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
;
573 * There are two eviction callbacks - one that we call synchronously
574 * and one that we invoke via a taskq. The async one is useful for
575 * avoiding lock order reversals and limiting stack depth.
577 * Note that if we have a sync callback but no async callback,
578 * it's likely that the sync callback will free the structure
579 * containing the dbu. In that case we need to take care to not
580 * dereference dbu after calling the sync evict func.
582 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
584 if (dbu
->dbu_evict_func_sync
!= NULL
)
585 dbu
->dbu_evict_func_sync(dbu
);
588 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
589 dbu
, 0, &dbu
->dbu_tqent
);
594 dbuf_is_metadata(dmu_buf_impl_t
*db
)
597 * Consider indirect blocks and spill blocks to be meta data.
599 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
602 boolean_t is_metadata
;
605 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
608 return (is_metadata
);
613 * We want to exclude buffers that are on a special allocation class from
617 dbuf_is_l2cacheable(dmu_buf_impl_t
*db
)
619 if (db
->db_objset
->os_secondary_cache
== ZFS_CACHE_ALL
||
620 (db
->db_objset
->os_secondary_cache
==
621 ZFS_CACHE_METADATA
&& dbuf_is_metadata(db
))) {
622 if (l2arc_exclude_special
== 0)
625 blkptr_t
*bp
= db
->db_blkptr
;
626 if (bp
== NULL
|| BP_IS_HOLE(bp
))
628 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
629 vdev_t
*rvd
= db
->db_objset
->os_spa
->spa_root_vdev
;
632 if (vdev
< rvd
->vdev_children
)
633 vd
= rvd
->vdev_child
[vdev
];
638 if (vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
639 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
)
645 static inline boolean_t
646 dnode_level_is_l2cacheable(blkptr_t
*bp
, dnode_t
*dn
, int64_t level
)
648 if (dn
->dn_objset
->os_secondary_cache
== ZFS_CACHE_ALL
||
649 (dn
->dn_objset
->os_secondary_cache
== ZFS_CACHE_METADATA
&&
651 DMU_OT_IS_METADATA(dn
->dn_handle
->dnh_dnode
->dn_type
)))) {
652 if (l2arc_exclude_special
== 0)
655 if (bp
== NULL
|| BP_IS_HOLE(bp
))
657 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
658 vdev_t
*rvd
= dn
->dn_objset
->os_spa
->spa_root_vdev
;
661 if (vdev
< rvd
->vdev_children
)
662 vd
= rvd
->vdev_child
[vdev
];
667 if (vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
668 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
)
676 * This function *must* return indices evenly distributed between all
677 * sublists of the multilist. This is needed due to how the dbuf eviction
678 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
679 * distributed between all sublists and uses this assumption when
680 * deciding which sublist to evict from and how much to evict from it.
683 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
685 dmu_buf_impl_t
*db
= obj
;
688 * The assumption here, is the hash value for a given
689 * dmu_buf_impl_t will remain constant throughout it's lifetime
690 * (i.e. it's objset, object, level and blkid fields don't change).
691 * Thus, we don't need to store the dbuf's sublist index
692 * on insertion, as this index can be recalculated on removal.
694 * Also, the low order bits of the hash value are thought to be
695 * distributed evenly. Otherwise, in the case that the multilist
696 * has a power of two number of sublists, each sublists' usage
697 * would not be evenly distributed. In this context full 64bit
698 * division would be a waste of time, so limit it to 32 bits.
700 return ((unsigned int)dbuf_hash(db
->db_objset
, db
->db
.db_object
,
701 db
->db_level
, db
->db_blkid
) %
702 multilist_get_num_sublists(ml
));
706 * The target size of the dbuf cache can grow with the ARC target,
707 * unless limited by the tunable dbuf_cache_max_bytes.
709 static inline unsigned long
710 dbuf_cache_target_bytes(void)
712 return (MIN(dbuf_cache_max_bytes
,
713 arc_target_bytes() >> dbuf_cache_shift
));
717 * The target size of the dbuf metadata cache can grow with the ARC target,
718 * unless limited by the tunable dbuf_metadata_cache_max_bytes.
720 static inline unsigned long
721 dbuf_metadata_cache_target_bytes(void)
723 return (MIN(dbuf_metadata_cache_max_bytes
,
724 arc_target_bytes() >> dbuf_metadata_cache_shift
));
727 static inline uint64_t
728 dbuf_cache_hiwater_bytes(void)
730 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
731 return (dbuf_cache_target
+
732 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100);
735 static inline uint64_t
736 dbuf_cache_lowater_bytes(void)
738 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
739 return (dbuf_cache_target
-
740 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100);
743 static inline boolean_t
744 dbuf_cache_above_lowater(void)
746 return (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
747 dbuf_cache_lowater_bytes());
751 * Evict the oldest eligible dbuf from the dbuf cache.
756 int idx
= multilist_get_random_index(&dbuf_caches
[DB_DBUF_CACHE
].cache
);
757 multilist_sublist_t
*mls
= multilist_sublist_lock(
758 &dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
760 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
762 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
763 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
764 db
= multilist_sublist_prev(mls
, db
);
767 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
768 multilist_sublist_t
*, mls
);
771 multilist_sublist_remove(mls
, db
);
772 multilist_sublist_unlock(mls
);
773 (void) zfs_refcount_remove_many(
774 &dbuf_caches
[DB_DBUF_CACHE
].size
, db
->db
.db_size
, db
);
775 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
776 DBUF_STAT_BUMPDOWN(cache_count
);
777 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
779 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
780 db
->db_caching_status
= DB_NO_CACHE
;
782 DBUF_STAT_BUMP(cache_total_evicts
);
784 multilist_sublist_unlock(mls
);
789 * The dbuf evict thread is responsible for aging out dbufs from the
790 * cache. Once the cache has reached it's maximum size, dbufs are removed
791 * and destroyed. The eviction thread will continue running until the size
792 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
793 * out of the cache it is destroyed and becomes eligible for arc eviction.
795 static __attribute__((noreturn
)) void
796 dbuf_evict_thread(void *unused
)
801 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
803 mutex_enter(&dbuf_evict_lock
);
804 while (!dbuf_evict_thread_exit
) {
805 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
806 CALLB_CPR_SAFE_BEGIN(&cpr
);
807 (void) cv_timedwait_idle_hires(&dbuf_evict_cv
,
808 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
809 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
811 mutex_exit(&dbuf_evict_lock
);
814 * Keep evicting as long as we're above the low water mark
815 * for the cache. We do this without holding the locks to
816 * minimize lock contention.
818 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
822 mutex_enter(&dbuf_evict_lock
);
825 dbuf_evict_thread_exit
= B_FALSE
;
826 cv_broadcast(&dbuf_evict_cv
);
827 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
832 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
833 * If the dbuf cache is at its high water mark, then evict a dbuf from the
834 * dbuf cache using the caller's context.
837 dbuf_evict_notify(uint64_t size
)
840 * We check if we should evict without holding the dbuf_evict_lock,
841 * because it's OK to occasionally make the wrong decision here,
842 * and grabbing the lock results in massive lock contention.
844 if (size
> dbuf_cache_target_bytes()) {
845 if (size
> dbuf_cache_hiwater_bytes())
847 cv_signal(&dbuf_evict_cv
);
852 dbuf_kstat_update(kstat_t
*ksp
, int rw
)
854 dbuf_stats_t
*ds
= ksp
->ks_data
;
855 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
857 if (rw
== KSTAT_WRITE
)
858 return (SET_ERROR(EACCES
));
860 ds
->cache_count
.value
.ui64
=
861 wmsum_value(&dbuf_sums
.cache_count
);
862 ds
->cache_size_bytes
.value
.ui64
=
863 zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
);
864 ds
->cache_target_bytes
.value
.ui64
= dbuf_cache_target_bytes();
865 ds
->cache_hiwater_bytes
.value
.ui64
= dbuf_cache_hiwater_bytes();
866 ds
->cache_lowater_bytes
.value
.ui64
= dbuf_cache_lowater_bytes();
867 ds
->cache_total_evicts
.value
.ui64
=
868 wmsum_value(&dbuf_sums
.cache_total_evicts
);
869 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
870 ds
->cache_levels
[i
].value
.ui64
=
871 wmsum_value(&dbuf_sums
.cache_levels
[i
]);
872 ds
->cache_levels_bytes
[i
].value
.ui64
=
873 wmsum_value(&dbuf_sums
.cache_levels_bytes
[i
]);
875 ds
->hash_hits
.value
.ui64
=
876 wmsum_value(&dbuf_sums
.hash_hits
);
877 ds
->hash_misses
.value
.ui64
=
878 wmsum_value(&dbuf_sums
.hash_misses
);
879 ds
->hash_collisions
.value
.ui64
=
880 wmsum_value(&dbuf_sums
.hash_collisions
);
881 ds
->hash_chains
.value
.ui64
=
882 wmsum_value(&dbuf_sums
.hash_chains
);
883 ds
->hash_insert_race
.value
.ui64
=
884 wmsum_value(&dbuf_sums
.hash_insert_race
);
885 ds
->hash_table_count
.value
.ui64
= h
->hash_table_mask
+ 1;
886 ds
->hash_mutex_count
.value
.ui64
= h
->hash_mutex_mask
+ 1;
887 ds
->metadata_cache_count
.value
.ui64
=
888 wmsum_value(&dbuf_sums
.metadata_cache_count
);
889 ds
->metadata_cache_size_bytes
.value
.ui64
= zfs_refcount_count(
890 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
);
891 ds
->metadata_cache_overflow
.value
.ui64
=
892 wmsum_value(&dbuf_sums
.metadata_cache_overflow
);
899 uint64_t hmsize
, hsize
= 1ULL << 16;
900 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
903 * The hash table is big enough to fill one eighth of physical memory
904 * with an average block size of zfs_arc_average_blocksize (default 8K).
905 * By default, the table will take up
906 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
908 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory() / 8)
911 h
->hash_table
= NULL
;
912 while (h
->hash_table
== NULL
) {
913 h
->hash_table_mask
= hsize
- 1;
915 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
916 if (h
->hash_table
== NULL
)
919 ASSERT3U(hsize
, >=, 1ULL << 10);
923 * The hash table buckets are protected by an array of mutexes where
924 * each mutex is reponsible for protecting 128 buckets. A minimum
925 * array size of 8192 is targeted to avoid contention.
927 if (dbuf_mutex_cache_shift
== 0)
928 hmsize
= MAX(hsize
>> 7, 1ULL << 13);
930 hmsize
= 1ULL << MIN(dbuf_mutex_cache_shift
, 24);
932 h
->hash_mutexes
= NULL
;
933 while (h
->hash_mutexes
== NULL
) {
934 h
->hash_mutex_mask
= hmsize
- 1;
936 h
->hash_mutexes
= vmem_zalloc(hmsize
* sizeof (kmutex_t
),
938 if (h
->hash_mutexes
== NULL
)
942 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
943 sizeof (dmu_buf_impl_t
),
944 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
946 for (int i
= 0; i
< hmsize
; i
++)
947 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
952 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
953 * configuration is not required.
955 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
957 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
958 multilist_create(&dbuf_caches
[dcs
].cache
,
959 sizeof (dmu_buf_impl_t
),
960 offsetof(dmu_buf_impl_t
, db_cache_link
),
961 dbuf_cache_multilist_index_func
);
962 zfs_refcount_create(&dbuf_caches
[dcs
].size
);
965 dbuf_evict_thread_exit
= B_FALSE
;
966 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
967 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
968 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
969 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
971 wmsum_init(&dbuf_sums
.cache_count
, 0);
972 wmsum_init(&dbuf_sums
.cache_total_evicts
, 0);
973 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
974 wmsum_init(&dbuf_sums
.cache_levels
[i
], 0);
975 wmsum_init(&dbuf_sums
.cache_levels_bytes
[i
], 0);
977 wmsum_init(&dbuf_sums
.hash_hits
, 0);
978 wmsum_init(&dbuf_sums
.hash_misses
, 0);
979 wmsum_init(&dbuf_sums
.hash_collisions
, 0);
980 wmsum_init(&dbuf_sums
.hash_chains
, 0);
981 wmsum_init(&dbuf_sums
.hash_insert_race
, 0);
982 wmsum_init(&dbuf_sums
.metadata_cache_count
, 0);
983 wmsum_init(&dbuf_sums
.metadata_cache_overflow
, 0);
985 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
986 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
988 if (dbuf_ksp
!= NULL
) {
989 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
990 snprintf(dbuf_stats
.cache_levels
[i
].name
,
991 KSTAT_STRLEN
, "cache_level_%d", i
);
992 dbuf_stats
.cache_levels
[i
].data_type
=
994 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
995 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
996 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
999 dbuf_ksp
->ks_data
= &dbuf_stats
;
1000 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
1001 kstat_install(dbuf_ksp
);
1008 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
1010 dbuf_stats_destroy();
1012 for (int i
= 0; i
< (h
->hash_mutex_mask
+ 1); i
++)
1013 mutex_destroy(&h
->hash_mutexes
[i
]);
1015 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
1016 vmem_free(h
->hash_mutexes
, (h
->hash_mutex_mask
+ 1) *
1019 kmem_cache_destroy(dbuf_kmem_cache
);
1020 taskq_destroy(dbu_evict_taskq
);
1022 mutex_enter(&dbuf_evict_lock
);
1023 dbuf_evict_thread_exit
= B_TRUE
;
1024 while (dbuf_evict_thread_exit
) {
1025 cv_signal(&dbuf_evict_cv
);
1026 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
1028 mutex_exit(&dbuf_evict_lock
);
1030 mutex_destroy(&dbuf_evict_lock
);
1031 cv_destroy(&dbuf_evict_cv
);
1033 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
1034 zfs_refcount_destroy(&dbuf_caches
[dcs
].size
);
1035 multilist_destroy(&dbuf_caches
[dcs
].cache
);
1038 if (dbuf_ksp
!= NULL
) {
1039 kstat_delete(dbuf_ksp
);
1043 wmsum_fini(&dbuf_sums
.cache_count
);
1044 wmsum_fini(&dbuf_sums
.cache_total_evicts
);
1045 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
1046 wmsum_fini(&dbuf_sums
.cache_levels
[i
]);
1047 wmsum_fini(&dbuf_sums
.cache_levels_bytes
[i
]);
1049 wmsum_fini(&dbuf_sums
.hash_hits
);
1050 wmsum_fini(&dbuf_sums
.hash_misses
);
1051 wmsum_fini(&dbuf_sums
.hash_collisions
);
1052 wmsum_fini(&dbuf_sums
.hash_chains
);
1053 wmsum_fini(&dbuf_sums
.hash_insert_race
);
1054 wmsum_fini(&dbuf_sums
.metadata_cache_count
);
1055 wmsum_fini(&dbuf_sums
.metadata_cache_overflow
);
1064 dbuf_verify(dmu_buf_impl_t
*db
)
1067 dbuf_dirty_record_t
*dr
;
1070 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1072 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
1075 ASSERT(db
->db_objset
!= NULL
);
1079 ASSERT(db
->db_parent
== NULL
);
1080 ASSERT(db
->db_blkptr
== NULL
);
1082 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
1083 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
1084 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
1085 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
1086 db
->db_blkid
== DMU_SPILL_BLKID
||
1087 !avl_is_empty(&dn
->dn_dbufs
));
1089 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1091 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1092 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
1093 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
1095 ASSERT0(db
->db
.db_offset
);
1097 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
1100 if ((dr
= list_head(&db
->db_dirty_records
)) != NULL
) {
1101 ASSERT(dr
->dr_dbuf
== db
);
1102 txg_prev
= dr
->dr_txg
;
1103 for (dr
= list_next(&db
->db_dirty_records
, dr
); dr
!= NULL
;
1104 dr
= list_next(&db
->db_dirty_records
, dr
)) {
1105 ASSERT(dr
->dr_dbuf
== db
);
1106 ASSERT(txg_prev
> dr
->dr_txg
);
1107 txg_prev
= dr
->dr_txg
;
1112 * We can't assert that db_size matches dn_datablksz because it
1113 * can be momentarily different when another thread is doing
1114 * dnode_set_blksz().
1116 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
1117 dr
= db
->db_data_pending
;
1119 * It should only be modified in syncing context, so
1120 * make sure we only have one copy of the data.
1122 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
1125 /* verify db->db_blkptr */
1126 if (db
->db_blkptr
) {
1127 if (db
->db_parent
== dn
->dn_dbuf
) {
1128 /* db is pointed to by the dnode */
1129 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1130 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
1131 ASSERT(db
->db_parent
== NULL
);
1133 ASSERT(db
->db_parent
!= NULL
);
1134 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
1135 ASSERT3P(db
->db_blkptr
, ==,
1136 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
1138 /* db is pointed to by an indirect block */
1139 int epb __maybe_unused
= db
->db_parent
->db
.db_size
>>
1141 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
1142 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
1145 * dnode_grow_indblksz() can make this fail if we don't
1146 * have the parent's rwlock. XXX indblksz no longer
1147 * grows. safe to do this now?
1149 if (RW_LOCK_HELD(&db
->db_parent
->db_rwlock
)) {
1150 ASSERT3P(db
->db_blkptr
, ==,
1151 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
1152 db
->db_blkid
% epb
));
1156 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
1157 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
1158 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
1159 db
->db_state
!= DB_FILL
&& (dn
== NULL
|| !dn
->dn_free_txg
)) {
1161 * If the blkptr isn't set but they have nonzero data,
1162 * it had better be dirty, otherwise we'll lose that
1163 * data when we evict this buffer.
1165 * There is an exception to this rule for indirect blocks; in
1166 * this case, if the indirect block is a hole, we fill in a few
1167 * fields on each of the child blocks (importantly, birth time)
1168 * to prevent hole birth times from being lost when you
1169 * partially fill in a hole.
1171 if (db
->db_dirtycnt
== 0) {
1172 if (db
->db_level
== 0) {
1173 uint64_t *buf
= db
->db
.db_data
;
1176 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
1177 ASSERT(buf
[i
] == 0);
1180 blkptr_t
*bps
= db
->db
.db_data
;
1181 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1184 * We want to verify that all the blkptrs in the
1185 * indirect block are holes, but we may have
1186 * automatically set up a few fields for them.
1187 * We iterate through each blkptr and verify
1188 * they only have those fields set.
1191 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1193 blkptr_t
*bp
= &bps
[i
];
1194 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1197 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1198 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1199 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1200 ASSERT0(bp
->blk_fill
);
1201 ASSERT0(bp
->blk_pad
[0]);
1202 ASSERT0(bp
->blk_pad
[1]);
1203 ASSERT(!BP_IS_EMBEDDED(bp
));
1204 ASSERT(BP_IS_HOLE(bp
));
1205 ASSERT0(bp
->blk_phys_birth
);
1215 dbuf_clear_data(dmu_buf_impl_t
*db
)
1217 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1218 dbuf_evict_user(db
);
1219 ASSERT3P(db
->db_buf
, ==, NULL
);
1220 db
->db
.db_data
= NULL
;
1221 if (db
->db_state
!= DB_NOFILL
) {
1222 db
->db_state
= DB_UNCACHED
;
1223 DTRACE_SET_STATE(db
, "clear data");
1228 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1230 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1231 ASSERT(buf
!= NULL
);
1234 ASSERT(buf
->b_data
!= NULL
);
1235 db
->db
.db_data
= buf
->b_data
;
1239 dbuf_alloc_arcbuf(dmu_buf_impl_t
*db
)
1241 spa_t
*spa
= db
->db_objset
->os_spa
;
1243 return (arc_alloc_buf(spa
, db
, DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
1247 * Loan out an arc_buf for read. Return the loaned arc_buf.
1250 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1254 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1255 mutex_enter(&db
->db_mtx
);
1256 if (arc_released(db
->db_buf
) || zfs_refcount_count(&db
->db_holds
) > 1) {
1257 int blksz
= db
->db
.db_size
;
1258 spa_t
*spa
= db
->db_objset
->os_spa
;
1260 mutex_exit(&db
->db_mtx
);
1261 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1262 memcpy(abuf
->b_data
, db
->db
.db_data
, blksz
);
1265 arc_loan_inuse_buf(abuf
, db
);
1267 dbuf_clear_data(db
);
1268 mutex_exit(&db
->db_mtx
);
1274 * Calculate which level n block references the data at the level 0 offset
1278 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1280 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1282 * The level n blkid is equal to the level 0 blkid divided by
1283 * the number of level 0s in a level n block.
1285 * The level 0 blkid is offset >> datablkshift =
1286 * offset / 2^datablkshift.
1288 * The number of level 0s in a level n is the number of block
1289 * pointers in an indirect block, raised to the power of level.
1290 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1291 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1293 * Thus, the level n blkid is: offset /
1294 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1295 * = offset / 2^(datablkshift + level *
1296 * (indblkshift - SPA_BLKPTRSHIFT))
1297 * = offset >> (datablkshift + level *
1298 * (indblkshift - SPA_BLKPTRSHIFT))
1301 const unsigned exp
= dn
->dn_datablkshift
+
1302 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1304 if (exp
>= 8 * sizeof (offset
)) {
1305 /* This only happens on the highest indirection level */
1306 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1310 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1312 return (offset
>> exp
);
1314 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1320 * This function is used to lock the parent of the provided dbuf. This should be
1321 * used when modifying or reading db_blkptr.
1324 dmu_buf_lock_parent(dmu_buf_impl_t
*db
, krw_t rw
, const void *tag
)
1326 enum db_lock_type ret
= DLT_NONE
;
1327 if (db
->db_parent
!= NULL
) {
1328 rw_enter(&db
->db_parent
->db_rwlock
, rw
);
1330 } else if (dmu_objset_ds(db
->db_objset
) != NULL
) {
1331 rrw_enter(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, rw
,
1336 * We only return a DLT_NONE lock when it's the top-most indirect block
1337 * of the meta-dnode of the MOS.
1343 * We need to pass the lock type in because it's possible that the block will
1344 * move from being the topmost indirect block in a dnode (and thus, have no
1345 * parent) to not the top-most via an indirection increase. This would cause a
1346 * panic if we didn't pass the lock type in.
1349 dmu_buf_unlock_parent(dmu_buf_impl_t
*db
, db_lock_type_t type
, const void *tag
)
1351 if (type
== DLT_PARENT
)
1352 rw_exit(&db
->db_parent
->db_rwlock
);
1353 else if (type
== DLT_OBJSET
)
1354 rrw_exit(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, tag
);
1358 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1359 arc_buf_t
*buf
, void *vdb
)
1361 (void) zb
, (void) bp
;
1362 dmu_buf_impl_t
*db
= vdb
;
1364 mutex_enter(&db
->db_mtx
);
1365 ASSERT3U(db
->db_state
, ==, DB_READ
);
1367 * All reads are synchronous, so we must have a hold on the dbuf
1369 ASSERT(zfs_refcount_count(&db
->db_holds
) > 0);
1370 ASSERT(db
->db_buf
== NULL
);
1371 ASSERT(db
->db
.db_data
== NULL
);
1374 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
1375 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1376 ASSERT3P(db
->db_buf
, ==, NULL
);
1377 db
->db_state
= DB_UNCACHED
;
1378 DTRACE_SET_STATE(db
, "i/o error");
1379 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1380 /* freed in flight */
1381 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1382 arc_release(buf
, db
);
1383 memset(buf
->b_data
, 0, db
->db
.db_size
);
1384 arc_buf_freeze(buf
);
1385 db
->db_freed_in_flight
= FALSE
;
1386 dbuf_set_data(db
, buf
);
1387 db
->db_state
= DB_CACHED
;
1388 DTRACE_SET_STATE(db
, "freed in flight");
1391 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1392 dbuf_set_data(db
, buf
);
1393 db
->db_state
= DB_CACHED
;
1394 DTRACE_SET_STATE(db
, "successful read");
1396 cv_broadcast(&db
->db_changed
);
1397 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
1401 * Shortcut for performing reads on bonus dbufs. Returns
1402 * an error if we fail to verify the dnode associated with
1403 * a decrypted block. Otherwise success.
1406 dbuf_read_bonus(dmu_buf_impl_t
*db
, dnode_t
*dn
, uint32_t flags
)
1408 int bonuslen
, max_bonuslen
, err
;
1410 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1414 bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1415 max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1416 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1417 ASSERT(DB_DNODE_HELD(db
));
1418 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1419 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1420 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1421 if (bonuslen
< max_bonuslen
)
1422 memset(db
->db
.db_data
, 0, max_bonuslen
);
1424 memcpy(db
->db
.db_data
, DN_BONUS(dn
->dn_phys
), bonuslen
);
1425 db
->db_state
= DB_CACHED
;
1426 DTRACE_SET_STATE(db
, "bonus buffer filled");
1431 dbuf_handle_indirect_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
, blkptr_t
*dbbp
)
1433 blkptr_t
*bps
= db
->db
.db_data
;
1434 uint32_t indbs
= 1ULL << dn
->dn_indblkshift
;
1435 int n_bps
= indbs
>> SPA_BLKPTRSHIFT
;
1437 for (int i
= 0; i
< n_bps
; i
++) {
1438 blkptr_t
*bp
= &bps
[i
];
1440 ASSERT3U(BP_GET_LSIZE(dbbp
), ==, indbs
);
1441 BP_SET_LSIZE(bp
, BP_GET_LEVEL(dbbp
) == 1 ?
1442 dn
->dn_datablksz
: BP_GET_LSIZE(dbbp
));
1443 BP_SET_TYPE(bp
, BP_GET_TYPE(dbbp
));
1444 BP_SET_LEVEL(bp
, BP_GET_LEVEL(dbbp
) - 1);
1445 BP_SET_BIRTH(bp
, dbbp
->blk_birth
, 0);
1450 * Handle reads on dbufs that are holes, if necessary. This function
1451 * requires that the dbuf's mutex is held. Returns success (0) if action
1452 * was taken, ENOENT if no action was taken.
1455 dbuf_read_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
, blkptr_t
*bp
)
1457 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1459 int is_hole
= bp
== NULL
|| BP_IS_HOLE(bp
);
1461 * For level 0 blocks only, if the above check fails:
1462 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1463 * processes the delete record and clears the bp while we are waiting
1464 * for the dn_mtx (resulting in a "no" from block_freed).
1466 if (!is_hole
&& db
->db_level
== 0)
1467 is_hole
= dnode_block_freed(dn
, db
->db_blkid
) || BP_IS_HOLE(bp
);
1470 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1471 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
1473 if (bp
!= NULL
&& db
->db_level
> 0 && BP_IS_HOLE(bp
) &&
1474 bp
->blk_birth
!= 0) {
1475 dbuf_handle_indirect_hole(db
, dn
, bp
);
1477 db
->db_state
= DB_CACHED
;
1478 DTRACE_SET_STATE(db
, "hole read satisfied");
1485 * This function ensures that, when doing a decrypting read of a block,
1486 * we make sure we have decrypted the dnode associated with it. We must do
1487 * this so that we ensure we are fully authenticating the checksum-of-MACs
1488 * tree from the root of the objset down to this block. Indirect blocks are
1489 * always verified against their secure checksum-of-MACs assuming that the
1490 * dnode containing them is correct. Now that we are doing a decrypting read,
1491 * we can be sure that the key is loaded and verify that assumption. This is
1492 * especially important considering that we always read encrypted dnode
1493 * blocks as raw data (without verifying their MACs) to start, and
1494 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1497 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
)
1500 objset_t
*os
= db
->db_objset
;
1501 arc_buf_t
*dnode_abuf
;
1503 zbookmark_phys_t zb
;
1505 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1507 if ((flags
& DB_RF_NO_DECRYPT
) != 0 ||
1508 !os
->os_encrypted
|| os
->os_raw_receive
)
1513 dnode_abuf
= (dn
->dn_dbuf
!= NULL
) ? dn
->dn_dbuf
->db_buf
: NULL
;
1515 if (dnode_abuf
== NULL
|| !arc_is_encrypted(dnode_abuf
)) {
1520 SET_BOOKMARK(&zb
, dmu_objset_id(os
),
1521 DMU_META_DNODE_OBJECT
, 0, dn
->dn_dbuf
->db_blkid
);
1522 err
= arc_untransform(dnode_abuf
, os
->os_spa
, &zb
, B_TRUE
);
1525 * An error code of EACCES tells us that the key is still not
1526 * available. This is ok if we are only reading authenticated
1527 * (and therefore non-encrypted) blocks.
1529 if (err
== EACCES
&& ((db
->db_blkid
!= DMU_BONUS_BLKID
&&
1530 !DMU_OT_IS_ENCRYPTED(dn
->dn_type
)) ||
1531 (db
->db_blkid
== DMU_BONUS_BLKID
&&
1532 !DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))))
1541 * Drops db_mtx and the parent lock specified by dblt and tag before
1545 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
,
1546 db_lock_type_t dblt
, const void *tag
)
1549 zbookmark_phys_t zb
;
1550 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1556 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1557 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1558 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1559 ASSERT(db
->db_buf
== NULL
);
1560 ASSERT(db
->db_parent
== NULL
||
1561 RW_LOCK_HELD(&db
->db_parent
->db_rwlock
));
1563 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1564 err
= dbuf_read_bonus(db
, dn
, flags
);
1568 if (db
->db_state
== DB_UNCACHED
) {
1569 if (db
->db_blkptr
== NULL
) {
1572 bp
= *db
->db_blkptr
;
1576 dbuf_dirty_record_t
*dr
;
1578 ASSERT3S(db
->db_state
, ==, DB_NOFILL
);
1581 * Block cloning: If we have a pending block clone,
1582 * we don't want to read the underlying block, but the content
1583 * of the block being cloned, so we have the most recent data.
1585 dr
= list_head(&db
->db_dirty_records
);
1586 if (dr
== NULL
|| !dr
->dt
.dl
.dr_brtwrite
) {
1590 bp
= dr
->dt
.dl
.dr_overridden_by
;
1594 err
= dbuf_read_hole(db
, dn
, bpp
);
1598 ASSERT(bpp
!= NULL
);
1601 * Any attempt to read a redacted block should result in an error. This
1602 * will never happen under normal conditions, but can be useful for
1603 * debugging purposes.
1605 if (BP_IS_REDACTED(bpp
)) {
1606 ASSERT(dsl_dataset_feature_is_active(
1607 db
->db_objset
->os_dsl_dataset
,
1608 SPA_FEATURE_REDACTED_DATASETS
));
1609 err
= SET_ERROR(EIO
);
1613 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1614 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1617 * All bps of an encrypted os should have the encryption bit set.
1618 * If this is not true it indicates tampering and we report an error.
1620 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(bpp
)) {
1621 spa_log_error(db
->db_objset
->os_spa
, &zb
, &bpp
->blk_birth
);
1622 err
= SET_ERROR(EIO
);
1626 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1632 db
->db_state
= DB_READ
;
1633 DTRACE_SET_STATE(db
, "read issued");
1634 mutex_exit(&db
->db_mtx
);
1636 if (!DBUF_IS_CACHEABLE(db
))
1637 aflags
|= ARC_FLAG_UNCACHED
;
1638 else if (dbuf_is_l2cacheable(db
))
1639 aflags
|= ARC_FLAG_L2CACHE
;
1641 dbuf_add_ref(db
, NULL
);
1643 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1644 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1646 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1647 zio_flags
|= ZIO_FLAG_RAW
;
1649 * The zio layer will copy the provided blkptr later, but we have our
1650 * own copy so that we can release the parent's rwlock. We have to
1651 * do that so that if dbuf_read_done is called synchronously (on
1652 * an l1 cache hit) we don't acquire the db_mtx while holding the
1653 * parent's rwlock, which would be a lock ordering violation.
1655 dmu_buf_unlock_parent(db
, dblt
, tag
);
1656 (void) arc_read(zio
, db
->db_objset
->os_spa
, bpp
,
1657 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1662 mutex_exit(&db
->db_mtx
);
1663 dmu_buf_unlock_parent(db
, dblt
, tag
);
1668 * This is our just-in-time copy function. It makes a copy of buffers that
1669 * have been modified in a previous transaction group before we access them in
1670 * the current active group.
1672 * This function is used in three places: when we are dirtying a buffer for the
1673 * first time in a txg, when we are freeing a range in a dnode that includes
1674 * this buffer, and when we are accessing a buffer which was received compressed
1675 * and later referenced in a WRITE_BYREF record.
1677 * Note that when we are called from dbuf_free_range() we do not put a hold on
1678 * the buffer, we just traverse the active dbuf list for the dnode.
1681 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1683 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
1685 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1686 ASSERT(db
->db
.db_data
!= NULL
);
1687 ASSERT(db
->db_level
== 0);
1688 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1691 (dr
->dt
.dl
.dr_data
!=
1692 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1696 * If the last dirty record for this dbuf has not yet synced
1697 * and its referencing the dbuf data, either:
1698 * reset the reference to point to a new copy,
1699 * or (if there a no active holders)
1700 * just null out the current db_data pointer.
1702 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1703 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1704 dnode_t
*dn
= DB_DNODE(db
);
1705 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1706 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1707 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1708 memcpy(dr
->dt
.dl
.dr_data
, db
->db
.db_data
, bonuslen
);
1709 } else if (zfs_refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1710 dnode_t
*dn
= DB_DNODE(db
);
1711 int size
= arc_buf_size(db
->db_buf
);
1712 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1713 spa_t
*spa
= db
->db_objset
->os_spa
;
1714 enum zio_compress compress_type
=
1715 arc_get_compression(db
->db_buf
);
1716 uint8_t complevel
= arc_get_complevel(db
->db_buf
);
1718 if (arc_is_encrypted(db
->db_buf
)) {
1719 boolean_t byteorder
;
1720 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1721 uint8_t iv
[ZIO_DATA_IV_LEN
];
1722 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1724 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1726 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1727 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1728 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1729 compress_type
, complevel
);
1730 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1731 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1732 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1733 size
, arc_buf_lsize(db
->db_buf
), compress_type
,
1736 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1738 memcpy(dr
->dt
.dl
.dr_data
->b_data
, db
->db
.db_data
, size
);
1741 dbuf_clear_data(db
);
1746 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1753 * We don't have to hold the mutex to check db_state because it
1754 * can't be freed while we have a hold on the buffer.
1756 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1761 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1762 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
;
1764 mutex_enter(&db
->db_mtx
);
1765 if (flags
& DB_RF_PARTIAL_FIRST
)
1766 db
->db_partial_read
= B_TRUE
;
1767 else if (!(flags
& DB_RF_PARTIAL_MORE
))
1768 db
->db_partial_read
= B_FALSE
;
1769 if (db
->db_state
== DB_CACHED
) {
1771 * Ensure that this block's dnode has been decrypted if
1772 * the caller has requested decrypted data.
1774 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1777 * If the arc buf is compressed or encrypted and the caller
1778 * requested uncompressed data, we need to untransform it
1779 * before returning. We also call arc_untransform() on any
1780 * unauthenticated blocks, which will verify their MAC if
1781 * the key is now available.
1783 if (err
== 0 && db
->db_buf
!= NULL
&&
1784 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1785 (arc_is_encrypted(db
->db_buf
) ||
1786 arc_is_unauthenticated(db
->db_buf
) ||
1787 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1788 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1789 zbookmark_phys_t zb
;
1791 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1792 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1793 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1794 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1795 dbuf_set_data(db
, db
->db_buf
);
1797 mutex_exit(&db
->db_mtx
);
1798 if (err
== 0 && prefetch
) {
1799 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1800 B_FALSE
, flags
& DB_RF_HAVESTRUCT
);
1803 DBUF_STAT_BUMP(hash_hits
);
1804 } else if (db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
) {
1805 boolean_t need_wait
= B_FALSE
;
1807 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
1809 if (zio
== NULL
&& (db
->db_state
== DB_NOFILL
||
1810 (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)))) {
1811 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1812 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1815 err
= dbuf_read_impl(db
, zio
, flags
, dblt
, FTAG
);
1817 * dbuf_read_impl has dropped db_mtx and our parent's rwlock
1820 if (!err
&& prefetch
) {
1821 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1822 db
->db_state
!= DB_CACHED
,
1823 flags
& DB_RF_HAVESTRUCT
);
1827 DBUF_STAT_BUMP(hash_misses
);
1830 * If we created a zio_root we must execute it to avoid
1831 * leaking it, even if it isn't attached to any work due
1832 * to an error in dbuf_read_impl().
1836 err
= zio_wait(zio
);
1838 VERIFY0(zio_wait(zio
));
1842 * Another reader came in while the dbuf was in flight
1843 * between UNCACHED and CACHED. Either a writer will finish
1844 * writing the buffer (sending the dbuf to CACHED) or the
1845 * first reader's request will reach the read_done callback
1846 * and send the dbuf to CACHED. Otherwise, a failure
1847 * occurred and the dbuf went to UNCACHED.
1849 mutex_exit(&db
->db_mtx
);
1851 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1852 B_TRUE
, flags
& DB_RF_HAVESTRUCT
);
1855 DBUF_STAT_BUMP(hash_misses
);
1857 /* Skip the wait per the caller's request. */
1858 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1859 mutex_enter(&db
->db_mtx
);
1860 while (db
->db_state
== DB_READ
||
1861 db
->db_state
== DB_FILL
) {
1862 ASSERT(db
->db_state
== DB_READ
||
1863 (flags
& DB_RF_HAVESTRUCT
) == 0);
1864 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1866 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1868 if (db
->db_state
== DB_UNCACHED
)
1869 err
= SET_ERROR(EIO
);
1870 mutex_exit(&db
->db_mtx
);
1878 dbuf_noread(dmu_buf_impl_t
*db
)
1880 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1881 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1882 mutex_enter(&db
->db_mtx
);
1883 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1884 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1885 if (db
->db_state
== DB_UNCACHED
) {
1886 ASSERT(db
->db_buf
== NULL
);
1887 ASSERT(db
->db
.db_data
== NULL
);
1888 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1889 db
->db_state
= DB_FILL
;
1890 DTRACE_SET_STATE(db
, "assigning filled buffer");
1891 } else if (db
->db_state
== DB_NOFILL
) {
1892 dbuf_clear_data(db
);
1894 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1896 mutex_exit(&db
->db_mtx
);
1900 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1902 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1903 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1904 uint64_t txg
= dr
->dr_txg
;
1906 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1908 * This assert is valid because dmu_sync() expects to be called by
1909 * a zilog's get_data while holding a range lock. This call only
1910 * comes from dbuf_dirty() callers who must also hold a range lock.
1912 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1913 ASSERT(db
->db_level
== 0);
1915 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1916 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1919 ASSERT(db
->db_data_pending
!= dr
);
1921 /* free this block */
1922 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1923 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1925 if (dr
->dt
.dl
.dr_brtwrite
) {
1926 ASSERT0(dr
->dt
.dl
.dr_data
);
1927 dr
->dt
.dl
.dr_data
= db
->db_buf
;
1929 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1930 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1931 dr
->dt
.dl
.dr_brtwrite
= B_FALSE
;
1932 dr
->dt
.dl
.dr_has_raw_params
= B_FALSE
;
1935 * Release the already-written buffer, so we leave it in
1936 * a consistent dirty state. Note that all callers are
1937 * modifying the buffer, so they will immediately do
1938 * another (redundant) arc_release(). Therefore, leave
1939 * the buf thawed to save the effort of freezing &
1940 * immediately re-thawing it.
1942 if (dr
->dt
.dl
.dr_data
)
1943 arc_release(dr
->dt
.dl
.dr_data
, db
);
1947 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1948 * data blocks in the free range, so that any future readers will find
1952 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1955 dmu_buf_impl_t
*db_search
;
1956 dmu_buf_impl_t
*db
, *db_next
;
1957 uint64_t txg
= tx
->tx_txg
;
1959 dbuf_dirty_record_t
*dr
;
1961 if (end_blkid
> dn
->dn_maxblkid
&&
1962 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1963 end_blkid
= dn
->dn_maxblkid
;
1964 dprintf_dnode(dn
, "start=%llu end=%llu\n", (u_longlong_t
)start_blkid
,
1965 (u_longlong_t
)end_blkid
);
1967 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1968 db_search
->db_level
= 0;
1969 db_search
->db_blkid
= start_blkid
;
1970 db_search
->db_state
= DB_SEARCH
;
1972 mutex_enter(&dn
->dn_dbufs_mtx
);
1973 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1974 ASSERT3P(db
, ==, NULL
);
1976 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1978 for (; db
!= NULL
; db
= db_next
) {
1979 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1980 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1982 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1985 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1987 /* found a level 0 buffer in the range */
1988 mutex_enter(&db
->db_mtx
);
1989 if (dbuf_undirty(db
, tx
)) {
1990 /* mutex has been dropped and dbuf destroyed */
1994 if (db
->db_state
== DB_UNCACHED
||
1995 db
->db_state
== DB_NOFILL
||
1996 db
->db_state
== DB_EVICTING
) {
1997 ASSERT(db
->db
.db_data
== NULL
);
1998 mutex_exit(&db
->db_mtx
);
2001 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
2002 /* will be handled in dbuf_read_done or dbuf_rele */
2003 db
->db_freed_in_flight
= TRUE
;
2004 mutex_exit(&db
->db_mtx
);
2007 if (zfs_refcount_count(&db
->db_holds
) == 0) {
2012 /* The dbuf is referenced */
2014 dr
= list_head(&db
->db_dirty_records
);
2016 if (dr
->dr_txg
== txg
) {
2018 * This buffer is "in-use", re-adjust the file
2019 * size to reflect that this buffer may
2020 * contain new data when we sync.
2022 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
2023 db
->db_blkid
> dn
->dn_maxblkid
)
2024 dn
->dn_maxblkid
= db
->db_blkid
;
2025 dbuf_unoverride(dr
);
2028 * This dbuf is not dirty in the open context.
2029 * Either uncache it (if its not referenced in
2030 * the open context) or reset its contents to
2033 dbuf_fix_old_data(db
, txg
);
2036 /* clear the contents if its cached */
2037 if (db
->db_state
== DB_CACHED
) {
2038 ASSERT(db
->db
.db_data
!= NULL
);
2039 arc_release(db
->db_buf
, db
);
2040 rw_enter(&db
->db_rwlock
, RW_WRITER
);
2041 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2042 rw_exit(&db
->db_rwlock
);
2043 arc_buf_freeze(db
->db_buf
);
2046 mutex_exit(&db
->db_mtx
);
2049 mutex_exit(&dn
->dn_dbufs_mtx
);
2050 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
2054 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
2056 arc_buf_t
*buf
, *old_buf
;
2057 dbuf_dirty_record_t
*dr
;
2058 int osize
= db
->db
.db_size
;
2059 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2062 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2068 * XXX we should be doing a dbuf_read, checking the return
2069 * value and returning that up to our callers
2071 dmu_buf_will_dirty(&db
->db
, tx
);
2073 /* create the data buffer for the new block */
2074 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
2076 /* copy old block data to the new block */
2077 old_buf
= db
->db_buf
;
2078 memcpy(buf
->b_data
, old_buf
->b_data
, MIN(osize
, size
));
2079 /* zero the remainder */
2081 memset((uint8_t *)buf
->b_data
+ osize
, 0, size
- osize
);
2083 mutex_enter(&db
->db_mtx
);
2084 dbuf_set_data(db
, buf
);
2085 arc_buf_destroy(old_buf
, db
);
2086 db
->db
.db_size
= size
;
2088 dr
= list_head(&db
->db_dirty_records
);
2089 /* dirty record added by dmu_buf_will_dirty() */
2091 if (db
->db_level
== 0)
2092 dr
->dt
.dl
.dr_data
= buf
;
2093 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2094 ASSERT3U(dr
->dr_accounted
, ==, osize
);
2095 dr
->dr_accounted
= size
;
2096 mutex_exit(&db
->db_mtx
);
2098 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
2103 dbuf_release_bp(dmu_buf_impl_t
*db
)
2105 objset_t
*os __maybe_unused
= db
->db_objset
;
2107 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
2108 ASSERT(arc_released(os
->os_phys_buf
) ||
2109 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
2110 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
2112 (void) arc_release(db
->db_buf
, db
);
2116 * We already have a dirty record for this TXG, and we are being
2120 dbuf_redirty(dbuf_dirty_record_t
*dr
)
2122 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2124 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2126 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
2128 * If this buffer has already been written out,
2129 * we now need to reset its state.
2131 dbuf_unoverride(dr
);
2132 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
2133 db
->db_state
!= DB_NOFILL
) {
2134 /* Already released on initial dirty, so just thaw. */
2135 ASSERT(arc_released(db
->db_buf
));
2136 arc_buf_thaw(db
->db_buf
);
2141 dbuf_dirty_record_t
*
2142 dbuf_dirty_lightweight(dnode_t
*dn
, uint64_t blkid
, dmu_tx_t
*tx
)
2144 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2145 IMPLY(dn
->dn_objset
->os_raw_receive
, dn
->dn_maxblkid
>= blkid
);
2146 dnode_new_blkid(dn
, blkid
, tx
, B_TRUE
, B_FALSE
);
2147 ASSERT(dn
->dn_maxblkid
>= blkid
);
2149 dbuf_dirty_record_t
*dr
= kmem_zalloc(sizeof (*dr
), KM_SLEEP
);
2150 list_link_init(&dr
->dr_dirty_node
);
2151 list_link_init(&dr
->dr_dbuf_node
);
2153 dr
->dr_txg
= tx
->tx_txg
;
2154 dr
->dt
.dll
.dr_blkid
= blkid
;
2155 dr
->dr_accounted
= dn
->dn_datablksz
;
2158 * There should not be any dbuf for the block that we're dirtying.
2159 * Otherwise the buffer contents could be inconsistent between the
2160 * dbuf and the lightweight dirty record.
2162 ASSERT3P(NULL
, ==, dbuf_find(dn
->dn_objset
, dn
->dn_object
, 0, blkid
,
2165 mutex_enter(&dn
->dn_mtx
);
2166 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2167 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2168 range_tree_clear(dn
->dn_free_ranges
[txgoff
], blkid
, 1);
2171 if (dn
->dn_nlevels
== 1) {
2172 ASSERT3U(blkid
, <, dn
->dn_nblkptr
);
2173 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2174 mutex_exit(&dn
->dn_mtx
);
2175 rw_exit(&dn
->dn_struct_rwlock
);
2176 dnode_setdirty(dn
, tx
);
2178 mutex_exit(&dn
->dn_mtx
);
2180 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2181 dmu_buf_impl_t
*parent_db
= dbuf_hold_level(dn
,
2182 1, blkid
>> epbs
, FTAG
);
2183 rw_exit(&dn
->dn_struct_rwlock
);
2184 if (parent_db
== NULL
) {
2185 kmem_free(dr
, sizeof (*dr
));
2188 int err
= dbuf_read(parent_db
, NULL
,
2189 (DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2191 dbuf_rele(parent_db
, FTAG
);
2192 kmem_free(dr
, sizeof (*dr
));
2196 dbuf_dirty_record_t
*parent_dr
= dbuf_dirty(parent_db
, tx
);
2197 dbuf_rele(parent_db
, FTAG
);
2198 mutex_enter(&parent_dr
->dt
.di
.dr_mtx
);
2199 ASSERT3U(parent_dr
->dr_txg
, ==, tx
->tx_txg
);
2200 list_insert_tail(&parent_dr
->dt
.di
.dr_children
, dr
);
2201 mutex_exit(&parent_dr
->dt
.di
.dr_mtx
);
2202 dr
->dr_parent
= parent_dr
;
2205 dmu_objset_willuse_space(dn
->dn_objset
, dr
->dr_accounted
, tx
);
2210 dbuf_dirty_record_t
*
2211 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2215 dbuf_dirty_record_t
*dr
, *dr_next
, *dr_head
;
2216 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2217 boolean_t drop_struct_rwlock
= B_FALSE
;
2219 ASSERT(tx
->tx_txg
!= 0);
2220 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2221 DMU_TX_DIRTY_BUF(tx
, db
);
2226 * Shouldn't dirty a regular buffer in syncing context. Private
2227 * objects may be dirtied in syncing context, but only if they
2228 * were already pre-dirtied in open context.
2231 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
2232 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
2235 ASSERT(!dmu_tx_is_syncing(tx
) ||
2236 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
2237 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2238 dn
->dn_objset
->os_dsl_dataset
== NULL
);
2239 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2240 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2243 * We make this assert for private objects as well, but after we
2244 * check if we're already dirty. They are allowed to re-dirty
2245 * in syncing context.
2247 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2248 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2249 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2251 mutex_enter(&db
->db_mtx
);
2253 * XXX make this true for indirects too? The problem is that
2254 * transactions created with dmu_tx_create_assigned() from
2255 * syncing context don't bother holding ahead.
2257 ASSERT(db
->db_level
!= 0 ||
2258 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
2259 db
->db_state
== DB_NOFILL
);
2261 mutex_enter(&dn
->dn_mtx
);
2262 dnode_set_dirtyctx(dn
, tx
, db
);
2263 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
2264 dn
->dn_dirty_txg
= tx
->tx_txg
;
2265 mutex_exit(&dn
->dn_mtx
);
2267 if (db
->db_blkid
== DMU_SPILL_BLKID
)
2268 dn
->dn_have_spill
= B_TRUE
;
2271 * If this buffer is already dirty, we're done.
2273 dr_head
= list_head(&db
->db_dirty_records
);
2274 ASSERT(dr_head
== NULL
|| dr_head
->dr_txg
<= tx
->tx_txg
||
2275 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
2276 dr_next
= dbuf_find_dirty_lte(db
, tx
->tx_txg
);
2277 if (dr_next
&& dr_next
->dr_txg
== tx
->tx_txg
) {
2280 dbuf_redirty(dr_next
);
2281 mutex_exit(&db
->db_mtx
);
2286 * Only valid if not already dirty.
2288 ASSERT(dn
->dn_object
== 0 ||
2289 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2290 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2292 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
2295 * We should only be dirtying in syncing context if it's the
2296 * mos or we're initializing the os or it's a special object.
2297 * However, we are allowed to dirty in syncing context provided
2298 * we already dirtied it in open context. Hence we must make
2299 * this assertion only if we're not already dirty.
2302 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
2304 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2305 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
2306 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2307 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
2308 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2309 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2311 ASSERT(db
->db
.db_size
!= 0);
2313 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2315 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& db
->db_state
!= DB_NOFILL
) {
2316 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
2320 * If this buffer is dirty in an old transaction group we need
2321 * to make a copy of it so that the changes we make in this
2322 * transaction group won't leak out when we sync the older txg.
2324 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
2325 list_link_init(&dr
->dr_dirty_node
);
2326 list_link_init(&dr
->dr_dbuf_node
);
2328 if (db
->db_level
== 0) {
2329 void *data_old
= db
->db_buf
;
2331 if (db
->db_state
!= DB_NOFILL
) {
2332 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2333 dbuf_fix_old_data(db
, tx
->tx_txg
);
2334 data_old
= db
->db
.db_data
;
2335 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
2337 * Release the data buffer from the cache so
2338 * that we can modify it without impacting
2339 * possible other users of this cached data
2340 * block. Note that indirect blocks and
2341 * private objects are not released until the
2342 * syncing state (since they are only modified
2345 arc_release(db
->db_buf
, db
);
2346 dbuf_fix_old_data(db
, tx
->tx_txg
);
2347 data_old
= db
->db_buf
;
2349 ASSERT(data_old
!= NULL
);
2351 dr
->dt
.dl
.dr_data
= data_old
;
2353 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
2354 list_create(&dr
->dt
.di
.dr_children
,
2355 sizeof (dbuf_dirty_record_t
),
2356 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
2358 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& db
->db_state
!= DB_NOFILL
) {
2359 dr
->dr_accounted
= db
->db
.db_size
;
2362 dr
->dr_txg
= tx
->tx_txg
;
2363 list_insert_before(&db
->db_dirty_records
, dr_next
, dr
);
2366 * We could have been freed_in_flight between the dbuf_noread
2367 * and dbuf_dirty. We win, as though the dbuf_noread() had
2368 * happened after the free.
2370 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2371 db
->db_blkid
!= DMU_SPILL_BLKID
) {
2372 mutex_enter(&dn
->dn_mtx
);
2373 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2374 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
2377 mutex_exit(&dn
->dn_mtx
);
2378 db
->db_freed_in_flight
= FALSE
;
2382 * This buffer is now part of this txg
2384 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
2385 db
->db_dirtycnt
+= 1;
2386 ASSERT3U(db
->db_dirtycnt
, <=, 3);
2388 mutex_exit(&db
->db_mtx
);
2390 if (db
->db_blkid
== DMU_BONUS_BLKID
||
2391 db
->db_blkid
== DMU_SPILL_BLKID
) {
2392 mutex_enter(&dn
->dn_mtx
);
2393 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2394 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2395 mutex_exit(&dn
->dn_mtx
);
2396 dnode_setdirty(dn
, tx
);
2401 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
2402 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2403 drop_struct_rwlock
= B_TRUE
;
2407 * If we are overwriting a dedup BP, then unless it is snapshotted,
2408 * when we get to syncing context we will need to decrement its
2409 * refcount in the DDT. Prefetch the relevant DDT block so that
2410 * syncing context won't have to wait for the i/o.
2412 if (db
->db_blkptr
!= NULL
) {
2413 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
2414 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
2415 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
2419 * We need to hold the dn_struct_rwlock to make this assertion,
2420 * because it protects dn_phys / dn_next_nlevels from changing.
2422 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
2423 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
2424 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
2425 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
2426 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
2429 if (db
->db_level
== 0) {
2430 ASSERT(!db
->db_objset
->os_raw_receive
||
2431 dn
->dn_maxblkid
>= db
->db_blkid
);
2432 dnode_new_blkid(dn
, db
->db_blkid
, tx
,
2433 drop_struct_rwlock
, B_FALSE
);
2434 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
2437 if (db
->db_level
+1 < dn
->dn_nlevels
) {
2438 dmu_buf_impl_t
*parent
= db
->db_parent
;
2439 dbuf_dirty_record_t
*di
;
2440 int parent_held
= FALSE
;
2442 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
2443 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2444 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2445 db
->db_blkid
>> epbs
, FTAG
);
2446 ASSERT(parent
!= NULL
);
2449 if (drop_struct_rwlock
)
2450 rw_exit(&dn
->dn_struct_rwlock
);
2451 ASSERT3U(db
->db_level
+ 1, ==, parent
->db_level
);
2452 di
= dbuf_dirty(parent
, tx
);
2454 dbuf_rele(parent
, FTAG
);
2456 mutex_enter(&db
->db_mtx
);
2458 * Since we've dropped the mutex, it's possible that
2459 * dbuf_undirty() might have changed this out from under us.
2461 if (list_head(&db
->db_dirty_records
) == dr
||
2462 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2463 mutex_enter(&di
->dt
.di
.dr_mtx
);
2464 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2465 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2466 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2467 mutex_exit(&di
->dt
.di
.dr_mtx
);
2470 mutex_exit(&db
->db_mtx
);
2472 ASSERT(db
->db_level
+ 1 == dn
->dn_nlevels
);
2473 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2474 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2475 mutex_enter(&dn
->dn_mtx
);
2476 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2477 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2478 mutex_exit(&dn
->dn_mtx
);
2479 if (drop_struct_rwlock
)
2480 rw_exit(&dn
->dn_struct_rwlock
);
2483 dnode_setdirty(dn
, tx
);
2489 dbuf_undirty_bonus(dbuf_dirty_record_t
*dr
)
2491 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2493 if (dr
->dt
.dl
.dr_data
!= db
->db
.db_data
) {
2494 struct dnode
*dn
= dr
->dr_dnode
;
2495 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
2497 kmem_free(dr
->dt
.dl
.dr_data
, max_bonuslen
);
2498 arc_space_return(max_bonuslen
, ARC_SPACE_BONUS
);
2500 db
->db_data_pending
= NULL
;
2501 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
2502 list_remove(&db
->db_dirty_records
, dr
);
2503 if (dr
->dr_dbuf
->db_level
!= 0) {
2504 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2505 list_destroy(&dr
->dt
.di
.dr_children
);
2507 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2508 ASSERT3U(db
->db_dirtycnt
, >, 0);
2509 db
->db_dirtycnt
-= 1;
2513 * Undirty a buffer in the transaction group referenced by the given
2514 * transaction. Return whether this evicted the dbuf.
2517 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2519 uint64_t txg
= tx
->tx_txg
;
2525 * Due to our use of dn_nlevels below, this can only be called
2526 * in open context, unless we are operating on the MOS.
2527 * From syncing context, dn_nlevels may be different from the
2528 * dn_nlevels used when dbuf was dirtied.
2530 ASSERT(db
->db_objset
==
2531 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2532 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2533 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2534 ASSERT0(db
->db_level
);
2535 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2538 * If this buffer is not dirty, we're done.
2540 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, txg
);
2543 ASSERT(dr
->dr_dbuf
== db
);
2545 brtwrite
= dr
->dt
.dl
.dr_brtwrite
;
2548 * We are freeing a block that we cloned in the same
2549 * transaction group.
2551 brt_pending_remove(dmu_objset_spa(db
->db_objset
),
2552 &dr
->dt
.dl
.dr_overridden_by
, tx
);
2555 dnode_t
*dn
= dr
->dr_dnode
;
2557 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2559 ASSERT(db
->db
.db_size
!= 0);
2561 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2562 dr
->dr_accounted
, txg
);
2564 list_remove(&db
->db_dirty_records
, dr
);
2567 * Note that there are three places in dbuf_dirty()
2568 * where this dirty record may be put on a list.
2569 * Make sure to do a list_remove corresponding to
2570 * every one of those list_insert calls.
2572 if (dr
->dr_parent
) {
2573 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2574 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2575 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2576 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2577 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2578 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2579 mutex_enter(&dn
->dn_mtx
);
2580 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2581 mutex_exit(&dn
->dn_mtx
);
2584 if (db
->db_state
!= DB_NOFILL
&& !brtwrite
) {
2585 dbuf_unoverride(dr
);
2587 ASSERT(db
->db_buf
!= NULL
);
2588 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2589 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2590 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2593 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2595 ASSERT(db
->db_dirtycnt
> 0);
2596 db
->db_dirtycnt
-= 1;
2598 if (zfs_refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2599 ASSERT(db
->db_state
== DB_NOFILL
|| brtwrite
||
2600 arc_released(db
->db_buf
));
2609 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2611 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2612 boolean_t undirty
= B_FALSE
;
2614 ASSERT(tx
->tx_txg
!= 0);
2615 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2618 * Quick check for dirtiness. For already dirty blocks, this
2619 * reduces runtime of this function by >90%, and overall performance
2620 * by 50% for some workloads (e.g. file deletion with indirect blocks
2623 mutex_enter(&db
->db_mtx
);
2625 if (db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
) {
2626 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2628 * It's possible that it is already dirty but not cached,
2629 * because there are some calls to dbuf_dirty() that don't
2630 * go through dmu_buf_will_dirty().
2633 if (dr
->dt
.dl
.dr_brtwrite
) {
2635 * Block cloning: If we are dirtying a cloned
2636 * block, we cannot simply redirty it, because
2637 * this dr has no data associated with it.
2638 * We will go through a full undirtying below,
2639 * before dirtying it again.
2643 /* This dbuf is already dirty and cached. */
2645 mutex_exit(&db
->db_mtx
);
2650 mutex_exit(&db
->db_mtx
);
2653 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2654 flags
|= DB_RF_HAVESTRUCT
;
2658 * Block cloning: Do the dbuf_read() before undirtying the dbuf, as we
2659 * want to make sure dbuf_read() will read the pending cloned block and
2660 * not the uderlying block that is being replaced. dbuf_undirty() will
2661 * do dbuf_unoverride(), so we will end up with cloned block content,
2662 * without overridden BP.
2664 (void) dbuf_read(db
, NULL
, flags
);
2666 mutex_enter(&db
->db_mtx
);
2667 VERIFY(!dbuf_undirty(db
, tx
));
2668 mutex_exit(&db
->db_mtx
);
2670 (void) dbuf_dirty(db
, tx
);
2674 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2676 dmu_buf_will_dirty_impl(db_fake
,
2677 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2681 dmu_buf_is_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2683 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2684 dbuf_dirty_record_t
*dr
;
2686 mutex_enter(&db
->db_mtx
);
2687 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2688 mutex_exit(&db
->db_mtx
);
2689 return (dr
!= NULL
);
2693 dmu_buf_will_clone(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2695 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2698 * Block cloning: We are going to clone into this block, so undirty
2699 * modifications done to this block so far in this txg. This includes
2700 * writes and clones into this block.
2702 mutex_enter(&db
->db_mtx
);
2704 VERIFY(!dbuf_undirty(db
, tx
));
2705 ASSERT3P(dbuf_find_dirty_eq(db
, tx
->tx_txg
), ==, NULL
);
2706 if (db
->db_buf
!= NULL
) {
2707 arc_buf_destroy(db
->db_buf
, db
);
2709 dbuf_clear_data(db
);
2712 db
->db_state
= DB_NOFILL
;
2713 DTRACE_SET_STATE(db
, "allocating NOFILL buffer for clone");
2716 mutex_exit(&db
->db_mtx
);
2719 (void) dbuf_dirty(db
, tx
);
2723 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2725 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2727 mutex_enter(&db
->db_mtx
);
2728 db
->db_state
= DB_NOFILL
;
2729 DTRACE_SET_STATE(db
, "allocating NOFILL buffer");
2730 mutex_exit(&db
->db_mtx
);
2733 (void) dbuf_dirty(db
, tx
);
2737 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
, boolean_t canfail
)
2739 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2741 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2742 ASSERT(tx
->tx_txg
!= 0);
2743 ASSERT(db
->db_level
== 0);
2744 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2746 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2747 dmu_tx_private_ok(tx
));
2749 mutex_enter(&db
->db_mtx
);
2750 if (db
->db_state
== DB_NOFILL
) {
2752 * Block cloning: We will be completely overwriting a block
2753 * cloned in this transaction group, so let's undirty the
2754 * pending clone and mark the block as uncached. This will be
2755 * as if the clone was never done. But if the fill can fail
2756 * we should have a way to return back to the cloned data.
2758 if (canfail
&& dbuf_find_dirty_eq(db
, tx
->tx_txg
) != NULL
) {
2759 mutex_exit(&db
->db_mtx
);
2760 dmu_buf_will_dirty(db_fake
, tx
);
2763 VERIFY(!dbuf_undirty(db
, tx
));
2764 db
->db_state
= DB_UNCACHED
;
2766 mutex_exit(&db
->db_mtx
);
2769 (void) dbuf_dirty(db
, tx
);
2773 * This function is effectively the same as dmu_buf_will_dirty(), but
2774 * indicates the caller expects raw encrypted data in the db, and provides
2775 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2776 * blkptr_t when this dbuf is written. This is only used for blocks of
2777 * dnodes, during raw receive.
2780 dmu_buf_set_crypt_params(dmu_buf_t
*db_fake
, boolean_t byteorder
,
2781 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
, dmu_tx_t
*tx
)
2783 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2784 dbuf_dirty_record_t
*dr
;
2787 * dr_has_raw_params is only processed for blocks of dnodes
2788 * (see dbuf_sync_dnode_leaf_crypt()).
2790 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
2791 ASSERT3U(db
->db_level
, ==, 0);
2792 ASSERT(db
->db_objset
->os_raw_receive
);
2794 dmu_buf_will_dirty_impl(db_fake
,
2795 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2797 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2799 ASSERT3P(dr
, !=, NULL
);
2801 dr
->dt
.dl
.dr_has_raw_params
= B_TRUE
;
2802 dr
->dt
.dl
.dr_byteorder
= byteorder
;
2803 memcpy(dr
->dt
.dl
.dr_salt
, salt
, ZIO_DATA_SALT_LEN
);
2804 memcpy(dr
->dt
.dl
.dr_iv
, iv
, ZIO_DATA_IV_LEN
);
2805 memcpy(dr
->dt
.dl
.dr_mac
, mac
, ZIO_DATA_MAC_LEN
);
2809 dbuf_override_impl(dmu_buf_impl_t
*db
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
2811 struct dirty_leaf
*dl
;
2812 dbuf_dirty_record_t
*dr
;
2814 dr
= list_head(&db
->db_dirty_records
);
2815 ASSERT3P(dr
, !=, NULL
);
2816 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2818 dl
->dr_overridden_by
= *bp
;
2819 dl
->dr_override_state
= DR_OVERRIDDEN
;
2820 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2824 dmu_buf_fill_done(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
, boolean_t failed
)
2827 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2828 mutex_enter(&db
->db_mtx
);
2831 if (db
->db_state
== DB_FILL
) {
2832 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2833 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2834 /* we were freed while filling */
2835 /* XXX dbuf_undirty? */
2836 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2837 db
->db_freed_in_flight
= FALSE
;
2838 db
->db_state
= DB_CACHED
;
2839 DTRACE_SET_STATE(db
,
2840 "fill done handling freed in flight");
2842 } else if (failed
) {
2843 VERIFY(!dbuf_undirty(db
, tx
));
2845 dbuf_clear_data(db
);
2846 DTRACE_SET_STATE(db
, "fill failed");
2848 db
->db_state
= DB_CACHED
;
2849 DTRACE_SET_STATE(db
, "fill done");
2851 cv_broadcast(&db
->db_changed
);
2853 db
->db_state
= DB_CACHED
;
2856 mutex_exit(&db
->db_mtx
);
2861 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2862 bp_embedded_type_t etype
, enum zio_compress comp
,
2863 int uncompressed_size
, int compressed_size
, int byteorder
,
2866 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2867 struct dirty_leaf
*dl
;
2868 dmu_object_type_t type
;
2869 dbuf_dirty_record_t
*dr
;
2871 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2872 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2873 SPA_FEATURE_EMBEDDED_DATA
));
2877 type
= DB_DNODE(db
)->dn_type
;
2880 ASSERT0(db
->db_level
);
2881 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2883 dmu_buf_will_not_fill(dbuf
, tx
);
2885 dr
= list_head(&db
->db_dirty_records
);
2886 ASSERT3P(dr
, !=, NULL
);
2887 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2889 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2890 data
, comp
, uncompressed_size
, compressed_size
);
2891 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2892 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2893 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2894 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2896 dl
->dr_override_state
= DR_OVERRIDDEN
;
2897 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2901 dmu_buf_redact(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2903 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2904 dmu_object_type_t type
;
2905 ASSERT(dsl_dataset_feature_is_active(db
->db_objset
->os_dsl_dataset
,
2906 SPA_FEATURE_REDACTED_DATASETS
));
2909 type
= DB_DNODE(db
)->dn_type
;
2912 ASSERT0(db
->db_level
);
2913 dmu_buf_will_not_fill(dbuf
, tx
);
2915 blkptr_t bp
= { { { {0} } } };
2916 BP_SET_TYPE(&bp
, type
);
2917 BP_SET_LEVEL(&bp
, 0);
2918 BP_SET_BIRTH(&bp
, tx
->tx_txg
, 0);
2919 BP_SET_REDACTED(&bp
);
2920 BPE_SET_LSIZE(&bp
, dbuf
->db_size
);
2922 dbuf_override_impl(db
, &bp
, tx
);
2926 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2927 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2930 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2932 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2933 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2934 ASSERT(db
->db_level
== 0);
2935 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2936 ASSERT(buf
!= NULL
);
2937 ASSERT3U(arc_buf_lsize(buf
), ==, db
->db
.db_size
);
2938 ASSERT(tx
->tx_txg
!= 0);
2940 arc_return_buf(buf
, db
);
2941 ASSERT(arc_released(buf
));
2943 mutex_enter(&db
->db_mtx
);
2945 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2946 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2948 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
||
2949 db
->db_state
== DB_NOFILL
);
2951 if (db
->db_state
== DB_CACHED
&&
2952 zfs_refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2954 * In practice, we will never have a case where we have an
2955 * encrypted arc buffer while additional holds exist on the
2956 * dbuf. We don't handle this here so we simply assert that
2959 ASSERT(!arc_is_encrypted(buf
));
2960 mutex_exit(&db
->db_mtx
);
2961 (void) dbuf_dirty(db
, tx
);
2962 memcpy(db
->db
.db_data
, buf
->b_data
, db
->db
.db_size
);
2963 arc_buf_destroy(buf
, db
);
2967 if (db
->db_state
== DB_CACHED
) {
2968 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
2970 ASSERT(db
->db_buf
!= NULL
);
2971 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2972 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2974 if (!arc_released(db
->db_buf
)) {
2975 ASSERT(dr
->dt
.dl
.dr_override_state
==
2977 arc_release(db
->db_buf
, db
);
2979 dr
->dt
.dl
.dr_data
= buf
;
2980 arc_buf_destroy(db
->db_buf
, db
);
2981 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2982 arc_release(db
->db_buf
, db
);
2983 arc_buf_destroy(db
->db_buf
, db
);
2986 } else if (db
->db_state
== DB_NOFILL
) {
2988 * We will be completely replacing the cloned block. In case
2989 * it was cloned in this transaction group, let's undirty the
2990 * pending clone and mark the block as uncached. This will be
2991 * as if the clone was never done.
2993 VERIFY(!dbuf_undirty(db
, tx
));
2994 db
->db_state
= DB_UNCACHED
;
2996 ASSERT(db
->db_buf
== NULL
);
2997 dbuf_set_data(db
, buf
);
2998 db
->db_state
= DB_FILL
;
2999 DTRACE_SET_STATE(db
, "filling assigned arcbuf");
3000 mutex_exit(&db
->db_mtx
);
3001 (void) dbuf_dirty(db
, tx
);
3002 dmu_buf_fill_done(&db
->db
, tx
, B_FALSE
);
3006 dbuf_destroy(dmu_buf_impl_t
*db
)
3009 dmu_buf_impl_t
*parent
= db
->db_parent
;
3010 dmu_buf_impl_t
*dndb
;
3012 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3013 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3015 if (db
->db_buf
!= NULL
) {
3016 arc_buf_destroy(db
->db_buf
, db
);
3020 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3021 int slots
= DB_DNODE(db
)->dn_num_slots
;
3022 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3023 if (db
->db
.db_data
!= NULL
) {
3024 kmem_free(db
->db
.db_data
, bonuslen
);
3025 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3026 db
->db_state
= DB_UNCACHED
;
3027 DTRACE_SET_STATE(db
, "buffer cleared");
3031 dbuf_clear_data(db
);
3033 if (multilist_link_active(&db
->db_cache_link
)) {
3034 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3035 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3037 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
3038 (void) zfs_refcount_remove_many(
3039 &dbuf_caches
[db
->db_caching_status
].size
,
3040 db
->db
.db_size
, db
);
3042 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3043 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3045 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3046 DBUF_STAT_BUMPDOWN(cache_count
);
3047 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3050 db
->db_caching_status
= DB_NO_CACHE
;
3053 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
3054 ASSERT(db
->db_data_pending
== NULL
);
3055 ASSERT(list_is_empty(&db
->db_dirty_records
));
3057 db
->db_state
= DB_EVICTING
;
3058 DTRACE_SET_STATE(db
, "buffer eviction started");
3059 db
->db_blkptr
= NULL
;
3062 * Now that db_state is DB_EVICTING, nobody else can find this via
3063 * the hash table. We can now drop db_mtx, which allows us to
3064 * acquire the dn_dbufs_mtx.
3066 mutex_exit(&db
->db_mtx
);
3071 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
3072 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
3074 mutex_enter_nested(&dn
->dn_dbufs_mtx
,
3076 avl_remove(&dn
->dn_dbufs
, db
);
3080 mutex_exit(&dn
->dn_dbufs_mtx
);
3082 * Decrementing the dbuf count means that the hold corresponding
3083 * to the removed dbuf is no longer discounted in dnode_move(),
3084 * so the dnode cannot be moved until after we release the hold.
3085 * The membar_producer() ensures visibility of the decremented
3086 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
3089 mutex_enter(&dn
->dn_mtx
);
3090 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
3091 db
->db_dnode_handle
= NULL
;
3093 dbuf_hash_remove(db
);
3098 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3100 db
->db_parent
= NULL
;
3102 ASSERT(db
->db_buf
== NULL
);
3103 ASSERT(db
->db
.db_data
== NULL
);
3104 ASSERT(db
->db_hash_next
== NULL
);
3105 ASSERT(db
->db_blkptr
== NULL
);
3106 ASSERT(db
->db_data_pending
== NULL
);
3107 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
3108 ASSERT(!multilist_link_active(&db
->db_cache_link
));
3111 * If this dbuf is referenced from an indirect dbuf,
3112 * decrement the ref count on the indirect dbuf.
3114 if (parent
&& parent
!= dndb
) {
3115 mutex_enter(&parent
->db_mtx
);
3116 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
3119 kmem_cache_free(dbuf_kmem_cache
, db
);
3120 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3124 * Note: While bpp will always be updated if the function returns success,
3125 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
3126 * this happens when the dnode is the meta-dnode, or {user|group|project}used
3129 __attribute__((always_inline
))
3131 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
3132 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
3137 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3139 if (blkid
== DMU_SPILL_BLKID
) {
3140 mutex_enter(&dn
->dn_mtx
);
3141 if (dn
->dn_have_spill
&&
3142 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
3143 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3146 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3147 *parentp
= dn
->dn_dbuf
;
3148 mutex_exit(&dn
->dn_mtx
);
3153 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
3154 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3156 ASSERT3U(level
* epbs
, <, 64);
3157 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3159 * This assertion shouldn't trip as long as the max indirect block size
3160 * is less than 1M. The reason for this is that up to that point,
3161 * the number of levels required to address an entire object with blocks
3162 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
3163 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
3164 * (i.e. we can address the entire object), objects will all use at most
3165 * N-1 levels and the assertion won't overflow. However, once epbs is
3166 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
3167 * enough to address an entire object, so objects will have 5 levels,
3168 * but then this assertion will overflow.
3170 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
3171 * need to redo this logic to handle overflows.
3173 ASSERT(level
>= nlevels
||
3174 ((nlevels
- level
- 1) * epbs
) +
3175 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
3176 if (level
>= nlevels
||
3177 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
3178 ((nlevels
- level
- 1) * epbs
)) ||
3180 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
3181 /* the buffer has no parent yet */
3182 return (SET_ERROR(ENOENT
));
3183 } else if (level
< nlevels
-1) {
3184 /* this block is referenced from an indirect block */
3187 err
= dbuf_hold_impl(dn
, level
+ 1,
3188 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
3192 err
= dbuf_read(*parentp
, NULL
,
3193 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
3195 dbuf_rele(*parentp
, NULL
);
3199 rw_enter(&(*parentp
)->db_rwlock
, RW_READER
);
3200 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
3201 (blkid
& ((1ULL << epbs
) - 1));
3202 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
3203 ASSERT(BP_IS_HOLE(*bpp
));
3204 rw_exit(&(*parentp
)->db_rwlock
);
3207 /* the block is referenced from the dnode */
3208 ASSERT3U(level
, ==, nlevels
-1);
3209 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
3210 blkid
< dn
->dn_phys
->dn_nblkptr
);
3212 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3213 *parentp
= dn
->dn_dbuf
;
3215 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
3220 static dmu_buf_impl_t
*
3221 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3222 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
, uint64_t hash
)
3224 objset_t
*os
= dn
->dn_objset
;
3225 dmu_buf_impl_t
*db
, *odb
;
3227 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3228 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
3230 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
3232 list_create(&db
->db_dirty_records
, sizeof (dbuf_dirty_record_t
),
3233 offsetof(dbuf_dirty_record_t
, dr_dbuf_node
));
3236 db
->db
.db_object
= dn
->dn_object
;
3237 db
->db_level
= level
;
3238 db
->db_blkid
= blkid
;
3239 db
->db_dirtycnt
= 0;
3240 db
->db_dnode_handle
= dn
->dn_handle
;
3241 db
->db_parent
= parent
;
3242 db
->db_blkptr
= blkptr
;
3246 db
->db_user_immediate_evict
= FALSE
;
3247 db
->db_freed_in_flight
= FALSE
;
3248 db
->db_pending_evict
= FALSE
;
3250 if (blkid
== DMU_BONUS_BLKID
) {
3251 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
3252 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
3253 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
3254 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
3255 db
->db
.db_offset
= DMU_BONUS_BLKID
;
3256 db
->db_state
= DB_UNCACHED
;
3257 DTRACE_SET_STATE(db
, "bonus buffer created");
3258 db
->db_caching_status
= DB_NO_CACHE
;
3259 /* the bonus dbuf is not placed in the hash table */
3260 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3262 } else if (blkid
== DMU_SPILL_BLKID
) {
3263 db
->db
.db_size
= (blkptr
!= NULL
) ?
3264 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
3265 db
->db
.db_offset
= 0;
3268 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
3269 db
->db
.db_size
= blocksize
;
3270 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
3274 * Hold the dn_dbufs_mtx while we get the new dbuf
3275 * in the hash table *and* added to the dbufs list.
3276 * This prevents a possible deadlock with someone
3277 * trying to look up this dbuf before it's added to the
3280 mutex_enter(&dn
->dn_dbufs_mtx
);
3281 db
->db_state
= DB_EVICTING
; /* not worth logging this state change */
3282 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
3283 /* someone else inserted it first */
3284 mutex_exit(&dn
->dn_dbufs_mtx
);
3285 kmem_cache_free(dbuf_kmem_cache
, db
);
3286 DBUF_STAT_BUMP(hash_insert_race
);
3289 avl_add(&dn
->dn_dbufs
, db
);
3291 db
->db_state
= DB_UNCACHED
;
3292 DTRACE_SET_STATE(db
, "regular buffer created");
3293 db
->db_caching_status
= DB_NO_CACHE
;
3294 mutex_exit(&dn
->dn_dbufs_mtx
);
3295 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3297 if (parent
&& parent
!= dn
->dn_dbuf
)
3298 dbuf_add_ref(parent
, db
);
3300 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
3301 zfs_refcount_count(&dn
->dn_holds
) > 0);
3302 (void) zfs_refcount_add(&dn
->dn_holds
, db
);
3304 dprintf_dbuf(db
, "db=%p\n", db
);
3310 * This function returns a block pointer and information about the object,
3311 * given a dnode and a block. This is a publicly accessible version of
3312 * dbuf_findbp that only returns some information, rather than the
3313 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
3314 * should be locked as (at least) a reader.
3317 dbuf_dnode_findbp(dnode_t
*dn
, uint64_t level
, uint64_t blkid
,
3318 blkptr_t
*bp
, uint16_t *datablkszsec
, uint8_t *indblkshift
)
3320 dmu_buf_impl_t
*dbp
= NULL
;
3323 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3325 err
= dbuf_findbp(dn
, level
, blkid
, B_FALSE
, &dbp
, &bp2
);
3327 ASSERT3P(bp2
, !=, NULL
);
3330 dbuf_rele(dbp
, NULL
);
3331 if (datablkszsec
!= NULL
)
3332 *datablkszsec
= dn
->dn_phys
->dn_datablkszsec
;
3333 if (indblkshift
!= NULL
)
3334 *indblkshift
= dn
->dn_phys
->dn_indblkshift
;
3340 typedef struct dbuf_prefetch_arg
{
3341 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
3342 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
3343 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
3344 int dpa_curlevel
; /* The current level that we're reading */
3345 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
3346 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
3347 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
3348 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
3349 dbuf_prefetch_fn dpa_cb
; /* prefetch completion callback */
3350 void *dpa_arg
; /* prefetch completion arg */
3351 } dbuf_prefetch_arg_t
;
3354 dbuf_prefetch_fini(dbuf_prefetch_arg_t
*dpa
, boolean_t io_done
)
3356 if (dpa
->dpa_cb
!= NULL
) {
3357 dpa
->dpa_cb(dpa
->dpa_arg
, dpa
->dpa_zb
.zb_level
,
3358 dpa
->dpa_zb
.zb_blkid
, io_done
);
3360 kmem_free(dpa
, sizeof (*dpa
));
3364 dbuf_issue_final_prefetch_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3365 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3367 (void) zio
, (void) zb
, (void) iobp
;
3368 dbuf_prefetch_arg_t
*dpa
= private;
3371 arc_buf_destroy(abuf
, private);
3373 dbuf_prefetch_fini(dpa
, B_TRUE
);
3377 * Actually issue the prefetch read for the block given.
3380 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
3382 ASSERT(!BP_IS_REDACTED(bp
) ||
3383 dsl_dataset_feature_is_active(
3384 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3385 SPA_FEATURE_REDACTED_DATASETS
));
3387 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
) || BP_IS_REDACTED(bp
))
3388 return (dbuf_prefetch_fini(dpa
, B_FALSE
));
3390 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
3391 arc_flags_t aflags
=
3392 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
|
3395 /* dnodes are always read as raw and then converted later */
3396 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
3397 dpa
->dpa_curlevel
== 0)
3398 zio_flags
|= ZIO_FLAG_RAW
;
3400 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3401 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
3402 ASSERT(dpa
->dpa_zio
!= NULL
);
3403 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
,
3404 dbuf_issue_final_prefetch_done
, dpa
,
3405 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
3409 * Called when an indirect block above our prefetch target is read in. This
3410 * will either read in the next indirect block down the tree or issue the actual
3411 * prefetch if the next block down is our target.
3414 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3415 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3417 (void) zb
, (void) iobp
;
3418 dbuf_prefetch_arg_t
*dpa
= private;
3420 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
3421 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
3424 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
3425 dbuf_prefetch_fini(dpa
, B_TRUE
);
3428 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
3431 * The dpa_dnode is only valid if we are called with a NULL
3432 * zio. This indicates that the arc_read() returned without
3433 * first calling zio_read() to issue a physical read. Once
3434 * a physical read is made the dpa_dnode must be invalidated
3435 * as the locks guarding it may have been dropped. If the
3436 * dpa_dnode is still valid, then we want to add it to the dbuf
3437 * cache. To do so, we must hold the dbuf associated with the block
3438 * we just prefetched, read its contents so that we associate it
3439 * with an arc_buf_t, and then release it.
3442 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
3443 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
3444 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
3446 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
3448 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
3450 dpa
->dpa_dnode
= NULL
;
3451 } else if (dpa
->dpa_dnode
!= NULL
) {
3452 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
3453 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
3454 dpa
->dpa_zb
.zb_level
));
3455 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
3456 dpa
->dpa_curlevel
, curblkid
, FTAG
);
3458 arc_buf_destroy(abuf
, private);
3459 dbuf_prefetch_fini(dpa
, B_TRUE
);
3462 (void) dbuf_read(db
, NULL
,
3463 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
3464 dbuf_rele(db
, FTAG
);
3467 dpa
->dpa_curlevel
--;
3468 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
3469 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
3470 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
3471 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
3473 ASSERT(!BP_IS_REDACTED(bp
) || (dpa
->dpa_dnode
&&
3474 dsl_dataset_feature_is_active(
3475 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3476 SPA_FEATURE_REDACTED_DATASETS
)));
3477 if (BP_IS_HOLE(bp
) || BP_IS_REDACTED(bp
)) {
3478 arc_buf_destroy(abuf
, private);
3479 dbuf_prefetch_fini(dpa
, B_TRUE
);
3481 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
3482 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
3483 dbuf_issue_final_prefetch(dpa
, bp
);
3485 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3486 zbookmark_phys_t zb
;
3488 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3489 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
3490 iter_aflags
|= ARC_FLAG_L2CACHE
;
3492 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3494 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
3495 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
3497 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3498 bp
, dbuf_prefetch_indirect_done
, dpa
,
3499 ZIO_PRIORITY_SYNC_READ
,
3500 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3504 arc_buf_destroy(abuf
, private);
3508 * Issue prefetch reads for the given block on the given level. If the indirect
3509 * blocks above that block are not in memory, we will read them in
3510 * asynchronously. As a result, this call never blocks waiting for a read to
3511 * complete. Note that the prefetch might fail if the dataset is encrypted and
3512 * the encryption key is unmapped before the IO completes.
3515 dbuf_prefetch_impl(dnode_t
*dn
, int64_t level
, uint64_t blkid
,
3516 zio_priority_t prio
, arc_flags_t aflags
, dbuf_prefetch_fn cb
,
3520 int epbs
, nlevels
, curlevel
;
3523 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3524 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3526 if (blkid
> dn
->dn_maxblkid
)
3529 if (level
== 0 && dnode_block_freed(dn
, blkid
))
3533 * This dnode hasn't been written to disk yet, so there's nothing to
3536 nlevels
= dn
->dn_phys
->dn_nlevels
;
3537 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
3540 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3541 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
3544 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
3545 level
, blkid
, NULL
);
3547 mutex_exit(&db
->db_mtx
);
3549 * This dbuf already exists. It is either CACHED, or
3550 * (we assume) about to be read or filled.
3556 * Find the closest ancestor (indirect block) of the target block
3557 * that is present in the cache. In this indirect block, we will
3558 * find the bp that is at curlevel, curblkid.
3562 while (curlevel
< nlevels
- 1) {
3563 int parent_level
= curlevel
+ 1;
3564 uint64_t parent_blkid
= curblkid
>> epbs
;
3567 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
3568 FALSE
, TRUE
, FTAG
, &db
) == 0) {
3569 blkptr_t
*bpp
= db
->db_buf
->b_data
;
3570 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
3571 dbuf_rele(db
, FTAG
);
3575 curlevel
= parent_level
;
3576 curblkid
= parent_blkid
;
3579 if (curlevel
== nlevels
- 1) {
3580 /* No cached indirect blocks found. */
3581 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
3582 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
3584 ASSERT(!BP_IS_REDACTED(&bp
) ||
3585 dsl_dataset_feature_is_active(dn
->dn_objset
->os_dsl_dataset
,
3586 SPA_FEATURE_REDACTED_DATASETS
));
3587 if (BP_IS_HOLE(&bp
) || BP_IS_REDACTED(&bp
))
3590 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
3592 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
3595 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
3596 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
3597 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3598 dn
->dn_object
, level
, blkid
);
3599 dpa
->dpa_curlevel
= curlevel
;
3600 dpa
->dpa_prio
= prio
;
3601 dpa
->dpa_aflags
= aflags
;
3602 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
3603 dpa
->dpa_dnode
= dn
;
3604 dpa
->dpa_epbs
= epbs
;
3609 if (!DNODE_LEVEL_IS_CACHEABLE(dn
, level
))
3610 dpa
->dpa_aflags
|= ARC_FLAG_UNCACHED
;
3611 else if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3612 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
3615 * If we have the indirect just above us, no need to do the asynchronous
3616 * prefetch chain; we'll just run the last step ourselves. If we're at
3617 * a higher level, though, we want to issue the prefetches for all the
3618 * indirect blocks asynchronously, so we can go on with whatever we were
3621 if (curlevel
== level
) {
3622 ASSERT3U(curblkid
, ==, blkid
);
3623 dbuf_issue_final_prefetch(dpa
, &bp
);
3625 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3626 zbookmark_phys_t zb
;
3628 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3629 if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3630 iter_aflags
|= ARC_FLAG_L2CACHE
;
3632 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3633 dn
->dn_object
, curlevel
, curblkid
);
3634 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3635 &bp
, dbuf_prefetch_indirect_done
, dpa
,
3636 ZIO_PRIORITY_SYNC_READ
,
3637 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3641 * We use pio here instead of dpa_zio since it's possible that
3642 * dpa may have already been freed.
3648 cb(arg
, level
, blkid
, B_FALSE
);
3653 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
3657 return (dbuf_prefetch_impl(dn
, level
, blkid
, prio
, aflags
, NULL
, NULL
));
3661 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3662 * the case of encrypted, compressed and uncompressed buffers by
3663 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3664 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3666 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3668 noinline
static void
3669 dbuf_hold_copy(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3671 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3672 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
3673 enum zio_compress compress_type
= arc_get_compression(data
);
3674 uint8_t complevel
= arc_get_complevel(data
);
3676 if (arc_is_encrypted(data
)) {
3677 boolean_t byteorder
;
3678 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3679 uint8_t iv
[ZIO_DATA_IV_LEN
];
3680 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3682 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
3683 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
3684 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
3685 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
3686 compress_type
, complevel
));
3687 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3688 dbuf_set_data(db
, arc_alloc_compressed_buf(
3689 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
3690 arc_buf_lsize(data
), compress_type
, complevel
));
3692 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
3693 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
3696 rw_enter(&db
->db_rwlock
, RW_WRITER
);
3697 memcpy(db
->db
.db_data
, data
->b_data
, arc_buf_size(data
));
3698 rw_exit(&db
->db_rwlock
);
3702 * Returns with db_holds incremented, and db_mtx not held.
3703 * Note: dn_struct_rwlock must be held.
3706 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3707 boolean_t fail_sparse
, boolean_t fail_uncached
,
3708 const void *tag
, dmu_buf_impl_t
**dbp
)
3710 dmu_buf_impl_t
*db
, *parent
= NULL
;
3713 /* If the pool has been created, verify the tx_sync_lock is not held */
3714 spa_t
*spa
= dn
->dn_objset
->os_spa
;
3715 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
3717 ASSERT(!MUTEX_HELD(&dp
->dp_tx
.tx_sync_lock
));
3720 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3721 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3722 ASSERT3U(dn
->dn_nlevels
, >, level
);
3726 /* dbuf_find() returns with db_mtx held */
3727 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
, &hv
);
3730 blkptr_t
*bp
= NULL
;
3734 return (SET_ERROR(ENOENT
));
3736 ASSERT3P(parent
, ==, NULL
);
3737 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
3739 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
3740 err
= SET_ERROR(ENOENT
);
3743 dbuf_rele(parent
, NULL
);
3747 if (err
&& err
!= ENOENT
)
3749 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
, hv
);
3752 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
3753 mutex_exit(&db
->db_mtx
);
3754 return (SET_ERROR(ENOENT
));
3757 if (db
->db_buf
!= NULL
) {
3758 arc_buf_access(db
->db_buf
);
3759 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
3762 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
3765 * If this buffer is currently syncing out, and we are
3766 * still referencing it from db_data, we need to make a copy
3767 * of it in case we decide we want to dirty it again in this txg.
3769 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
3770 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3771 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
3772 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3773 if (dr
->dt
.dl
.dr_data
== db
->db_buf
) {
3774 ASSERT3P(db
->db_buf
, !=, NULL
);
3775 dbuf_hold_copy(dn
, db
);
3779 if (multilist_link_active(&db
->db_cache_link
)) {
3780 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3781 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3782 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3784 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
3785 (void) zfs_refcount_remove_many(
3786 &dbuf_caches
[db
->db_caching_status
].size
,
3787 db
->db
.db_size
, db
);
3789 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3790 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3792 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3793 DBUF_STAT_BUMPDOWN(cache_count
);
3794 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3797 db
->db_caching_status
= DB_NO_CACHE
;
3799 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3801 mutex_exit(&db
->db_mtx
);
3803 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3805 dbuf_rele(parent
, NULL
);
3807 ASSERT3P(DB_DNODE(db
), ==, dn
);
3808 ASSERT3U(db
->db_blkid
, ==, blkid
);
3809 ASSERT3U(db
->db_level
, ==, level
);
3816 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, const void *tag
)
3818 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3822 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, const void *tag
)
3825 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3826 return (err
? NULL
: db
);
3830 dbuf_create_bonus(dnode_t
*dn
)
3832 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3834 ASSERT(dn
->dn_bonus
== NULL
);
3835 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
,
3836 dbuf_hash(dn
->dn_objset
, dn
->dn_object
, 0, DMU_BONUS_BLKID
));
3840 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3842 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3844 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3845 return (SET_ERROR(ENOTSUP
));
3847 blksz
= SPA_MINBLOCKSIZE
;
3848 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3849 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3851 dbuf_new_size(db
, blksz
, tx
);
3857 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3859 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3862 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3864 dbuf_add_ref(dmu_buf_impl_t
*db
, const void *tag
)
3866 int64_t holds
= zfs_refcount_add(&db
->db_holds
, tag
);
3867 VERIFY3S(holds
, >, 1);
3870 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3872 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3875 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3876 dmu_buf_impl_t
*found_db
;
3877 boolean_t result
= B_FALSE
;
3879 if (blkid
== DMU_BONUS_BLKID
)
3880 found_db
= dbuf_find_bonus(os
, obj
);
3882 found_db
= dbuf_find(os
, obj
, 0, blkid
, NULL
);
3884 if (found_db
!= NULL
) {
3885 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3886 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3889 mutex_exit(&found_db
->db_mtx
);
3895 * If you call dbuf_rele() you had better not be referencing the dnode handle
3896 * unless you have some other direct or indirect hold on the dnode. (An indirect
3897 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3898 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3899 * dnode's parent dbuf evicting its dnode handles.
3902 dbuf_rele(dmu_buf_impl_t
*db
, const void *tag
)
3904 mutex_enter(&db
->db_mtx
);
3905 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
3909 dmu_buf_rele(dmu_buf_t
*db
, const void *tag
)
3911 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3915 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3916 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3917 * argument should be set if we are already in the dbuf-evicting code
3918 * path, in which case we don't want to recursively evict. This allows us to
3919 * avoid deeply nested stacks that would have a call flow similar to this:
3921 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3924 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3928 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, const void *tag
, boolean_t evicting
)
3933 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3937 * Remove the reference to the dbuf before removing its hold on the
3938 * dnode so we can guarantee in dnode_move() that a referenced bonus
3939 * buffer has a corresponding dnode hold.
3941 holds
= zfs_refcount_remove(&db
->db_holds
, tag
);
3945 * We can't freeze indirects if there is a possibility that they
3946 * may be modified in the current syncing context.
3948 if (db
->db_buf
!= NULL
&&
3949 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3950 arc_buf_freeze(db
->db_buf
);
3953 if (holds
== db
->db_dirtycnt
&&
3954 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3955 dbuf_evict_user(db
);
3958 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3960 boolean_t evict_dbuf
= db
->db_pending_evict
;
3963 * If the dnode moves here, we cannot cross this
3964 * barrier until the move completes.
3969 atomic_dec_32(&dn
->dn_dbufs_count
);
3972 * Decrementing the dbuf count means that the bonus
3973 * buffer's dnode hold is no longer discounted in
3974 * dnode_move(). The dnode cannot move until after
3975 * the dnode_rele() below.
3980 * Do not reference db after its lock is dropped.
3981 * Another thread may evict it.
3983 mutex_exit(&db
->db_mtx
);
3986 dnode_evict_bonus(dn
);
3989 } else if (db
->db_buf
== NULL
) {
3991 * This is a special case: we never associated this
3992 * dbuf with any data allocated from the ARC.
3994 ASSERT(db
->db_state
== DB_UNCACHED
||
3995 db
->db_state
== DB_NOFILL
);
3997 } else if (arc_released(db
->db_buf
)) {
3999 * This dbuf has anonymous data associated with it.
4002 } else if (!(DBUF_IS_CACHEABLE(db
) || db
->db_partial_read
) ||
4003 db
->db_pending_evict
) {
4005 } else if (!multilist_link_active(&db
->db_cache_link
)) {
4006 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
4008 dbuf_cached_state_t dcs
=
4009 dbuf_include_in_metadata_cache(db
) ?
4010 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
4011 db
->db_caching_status
= dcs
;
4013 multilist_insert(&dbuf_caches
[dcs
].cache
, db
);
4014 uint64_t db_size
= db
->db
.db_size
;
4015 size
= zfs_refcount_add_many(
4016 &dbuf_caches
[dcs
].size
, db_size
, db
);
4017 uint8_t db_level
= db
->db_level
;
4018 mutex_exit(&db
->db_mtx
);
4020 if (dcs
== DB_DBUF_METADATA_CACHE
) {
4021 DBUF_STAT_BUMP(metadata_cache_count
);
4022 DBUF_STAT_MAX(metadata_cache_size_bytes_max
,
4025 DBUF_STAT_BUMP(cache_count
);
4026 DBUF_STAT_MAX(cache_size_bytes_max
, size
);
4027 DBUF_STAT_BUMP(cache_levels
[db_level
]);
4028 DBUF_STAT_INCR(cache_levels_bytes
[db_level
],
4032 if (dcs
== DB_DBUF_CACHE
&& !evicting
)
4033 dbuf_evict_notify(size
);
4036 mutex_exit(&db
->db_mtx
);
4041 #pragma weak dmu_buf_refcount = dbuf_refcount
4043 dbuf_refcount(dmu_buf_impl_t
*db
)
4045 return (zfs_refcount_count(&db
->db_holds
));
4049 dmu_buf_user_refcount(dmu_buf_t
*db_fake
)
4052 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4054 mutex_enter(&db
->db_mtx
);
4055 ASSERT3U(zfs_refcount_count(&db
->db_holds
), >=, db
->db_dirtycnt
);
4056 holds
= zfs_refcount_count(&db
->db_holds
) - db
->db_dirtycnt
;
4057 mutex_exit(&db
->db_mtx
);
4063 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
4064 dmu_buf_user_t
*new_user
)
4066 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4068 mutex_enter(&db
->db_mtx
);
4069 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4070 if (db
->db_user
== old_user
)
4071 db
->db_user
= new_user
;
4073 old_user
= db
->db_user
;
4074 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4075 mutex_exit(&db
->db_mtx
);
4081 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4083 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
4087 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4089 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4091 db
->db_user_immediate_evict
= TRUE
;
4092 return (dmu_buf_set_user(db_fake
, user
));
4096 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
4098 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
4102 dmu_buf_get_user(dmu_buf_t
*db_fake
)
4104 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
4106 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
4107 return (db
->db_user
);
4111 dmu_buf_user_evict_wait(void)
4113 taskq_wait(dbu_evict_taskq
);
4117 dmu_buf_get_blkptr(dmu_buf_t
*db
)
4119 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4120 return (dbi
->db_blkptr
);
4124 dmu_buf_get_objset(dmu_buf_t
*db
)
4126 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4127 return (dbi
->db_objset
);
4131 dmu_buf_dnode_enter(dmu_buf_t
*db
)
4133 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4134 DB_DNODE_ENTER(dbi
);
4135 return (DB_DNODE(dbi
));
4139 dmu_buf_dnode_exit(dmu_buf_t
*db
)
4141 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4146 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
4148 /* ASSERT(dmu_tx_is_syncing(tx) */
4149 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4151 if (db
->db_blkptr
!= NULL
)
4154 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4155 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
4156 BP_ZERO(db
->db_blkptr
);
4159 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
4161 * This buffer was allocated at a time when there was
4162 * no available blkptrs from the dnode, or it was
4163 * inappropriate to hook it in (i.e., nlevels mismatch).
4165 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
4166 ASSERT(db
->db_parent
== NULL
);
4167 db
->db_parent
= dn
->dn_dbuf
;
4168 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
4171 dmu_buf_impl_t
*parent
= db
->db_parent
;
4172 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4174 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
4175 if (parent
== NULL
) {
4176 mutex_exit(&db
->db_mtx
);
4177 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
4178 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
4179 db
->db_blkid
>> epbs
, db
);
4180 rw_exit(&dn
->dn_struct_rwlock
);
4181 mutex_enter(&db
->db_mtx
);
4182 db
->db_parent
= parent
;
4184 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
4185 (db
->db_blkid
& ((1ULL << epbs
) - 1));
4191 dbuf_sync_bonus(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4193 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4194 void *data
= dr
->dt
.dl
.dr_data
;
4196 ASSERT0(db
->db_level
);
4197 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4198 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
);
4199 ASSERT(data
!= NULL
);
4201 dnode_t
*dn
= dr
->dr_dnode
;
4202 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
4203 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
4204 memcpy(DN_BONUS(dn
->dn_phys
), data
, DN_MAX_BONUS_LEN(dn
->dn_phys
));
4206 dbuf_sync_leaf_verify_bonus_dnode(dr
);
4208 dbuf_undirty_bonus(dr
);
4209 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4213 * When syncing out a blocks of dnodes, adjust the block to deal with
4214 * encryption. Normally, we make sure the block is decrypted before writing
4215 * it. If we have crypt params, then we are writing a raw (encrypted) block,
4216 * from a raw receive. In this case, set the ARC buf's crypt params so
4217 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
4220 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t
*dr
)
4223 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4225 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4226 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
4227 ASSERT3U(db
->db_level
, ==, 0);
4229 if (!db
->db_objset
->os_raw_receive
&& arc_is_encrypted(db
->db_buf
)) {
4230 zbookmark_phys_t zb
;
4233 * Unfortunately, there is currently no mechanism for
4234 * syncing context to handle decryption errors. An error
4235 * here is only possible if an attacker maliciously
4236 * changed a dnode block and updated the associated
4237 * checksums going up the block tree.
4239 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
4240 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4241 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
4244 panic("Invalid dnode block MAC");
4245 } else if (dr
->dt
.dl
.dr_has_raw_params
) {
4246 (void) arc_release(dr
->dt
.dl
.dr_data
, db
);
4247 arc_convert_to_raw(dr
->dt
.dl
.dr_data
,
4248 dmu_objset_id(db
->db_objset
),
4249 dr
->dt
.dl
.dr_byteorder
, DMU_OT_DNODE
,
4250 dr
->dt
.dl
.dr_salt
, dr
->dt
.dl
.dr_iv
, dr
->dt
.dl
.dr_mac
);
4255 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
4256 * is critical the we not allow the compiler to inline this function in to
4257 * dbuf_sync_list() thereby drastically bloating the stack usage.
4259 noinline
static void
4260 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4262 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4263 dnode_t
*dn
= dr
->dr_dnode
;
4265 ASSERT(dmu_tx_is_syncing(tx
));
4267 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4269 mutex_enter(&db
->db_mtx
);
4271 ASSERT(db
->db_level
> 0);
4274 /* Read the block if it hasn't been read yet. */
4275 if (db
->db_buf
== NULL
) {
4276 mutex_exit(&db
->db_mtx
);
4277 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
4278 mutex_enter(&db
->db_mtx
);
4280 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
4281 ASSERT(db
->db_buf
!= NULL
);
4283 /* Indirect block size must match what the dnode thinks it is. */
4284 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4285 dbuf_check_blkptr(dn
, db
);
4287 /* Provide the pending dirty record to child dbufs */
4288 db
->db_data_pending
= dr
;
4290 mutex_exit(&db
->db_mtx
);
4292 dbuf_write(dr
, db
->db_buf
, tx
);
4294 zio_t
*zio
= dr
->dr_zio
;
4295 mutex_enter(&dr
->dt
.di
.dr_mtx
);
4296 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
4297 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4298 mutex_exit(&dr
->dt
.di
.dr_mtx
);
4303 * Verify that the size of the data in our bonus buffer does not exceed
4304 * its recorded size.
4306 * The purpose of this verification is to catch any cases in development
4307 * where the size of a phys structure (i.e space_map_phys_t) grows and,
4308 * due to incorrect feature management, older pools expect to read more
4309 * data even though they didn't actually write it to begin with.
4311 * For a example, this would catch an error in the feature logic where we
4312 * open an older pool and we expect to write the space map histogram of
4313 * a space map with size SPACE_MAP_SIZE_V0.
4316 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
)
4319 dnode_t
*dn
= dr
->dr_dnode
;
4322 * Encrypted bonus buffers can have data past their bonuslen.
4323 * Skip the verification of these blocks.
4325 if (DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))
4328 uint16_t bonuslen
= dn
->dn_phys
->dn_bonuslen
;
4329 uint16_t maxbonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
4330 ASSERT3U(bonuslen
, <=, maxbonuslen
);
4332 arc_buf_t
*datap
= dr
->dt
.dl
.dr_data
;
4333 char *datap_end
= ((char *)datap
) + bonuslen
;
4334 char *datap_max
= ((char *)datap
) + maxbonuslen
;
4336 /* ensure that everything is zero after our data */
4337 for (; datap_end
< datap_max
; datap_end
++)
4338 ASSERT(*datap_end
== 0);
4343 dbuf_lightweight_bp(dbuf_dirty_record_t
*dr
)
4345 /* This must be a lightweight dirty record. */
4346 ASSERT3P(dr
->dr_dbuf
, ==, NULL
);
4347 dnode_t
*dn
= dr
->dr_dnode
;
4349 if (dn
->dn_phys
->dn_nlevels
== 1) {
4350 VERIFY3U(dr
->dt
.dll
.dr_blkid
, <, dn
->dn_phys
->dn_nblkptr
);
4351 return (&dn
->dn_phys
->dn_blkptr
[dr
->dt
.dll
.dr_blkid
]);
4353 dmu_buf_impl_t
*parent_db
= dr
->dr_parent
->dr_dbuf
;
4354 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4355 VERIFY3U(parent_db
->db_level
, ==, 1);
4356 VERIFY3P(parent_db
->db_dnode_handle
->dnh_dnode
, ==, dn
);
4357 VERIFY3U(dr
->dt
.dll
.dr_blkid
>> epbs
, ==, parent_db
->db_blkid
);
4358 blkptr_t
*bp
= parent_db
->db
.db_data
;
4359 return (&bp
[dr
->dt
.dll
.dr_blkid
& ((1 << epbs
) - 1)]);
4364 dbuf_lightweight_ready(zio_t
*zio
)
4366 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4367 blkptr_t
*bp
= zio
->io_bp
;
4369 if (zio
->io_error
!= 0)
4372 dnode_t
*dn
= dr
->dr_dnode
;
4374 blkptr_t
*bp_orig
= dbuf_lightweight_bp(dr
);
4375 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4376 int64_t delta
= bp_get_dsize_sync(spa
, bp
) -
4377 bp_get_dsize_sync(spa
, bp_orig
);
4378 dnode_diduse_space(dn
, delta
);
4380 uint64_t blkid
= dr
->dt
.dll
.dr_blkid
;
4381 mutex_enter(&dn
->dn_mtx
);
4382 if (blkid
> dn
->dn_phys
->dn_maxblkid
) {
4383 ASSERT0(dn
->dn_objset
->os_raw_receive
);
4384 dn
->dn_phys
->dn_maxblkid
= blkid
;
4386 mutex_exit(&dn
->dn_mtx
);
4388 if (!BP_IS_EMBEDDED(bp
)) {
4389 uint64_t fill
= BP_IS_HOLE(bp
) ? 0 : 1;
4390 BP_SET_FILL(bp
, fill
);
4393 dmu_buf_impl_t
*parent_db
;
4394 EQUIV(dr
->dr_parent
== NULL
, dn
->dn_phys
->dn_nlevels
== 1);
4395 if (dr
->dr_parent
== NULL
) {
4396 parent_db
= dn
->dn_dbuf
;
4398 parent_db
= dr
->dr_parent
->dr_dbuf
;
4400 rw_enter(&parent_db
->db_rwlock
, RW_WRITER
);
4402 rw_exit(&parent_db
->db_rwlock
);
4406 dbuf_lightweight_done(zio_t
*zio
)
4408 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4410 VERIFY0(zio
->io_error
);
4412 objset_t
*os
= dr
->dr_dnode
->dn_objset
;
4413 dmu_tx_t
*tx
= os
->os_synctx
;
4415 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4416 ASSERT(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4418 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4419 (void) dsl_dataset_block_kill(ds
, &zio
->io_bp_orig
, tx
, B_TRUE
);
4420 dsl_dataset_block_born(ds
, zio
->io_bp
, tx
);
4423 dsl_pool_undirty_space(dmu_objset_pool(os
), dr
->dr_accounted
,
4426 abd_free(dr
->dt
.dll
.dr_abd
);
4427 kmem_free(dr
, sizeof (*dr
));
4430 noinline
static void
4431 dbuf_sync_lightweight(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4433 dnode_t
*dn
= dr
->dr_dnode
;
4435 if (dn
->dn_phys
->dn_nlevels
== 1) {
4438 pio
= dr
->dr_parent
->dr_zio
;
4441 zbookmark_phys_t zb
= {
4442 .zb_objset
= dmu_objset_id(dn
->dn_objset
),
4443 .zb_object
= dn
->dn_object
,
4445 .zb_blkid
= dr
->dt
.dll
.dr_blkid
,
4449 * See comment in dbuf_write(). This is so that zio->io_bp_orig
4450 * will have the old BP in dbuf_lightweight_done().
4452 dr
->dr_bp_copy
= *dbuf_lightweight_bp(dr
);
4454 dr
->dr_zio
= zio_write(pio
, dmu_objset_spa(dn
->dn_objset
),
4455 dmu_tx_get_txg(tx
), &dr
->dr_bp_copy
, dr
->dt
.dll
.dr_abd
,
4456 dn
->dn_datablksz
, abd_get_size(dr
->dt
.dll
.dr_abd
),
4457 &dr
->dt
.dll
.dr_props
, dbuf_lightweight_ready
, NULL
,
4458 dbuf_lightweight_done
, dr
, ZIO_PRIORITY_ASYNC_WRITE
,
4459 ZIO_FLAG_MUSTSUCCEED
| dr
->dt
.dll
.dr_flags
, &zb
);
4461 zio_nowait(dr
->dr_zio
);
4465 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
4466 * critical the we not allow the compiler to inline this function in to
4467 * dbuf_sync_list() thereby drastically bloating the stack usage.
4469 noinline
static void
4470 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4472 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
4473 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4474 dnode_t
*dn
= dr
->dr_dnode
;
4476 uint64_t txg
= tx
->tx_txg
;
4478 ASSERT(dmu_tx_is_syncing(tx
));
4480 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4482 mutex_enter(&db
->db_mtx
);
4484 * To be synced, we must be dirtied. But we
4485 * might have been freed after the dirty.
4487 if (db
->db_state
== DB_UNCACHED
) {
4488 /* This buffer has been freed since it was dirtied */
4489 ASSERT(db
->db
.db_data
== NULL
);
4490 } else if (db
->db_state
== DB_FILL
) {
4491 /* This buffer was freed and is now being re-filled */
4492 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
4493 } else if (db
->db_state
== DB_READ
) {
4495 * This buffer has a clone we need to write, and an in-flight
4496 * read on the BP we're about to clone. Its safe to issue the
4497 * write here because the read has already been issued and the
4498 * contents won't change.
4500 ASSERT(dr
->dt
.dl
.dr_brtwrite
&&
4501 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
);
4503 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
4507 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4508 mutex_enter(&dn
->dn_mtx
);
4509 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
4511 * In the previous transaction group, the bonus buffer
4512 * was entirely used to store the attributes for the
4513 * dnode which overrode the dn_spill field. However,
4514 * when adding more attributes to the file a spill
4515 * block was required to hold the extra attributes.
4517 * Make sure to clear the garbage left in the dn_spill
4518 * field from the previous attributes in the bonus
4519 * buffer. Otherwise, after writing out the spill
4520 * block to the new allocated dva, it will free
4521 * the old block pointed to by the invalid dn_spill.
4523 db
->db_blkptr
= NULL
;
4525 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
4526 mutex_exit(&dn
->dn_mtx
);
4530 * If this is a bonus buffer, simply copy the bonus data into the
4531 * dnode. It will be written out when the dnode is synced (and it
4532 * will be synced, since it must have been dirty for dbuf_sync to
4535 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
4536 ASSERT(dr
->dr_dbuf
== db
);
4537 dbuf_sync_bonus(dr
, tx
);
4544 * This function may have dropped the db_mtx lock allowing a dmu_sync
4545 * operation to sneak in. As a result, we need to ensure that we
4546 * don't check the dr_override_state until we have returned from
4547 * dbuf_check_blkptr.
4549 dbuf_check_blkptr(dn
, db
);
4552 * If this buffer is in the middle of an immediate write,
4553 * wait for the synchronous IO to complete.
4555 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
4556 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
4557 cv_wait(&db
->db_changed
, &db
->db_mtx
);
4561 * If this is a dnode block, ensure it is appropriately encrypted
4562 * or decrypted, depending on what we are writing to it this txg.
4564 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
4565 dbuf_prepare_encrypted_dnode_leaf(dr
);
4567 if (db
->db_state
!= DB_NOFILL
&&
4568 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
4569 zfs_refcount_count(&db
->db_holds
) > 1 &&
4570 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
4571 *datap
== db
->db_buf
) {
4573 * If this buffer is currently "in use" (i.e., there
4574 * are active holds and db_data still references it),
4575 * then make a copy before we start the write so that
4576 * any modifications from the open txg will not leak
4579 * NOTE: this copy does not need to be made for
4580 * objects only modified in the syncing context (e.g.
4581 * DNONE_DNODE blocks).
4583 int psize
= arc_buf_size(*datap
);
4584 int lsize
= arc_buf_lsize(*datap
);
4585 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
4586 enum zio_compress compress_type
= arc_get_compression(*datap
);
4587 uint8_t complevel
= arc_get_complevel(*datap
);
4589 if (arc_is_encrypted(*datap
)) {
4590 boolean_t byteorder
;
4591 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4592 uint8_t iv
[ZIO_DATA_IV_LEN
];
4593 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4595 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
4596 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
4597 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
4598 dn
->dn_type
, psize
, lsize
, compress_type
,
4600 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
4601 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
4602 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
4603 psize
, lsize
, compress_type
, complevel
);
4605 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
4607 memcpy((*datap
)->b_data
, db
->db
.db_data
, psize
);
4609 db
->db_data_pending
= dr
;
4611 mutex_exit(&db
->db_mtx
);
4613 dbuf_write(dr
, *datap
, tx
);
4615 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4616 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
4617 list_insert_tail(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
4619 zio_nowait(dr
->dr_zio
);
4624 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
4626 dbuf_dirty_record_t
*dr
;
4628 while ((dr
= list_head(list
))) {
4629 if (dr
->dr_zio
!= NULL
) {
4631 * If we find an already initialized zio then we
4632 * are processing the meta-dnode, and we have finished.
4633 * The dbufs for all dnodes are put back on the list
4634 * during processing, so that we can zio_wait()
4635 * these IOs after initiating all child IOs.
4637 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
4638 DMU_META_DNODE_OBJECT
);
4641 list_remove(list
, dr
);
4642 if (dr
->dr_dbuf
== NULL
) {
4643 dbuf_sync_lightweight(dr
, tx
);
4645 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
4646 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
4647 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
4649 if (dr
->dr_dbuf
->db_level
> 0)
4650 dbuf_sync_indirect(dr
, tx
);
4652 dbuf_sync_leaf(dr
, tx
);
4658 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4661 dmu_buf_impl_t
*db
= vdb
;
4663 blkptr_t
*bp
= zio
->io_bp
;
4664 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4665 spa_t
*spa
= zio
->io_spa
;
4670 ASSERT3P(db
->db_blkptr
, !=, NULL
);
4671 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
4675 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
4676 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
4677 zio
->io_prev_space_delta
= delta
;
4679 if (bp
->blk_birth
!= 0) {
4680 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
4681 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
4682 (db
->db_blkid
== DMU_SPILL_BLKID
&&
4683 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
4684 BP_IS_EMBEDDED(bp
));
4685 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
4688 mutex_enter(&db
->db_mtx
);
4691 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4692 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4693 ASSERT(!(BP_IS_HOLE(bp
)) &&
4694 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4698 if (db
->db_level
== 0) {
4699 mutex_enter(&dn
->dn_mtx
);
4700 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
4701 db
->db_blkid
!= DMU_SPILL_BLKID
) {
4702 ASSERT0(db
->db_objset
->os_raw_receive
);
4703 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
4705 mutex_exit(&dn
->dn_mtx
);
4707 if (dn
->dn_type
== DMU_OT_DNODE
) {
4709 while (i
< db
->db
.db_size
) {
4711 (void *)(((char *)db
->db
.db_data
) + i
);
4713 i
+= DNODE_MIN_SIZE
;
4714 if (dnp
->dn_type
!= DMU_OT_NONE
) {
4716 for (int j
= 0; j
< dnp
->dn_nblkptr
;
4718 (void) zfs_blkptr_verify(spa
,
4724 DNODE_FLAG_SPILL_BLKPTR
) {
4725 (void) zfs_blkptr_verify(spa
,
4726 DN_SPILL_BLKPTR(dnp
),
4730 i
+= dnp
->dn_extra_slots
*
4735 if (BP_IS_HOLE(bp
)) {
4742 blkptr_t
*ibp
= db
->db
.db_data
;
4743 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4744 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
4745 if (BP_IS_HOLE(ibp
))
4747 (void) zfs_blkptr_verify(spa
, ibp
,
4748 BLK_CONFIG_SKIP
, BLK_VERIFY_HALT
);
4749 fill
+= BP_GET_FILL(ibp
);
4754 if (!BP_IS_EMBEDDED(bp
))
4755 BP_SET_FILL(bp
, fill
);
4757 mutex_exit(&db
->db_mtx
);
4759 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_WRITER
, FTAG
);
4760 *db
->db_blkptr
= *bp
;
4761 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
4765 * This function gets called just prior to running through the compression
4766 * stage of the zio pipeline. If we're an indirect block comprised of only
4767 * holes, then we want this indirect to be compressed away to a hole. In
4768 * order to do that we must zero out any information about the holes that
4769 * this indirect points to prior to before we try to compress it.
4772 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4774 (void) zio
, (void) buf
;
4775 dmu_buf_impl_t
*db
= vdb
;
4778 unsigned int epbs
, i
;
4780 ASSERT3U(db
->db_level
, >, 0);
4783 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4784 ASSERT3U(epbs
, <, 31);
4786 /* Determine if all our children are holes */
4787 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
4788 if (!BP_IS_HOLE(bp
))
4793 * If all the children are holes, then zero them all out so that
4794 * we may get compressed away.
4796 if (i
== 1ULL << epbs
) {
4798 * We only found holes. Grab the rwlock to prevent
4799 * anybody from reading the blocks we're about to
4802 rw_enter(&db
->db_rwlock
, RW_WRITER
);
4803 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
4804 rw_exit(&db
->db_rwlock
);
4810 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4813 dmu_buf_impl_t
*db
= vdb
;
4814 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4815 blkptr_t
*bp
= db
->db_blkptr
;
4816 objset_t
*os
= db
->db_objset
;
4817 dmu_tx_t
*tx
= os
->os_synctx
;
4819 ASSERT0(zio
->io_error
);
4820 ASSERT(db
->db_blkptr
== bp
);
4823 * For nopwrites and rewrites we ensure that the bp matches our
4824 * original and bypass all the accounting.
4826 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4827 ASSERT(BP_EQUAL(bp
, bp_orig
));
4829 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4830 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
4831 dsl_dataset_block_born(ds
, bp
, tx
);
4834 mutex_enter(&db
->db_mtx
);
4838 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
4839 dnode_t
*dn
= dr
->dr_dnode
;
4840 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4841 ASSERT(dr
->dr_dbuf
== db
);
4842 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
4843 list_remove(&db
->db_dirty_records
, dr
);
4846 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4847 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4848 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
4849 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4853 if (db
->db_level
== 0) {
4854 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
4855 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
4856 if (db
->db_state
!= DB_NOFILL
) {
4857 if (dr
->dt
.dl
.dr_data
!= NULL
&&
4858 dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
4859 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
4863 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4864 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
4865 if (!BP_IS_HOLE(db
->db_blkptr
)) {
4866 int epbs __maybe_unused
= dn
->dn_phys
->dn_indblkshift
-
4868 ASSERT3U(db
->db_blkid
, <=,
4869 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
4870 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4873 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4874 list_destroy(&dr
->dt
.di
.dr_children
);
4877 cv_broadcast(&db
->db_changed
);
4878 ASSERT(db
->db_dirtycnt
> 0);
4879 db
->db_dirtycnt
-= 1;
4880 db
->db_data_pending
= NULL
;
4881 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4883 dsl_pool_undirty_space(dmu_objset_pool(os
), dr
->dr_accounted
,
4886 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4890 dbuf_write_nofill_ready(zio_t
*zio
)
4892 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4896 dbuf_write_nofill_done(zio_t
*zio
)
4898 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4902 dbuf_write_override_ready(zio_t
*zio
)
4904 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4905 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4907 dbuf_write_ready(zio
, NULL
, db
);
4911 dbuf_write_override_done(zio_t
*zio
)
4913 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4914 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4915 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4917 mutex_enter(&db
->db_mtx
);
4918 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4919 if (!BP_IS_HOLE(obp
))
4920 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4921 arc_release(dr
->dt
.dl
.dr_data
, db
);
4923 mutex_exit(&db
->db_mtx
);
4925 dbuf_write_done(zio
, NULL
, db
);
4927 if (zio
->io_abd
!= NULL
)
4928 abd_free(zio
->io_abd
);
4931 typedef struct dbuf_remap_impl_callback_arg
{
4933 uint64_t drica_blk_birth
;
4935 } dbuf_remap_impl_callback_arg_t
;
4938 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
4941 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
4942 objset_t
*os
= drica
->drica_os
;
4943 spa_t
*spa
= dmu_objset_spa(os
);
4944 dmu_tx_t
*tx
= drica
->drica_tx
;
4946 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4948 if (os
== spa_meta_objset(spa
)) {
4949 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
4951 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
4952 size
, drica
->drica_blk_birth
, tx
);
4957 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, krwlock_t
*rw
, dmu_tx_t
*tx
)
4959 blkptr_t bp_copy
= *bp
;
4960 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4961 dbuf_remap_impl_callback_arg_t drica
;
4963 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4965 drica
.drica_os
= dn
->dn_objset
;
4966 drica
.drica_blk_birth
= bp
->blk_birth
;
4967 drica
.drica_tx
= tx
;
4968 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
4971 * If the blkptr being remapped is tracked by a livelist,
4972 * then we need to make sure the livelist reflects the update.
4973 * First, cancel out the old blkptr by appending a 'FREE'
4974 * entry. Next, add an 'ALLOC' to track the new version. This
4975 * way we avoid trying to free an inaccurate blkptr at delete.
4976 * Note that embedded blkptrs are not tracked in livelists.
4978 if (dn
->dn_objset
!= spa_meta_objset(spa
)) {
4979 dsl_dataset_t
*ds
= dmu_objset_ds(dn
->dn_objset
);
4980 if (dsl_deadlist_is_open(&ds
->ds_dir
->dd_livelist
) &&
4981 bp
->blk_birth
> ds
->ds_dir
->dd_origin_txg
) {
4982 ASSERT(!BP_IS_EMBEDDED(bp
));
4983 ASSERT(dsl_dir_is_clone(ds
->ds_dir
));
4984 ASSERT(spa_feature_is_enabled(spa
,
4985 SPA_FEATURE_LIVELIST
));
4986 bplist_append(&ds
->ds_dir
->dd_pending_frees
,
4988 bplist_append(&ds
->ds_dir
->dd_pending_allocs
,
4994 * The db_rwlock prevents dbuf_read_impl() from
4995 * dereferencing the BP while we are changing it. To
4996 * avoid lock contention, only grab it when we are actually
5000 rw_enter(rw
, RW_WRITER
);
5008 * Remap any existing BP's to concrete vdevs, if possible.
5011 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
5013 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
5014 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
5016 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
5019 if (db
->db_level
> 0) {
5020 blkptr_t
*bp
= db
->db
.db_data
;
5021 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
5022 dbuf_remap_impl(dn
, &bp
[i
], &db
->db_rwlock
, tx
);
5024 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
5025 dnode_phys_t
*dnp
= db
->db
.db_data
;
5026 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
5028 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
;
5029 i
+= dnp
[i
].dn_extra_slots
+ 1) {
5030 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
5031 krwlock_t
*lock
= (dn
->dn_dbuf
== NULL
? NULL
:
5032 &dn
->dn_dbuf
->db_rwlock
);
5033 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], lock
,
5041 /* Issue I/O to commit a dirty buffer to disk. */
5043 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
5045 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
5046 dnode_t
*dn
= dr
->dr_dnode
;
5048 dmu_buf_impl_t
*parent
= db
->db_parent
;
5049 uint64_t txg
= tx
->tx_txg
;
5050 zbookmark_phys_t zb
;
5052 zio_t
*pio
; /* parent I/O */
5055 ASSERT(dmu_tx_is_syncing(tx
));
5059 if (db
->db_state
!= DB_NOFILL
) {
5060 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
5062 * Private object buffers are released here rather
5063 * than in dbuf_dirty() since they are only modified
5064 * in the syncing context and we don't want the
5065 * overhead of making multiple copies of the data.
5067 if (BP_IS_HOLE(db
->db_blkptr
)) {
5070 dbuf_release_bp(db
);
5072 dbuf_remap(dn
, db
, tx
);
5076 if (parent
!= dn
->dn_dbuf
) {
5077 /* Our parent is an indirect block. */
5078 /* We have a dirty parent that has been scheduled for write. */
5079 ASSERT(parent
&& parent
->db_data_pending
);
5080 /* Our parent's buffer is one level closer to the dnode. */
5081 ASSERT(db
->db_level
== parent
->db_level
-1);
5083 * We're about to modify our parent's db_data by modifying
5084 * our block pointer, so the parent must be released.
5086 ASSERT(arc_released(parent
->db_buf
));
5087 pio
= parent
->db_data_pending
->dr_zio
;
5089 /* Our parent is the dnode itself. */
5090 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
5091 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
5092 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
5093 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
5094 ASSERT3P(db
->db_blkptr
, ==,
5095 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
5099 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
5100 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
5103 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
5104 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
5105 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
5107 if (db
->db_blkid
== DMU_SPILL_BLKID
)
5109 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
5111 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
5114 * We copy the blkptr now (rather than when we instantiate the dirty
5115 * record), because its value can change between open context and
5116 * syncing context. We do not need to hold dn_struct_rwlock to read
5117 * db_blkptr because we are in syncing context.
5119 dr
->dr_bp_copy
= *db
->db_blkptr
;
5121 if (db
->db_level
== 0 &&
5122 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
5124 * The BP for this block has been provided by open context
5125 * (by dmu_sync() or dmu_buf_write_embedded()).
5127 abd_t
*contents
= (data
!= NULL
) ?
5128 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
5130 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
, &dr
->dr_bp_copy
,
5131 contents
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5132 dbuf_write_override_ready
, NULL
,
5133 dbuf_write_override_done
,
5134 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
5135 mutex_enter(&db
->db_mtx
);
5136 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
5137 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
5138 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
,
5139 dr
->dt
.dl
.dr_brtwrite
);
5140 mutex_exit(&db
->db_mtx
);
5141 } else if (db
->db_state
== DB_NOFILL
) {
5142 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
5143 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
5144 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
,
5145 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5146 dbuf_write_nofill_ready
, NULL
,
5147 dbuf_write_nofill_done
, db
,
5148 ZIO_PRIORITY_ASYNC_WRITE
,
5149 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
5151 ASSERT(arc_released(data
));
5154 * For indirect blocks, we want to setup the children
5155 * ready callback so that we can properly handle an indirect
5156 * block that only contains holes.
5158 arc_write_done_func_t
*children_ready_cb
= NULL
;
5159 if (db
->db_level
!= 0)
5160 children_ready_cb
= dbuf_write_children_ready
;
5162 dr
->dr_zio
= arc_write(pio
, os
->os_spa
, txg
,
5163 &dr
->dr_bp_copy
, data
, !DBUF_IS_CACHEABLE(db
),
5164 dbuf_is_l2cacheable(db
), &zp
, dbuf_write_ready
,
5165 children_ready_cb
, dbuf_write_done
, db
,
5166 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
5170 EXPORT_SYMBOL(dbuf_find
);
5171 EXPORT_SYMBOL(dbuf_is_metadata
);
5172 EXPORT_SYMBOL(dbuf_destroy
);
5173 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
5174 EXPORT_SYMBOL(dbuf_whichblock
);
5175 EXPORT_SYMBOL(dbuf_read
);
5176 EXPORT_SYMBOL(dbuf_unoverride
);
5177 EXPORT_SYMBOL(dbuf_free_range
);
5178 EXPORT_SYMBOL(dbuf_new_size
);
5179 EXPORT_SYMBOL(dbuf_release_bp
);
5180 EXPORT_SYMBOL(dbuf_dirty
);
5181 EXPORT_SYMBOL(dmu_buf_set_crypt_params
);
5182 EXPORT_SYMBOL(dmu_buf_will_dirty
);
5183 EXPORT_SYMBOL(dmu_buf_is_dirty
);
5184 EXPORT_SYMBOL(dmu_buf_will_clone
);
5185 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
5186 EXPORT_SYMBOL(dmu_buf_will_fill
);
5187 EXPORT_SYMBOL(dmu_buf_fill_done
);
5188 EXPORT_SYMBOL(dmu_buf_rele
);
5189 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
5190 EXPORT_SYMBOL(dbuf_prefetch
);
5191 EXPORT_SYMBOL(dbuf_hold_impl
);
5192 EXPORT_SYMBOL(dbuf_hold
);
5193 EXPORT_SYMBOL(dbuf_hold_level
);
5194 EXPORT_SYMBOL(dbuf_create_bonus
);
5195 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
5196 EXPORT_SYMBOL(dbuf_rm_spill
);
5197 EXPORT_SYMBOL(dbuf_add_ref
);
5198 EXPORT_SYMBOL(dbuf_rele
);
5199 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
5200 EXPORT_SYMBOL(dbuf_refcount
);
5201 EXPORT_SYMBOL(dbuf_sync_list
);
5202 EXPORT_SYMBOL(dmu_buf_set_user
);
5203 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
5204 EXPORT_SYMBOL(dmu_buf_get_user
);
5205 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
5207 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, max_bytes
, U64
, ZMOD_RW
,
5208 "Maximum size in bytes of the dbuf cache.");
5210 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, hiwater_pct
, UINT
, ZMOD_RW
,
5211 "Percentage over dbuf_cache_max_bytes for direct dbuf eviction.");
5213 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, lowater_pct
, UINT
, ZMOD_RW
,
5214 "Percentage below dbuf_cache_max_bytes when dbuf eviction stops.");
5216 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_max_bytes
, U64
, ZMOD_RW
,
5217 "Maximum size in bytes of dbuf metadata cache.");
5219 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, cache_shift
, UINT
, ZMOD_RW
,
5220 "Set size of dbuf cache to log2 fraction of arc size.");
5222 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_shift
, UINT
, ZMOD_RW
,
5223 "Set size of dbuf metadata cache to log2 fraction of arc size.");
5225 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, mutex_cache_shift
, UINT
, ZMOD_RD
,
5226 "Set size of dbuf cache mutex array as log2 shift.");