERT3U(dn->dn_nlevels, >, level);
3552 /* dbuf_find() returns with db_mtx held */
3553 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
3556 blkptr_t *bp = NULL;
3560 return (SET_ERROR(ENOENT));
3562 ASSERT3P(parent, ==, NULL);
3563 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
3565 if (err == 0 && bp && BP_IS_HOLE(bp))
3566 err = SET_ERROR(ENOENT);
3569 dbuf_rele(parent, NULL);
3573 if (err && err != ENOENT)
3575 db = dbuf_create(dn, level, blkid, parent, bp);
3578 if (fail_uncached && db->db_state != DB_CACHED) {
3579 mutex_exit(&db->db_mtx);
3580 return (SET_ERROR(ENOENT));
3583 if (db->db_buf != NULL) {
3584 arc_buf_access(db->db_buf);
3585 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
3588 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
3591 * If this buffer is currently syncing out, and we are
3592 * still referencing it from db_data, we need to make a copy
3593 * of it in case we decide we want to dirty it again in this txg.
3595 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
3596 dn->dn_object != DMU_META_DNODE_OBJECT &&
3597 db->db_state == DB_CACHED && db->db_data_pending) {
3598 dbuf_dirty_record_t *dr = db->db_data_pending;
3599 if (dr->dt.dl.dr_data == db->db_buf)
3600 dbuf_hold_copy(dn, db);
3603 if (multilist_link_active(&db->db_cache_link)) {
3604 ASSERT(zfs_refcount_is_zero(&db->db_holds));
3605 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
3606 db->db_caching_status == DB_DBUF_METADATA_CACHE);
3608 multilist_remove(&dbuf_caches[db->db_caching_status].cache, db);
3609 (void) zfs_refcount_remove_many(
3610 &dbuf_caches[db->db_caching_status].size,
3611 db->db.db_size, db);
3613 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
3614 DBUF_STAT_BUMPDOWN(metadata_cache_count);
3616 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
3617 DBUF_STAT_BUMPDOWN(cache_count);
3618 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
3621 db->db_caching_status = DB_NO_CACHE;
3623 (void) zfs_refcount_add(&db->db_holds, tag);
3625 mutex_exit(&db->db_mtx);
3627 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3629 dbuf_rele(parent, NULL);
3631 ASSERT3P(DB_DNODE(db), ==, dn);
3632 ASSERT3U(db->db_blkid, ==, blkid);
3633 ASSERT3U(db->db_level, ==, level);
3640 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3642 return (dbuf_hold_level(dn, 0, blkid, tag));
3646 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3649 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3650 return (err ? NULL : db);
3654 dbuf_create_bonus(dnode_t *dn)
3656 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3658 ASSERT(dn->dn_bonus == NULL);
3659 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3663 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3665 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3667 if (db->db_blkid != DMU_SPILL_BLKID)
3668 return (SET_ERROR(ENOTSUP));
3670 blksz = SPA_MINBLOCKSIZE;
3671 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3672 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3674 dbuf_new_size(db, blksz, tx);
3680 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3682 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3685 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3687 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3689 int64_t holds = zfs_refcount_add(&db->db_holds, tag);
3690 VERIFY3S(holds, >, 1);
3693 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3695 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3698 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3699 dmu_buf_impl_t *found_db;
3700 boolean_t result = B_FALSE;
3702 if (blkid == DMU_BONUS_BLKID)
3703 found_db = dbuf_find_bonus(os, obj);
3705 found_db = dbuf_find(os, obj, 0, blkid);
3707 if (found_db != NULL) {
3708 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3709 (void) zfs_refcount_add(&db->db_holds, tag);
3712 mutex_exit(&found_db->db_mtx);
3718 * If you call dbuf_rele() you had better not be referencing the dnode handle
3719 * unless you have some other direct or indirect hold on the dnode. (An indirect
3720 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3721 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3722 * dnode's parent dbuf evicting its dnode handles.
3725 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3727 mutex_enter(&db->db_mtx);
3728 dbuf_rele_and_unlock(db, tag, B_FALSE);
3732 dmu_buf_rele(dmu_buf_t *db, void *tag)
3734 dbuf_rele((dmu_buf_impl_t *)db, tag);
3738 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3739 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3740 * argument should be set if we are already in the dbuf-evicting code
3741 * path, in which case we don't want to recursively evict. This allows us to
3742 * avoid deeply nested stacks that would have a call flow similar to this:
3744 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3747 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3751 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
3756 ASSERT(MUTEX_HELD(&db->db_mtx));
3760 * Remove the reference to the dbuf before removing its hold on the
3761 * dnode so we can guarantee in dnode_move() that a referenced bonus
3762 * buffer has a corresponding dnode hold.
3764 holds = zfs_refcount_remove(&db->db_holds, tag);
3768 * We can't freeze indirects if there is a possibility that they
3769 * may be modified in the current syncing context.
3771 if (db->db_buf != NULL &&
3772 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3773 arc_buf_freeze(db->db_buf);
3776 if (holds == db->db_dirtycnt &&
3777 db->db_level == 0 && db->db_user_immediate_evict)
3778 dbuf_evict_user(db);
3781 if (db->db_blkid == DMU_BONUS_BLKID) {
3783 boolean_t evict_dbuf = db->db_pending_evict;
3786 * If the dnode moves here, we cannot cross this
3787 * barrier until the move completes.
3792 atomic_dec_32(&dn->dn_dbufs_count);
3795 * Decrementing the dbuf count means that the bonus
3796 * buffer's dnode hold is no longer discounted in
3797 * dnode_move(). The dnode cannot move until after
3798 * the dnode_rele() below.
3803 * Do not reference db after its lock is dropped.
3804 * Another thread may evict it.
3806 mutex_exit(&db->db_mtx);
3809 dnode_evict_bonus(dn);
3812 } else if (db->db_buf == NULL) {
3814 * This is a special case: we never associated this
3815 * dbuf with any data allocated from the ARC.
3817 ASSERT(db->db_state == DB_UNCACHED ||
3818 db->db_state == DB_NOFILL);
3820 } else if (arc_released(db->db_buf)) {
3822 * This dbuf has anonymous data associated with it.
3826 boolean_t do_arc_evict = B_FALSE;
3828 spa_t *spa = dmu_objset_spa(db->db_objset);
3830 if (!DBUF_IS_CACHEABLE(db) &&
3831 db->db_blkptr != NULL &&
3832 !BP_IS_HOLE(db->db_blkptr) &&
3833 !BP_IS_EMBEDDED(db->db_blkptr)) {
3834 do_arc_evict = B_TRUE;
3835 bp = *db->db_blkptr;
3838 if (!DBUF_IS_CACHEABLE(db) ||
3839 db->db_pending_evict) {
3841 } else if (!multilist_link_active(&db->db_cache_link)) {
3842 ASSERT3U(db->db_caching_status, ==,
3845 dbuf_cached_state_t dcs =
3846 dbuf_include_in_metadata_cache(db) ?
3847 DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
3848 db->db_caching_status = dcs;
3850 multilist_insert(&dbuf_caches[dcs].cache, db);
3851 uint64_t db_size = db->db.db_size;
3852 size = zfs_refcount_add_many(
3853 &dbuf_caches[dcs].size, db_size, db);
3854 uint8_t db_level = db->db_level;
3855 mutex_exit(&db->db_mtx);
3857 if (dcs == DB_DBUF_METADATA_CACHE) {
3858 DBUF_STAT_BUMP(metadata_cache_count);
3860 metadata_cache_size_bytes_max,
3863 DBUF_STAT_BUMP(cache_count);
3864 DBUF_STAT_MAX(cache_size_bytes_max,
3866 DBUF_STAT_BUMP(cache_levels[db_level]);
3868 cache_levels_bytes[db_level],
3872 if (dcs == DB_DBUF_CACHE && !evicting)
3873 dbuf_evict_notify(size);
3877 arc_freed(spa, &bp);
3880 mutex_exit(&db->db_mtx);
3885 #pragma weak dmu_buf_refcount = dbuf_refcount
3887 dbuf_refcount(dmu_buf_impl_t *db)
3889 return (zfs_refcount_count(&db->db_holds));
3893 dmu_buf_user_refcount(dmu_buf_t *db_fake)
3896 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3898 mutex_enter(&db->db_mtx);
3899 ASSERT3U(zfs_refcount_count(&db->db_holds), >=, db->db_dirtycnt);
3900 holds = zfs_refcount_count(&db->db_holds) - db->db_dirtycnt;
3901 mutex_exit(&db->db_mtx);
3907 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3908 dmu_buf_user_t *new_user)
3910 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3912 mutex_enter(&db->db_mtx);
3913 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3914 if (db->db_user == old_user)
3915 db->db_user = new_user;
3917 old_user = db->db_user;
3918 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3919 mutex_exit(&db->db_mtx);
3925 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3927 return (dmu_buf_replace_user(db_fake, NULL, user));
3931 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3933 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3935 db->db_user_immediate_evict = TRUE;
3936 return (dmu_buf_set_user(db_fake, user));
3940 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3942 return (dmu_buf_replace_user(db_fake, user, NULL));
3946 dmu_buf_get_user(dmu_buf_t *db_fake)
3948 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3950 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3951 return (db->db_user);
3955 dmu_buf_user_evict_wait()
3957 taskq_wait(dbu_evict_taskq);
3961 dmu_buf_get_blkptr(dmu_buf_t *db)
3963 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3964 return (dbi->db_blkptr);
3968 dmu_buf_get_objset(dmu_buf_t *db)
3970 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3971 return (dbi->db_objset);
3975 dmu_buf_dnode_enter(dmu_buf_t *db)
3977 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3978 DB_DNODE_ENTER(dbi);
3979 return (DB_DNODE(dbi));
3983 dmu_buf_dnode_exit(dmu_buf_t *db)
3985 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3990 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3992 /* ASSERT(dmu_tx_is_syncing(tx) */
3993 ASSERT(MUTEX_HELD(&db->db_mtx));
3995 if (db->db_blkptr != NULL)
3998 if (db->db_blkid == DMU_SPILL_BLKID) {
3999 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
4000 BP_ZERO(db->db_blkptr);
4003 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
4005 * This buffer was allocated at a time when there was
4006 * no available blkptrs from the dnode, or it was
4007 * inappropriate to hook it in (i.e., nlevels mismatch).
4009 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
4010 ASSERT(db->db_parent == NULL);
4011 db->db_parent = dn->dn_dbuf;
4012 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
4015 dmu_buf_impl_t *parent = db->db_parent;
4016 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
4018 ASSERT(dn->dn_phys->dn_nlevels > 1);
4019 if (parent == NULL) {
4020 mutex_exit(&db->db_mtx);
4021 rw_enter(&dn->dn_struct_rwlock, RW_READER);
4022 parent = dbuf_hold_level(dn, db->db_level + 1,
4023 db->db_blkid >> epbs, db);
4024 rw_exit(&dn->dn_struct_rwlock);
4025 mutex_enter(&db->db_mtx);
4026 db->db_parent = parent;
4028 db->db_blkptr = (blkptr_t *)parent->db.db_data +
4029 (db->db_blkid & ((1ULL << epbs) - 1));
4035 dbuf_sync_bonus(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
4037 dmu_buf_impl_t *db = dr->dr_dbuf;
4038 void *data = dr->dt.dl.dr_data;
4040 ASSERT0(db->db_level);
4041 ASSERT(MUTEX_HELD(&db->db_mtx));
4042 ASSERT(db->db_blkid == DMU_BONUS_BLKID);
4043 ASSERT(data != NULL);
4045 dnode_t *dn = dr->dr_dnode;
4046 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
4047 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
4048 bcopy(data, DN_BONUS(dn->dn_phys), DN_MAX_BONUS_LEN(dn->dn_phys));
4050 dbuf_sync_leaf_verify_bonus_dnode(dr);
4052 dbuf_undirty_bonus(dr);
4053 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
4057 * When syncing out a blocks of dnodes, adjust the block to deal with
4058 * encryption. Normally, we make sure the block is decrypted before writing
4059 * it. If we have crypt params, then we are writing a raw (encrypted) block,
4060 * from a raw receive. In this case, set the ARC buf's crypt params so
4061 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
4064 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t *dr)
4067 dmu_buf_impl_t *db = dr->dr_dbuf;
4069 ASSERT(MUTEX_HELD(&db->db_mtx));
4070 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT);
4071 ASSERT3U(db->db_level, ==, 0);
4073 if (!db->db_objset->os_raw_receive && arc_is_encrypted(db->db_buf)) {
4074 zbookmark_phys_t zb;
4077 * Unfortunately, there is currently no mechanism for
4078 * syncing context to handle decryption errors. An error
4079 * here is only possible if an attacker maliciously
4080 * changed a dnode block and updated the associated
4081 * checksums going up the block tree.
4083 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
4084 db->db.db_object, db->db_level, db->db_blkid);
4085 err = arc_untransform(db->db_buf, db->db_objset->os_spa,
4088 panic("Invalid dnode block MAC");
4089 } else if (dr->dt.dl.dr_has_raw_params) {
4090 (void) arc_release(dr->dt.dl.dr_data, db);
4091 arc_convert_to_raw(dr->dt.dl.dr_data,
4092 dmu_objset_id(db->db_objset),
4093 dr->dt.dl.dr_byteorder, DMU_OT_DNODE,
4094 dr->dt.dl.dr_salt, dr->dt.dl.dr_iv, dr->dt.dl.dr_mac);
4099 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
4100 * is critical the we not allow the compiler to inline this function in to
4101 * dbuf_sync_list() thereby drastically bloating the stack usage.
4103 noinline static void
4104 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
4106 dmu_buf_impl_t *db = dr->dr_dbuf;
4107 dnode_t *dn = dr->dr_dnode;
4109 ASSERT(dmu_tx_is_syncing(tx));
4111 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
4113 mutex_enter(&db->db_mtx);
4115 ASSERT(db->db_level > 0);
4118 /* Read the block if it hasn't been read yet. */
4119 if (db->db_buf == NULL) {
4120 mutex_exit(&db->db_mtx);
4121 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
4122 mutex_enter(&db->db_mtx);
4124 ASSERT3U(db->db_state, ==, DB_CACHED);
4125 ASSERT(db->db_buf != NULL);
4127 /* Indirect block size must match what the dnode thinks it is. */
4128 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
4129 dbuf_check_blkptr(dn, db);
4131 /* Provide the pending dirty record to child dbufs */
4132 db->db_data_pending = dr;
4134 mutex_exit(&db->db_mtx);
4136 dbuf_write(dr, db->db_buf, tx);
4138 zio_t *zio = dr->dr_zio;
4139 mutex_enter(&dr->dt.di.dr_mtx);
4140 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
4141 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
4142 mutex_exit(&dr->dt.di.dr_mtx);
4147 * Verify that the size of the data in our bonus buffer does not exceed
4148 * its recorded size.
4150 * The purpose of this verification is to catch any cases in development
4151 * where the size of a phys structure (i.e space_map_phys_t) grows and,
4152 * due to incorrect feature management, older pools expect to read more
4153 * data even though they didn't actually write it to begin with.
4155 * For a example, this would catch an error in the feature logic where we
4156 * open an older pool and we expect to write the space map histogram of
4157 * a space map with size SPACE_MAP_SIZE_V0.
4160 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t *dr)
4163 dnode_t *dn = dr->dr_dnode;
4166 * Encrypted bonus buffers can have data past their bonuslen.
4167 * Skip the verification of these blocks.
4169 if (DMU_OT_IS_ENCRYPTED(dn->dn_bonustype))
4172 uint16_t bonuslen = dn->dn_phys->dn_bonuslen;
4173 uint16_t maxbonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
4174 ASSERT3U(bonuslen, <=, maxbonuslen);
4176 arc_buf_t *datap = dr->dt.dl.dr_data;
4177 char *datap_end = ((char *)datap) + bonuslen;
4178 char *datap_max = ((char *)datap) + maxbonuslen;
4180 /* ensure that everything is zero after our data */
4181 for (; datap_end < datap_max; datap_end++)
4182 ASSERT(*datap_end == 0);
4187 dbuf_lightweight_bp(dbuf_dirty_record_t *dr)
4189 /* This must be a lightweight dirty record. */
4190 ASSERT3P(dr->dr_dbuf, ==, NULL);
4191 dnode_t *dn = dr->dr_dnode;
4193 if (dn->dn_phys->dn_nlevels == 1) {
4194 VERIFY3U(dr->dt.dll.dr_blkid, <, dn->dn_phys->dn_nblkptr);
4195 return (&dn->dn_phys->dn_blkptr[dr->dt.dll.dr_blkid]);
4197 dmu_buf_impl_t *parent_db = dr->dr_parent->dr_dbuf;
4198 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
4199 VERIFY3U(parent_db->db_level, ==, 1);
4200 VERIFY3P(parent_db->db_dnode_handle->dnh_dnode, ==, dn);
4201 VERIFY3U(dr->dt.dll.dr_blkid >> epbs, ==, parent_db->db_blkid);
4202 blkptr_t *bp = parent_db->db.db_data;
4203 return (&bp[dr->dt.dll.dr_blkid & ((1 << epbs) - 1)]);
4208 dbuf_lightweight_ready(zio_t *zio)
4210 dbuf_dirty_record_t *dr = zio->io_private;
4211 blkptr_t *bp = zio->io_bp;
4213 if (zio->io_error != 0)
4216 dnode_t *dn = dr->dr_dnode;
4218 blkptr_t *bp_orig = dbuf_lightweight_bp(dr);
4219 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4220 int64_t delta = bp_get_dsize_sync(spa, bp) -
4221 bp_get_dsize_sync(spa, bp_orig);
4222 dnode_diduse_space(dn, delta);
4224 uint64_t blkid = dr->dt.dll.dr_blkid;
4225 mutex_enter(&dn->dn_mtx);
4226 if (blkid > dn->dn_phys->dn_maxblkid) {
4227 ASSERT0(dn->dn_objset->os_raw_receive);
4228 dn->dn_phys->dn_maxblkid = blkid;
4230 mutex_exit(&dn->dn_mtx);
4232 if (!BP_IS_EMBEDDED(bp)) {
4233 uint64_t fill = BP_IS_HOLE(bp) ? 0 : 1;
4234 BP_SET_FILL(bp, fill);
4237 dmu_buf_impl_t *parent_db;
4238 EQUIV(dr->dr_parent == NULL, dn->dn_phys->dn_nlevels == 1);
4239 if (dr->dr_parent == NULL) {
4240 parent_db = dn->dn_dbuf;
4242 parent_db = dr->dr_parent->dr_dbuf;
4244 rw_enter(&parent_db->db_rwlock, RW_WRITER);
4246 rw_exit(&parent_db->db_rwlock);
4250 dbuf_lightweight_physdone(zio_t *zio)
4252 dbuf_dirty_record_t *dr = zio->io_private;
4253 dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
4254 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
4257 * The callback will be called io_phys_children times. Retire one
4258 * portion of our dirty space each time we are called. Any rounding
4259 * error will be cleaned up by dbuf_lightweight_done().
4261 int delta = dr->dr_accounted / zio->io_phys_children;
4262 dsl_pool_undirty_space(dp, delta, zio->io_txg);
4266 dbuf_lightweight_done(zio_t *zio)
4268 dbuf_dirty_record_t *dr = zio->io_private;
4270 VERIFY0(zio->io_error);
4272 objset_t *os = dr->dr_dnode->dn_objset;
4273 dmu_tx_t *tx = os->os_synctx;
4275 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
4276 ASSERT(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4278 dsl_dataset_t *ds = os->os_dsl_dataset;
4279 (void) dsl_dataset_block_kill(ds, &zio->io_bp_orig, tx, B_TRUE);
4280 dsl_dataset_block_born(ds, zio->io_bp, tx);
4284 * See comment in dbuf_write_done().
4286 if (zio->io_phys_children == 0) {
4287 dsl_pool_undirty_space(dmu_objset_pool(os),
4288 dr->dr_accounted, zio->io_txg);
4290 dsl_pool_undirty_space(dmu_objset_pool(os),
4291 dr->dr_accounted % zio->io_phys_children, zio->io_txg);
4294 abd_free(dr->dt.dll.dr_abd);
4295 kmem_free(dr, sizeof (*dr));
4298 noinline static void
4299 dbuf_sync_lightweight(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
4301 dnode_t *dn = dr->dr_dnode;
4303 if (dn->dn_phys->dn_nlevels == 1) {
4306 pio = dr->dr_parent->dr_zio;
4309 zbookmark_phys_t zb = {
4310 .zb_objset = dmu_objset_id(dn->dn_objset),
4311 .zb_object = dn->dn_object,
4313 .zb_blkid = dr->dt.dll.dr_blkid,
4317 * See comment in dbuf_write(). This is so that zio->io_bp_orig
4318 * will have the old BP in dbuf_lightweight_done().
4320 dr->dr_bp_copy = *dbuf_lightweight_bp(dr);
4322 dr->dr_zio = zio_write(pio, dmu_objset_spa(dn->dn_objset),
4323 dmu_tx_get_txg(tx), &dr->dr_bp_copy, dr->dt.dll.dr_abd,
4324 dn->dn_datablksz, abd_get_size(dr->dt.dll.dr_abd),
4325 &dr->dt.dll.dr_props, dbuf_lightweight_ready, NULL,
4326 dbuf_lightweight_physdone, dbuf_lightweight_done, dr,
4327 ZIO_PRIORITY_ASYNC_WRITE,
4328 ZIO_FLAG_MUSTSUCCEED | dr->dt.dll.dr_flags, &zb);
4330 zio_nowait(dr->dr_zio);
4334 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
4335 * critical the we not allow the compiler to inline this function in to
4336 * dbuf_sync_list() thereby drastically bloating the stack usage.
4338 noinline static void
4339 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
4341 arc_buf_t **datap = &dr->dt.dl.dr_data;
4342 dmu_buf_impl_t *db = dr->dr_dbuf;
4343 dnode_t *dn = dr->dr_dnode;
4345 uint64_t txg = tx->tx_txg;
4347 ASSERT(dmu_tx_is_syncing(tx));
4349 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
4351 mutex_enter(&db->db_mtx);
4353 * To be synced, we must be dirtied. But we
4354 * might have been freed after the dirty.
4356 if (db->db_state == DB_UNCACHED) {
4357 /* This buffer has been freed since it was dirtied */
4358 ASSERT(db->db.db_data == NULL);
4359 } else if (db->db_state == DB_FILL) {
4360 /* This buffer was freed and is now being re-filled */
4361 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
4363 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
4367 if (db->db_blkid == DMU_SPILL_BLKID) {
4368 mutex_enter(&dn->dn_mtx);
4369 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
4371 * In the previous transaction group, the bonus buffer
4372 * was entirely used to store the attributes for the
4373 * dnode which overrode the dn_spill field. However,
4374 * when adding more attributes to the file a spill
4375 * block was required to hold the extra attributes.
4377 * Make sure to clear the garbage left in the dn_spill
4378 * field from the previous attributes in the bonus
4379 * buffer. Otherwise, after writing out the spill
4380 * block to the new allocated dva, it will free
4381 * the old block pointed to by the invalid dn_spill.
4383 db->db_blkptr = NULL;
4385 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
4386 mutex_exit(&dn->dn_mtx);
4390 * If this is a bonus buffer, simply copy the bonus data into the
4391 * dnode. It will be written out when the dnode is synced (and it
4392 * will be synced, since it must have been dirty for dbuf_sync to
4395 if (db->db_blkid == DMU_BONUS_BLKID) {
4396 ASSERT(dr->dr_dbuf == db);
4397 dbuf_sync_bonus(dr, tx);
4404 * This function may have dropped the db_mtx lock allowing a dmu_sync
4405 * operation to sneak in. As a result, we need to ensure that we
4406 * don't check the dr_override_state until we have returned from
4407 * dbuf_check_blkptr.
4409 dbuf_check_blkptr(dn, db);
4412 * If this buffer is in the middle of an immediate write,
4413 * wait for the synchronous IO to complete.
4415 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
4416 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
4417 cv_wait(&db->db_changed, &db->db_mtx);
4418 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
4422 * If this is a dnode block, ensure it is appropriately encrypted
4423 * or decrypted, depending on what we are writing to it this txg.
4425 if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT)
4426 dbuf_prepare_encrypted_dnode_leaf(dr);
4428 if (db->db_state != DB_NOFILL &&
4429 dn->dn_object != DMU_META_DNODE_OBJECT &&
4430 zfs_refcount_count(&db->db_holds) > 1 &&
4431 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
4432 *datap == db->db_buf) {
4434 * If this buffer is currently "in use" (i.e., there
4435 * are active holds and db_data still references it),
4436 * then make a copy before we start the write so that
4437 * any modifications from the open txg will not leak
4440 * NOTE: this copy does not need to be made for
4441 * objects only modified in the syncing context (e.g.
4442 * DNONE_DNODE blocks).
4444 int psize = arc_buf_size(*datap);
4445 int lsize = arc_buf_lsize(*datap);
4446 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
4447 enum zio_compress compress_type = arc_get_compression(*datap);
4448 uint8_t complevel = arc_get_complevel(*datap);
4450 if (arc_is_encrypted(*datap)) {
4451 boolean_t byteorder;
4452 uint8_t salt[ZIO_DATA_SALT_LEN];
4453 uint8_t iv[ZIO_DATA_IV_LEN];
4454 uint8_t mac[ZIO_DATA_MAC_LEN];
4456 arc_get_raw_params(*datap, &byteorder, salt, iv, mac);
4457 *datap = arc_alloc_raw_buf(os->os_spa, db,
4458 dmu_objset_id(os), byteorder, salt, iv, mac,
4459 dn->dn_type, psize, lsize, compress_type,
4461 } else if (compress_type != ZIO_COMPRESS_OFF) {
4462 ASSERT3U(type, ==, ARC_BUFC_DATA);
4463 *datap = arc_alloc_compressed_buf(os->os_spa, db,
4464 psize, lsize, compress_type, complevel);
4466 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
4468 bcopy(db->db.db_data, (*datap)->b_data, psize);
4470 db->db_data_pending = dr;
4472 mutex_exit(&db->db_mtx);
4474 dbuf_write(dr, *datap, tx);
4476 ASSERT(!list_link_active(&dr->dr_dirty_node));
4477 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
4478 list_insert_tail(&dn->dn_dirty_records[txg & TXG_MASK], dr);
4480 zio_nowait(dr->dr_zio);
4485 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
4487 dbuf_dirty_record_t *dr;
4489 while ((dr = list_head(list))) {
4490 if (dr->dr_zio != NULL) {
4492 * If we find an already initialized zio then we
4493 * are processing the meta-dnode, and we have finished.
4494 * The dbufs for all dnodes are put back on the list
4495 * during processing, so that we can zio_wait()
4496 * these IOs after initiating all child IOs.
4498 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
4499 DMU_META_DNODE_OBJECT);
4502 list_remove(list, dr);
4503 if (dr->dr_dbuf == NULL) {
4504 dbuf_sync_lightweight(dr, tx);
4506 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
4507 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
4508 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
4510 if (dr->dr_dbuf->db_level > 0)
4511 dbuf_sync_indirect(dr, tx);
4513 dbuf_sync_leaf(dr, tx);
4519 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
4522 dmu_buf_impl_t *db = vdb;
4524 blkptr_t *bp = zio->io_bp;
4525 blkptr_t *bp_orig = &zio->io_bp_orig;
4526 spa_t *spa = zio->io_spa;
4531 ASSERT3P(db->db_blkptr, !=, NULL);
4532 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
4536 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
4537 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
4538 zio->io_prev_space_delta = delta;
4540 if (bp->blk_birth != 0) {
4541 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
4542 BP_GET_TYPE(bp) == dn->dn_type) ||
4543 (db->db_blkid == DMU_SPILL_BLKID &&
4544 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
4545 BP_IS_EMBEDDED(bp));
4546 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
4549 mutex_enter(&db->db_mtx);
4552 if (db->db_blkid == DMU_SPILL_BLKID) {
4553 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
4554 ASSERT(!(BP_IS_HOLE(bp)) &&
4555 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
4559 if (db->db_level == 0) {
4560 mutex_enter(&dn->dn_mtx);
4561 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
4562 db->db_blkid != DMU_SPILL_BLKID) {
4563 ASSERT0(db->db_objset->os_raw_receive);
4564 dn->dn_phys->dn_maxblkid = db->db_blkid;
4566 mutex_exit(&dn->dn_mtx);
4568 if (dn->dn_type == DMU_OT_DNODE) {
4570 while (i < db->db.db_size) {
4572 (void *)(((char *)db->db.db_data) + i);
4574 i += DNODE_MIN_SIZE;
4575 if (dnp->dn_type != DMU_OT_NONE) {
4577 i += dnp->dn_extra_slots *
4582 if (BP_IS_HOLE(bp)) {
4589 blkptr_t *ibp = db->db.db_data;
4590 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
4591 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
4592 if (BP_IS_HOLE(ibp))
4594 fill += BP_GET_FILL(ibp);
4599 if (!BP_IS_EMBEDDED(bp))
4600 BP_SET_FILL(bp, fill);
4602 mutex_exit(&db->db_mtx);
4604 db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_WRITER, FTAG);
4605 *db->db_blkptr = *bp;
4606 dmu_buf_unlock_parent(db, dblt, FTAG);
4610 * This function gets called just prior to running through the compression
4611 * stage of the zio pipeline. If we're an indirect block comprised of only
4612 * holes, then we want this indirect to be compressed away to a hole. In
4613 * order to do that we must zero out any information about the holes that
4614 * this indirect points to prior to before we try to compress it.
4617 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
4619 (void) zio, (void) buf;
4620 dmu_buf_impl_t *db = vdb;
4623 unsigned int epbs, i;
4625 ASSERT3U(db->db_level, >, 0);
4628 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
4629 ASSERT3U(epbs, <, 31);
4631 /* Determine if all our children are holes */
4632 for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) {
4633 if (!BP_IS_HOLE(bp))
4638 * If all the children are holes, then zero them all out so that
4639 * we may get compressed away.
4641 if (i == 1ULL << epbs) {
4643 * We only found holes. Grab the rwlock to prevent
4644 * anybody from reading the blocks we're about to
4647 rw_enter(&db->db_rwlock, RW_WRITER);
4648 bzero(db->db.db_data, db->db.db_size);
4649 rw_exit(&db->db_rwlock);
4655 * The SPA will call this callback several times for each zio - once
4656 * for every physical child i/o (zio->io_phys_children times). This
4657 * allows the DMU to monitor the progress of each logical i/o. For example,
4658 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
4659 * block. There may be a long delay before all copies/fragments are completed,
4660 * so this callback allows us to retire dirty space gradually, as the physical
4664 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
4667 dmu_buf_impl_t *db = arg;
4668 objset_t *os = db->db_objset;
4669 dsl_pool_t *dp = dmu_objset_pool(os);
4670 dbuf_dirty_record_t *dr;
4673 dr = db->db_data_pending;
4674 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
4677 * The callback will be called io_phys_children times. Retire one
4678 * portion of our dirty space each time we are called. Any rounding
4679 * error will be cleaned up by dbuf_write_done().
4681 delta = dr->dr_accounted / zio->io_phys_children;
4682 dsl_pool_undirty_space(dp, delta, zio->io_txg);
4686 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
4689 dmu_buf_impl_t *db = vdb;
4690 blkptr_t *bp_orig = &zio->io_bp_orig;
4691 blkptr_t *bp = db->db_blkptr;
4692 objset_t *os = db->db_objset;
4693 dmu_tx_t *tx = os->os_synctx;
4695 ASSERT0(zio->io_error);
4696 ASSERT(db->db_blkptr == bp);
4699 * For nopwrites and rewrites we ensure that the bp matches our
4700 * original and bypass all the accounting.
4702 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
4703 ASSERT(BP_EQUAL(bp, bp_orig));
4705 dsl_dataset_t *ds = os->os_dsl_dataset;
4706 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
4707 dsl_dataset_block_born(ds, bp, tx);
4710 mutex_enter(&db->db_mtx);
4714 dbuf_dirty_record_t *dr = db->db_data_pending;
4715 dnode_t *dn = dr->dr_dnode;
4716 ASSERT(!list_link_active(&dr->dr_dirty_node));
4717 ASSERT(dr->dr_dbuf == db);
4718 ASSERT(list_next(&db->db_dirty_records, dr) == NULL);
4719 list_remove(&db->db_dirty_records, dr);
4722 if (db->db_blkid == DMU_SPILL_BLKID) {
4723 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
4724 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
4725 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
4729 if (db->db_level == 0) {
4730 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
4731 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
4732 if (db->db_state != DB_NOFILL) {
4733 if (dr->dt.dl.dr_data != db->db_buf)
4734 arc_buf_destroy(dr->dt.dl.dr_data, db);
4737 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
4738 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
4739 if (!BP_IS_HOLE(db->db_blkptr)) {
4740 int epbs __maybe_unused = dn->dn_phys->dn_indblkshift -
4742 ASSERT3U(db->db_blkid, <=,
4743 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
4744 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
4747 mutex_destroy(&dr->dt.di.dr_mtx);
4748 list_destroy(&dr->dt.di.dr_children);
4751 cv_broadcast(&db->db_changed);
4752 ASSERT(db->db_dirtycnt > 0);
4753 db->db_dirtycnt -= 1;
4754 db->db_data_pending = NULL;
4755 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
4758 * If we didn't do a physical write in this ZIO and we
4759 * still ended up here, it means that the space of the
4760 * dbuf that we just released (and undirtied) above hasn't
4761 * been marked as undirtied in the pool's accounting.
4763 * Thus, we undirty that space in the pool's view of the
4764 * world here. For physical writes this type of update
4765 * happens in dbuf_write_physdone().
4767 * If we did a physical write, cleanup any rounding errors
4768 * that came up due to writing multiple copies of a block
4769 * on disk [see dbuf_write_physdone()].
4771 if (zio->io_phys_children == 0) {
4772 dsl_pool_undirty_space(dmu_objset_pool(os),
4773 dr->dr_accounted, zio->io_txg);
4775 dsl_pool_undirty_space(dmu_objset_pool(os),
4776 dr->dr_accounted % zio->io_phys_children, zio->io_txg);
4779 kmem_free(dr, sizeof (dbuf_dirty_record_t));
4783 dbuf_write_nofill_ready(zio_t *zio)
4785 dbuf_write_ready(zio, NULL, zio->io_private);
4789 dbuf_write_nofill_done(zio_t *zio)
4791 dbuf_write_done(zio, NULL, zio->io_private);
4795 dbuf_write_override_ready(zio_t *zio)
4797 dbuf_dirty_record_t *dr = zio->io_private;
4798 dmu_buf_impl_t *db = dr->dr_dbuf;
4800 dbuf_write_ready(zio, NULL, db);
4804 dbuf_write_override_done(zio_t *zio)
4806 dbuf_dirty_record_t *dr = zio->io_private;
4807 dmu_buf_impl_t *db = dr->dr_dbuf;
4808 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
4810 mutex_enter(&db->db_mtx);
4811 if (!BP_EQUAL(zio->io_bp, obp)) {
4812 if (!BP_IS_HOLE(obp))
4813 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
4814 arc_release(dr->dt.dl.dr_data, db);
4816 mutex_exit(&db->db_mtx);
4818 dbuf_write_done(zio, NULL, db);
4820 if (zio->io_abd != NULL)
4821 abd_free(zio->io_abd);
4824 typedef struct dbuf_remap_impl_callback_arg {
4826 uint64_t drica_blk_birth;
4828 } dbuf_remap_impl_callback_arg_t;
4831 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
4834 dbuf_remap_impl_callback_arg_t *drica = arg;
4835 objset_t *os = drica->drica_os;
4836 spa_t *spa = dmu_objset_spa(os);
4837 dmu_tx_t *tx = drica->drica_tx;
4839 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4841 if (os == spa_meta_objset(spa)) {
4842 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
4844 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
4845 size, drica->drica_blk_birth, tx);
4850 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, krwlock_t *rw, dmu_tx_t *tx)
4852 blkptr_t bp_copy = *bp;
4853 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4854 dbuf_remap_impl_callback_arg_t drica;
4856 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4858 drica.drica_os = dn->dn_objset;
4859 drica.drica_blk_birth = bp->blk_birth;
4860 drica.drica_tx = tx;
4861 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
4864 * If the blkptr being remapped is tracked by a livelist,
4865 * then we need to make sure the livelist reflects the update.
4866 * First, cancel out the old blkptr by appending a 'FREE'
4867 * entry. Next, add an 'ALLOC' to track the new version. This
4868 * way we avoid trying to free an inaccurate blkptr at delete.
4869 * Note that embedded blkptrs are not tracked in livelists.
4871 if (dn->dn_objset != spa_meta_objset(spa)) {
4872 dsl_dataset_t *ds = dmu_objset_ds(dn->dn_objset);
4873 if (dsl_deadlist_is_open(&ds->ds_dir->dd_livelist) &&
4874 bp->blk_birth > ds->ds_dir->dd_origin_txg) {
4875 ASSERT(!BP_IS_EMBEDDED(bp));
4876 ASSERT(dsl_dir_is_clone(ds->ds_dir));
4877 ASSERT(spa_feature_is_enabled(spa,
4878 SPA_FEATURE_LIVELIST));
4879 bplist_append(&ds->ds_dir->dd_pending_frees,
4881 bplist_append(&ds->ds_dir->dd_pending_allocs,
4887 * The db_rwlock prevents dbuf_read_impl() from
4888 * dereferencing the BP while we are changing it. To
4889 * avoid lock contention, only grab it when we are actually
4893 rw_enter(rw, RW_WRITER);
4901 * Remap any existing BP's to concrete vdevs, if possible.
4904 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
4906 spa_t *spa = dmu_objset_spa(db->db_objset);
4907 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4909 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
4912 if (db->db_level > 0) {
4913 blkptr_t *bp = db->db.db_data;
4914 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4915 dbuf_remap_impl(dn, &bp[i], &db->db_rwlock, tx);
4917 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
4918 dnode_phys_t *dnp = db->db.db_data;
4919 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
4921 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT;
4922 i += dnp[i].dn_extra_slots + 1) {
4923 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
4924 krwlock_t *lock = (dn->dn_dbuf == NULL ? NULL :
4925 &dn->dn_dbuf->db_rwlock);
4926 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], lock,
4934 /* Issue I/O to commit a dirty buffer to disk. */
4936 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4938 dmu_buf_impl_t *db = dr->dr_dbuf;
4939 dnode_t *dn = dr->dr_dnode;
4941 dmu_buf_impl_t *parent = db->db_parent;
4942 uint64_t txg = tx->tx_txg;
4943 zbookmark_phys_t zb;
4945 zio_t *pio; /* parent I/O */
4948 ASSERT(dmu_tx_is_syncing(tx));
4952 if (db->db_state != DB_NOFILL) {
4953 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4955 * Private object buffers are released here rather
4956 * than in dbuf_dirty() since they are only modified
4957 * in the syncing context and we don't want the
4958 * overhead of making multiple copies of the data.
4960 if (BP_IS_HOLE(db->db_blkptr)) {
4963 dbuf_release_bp(db);
4965 dbuf_remap(dn, db, tx);
4969 if (parent != dn->dn_dbuf) {
4970 /* Our parent is an indirect block. */
4971 /* We have a dirty parent that has been scheduled for write. */
4972 ASSERT(parent && parent->db_data_pending);
4973 /* Our parent's buffer is one level closer to the dnode. */
4974 ASSERT(db->db_level == parent->db_level-1);
4976 * We're about to modify our parent's db_data by modifying
4977 * our block pointer, so the parent must be released.
4979 ASSERT(arc_released(parent->db_buf));
4980 pio = parent->db_data_pending->dr_zio;
4982 /* Our parent is the dnode itself. */
4983 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4984 db->db_blkid != DMU_SPILL_BLKID) ||
4985 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4986 if (db->db_blkid != DMU_SPILL_BLKID)
4987 ASSERT3P(db->db_blkptr, ==,
4988 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4992 ASSERT(db->db_level == 0 || data == db->db_buf);
4993 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4996 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4997 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4998 db->db.db_object, db->db_level, db->db_blkid);
5000 if (db->db_blkid == DMU_SPILL_BLKID)
5002 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
5004 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
5007 * We copy the blkptr now (rather than when we instantiate the dirty
5008 * record), because its value can change between open context and
5009 * syncing context. We do not need to hold dn_struct_rwlock to read
5010 * db_blkptr because we are in syncing context.
5012 dr->dr_bp_copy = *db->db_blkptr;
5014 if (db->db_level == 0 &&
5015 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
5017 * The BP for this block has been provided by open context
5018 * (by dmu_sync() or dmu_buf_write_embedded()).
5020 abd_t *contents = (data != NULL) ?
5021 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
5023 dr->dr_zio = zio_write(pio, os->os_spa, txg, &dr->dr_bp_copy,
5024 contents, db->db.db_size, db->db.db_size, &zp,
5025 dbuf_write_override_ready, NULL, NULL,
5026 dbuf_write_override_done,
5027 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
5028 mutex_enter(&db->db_mtx);
5029 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
5030 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
5031 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
5032 mutex_exit(&db->db_mtx);
5033 } else if (db->db_state == DB_NOFILL) {
5034 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
5035 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
5036 dr->dr_zio = zio_write(pio, os->os_spa, txg,
5037 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
5038 dbuf_write_nofill_ready, NULL, NULL,
5039 dbuf_write_nofill_done, db,
5040 ZIO_PRIORITY_ASYNC_WRITE,
5041 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
5043 ASSERT(arc_released(data));
5046 * For indirect blocks, we want to setup the children
5047 * ready callback so that we can properly handle an indirect
5048 * block that only contains holes.
5050 arc_write_done_func_t *children_ready_cb = NULL;
5051 if (db->db_level != 0)
5052 children_ready_cb = dbuf_write_children_ready;
5054 dr->dr_zio = arc_write(pio, os->os_spa, txg,
5055 &dr->dr_bp_copy, data, dbuf_is_l2cacheable(db),
5056 &zp, dbuf_write_ready,
5057 children_ready_cb, dbuf_write_physdone,
5058 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
5059 ZIO_FLAG_MUSTSUCCEED, &zb);
5063 EXPORT_SYMBOL(dbuf_find);
5064 EXPORT_SYMBOL(dbuf_is_metadata);
5065 EXPORT_SYMBOL(dbuf_destroy);
5066 EXPORT_SYMBOL(dbuf_loan_arcbuf);
5067 EXPORT_SYMBOL(dbuf_whichblock);
5068 EXPORT_SYMBOL(dbuf_read);
5069 EXPORT_SYMBOL(dbuf_unoverride);
5070 EXPORT_SYMBOL(dbuf_free_range);
5071 EXPORT_SYMBOL(dbuf_new_size);
5072 EXPORT_SYMBOL(dbuf_release_bp);
5073 EXPORT_SYMBOL(dbuf_dirty);
5074 EXPORT_SYMBOL(dmu_buf_set_crypt_params);
5075 EXPORT_SYMBOL(dmu_buf_will_dirty);
5076 EXPORT_SYMBOL(dmu_buf_is_dirty);
5077 EXPORT_SYMBOL(dmu_buf_will_not_fill);
5078 EXPORT_SYMBOL(dmu_buf_will_fill);
5079 EXPORT_SYMBOL(dmu_buf_fill_done);
5080 EXPORT_SYMBOL(dmu_buf_rele);
5081 EXPORT_SYMBOL(dbuf_assign_arcbuf);
5082 EXPORT_SYMBOL(dbuf_prefetch);
5083 EXPORT_SYMBOL(dbuf_hold_impl);
5084 EXPORT_SYMBOL(dbuf_hold);
5085 EXPORT_SYMBOL(dbuf_hold_level);
5086 EXPORT_SYMBOL(dbuf_create_bonus);
5087 EXPORT_SYMBOL(dbuf_spill_set_blksz);
5088 EXPORT_SYMBOL(dbuf_rm_spill);
5089 EXPORT_SYMBOL(dbuf_add_ref);
5090 EXPORT_SYMBOL(dbuf_rele);
5091 EXPORT_SYMBOL(dbuf_rele_and_unlock);
5092 EXPORT_SYMBOL(dbuf_refcount);
5093 EXPORT_SYMBOL(dbuf_sync_list);
5094 EXPORT_SYMBOL(dmu_buf_set_user);
5095 EXPORT_SYMBOL(dmu_buf_set_user_ie);
5096 EXPORT_SYMBOL(dmu_buf_get_user);
5097 EXPORT_SYMBOL(dmu_buf_get_blkptr);
5100 ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, max_bytes, ULONG, ZMOD_RW,
5101 "Maximum size in bytes of the dbuf cache.");
5103 ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, hiwater_pct, UINT, ZMOD_RW,
5104 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
5107 ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, lowater_pct, UINT, ZMOD_RW,
5108 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
5111 ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_max_bytes, ULONG, ZMOD_RW,
5112 "Maximum size in bytes of the dbuf metadata cache.");
5114 ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, cache_shift, INT, ZMOD_RW,
5115 "Set the size of the dbuf cache to a log2 fraction of arc size.");
5117 ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_shift, INT, ZMOD_RW,
5118 "Set the size of the dbuf metadata cache to a log2 fraction of arc "