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 https://opensource.org/licenses/CDDL-1.0.
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
29 * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
32 #include <sys/zfs_context.h>
35 #include <sys/dmu_send.h>
36 #include <sys/dmu_impl.h>
38 #include <sys/dmu_objset.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/dsl_dir.h>
41 #include <sys/dmu_tx.h>
44 #include <sys/dmu_zfetch.h>
46 #include <sys/sa_impl.h>
47 #include <sys/zfeature.h>
48 #include <sys/blkptr.h>
49 #include <sys/range_tree.h>
50 #include <sys/trace_zfs.h>
51 #include <sys/callb.h>
56 #include <sys/spa_impl.h>
57 #include <sys/wmsum.h>
58 #include <sys/vdev_impl.h>
60 static kstat_t
*dbuf_ksp
;
62 typedef struct dbuf_stats
{
64 * Various statistics about the size of the dbuf cache.
66 kstat_named_t cache_count
;
67 kstat_named_t cache_size_bytes
;
68 kstat_named_t cache_size_bytes_max
;
70 * Statistics regarding the bounds on the dbuf cache size.
72 kstat_named_t cache_target_bytes
;
73 kstat_named_t cache_lowater_bytes
;
74 kstat_named_t cache_hiwater_bytes
;
76 * Total number of dbuf cache evictions that have occurred.
78 kstat_named_t cache_total_evicts
;
80 * The distribution of dbuf levels in the dbuf cache and
81 * the total size of all dbufs at each level.
83 kstat_named_t cache_levels
[DN_MAX_LEVELS
];
84 kstat_named_t cache_levels_bytes
[DN_MAX_LEVELS
];
86 * Statistics about the dbuf hash table.
88 kstat_named_t hash_hits
;
89 kstat_named_t hash_misses
;
90 kstat_named_t hash_collisions
;
91 kstat_named_t hash_elements
;
92 kstat_named_t hash_elements_max
;
94 * Number of sublists containing more than one dbuf in the dbuf
95 * hash table. Keep track of the longest hash chain.
97 kstat_named_t hash_chains
;
98 kstat_named_t hash_chain_max
;
100 * Number of times a dbuf_create() discovers that a dbuf was
101 * already created and in the dbuf hash table.
103 kstat_named_t hash_insert_race
;
105 * Number of entries in the hash table dbuf and mutex arrays.
107 kstat_named_t hash_table_count
;
108 kstat_named_t hash_mutex_count
;
110 * Statistics about the size of the metadata dbuf cache.
112 kstat_named_t metadata_cache_count
;
113 kstat_named_t metadata_cache_size_bytes
;
114 kstat_named_t metadata_cache_size_bytes_max
;
116 * For diagnostic purposes, this is incremented whenever we can't add
117 * something to the metadata cache because it's full, and instead put
118 * the data in the regular dbuf cache.
120 kstat_named_t metadata_cache_overflow
;
123 dbuf_stats_t dbuf_stats
= {
124 { "cache_count", KSTAT_DATA_UINT64
},
125 { "cache_size_bytes", KSTAT_DATA_UINT64
},
126 { "cache_size_bytes_max", KSTAT_DATA_UINT64
},
127 { "cache_target_bytes", KSTAT_DATA_UINT64
},
128 { "cache_lowater_bytes", KSTAT_DATA_UINT64
},
129 { "cache_hiwater_bytes", KSTAT_DATA_UINT64
},
130 { "cache_total_evicts", KSTAT_DATA_UINT64
},
131 { { "cache_levels_N", KSTAT_DATA_UINT64
} },
132 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64
} },
133 { "hash_hits", KSTAT_DATA_UINT64
},
134 { "hash_misses", KSTAT_DATA_UINT64
},
135 { "hash_collisions", KSTAT_DATA_UINT64
},
136 { "hash_elements", KSTAT_DATA_UINT64
},
137 { "hash_elements_max", KSTAT_DATA_UINT64
},
138 { "hash_chains", KSTAT_DATA_UINT64
},
139 { "hash_chain_max", KSTAT_DATA_UINT64
},
140 { "hash_insert_race", KSTAT_DATA_UINT64
},
141 { "hash_table_count", KSTAT_DATA_UINT64
},
142 { "hash_mutex_count", KSTAT_DATA_UINT64
},
143 { "metadata_cache_count", KSTAT_DATA_UINT64
},
144 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64
},
145 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64
},
146 { "metadata_cache_overflow", KSTAT_DATA_UINT64
}
151 wmsum_t cache_total_evicts
;
152 wmsum_t cache_levels
[DN_MAX_LEVELS
];
153 wmsum_t cache_levels_bytes
[DN_MAX_LEVELS
];
156 wmsum_t hash_collisions
;
158 wmsum_t hash_insert_race
;
159 wmsum_t metadata_cache_count
;
160 wmsum_t metadata_cache_overflow
;
163 #define DBUF_STAT_INCR(stat, val) \
164 wmsum_add(&dbuf_sums.stat, val);
165 #define DBUF_STAT_DECR(stat, val) \
166 DBUF_STAT_INCR(stat, -(val));
167 #define DBUF_STAT_BUMP(stat) \
168 DBUF_STAT_INCR(stat, 1);
169 #define DBUF_STAT_BUMPDOWN(stat) \
170 DBUF_STAT_INCR(stat, -1);
171 #define DBUF_STAT_MAX(stat, v) { \
173 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
174 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
178 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
179 static void dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
);
180 static int dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
);
183 * Global data structures and functions for the dbuf cache.
185 static kmem_cache_t
*dbuf_kmem_cache
;
186 static taskq_t
*dbu_evict_taskq
;
188 static kthread_t
*dbuf_cache_evict_thread
;
189 static kmutex_t dbuf_evict_lock
;
190 static kcondvar_t dbuf_evict_cv
;
191 static boolean_t dbuf_evict_thread_exit
;
194 * There are two dbuf caches; each dbuf can only be in one of them at a time.
196 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
197 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
198 * that represent the metadata that describes filesystems/snapshots/
199 * bookmarks/properties/etc. We only evict from this cache when we export a
200 * pool, to short-circuit as much I/O as possible for all administrative
201 * commands that need the metadata. There is no eviction policy for this
202 * cache, because we try to only include types in it which would occupy a
203 * very small amount of space per object but create a large impact on the
204 * performance of these commands. Instead, after it reaches a maximum size
205 * (which should only happen on very small memory systems with a very large
206 * number of filesystem objects), we stop taking new dbufs into the
207 * metadata cache, instead putting them in the normal dbuf cache.
209 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
210 * are not currently held but have been recently released. These dbufs
211 * are not eligible for arc eviction until they are aged out of the cache.
212 * Dbufs that are aged out of the cache will be immediately destroyed and
213 * become eligible for arc eviction.
215 * Dbufs are added to these caches once the last hold is released. If a dbuf is
216 * later accessed and still exists in the dbuf cache, then it will be removed
217 * from the cache and later re-added to the head of the cache.
219 * If a given dbuf meets the requirements for the metadata cache, it will go
220 * there, otherwise it will be considered for the generic LRU dbuf cache. The
221 * caches and the refcounts tracking their sizes are stored in an array indexed
222 * by those caches' matching enum values (from dbuf_cached_state_t).
224 typedef struct dbuf_cache
{
226 zfs_refcount_t size ____cacheline_aligned
;
228 dbuf_cache_t dbuf_caches
[DB_CACHE_MAX
];
230 /* Size limits for the caches */
231 static uint64_t dbuf_cache_max_bytes
= UINT64_MAX
;
232 static uint64_t dbuf_metadata_cache_max_bytes
= UINT64_MAX
;
234 /* Set the default sizes of the caches to log2 fraction of arc size */
235 static uint_t dbuf_cache_shift
= 5;
236 static uint_t dbuf_metadata_cache_shift
= 6;
238 /* Set the dbuf hash mutex count as log2 shift (dynamic by default) */
239 static uint_t dbuf_mutex_cache_shift
= 0;
241 static unsigned long dbuf_cache_target_bytes(void);
242 static unsigned long dbuf_metadata_cache_target_bytes(void);
245 * The LRU dbuf cache uses a three-stage eviction policy:
246 * - A low water marker designates when the dbuf eviction thread
247 * should stop evicting from the dbuf cache.
248 * - When we reach the maximum size (aka mid water mark), we
249 * signal the eviction thread to run.
250 * - The high water mark indicates when the eviction thread
251 * is unable to keep up with the incoming load and eviction must
252 * happen in the context of the calling thread.
256 * low water mid water hi water
257 * +----------------------------------------+----------+----------+
262 * +----------------------------------------+----------+----------+
264 * evicting eviction directly
267 * The high and low water marks indicate the operating range for the eviction
268 * thread. The low water mark is, by default, 90% of the total size of the
269 * cache and the high water mark is at 110% (both of these percentages can be
270 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
271 * respectively). The eviction thread will try to ensure that the cache remains
272 * within this range by waking up every second and checking if the cache is
273 * above the low water mark. The thread can also be woken up by callers adding
274 * elements into the cache if the cache is larger than the mid water (i.e max
275 * cache size). Once the eviction thread is woken up and eviction is required,
276 * it will continue evicting buffers until it's able to reduce the cache size
277 * to the low water mark. If the cache size continues to grow and hits the high
278 * water mark, then callers adding elements to the cache will begin to evict
279 * directly from the cache until the cache is no longer above the high water
284 * The percentage above and below the maximum cache size.
286 static uint_t dbuf_cache_hiwater_pct
= 10;
287 static uint_t dbuf_cache_lowater_pct
= 10;
290 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
292 (void) unused
, (void) kmflag
;
293 dmu_buf_impl_t
*db
= vdb
;
294 memset(db
, 0, sizeof (dmu_buf_impl_t
));
296 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
297 rw_init(&db
->db_rwlock
, NULL
, RW_DEFAULT
, NULL
);
298 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
299 multilist_link_init(&db
->db_cache_link
);
300 zfs_refcount_create(&db
->db_holds
);
306 dbuf_dest(void *vdb
, void *unused
)
309 dmu_buf_impl_t
*db
= vdb
;
310 mutex_destroy(&db
->db_mtx
);
311 rw_destroy(&db
->db_rwlock
);
312 cv_destroy(&db
->db_changed
);
313 ASSERT(!multilist_link_active(&db
->db_cache_link
));
314 zfs_refcount_destroy(&db
->db_holds
);
318 * dbuf hash table routines
320 static dbuf_hash_table_t dbuf_hash_table
;
323 * We use Cityhash for this. It's fast, and has good hash properties without
324 * requiring any large static buffers.
327 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
329 return (cityhash4((uintptr_t)os
, obj
, (uint64_t)lvl
, blkid
));
332 #define DTRACE_SET_STATE(db, why) \
333 DTRACE_PROBE2(dbuf__state_change, dmu_buf_impl_t *, db, \
336 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
337 ((dbuf)->db.db_object == (obj) && \
338 (dbuf)->db_objset == (os) && \
339 (dbuf)->db_level == (level) && \
340 (dbuf)->db_blkid == (blkid))
343 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
,
346 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
351 hv
= dbuf_hash(os
, obj
, level
, blkid
);
352 idx
= hv
& h
->hash_table_mask
;
354 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
355 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
356 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
357 mutex_enter(&db
->db_mtx
);
358 if (db
->db_state
!= DB_EVICTING
) {
359 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
362 mutex_exit(&db
->db_mtx
);
365 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
366 if (hash_out
!= NULL
)
371 static dmu_buf_impl_t
*
372 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
375 dmu_buf_impl_t
*db
= NULL
;
377 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
378 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
379 if (dn
->dn_bonus
!= NULL
) {
381 mutex_enter(&db
->db_mtx
);
383 rw_exit(&dn
->dn_struct_rwlock
);
384 dnode_rele(dn
, FTAG
);
390 * Insert an entry into the hash table. If there is already an element
391 * equal to elem in the hash table, then the already existing element
392 * will be returned and the new element will not be inserted.
393 * Otherwise returns NULL.
395 static dmu_buf_impl_t
*
396 dbuf_hash_insert(dmu_buf_impl_t
*db
)
398 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
399 objset_t
*os
= db
->db_objset
;
400 uint64_t obj
= db
->db
.db_object
;
401 int level
= db
->db_level
;
406 blkid
= db
->db_blkid
;
407 ASSERT3U(dbuf_hash(os
, obj
, level
, blkid
), ==, db
->db_hash
);
408 idx
= db
->db_hash
& h
->hash_table_mask
;
410 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
411 for (dbf
= h
->hash_table
[idx
], i
= 0; dbf
!= NULL
;
412 dbf
= dbf
->db_hash_next
, i
++) {
413 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
414 mutex_enter(&dbf
->db_mtx
);
415 if (dbf
->db_state
!= DB_EVICTING
) {
416 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
419 mutex_exit(&dbf
->db_mtx
);
424 DBUF_STAT_BUMP(hash_collisions
);
426 DBUF_STAT_BUMP(hash_chains
);
428 DBUF_STAT_MAX(hash_chain_max
, i
);
431 mutex_enter(&db
->db_mtx
);
432 db
->db_hash_next
= h
->hash_table
[idx
];
433 h
->hash_table
[idx
] = db
;
434 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
435 uint64_t he
= atomic_inc_64_nv(&dbuf_stats
.hash_elements
.value
.ui64
);
436 DBUF_STAT_MAX(hash_elements_max
, he
);
442 * This returns whether this dbuf should be stored in the metadata cache, which
443 * is based on whether it's from one of the dnode types that store data related
444 * to traversing dataset hierarchies.
447 dbuf_include_in_metadata_cache(dmu_buf_impl_t
*db
)
450 dmu_object_type_t type
= DB_DNODE(db
)->dn_type
;
453 /* Check if this dbuf is one of the types we care about */
454 if (DMU_OT_IS_METADATA_CACHED(type
)) {
455 /* If we hit this, then we set something up wrong in dmu_ot */
456 ASSERT(DMU_OT_IS_METADATA(type
));
459 * Sanity check for small-memory systems: don't allocate too
460 * much memory for this purpose.
462 if (zfs_refcount_count(
463 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
) >
464 dbuf_metadata_cache_target_bytes()) {
465 DBUF_STAT_BUMP(metadata_cache_overflow
);
476 * Remove an entry from the hash table. It must be in the EVICTING state.
479 dbuf_hash_remove(dmu_buf_impl_t
*db
)
481 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
483 dmu_buf_impl_t
*dbf
, **dbp
;
485 ASSERT3U(dbuf_hash(db
->db_objset
, db
->db
.db_object
, db
->db_level
,
486 db
->db_blkid
), ==, db
->db_hash
);
487 idx
= db
->db_hash
& h
->hash_table_mask
;
490 * We mustn't hold db_mtx to maintain lock ordering:
491 * DBUF_HASH_MUTEX > db_mtx.
493 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
494 ASSERT(db
->db_state
== DB_EVICTING
);
495 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
497 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
498 dbp
= &h
->hash_table
[idx
];
499 while ((dbf
= *dbp
) != db
) {
500 dbp
= &dbf
->db_hash_next
;
503 *dbp
= db
->db_hash_next
;
504 db
->db_hash_next
= NULL
;
505 if (h
->hash_table
[idx
] &&
506 h
->hash_table
[idx
]->db_hash_next
== NULL
)
507 DBUF_STAT_BUMPDOWN(hash_chains
);
508 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
509 atomic_dec_64(&dbuf_stats
.hash_elements
.value
.ui64
);
515 } dbvu_verify_type_t
;
518 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
523 if (db
->db_user
== NULL
)
526 /* Only data blocks support the attachment of user data. */
527 ASSERT(db
->db_level
== 0);
529 /* Clients must resolve a dbuf before attaching user data. */
530 ASSERT(db
->db
.db_data
!= NULL
);
531 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
533 holds
= zfs_refcount_count(&db
->db_holds
);
534 if (verify_type
== DBVU_EVICTING
) {
536 * Immediate eviction occurs when holds == dirtycnt.
537 * For normal eviction buffers, holds is zero on
538 * eviction, except when dbuf_fix_old_data() calls
539 * dbuf_clear_data(). However, the hold count can grow
540 * during eviction even though db_mtx is held (see
541 * dmu_bonus_hold() for an example), so we can only
542 * test the generic invariant that holds >= dirtycnt.
544 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
546 if (db
->db_user_immediate_evict
== TRUE
)
547 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
549 ASSERT3U(holds
, >, 0);
555 dbuf_evict_user(dmu_buf_impl_t
*db
)
557 dmu_buf_user_t
*dbu
= db
->db_user
;
559 ASSERT(MUTEX_HELD(&db
->db_mtx
));
564 dbuf_verify_user(db
, DBVU_EVICTING
);
568 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
569 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
572 if (db
->db_caching_status
!= DB_NO_CACHE
) {
574 * This is a cached dbuf, so the size of the user data is
575 * included in its cached amount. We adjust it here because the
576 * user data has already been detached from the dbuf, and the
577 * sync functions are not supposed to touch it (the dbuf might
578 * not exist anymore by the time the sync functions run.
580 uint64_t size
= dbu
->dbu_size
;
581 (void) zfs_refcount_remove_many(
582 &dbuf_caches
[db
->db_caching_status
].size
, size
, db
);
583 if (db
->db_caching_status
== DB_DBUF_CACHE
)
584 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
], size
);
588 * There are two eviction callbacks - one that we call synchronously
589 * and one that we invoke via a taskq. The async one is useful for
590 * avoiding lock order reversals and limiting stack depth.
592 * Note that if we have a sync callback but no async callback,
593 * it's likely that the sync callback will free the structure
594 * containing the dbu. In that case we need to take care to not
595 * dereference dbu after calling the sync evict func.
597 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
599 if (dbu
->dbu_evict_func_sync
!= NULL
)
600 dbu
->dbu_evict_func_sync(dbu
);
603 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
604 dbu
, 0, &dbu
->dbu_tqent
);
609 dbuf_is_metadata(dmu_buf_impl_t
*db
)
612 * Consider indirect blocks and spill blocks to be meta data.
614 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
617 boolean_t is_metadata
;
620 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
623 return (is_metadata
);
628 * We want to exclude buffers that are on a special allocation class from
632 dbuf_is_l2cacheable(dmu_buf_impl_t
*db
)
634 if (db
->db_objset
->os_secondary_cache
== ZFS_CACHE_ALL
||
635 (db
->db_objset
->os_secondary_cache
==
636 ZFS_CACHE_METADATA
&& dbuf_is_metadata(db
))) {
637 if (l2arc_exclude_special
== 0)
640 blkptr_t
*bp
= db
->db_blkptr
;
641 if (bp
== NULL
|| BP_IS_HOLE(bp
))
643 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
644 vdev_t
*rvd
= db
->db_objset
->os_spa
->spa_root_vdev
;
647 if (vdev
< rvd
->vdev_children
)
648 vd
= rvd
->vdev_child
[vdev
];
653 if (vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
654 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
)
660 static inline boolean_t
661 dnode_level_is_l2cacheable(blkptr_t
*bp
, dnode_t
*dn
, int64_t level
)
663 if (dn
->dn_objset
->os_secondary_cache
== ZFS_CACHE_ALL
||
664 (dn
->dn_objset
->os_secondary_cache
== ZFS_CACHE_METADATA
&&
666 DMU_OT_IS_METADATA(dn
->dn_handle
->dnh_dnode
->dn_type
)))) {
667 if (l2arc_exclude_special
== 0)
670 if (bp
== NULL
|| BP_IS_HOLE(bp
))
672 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
673 vdev_t
*rvd
= dn
->dn_objset
->os_spa
->spa_root_vdev
;
676 if (vdev
< rvd
->vdev_children
)
677 vd
= rvd
->vdev_child
[vdev
];
682 if (vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
683 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
)
691 * This function *must* return indices evenly distributed between all
692 * sublists of the multilist. This is needed due to how the dbuf eviction
693 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
694 * distributed between all sublists and uses this assumption when
695 * deciding which sublist to evict from and how much to evict from it.
698 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
700 dmu_buf_impl_t
*db
= obj
;
703 * The assumption here, is the hash value for a given
704 * dmu_buf_impl_t will remain constant throughout it's lifetime
705 * (i.e. it's objset, object, level and blkid fields don't change).
706 * Thus, we don't need to store the dbuf's sublist index
707 * on insertion, as this index can be recalculated on removal.
709 * Also, the low order bits of the hash value are thought to be
710 * distributed evenly. Otherwise, in the case that the multilist
711 * has a power of two number of sublists, each sublists' usage
712 * would not be evenly distributed. In this context full 64bit
713 * division would be a waste of time, so limit it to 32 bits.
715 return ((unsigned int)dbuf_hash(db
->db_objset
, db
->db
.db_object
,
716 db
->db_level
, db
->db_blkid
) %
717 multilist_get_num_sublists(ml
));
721 * The target size of the dbuf cache can grow with the ARC target,
722 * unless limited by the tunable dbuf_cache_max_bytes.
724 static inline unsigned long
725 dbuf_cache_target_bytes(void)
727 return (MIN(dbuf_cache_max_bytes
,
728 arc_target_bytes() >> dbuf_cache_shift
));
732 * The target size of the dbuf metadata cache can grow with the ARC target,
733 * unless limited by the tunable dbuf_metadata_cache_max_bytes.
735 static inline unsigned long
736 dbuf_metadata_cache_target_bytes(void)
738 return (MIN(dbuf_metadata_cache_max_bytes
,
739 arc_target_bytes() >> dbuf_metadata_cache_shift
));
742 static inline uint64_t
743 dbuf_cache_hiwater_bytes(void)
745 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
746 return (dbuf_cache_target
+
747 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100);
750 static inline uint64_t
751 dbuf_cache_lowater_bytes(void)
753 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
754 return (dbuf_cache_target
-
755 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100);
758 static inline boolean_t
759 dbuf_cache_above_lowater(void)
761 return (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
762 dbuf_cache_lowater_bytes());
766 * Evict the oldest eligible dbuf from the dbuf cache.
771 int idx
= multilist_get_random_index(&dbuf_caches
[DB_DBUF_CACHE
].cache
);
772 multilist_sublist_t
*mls
= multilist_sublist_lock(
773 &dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
775 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
777 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
778 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
779 db
= multilist_sublist_prev(mls
, db
);
782 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
783 multilist_sublist_t
*, mls
);
786 multilist_sublist_remove(mls
, db
);
787 multilist_sublist_unlock(mls
);
788 uint64_t size
= db
->db
.db_size
+ dmu_buf_user_size(&db
->db
);
789 (void) zfs_refcount_remove_many(
790 &dbuf_caches
[DB_DBUF_CACHE
].size
, size
, db
);
791 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
792 DBUF_STAT_BUMPDOWN(cache_count
);
793 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
], size
);
794 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
795 db
->db_caching_status
= DB_NO_CACHE
;
797 DBUF_STAT_BUMP(cache_total_evicts
);
799 multilist_sublist_unlock(mls
);
804 * The dbuf evict thread is responsible for aging out dbufs from the
805 * cache. Once the cache has reached it's maximum size, dbufs are removed
806 * and destroyed. The eviction thread will continue running until the size
807 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
808 * out of the cache it is destroyed and becomes eligible for arc eviction.
810 static __attribute__((noreturn
)) void
811 dbuf_evict_thread(void *unused
)
816 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
818 mutex_enter(&dbuf_evict_lock
);
819 while (!dbuf_evict_thread_exit
) {
820 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
821 CALLB_CPR_SAFE_BEGIN(&cpr
);
822 (void) cv_timedwait_idle_hires(&dbuf_evict_cv
,
823 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
824 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
826 mutex_exit(&dbuf_evict_lock
);
829 * Keep evicting as long as we're above the low water mark
830 * for the cache. We do this without holding the locks to
831 * minimize lock contention.
833 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
837 mutex_enter(&dbuf_evict_lock
);
840 dbuf_evict_thread_exit
= B_FALSE
;
841 cv_broadcast(&dbuf_evict_cv
);
842 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
847 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
848 * If the dbuf cache is at its high water mark, then evict a dbuf from the
849 * dbuf cache using the caller's context.
852 dbuf_evict_notify(uint64_t size
)
855 * We check if we should evict without holding the dbuf_evict_lock,
856 * because it's OK to occasionally make the wrong decision here,
857 * and grabbing the lock results in massive lock contention.
859 if (size
> dbuf_cache_target_bytes()) {
860 if (size
> dbuf_cache_hiwater_bytes())
862 cv_signal(&dbuf_evict_cv
);
867 dbuf_kstat_update(kstat_t
*ksp
, int rw
)
869 dbuf_stats_t
*ds
= ksp
->ks_data
;
870 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
872 if (rw
== KSTAT_WRITE
)
873 return (SET_ERROR(EACCES
));
875 ds
->cache_count
.value
.ui64
=
876 wmsum_value(&dbuf_sums
.cache_count
);
877 ds
->cache_size_bytes
.value
.ui64
=
878 zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
);
879 ds
->cache_target_bytes
.value
.ui64
= dbuf_cache_target_bytes();
880 ds
->cache_hiwater_bytes
.value
.ui64
= dbuf_cache_hiwater_bytes();
881 ds
->cache_lowater_bytes
.value
.ui64
= dbuf_cache_lowater_bytes();
882 ds
->cache_total_evicts
.value
.ui64
=
883 wmsum_value(&dbuf_sums
.cache_total_evicts
);
884 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
885 ds
->cache_levels
[i
].value
.ui64
=
886 wmsum_value(&dbuf_sums
.cache_levels
[i
]);
887 ds
->cache_levels_bytes
[i
].value
.ui64
=
888 wmsum_value(&dbuf_sums
.cache_levels_bytes
[i
]);
890 ds
->hash_hits
.value
.ui64
=
891 wmsum_value(&dbuf_sums
.hash_hits
);
892 ds
->hash_misses
.value
.ui64
=
893 wmsum_value(&dbuf_sums
.hash_misses
);
894 ds
->hash_collisions
.value
.ui64
=
895 wmsum_value(&dbuf_sums
.hash_collisions
);
896 ds
->hash_chains
.value
.ui64
=
897 wmsum_value(&dbuf_sums
.hash_chains
);
898 ds
->hash_insert_race
.value
.ui64
=
899 wmsum_value(&dbuf_sums
.hash_insert_race
);
900 ds
->hash_table_count
.value
.ui64
= h
->hash_table_mask
+ 1;
901 ds
->hash_mutex_count
.value
.ui64
= h
->hash_mutex_mask
+ 1;
902 ds
->metadata_cache_count
.value
.ui64
=
903 wmsum_value(&dbuf_sums
.metadata_cache_count
);
904 ds
->metadata_cache_size_bytes
.value
.ui64
= zfs_refcount_count(
905 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
);
906 ds
->metadata_cache_overflow
.value
.ui64
=
907 wmsum_value(&dbuf_sums
.metadata_cache_overflow
);
914 uint64_t hmsize
, hsize
= 1ULL << 16;
915 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
918 * The hash table is big enough to fill one eighth of physical memory
919 * with an average block size of zfs_arc_average_blocksize (default 8K).
920 * By default, the table will take up
921 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
923 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory() / 8)
926 h
->hash_table
= NULL
;
927 while (h
->hash_table
== NULL
) {
928 h
->hash_table_mask
= hsize
- 1;
930 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
931 if (h
->hash_table
== NULL
)
934 ASSERT3U(hsize
, >=, 1ULL << 10);
938 * The hash table buckets are protected by an array of mutexes where
939 * each mutex is reponsible for protecting 128 buckets. A minimum
940 * array size of 8192 is targeted to avoid contention.
942 if (dbuf_mutex_cache_shift
== 0)
943 hmsize
= MAX(hsize
>> 7, 1ULL << 13);
945 hmsize
= 1ULL << MIN(dbuf_mutex_cache_shift
, 24);
947 h
->hash_mutexes
= NULL
;
948 while (h
->hash_mutexes
== NULL
) {
949 h
->hash_mutex_mask
= hmsize
- 1;
951 h
->hash_mutexes
= vmem_zalloc(hmsize
* sizeof (kmutex_t
),
953 if (h
->hash_mutexes
== NULL
)
957 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
958 sizeof (dmu_buf_impl_t
),
959 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
961 for (int i
= 0; i
< hmsize
; i
++)
962 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
967 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
968 * configuration is not required.
970 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
972 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
973 multilist_create(&dbuf_caches
[dcs
].cache
,
974 sizeof (dmu_buf_impl_t
),
975 offsetof(dmu_buf_impl_t
, db_cache_link
),
976 dbuf_cache_multilist_index_func
);
977 zfs_refcount_create(&dbuf_caches
[dcs
].size
);
980 dbuf_evict_thread_exit
= B_FALSE
;
981 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
982 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
983 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
984 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
986 wmsum_init(&dbuf_sums
.cache_count
, 0);
987 wmsum_init(&dbuf_sums
.cache_total_evicts
, 0);
988 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
989 wmsum_init(&dbuf_sums
.cache_levels
[i
], 0);
990 wmsum_init(&dbuf_sums
.cache_levels_bytes
[i
], 0);
992 wmsum_init(&dbuf_sums
.hash_hits
, 0);
993 wmsum_init(&dbuf_sums
.hash_misses
, 0);
994 wmsum_init(&dbuf_sums
.hash_collisions
, 0);
995 wmsum_init(&dbuf_sums
.hash_chains
, 0);
996 wmsum_init(&dbuf_sums
.hash_insert_race
, 0);
997 wmsum_init(&dbuf_sums
.metadata_cache_count
, 0);
998 wmsum_init(&dbuf_sums
.metadata_cache_overflow
, 0);
1000 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
1001 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
1002 KSTAT_FLAG_VIRTUAL
);
1003 if (dbuf_ksp
!= NULL
) {
1004 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
1005 snprintf(dbuf_stats
.cache_levels
[i
].name
,
1006 KSTAT_STRLEN
, "cache_level_%d", i
);
1007 dbuf_stats
.cache_levels
[i
].data_type
=
1009 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
1010 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
1011 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
1014 dbuf_ksp
->ks_data
= &dbuf_stats
;
1015 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
1016 kstat_install(dbuf_ksp
);
1023 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
1025 dbuf_stats_destroy();
1027 for (int i
= 0; i
< (h
->hash_mutex_mask
+ 1); i
++)
1028 mutex_destroy(&h
->hash_mutexes
[i
]);
1030 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
1031 vmem_free(h
->hash_mutexes
, (h
->hash_mutex_mask
+ 1) *
1034 kmem_cache_destroy(dbuf_kmem_cache
);
1035 taskq_destroy(dbu_evict_taskq
);
1037 mutex_enter(&dbuf_evict_lock
);
1038 dbuf_evict_thread_exit
= B_TRUE
;
1039 while (dbuf_evict_thread_exit
) {
1040 cv_signal(&dbuf_evict_cv
);
1041 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
1043 mutex_exit(&dbuf_evict_lock
);
1045 mutex_destroy(&dbuf_evict_lock
);
1046 cv_destroy(&dbuf_evict_cv
);
1048 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
1049 zfs_refcount_destroy(&dbuf_caches
[dcs
].size
);
1050 multilist_destroy(&dbuf_caches
[dcs
].cache
);
1053 if (dbuf_ksp
!= NULL
) {
1054 kstat_delete(dbuf_ksp
);
1058 wmsum_fini(&dbuf_sums
.cache_count
);
1059 wmsum_fini(&dbuf_sums
.cache_total_evicts
);
1060 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
1061 wmsum_fini(&dbuf_sums
.cache_levels
[i
]);
1062 wmsum_fini(&dbuf_sums
.cache_levels_bytes
[i
]);
1064 wmsum_fini(&dbuf_sums
.hash_hits
);
1065 wmsum_fini(&dbuf_sums
.hash_misses
);
1066 wmsum_fini(&dbuf_sums
.hash_collisions
);
1067 wmsum_fini(&dbuf_sums
.hash_chains
);
1068 wmsum_fini(&dbuf_sums
.hash_insert_race
);
1069 wmsum_fini(&dbuf_sums
.metadata_cache_count
);
1070 wmsum_fini(&dbuf_sums
.metadata_cache_overflow
);
1079 dbuf_verify(dmu_buf_impl_t
*db
)
1082 dbuf_dirty_record_t
*dr
;
1085 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1087 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
1090 ASSERT(db
->db_objset
!= NULL
);
1094 ASSERT(db
->db_parent
== NULL
);
1095 ASSERT(db
->db_blkptr
== NULL
);
1097 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
1098 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
1099 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
1100 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
1101 db
->db_blkid
== DMU_SPILL_BLKID
||
1102 !avl_is_empty(&dn
->dn_dbufs
));
1104 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1106 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1107 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
1108 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
1110 ASSERT0(db
->db
.db_offset
);
1112 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
1115 if ((dr
= list_head(&db
->db_dirty_records
)) != NULL
) {
1116 ASSERT(dr
->dr_dbuf
== db
);
1117 txg_prev
= dr
->dr_txg
;
1118 for (dr
= list_next(&db
->db_dirty_records
, dr
); dr
!= NULL
;
1119 dr
= list_next(&db
->db_dirty_records
, dr
)) {
1120 ASSERT(dr
->dr_dbuf
== db
);
1121 ASSERT(txg_prev
> dr
->dr_txg
);
1122 txg_prev
= dr
->dr_txg
;
1127 * We can't assert that db_size matches dn_datablksz because it
1128 * can be momentarily different when another thread is doing
1129 * dnode_set_blksz().
1131 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
1132 dr
= db
->db_data_pending
;
1134 * It should only be modified in syncing context, so
1135 * make sure we only have one copy of the data.
1137 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
1140 /* verify db->db_blkptr */
1141 if (db
->db_blkptr
) {
1142 if (db
->db_parent
== dn
->dn_dbuf
) {
1143 /* db is pointed to by the dnode */
1144 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1145 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
1146 ASSERT(db
->db_parent
== NULL
);
1148 ASSERT(db
->db_parent
!= NULL
);
1149 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
1150 ASSERT3P(db
->db_blkptr
, ==,
1151 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
1153 /* db is pointed to by an indirect block */
1154 int epb __maybe_unused
= db
->db_parent
->db
.db_size
>>
1156 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
1157 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
1160 * dnode_grow_indblksz() can make this fail if we don't
1161 * have the parent's rwlock. XXX indblksz no longer
1162 * grows. safe to do this now?
1164 if (RW_LOCK_HELD(&db
->db_parent
->db_rwlock
)) {
1165 ASSERT3P(db
->db_blkptr
, ==,
1166 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
1167 db
->db_blkid
% epb
));
1171 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
1172 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
1173 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
1174 db
->db_state
!= DB_FILL
&& (dn
== NULL
|| !dn
->dn_free_txg
)) {
1176 * If the blkptr isn't set but they have nonzero data,
1177 * it had better be dirty, otherwise we'll lose that
1178 * data when we evict this buffer.
1180 * There is an exception to this rule for indirect blocks; in
1181 * this case, if the indirect block is a hole, we fill in a few
1182 * fields on each of the child blocks (importantly, birth time)
1183 * to prevent hole birth times from being lost when you
1184 * partially fill in a hole.
1186 if (db
->db_dirtycnt
== 0) {
1187 if (db
->db_level
== 0) {
1188 uint64_t *buf
= db
->db
.db_data
;
1191 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
1192 ASSERT(buf
[i
] == 0);
1195 blkptr_t
*bps
= db
->db
.db_data
;
1196 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1199 * We want to verify that all the blkptrs in the
1200 * indirect block are holes, but we may have
1201 * automatically set up a few fields for them.
1202 * We iterate through each blkptr and verify
1203 * they only have those fields set.
1206 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1208 blkptr_t
*bp
= &bps
[i
];
1209 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1212 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1213 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1214 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1215 ASSERT0(bp
->blk_fill
);
1216 ASSERT0(bp
->blk_pad
[0]);
1217 ASSERT0(bp
->blk_pad
[1]);
1218 ASSERT(!BP_IS_EMBEDDED(bp
));
1219 ASSERT(BP_IS_HOLE(bp
));
1220 ASSERT0(bp
->blk_phys_birth
);
1230 dbuf_clear_data(dmu_buf_impl_t
*db
)
1232 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1233 dbuf_evict_user(db
);
1234 ASSERT3P(db
->db_buf
, ==, NULL
);
1235 db
->db
.db_data
= NULL
;
1236 if (db
->db_state
!= DB_NOFILL
) {
1237 db
->db_state
= DB_UNCACHED
;
1238 DTRACE_SET_STATE(db
, "clear data");
1243 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1245 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1246 ASSERT(buf
!= NULL
);
1249 ASSERT(buf
->b_data
!= NULL
);
1250 db
->db
.db_data
= buf
->b_data
;
1254 dbuf_alloc_arcbuf(dmu_buf_impl_t
*db
)
1256 spa_t
*spa
= db
->db_objset
->os_spa
;
1258 return (arc_alloc_buf(spa
, db
, DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
1262 * Loan out an arc_buf for read. Return the loaned arc_buf.
1265 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1269 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1270 mutex_enter(&db
->db_mtx
);
1271 if (arc_released(db
->db_buf
) || zfs_refcount_count(&db
->db_holds
) > 1) {
1272 int blksz
= db
->db
.db_size
;
1273 spa_t
*spa
= db
->db_objset
->os_spa
;
1275 mutex_exit(&db
->db_mtx
);
1276 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1277 memcpy(abuf
->b_data
, db
->db
.db_data
, blksz
);
1280 arc_loan_inuse_buf(abuf
, db
);
1282 dbuf_clear_data(db
);
1283 mutex_exit(&db
->db_mtx
);
1289 * Calculate which level n block references the data at the level 0 offset
1293 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1295 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1297 * The level n blkid is equal to the level 0 blkid divided by
1298 * the number of level 0s in a level n block.
1300 * The level 0 blkid is offset >> datablkshift =
1301 * offset / 2^datablkshift.
1303 * The number of level 0s in a level n is the number of block
1304 * pointers in an indirect block, raised to the power of level.
1305 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1306 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1308 * Thus, the level n blkid is: offset /
1309 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1310 * = offset / 2^(datablkshift + level *
1311 * (indblkshift - SPA_BLKPTRSHIFT))
1312 * = offset >> (datablkshift + level *
1313 * (indblkshift - SPA_BLKPTRSHIFT))
1316 const unsigned exp
= dn
->dn_datablkshift
+
1317 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1319 if (exp
>= 8 * sizeof (offset
)) {
1320 /* This only happens on the highest indirection level */
1321 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1325 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1327 return (offset
>> exp
);
1329 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1335 * This function is used to lock the parent of the provided dbuf. This should be
1336 * used when modifying or reading db_blkptr.
1339 dmu_buf_lock_parent(dmu_buf_impl_t
*db
, krw_t rw
, const void *tag
)
1341 enum db_lock_type ret
= DLT_NONE
;
1342 if (db
->db_parent
!= NULL
) {
1343 rw_enter(&db
->db_parent
->db_rwlock
, rw
);
1345 } else if (dmu_objset_ds(db
->db_objset
) != NULL
) {
1346 rrw_enter(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, rw
,
1351 * We only return a DLT_NONE lock when it's the top-most indirect block
1352 * of the meta-dnode of the MOS.
1358 * We need to pass the lock type in because it's possible that the block will
1359 * move from being the topmost indirect block in a dnode (and thus, have no
1360 * parent) to not the top-most via an indirection increase. This would cause a
1361 * panic if we didn't pass the lock type in.
1364 dmu_buf_unlock_parent(dmu_buf_impl_t
*db
, db_lock_type_t type
, const void *tag
)
1366 if (type
== DLT_PARENT
)
1367 rw_exit(&db
->db_parent
->db_rwlock
);
1368 else if (type
== DLT_OBJSET
)
1369 rrw_exit(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, tag
);
1373 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1374 arc_buf_t
*buf
, void *vdb
)
1376 (void) zb
, (void) bp
;
1377 dmu_buf_impl_t
*db
= vdb
;
1379 mutex_enter(&db
->db_mtx
);
1380 ASSERT3U(db
->db_state
, ==, DB_READ
);
1382 * All reads are synchronous, so we must have a hold on the dbuf
1384 ASSERT(zfs_refcount_count(&db
->db_holds
) > 0);
1385 ASSERT(db
->db_buf
== NULL
);
1386 ASSERT(db
->db
.db_data
== NULL
);
1389 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
1390 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1391 ASSERT3P(db
->db_buf
, ==, NULL
);
1392 db
->db_state
= DB_UNCACHED
;
1393 DTRACE_SET_STATE(db
, "i/o error");
1394 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1395 /* freed in flight */
1396 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1397 arc_release(buf
, db
);
1398 memset(buf
->b_data
, 0, db
->db
.db_size
);
1399 arc_buf_freeze(buf
);
1400 db
->db_freed_in_flight
= FALSE
;
1401 dbuf_set_data(db
, buf
);
1402 db
->db_state
= DB_CACHED
;
1403 DTRACE_SET_STATE(db
, "freed in flight");
1406 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1407 dbuf_set_data(db
, buf
);
1408 db
->db_state
= DB_CACHED
;
1409 DTRACE_SET_STATE(db
, "successful read");
1411 cv_broadcast(&db
->db_changed
);
1412 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
1416 * Shortcut for performing reads on bonus dbufs. Returns
1417 * an error if we fail to verify the dnode associated with
1418 * a decrypted block. Otherwise success.
1421 dbuf_read_bonus(dmu_buf_impl_t
*db
, dnode_t
*dn
, uint32_t flags
)
1423 int bonuslen
, max_bonuslen
, err
;
1425 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1429 bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1430 max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1431 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1432 ASSERT(DB_DNODE_HELD(db
));
1433 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1434 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1435 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1436 if (bonuslen
< max_bonuslen
)
1437 memset(db
->db
.db_data
, 0, max_bonuslen
);
1439 memcpy(db
->db
.db_data
, DN_BONUS(dn
->dn_phys
), bonuslen
);
1440 db
->db_state
= DB_CACHED
;
1441 DTRACE_SET_STATE(db
, "bonus buffer filled");
1446 dbuf_handle_indirect_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
, blkptr_t
*dbbp
)
1448 blkptr_t
*bps
= db
->db
.db_data
;
1449 uint32_t indbs
= 1ULL << dn
->dn_indblkshift
;
1450 int n_bps
= indbs
>> SPA_BLKPTRSHIFT
;
1452 for (int i
= 0; i
< n_bps
; i
++) {
1453 blkptr_t
*bp
= &bps
[i
];
1455 ASSERT3U(BP_GET_LSIZE(dbbp
), ==, indbs
);
1456 BP_SET_LSIZE(bp
, BP_GET_LEVEL(dbbp
) == 1 ?
1457 dn
->dn_datablksz
: BP_GET_LSIZE(dbbp
));
1458 BP_SET_TYPE(bp
, BP_GET_TYPE(dbbp
));
1459 BP_SET_LEVEL(bp
, BP_GET_LEVEL(dbbp
) - 1);
1460 BP_SET_BIRTH(bp
, dbbp
->blk_birth
, 0);
1465 * Handle reads on dbufs that are holes, if necessary. This function
1466 * requires that the dbuf's mutex is held. Returns success (0) if action
1467 * was taken, ENOENT if no action was taken.
1470 dbuf_read_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
, blkptr_t
*bp
)
1472 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1474 int is_hole
= bp
== NULL
|| BP_IS_HOLE(bp
);
1476 * For level 0 blocks only, if the above check fails:
1477 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1478 * processes the delete record and clears the bp while we are waiting
1479 * for the dn_mtx (resulting in a "no" from block_freed).
1481 if (!is_hole
&& db
->db_level
== 0)
1482 is_hole
= dnode_block_freed(dn
, db
->db_blkid
) || BP_IS_HOLE(bp
);
1485 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1486 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
1488 if (bp
!= NULL
&& db
->db_level
> 0 && BP_IS_HOLE(bp
) &&
1489 bp
->blk_birth
!= 0) {
1490 dbuf_handle_indirect_hole(db
, dn
, bp
);
1492 db
->db_state
= DB_CACHED
;
1493 DTRACE_SET_STATE(db
, "hole read satisfied");
1500 * This function ensures that, when doing a decrypting read of a block,
1501 * we make sure we have decrypted the dnode associated with it. We must do
1502 * this so that we ensure we are fully authenticating the checksum-of-MACs
1503 * tree from the root of the objset down to this block. Indirect blocks are
1504 * always verified against their secure checksum-of-MACs assuming that the
1505 * dnode containing them is correct. Now that we are doing a decrypting read,
1506 * we can be sure that the key is loaded and verify that assumption. This is
1507 * especially important considering that we always read encrypted dnode
1508 * blocks as raw data (without verifying their MACs) to start, and
1509 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1512 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
)
1515 objset_t
*os
= db
->db_objset
;
1516 arc_buf_t
*dnode_abuf
;
1518 zbookmark_phys_t zb
;
1520 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1522 if ((flags
& DB_RF_NO_DECRYPT
) != 0 ||
1523 !os
->os_encrypted
|| os
->os_raw_receive
)
1528 dnode_abuf
= (dn
->dn_dbuf
!= NULL
) ? dn
->dn_dbuf
->db_buf
: NULL
;
1530 if (dnode_abuf
== NULL
|| !arc_is_encrypted(dnode_abuf
)) {
1535 SET_BOOKMARK(&zb
, dmu_objset_id(os
),
1536 DMU_META_DNODE_OBJECT
, 0, dn
->dn_dbuf
->db_blkid
);
1537 err
= arc_untransform(dnode_abuf
, os
->os_spa
, &zb
, B_TRUE
);
1540 * An error code of EACCES tells us that the key is still not
1541 * available. This is ok if we are only reading authenticated
1542 * (and therefore non-encrypted) blocks.
1544 if (err
== EACCES
&& ((db
->db_blkid
!= DMU_BONUS_BLKID
&&
1545 !DMU_OT_IS_ENCRYPTED(dn
->dn_type
)) ||
1546 (db
->db_blkid
== DMU_BONUS_BLKID
&&
1547 !DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))))
1556 * Drops db_mtx and the parent lock specified by dblt and tag before
1560 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
,
1561 db_lock_type_t dblt
, const void *tag
)
1564 zbookmark_phys_t zb
;
1565 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1571 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1572 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1573 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1574 ASSERT(db
->db_buf
== NULL
);
1575 ASSERT(db
->db_parent
== NULL
||
1576 RW_LOCK_HELD(&db
->db_parent
->db_rwlock
));
1578 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1579 err
= dbuf_read_bonus(db
, dn
, flags
);
1583 if (db
->db_state
== DB_UNCACHED
) {
1584 if (db
->db_blkptr
== NULL
) {
1587 bp
= *db
->db_blkptr
;
1591 dbuf_dirty_record_t
*dr
;
1593 ASSERT3S(db
->db_state
, ==, DB_NOFILL
);
1596 * Block cloning: If we have a pending block clone,
1597 * we don't want to read the underlying block, but the content
1598 * of the block being cloned, so we have the most recent data.
1600 dr
= list_head(&db
->db_dirty_records
);
1601 if (dr
== NULL
|| !dr
->dt
.dl
.dr_brtwrite
) {
1605 bp
= dr
->dt
.dl
.dr_overridden_by
;
1609 err
= dbuf_read_hole(db
, dn
, bpp
);
1613 ASSERT(bpp
!= NULL
);
1616 * Any attempt to read a redacted block should result in an error. This
1617 * will never happen under normal conditions, but can be useful for
1618 * debugging purposes.
1620 if (BP_IS_REDACTED(bpp
)) {
1621 ASSERT(dsl_dataset_feature_is_active(
1622 db
->db_objset
->os_dsl_dataset
,
1623 SPA_FEATURE_REDACTED_DATASETS
));
1624 err
= SET_ERROR(EIO
);
1628 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1629 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1632 * All bps of an encrypted os should have the encryption bit set.
1633 * If this is not true it indicates tampering and we report an error.
1635 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(bpp
)) {
1636 spa_log_error(db
->db_objset
->os_spa
, &zb
, &bpp
->blk_birth
);
1637 err
= SET_ERROR(EIO
);
1641 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1647 db
->db_state
= DB_READ
;
1648 DTRACE_SET_STATE(db
, "read issued");
1649 mutex_exit(&db
->db_mtx
);
1651 if (!DBUF_IS_CACHEABLE(db
))
1652 aflags
|= ARC_FLAG_UNCACHED
;
1653 else if (dbuf_is_l2cacheable(db
))
1654 aflags
|= ARC_FLAG_L2CACHE
;
1656 dbuf_add_ref(db
, NULL
);
1658 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1659 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1661 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1662 zio_flags
|= ZIO_FLAG_RAW
;
1664 * The zio layer will copy the provided blkptr later, but we have our
1665 * own copy so that we can release the parent's rwlock. We have to
1666 * do that so that if dbuf_read_done is called synchronously (on
1667 * an l1 cache hit) we don't acquire the db_mtx while holding the
1668 * parent's rwlock, which would be a lock ordering violation.
1670 dmu_buf_unlock_parent(db
, dblt
, tag
);
1671 (void) arc_read(zio
, db
->db_objset
->os_spa
, bpp
,
1672 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1677 mutex_exit(&db
->db_mtx
);
1678 dmu_buf_unlock_parent(db
, dblt
, tag
);
1683 * This is our just-in-time copy function. It makes a copy of buffers that
1684 * have been modified in a previous transaction group before we access them in
1685 * the current active group.
1687 * This function is used in three places: when we are dirtying a buffer for the
1688 * first time in a txg, when we are freeing a range in a dnode that includes
1689 * this buffer, and when we are accessing a buffer which was received compressed
1690 * and later referenced in a WRITE_BYREF record.
1692 * Note that when we are called from dbuf_free_range() we do not put a hold on
1693 * the buffer, we just traverse the active dbuf list for the dnode.
1696 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1698 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
1700 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1701 ASSERT(db
->db
.db_data
!= NULL
);
1702 ASSERT(db
->db_level
== 0);
1703 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1706 (dr
->dt
.dl
.dr_data
!=
1707 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1711 * If the last dirty record for this dbuf has not yet synced
1712 * and its referencing the dbuf data, either:
1713 * reset the reference to point to a new copy,
1714 * or (if there a no active holders)
1715 * just null out the current db_data pointer.
1717 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1718 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1719 dnode_t
*dn
= DB_DNODE(db
);
1720 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1721 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1722 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1723 memcpy(dr
->dt
.dl
.dr_data
, db
->db
.db_data
, bonuslen
);
1724 } else if (zfs_refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1725 dnode_t
*dn
= DB_DNODE(db
);
1726 int size
= arc_buf_size(db
->db_buf
);
1727 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1728 spa_t
*spa
= db
->db_objset
->os_spa
;
1729 enum zio_compress compress_type
=
1730 arc_get_compression(db
->db_buf
);
1731 uint8_t complevel
= arc_get_complevel(db
->db_buf
);
1733 if (arc_is_encrypted(db
->db_buf
)) {
1734 boolean_t byteorder
;
1735 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1736 uint8_t iv
[ZIO_DATA_IV_LEN
];
1737 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1739 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1741 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1742 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1743 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1744 compress_type
, complevel
);
1745 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1746 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1747 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1748 size
, arc_buf_lsize(db
->db_buf
), compress_type
,
1751 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1753 memcpy(dr
->dt
.dl
.dr_data
->b_data
, db
->db
.db_data
, size
);
1756 dbuf_clear_data(db
);
1761 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1768 * We don't have to hold the mutex to check db_state because it
1769 * can't be freed while we have a hold on the buffer.
1771 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1776 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1777 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
;
1779 mutex_enter(&db
->db_mtx
);
1780 if (flags
& DB_RF_PARTIAL_FIRST
)
1781 db
->db_partial_read
= B_TRUE
;
1782 else if (!(flags
& DB_RF_PARTIAL_MORE
))
1783 db
->db_partial_read
= B_FALSE
;
1784 if (db
->db_state
== DB_CACHED
) {
1786 * Ensure that this block's dnode has been decrypted if
1787 * the caller has requested decrypted data.
1789 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1792 * If the arc buf is compressed or encrypted and the caller
1793 * requested uncompressed data, we need to untransform it
1794 * before returning. We also call arc_untransform() on any
1795 * unauthenticated blocks, which will verify their MAC if
1796 * the key is now available.
1798 if (err
== 0 && db
->db_buf
!= NULL
&&
1799 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1800 (arc_is_encrypted(db
->db_buf
) ||
1801 arc_is_unauthenticated(db
->db_buf
) ||
1802 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1803 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1804 zbookmark_phys_t zb
;
1806 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1807 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1808 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1809 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1810 dbuf_set_data(db
, db
->db_buf
);
1812 mutex_exit(&db
->db_mtx
);
1813 if (err
== 0 && prefetch
) {
1814 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1815 B_FALSE
, flags
& DB_RF_HAVESTRUCT
);
1818 DBUF_STAT_BUMP(hash_hits
);
1819 } else if (db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
) {
1820 boolean_t need_wait
= B_FALSE
;
1822 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
1824 if (zio
== NULL
&& (db
->db_state
== DB_NOFILL
||
1825 (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)))) {
1826 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1827 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1830 err
= dbuf_read_impl(db
, zio
, flags
, dblt
, FTAG
);
1832 * dbuf_read_impl has dropped db_mtx and our parent's rwlock
1835 if (!err
&& prefetch
) {
1836 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1837 db
->db_state
!= DB_CACHED
,
1838 flags
& DB_RF_HAVESTRUCT
);
1842 DBUF_STAT_BUMP(hash_misses
);
1845 * If we created a zio_root we must execute it to avoid
1846 * leaking it, even if it isn't attached to any work due
1847 * to an error in dbuf_read_impl().
1851 err
= zio_wait(zio
);
1853 VERIFY0(zio_wait(zio
));
1857 * Another reader came in while the dbuf was in flight
1858 * between UNCACHED and CACHED. Either a writer will finish
1859 * writing the buffer (sending the dbuf to CACHED) or the
1860 * first reader's request will reach the read_done callback
1861 * and send the dbuf to CACHED. Otherwise, a failure
1862 * occurred and the dbuf went to UNCACHED.
1864 mutex_exit(&db
->db_mtx
);
1866 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1867 B_TRUE
, flags
& DB_RF_HAVESTRUCT
);
1870 DBUF_STAT_BUMP(hash_misses
);
1872 /* Skip the wait per the caller's request. */
1873 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1874 mutex_enter(&db
->db_mtx
);
1875 while (db
->db_state
== DB_READ
||
1876 db
->db_state
== DB_FILL
) {
1877 ASSERT(db
->db_state
== DB_READ
||
1878 (flags
& DB_RF_HAVESTRUCT
) == 0);
1879 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1881 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1883 if (db
->db_state
== DB_UNCACHED
)
1884 err
= SET_ERROR(EIO
);
1885 mutex_exit(&db
->db_mtx
);
1893 dbuf_noread(dmu_buf_impl_t
*db
)
1895 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1896 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1897 mutex_enter(&db
->db_mtx
);
1898 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1899 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1900 if (db
->db_state
== DB_UNCACHED
) {
1901 ASSERT(db
->db_buf
== NULL
);
1902 ASSERT(db
->db
.db_data
== NULL
);
1903 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1904 db
->db_state
= DB_FILL
;
1905 DTRACE_SET_STATE(db
, "assigning filled buffer");
1906 } else if (db
->db_state
== DB_NOFILL
) {
1907 dbuf_clear_data(db
);
1909 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1911 mutex_exit(&db
->db_mtx
);
1915 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1917 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1918 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1919 uint64_t txg
= dr
->dr_txg
;
1921 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1923 * This assert is valid because dmu_sync() expects to be called by
1924 * a zilog's get_data while holding a range lock. This call only
1925 * comes from dbuf_dirty() callers who must also hold a range lock.
1927 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1928 ASSERT(db
->db_level
== 0);
1930 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1931 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1934 ASSERT(db
->db_data_pending
!= dr
);
1936 /* free this block */
1937 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1938 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1940 if (dr
->dt
.dl
.dr_brtwrite
) {
1941 ASSERT0P(dr
->dt
.dl
.dr_data
);
1942 dr
->dt
.dl
.dr_data
= db
->db_buf
;
1944 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1945 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1946 dr
->dt
.dl
.dr_brtwrite
= B_FALSE
;
1947 dr
->dt
.dl
.dr_has_raw_params
= B_FALSE
;
1950 * Release the already-written buffer, so we leave it in
1951 * a consistent dirty state. Note that all callers are
1952 * modifying the buffer, so they will immediately do
1953 * another (redundant) arc_release(). Therefore, leave
1954 * the buf thawed to save the effort of freezing &
1955 * immediately re-thawing it.
1957 if (dr
->dt
.dl
.dr_data
)
1958 arc_release(dr
->dt
.dl
.dr_data
, db
);
1962 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1963 * data blocks in the free range, so that any future readers will find
1967 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1970 dmu_buf_impl_t
*db_search
;
1971 dmu_buf_impl_t
*db
, *db_next
;
1972 uint64_t txg
= tx
->tx_txg
;
1974 dbuf_dirty_record_t
*dr
;
1976 if (end_blkid
> dn
->dn_maxblkid
&&
1977 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1978 end_blkid
= dn
->dn_maxblkid
;
1979 dprintf_dnode(dn
, "start=%llu end=%llu\n", (u_longlong_t
)start_blkid
,
1980 (u_longlong_t
)end_blkid
);
1982 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1983 db_search
->db_level
= 0;
1984 db_search
->db_blkid
= start_blkid
;
1985 db_search
->db_state
= DB_SEARCH
;
1987 mutex_enter(&dn
->dn_dbufs_mtx
);
1988 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1989 ASSERT3P(db
, ==, NULL
);
1991 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1993 for (; db
!= NULL
; db
= db_next
) {
1994 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1995 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1997 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
2000 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
2002 /* found a level 0 buffer in the range */
2003 mutex_enter(&db
->db_mtx
);
2004 if (dbuf_undirty(db
, tx
)) {
2005 /* mutex has been dropped and dbuf destroyed */
2009 if (db
->db_state
== DB_UNCACHED
||
2010 db
->db_state
== DB_NOFILL
||
2011 db
->db_state
== DB_EVICTING
) {
2012 ASSERT(db
->db
.db_data
== NULL
);
2013 mutex_exit(&db
->db_mtx
);
2016 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
2017 /* will be handled in dbuf_read_done or dbuf_rele */
2018 db
->db_freed_in_flight
= TRUE
;
2019 mutex_exit(&db
->db_mtx
);
2022 if (zfs_refcount_count(&db
->db_holds
) == 0) {
2027 /* The dbuf is referenced */
2029 dr
= list_head(&db
->db_dirty_records
);
2031 if (dr
->dr_txg
== txg
) {
2033 * This buffer is "in-use", re-adjust the file
2034 * size to reflect that this buffer may
2035 * contain new data when we sync.
2037 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
2038 db
->db_blkid
> dn
->dn_maxblkid
)
2039 dn
->dn_maxblkid
= db
->db_blkid
;
2040 dbuf_unoverride(dr
);
2043 * This dbuf is not dirty in the open context.
2044 * Either uncache it (if its not referenced in
2045 * the open context) or reset its contents to
2048 dbuf_fix_old_data(db
, txg
);
2051 /* clear the contents if its cached */
2052 if (db
->db_state
== DB_CACHED
) {
2053 ASSERT(db
->db
.db_data
!= NULL
);
2054 arc_release(db
->db_buf
, db
);
2055 rw_enter(&db
->db_rwlock
, RW_WRITER
);
2056 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2057 rw_exit(&db
->db_rwlock
);
2058 arc_buf_freeze(db
->db_buf
);
2061 mutex_exit(&db
->db_mtx
);
2064 mutex_exit(&dn
->dn_dbufs_mtx
);
2065 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
2069 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
2071 arc_buf_t
*buf
, *old_buf
;
2072 dbuf_dirty_record_t
*dr
;
2073 int osize
= db
->db
.db_size
;
2074 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2077 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2083 * XXX we should be doing a dbuf_read, checking the return
2084 * value and returning that up to our callers
2086 dmu_buf_will_dirty(&db
->db
, tx
);
2088 /* create the data buffer for the new block */
2089 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
2091 /* copy old block data to the new block */
2092 old_buf
= db
->db_buf
;
2093 memcpy(buf
->b_data
, old_buf
->b_data
, MIN(osize
, size
));
2094 /* zero the remainder */
2096 memset((uint8_t *)buf
->b_data
+ osize
, 0, size
- osize
);
2098 mutex_enter(&db
->db_mtx
);
2099 dbuf_set_data(db
, buf
);
2100 arc_buf_destroy(old_buf
, db
);
2101 db
->db
.db_size
= size
;
2103 dr
= list_head(&db
->db_dirty_records
);
2104 /* dirty record added by dmu_buf_will_dirty() */
2106 if (db
->db_level
== 0)
2107 dr
->dt
.dl
.dr_data
= buf
;
2108 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2109 ASSERT3U(dr
->dr_accounted
, ==, osize
);
2110 dr
->dr_accounted
= size
;
2111 mutex_exit(&db
->db_mtx
);
2113 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
2118 dbuf_release_bp(dmu_buf_impl_t
*db
)
2120 objset_t
*os __maybe_unused
= db
->db_objset
;
2122 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
2123 ASSERT(arc_released(os
->os_phys_buf
) ||
2124 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
2125 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
2127 (void) arc_release(db
->db_buf
, db
);
2131 * We already have a dirty record for this TXG, and we are being
2135 dbuf_redirty(dbuf_dirty_record_t
*dr
)
2137 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2139 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2141 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
2143 * If this buffer has already been written out,
2144 * we now need to reset its state.
2146 dbuf_unoverride(dr
);
2147 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
2148 db
->db_state
!= DB_NOFILL
) {
2149 /* Already released on initial dirty, so just thaw. */
2150 ASSERT(arc_released(db
->db_buf
));
2151 arc_buf_thaw(db
->db_buf
);
2156 dbuf_dirty_record_t
*
2157 dbuf_dirty_lightweight(dnode_t
*dn
, uint64_t blkid
, dmu_tx_t
*tx
)
2159 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2160 IMPLY(dn
->dn_objset
->os_raw_receive
, dn
->dn_maxblkid
>= blkid
);
2161 dnode_new_blkid(dn
, blkid
, tx
, B_TRUE
, B_FALSE
);
2162 ASSERT(dn
->dn_maxblkid
>= blkid
);
2164 dbuf_dirty_record_t
*dr
= kmem_zalloc(sizeof (*dr
), KM_SLEEP
);
2165 list_link_init(&dr
->dr_dirty_node
);
2166 list_link_init(&dr
->dr_dbuf_node
);
2168 dr
->dr_txg
= tx
->tx_txg
;
2169 dr
->dt
.dll
.dr_blkid
= blkid
;
2170 dr
->dr_accounted
= dn
->dn_datablksz
;
2173 * There should not be any dbuf for the block that we're dirtying.
2174 * Otherwise the buffer contents could be inconsistent between the
2175 * dbuf and the lightweight dirty record.
2177 ASSERT3P(NULL
, ==, dbuf_find(dn
->dn_objset
, dn
->dn_object
, 0, blkid
,
2180 mutex_enter(&dn
->dn_mtx
);
2181 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2182 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2183 range_tree_clear(dn
->dn_free_ranges
[txgoff
], blkid
, 1);
2186 if (dn
->dn_nlevels
== 1) {
2187 ASSERT3U(blkid
, <, dn
->dn_nblkptr
);
2188 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2189 mutex_exit(&dn
->dn_mtx
);
2190 rw_exit(&dn
->dn_struct_rwlock
);
2191 dnode_setdirty(dn
, tx
);
2193 mutex_exit(&dn
->dn_mtx
);
2195 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2196 dmu_buf_impl_t
*parent_db
= dbuf_hold_level(dn
,
2197 1, blkid
>> epbs
, FTAG
);
2198 rw_exit(&dn
->dn_struct_rwlock
);
2199 if (parent_db
== NULL
) {
2200 kmem_free(dr
, sizeof (*dr
));
2203 int err
= dbuf_read(parent_db
, NULL
,
2204 (DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2206 dbuf_rele(parent_db
, FTAG
);
2207 kmem_free(dr
, sizeof (*dr
));
2211 dbuf_dirty_record_t
*parent_dr
= dbuf_dirty(parent_db
, tx
);
2212 dbuf_rele(parent_db
, FTAG
);
2213 mutex_enter(&parent_dr
->dt
.di
.dr_mtx
);
2214 ASSERT3U(parent_dr
->dr_txg
, ==, tx
->tx_txg
);
2215 list_insert_tail(&parent_dr
->dt
.di
.dr_children
, dr
);
2216 mutex_exit(&parent_dr
->dt
.di
.dr_mtx
);
2217 dr
->dr_parent
= parent_dr
;
2220 dmu_objset_willuse_space(dn
->dn_objset
, dr
->dr_accounted
, tx
);
2225 dbuf_dirty_record_t
*
2226 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2230 dbuf_dirty_record_t
*dr
, *dr_next
, *dr_head
;
2231 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2232 boolean_t drop_struct_rwlock
= B_FALSE
;
2234 ASSERT(tx
->tx_txg
!= 0);
2235 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2236 DMU_TX_DIRTY_BUF(tx
, db
);
2241 * Shouldn't dirty a regular buffer in syncing context. Private
2242 * objects may be dirtied in syncing context, but only if they
2243 * were already pre-dirtied in open context.
2246 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
2247 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
2250 ASSERT(!dmu_tx_is_syncing(tx
) ||
2251 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
2252 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2253 dn
->dn_objset
->os_dsl_dataset
== NULL
);
2254 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2255 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2258 * We make this assert for private objects as well, but after we
2259 * check if we're already dirty. They are allowed to re-dirty
2260 * in syncing context.
2262 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2263 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2264 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2266 mutex_enter(&db
->db_mtx
);
2268 * XXX make this true for indirects too? The problem is that
2269 * transactions created with dmu_tx_create_assigned() from
2270 * syncing context don't bother holding ahead.
2272 ASSERT(db
->db_level
!= 0 ||
2273 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
2274 db
->db_state
== DB_NOFILL
);
2276 mutex_enter(&dn
->dn_mtx
);
2277 dnode_set_dirtyctx(dn
, tx
, db
);
2278 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
2279 dn
->dn_dirty_txg
= tx
->tx_txg
;
2280 mutex_exit(&dn
->dn_mtx
);
2282 if (db
->db_blkid
== DMU_SPILL_BLKID
)
2283 dn
->dn_have_spill
= B_TRUE
;
2286 * If this buffer is already dirty, we're done.
2288 dr_head
= list_head(&db
->db_dirty_records
);
2289 ASSERT(dr_head
== NULL
|| dr_head
->dr_txg
<= tx
->tx_txg
||
2290 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
2291 dr_next
= dbuf_find_dirty_lte(db
, tx
->tx_txg
);
2292 if (dr_next
&& dr_next
->dr_txg
== tx
->tx_txg
) {
2295 dbuf_redirty(dr_next
);
2296 mutex_exit(&db
->db_mtx
);
2301 * Only valid if not already dirty.
2303 ASSERT(dn
->dn_object
== 0 ||
2304 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2305 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2307 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
2310 * We should only be dirtying in syncing context if it's the
2311 * mos or we're initializing the os or it's a special object.
2312 * However, we are allowed to dirty in syncing context provided
2313 * we already dirtied it in open context. Hence we must make
2314 * this assertion only if we're not already dirty.
2317 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
2319 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2320 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
2321 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2322 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
2323 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2324 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2326 ASSERT(db
->db
.db_size
!= 0);
2328 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2330 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& db
->db_state
!= DB_NOFILL
) {
2331 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
2335 * If this buffer is dirty in an old transaction group we need
2336 * to make a copy of it so that the changes we make in this
2337 * transaction group won't leak out when we sync the older txg.
2339 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
2340 list_link_init(&dr
->dr_dirty_node
);
2341 list_link_init(&dr
->dr_dbuf_node
);
2343 if (db
->db_level
== 0) {
2344 void *data_old
= db
->db_buf
;
2346 if (db
->db_state
!= DB_NOFILL
) {
2347 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2348 dbuf_fix_old_data(db
, tx
->tx_txg
);
2349 data_old
= db
->db
.db_data
;
2350 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
2352 * Release the data buffer from the cache so
2353 * that we can modify it without impacting
2354 * possible other users of this cached data
2355 * block. Note that indirect blocks and
2356 * private objects are not released until the
2357 * syncing state (since they are only modified
2360 arc_release(db
->db_buf
, db
);
2361 dbuf_fix_old_data(db
, tx
->tx_txg
);
2362 data_old
= db
->db_buf
;
2364 ASSERT(data_old
!= NULL
);
2366 dr
->dt
.dl
.dr_data
= data_old
;
2368 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
2369 list_create(&dr
->dt
.di
.dr_children
,
2370 sizeof (dbuf_dirty_record_t
),
2371 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
2373 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& db
->db_state
!= DB_NOFILL
) {
2374 dr
->dr_accounted
= db
->db
.db_size
;
2377 dr
->dr_txg
= tx
->tx_txg
;
2378 list_insert_before(&db
->db_dirty_records
, dr_next
, dr
);
2381 * We could have been freed_in_flight between the dbuf_noread
2382 * and dbuf_dirty. We win, as though the dbuf_noread() had
2383 * happened after the free.
2385 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2386 db
->db_blkid
!= DMU_SPILL_BLKID
) {
2387 mutex_enter(&dn
->dn_mtx
);
2388 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2389 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
2392 mutex_exit(&dn
->dn_mtx
);
2393 db
->db_freed_in_flight
= FALSE
;
2397 * This buffer is now part of this txg
2399 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
2400 db
->db_dirtycnt
+= 1;
2401 ASSERT3U(db
->db_dirtycnt
, <=, 3);
2403 mutex_exit(&db
->db_mtx
);
2405 if (db
->db_blkid
== DMU_BONUS_BLKID
||
2406 db
->db_blkid
== DMU_SPILL_BLKID
) {
2407 mutex_enter(&dn
->dn_mtx
);
2408 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2409 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2410 mutex_exit(&dn
->dn_mtx
);
2411 dnode_setdirty(dn
, tx
);
2416 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
2417 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2418 drop_struct_rwlock
= B_TRUE
;
2422 * If we are overwriting a dedup BP, then unless it is snapshotted,
2423 * when we get to syncing context we will need to decrement its
2424 * refcount in the DDT. Prefetch the relevant DDT block so that
2425 * syncing context won't have to wait for the i/o.
2427 if (db
->db_blkptr
!= NULL
) {
2428 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
2429 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
2430 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
2434 * We need to hold the dn_struct_rwlock to make this assertion,
2435 * because it protects dn_phys / dn_next_nlevels from changing.
2437 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
2438 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
2439 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
2440 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
2441 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
2444 if (db
->db_level
== 0) {
2445 ASSERT(!db
->db_objset
->os_raw_receive
||
2446 dn
->dn_maxblkid
>= db
->db_blkid
);
2447 dnode_new_blkid(dn
, db
->db_blkid
, tx
,
2448 drop_struct_rwlock
, B_FALSE
);
2449 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
2452 if (db
->db_level
+1 < dn
->dn_nlevels
) {
2453 dmu_buf_impl_t
*parent
= db
->db_parent
;
2454 dbuf_dirty_record_t
*di
;
2455 int parent_held
= FALSE
;
2457 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
2458 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2459 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2460 db
->db_blkid
>> epbs
, FTAG
);
2461 ASSERT(parent
!= NULL
);
2464 if (drop_struct_rwlock
)
2465 rw_exit(&dn
->dn_struct_rwlock
);
2466 ASSERT3U(db
->db_level
+ 1, ==, parent
->db_level
);
2467 di
= dbuf_dirty(parent
, tx
);
2469 dbuf_rele(parent
, FTAG
);
2471 mutex_enter(&db
->db_mtx
);
2473 * Since we've dropped the mutex, it's possible that
2474 * dbuf_undirty() might have changed this out from under us.
2476 if (list_head(&db
->db_dirty_records
) == dr
||
2477 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2478 mutex_enter(&di
->dt
.di
.dr_mtx
);
2479 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2480 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2481 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2482 mutex_exit(&di
->dt
.di
.dr_mtx
);
2485 mutex_exit(&db
->db_mtx
);
2487 ASSERT(db
->db_level
+ 1 == dn
->dn_nlevels
);
2488 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2489 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2490 mutex_enter(&dn
->dn_mtx
);
2491 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2492 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2493 mutex_exit(&dn
->dn_mtx
);
2494 if (drop_struct_rwlock
)
2495 rw_exit(&dn
->dn_struct_rwlock
);
2498 dnode_setdirty(dn
, tx
);
2504 dbuf_undirty_bonus(dbuf_dirty_record_t
*dr
)
2506 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2508 if (dr
->dt
.dl
.dr_data
!= db
->db
.db_data
) {
2509 struct dnode
*dn
= dr
->dr_dnode
;
2510 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
2512 kmem_free(dr
->dt
.dl
.dr_data
, max_bonuslen
);
2513 arc_space_return(max_bonuslen
, ARC_SPACE_BONUS
);
2515 db
->db_data_pending
= NULL
;
2516 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
2517 list_remove(&db
->db_dirty_records
, dr
);
2518 if (dr
->dr_dbuf
->db_level
!= 0) {
2519 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2520 list_destroy(&dr
->dt
.di
.dr_children
);
2522 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2523 ASSERT3U(db
->db_dirtycnt
, >, 0);
2524 db
->db_dirtycnt
-= 1;
2528 * Undirty a buffer in the transaction group referenced by the given
2529 * transaction. Return whether this evicted the dbuf.
2532 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2534 uint64_t txg
= tx
->tx_txg
;
2540 * Due to our use of dn_nlevels below, this can only be called
2541 * in open context, unless we are operating on the MOS.
2542 * From syncing context, dn_nlevels may be different from the
2543 * dn_nlevels used when dbuf was dirtied.
2545 ASSERT(db
->db_objset
==
2546 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2547 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2548 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2549 ASSERT0(db
->db_level
);
2550 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2553 * If this buffer is not dirty, we're done.
2555 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, txg
);
2558 ASSERT(dr
->dr_dbuf
== db
);
2560 brtwrite
= dr
->dt
.dl
.dr_brtwrite
;
2563 * We are freeing a block that we cloned in the same
2564 * transaction group.
2566 brt_pending_remove(dmu_objset_spa(db
->db_objset
),
2567 &dr
->dt
.dl
.dr_overridden_by
, tx
);
2570 dnode_t
*dn
= dr
->dr_dnode
;
2572 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2574 ASSERT(db
->db
.db_size
!= 0);
2576 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2577 dr
->dr_accounted
, txg
);
2579 list_remove(&db
->db_dirty_records
, dr
);
2582 * Note that there are three places in dbuf_dirty()
2583 * where this dirty record may be put on a list.
2584 * Make sure to do a list_remove corresponding to
2585 * every one of those list_insert calls.
2587 if (dr
->dr_parent
) {
2588 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2589 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2590 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2591 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2592 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2593 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2594 mutex_enter(&dn
->dn_mtx
);
2595 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2596 mutex_exit(&dn
->dn_mtx
);
2599 if (db
->db_state
!= DB_NOFILL
&& !brtwrite
) {
2600 dbuf_unoverride(dr
);
2602 ASSERT(db
->db_buf
!= NULL
);
2603 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2604 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2605 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2608 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2610 ASSERT(db
->db_dirtycnt
> 0);
2611 db
->db_dirtycnt
-= 1;
2613 if (zfs_refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2614 ASSERT(db
->db_state
== DB_NOFILL
|| brtwrite
||
2615 arc_released(db
->db_buf
));
2624 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2626 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2627 boolean_t undirty
= B_FALSE
;
2629 ASSERT(tx
->tx_txg
!= 0);
2630 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2633 * Quick check for dirtiness. For already dirty blocks, this
2634 * reduces runtime of this function by >90%, and overall performance
2635 * by 50% for some workloads (e.g. file deletion with indirect blocks
2638 mutex_enter(&db
->db_mtx
);
2640 if (db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
) {
2641 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2643 * It's possible that it is already dirty but not cached,
2644 * because there are some calls to dbuf_dirty() that don't
2645 * go through dmu_buf_will_dirty().
2648 if (dr
->dt
.dl
.dr_brtwrite
) {
2650 * Block cloning: If we are dirtying a cloned
2651 * block, we cannot simply redirty it, because
2652 * this dr has no data associated with it.
2653 * We will go through a full undirtying below,
2654 * before dirtying it again.
2658 /* This dbuf is already dirty and cached. */
2660 mutex_exit(&db
->db_mtx
);
2665 mutex_exit(&db
->db_mtx
);
2668 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2669 flags
|= DB_RF_HAVESTRUCT
;
2673 * Block cloning: Do the dbuf_read() before undirtying the dbuf, as we
2674 * want to make sure dbuf_read() will read the pending cloned block and
2675 * not the uderlying block that is being replaced. dbuf_undirty() will
2676 * do dbuf_unoverride(), so we will end up with cloned block content,
2677 * without overridden BP.
2679 (void) dbuf_read(db
, NULL
, flags
);
2681 mutex_enter(&db
->db_mtx
);
2682 VERIFY(!dbuf_undirty(db
, tx
));
2683 mutex_exit(&db
->db_mtx
);
2685 (void) dbuf_dirty(db
, tx
);
2689 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2691 dmu_buf_will_dirty_impl(db_fake
,
2692 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2696 dmu_buf_is_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2698 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2699 dbuf_dirty_record_t
*dr
;
2701 mutex_enter(&db
->db_mtx
);
2702 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2703 mutex_exit(&db
->db_mtx
);
2704 return (dr
!= NULL
);
2708 dmu_buf_will_clone(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2710 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2713 * Block cloning: We are going to clone into this block, so undirty
2714 * modifications done to this block so far in this txg. This includes
2715 * writes and clones into this block.
2717 mutex_enter(&db
->db_mtx
);
2719 VERIFY(!dbuf_undirty(db
, tx
));
2720 ASSERT0P(dbuf_find_dirty_eq(db
, tx
->tx_txg
));
2721 if (db
->db_buf
!= NULL
) {
2722 arc_buf_destroy(db
->db_buf
, db
);
2724 dbuf_clear_data(db
);
2727 db
->db_state
= DB_NOFILL
;
2728 DTRACE_SET_STATE(db
, "allocating NOFILL buffer for clone");
2731 mutex_exit(&db
->db_mtx
);
2734 (void) dbuf_dirty(db
, tx
);
2738 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2740 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2742 mutex_enter(&db
->db_mtx
);
2743 db
->db_state
= DB_NOFILL
;
2744 DTRACE_SET_STATE(db
, "allocating NOFILL buffer");
2745 mutex_exit(&db
->db_mtx
);
2748 (void) dbuf_dirty(db
, tx
);
2752 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
, boolean_t canfail
)
2754 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2756 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2757 ASSERT(tx
->tx_txg
!= 0);
2758 ASSERT(db
->db_level
== 0);
2759 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2761 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2762 dmu_tx_private_ok(tx
));
2764 mutex_enter(&db
->db_mtx
);
2765 if (db
->db_state
== DB_NOFILL
) {
2767 * Block cloning: We will be completely overwriting a block
2768 * cloned in this transaction group, so let's undirty the
2769 * pending clone and mark the block as uncached. This will be
2770 * as if the clone was never done. But if the fill can fail
2771 * we should have a way to return back to the cloned data.
2773 if (canfail
&& dbuf_find_dirty_eq(db
, tx
->tx_txg
) != NULL
) {
2774 mutex_exit(&db
->db_mtx
);
2775 dmu_buf_will_dirty(db_fake
, tx
);
2778 VERIFY(!dbuf_undirty(db
, tx
));
2779 db
->db_state
= DB_UNCACHED
;
2781 mutex_exit(&db
->db_mtx
);
2784 (void) dbuf_dirty(db
, tx
);
2788 * This function is effectively the same as dmu_buf_will_dirty(), but
2789 * indicates the caller expects raw encrypted data in the db, and provides
2790 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2791 * blkptr_t when this dbuf is written. This is only used for blocks of
2792 * dnodes, during raw receive.
2795 dmu_buf_set_crypt_params(dmu_buf_t
*db_fake
, boolean_t byteorder
,
2796 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
, dmu_tx_t
*tx
)
2798 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2799 dbuf_dirty_record_t
*dr
;
2802 * dr_has_raw_params is only processed for blocks of dnodes
2803 * (see dbuf_sync_dnode_leaf_crypt()).
2805 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
2806 ASSERT3U(db
->db_level
, ==, 0);
2807 ASSERT(db
->db_objset
->os_raw_receive
);
2809 dmu_buf_will_dirty_impl(db_fake
,
2810 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2812 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2814 ASSERT3P(dr
, !=, NULL
);
2816 dr
->dt
.dl
.dr_has_raw_params
= B_TRUE
;
2817 dr
->dt
.dl
.dr_byteorder
= byteorder
;
2818 memcpy(dr
->dt
.dl
.dr_salt
, salt
, ZIO_DATA_SALT_LEN
);
2819 memcpy(dr
->dt
.dl
.dr_iv
, iv
, ZIO_DATA_IV_LEN
);
2820 memcpy(dr
->dt
.dl
.dr_mac
, mac
, ZIO_DATA_MAC_LEN
);
2824 dbuf_override_impl(dmu_buf_impl_t
*db
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
2826 struct dirty_leaf
*dl
;
2827 dbuf_dirty_record_t
*dr
;
2829 dr
= list_head(&db
->db_dirty_records
);
2830 ASSERT3P(dr
, !=, NULL
);
2831 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2833 dl
->dr_overridden_by
= *bp
;
2834 dl
->dr_override_state
= DR_OVERRIDDEN
;
2835 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2839 dmu_buf_fill_done(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
, boolean_t failed
)
2842 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2843 mutex_enter(&db
->db_mtx
);
2846 if (db
->db_state
== DB_FILL
) {
2847 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2848 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2849 /* we were freed while filling */
2850 /* XXX dbuf_undirty? */
2851 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2852 db
->db_freed_in_flight
= FALSE
;
2853 db
->db_state
= DB_CACHED
;
2854 DTRACE_SET_STATE(db
,
2855 "fill done handling freed in flight");
2857 } else if (failed
) {
2858 VERIFY(!dbuf_undirty(db
, tx
));
2860 dbuf_clear_data(db
);
2861 DTRACE_SET_STATE(db
, "fill failed");
2863 db
->db_state
= DB_CACHED
;
2864 DTRACE_SET_STATE(db
, "fill done");
2866 cv_broadcast(&db
->db_changed
);
2868 db
->db_state
= DB_CACHED
;
2871 mutex_exit(&db
->db_mtx
);
2876 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2877 bp_embedded_type_t etype
, enum zio_compress comp
,
2878 int uncompressed_size
, int compressed_size
, int byteorder
,
2881 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2882 struct dirty_leaf
*dl
;
2883 dmu_object_type_t type
;
2884 dbuf_dirty_record_t
*dr
;
2886 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2887 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2888 SPA_FEATURE_EMBEDDED_DATA
));
2892 type
= DB_DNODE(db
)->dn_type
;
2895 ASSERT0(db
->db_level
);
2896 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2898 dmu_buf_will_not_fill(dbuf
, tx
);
2900 dr
= list_head(&db
->db_dirty_records
);
2901 ASSERT3P(dr
, !=, NULL
);
2902 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2904 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2905 data
, comp
, uncompressed_size
, compressed_size
);
2906 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2907 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2908 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2909 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2911 dl
->dr_override_state
= DR_OVERRIDDEN
;
2912 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2916 dmu_buf_redact(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2918 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2919 dmu_object_type_t type
;
2920 ASSERT(dsl_dataset_feature_is_active(db
->db_objset
->os_dsl_dataset
,
2921 SPA_FEATURE_REDACTED_DATASETS
));
2924 type
= DB_DNODE(db
)->dn_type
;
2927 ASSERT0(db
->db_level
);
2928 dmu_buf_will_not_fill(dbuf
, tx
);
2930 blkptr_t bp
= { { { {0} } } };
2931 BP_SET_TYPE(&bp
, type
);
2932 BP_SET_LEVEL(&bp
, 0);
2933 BP_SET_BIRTH(&bp
, tx
->tx_txg
, 0);
2934 BP_SET_REDACTED(&bp
);
2935 BPE_SET_LSIZE(&bp
, dbuf
->db_size
);
2937 dbuf_override_impl(db
, &bp
, tx
);
2941 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2942 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2945 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2947 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2948 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2949 ASSERT(db
->db_level
== 0);
2950 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2951 ASSERT(buf
!= NULL
);
2952 ASSERT3U(arc_buf_lsize(buf
), ==, db
->db
.db_size
);
2953 ASSERT(tx
->tx_txg
!= 0);
2955 arc_return_buf(buf
, db
);
2956 ASSERT(arc_released(buf
));
2958 mutex_enter(&db
->db_mtx
);
2960 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2961 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2963 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
||
2964 db
->db_state
== DB_NOFILL
);
2966 if (db
->db_state
== DB_CACHED
&&
2967 zfs_refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2969 * In practice, we will never have a case where we have an
2970 * encrypted arc buffer while additional holds exist on the
2971 * dbuf. We don't handle this here so we simply assert that
2974 ASSERT(!arc_is_encrypted(buf
));
2975 mutex_exit(&db
->db_mtx
);
2976 (void) dbuf_dirty(db
, tx
);
2977 memcpy(db
->db
.db_data
, buf
->b_data
, db
->db
.db_size
);
2978 arc_buf_destroy(buf
, db
);
2982 if (db
->db_state
== DB_CACHED
) {
2983 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
2985 ASSERT(db
->db_buf
!= NULL
);
2986 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2987 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2989 if (!arc_released(db
->db_buf
)) {
2990 ASSERT(dr
->dt
.dl
.dr_override_state
==
2992 arc_release(db
->db_buf
, db
);
2994 dr
->dt
.dl
.dr_data
= buf
;
2995 arc_buf_destroy(db
->db_buf
, db
);
2996 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2997 arc_release(db
->db_buf
, db
);
2998 arc_buf_destroy(db
->db_buf
, db
);
3001 } else if (db
->db_state
== DB_NOFILL
) {
3003 * We will be completely replacing the cloned block. In case
3004 * it was cloned in this transaction group, let's undirty the
3005 * pending clone and mark the block as uncached. This will be
3006 * as if the clone was never done.
3008 VERIFY(!dbuf_undirty(db
, tx
));
3009 db
->db_state
= DB_UNCACHED
;
3011 ASSERT(db
->db_buf
== NULL
);
3012 dbuf_set_data(db
, buf
);
3013 db
->db_state
= DB_FILL
;
3014 DTRACE_SET_STATE(db
, "filling assigned arcbuf");
3015 mutex_exit(&db
->db_mtx
);
3016 (void) dbuf_dirty(db
, tx
);
3017 dmu_buf_fill_done(&db
->db
, tx
, B_FALSE
);
3021 dbuf_destroy(dmu_buf_impl_t
*db
)
3024 dmu_buf_impl_t
*parent
= db
->db_parent
;
3025 dmu_buf_impl_t
*dndb
;
3027 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3028 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3030 if (db
->db_buf
!= NULL
) {
3031 arc_buf_destroy(db
->db_buf
, db
);
3035 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3036 int slots
= DB_DNODE(db
)->dn_num_slots
;
3037 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3038 if (db
->db
.db_data
!= NULL
) {
3039 kmem_free(db
->db
.db_data
, bonuslen
);
3040 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3041 db
->db_state
= DB_UNCACHED
;
3042 DTRACE_SET_STATE(db
, "buffer cleared");
3046 dbuf_clear_data(db
);
3048 if (multilist_link_active(&db
->db_cache_link
)) {
3049 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3050 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3052 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
3054 ASSERT0(dmu_buf_user_size(&db
->db
));
3055 (void) zfs_refcount_remove_many(
3056 &dbuf_caches
[db
->db_caching_status
].size
,
3057 db
->db
.db_size
, db
);
3059 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3060 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3062 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3063 DBUF_STAT_BUMPDOWN(cache_count
);
3064 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3067 db
->db_caching_status
= DB_NO_CACHE
;
3070 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
3071 ASSERT(db
->db_data_pending
== NULL
);
3072 ASSERT(list_is_empty(&db
->db_dirty_records
));
3074 db
->db_state
= DB_EVICTING
;
3075 DTRACE_SET_STATE(db
, "buffer eviction started");
3076 db
->db_blkptr
= NULL
;
3079 * Now that db_state is DB_EVICTING, nobody else can find this via
3080 * the hash table. We can now drop db_mtx, which allows us to
3081 * acquire the dn_dbufs_mtx.
3083 mutex_exit(&db
->db_mtx
);
3088 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
3089 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
3091 mutex_enter_nested(&dn
->dn_dbufs_mtx
,
3093 avl_remove(&dn
->dn_dbufs
, db
);
3097 mutex_exit(&dn
->dn_dbufs_mtx
);
3099 * Decrementing the dbuf count means that the hold corresponding
3100 * to the removed dbuf is no longer discounted in dnode_move(),
3101 * so the dnode cannot be moved until after we release the hold.
3102 * The membar_producer() ensures visibility of the decremented
3103 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
3106 mutex_enter(&dn
->dn_mtx
);
3107 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
3108 db
->db_dnode_handle
= NULL
;
3110 dbuf_hash_remove(db
);
3115 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3117 db
->db_parent
= NULL
;
3119 ASSERT(db
->db_buf
== NULL
);
3120 ASSERT(db
->db
.db_data
== NULL
);
3121 ASSERT(db
->db_hash_next
== NULL
);
3122 ASSERT(db
->db_blkptr
== NULL
);
3123 ASSERT(db
->db_data_pending
== NULL
);
3124 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
3125 ASSERT(!multilist_link_active(&db
->db_cache_link
));
3128 * If this dbuf is referenced from an indirect dbuf,
3129 * decrement the ref count on the indirect dbuf.
3131 if (parent
&& parent
!= dndb
) {
3132 mutex_enter(&parent
->db_mtx
);
3133 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
3136 kmem_cache_free(dbuf_kmem_cache
, db
);
3137 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3141 * Note: While bpp will always be updated if the function returns success,
3142 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
3143 * this happens when the dnode is the meta-dnode, or {user|group|project}used
3146 __attribute__((always_inline
))
3148 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
3149 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
3154 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3156 if (blkid
== DMU_SPILL_BLKID
) {
3157 mutex_enter(&dn
->dn_mtx
);
3158 if (dn
->dn_have_spill
&&
3159 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
3160 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3163 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3164 *parentp
= dn
->dn_dbuf
;
3165 mutex_exit(&dn
->dn_mtx
);
3170 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
3171 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3173 ASSERT3U(level
* epbs
, <, 64);
3174 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3176 * This assertion shouldn't trip as long as the max indirect block size
3177 * is less than 1M. The reason for this is that up to that point,
3178 * the number of levels required to address an entire object with blocks
3179 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
3180 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
3181 * (i.e. we can address the entire object), objects will all use at most
3182 * N-1 levels and the assertion won't overflow. However, once epbs is
3183 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
3184 * enough to address an entire object, so objects will have 5 levels,
3185 * but then this assertion will overflow.
3187 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
3188 * need to redo this logic to handle overflows.
3190 ASSERT(level
>= nlevels
||
3191 ((nlevels
- level
- 1) * epbs
) +
3192 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
3193 if (level
>= nlevels
||
3194 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
3195 ((nlevels
- level
- 1) * epbs
)) ||
3197 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
3198 /* the buffer has no parent yet */
3199 return (SET_ERROR(ENOENT
));
3200 } else if (level
< nlevels
-1) {
3201 /* this block is referenced from an indirect block */
3204 err
= dbuf_hold_impl(dn
, level
+ 1,
3205 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
3209 err
= dbuf_read(*parentp
, NULL
,
3210 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
3212 dbuf_rele(*parentp
, NULL
);
3216 rw_enter(&(*parentp
)->db_rwlock
, RW_READER
);
3217 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
3218 (blkid
& ((1ULL << epbs
) - 1));
3219 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
3220 ASSERT(BP_IS_HOLE(*bpp
));
3221 rw_exit(&(*parentp
)->db_rwlock
);
3224 /* the block is referenced from the dnode */
3225 ASSERT3U(level
, ==, nlevels
-1);
3226 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
3227 blkid
< dn
->dn_phys
->dn_nblkptr
);
3229 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3230 *parentp
= dn
->dn_dbuf
;
3232 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
3237 static dmu_buf_impl_t
*
3238 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3239 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
, uint64_t hash
)
3241 objset_t
*os
= dn
->dn_objset
;
3242 dmu_buf_impl_t
*db
, *odb
;
3244 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3245 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
3247 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
3249 list_create(&db
->db_dirty_records
, sizeof (dbuf_dirty_record_t
),
3250 offsetof(dbuf_dirty_record_t
, dr_dbuf_node
));
3253 db
->db
.db_object
= dn
->dn_object
;
3254 db
->db_level
= level
;
3255 db
->db_blkid
= blkid
;
3256 db
->db_dirtycnt
= 0;
3257 db
->db_dnode_handle
= dn
->dn_handle
;
3258 db
->db_parent
= parent
;
3259 db
->db_blkptr
= blkptr
;
3263 db
->db_user_immediate_evict
= FALSE
;
3264 db
->db_freed_in_flight
= FALSE
;
3265 db
->db_pending_evict
= FALSE
;
3267 if (blkid
== DMU_BONUS_BLKID
) {
3268 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
3269 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
3270 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
3271 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
3272 db
->db
.db_offset
= DMU_BONUS_BLKID
;
3273 db
->db_state
= DB_UNCACHED
;
3274 DTRACE_SET_STATE(db
, "bonus buffer created");
3275 db
->db_caching_status
= DB_NO_CACHE
;
3276 /* the bonus dbuf is not placed in the hash table */
3277 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3279 } else if (blkid
== DMU_SPILL_BLKID
) {
3280 db
->db
.db_size
= (blkptr
!= NULL
) ?
3281 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
3282 db
->db
.db_offset
= 0;
3285 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
3286 db
->db
.db_size
= blocksize
;
3287 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
3291 * Hold the dn_dbufs_mtx while we get the new dbuf
3292 * in the hash table *and* added to the dbufs list.
3293 * This prevents a possible deadlock with someone
3294 * trying to look up this dbuf before it's added to the
3297 mutex_enter(&dn
->dn_dbufs_mtx
);
3298 db
->db_state
= DB_EVICTING
; /* not worth logging this state change */
3299 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
3300 /* someone else inserted it first */
3301 mutex_exit(&dn
->dn_dbufs_mtx
);
3302 kmem_cache_free(dbuf_kmem_cache
, db
);
3303 DBUF_STAT_BUMP(hash_insert_race
);
3306 avl_add(&dn
->dn_dbufs
, db
);
3308 db
->db_state
= DB_UNCACHED
;
3309 DTRACE_SET_STATE(db
, "regular buffer created");
3310 db
->db_caching_status
= DB_NO_CACHE
;
3311 mutex_exit(&dn
->dn_dbufs_mtx
);
3312 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3314 if (parent
&& parent
!= dn
->dn_dbuf
)
3315 dbuf_add_ref(parent
, db
);
3317 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
3318 zfs_refcount_count(&dn
->dn_holds
) > 0);
3319 (void) zfs_refcount_add(&dn
->dn_holds
, db
);
3321 dprintf_dbuf(db
, "db=%p\n", db
);
3327 * This function returns a block pointer and information about the object,
3328 * given a dnode and a block. This is a publicly accessible version of
3329 * dbuf_findbp that only returns some information, rather than the
3330 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
3331 * should be locked as (at least) a reader.
3334 dbuf_dnode_findbp(dnode_t
*dn
, uint64_t level
, uint64_t blkid
,
3335 blkptr_t
*bp
, uint16_t *datablkszsec
, uint8_t *indblkshift
)
3337 dmu_buf_impl_t
*dbp
= NULL
;
3340 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3342 err
= dbuf_findbp(dn
, level
, blkid
, B_FALSE
, &dbp
, &bp2
);
3344 ASSERT3P(bp2
, !=, NULL
);
3347 dbuf_rele(dbp
, NULL
);
3348 if (datablkszsec
!= NULL
)
3349 *datablkszsec
= dn
->dn_phys
->dn_datablkszsec
;
3350 if (indblkshift
!= NULL
)
3351 *indblkshift
= dn
->dn_phys
->dn_indblkshift
;
3357 typedef struct dbuf_prefetch_arg
{
3358 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
3359 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
3360 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
3361 int dpa_curlevel
; /* The current level that we're reading */
3362 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
3363 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
3364 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
3365 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
3366 dbuf_prefetch_fn dpa_cb
; /* prefetch completion callback */
3367 void *dpa_arg
; /* prefetch completion arg */
3368 } dbuf_prefetch_arg_t
;
3371 dbuf_prefetch_fini(dbuf_prefetch_arg_t
*dpa
, boolean_t io_done
)
3373 if (dpa
->dpa_cb
!= NULL
) {
3374 dpa
->dpa_cb(dpa
->dpa_arg
, dpa
->dpa_zb
.zb_level
,
3375 dpa
->dpa_zb
.zb_blkid
, io_done
);
3377 kmem_free(dpa
, sizeof (*dpa
));
3381 dbuf_issue_final_prefetch_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3382 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3384 (void) zio
, (void) zb
, (void) iobp
;
3385 dbuf_prefetch_arg_t
*dpa
= private;
3388 arc_buf_destroy(abuf
, private);
3390 dbuf_prefetch_fini(dpa
, B_TRUE
);
3394 * Actually issue the prefetch read for the block given.
3397 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
3399 ASSERT(!BP_IS_REDACTED(bp
) ||
3400 dsl_dataset_feature_is_active(
3401 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3402 SPA_FEATURE_REDACTED_DATASETS
));
3404 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
) || BP_IS_REDACTED(bp
))
3405 return (dbuf_prefetch_fini(dpa
, B_FALSE
));
3407 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
3408 arc_flags_t aflags
=
3409 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
|
3412 /* dnodes are always read as raw and then converted later */
3413 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
3414 dpa
->dpa_curlevel
== 0)
3415 zio_flags
|= ZIO_FLAG_RAW
;
3417 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3418 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
3419 ASSERT(dpa
->dpa_zio
!= NULL
);
3420 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
,
3421 dbuf_issue_final_prefetch_done
, dpa
,
3422 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
3426 * Called when an indirect block above our prefetch target is read in. This
3427 * will either read in the next indirect block down the tree or issue the actual
3428 * prefetch if the next block down is our target.
3431 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3432 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3434 (void) zb
, (void) iobp
;
3435 dbuf_prefetch_arg_t
*dpa
= private;
3437 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
3438 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
3441 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
3442 dbuf_prefetch_fini(dpa
, B_TRUE
);
3445 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
3448 * The dpa_dnode is only valid if we are called with a NULL
3449 * zio. This indicates that the arc_read() returned without
3450 * first calling zio_read() to issue a physical read. Once
3451 * a physical read is made the dpa_dnode must be invalidated
3452 * as the locks guarding it may have been dropped. If the
3453 * dpa_dnode is still valid, then we want to add it to the dbuf
3454 * cache. To do so, we must hold the dbuf associated with the block
3455 * we just prefetched, read its contents so that we associate it
3456 * with an arc_buf_t, and then release it.
3459 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
3460 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
3461 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
3463 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
3465 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
3467 dpa
->dpa_dnode
= NULL
;
3468 } else if (dpa
->dpa_dnode
!= NULL
) {
3469 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
3470 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
3471 dpa
->dpa_zb
.zb_level
));
3472 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
3473 dpa
->dpa_curlevel
, curblkid
, FTAG
);
3475 arc_buf_destroy(abuf
, private);
3476 dbuf_prefetch_fini(dpa
, B_TRUE
);
3479 (void) dbuf_read(db
, NULL
,
3480 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
3481 dbuf_rele(db
, FTAG
);
3484 dpa
->dpa_curlevel
--;
3485 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
3486 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
3487 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
3488 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
3490 ASSERT(!BP_IS_REDACTED(bp
) || (dpa
->dpa_dnode
&&
3491 dsl_dataset_feature_is_active(
3492 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3493 SPA_FEATURE_REDACTED_DATASETS
)));
3494 if (BP_IS_HOLE(bp
) || BP_IS_REDACTED(bp
)) {
3495 arc_buf_destroy(abuf
, private);
3496 dbuf_prefetch_fini(dpa
, B_TRUE
);
3498 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
3499 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
3500 dbuf_issue_final_prefetch(dpa
, bp
);
3502 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3503 zbookmark_phys_t zb
;
3505 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3506 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
3507 iter_aflags
|= ARC_FLAG_L2CACHE
;
3509 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3511 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
3512 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
3514 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3515 bp
, dbuf_prefetch_indirect_done
, dpa
,
3516 ZIO_PRIORITY_SYNC_READ
,
3517 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3521 arc_buf_destroy(abuf
, private);
3525 * Issue prefetch reads for the given block on the given level. If the indirect
3526 * blocks above that block are not in memory, we will read them in
3527 * asynchronously. As a result, this call never blocks waiting for a read to
3528 * complete. Note that the prefetch might fail if the dataset is encrypted and
3529 * the encryption key is unmapped before the IO completes.
3532 dbuf_prefetch_impl(dnode_t
*dn
, int64_t level
, uint64_t blkid
,
3533 zio_priority_t prio
, arc_flags_t aflags
, dbuf_prefetch_fn cb
,
3537 int epbs
, nlevels
, curlevel
;
3540 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3541 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3543 if (blkid
> dn
->dn_maxblkid
)
3546 if (level
== 0 && dnode_block_freed(dn
, blkid
))
3550 * This dnode hasn't been written to disk yet, so there's nothing to
3553 nlevels
= dn
->dn_phys
->dn_nlevels
;
3554 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
3557 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3558 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
3561 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
3562 level
, blkid
, NULL
);
3564 mutex_exit(&db
->db_mtx
);
3566 * This dbuf already exists. It is either CACHED, or
3567 * (we assume) about to be read or filled.
3573 * Find the closest ancestor (indirect block) of the target block
3574 * that is present in the cache. In this indirect block, we will
3575 * find the bp that is at curlevel, curblkid.
3579 while (curlevel
< nlevels
- 1) {
3580 int parent_level
= curlevel
+ 1;
3581 uint64_t parent_blkid
= curblkid
>> epbs
;
3584 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
3585 FALSE
, TRUE
, FTAG
, &db
) == 0) {
3586 blkptr_t
*bpp
= db
->db_buf
->b_data
;
3587 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
3588 dbuf_rele(db
, FTAG
);
3592 curlevel
= parent_level
;
3593 curblkid
= parent_blkid
;
3596 if (curlevel
== nlevels
- 1) {
3597 /* No cached indirect blocks found. */
3598 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
3599 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
3601 ASSERT(!BP_IS_REDACTED(&bp
) ||
3602 dsl_dataset_feature_is_active(dn
->dn_objset
->os_dsl_dataset
,
3603 SPA_FEATURE_REDACTED_DATASETS
));
3604 if (BP_IS_HOLE(&bp
) || BP_IS_REDACTED(&bp
))
3607 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
3609 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
3612 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
3613 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
3614 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3615 dn
->dn_object
, level
, blkid
);
3616 dpa
->dpa_curlevel
= curlevel
;
3617 dpa
->dpa_prio
= prio
;
3618 dpa
->dpa_aflags
= aflags
;
3619 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
3620 dpa
->dpa_dnode
= dn
;
3621 dpa
->dpa_epbs
= epbs
;
3626 if (!DNODE_LEVEL_IS_CACHEABLE(dn
, level
))
3627 dpa
->dpa_aflags
|= ARC_FLAG_UNCACHED
;
3628 else if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3629 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
3632 * If we have the indirect just above us, no need to do the asynchronous
3633 * prefetch chain; we'll just run the last step ourselves. If we're at
3634 * a higher level, though, we want to issue the prefetches for all the
3635 * indirect blocks asynchronously, so we can go on with whatever we were
3638 if (curlevel
== level
) {
3639 ASSERT3U(curblkid
, ==, blkid
);
3640 dbuf_issue_final_prefetch(dpa
, &bp
);
3642 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3643 zbookmark_phys_t zb
;
3645 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3646 if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3647 iter_aflags
|= ARC_FLAG_L2CACHE
;
3649 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3650 dn
->dn_object
, curlevel
, curblkid
);
3651 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3652 &bp
, dbuf_prefetch_indirect_done
, dpa
,
3653 ZIO_PRIORITY_SYNC_READ
,
3654 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3658 * We use pio here instead of dpa_zio since it's possible that
3659 * dpa may have already been freed.
3665 cb(arg
, level
, blkid
, B_FALSE
);
3670 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
3674 return (dbuf_prefetch_impl(dn
, level
, blkid
, prio
, aflags
, NULL
, NULL
));
3678 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3679 * the case of encrypted, compressed and uncompressed buffers by
3680 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3681 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3683 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3685 noinline
static void
3686 dbuf_hold_copy(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3688 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3689 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
3690 enum zio_compress compress_type
= arc_get_compression(data
);
3691 uint8_t complevel
= arc_get_complevel(data
);
3693 if (arc_is_encrypted(data
)) {
3694 boolean_t byteorder
;
3695 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3696 uint8_t iv
[ZIO_DATA_IV_LEN
];
3697 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3699 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
3700 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
3701 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
3702 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
3703 compress_type
, complevel
));
3704 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3705 dbuf_set_data(db
, arc_alloc_compressed_buf(
3706 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
3707 arc_buf_lsize(data
), compress_type
, complevel
));
3709 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
3710 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
3713 rw_enter(&db
->db_rwlock
, RW_WRITER
);
3714 memcpy(db
->db
.db_data
, data
->b_data
, arc_buf_size(data
));
3715 rw_exit(&db
->db_rwlock
);
3719 * Returns with db_holds incremented, and db_mtx not held.
3720 * Note: dn_struct_rwlock must be held.
3723 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3724 boolean_t fail_sparse
, boolean_t fail_uncached
,
3725 const void *tag
, dmu_buf_impl_t
**dbp
)
3727 dmu_buf_impl_t
*db
, *parent
= NULL
;
3730 /* If the pool has been created, verify the tx_sync_lock is not held */
3731 spa_t
*spa
= dn
->dn_objset
->os_spa
;
3732 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
3734 ASSERT(!MUTEX_HELD(&dp
->dp_tx
.tx_sync_lock
));
3737 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3738 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3739 ASSERT3U(dn
->dn_nlevels
, >, level
);
3743 /* dbuf_find() returns with db_mtx held */
3744 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
, &hv
);
3747 blkptr_t
*bp
= NULL
;
3751 return (SET_ERROR(ENOENT
));
3753 ASSERT3P(parent
, ==, NULL
);
3754 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
3756 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
3757 err
= SET_ERROR(ENOENT
);
3760 dbuf_rele(parent
, NULL
);
3764 if (err
&& err
!= ENOENT
)
3766 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
, hv
);
3769 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
3770 mutex_exit(&db
->db_mtx
);
3771 return (SET_ERROR(ENOENT
));
3774 if (db
->db_buf
!= NULL
) {
3775 arc_buf_access(db
->db_buf
);
3776 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
3779 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
3782 * If this buffer is currently syncing out, and we are
3783 * still referencing it from db_data, we need to make a copy
3784 * of it in case we decide we want to dirty it again in this txg.
3786 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
3787 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3788 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
3789 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3790 if (dr
->dt
.dl
.dr_data
== db
->db_buf
) {
3791 ASSERT3P(db
->db_buf
, !=, NULL
);
3792 dbuf_hold_copy(dn
, db
);
3796 if (multilist_link_active(&db
->db_cache_link
)) {
3797 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3798 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3799 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3801 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
3803 uint64_t size
= db
->db
.db_size
+ dmu_buf_user_size(&db
->db
);
3804 (void) zfs_refcount_remove_many(
3805 &dbuf_caches
[db
->db_caching_status
].size
, size
, db
);
3807 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3808 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3810 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3811 DBUF_STAT_BUMPDOWN(cache_count
);
3812 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
], size
);
3814 db
->db_caching_status
= DB_NO_CACHE
;
3816 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3818 mutex_exit(&db
->db_mtx
);
3820 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3822 dbuf_rele(parent
, NULL
);
3824 ASSERT3P(DB_DNODE(db
), ==, dn
);
3825 ASSERT3U(db
->db_blkid
, ==, blkid
);
3826 ASSERT3U(db
->db_level
, ==, level
);
3833 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, const void *tag
)
3835 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3839 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, const void *tag
)
3842 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3843 return (err
? NULL
: db
);
3847 dbuf_create_bonus(dnode_t
*dn
)
3849 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3851 ASSERT(dn
->dn_bonus
== NULL
);
3852 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
,
3853 dbuf_hash(dn
->dn_objset
, dn
->dn_object
, 0, DMU_BONUS_BLKID
));
3857 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3859 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3861 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3862 return (SET_ERROR(ENOTSUP
));
3864 blksz
= SPA_MINBLOCKSIZE
;
3865 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3866 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3868 dbuf_new_size(db
, blksz
, tx
);
3874 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3876 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3879 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3881 dbuf_add_ref(dmu_buf_impl_t
*db
, const void *tag
)
3883 int64_t holds
= zfs_refcount_add(&db
->db_holds
, tag
);
3884 VERIFY3S(holds
, >, 1);
3887 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3889 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3892 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3893 dmu_buf_impl_t
*found_db
;
3894 boolean_t result
= B_FALSE
;
3896 if (blkid
== DMU_BONUS_BLKID
)
3897 found_db
= dbuf_find_bonus(os
, obj
);
3899 found_db
= dbuf_find(os
, obj
, 0, blkid
, NULL
);
3901 if (found_db
!= NULL
) {
3902 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3903 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3906 mutex_exit(&found_db
->db_mtx
);
3912 * If you call dbuf_rele() you had better not be referencing the dnode handle
3913 * unless you have some other direct or indirect hold on the dnode. (An indirect
3914 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3915 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3916 * dnode's parent dbuf evicting its dnode handles.
3919 dbuf_rele(dmu_buf_impl_t
*db
, const void *tag
)
3921 mutex_enter(&db
->db_mtx
);
3922 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
3926 dmu_buf_rele(dmu_buf_t
*db
, const void *tag
)
3928 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3932 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3933 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3934 * argument should be set if we are already in the dbuf-evicting code
3935 * path, in which case we don't want to recursively evict. This allows us to
3936 * avoid deeply nested stacks that would have a call flow similar to this:
3938 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3941 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3945 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, const void *tag
, boolean_t evicting
)
3950 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3954 * Remove the reference to the dbuf before removing its hold on the
3955 * dnode so we can guarantee in dnode_move() that a referenced bonus
3956 * buffer has a corresponding dnode hold.
3958 holds
= zfs_refcount_remove(&db
->db_holds
, tag
);
3962 * We can't freeze indirects if there is a possibility that they
3963 * may be modified in the current syncing context.
3965 if (db
->db_buf
!= NULL
&&
3966 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3967 arc_buf_freeze(db
->db_buf
);
3970 if (holds
== db
->db_dirtycnt
&&
3971 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3972 dbuf_evict_user(db
);
3975 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3977 boolean_t evict_dbuf
= db
->db_pending_evict
;
3980 * If the dnode moves here, we cannot cross this
3981 * barrier until the move completes.
3986 atomic_dec_32(&dn
->dn_dbufs_count
);
3989 * Decrementing the dbuf count means that the bonus
3990 * buffer's dnode hold is no longer discounted in
3991 * dnode_move(). The dnode cannot move until after
3992 * the dnode_rele() below.
3997 * Do not reference db after its lock is dropped.
3998 * Another thread may evict it.
4000 mutex_exit(&db
->db_mtx
);
4003 dnode_evict_bonus(dn
);
4006 } else if (db
->db_buf
== NULL
) {
4008 * This is a special case: we never associated this
4009 * dbuf with any data allocated from the ARC.
4011 ASSERT(db
->db_state
== DB_UNCACHED
||
4012 db
->db_state
== DB_NOFILL
);
4014 } else if (arc_released(db
->db_buf
)) {
4016 * This dbuf has anonymous data associated with it.
4019 } else if (!(DBUF_IS_CACHEABLE(db
) || db
->db_partial_read
) ||
4020 db
->db_pending_evict
) {
4022 } else if (!multilist_link_active(&db
->db_cache_link
)) {
4023 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
4025 dbuf_cached_state_t dcs
=
4026 dbuf_include_in_metadata_cache(db
) ?
4027 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
4028 db
->db_caching_status
= dcs
;
4030 multilist_insert(&dbuf_caches
[dcs
].cache
, db
);
4031 uint64_t db_size
= db
->db
.db_size
+
4032 dmu_buf_user_size(&db
->db
);
4033 size
= zfs_refcount_add_many(
4034 &dbuf_caches
[dcs
].size
, db_size
, db
);
4035 uint8_t db_level
= db
->db_level
;
4036 mutex_exit(&db
->db_mtx
);
4038 if (dcs
== DB_DBUF_METADATA_CACHE
) {
4039 DBUF_STAT_BUMP(metadata_cache_count
);
4040 DBUF_STAT_MAX(metadata_cache_size_bytes_max
,
4043 DBUF_STAT_BUMP(cache_count
);
4044 DBUF_STAT_MAX(cache_size_bytes_max
, size
);
4045 DBUF_STAT_BUMP(cache_levels
[db_level
]);
4046 DBUF_STAT_INCR(cache_levels_bytes
[db_level
],
4050 if (dcs
== DB_DBUF_CACHE
&& !evicting
)
4051 dbuf_evict_notify(size
);
4054 mutex_exit(&db
->db_mtx
);
4059 #pragma weak dmu_buf_refcount = dbuf_refcount
4061 dbuf_refcount(dmu_buf_impl_t
*db
)
4063 return (zfs_refcount_count(&db
->db_holds
));
4067 dmu_buf_user_refcount(dmu_buf_t
*db_fake
)
4070 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4072 mutex_enter(&db
->db_mtx
);
4073 ASSERT3U(zfs_refcount_count(&db
->db_holds
), >=, db
->db_dirtycnt
);
4074 holds
= zfs_refcount_count(&db
->db_holds
) - db
->db_dirtycnt
;
4075 mutex_exit(&db
->db_mtx
);
4081 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
4082 dmu_buf_user_t
*new_user
)
4084 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4086 mutex_enter(&db
->db_mtx
);
4087 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4088 if (db
->db_user
== old_user
)
4089 db
->db_user
= new_user
;
4091 old_user
= db
->db_user
;
4092 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4093 mutex_exit(&db
->db_mtx
);
4099 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4101 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
4105 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4107 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4109 db
->db_user_immediate_evict
= TRUE
;
4110 return (dmu_buf_set_user(db_fake
, user
));
4114 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4116 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
4120 dmu_buf_get_user(dmu_buf_t
*db_fake
)
4122 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4124 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4125 return (db
->db_user
);
4129 dmu_buf_user_size(dmu_buf_t
*db_fake
)
4131 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4132 if (db
->db_user
== NULL
)
4134 return (atomic_load_64(&db
->db_user
->dbu_size
));
4138 dmu_buf_add_user_size(dmu_buf_t
*db_fake
, uint64_t nadd
)
4140 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4141 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
4142 ASSERT3P(db
->db_user
, !=, NULL
);
4143 ASSERT3U(atomic_load_64(&db
->db_user
->dbu_size
), <, UINT64_MAX
- nadd
);
4144 atomic_add_64(&db
->db_user
->dbu_size
, nadd
);
4148 dmu_buf_sub_user_size(dmu_buf_t
*db_fake
, uint64_t nsub
)
4150 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4151 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
4152 ASSERT3P(db
->db_user
, !=, NULL
);
4153 ASSERT3U(atomic_load_64(&db
->db_user
->dbu_size
), >=, nsub
);
4154 atomic_sub_64(&db
->db_user
->dbu_size
, nsub
);
4158 dmu_buf_user_evict_wait(void)
4160 taskq_wait(dbu_evict_taskq
);
4164 dmu_buf_get_blkptr(dmu_buf_t
*db
)
4166 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4167 return (dbi
->db_blkptr
);
4171 dmu_buf_get_objset(dmu_buf_t
*db
)
4173 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4174 return (dbi
->db_objset
);
4178 dmu_buf_dnode_enter(dmu_buf_t
*db
)
4180 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4181 DB_DNODE_ENTER(dbi
);
4182 return (DB_DNODE(dbi
));
4186 dmu_buf_dnode_exit(dmu_buf_t
*db
)
4188 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4193 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
4195 /* ASSERT(dmu_tx_is_syncing(tx) */
4196 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4198 if (db
->db_blkptr
!= NULL
)
4201 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4202 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
4203 BP_ZERO(db
->db_blkptr
);
4206 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
4208 * This buffer was allocated at a time when there was
4209 * no available blkptrs from the dnode, or it was
4210 * inappropriate to hook it in (i.e., nlevels mismatch).
4212 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
4213 ASSERT(db
->db_parent
== NULL
);
4214 db
->db_parent
= dn
->dn_dbuf
;
4215 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
4218 dmu_buf_impl_t
*parent
= db
->db_parent
;
4219 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4221 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
4222 if (parent
== NULL
) {
4223 mutex_exit(&db
->db_mtx
);
4224 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
4225 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
4226 db
->db_blkid
>> epbs
, db
);
4227 rw_exit(&dn
->dn_struct_rwlock
);
4228 mutex_enter(&db
->db_mtx
);
4229 db
->db_parent
= parent
;
4231 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
4232 (db
->db_blkid
& ((1ULL << epbs
) - 1));
4238 dbuf_sync_bonus(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4240 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4241 void *data
= dr
->dt
.dl
.dr_data
;
4243 ASSERT0(db
->db_level
);
4244 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4245 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
);
4246 ASSERT(data
!= NULL
);
4248 dnode_t
*dn
= dr
->dr_dnode
;
4249 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
4250 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
4251 memcpy(DN_BONUS(dn
->dn_phys
), data
, DN_MAX_BONUS_LEN(dn
->dn_phys
));
4253 dbuf_sync_leaf_verify_bonus_dnode(dr
);
4255 dbuf_undirty_bonus(dr
);
4256 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4260 * When syncing out a blocks of dnodes, adjust the block to deal with
4261 * encryption. Normally, we make sure the block is decrypted before writing
4262 * it. If we have crypt params, then we are writing a raw (encrypted) block,
4263 * from a raw receive. In this case, set the ARC buf's crypt params so
4264 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
4267 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t
*dr
)
4270 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4272 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4273 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
4274 ASSERT3U(db
->db_level
, ==, 0);
4276 if (!db
->db_objset
->os_raw_receive
&& arc_is_encrypted(db
->db_buf
)) {
4277 zbookmark_phys_t zb
;
4280 * Unfortunately, there is currently no mechanism for
4281 * syncing context to handle decryption errors. An error
4282 * here is only possible if an attacker maliciously
4283 * changed a dnode block and updated the associated
4284 * checksums going up the block tree.
4286 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
4287 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4288 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
4291 panic("Invalid dnode block MAC");
4292 } else if (dr
->dt
.dl
.dr_has_raw_params
) {
4293 (void) arc_release(dr
->dt
.dl
.dr_data
, db
);
4294 arc_convert_to_raw(dr
->dt
.dl
.dr_data
,
4295 dmu_objset_id(db
->db_objset
),
4296 dr
->dt
.dl
.dr_byteorder
, DMU_OT_DNODE
,
4297 dr
->dt
.dl
.dr_salt
, dr
->dt
.dl
.dr_iv
, dr
->dt
.dl
.dr_mac
);
4302 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
4303 * is critical the we not allow the compiler to inline this function in to
4304 * dbuf_sync_list() thereby drastically bloating the stack usage.
4306 noinline
static void
4307 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4309 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4310 dnode_t
*dn
= dr
->dr_dnode
;
4312 ASSERT(dmu_tx_is_syncing(tx
));
4314 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4316 mutex_enter(&db
->db_mtx
);
4318 ASSERT(db
->db_level
> 0);
4321 /* Read the block if it hasn't been read yet. */
4322 if (db
->db_buf
== NULL
) {
4323 mutex_exit(&db
->db_mtx
);
4324 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
4325 mutex_enter(&db
->db_mtx
);
4327 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
4328 ASSERT(db
->db_buf
!= NULL
);
4330 /* Indirect block size must match what the dnode thinks it is. */
4331 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4332 dbuf_check_blkptr(dn
, db
);
4334 /* Provide the pending dirty record to child dbufs */
4335 db
->db_data_pending
= dr
;
4337 mutex_exit(&db
->db_mtx
);
4339 dbuf_write(dr
, db
->db_buf
, tx
);
4341 zio_t
*zio
= dr
->dr_zio
;
4342 mutex_enter(&dr
->dt
.di
.dr_mtx
);
4343 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
4344 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4345 mutex_exit(&dr
->dt
.di
.dr_mtx
);
4350 * Verify that the size of the data in our bonus buffer does not exceed
4351 * its recorded size.
4353 * The purpose of this verification is to catch any cases in development
4354 * where the size of a phys structure (i.e space_map_phys_t) grows and,
4355 * due to incorrect feature management, older pools expect to read more
4356 * data even though they didn't actually write it to begin with.
4358 * For a example, this would catch an error in the feature logic where we
4359 * open an older pool and we expect to write the space map histogram of
4360 * a space map with size SPACE_MAP_SIZE_V0.
4363 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
)
4366 dnode_t
*dn
= dr
->dr_dnode
;
4369 * Encrypted bonus buffers can have data past their bonuslen.
4370 * Skip the verification of these blocks.
4372 if (DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))
4375 uint16_t bonuslen
= dn
->dn_phys
->dn_bonuslen
;
4376 uint16_t maxbonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
4377 ASSERT3U(bonuslen
, <=, maxbonuslen
);
4379 arc_buf_t
*datap
= dr
->dt
.dl
.dr_data
;
4380 char *datap_end
= ((char *)datap
) + bonuslen
;
4381 char *datap_max
= ((char *)datap
) + maxbonuslen
;
4383 /* ensure that everything is zero after our data */
4384 for (; datap_end
< datap_max
; datap_end
++)
4385 ASSERT(*datap_end
== 0);
4390 dbuf_lightweight_bp(dbuf_dirty_record_t
*dr
)
4392 /* This must be a lightweight dirty record. */
4393 ASSERT3P(dr
->dr_dbuf
, ==, NULL
);
4394 dnode_t
*dn
= dr
->dr_dnode
;
4396 if (dn
->dn_phys
->dn_nlevels
== 1) {
4397 VERIFY3U(dr
->dt
.dll
.dr_blkid
, <, dn
->dn_phys
->dn_nblkptr
);
4398 return (&dn
->dn_phys
->dn_blkptr
[dr
->dt
.dll
.dr_blkid
]);
4400 dmu_buf_impl_t
*parent_db
= dr
->dr_parent
->dr_dbuf
;
4401 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4402 VERIFY3U(parent_db
->db_level
, ==, 1);
4403 VERIFY3P(parent_db
->db_dnode_handle
->dnh_dnode
, ==, dn
);
4404 VERIFY3U(dr
->dt
.dll
.dr_blkid
>> epbs
, ==, parent_db
->db_blkid
);
4405 blkptr_t
*bp
= parent_db
->db
.db_data
;
4406 return (&bp
[dr
->dt
.dll
.dr_blkid
& ((1 << epbs
) - 1)]);
4411 dbuf_lightweight_ready(zio_t
*zio
)
4413 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4414 blkptr_t
*bp
= zio
->io_bp
;
4416 if (zio
->io_error
!= 0)
4419 dnode_t
*dn
= dr
->dr_dnode
;
4421 blkptr_t
*bp_orig
= dbuf_lightweight_bp(dr
);
4422 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4423 int64_t delta
= bp_get_dsize_sync(spa
, bp
) -
4424 bp_get_dsize_sync(spa
, bp_orig
);
4425 dnode_diduse_space(dn
, delta
);
4427 uint64_t blkid
= dr
->dt
.dll
.dr_blkid
;
4428 mutex_enter(&dn
->dn_mtx
);
4429 if (blkid
> dn
->dn_phys
->dn_maxblkid
) {
4430 ASSERT0(dn
->dn_objset
->os_raw_receive
);
4431 dn
->dn_phys
->dn_maxblkid
= blkid
;
4433 mutex_exit(&dn
->dn_mtx
);
4435 if (!BP_IS_EMBEDDED(bp
)) {
4436 uint64_t fill
= BP_IS_HOLE(bp
) ? 0 : 1;
4437 BP_SET_FILL(bp
, fill
);
4440 dmu_buf_impl_t
*parent_db
;
4441 EQUIV(dr
->dr_parent
== NULL
, dn
->dn_phys
->dn_nlevels
== 1);
4442 if (dr
->dr_parent
== NULL
) {
4443 parent_db
= dn
->dn_dbuf
;
4445 parent_db
= dr
->dr_parent
->dr_dbuf
;
4447 rw_enter(&parent_db
->db_rwlock
, RW_WRITER
);
4449 rw_exit(&parent_db
->db_rwlock
);
4453 dbuf_lightweight_done(zio_t
*zio
)
4455 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4457 VERIFY0(zio
->io_error
);
4459 objset_t
*os
= dr
->dr_dnode
->dn_objset
;
4460 dmu_tx_t
*tx
= os
->os_synctx
;
4462 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4463 ASSERT(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4465 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4466 (void) dsl_dataset_block_kill(ds
, &zio
->io_bp_orig
, tx
, B_TRUE
);
4467 dsl_dataset_block_born(ds
, zio
->io_bp
, tx
);
4470 dsl_pool_undirty_space(dmu_objset_pool(os
), dr
->dr_accounted
,
4473 abd_free(dr
->dt
.dll
.dr_abd
);
4474 kmem_free(dr
, sizeof (*dr
));
4477 noinline
static void
4478 dbuf_sync_lightweight(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4480 dnode_t
*dn
= dr
->dr_dnode
;
4482 if (dn
->dn_phys
->dn_nlevels
== 1) {
4485 pio
= dr
->dr_parent
->dr_zio
;
4488 zbookmark_phys_t zb
= {
4489 .zb_objset
= dmu_objset_id(dn
->dn_objset
),
4490 .zb_object
= dn
->dn_object
,
4492 .zb_blkid
= dr
->dt
.dll
.dr_blkid
,
4496 * See comment in dbuf_write(). This is so that zio->io_bp_orig
4497 * will have the old BP in dbuf_lightweight_done().
4499 dr
->dr_bp_copy
= *dbuf_lightweight_bp(dr
);
4501 dr
->dr_zio
= zio_write(pio
, dmu_objset_spa(dn
->dn_objset
),
4502 dmu_tx_get_txg(tx
), &dr
->dr_bp_copy
, dr
->dt
.dll
.dr_abd
,
4503 dn
->dn_datablksz
, abd_get_size(dr
->dt
.dll
.dr_abd
),
4504 &dr
->dt
.dll
.dr_props
, dbuf_lightweight_ready
, NULL
,
4505 dbuf_lightweight_done
, dr
, ZIO_PRIORITY_ASYNC_WRITE
,
4506 ZIO_FLAG_MUSTSUCCEED
| dr
->dt
.dll
.dr_flags
, &zb
);
4508 zio_nowait(dr
->dr_zio
);
4512 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
4513 * critical the we not allow the compiler to inline this function in to
4514 * dbuf_sync_list() thereby drastically bloating the stack usage.
4516 noinline
static void
4517 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4519 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
4520 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4521 dnode_t
*dn
= dr
->dr_dnode
;
4523 uint64_t txg
= tx
->tx_txg
;
4525 ASSERT(dmu_tx_is_syncing(tx
));
4527 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4529 mutex_enter(&db
->db_mtx
);
4531 * To be synced, we must be dirtied. But we
4532 * might have been freed after the dirty.
4534 if (db
->db_state
== DB_UNCACHED
) {
4535 /* This buffer has been freed since it was dirtied */
4536 ASSERT(db
->db
.db_data
== NULL
);
4537 } else if (db
->db_state
== DB_FILL
) {
4538 /* This buffer was freed and is now being re-filled */
4539 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
4540 } else if (db
->db_state
== DB_READ
) {
4542 * This buffer has a clone we need to write, and an in-flight
4543 * read on the BP we're about to clone. Its safe to issue the
4544 * write here because the read has already been issued and the
4545 * contents won't change.
4547 ASSERT(dr
->dt
.dl
.dr_brtwrite
&&
4548 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
);
4550 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
4554 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4555 mutex_enter(&dn
->dn_mtx
);
4556 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
4558 * In the previous transaction group, the bonus buffer
4559 * was entirely used to store the attributes for the
4560 * dnode which overrode the dn_spill field. However,
4561 * when adding more attributes to the file a spill
4562 * block was required to hold the extra attributes.
4564 * Make sure to clear the garbage left in the dn_spill
4565 * field from the previous attributes in the bonus
4566 * buffer. Otherwise, after writing out the spill
4567 * block to the new allocated dva, it will free
4568 * the old block pointed to by the invalid dn_spill.
4570 db
->db_blkptr
= NULL
;
4572 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
4573 mutex_exit(&dn
->dn_mtx
);
4577 * If this is a bonus buffer, simply copy the bonus data into the
4578 * dnode. It will be written out when the dnode is synced (and it
4579 * will be synced, since it must have been dirty for dbuf_sync to
4582 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
4583 ASSERT(dr
->dr_dbuf
== db
);
4584 dbuf_sync_bonus(dr
, tx
);
4591 * This function may have dropped the db_mtx lock allowing a dmu_sync
4592 * operation to sneak in. As a result, we need to ensure that we
4593 * don't check the dr_override_state until we have returned from
4594 * dbuf_check_blkptr.
4596 dbuf_check_blkptr(dn
, db
);
4599 * If this buffer is in the middle of an immediate write,
4600 * wait for the synchronous IO to complete.
4602 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
4603 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
4604 cv_wait(&db
->db_changed
, &db
->db_mtx
);
4608 * If this is a dnode block, ensure it is appropriately encrypted
4609 * or decrypted, depending on what we are writing to it this txg.
4611 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
4612 dbuf_prepare_encrypted_dnode_leaf(dr
);
4614 if (db
->db_state
!= DB_NOFILL
&&
4615 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
4616 zfs_refcount_count(&db
->db_holds
) > 1 &&
4617 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
4618 *datap
== db
->db_buf
) {
4620 * If this buffer is currently "in use" (i.e., there
4621 * are active holds and db_data still references it),
4622 * then make a copy before we start the write so that
4623 * any modifications from the open txg will not leak
4626 * NOTE: this copy does not need to be made for
4627 * objects only modified in the syncing context (e.g.
4628 * DNONE_DNODE blocks).
4630 int psize
= arc_buf_size(*datap
);
4631 int lsize
= arc_buf_lsize(*datap
);
4632 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
4633 enum zio_compress compress_type
= arc_get_compression(*datap
);
4634 uint8_t complevel
= arc_get_complevel(*datap
);
4636 if (arc_is_encrypted(*datap
)) {
4637 boolean_t byteorder
;
4638 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4639 uint8_t iv
[ZIO_DATA_IV_LEN
];
4640 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4642 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
4643 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
4644 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
4645 dn
->dn_type
, psize
, lsize
, compress_type
,
4647 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
4648 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
4649 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
4650 psize
, lsize
, compress_type
, complevel
);
4652 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
4654 memcpy((*datap
)->b_data
, db
->db
.db_data
, psize
);
4656 db
->db_data_pending
= dr
;
4658 mutex_exit(&db
->db_mtx
);
4660 dbuf_write(dr
, *datap
, tx
);
4662 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4663 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
4664 list_insert_tail(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
4666 zio_nowait(dr
->dr_zio
);
4671 * Syncs out a range of dirty records for indirect or leaf dbufs. May be
4672 * called recursively from dbuf_sync_indirect().
4675 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
4677 dbuf_dirty_record_t
*dr
;
4679 while ((dr
= list_head(list
))) {
4680 if (dr
->dr_zio
!= NULL
) {
4682 * If we find an already initialized zio then we
4683 * are processing the meta-dnode, and we have finished.
4684 * The dbufs for all dnodes are put back on the list
4685 * during processing, so that we can zio_wait()
4686 * these IOs after initiating all child IOs.
4688 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
4689 DMU_META_DNODE_OBJECT
);
4692 list_remove(list
, dr
);
4693 if (dr
->dr_dbuf
== NULL
) {
4694 dbuf_sync_lightweight(dr
, tx
);
4696 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
4697 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
4698 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
4700 if (dr
->dr_dbuf
->db_level
> 0)
4701 dbuf_sync_indirect(dr
, tx
);
4703 dbuf_sync_leaf(dr
, tx
);
4709 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4712 dmu_buf_impl_t
*db
= vdb
;
4714 blkptr_t
*bp
= zio
->io_bp
;
4715 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4716 spa_t
*spa
= zio
->io_spa
;
4721 ASSERT3P(db
->db_blkptr
, !=, NULL
);
4722 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
4726 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
4727 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
4728 zio
->io_prev_space_delta
= delta
;
4730 if (bp
->blk_birth
!= 0) {
4731 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
4732 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
4733 (db
->db_blkid
== DMU_SPILL_BLKID
&&
4734 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
4735 BP_IS_EMBEDDED(bp
));
4736 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
4739 mutex_enter(&db
->db_mtx
);
4742 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4743 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4744 ASSERT(!(BP_IS_HOLE(bp
)) &&
4745 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4749 if (db
->db_level
== 0) {
4750 mutex_enter(&dn
->dn_mtx
);
4751 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
4752 db
->db_blkid
!= DMU_SPILL_BLKID
) {
4753 ASSERT0(db
->db_objset
->os_raw_receive
);
4754 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
4756 mutex_exit(&dn
->dn_mtx
);
4758 if (dn
->dn_type
== DMU_OT_DNODE
) {
4760 while (i
< db
->db
.db_size
) {
4762 (void *)(((char *)db
->db
.db_data
) + i
);
4764 i
+= DNODE_MIN_SIZE
;
4765 if (dnp
->dn_type
!= DMU_OT_NONE
) {
4767 for (int j
= 0; j
< dnp
->dn_nblkptr
;
4769 (void) zfs_blkptr_verify(spa
,
4775 DNODE_FLAG_SPILL_BLKPTR
) {
4776 (void) zfs_blkptr_verify(spa
,
4777 DN_SPILL_BLKPTR(dnp
),
4781 i
+= dnp
->dn_extra_slots
*
4786 if (BP_IS_HOLE(bp
)) {
4793 blkptr_t
*ibp
= db
->db
.db_data
;
4794 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4795 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
4796 if (BP_IS_HOLE(ibp
))
4798 (void) zfs_blkptr_verify(spa
, ibp
,
4799 BLK_CONFIG_SKIP
, BLK_VERIFY_HALT
);
4800 fill
+= BP_GET_FILL(ibp
);
4805 if (!BP_IS_EMBEDDED(bp
))
4806 BP_SET_FILL(bp
, fill
);
4808 mutex_exit(&db
->db_mtx
);
4810 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_WRITER
, FTAG
);
4811 *db
->db_blkptr
= *bp
;
4812 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
4816 * This function gets called just prior to running through the compression
4817 * stage of the zio pipeline. If we're an indirect block comprised of only
4818 * holes, then we want this indirect to be compressed away to a hole. In
4819 * order to do that we must zero out any information about the holes that
4820 * this indirect points to prior to before we try to compress it.
4823 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4825 (void) zio
, (void) buf
;
4826 dmu_buf_impl_t
*db
= vdb
;
4829 unsigned int epbs
, i
;
4831 ASSERT3U(db
->db_level
, >, 0);
4834 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4835 ASSERT3U(epbs
, <, 31);
4837 /* Determine if all our children are holes */
4838 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
4839 if (!BP_IS_HOLE(bp
))
4844 * If all the children are holes, then zero them all out so that
4845 * we may get compressed away.
4847 if (i
== 1ULL << epbs
) {
4849 * We only found holes. Grab the rwlock to prevent
4850 * anybody from reading the blocks we're about to
4853 rw_enter(&db
->db_rwlock
, RW_WRITER
);
4854 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
4855 rw_exit(&db
->db_rwlock
);
4861 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4864 dmu_buf_impl_t
*db
= vdb
;
4865 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4866 blkptr_t
*bp
= db
->db_blkptr
;
4867 objset_t
*os
= db
->db_objset
;
4868 dmu_tx_t
*tx
= os
->os_synctx
;
4870 ASSERT0(zio
->io_error
);
4871 ASSERT(db
->db_blkptr
== bp
);
4874 * For nopwrites and rewrites we ensure that the bp matches our
4875 * original and bypass all the accounting.
4877 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4878 ASSERT(BP_EQUAL(bp
, bp_orig
));
4880 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4881 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
4882 dsl_dataset_block_born(ds
, bp
, tx
);
4885 mutex_enter(&db
->db_mtx
);
4889 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
4890 dnode_t
*dn
= dr
->dr_dnode
;
4891 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4892 ASSERT(dr
->dr_dbuf
== db
);
4893 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
4894 list_remove(&db
->db_dirty_records
, dr
);
4897 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4898 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4899 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
4900 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4904 if (db
->db_level
== 0) {
4905 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
4906 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
4907 if (db
->db_state
!= DB_NOFILL
) {
4908 if (dr
->dt
.dl
.dr_data
!= NULL
&&
4909 dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
4910 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
4914 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4915 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
4916 if (!BP_IS_HOLE(db
->db_blkptr
)) {
4917 int epbs __maybe_unused
= dn
->dn_phys
->dn_indblkshift
-
4919 ASSERT3U(db
->db_blkid
, <=,
4920 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
4921 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4924 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4925 list_destroy(&dr
->dt
.di
.dr_children
);
4928 cv_broadcast(&db
->db_changed
);
4929 ASSERT(db
->db_dirtycnt
> 0);
4930 db
->db_dirtycnt
-= 1;
4931 db
->db_data_pending
= NULL
;
4932 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4934 dsl_pool_undirty_space(dmu_objset_pool(os
), dr
->dr_accounted
,
4937 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4941 dbuf_write_nofill_ready(zio_t
*zio
)
4943 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4947 dbuf_write_nofill_done(zio_t
*zio
)
4949 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4953 dbuf_write_override_ready(zio_t
*zio
)
4955 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4956 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4958 dbuf_write_ready(zio
, NULL
, db
);
4962 dbuf_write_override_done(zio_t
*zio
)
4964 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4965 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4966 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4968 mutex_enter(&db
->db_mtx
);
4969 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4970 if (!BP_IS_HOLE(obp
))
4971 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4972 arc_release(dr
->dt
.dl
.dr_data
, db
);
4974 mutex_exit(&db
->db_mtx
);
4976 dbuf_write_done(zio
, NULL
, db
);
4978 if (zio
->io_abd
!= NULL
)
4979 abd_free(zio
->io_abd
);
4982 typedef struct dbuf_remap_impl_callback_arg
{
4984 uint64_t drica_blk_birth
;
4986 } dbuf_remap_impl_callback_arg_t
;
4989 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
4992 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
4993 objset_t
*os
= drica
->drica_os
;
4994 spa_t
*spa
= dmu_objset_spa(os
);
4995 dmu_tx_t
*tx
= drica
->drica_tx
;
4997 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4999 if (os
== spa_meta_objset(spa
)) {
5000 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
5002 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
5003 size
, drica
->drica_blk_birth
, tx
);
5008 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, krwlock_t
*rw
, dmu_tx_t
*tx
)
5010 blkptr_t bp_copy
= *bp
;
5011 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
5012 dbuf_remap_impl_callback_arg_t drica
;
5014 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
5016 drica
.drica_os
= dn
->dn_objset
;
5017 drica
.drica_blk_birth
= bp
->blk_birth
;
5018 drica
.drica_tx
= tx
;
5019 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
5022 * If the blkptr being remapped is tracked by a livelist,
5023 * then we need to make sure the livelist reflects the update.
5024 * First, cancel out the old blkptr by appending a 'FREE'
5025 * entry. Next, add an 'ALLOC' to track the new version. This
5026 * way we avoid trying to free an inaccurate blkptr at delete.
5027 * Note that embedded blkptrs are not tracked in livelists.
5029 if (dn
->dn_objset
!= spa_meta_objset(spa
)) {
5030 dsl_dataset_t
*ds
= dmu_objset_ds(dn
->dn_objset
);
5031 if (dsl_deadlist_is_open(&ds
->ds_dir
->dd_livelist
) &&
5032 bp
->blk_birth
> ds
->ds_dir
->dd_origin_txg
) {
5033 ASSERT(!BP_IS_EMBEDDED(bp
));
5034 ASSERT(dsl_dir_is_clone(ds
->ds_dir
));
5035 ASSERT(spa_feature_is_enabled(spa
,
5036 SPA_FEATURE_LIVELIST
));
5037 bplist_append(&ds
->ds_dir
->dd_pending_frees
,
5039 bplist_append(&ds
->ds_dir
->dd_pending_allocs
,
5045 * The db_rwlock prevents dbuf_read_impl() from
5046 * dereferencing the BP while we are changing it. To
5047 * avoid lock contention, only grab it when we are actually
5051 rw_enter(rw
, RW_WRITER
);
5059 * Remap any existing BP's to concrete vdevs, if possible.
5062 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
5064 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
5065 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
5067 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
5070 if (db
->db_level
> 0) {
5071 blkptr_t
*bp
= db
->db
.db_data
;
5072 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
5073 dbuf_remap_impl(dn
, &bp
[i
], &db
->db_rwlock
, tx
);
5075 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
5076 dnode_phys_t
*dnp
= db
->db
.db_data
;
5077 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
5079 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
;
5080 i
+= dnp
[i
].dn_extra_slots
+ 1) {
5081 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
5082 krwlock_t
*lock
= (dn
->dn_dbuf
== NULL
? NULL
:
5083 &dn
->dn_dbuf
->db_rwlock
);
5084 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], lock
,
5093 * Populate dr->dr_zio with a zio to commit a dirty buffer to disk.
5094 * Caller is responsible for issuing the zio_[no]wait(dr->dr_zio).
5097 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
5099 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
5100 dnode_t
*dn
= dr
->dr_dnode
;
5102 dmu_buf_impl_t
*parent
= db
->db_parent
;
5103 uint64_t txg
= tx
->tx_txg
;
5104 zbookmark_phys_t zb
;
5106 zio_t
*pio
; /* parent I/O */
5109 ASSERT(dmu_tx_is_syncing(tx
));
5113 if (db
->db_state
!= DB_NOFILL
) {
5114 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
5116 * Private object buffers are released here rather
5117 * than in dbuf_dirty() since they are only modified
5118 * in the syncing context and we don't want the
5119 * overhead of making multiple copies of the data.
5121 if (BP_IS_HOLE(db
->db_blkptr
)) {
5124 dbuf_release_bp(db
);
5126 dbuf_remap(dn
, db
, tx
);
5130 if (parent
!= dn
->dn_dbuf
) {
5131 /* Our parent is an indirect block. */
5132 /* We have a dirty parent that has been scheduled for write. */
5133 ASSERT(parent
&& parent
->db_data_pending
);
5134 /* Our parent's buffer is one level closer to the dnode. */
5135 ASSERT(db
->db_level
== parent
->db_level
-1);
5137 * We're about to modify our parent's db_data by modifying
5138 * our block pointer, so the parent must be released.
5140 ASSERT(arc_released(parent
->db_buf
));
5141 pio
= parent
->db_data_pending
->dr_zio
;
5143 /* Our parent is the dnode itself. */
5144 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
5145 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
5146 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
5147 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
5148 ASSERT3P(db
->db_blkptr
, ==,
5149 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
5153 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
5154 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
5157 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
5158 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
5159 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
5161 if (db
->db_blkid
== DMU_SPILL_BLKID
)
5163 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
5165 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
5168 * We copy the blkptr now (rather than when we instantiate the dirty
5169 * record), because its value can change between open context and
5170 * syncing context. We do not need to hold dn_struct_rwlock to read
5171 * db_blkptr because we are in syncing context.
5173 dr
->dr_bp_copy
= *db
->db_blkptr
;
5175 if (db
->db_level
== 0 &&
5176 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
5178 * The BP for this block has been provided by open context
5179 * (by dmu_sync() or dmu_buf_write_embedded()).
5181 abd_t
*contents
= (data
!= NULL
) ?
5182 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
5184 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
, &dr
->dr_bp_copy
,
5185 contents
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5186 dbuf_write_override_ready
, NULL
,
5187 dbuf_write_override_done
,
5188 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
5189 mutex_enter(&db
->db_mtx
);
5190 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
5191 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
5192 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
,
5193 dr
->dt
.dl
.dr_brtwrite
);
5194 mutex_exit(&db
->db_mtx
);
5195 } else if (db
->db_state
== DB_NOFILL
) {
5196 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
5197 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
5198 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
,
5199 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5200 dbuf_write_nofill_ready
, NULL
,
5201 dbuf_write_nofill_done
, db
,
5202 ZIO_PRIORITY_ASYNC_WRITE
,
5203 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
5205 ASSERT(arc_released(data
));
5208 * For indirect blocks, we want to setup the children
5209 * ready callback so that we can properly handle an indirect
5210 * block that only contains holes.
5212 arc_write_done_func_t
*children_ready_cb
= NULL
;
5213 if (db
->db_level
!= 0)
5214 children_ready_cb
= dbuf_write_children_ready
;
5216 dr
->dr_zio
= arc_write(pio
, os
->os_spa
, txg
,
5217 &dr
->dr_bp_copy
, data
, !DBUF_IS_CACHEABLE(db
),
5218 dbuf_is_l2cacheable(db
), &zp
, dbuf_write_ready
,
5219 children_ready_cb
, dbuf_write_done
, db
,
5220 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
5224 EXPORT_SYMBOL(dbuf_find
);
5225 EXPORT_SYMBOL(dbuf_is_metadata
);
5226 EXPORT_SYMBOL(dbuf_destroy
);
5227 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
5228 EXPORT_SYMBOL(dbuf_whichblock
);
5229 EXPORT_SYMBOL(dbuf_read
);
5230 EXPORT_SYMBOL(dbuf_unoverride
);
5231 EXPORT_SYMBOL(dbuf_free_range
);
5232 EXPORT_SYMBOL(dbuf_new_size
);
5233 EXPORT_SYMBOL(dbuf_release_bp
);
5234 EXPORT_SYMBOL(dbuf_dirty
);
5235 EXPORT_SYMBOL(dmu_buf_set_crypt_params
);
5236 EXPORT_SYMBOL(dmu_buf_will_dirty
);
5237 EXPORT_SYMBOL(dmu_buf_is_dirty
);
5238 EXPORT_SYMBOL(dmu_buf_will_clone
);
5239 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
5240 EXPORT_SYMBOL(dmu_buf_will_fill
);
5241 EXPORT_SYMBOL(dmu_buf_fill_done
);
5242 EXPORT_SYMBOL(dmu_buf_rele
);
5243 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
5244 EXPORT_SYMBOL(dbuf_prefetch
);
5245 EXPORT_SYMBOL(dbuf_hold_impl
);
5246 EXPORT_SYMBOL(dbuf_hold
);
5247 EXPORT_SYMBOL(dbuf_hold_level
);
5248 EXPORT_SYMBOL(dbuf_create_bonus
);
5249 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
5250 EXPORT_SYMBOL(dbuf_rm_spill
);
5251 EXPORT_SYMBOL(dbuf_add_ref
);
5252 EXPORT_SYMBOL(dbuf_rele
);
5253 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
5254 EXPORT_SYMBOL(dbuf_refcount
);
5255 EXPORT_SYMBOL(dbuf_sync_list
);
5256 EXPORT_SYMBOL(dmu_buf_set_user
);
5257 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
5258 EXPORT_SYMBOL(dmu_buf_get_user
);
5259 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
5261 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, max_bytes
, U64
, ZMOD_RW
,
5262 "Maximum size in bytes of the dbuf cache.");
5264 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, hiwater_pct
, UINT
, ZMOD_RW
,
5265 "Percentage over dbuf_cache_max_bytes for direct dbuf eviction.");
5267 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, lowater_pct
, UINT
, ZMOD_RW
,
5268 "Percentage below dbuf_cache_max_bytes when dbuf eviction stops.");
5270 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_max_bytes
, U64
, ZMOD_RW
,
5271 "Maximum size in bytes of dbuf metadata cache.");
5273 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, cache_shift
, UINT
, ZMOD_RW
,
5274 "Set size of dbuf cache to log2 fraction of arc size.");
5276 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_shift
, UINT
, ZMOD_RW
,
5277 "Set size of dbuf metadata cache to log2 fraction of arc size.");
5279 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, mutex_cache_shift
, UINT
, ZMOD_RD
,
5280 "Set size of dbuf cache mutex array as log2 shift.");