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
31 #include <sys/zfs_context.h>
34 #include <sys/dmu_send.h>
35 #include <sys/dmu_impl.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dmu_tx.h>
43 #include <sys/dmu_zfetch.h>
45 #include <sys/sa_impl.h>
46 #include <sys/zfeature.h>
47 #include <sys/blkptr.h>
48 #include <sys/range_tree.h>
49 #include <sys/trace_zfs.h>
50 #include <sys/callb.h>
54 #include <sys/spa_impl.h>
55 #include <sys/wmsum.h>
56 #include <sys/vdev_impl.h>
58 static kstat_t
*dbuf_ksp
;
60 typedef struct dbuf_stats
{
62 * Various statistics about the size of the dbuf cache.
64 kstat_named_t cache_count
;
65 kstat_named_t cache_size_bytes
;
66 kstat_named_t cache_size_bytes_max
;
68 * Statistics regarding the bounds on the dbuf cache size.
70 kstat_named_t cache_target_bytes
;
71 kstat_named_t cache_lowater_bytes
;
72 kstat_named_t cache_hiwater_bytes
;
74 * Total number of dbuf cache evictions that have occurred.
76 kstat_named_t cache_total_evicts
;
78 * The distribution of dbuf levels in the dbuf cache and
79 * the total size of all dbufs at each level.
81 kstat_named_t cache_levels
[DN_MAX_LEVELS
];
82 kstat_named_t cache_levels_bytes
[DN_MAX_LEVELS
];
84 * Statistics about the dbuf hash table.
86 kstat_named_t hash_hits
;
87 kstat_named_t hash_misses
;
88 kstat_named_t hash_collisions
;
89 kstat_named_t hash_elements
;
90 kstat_named_t hash_elements_max
;
92 * Number of sublists containing more than one dbuf in the dbuf
93 * hash table. Keep track of the longest hash chain.
95 kstat_named_t hash_chains
;
96 kstat_named_t hash_chain_max
;
98 * Number of times a dbuf_create() discovers that a dbuf was
99 * already created and in the dbuf hash table.
101 kstat_named_t hash_insert_race
;
103 * Number of entries in the hash table dbuf and mutex arrays.
105 kstat_named_t hash_table_count
;
106 kstat_named_t hash_mutex_count
;
108 * Statistics about the size of the metadata dbuf cache.
110 kstat_named_t metadata_cache_count
;
111 kstat_named_t metadata_cache_size_bytes
;
112 kstat_named_t metadata_cache_size_bytes_max
;
114 * For diagnostic purposes, this is incremented whenever we can't add
115 * something to the metadata cache because it's full, and instead put
116 * the data in the regular dbuf cache.
118 kstat_named_t metadata_cache_overflow
;
121 dbuf_stats_t dbuf_stats
= {
122 { "cache_count", KSTAT_DATA_UINT64
},
123 { "cache_size_bytes", KSTAT_DATA_UINT64
},
124 { "cache_size_bytes_max", KSTAT_DATA_UINT64
},
125 { "cache_target_bytes", KSTAT_DATA_UINT64
},
126 { "cache_lowater_bytes", KSTAT_DATA_UINT64
},
127 { "cache_hiwater_bytes", KSTAT_DATA_UINT64
},
128 { "cache_total_evicts", KSTAT_DATA_UINT64
},
129 { { "cache_levels_N", KSTAT_DATA_UINT64
} },
130 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64
} },
131 { "hash_hits", KSTAT_DATA_UINT64
},
132 { "hash_misses", KSTAT_DATA_UINT64
},
133 { "hash_collisions", KSTAT_DATA_UINT64
},
134 { "hash_elements", KSTAT_DATA_UINT64
},
135 { "hash_elements_max", KSTAT_DATA_UINT64
},
136 { "hash_chains", KSTAT_DATA_UINT64
},
137 { "hash_chain_max", KSTAT_DATA_UINT64
},
138 { "hash_insert_race", KSTAT_DATA_UINT64
},
139 { "hash_table_count", KSTAT_DATA_UINT64
},
140 { "hash_mutex_count", KSTAT_DATA_UINT64
},
141 { "metadata_cache_count", KSTAT_DATA_UINT64
},
142 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64
},
143 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64
},
144 { "metadata_cache_overflow", KSTAT_DATA_UINT64
}
149 wmsum_t cache_total_evicts
;
150 wmsum_t cache_levels
[DN_MAX_LEVELS
];
151 wmsum_t cache_levels_bytes
[DN_MAX_LEVELS
];
154 wmsum_t hash_collisions
;
156 wmsum_t hash_insert_race
;
157 wmsum_t metadata_cache_count
;
158 wmsum_t metadata_cache_overflow
;
161 #define DBUF_STAT_INCR(stat, val) \
162 wmsum_add(&dbuf_sums.stat, val);
163 #define DBUF_STAT_DECR(stat, val) \
164 DBUF_STAT_INCR(stat, -(val));
165 #define DBUF_STAT_BUMP(stat) \
166 DBUF_STAT_INCR(stat, 1);
167 #define DBUF_STAT_BUMPDOWN(stat) \
168 DBUF_STAT_INCR(stat, -1);
169 #define DBUF_STAT_MAX(stat, v) { \
171 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
172 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
176 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
177 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
178 static void dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
);
179 static int dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
);
182 * Global data structures and functions for the dbuf cache.
184 static kmem_cache_t
*dbuf_kmem_cache
;
185 static taskq_t
*dbu_evict_taskq
;
187 static kthread_t
*dbuf_cache_evict_thread
;
188 static kmutex_t dbuf_evict_lock
;
189 static kcondvar_t dbuf_evict_cv
;
190 static boolean_t dbuf_evict_thread_exit
;
193 * There are two dbuf caches; each dbuf can only be in one of them at a time.
195 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
196 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
197 * that represent the metadata that describes filesystems/snapshots/
198 * bookmarks/properties/etc. We only evict from this cache when we export a
199 * pool, to short-circuit as much I/O as possible for all administrative
200 * commands that need the metadata. There is no eviction policy for this
201 * cache, because we try to only include types in it which would occupy a
202 * very small amount of space per object but create a large impact on the
203 * performance of these commands. Instead, after it reaches a maximum size
204 * (which should only happen on very small memory systems with a very large
205 * number of filesystem objects), we stop taking new dbufs into the
206 * metadata cache, instead putting them in the normal dbuf cache.
208 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
209 * are not currently held but have been recently released. These dbufs
210 * are not eligible for arc eviction until they are aged out of the cache.
211 * Dbufs that are aged out of the cache will be immediately destroyed and
212 * become eligible for arc eviction.
214 * Dbufs are added to these caches once the last hold is released. If a dbuf is
215 * later accessed and still exists in the dbuf cache, then it will be removed
216 * from the cache and later re-added to the head of the cache.
218 * If a given dbuf meets the requirements for the metadata cache, it will go
219 * there, otherwise it will be considered for the generic LRU dbuf cache. The
220 * caches and the refcounts tracking their sizes are stored in an array indexed
221 * by those caches' matching enum values (from dbuf_cached_state_t).
223 typedef struct dbuf_cache
{
225 zfs_refcount_t size ____cacheline_aligned
;
227 dbuf_cache_t dbuf_caches
[DB_CACHE_MAX
];
229 /* Size limits for the caches */
230 static unsigned long dbuf_cache_max_bytes
= ULONG_MAX
;
231 static unsigned long dbuf_metadata_cache_max_bytes
= ULONG_MAX
;
233 /* Set the default sizes of the caches to log2 fraction of arc size */
234 static uint_t dbuf_cache_shift
= 5;
235 static uint_t dbuf_metadata_cache_shift
= 6;
237 /* Set the dbuf hash mutex count as log2 shift (dynamic by default) */
238 static uint_t dbuf_mutex_cache_shift
= 0;
240 static unsigned long dbuf_cache_target_bytes(void);
241 static unsigned long dbuf_metadata_cache_target_bytes(void);
244 * The LRU dbuf cache uses a three-stage eviction policy:
245 * - A low water marker designates when the dbuf eviction thread
246 * should stop evicting from the dbuf cache.
247 * - When we reach the maximum size (aka mid water mark), we
248 * signal the eviction thread to run.
249 * - The high water mark indicates when the eviction thread
250 * is unable to keep up with the incoming load and eviction must
251 * happen in the context of the calling thread.
255 * low water mid water hi water
256 * +----------------------------------------+----------+----------+
261 * +----------------------------------------+----------+----------+
263 * evicting eviction directly
266 * The high and low water marks indicate the operating range for the eviction
267 * thread. The low water mark is, by default, 90% of the total size of the
268 * cache and the high water mark is at 110% (both of these percentages can be
269 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
270 * respectively). The eviction thread will try to ensure that the cache remains
271 * within this range by waking up every second and checking if the cache is
272 * above the low water mark. The thread can also be woken up by callers adding
273 * elements into the cache if the cache is larger than the mid water (i.e max
274 * cache size). Once the eviction thread is woken up and eviction is required,
275 * it will continue evicting buffers until it's able to reduce the cache size
276 * to the low water mark. If the cache size continues to grow and hits the high
277 * water mark, then callers adding elements to the cache will begin to evict
278 * directly from the cache until the cache is no longer above the high water
283 * The percentage above and below the maximum cache size.
285 static uint_t dbuf_cache_hiwater_pct
= 10;
286 static uint_t dbuf_cache_lowater_pct
= 10;
289 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
291 (void) unused
, (void) kmflag
;
292 dmu_buf_impl_t
*db
= vdb
;
293 memset(db
, 0, sizeof (dmu_buf_impl_t
));
295 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
296 rw_init(&db
->db_rwlock
, NULL
, RW_DEFAULT
, NULL
);
297 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
298 multilist_link_init(&db
->db_cache_link
);
299 zfs_refcount_create(&db
->db_holds
);
305 dbuf_dest(void *vdb
, void *unused
)
308 dmu_buf_impl_t
*db
= vdb
;
309 mutex_destroy(&db
->db_mtx
);
310 rw_destroy(&db
->db_rwlock
);
311 cv_destroy(&db
->db_changed
);
312 ASSERT(!multilist_link_active(&db
->db_cache_link
));
313 zfs_refcount_destroy(&db
->db_holds
);
317 * dbuf hash table routines
319 static dbuf_hash_table_t dbuf_hash_table
;
322 * We use Cityhash for this. It's fast, and has good hash properties without
323 * requiring any large static buffers.
326 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
328 return (cityhash4((uintptr_t)os
, obj
, (uint64_t)lvl
, blkid
));
331 #define DTRACE_SET_STATE(db, why) \
332 DTRACE_PROBE2(dbuf__state_change, dmu_buf_impl_t *, db, \
335 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
336 ((dbuf)->db.db_object == (obj) && \
337 (dbuf)->db_objset == (os) && \
338 (dbuf)->db_level == (level) && \
339 (dbuf)->db_blkid == (blkid))
342 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
344 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
349 hv
= dbuf_hash(os
, obj
, level
, blkid
);
350 idx
= hv
& h
->hash_table_mask
;
352 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
353 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
354 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
355 mutex_enter(&db
->db_mtx
);
356 if (db
->db_state
!= DB_EVICTING
) {
357 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
360 mutex_exit(&db
->db_mtx
);
363 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
367 static dmu_buf_impl_t
*
368 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
371 dmu_buf_impl_t
*db
= NULL
;
373 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
374 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
375 if (dn
->dn_bonus
!= NULL
) {
377 mutex_enter(&db
->db_mtx
);
379 rw_exit(&dn
->dn_struct_rwlock
);
380 dnode_rele(dn
, FTAG
);
386 * Insert an entry into the hash table. If there is already an element
387 * equal to elem in the hash table, then the already existing element
388 * will be returned and the new element will not be inserted.
389 * Otherwise returns NULL.
391 static dmu_buf_impl_t
*
392 dbuf_hash_insert(dmu_buf_impl_t
*db
)
394 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
395 objset_t
*os
= db
->db_objset
;
396 uint64_t obj
= db
->db
.db_object
;
397 int level
= db
->db_level
;
398 uint64_t blkid
, hv
, idx
;
402 blkid
= db
->db_blkid
;
403 hv
= dbuf_hash(os
, obj
, level
, blkid
);
404 idx
= hv
& h
->hash_table_mask
;
406 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
407 for (dbf
= h
->hash_table
[idx
], i
= 0; dbf
!= NULL
;
408 dbf
= dbf
->db_hash_next
, i
++) {
409 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
410 mutex_enter(&dbf
->db_mtx
);
411 if (dbf
->db_state
!= DB_EVICTING
) {
412 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
415 mutex_exit(&dbf
->db_mtx
);
420 DBUF_STAT_BUMP(hash_collisions
);
422 DBUF_STAT_BUMP(hash_chains
);
424 DBUF_STAT_MAX(hash_chain_max
, i
);
427 mutex_enter(&db
->db_mtx
);
428 db
->db_hash_next
= h
->hash_table
[idx
];
429 h
->hash_table
[idx
] = db
;
430 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
431 uint64_t he
= atomic_inc_64_nv(&dbuf_stats
.hash_elements
.value
.ui64
);
432 DBUF_STAT_MAX(hash_elements_max
, he
);
438 * This returns whether this dbuf should be stored in the metadata cache, which
439 * is based on whether it's from one of the dnode types that store data related
440 * to traversing dataset hierarchies.
443 dbuf_include_in_metadata_cache(dmu_buf_impl_t
*db
)
446 dmu_object_type_t type
= DB_DNODE(db
)->dn_type
;
449 /* Check if this dbuf is one of the types we care about */
450 if (DMU_OT_IS_METADATA_CACHED(type
)) {
451 /* If we hit this, then we set something up wrong in dmu_ot */
452 ASSERT(DMU_OT_IS_METADATA(type
));
455 * Sanity check for small-memory systems: don't allocate too
456 * much memory for this purpose.
458 if (zfs_refcount_count(
459 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
) >
460 dbuf_metadata_cache_target_bytes()) {
461 DBUF_STAT_BUMP(metadata_cache_overflow
);
472 * Remove an entry from the hash table. It must be in the EVICTING state.
475 dbuf_hash_remove(dmu_buf_impl_t
*db
)
477 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
479 dmu_buf_impl_t
*dbf
, **dbp
;
481 hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
482 db
->db_level
, db
->db_blkid
);
483 idx
= hv
& h
->hash_table_mask
;
486 * We mustn't hold db_mtx to maintain lock ordering:
487 * DBUF_HASH_MUTEX > db_mtx.
489 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
490 ASSERT(db
->db_state
== DB_EVICTING
);
491 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
493 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
494 dbp
= &h
->hash_table
[idx
];
495 while ((dbf
= *dbp
) != db
) {
496 dbp
= &dbf
->db_hash_next
;
499 *dbp
= db
->db_hash_next
;
500 db
->db_hash_next
= NULL
;
501 if (h
->hash_table
[idx
] &&
502 h
->hash_table
[idx
]->db_hash_next
== NULL
)
503 DBUF_STAT_BUMPDOWN(hash_chains
);
504 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
505 atomic_dec_64(&dbuf_stats
.hash_elements
.value
.ui64
);
511 } dbvu_verify_type_t
;
514 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
519 if (db
->db_user
== NULL
)
522 /* Only data blocks support the attachment of user data. */
523 ASSERT(db
->db_level
== 0);
525 /* Clients must resolve a dbuf before attaching user data. */
526 ASSERT(db
->db
.db_data
!= NULL
);
527 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
529 holds
= zfs_refcount_count(&db
->db_holds
);
530 if (verify_type
== DBVU_EVICTING
) {
532 * Immediate eviction occurs when holds == dirtycnt.
533 * For normal eviction buffers, holds is zero on
534 * eviction, except when dbuf_fix_old_data() calls
535 * dbuf_clear_data(). However, the hold count can grow
536 * during eviction even though db_mtx is held (see
537 * dmu_bonus_hold() for an example), so we can only
538 * test the generic invariant that holds >= dirtycnt.
540 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
542 if (db
->db_user_immediate_evict
== TRUE
)
543 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
545 ASSERT3U(holds
, >, 0);
551 dbuf_evict_user(dmu_buf_impl_t
*db
)
553 dmu_buf_user_t
*dbu
= db
->db_user
;
555 ASSERT(MUTEX_HELD(&db
->db_mtx
));
560 dbuf_verify_user(db
, DBVU_EVICTING
);
564 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
565 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
569 * There are two eviction callbacks - one that we call synchronously
570 * and one that we invoke via a taskq. The async one is useful for
571 * avoiding lock order reversals and limiting stack depth.
573 * Note that if we have a sync callback but no async callback,
574 * it's likely that the sync callback will free the structure
575 * containing the dbu. In that case we need to take care to not
576 * dereference dbu after calling the sync evict func.
578 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
580 if (dbu
->dbu_evict_func_sync
!= NULL
)
581 dbu
->dbu_evict_func_sync(dbu
);
584 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
585 dbu
, 0, &dbu
->dbu_tqent
);
590 dbuf_is_metadata(dmu_buf_impl_t
*db
)
593 * Consider indirect blocks and spill blocks to be meta data.
595 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
598 boolean_t is_metadata
;
601 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
604 return (is_metadata
);
609 * We want to exclude buffers that are on a special allocation class from
613 dbuf_is_l2cacheable(dmu_buf_impl_t
*db
)
616 zfs_cache_type_t cache
= db
->db_objset
->os_secondary_cache
;
617 blkptr_t
*bp
= db
->db_blkptr
;
619 if (bp
!= NULL
&& !BP_IS_HOLE(bp
)) {
620 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
621 vdev_t
*rvd
= db
->db_objset
->os_spa
->spa_root_vdev
;
623 if (vdev
< rvd
->vdev_children
)
624 vd
= rvd
->vdev_child
[vdev
];
626 if (cache
== ZFS_CACHE_ALL
||
627 (dbuf_is_metadata(db
) && cache
== ZFS_CACHE_METADATA
)) {
631 if ((vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
632 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
) ||
633 l2arc_exclude_special
== 0)
641 static inline boolean_t
642 dnode_level_is_l2cacheable(blkptr_t
*bp
, dnode_t
*dn
, int64_t level
)
645 zfs_cache_type_t cache
= dn
->dn_objset
->os_secondary_cache
;
647 if (bp
!= NULL
&& !BP_IS_HOLE(bp
)) {
648 uint64_t vdev
= DVA_GET_VDEV(bp
->blk_dva
);
649 vdev_t
*rvd
= dn
->dn_objset
->os_spa
->spa_root_vdev
;
651 if (vdev
< rvd
->vdev_children
)
652 vd
= rvd
->vdev_child
[vdev
];
654 if (cache
== ZFS_CACHE_ALL
|| ((level
> 0 ||
655 DMU_OT_IS_METADATA(dn
->dn_handle
->dnh_dnode
->dn_type
)) &&
656 cache
== ZFS_CACHE_METADATA
)) {
660 if ((vd
->vdev_alloc_bias
!= VDEV_BIAS_SPECIAL
&&
661 vd
->vdev_alloc_bias
!= VDEV_BIAS_DEDUP
) ||
662 l2arc_exclude_special
== 0)
672 * This function *must* return indices evenly distributed between all
673 * sublists of the multilist. This is needed due to how the dbuf eviction
674 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
675 * distributed between all sublists and uses this assumption when
676 * deciding which sublist to evict from and how much to evict from it.
679 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
681 dmu_buf_impl_t
*db
= obj
;
684 * The assumption here, is the hash value for a given
685 * dmu_buf_impl_t will remain constant throughout it's lifetime
686 * (i.e. it's objset, object, level and blkid fields don't change).
687 * Thus, we don't need to store the dbuf's sublist index
688 * on insertion, as this index can be recalculated on removal.
690 * Also, the low order bits of the hash value are thought to be
691 * distributed evenly. Otherwise, in the case that the multilist
692 * has a power of two number of sublists, each sublists' usage
693 * would not be evenly distributed. In this context full 64bit
694 * division would be a waste of time, so limit it to 32 bits.
696 return ((unsigned int)dbuf_hash(db
->db_objset
, db
->db
.db_object
,
697 db
->db_level
, db
->db_blkid
) %
698 multilist_get_num_sublists(ml
));
702 * The target size of the dbuf cache can grow with the ARC target,
703 * unless limited by the tunable dbuf_cache_max_bytes.
705 static inline unsigned long
706 dbuf_cache_target_bytes(void)
708 return (MIN(dbuf_cache_max_bytes
,
709 arc_target_bytes() >> dbuf_cache_shift
));
713 * The target size of the dbuf metadata cache can grow with the ARC target,
714 * unless limited by the tunable dbuf_metadata_cache_max_bytes.
716 static inline unsigned long
717 dbuf_metadata_cache_target_bytes(void)
719 return (MIN(dbuf_metadata_cache_max_bytes
,
720 arc_target_bytes() >> dbuf_metadata_cache_shift
));
723 static inline uint64_t
724 dbuf_cache_hiwater_bytes(void)
726 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
727 return (dbuf_cache_target
+
728 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100);
731 static inline uint64_t
732 dbuf_cache_lowater_bytes(void)
734 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
735 return (dbuf_cache_target
-
736 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100);
739 static inline boolean_t
740 dbuf_cache_above_lowater(void)
742 return (zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
743 dbuf_cache_lowater_bytes());
747 * Evict the oldest eligible dbuf from the dbuf cache.
752 int idx
= multilist_get_random_index(&dbuf_caches
[DB_DBUF_CACHE
].cache
);
753 multilist_sublist_t
*mls
= multilist_sublist_lock(
754 &dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
756 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
758 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
759 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
760 db
= multilist_sublist_prev(mls
, db
);
763 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
764 multilist_sublist_t
*, mls
);
767 multilist_sublist_remove(mls
, db
);
768 multilist_sublist_unlock(mls
);
769 (void) zfs_refcount_remove_many(
770 &dbuf_caches
[DB_DBUF_CACHE
].size
, db
->db
.db_size
, db
);
771 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
772 DBUF_STAT_BUMPDOWN(cache_count
);
773 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
775 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
776 db
->db_caching_status
= DB_NO_CACHE
;
778 DBUF_STAT_BUMP(cache_total_evicts
);
780 multilist_sublist_unlock(mls
);
785 * The dbuf evict thread is responsible for aging out dbufs from the
786 * cache. Once the cache has reached it's maximum size, dbufs are removed
787 * and destroyed. The eviction thread will continue running until the size
788 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
789 * out of the cache it is destroyed and becomes eligible for arc eviction.
791 static __attribute__((noreturn
)) void
792 dbuf_evict_thread(void *unused
)
797 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
799 mutex_enter(&dbuf_evict_lock
);
800 while (!dbuf_evict_thread_exit
) {
801 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
802 CALLB_CPR_SAFE_BEGIN(&cpr
);
803 (void) cv_timedwait_idle_hires(&dbuf_evict_cv
,
804 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
805 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
807 mutex_exit(&dbuf_evict_lock
);
810 * Keep evicting as long as we're above the low water mark
811 * for the cache. We do this without holding the locks to
812 * minimize lock contention.
814 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
818 mutex_enter(&dbuf_evict_lock
);
821 dbuf_evict_thread_exit
= B_FALSE
;
822 cv_broadcast(&dbuf_evict_cv
);
823 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
828 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
829 * If the dbuf cache is at its high water mark, then evict a dbuf from the
830 * dbuf cache using the caller's context.
833 dbuf_evict_notify(uint64_t size
)
836 * We check if we should evict without holding the dbuf_evict_lock,
837 * because it's OK to occasionally make the wrong decision here,
838 * and grabbing the lock results in massive lock contention.
840 if (size
> dbuf_cache_target_bytes()) {
841 if (size
> dbuf_cache_hiwater_bytes())
843 cv_signal(&dbuf_evict_cv
);
848 dbuf_kstat_update(kstat_t
*ksp
, int rw
)
850 dbuf_stats_t
*ds
= ksp
->ks_data
;
851 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
853 if (rw
== KSTAT_WRITE
)
854 return (SET_ERROR(EACCES
));
856 ds
->cache_count
.value
.ui64
=
857 wmsum_value(&dbuf_sums
.cache_count
);
858 ds
->cache_size_bytes
.value
.ui64
=
859 zfs_refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
);
860 ds
->cache_target_bytes
.value
.ui64
= dbuf_cache_target_bytes();
861 ds
->cache_hiwater_bytes
.value
.ui64
= dbuf_cache_hiwater_bytes();
862 ds
->cache_lowater_bytes
.value
.ui64
= dbuf_cache_lowater_bytes();
863 ds
->cache_total_evicts
.value
.ui64
=
864 wmsum_value(&dbuf_sums
.cache_total_evicts
);
865 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
866 ds
->cache_levels
[i
].value
.ui64
=
867 wmsum_value(&dbuf_sums
.cache_levels
[i
]);
868 ds
->cache_levels_bytes
[i
].value
.ui64
=
869 wmsum_value(&dbuf_sums
.cache_levels_bytes
[i
]);
871 ds
->hash_hits
.value
.ui64
=
872 wmsum_value(&dbuf_sums
.hash_hits
);
873 ds
->hash_misses
.value
.ui64
=
874 wmsum_value(&dbuf_sums
.hash_misses
);
875 ds
->hash_collisions
.value
.ui64
=
876 wmsum_value(&dbuf_sums
.hash_collisions
);
877 ds
->hash_chains
.value
.ui64
=
878 wmsum_value(&dbuf_sums
.hash_chains
);
879 ds
->hash_insert_race
.value
.ui64
=
880 wmsum_value(&dbuf_sums
.hash_insert_race
);
881 ds
->hash_table_count
.value
.ui64
= h
->hash_table_mask
+ 1;
882 ds
->hash_mutex_count
.value
.ui64
= h
->hash_mutex_mask
+ 1;
883 ds
->metadata_cache_count
.value
.ui64
=
884 wmsum_value(&dbuf_sums
.metadata_cache_count
);
885 ds
->metadata_cache_size_bytes
.value
.ui64
= zfs_refcount_count(
886 &dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
);
887 ds
->metadata_cache_overflow
.value
.ui64
=
888 wmsum_value(&dbuf_sums
.metadata_cache_overflow
);
895 uint64_t hmsize
, hsize
= 1ULL << 16;
896 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
899 * The hash table is big enough to fill one eighth of physical memory
900 * with an average block size of zfs_arc_average_blocksize (default 8K).
901 * By default, the table will take up
902 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
904 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory() / 8)
907 h
->hash_table
= NULL
;
908 while (h
->hash_table
== NULL
) {
909 h
->hash_table_mask
= hsize
- 1;
911 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
912 if (h
->hash_table
== NULL
)
915 ASSERT3U(hsize
, >=, 1ULL << 10);
919 * The hash table buckets are protected by an array of mutexes where
920 * each mutex is reponsible for protecting 128 buckets. A minimum
921 * array size of 8192 is targeted to avoid contention.
923 if (dbuf_mutex_cache_shift
== 0)
924 hmsize
= MAX(hsize
>> 7, 1ULL << 13);
926 hmsize
= 1ULL << MIN(dbuf_mutex_cache_shift
, 24);
928 h
->hash_mutexes
= NULL
;
929 while (h
->hash_mutexes
== NULL
) {
930 h
->hash_mutex_mask
= hmsize
- 1;
932 h
->hash_mutexes
= vmem_zalloc(hmsize
* sizeof (kmutex_t
),
934 if (h
->hash_mutexes
== NULL
)
938 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
939 sizeof (dmu_buf_impl_t
),
940 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
942 for (int i
= 0; i
< hmsize
; i
++)
943 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
948 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
949 * configuration is not required.
951 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
953 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
954 multilist_create(&dbuf_caches
[dcs
].cache
,
955 sizeof (dmu_buf_impl_t
),
956 offsetof(dmu_buf_impl_t
, db_cache_link
),
957 dbuf_cache_multilist_index_func
);
958 zfs_refcount_create(&dbuf_caches
[dcs
].size
);
961 dbuf_evict_thread_exit
= B_FALSE
;
962 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
963 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
964 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
965 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
967 wmsum_init(&dbuf_sums
.cache_count
, 0);
968 wmsum_init(&dbuf_sums
.cache_total_evicts
, 0);
969 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
970 wmsum_init(&dbuf_sums
.cache_levels
[i
], 0);
971 wmsum_init(&dbuf_sums
.cache_levels_bytes
[i
], 0);
973 wmsum_init(&dbuf_sums
.hash_hits
, 0);
974 wmsum_init(&dbuf_sums
.hash_misses
, 0);
975 wmsum_init(&dbuf_sums
.hash_collisions
, 0);
976 wmsum_init(&dbuf_sums
.hash_chains
, 0);
977 wmsum_init(&dbuf_sums
.hash_insert_race
, 0);
978 wmsum_init(&dbuf_sums
.metadata_cache_count
, 0);
979 wmsum_init(&dbuf_sums
.metadata_cache_overflow
, 0);
981 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
982 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
984 if (dbuf_ksp
!= NULL
) {
985 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
986 snprintf(dbuf_stats
.cache_levels
[i
].name
,
987 KSTAT_STRLEN
, "cache_level_%d", i
);
988 dbuf_stats
.cache_levels
[i
].data_type
=
990 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
991 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
992 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
995 dbuf_ksp
->ks_data
= &dbuf_stats
;
996 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
997 kstat_install(dbuf_ksp
);
1004 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
1006 dbuf_stats_destroy();
1008 for (int i
= 0; i
< (h
->hash_mutex_mask
+ 1); i
++)
1009 mutex_destroy(&h
->hash_mutexes
[i
]);
1011 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
1012 vmem_free(h
->hash_mutexes
, (h
->hash_mutex_mask
+ 1) *
1015 kmem_cache_destroy(dbuf_kmem_cache
);
1016 taskq_destroy(dbu_evict_taskq
);
1018 mutex_enter(&dbuf_evict_lock
);
1019 dbuf_evict_thread_exit
= B_TRUE
;
1020 while (dbuf_evict_thread_exit
) {
1021 cv_signal(&dbuf_evict_cv
);
1022 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
1024 mutex_exit(&dbuf_evict_lock
);
1026 mutex_destroy(&dbuf_evict_lock
);
1027 cv_destroy(&dbuf_evict_cv
);
1029 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
1030 zfs_refcount_destroy(&dbuf_caches
[dcs
].size
);
1031 multilist_destroy(&dbuf_caches
[dcs
].cache
);
1034 if (dbuf_ksp
!= NULL
) {
1035 kstat_delete(dbuf_ksp
);
1039 wmsum_fini(&dbuf_sums
.cache_count
);
1040 wmsum_fini(&dbuf_sums
.cache_total_evicts
);
1041 for (int i
= 0; i
< DN_MAX_LEVELS
; i
++) {
1042 wmsum_fini(&dbuf_sums
.cache_levels
[i
]);
1043 wmsum_fini(&dbuf_sums
.cache_levels_bytes
[i
]);
1045 wmsum_fini(&dbuf_sums
.hash_hits
);
1046 wmsum_fini(&dbuf_sums
.hash_misses
);
1047 wmsum_fini(&dbuf_sums
.hash_collisions
);
1048 wmsum_fini(&dbuf_sums
.hash_chains
);
1049 wmsum_fini(&dbuf_sums
.hash_insert_race
);
1050 wmsum_fini(&dbuf_sums
.metadata_cache_count
);
1051 wmsum_fini(&dbuf_sums
.metadata_cache_overflow
);
1060 dbuf_verify(dmu_buf_impl_t
*db
)
1063 dbuf_dirty_record_t
*dr
;
1066 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1068 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
1071 ASSERT(db
->db_objset
!= NULL
);
1075 ASSERT(db
->db_parent
== NULL
);
1076 ASSERT(db
->db_blkptr
== NULL
);
1078 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
1079 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
1080 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
1081 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
1082 db
->db_blkid
== DMU_SPILL_BLKID
||
1083 !avl_is_empty(&dn
->dn_dbufs
));
1085 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1087 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1088 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
1089 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
1091 ASSERT0(db
->db
.db_offset
);
1093 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
1096 if ((dr
= list_head(&db
->db_dirty_records
)) != NULL
) {
1097 ASSERT(dr
->dr_dbuf
== db
);
1098 txg_prev
= dr
->dr_txg
;
1099 for (dr
= list_next(&db
->db_dirty_records
, dr
); dr
!= NULL
;
1100 dr
= list_next(&db
->db_dirty_records
, dr
)) {
1101 ASSERT(dr
->dr_dbuf
== db
);
1102 ASSERT(txg_prev
> dr
->dr_txg
);
1103 txg_prev
= dr
->dr_txg
;
1108 * We can't assert that db_size matches dn_datablksz because it
1109 * can be momentarily different when another thread is doing
1110 * dnode_set_blksz().
1112 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
1113 dr
= db
->db_data_pending
;
1115 * It should only be modified in syncing context, so
1116 * make sure we only have one copy of the data.
1118 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
1121 /* verify db->db_blkptr */
1122 if (db
->db_blkptr
) {
1123 if (db
->db_parent
== dn
->dn_dbuf
) {
1124 /* db is pointed to by the dnode */
1125 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1126 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
1127 ASSERT(db
->db_parent
== NULL
);
1129 ASSERT(db
->db_parent
!= NULL
);
1130 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
1131 ASSERT3P(db
->db_blkptr
, ==,
1132 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
1134 /* db is pointed to by an indirect block */
1135 int epb __maybe_unused
= db
->db_parent
->db
.db_size
>>
1137 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
1138 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
1141 * dnode_grow_indblksz() can make this fail if we don't
1142 * have the parent's rwlock. XXX indblksz no longer
1143 * grows. safe to do this now?
1145 if (RW_LOCK_HELD(&db
->db_parent
->db_rwlock
)) {
1146 ASSERT3P(db
->db_blkptr
, ==,
1147 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
1148 db
->db_blkid
% epb
));
1152 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
1153 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
1154 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
1155 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
1157 * If the blkptr isn't set but they have nonzero data,
1158 * it had better be dirty, otherwise we'll lose that
1159 * data when we evict this buffer.
1161 * There is an exception to this rule for indirect blocks; in
1162 * this case, if the indirect block is a hole, we fill in a few
1163 * fields on each of the child blocks (importantly, birth time)
1164 * to prevent hole birth times from being lost when you
1165 * partially fill in a hole.
1167 if (db
->db_dirtycnt
== 0) {
1168 if (db
->db_level
== 0) {
1169 uint64_t *buf
= db
->db
.db_data
;
1172 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
1173 ASSERT(buf
[i
] == 0);
1176 blkptr_t
*bps
= db
->db
.db_data
;
1177 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1180 * We want to verify that all the blkptrs in the
1181 * indirect block are holes, but we may have
1182 * automatically set up a few fields for them.
1183 * We iterate through each blkptr and verify
1184 * they only have those fields set.
1187 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1189 blkptr_t
*bp
= &bps
[i
];
1190 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1193 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1194 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1195 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1196 ASSERT0(bp
->blk_fill
);
1197 ASSERT0(bp
->blk_pad
[0]);
1198 ASSERT0(bp
->blk_pad
[1]);
1199 ASSERT(!BP_IS_EMBEDDED(bp
));
1200 ASSERT(BP_IS_HOLE(bp
));
1201 ASSERT0(bp
->blk_phys_birth
);
1211 dbuf_clear_data(dmu_buf_impl_t
*db
)
1213 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1214 dbuf_evict_user(db
);
1215 ASSERT3P(db
->db_buf
, ==, NULL
);
1216 db
->db
.db_data
= NULL
;
1217 if (db
->db_state
!= DB_NOFILL
) {
1218 db
->db_state
= DB_UNCACHED
;
1219 DTRACE_SET_STATE(db
, "clear data");
1224 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1226 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1227 ASSERT(buf
!= NULL
);
1230 ASSERT(buf
->b_data
!= NULL
);
1231 db
->db
.db_data
= buf
->b_data
;
1235 dbuf_alloc_arcbuf(dmu_buf_impl_t
*db
)
1237 spa_t
*spa
= db
->db_objset
->os_spa
;
1239 return (arc_alloc_buf(spa
, db
, DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
1243 * Loan out an arc_buf for read. Return the loaned arc_buf.
1246 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1250 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1251 mutex_enter(&db
->db_mtx
);
1252 if (arc_released(db
->db_buf
) || zfs_refcount_count(&db
->db_holds
) > 1) {
1253 int blksz
= db
->db
.db_size
;
1254 spa_t
*spa
= db
->db_objset
->os_spa
;
1256 mutex_exit(&db
->db_mtx
);
1257 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1258 memcpy(abuf
->b_data
, db
->db
.db_data
, blksz
);
1261 arc_loan_inuse_buf(abuf
, db
);
1263 dbuf_clear_data(db
);
1264 mutex_exit(&db
->db_mtx
);
1270 * Calculate which level n block references the data at the level 0 offset
1274 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1276 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1278 * The level n blkid is equal to the level 0 blkid divided by
1279 * the number of level 0s in a level n block.
1281 * The level 0 blkid is offset >> datablkshift =
1282 * offset / 2^datablkshift.
1284 * The number of level 0s in a level n is the number of block
1285 * pointers in an indirect block, raised to the power of level.
1286 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1287 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1289 * Thus, the level n blkid is: offset /
1290 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1291 * = offset / 2^(datablkshift + level *
1292 * (indblkshift - SPA_BLKPTRSHIFT))
1293 * = offset >> (datablkshift + level *
1294 * (indblkshift - SPA_BLKPTRSHIFT))
1297 const unsigned exp
= dn
->dn_datablkshift
+
1298 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1300 if (exp
>= 8 * sizeof (offset
)) {
1301 /* This only happens on the highest indirection level */
1302 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1306 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1308 return (offset
>> exp
);
1310 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1316 * This function is used to lock the parent of the provided dbuf. This should be
1317 * used when modifying or reading db_blkptr.
1320 dmu_buf_lock_parent(dmu_buf_impl_t
*db
, krw_t rw
, const void *tag
)
1322 enum db_lock_type ret
= DLT_NONE
;
1323 if (db
->db_parent
!= NULL
) {
1324 rw_enter(&db
->db_parent
->db_rwlock
, rw
);
1326 } else if (dmu_objset_ds(db
->db_objset
) != NULL
) {
1327 rrw_enter(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, rw
,
1332 * We only return a DLT_NONE lock when it's the top-most indirect block
1333 * of the meta-dnode of the MOS.
1339 * We need to pass the lock type in because it's possible that the block will
1340 * move from being the topmost indirect block in a dnode (and thus, have no
1341 * parent) to not the top-most via an indirection increase. This would cause a
1342 * panic if we didn't pass the lock type in.
1345 dmu_buf_unlock_parent(dmu_buf_impl_t
*db
, db_lock_type_t type
, const void *tag
)
1347 if (type
== DLT_PARENT
)
1348 rw_exit(&db
->db_parent
->db_rwlock
);
1349 else if (type
== DLT_OBJSET
)
1350 rrw_exit(&dmu_objset_ds(db
->db_objset
)->ds_bp_rwlock
, tag
);
1354 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1355 arc_buf_t
*buf
, void *vdb
)
1357 (void) zb
, (void) bp
;
1358 dmu_buf_impl_t
*db
= vdb
;
1360 mutex_enter(&db
->db_mtx
);
1361 ASSERT3U(db
->db_state
, ==, DB_READ
);
1363 * All reads are synchronous, so we must have a hold on the dbuf
1365 ASSERT(zfs_refcount_count(&db
->db_holds
) > 0);
1366 ASSERT(db
->db_buf
== NULL
);
1367 ASSERT(db
->db
.db_data
== NULL
);
1370 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
1371 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1372 ASSERT3P(db
->db_buf
, ==, NULL
);
1373 db
->db_state
= DB_UNCACHED
;
1374 DTRACE_SET_STATE(db
, "i/o error");
1375 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1376 /* freed in flight */
1377 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1378 arc_release(buf
, db
);
1379 memset(buf
->b_data
, 0, db
->db
.db_size
);
1380 arc_buf_freeze(buf
);
1381 db
->db_freed_in_flight
= FALSE
;
1382 dbuf_set_data(db
, buf
);
1383 db
->db_state
= DB_CACHED
;
1384 DTRACE_SET_STATE(db
, "freed in flight");
1387 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1388 dbuf_set_data(db
, buf
);
1389 db
->db_state
= DB_CACHED
;
1390 DTRACE_SET_STATE(db
, "successful read");
1392 cv_broadcast(&db
->db_changed
);
1393 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
1397 * Shortcut for performing reads on bonus dbufs. Returns
1398 * an error if we fail to verify the dnode associated with
1399 * a decrypted block. Otherwise success.
1402 dbuf_read_bonus(dmu_buf_impl_t
*db
, dnode_t
*dn
, uint32_t flags
)
1404 int bonuslen
, max_bonuslen
, err
;
1406 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1410 bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1411 max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1412 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1413 ASSERT(DB_DNODE_HELD(db
));
1414 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1415 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1416 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1417 if (bonuslen
< max_bonuslen
)
1418 memset(db
->db
.db_data
, 0, max_bonuslen
);
1420 memcpy(db
->db
.db_data
, DN_BONUS(dn
->dn_phys
), bonuslen
);
1421 db
->db_state
= DB_CACHED
;
1422 DTRACE_SET_STATE(db
, "bonus buffer filled");
1427 dbuf_handle_indirect_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
)
1429 blkptr_t
*bps
= db
->db
.db_data
;
1430 uint32_t indbs
= 1ULL << dn
->dn_indblkshift
;
1431 int n_bps
= indbs
>> SPA_BLKPTRSHIFT
;
1433 for (int i
= 0; i
< n_bps
; i
++) {
1434 blkptr_t
*bp
= &bps
[i
];
1436 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==, indbs
);
1437 BP_SET_LSIZE(bp
, BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1438 dn
->dn_datablksz
: BP_GET_LSIZE(db
->db_blkptr
));
1439 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1440 BP_SET_LEVEL(bp
, BP_GET_LEVEL(db
->db_blkptr
) - 1);
1441 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1446 * Handle reads on dbufs that are holes, if necessary. This function
1447 * requires that the dbuf's mutex is held. Returns success (0) if action
1448 * was taken, ENOENT if no action was taken.
1451 dbuf_read_hole(dmu_buf_impl_t
*db
, dnode_t
*dn
)
1453 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1455 int is_hole
= db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
);
1457 * For level 0 blocks only, if the above check fails:
1458 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1459 * processes the delete record and clears the bp while we are waiting
1460 * for the dn_mtx (resulting in a "no" from block_freed).
1462 if (!is_hole
&& db
->db_level
== 0) {
1463 is_hole
= dnode_block_freed(dn
, db
->db_blkid
) ||
1464 BP_IS_HOLE(db
->db_blkptr
);
1468 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1469 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
1471 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1472 BP_IS_HOLE(db
->db_blkptr
) &&
1473 db
->db_blkptr
->blk_birth
!= 0) {
1474 dbuf_handle_indirect_hole(db
, dn
);
1476 db
->db_state
= DB_CACHED
;
1477 DTRACE_SET_STATE(db
, "hole read satisfied");
1484 * This function ensures that, when doing a decrypting read of a block,
1485 * we make sure we have decrypted the dnode associated with it. We must do
1486 * this so that we ensure we are fully authenticating the checksum-of-MACs
1487 * tree from the root of the objset down to this block. Indirect blocks are
1488 * always verified against their secure checksum-of-MACs assuming that the
1489 * dnode containing them is correct. Now that we are doing a decrypting read,
1490 * we can be sure that the key is loaded and verify that assumption. This is
1491 * especially important considering that we always read encrypted dnode
1492 * blocks as raw data (without verifying their MACs) to start, and
1493 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1496 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
)
1499 objset_t
*os
= db
->db_objset
;
1500 arc_buf_t
*dnode_abuf
;
1502 zbookmark_phys_t zb
;
1504 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1506 if (!os
->os_encrypted
|| os
->os_raw_receive
||
1507 (flags
& DB_RF_NO_DECRYPT
) != 0)
1512 dnode_abuf
= (dn
->dn_dbuf
!= NULL
) ? dn
->dn_dbuf
->db_buf
: NULL
;
1514 if (dnode_abuf
== NULL
|| !arc_is_encrypted(dnode_abuf
)) {
1519 SET_BOOKMARK(&zb
, dmu_objset_id(os
),
1520 DMU_META_DNODE_OBJECT
, 0, dn
->dn_dbuf
->db_blkid
);
1521 err
= arc_untransform(dnode_abuf
, os
->os_spa
, &zb
, B_TRUE
);
1524 * An error code of EACCES tells us that the key is still not
1525 * available. This is ok if we are only reading authenticated
1526 * (and therefore non-encrypted) blocks.
1528 if (err
== EACCES
&& ((db
->db_blkid
!= DMU_BONUS_BLKID
&&
1529 !DMU_OT_IS_ENCRYPTED(dn
->dn_type
)) ||
1530 (db
->db_blkid
== DMU_BONUS_BLKID
&&
1531 !DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))))
1540 * Drops db_mtx and the parent lock specified by dblt and tag before
1544 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
,
1545 db_lock_type_t dblt
, const void *tag
)
1548 zbookmark_phys_t zb
;
1549 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1552 err
= zio_flags
= 0;
1555 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1556 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1557 ASSERT(db
->db_state
== DB_UNCACHED
);
1558 ASSERT(db
->db_buf
== NULL
);
1559 ASSERT(db
->db_parent
== NULL
||
1560 RW_LOCK_HELD(&db
->db_parent
->db_rwlock
));
1562 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1563 err
= dbuf_read_bonus(db
, dn
, flags
);
1567 err
= dbuf_read_hole(db
, dn
);
1572 * Any attempt to read a redacted block should result in an error. This
1573 * will never happen under normal conditions, but can be useful for
1574 * debugging purposes.
1576 if (BP_IS_REDACTED(db
->db_blkptr
)) {
1577 ASSERT(dsl_dataset_feature_is_active(
1578 db
->db_objset
->os_dsl_dataset
,
1579 SPA_FEATURE_REDACTED_DATASETS
));
1580 err
= SET_ERROR(EIO
);
1584 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1585 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1588 * All bps of an encrypted os should have the encryption bit set.
1589 * If this is not true it indicates tampering and we report an error.
1591 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(db
->db_blkptr
)) {
1592 spa_log_error(db
->db_objset
->os_spa
, &zb
);
1593 zfs_panic_recover("unencrypted block in encrypted "
1594 "object set %llu", dmu_objset_id(db
->db_objset
));
1595 err
= SET_ERROR(EIO
);
1599 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1605 db
->db_state
= DB_READ
;
1606 DTRACE_SET_STATE(db
, "read issued");
1607 mutex_exit(&db
->db_mtx
);
1609 if (dbuf_is_l2cacheable(db
))
1610 aflags
|= ARC_FLAG_L2CACHE
;
1612 dbuf_add_ref(db
, NULL
);
1614 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1615 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1617 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1618 zio_flags
|= ZIO_FLAG_RAW
;
1620 * The zio layer will copy the provided blkptr later, but we need to
1621 * do this now so that we can release the parent's rwlock. We have to
1622 * do that now so that if dbuf_read_done is called synchronously (on
1623 * an l1 cache hit) we don't acquire the db_mtx while holding the
1624 * parent's rwlock, which would be a lock ordering violation.
1626 blkptr_t bp
= *db
->db_blkptr
;
1627 dmu_buf_unlock_parent(db
, dblt
, tag
);
1628 (void) arc_read(zio
, db
->db_objset
->os_spa
, &bp
,
1629 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1634 mutex_exit(&db
->db_mtx
);
1635 dmu_buf_unlock_parent(db
, dblt
, tag
);
1640 * This is our just-in-time copy function. It makes a copy of buffers that
1641 * have been modified in a previous transaction group before we access them in
1642 * the current active group.
1644 * This function is used in three places: when we are dirtying a buffer for the
1645 * first time in a txg, when we are freeing a range in a dnode that includes
1646 * this buffer, and when we are accessing a buffer which was received compressed
1647 * and later referenced in a WRITE_BYREF record.
1649 * Note that when we are called from dbuf_free_range() we do not put a hold on
1650 * the buffer, we just traverse the active dbuf list for the dnode.
1653 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1655 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
1657 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1658 ASSERT(db
->db
.db_data
!= NULL
);
1659 ASSERT(db
->db_level
== 0);
1660 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1663 (dr
->dt
.dl
.dr_data
!=
1664 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1668 * If the last dirty record for this dbuf has not yet synced
1669 * and its referencing the dbuf data, either:
1670 * reset the reference to point to a new copy,
1671 * or (if there a no active holders)
1672 * just null out the current db_data pointer.
1674 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1675 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1676 dnode_t
*dn
= DB_DNODE(db
);
1677 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1678 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1679 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1680 memcpy(dr
->dt
.dl
.dr_data
, db
->db
.db_data
, bonuslen
);
1681 } else if (zfs_refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1682 dnode_t
*dn
= DB_DNODE(db
);
1683 int size
= arc_buf_size(db
->db_buf
);
1684 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1685 spa_t
*spa
= db
->db_objset
->os_spa
;
1686 enum zio_compress compress_type
=
1687 arc_get_compression(db
->db_buf
);
1688 uint8_t complevel
= arc_get_complevel(db
->db_buf
);
1690 if (arc_is_encrypted(db
->db_buf
)) {
1691 boolean_t byteorder
;
1692 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1693 uint8_t iv
[ZIO_DATA_IV_LEN
];
1694 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1696 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1698 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1699 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1700 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1701 compress_type
, complevel
);
1702 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1703 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1704 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1705 size
, arc_buf_lsize(db
->db_buf
), compress_type
,
1708 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1710 memcpy(dr
->dt
.dl
.dr_data
->b_data
, db
->db
.db_data
, size
);
1713 dbuf_clear_data(db
);
1718 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1725 * We don't have to hold the mutex to check db_state because it
1726 * can't be freed while we have a hold on the buffer.
1728 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1730 if (db
->db_state
== DB_NOFILL
)
1731 return (SET_ERROR(EIO
));
1736 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1737 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1738 DBUF_IS_CACHEABLE(db
);
1740 mutex_enter(&db
->db_mtx
);
1741 if (db
->db_state
== DB_CACHED
) {
1742 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1745 * Ensure that this block's dnode has been decrypted if
1746 * the caller has requested decrypted data.
1748 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1751 * If the arc buf is compressed or encrypted and the caller
1752 * requested uncompressed data, we need to untransform it
1753 * before returning. We also call arc_untransform() on any
1754 * unauthenticated blocks, which will verify their MAC if
1755 * the key is now available.
1757 if (err
== 0 && db
->db_buf
!= NULL
&&
1758 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1759 (arc_is_encrypted(db
->db_buf
) ||
1760 arc_is_unauthenticated(db
->db_buf
) ||
1761 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1762 zbookmark_phys_t zb
;
1764 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1765 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1766 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1767 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1768 dbuf_set_data(db
, db
->db_buf
);
1770 mutex_exit(&db
->db_mtx
);
1771 if (err
== 0 && prefetch
) {
1772 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1773 B_FALSE
, flags
& DB_RF_HAVESTRUCT
);
1776 DBUF_STAT_BUMP(hash_hits
);
1777 } else if (db
->db_state
== DB_UNCACHED
) {
1778 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1779 boolean_t need_wait
= B_FALSE
;
1781 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
1784 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1785 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1788 err
= dbuf_read_impl(db
, zio
, flags
, dblt
, FTAG
);
1790 * dbuf_read_impl has dropped db_mtx and our parent's rwlock
1793 if (!err
&& prefetch
) {
1794 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1795 db
->db_state
!= DB_CACHED
,
1796 flags
& DB_RF_HAVESTRUCT
);
1800 DBUF_STAT_BUMP(hash_misses
);
1803 * If we created a zio_root we must execute it to avoid
1804 * leaking it, even if it isn't attached to any work due
1805 * to an error in dbuf_read_impl().
1809 err
= zio_wait(zio
);
1811 VERIFY0(zio_wait(zio
));
1815 * Another reader came in while the dbuf was in flight
1816 * between UNCACHED and CACHED. Either a writer will finish
1817 * writing the buffer (sending the dbuf to CACHED) or the
1818 * first reader's request will reach the read_done callback
1819 * and send the dbuf to CACHED. Otherwise, a failure
1820 * occurred and the dbuf went to UNCACHED.
1822 mutex_exit(&db
->db_mtx
);
1824 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
,
1825 B_TRUE
, flags
& DB_RF_HAVESTRUCT
);
1828 DBUF_STAT_BUMP(hash_misses
);
1830 /* Skip the wait per the caller's request. */
1831 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1832 mutex_enter(&db
->db_mtx
);
1833 while (db
->db_state
== DB_READ
||
1834 db
->db_state
== DB_FILL
) {
1835 ASSERT(db
->db_state
== DB_READ
||
1836 (flags
& DB_RF_HAVESTRUCT
) == 0);
1837 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1839 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1841 if (db
->db_state
== DB_UNCACHED
)
1842 err
= SET_ERROR(EIO
);
1843 mutex_exit(&db
->db_mtx
);
1851 dbuf_noread(dmu_buf_impl_t
*db
)
1853 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
1854 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1855 mutex_enter(&db
->db_mtx
);
1856 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1857 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1858 if (db
->db_state
== DB_UNCACHED
) {
1859 ASSERT(db
->db_buf
== NULL
);
1860 ASSERT(db
->db
.db_data
== NULL
);
1861 dbuf_set_data(db
, dbuf_alloc_arcbuf(db
));
1862 db
->db_state
= DB_FILL
;
1863 DTRACE_SET_STATE(db
, "assigning filled buffer");
1864 } else if (db
->db_state
== DB_NOFILL
) {
1865 dbuf_clear_data(db
);
1867 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1869 mutex_exit(&db
->db_mtx
);
1873 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1875 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1876 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1877 uint64_t txg
= dr
->dr_txg
;
1879 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1881 * This assert is valid because dmu_sync() expects to be called by
1882 * a zilog's get_data while holding a range lock. This call only
1883 * comes from dbuf_dirty() callers who must also hold a range lock.
1885 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1886 ASSERT(db
->db_level
== 0);
1888 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1889 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1892 ASSERT(db
->db_data_pending
!= dr
);
1894 /* free this block */
1895 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1896 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1898 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1899 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1900 dr
->dt
.dl
.dr_has_raw_params
= B_FALSE
;
1903 * Release the already-written buffer, so we leave it in
1904 * a consistent dirty state. Note that all callers are
1905 * modifying the buffer, so they will immediately do
1906 * another (redundant) arc_release(). Therefore, leave
1907 * the buf thawed to save the effort of freezing &
1908 * immediately re-thawing it.
1910 arc_release(dr
->dt
.dl
.dr_data
, db
);
1914 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1915 * data blocks in the free range, so that any future readers will find
1919 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1922 dmu_buf_impl_t
*db_search
;
1923 dmu_buf_impl_t
*db
, *db_next
;
1924 uint64_t txg
= tx
->tx_txg
;
1926 dbuf_dirty_record_t
*dr
;
1928 if (end_blkid
> dn
->dn_maxblkid
&&
1929 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1930 end_blkid
= dn
->dn_maxblkid
;
1931 dprintf_dnode(dn
, "start=%llu end=%llu\n", (u_longlong_t
)start_blkid
,
1932 (u_longlong_t
)end_blkid
);
1934 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1935 db_search
->db_level
= 0;
1936 db_search
->db_blkid
= start_blkid
;
1937 db_search
->db_state
= DB_SEARCH
;
1939 mutex_enter(&dn
->dn_dbufs_mtx
);
1940 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1941 ASSERT3P(db
, ==, NULL
);
1943 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1945 for (; db
!= NULL
; db
= db_next
) {
1946 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1947 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1949 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1952 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1954 /* found a level 0 buffer in the range */
1955 mutex_enter(&db
->db_mtx
);
1956 if (dbuf_undirty(db
, tx
)) {
1957 /* mutex has been dropped and dbuf destroyed */
1961 if (db
->db_state
== DB_UNCACHED
||
1962 db
->db_state
== DB_NOFILL
||
1963 db
->db_state
== DB_EVICTING
) {
1964 ASSERT(db
->db
.db_data
== NULL
);
1965 mutex_exit(&db
->db_mtx
);
1968 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1969 /* will be handled in dbuf_read_done or dbuf_rele */
1970 db
->db_freed_in_flight
= TRUE
;
1971 mutex_exit(&db
->db_mtx
);
1974 if (zfs_refcount_count(&db
->db_holds
) == 0) {
1979 /* The dbuf is referenced */
1981 dr
= list_head(&db
->db_dirty_records
);
1983 if (dr
->dr_txg
== txg
) {
1985 * This buffer is "in-use", re-adjust the file
1986 * size to reflect that this buffer may
1987 * contain new data when we sync.
1989 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1990 db
->db_blkid
> dn
->dn_maxblkid
)
1991 dn
->dn_maxblkid
= db
->db_blkid
;
1992 dbuf_unoverride(dr
);
1995 * This dbuf is not dirty in the open context.
1996 * Either uncache it (if its not referenced in
1997 * the open context) or reset its contents to
2000 dbuf_fix_old_data(db
, txg
);
2003 /* clear the contents if its cached */
2004 if (db
->db_state
== DB_CACHED
) {
2005 ASSERT(db
->db
.db_data
!= NULL
);
2006 arc_release(db
->db_buf
, db
);
2007 rw_enter(&db
->db_rwlock
, RW_WRITER
);
2008 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2009 rw_exit(&db
->db_rwlock
);
2010 arc_buf_freeze(db
->db_buf
);
2013 mutex_exit(&db
->db_mtx
);
2016 mutex_exit(&dn
->dn_dbufs_mtx
);
2017 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
2021 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
2023 arc_buf_t
*buf
, *old_buf
;
2024 dbuf_dirty_record_t
*dr
;
2025 int osize
= db
->db
.db_size
;
2026 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2029 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2035 * XXX we should be doing a dbuf_read, checking the return
2036 * value and returning that up to our callers
2038 dmu_buf_will_dirty(&db
->db
, tx
);
2040 /* create the data buffer for the new block */
2041 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
2043 /* copy old block data to the new block */
2044 old_buf
= db
->db_buf
;
2045 memcpy(buf
->b_data
, old_buf
->b_data
, MIN(osize
, size
));
2046 /* zero the remainder */
2048 memset((uint8_t *)buf
->b_data
+ osize
, 0, size
- osize
);
2050 mutex_enter(&db
->db_mtx
);
2051 dbuf_set_data(db
, buf
);
2052 arc_buf_destroy(old_buf
, db
);
2053 db
->db
.db_size
= size
;
2055 dr
= list_head(&db
->db_dirty_records
);
2056 /* dirty record added by dmu_buf_will_dirty() */
2058 if (db
->db_level
== 0)
2059 dr
->dt
.dl
.dr_data
= buf
;
2060 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2061 ASSERT3U(dr
->dr_accounted
, ==, osize
);
2062 dr
->dr_accounted
= size
;
2063 mutex_exit(&db
->db_mtx
);
2065 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
2070 dbuf_release_bp(dmu_buf_impl_t
*db
)
2072 objset_t
*os __maybe_unused
= db
->db_objset
;
2074 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
2075 ASSERT(arc_released(os
->os_phys_buf
) ||
2076 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
2077 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
2079 (void) arc_release(db
->db_buf
, db
);
2083 * We already have a dirty record for this TXG, and we are being
2087 dbuf_redirty(dbuf_dirty_record_t
*dr
)
2089 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2091 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2093 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
2095 * If this buffer has already been written out,
2096 * we now need to reset its state.
2098 dbuf_unoverride(dr
);
2099 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
2100 db
->db_state
!= DB_NOFILL
) {
2101 /* Already released on initial dirty, so just thaw. */
2102 ASSERT(arc_released(db
->db_buf
));
2103 arc_buf_thaw(db
->db_buf
);
2108 dbuf_dirty_record_t
*
2109 dbuf_dirty_lightweight(dnode_t
*dn
, uint64_t blkid
, dmu_tx_t
*tx
)
2111 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2112 IMPLY(dn
->dn_objset
->os_raw_receive
, dn
->dn_maxblkid
>= blkid
);
2113 dnode_new_blkid(dn
, blkid
, tx
, B_TRUE
, B_FALSE
);
2114 ASSERT(dn
->dn_maxblkid
>= blkid
);
2116 dbuf_dirty_record_t
*dr
= kmem_zalloc(sizeof (*dr
), KM_SLEEP
);
2117 list_link_init(&dr
->dr_dirty_node
);
2118 list_link_init(&dr
->dr_dbuf_node
);
2120 dr
->dr_txg
= tx
->tx_txg
;
2121 dr
->dt
.dll
.dr_blkid
= blkid
;
2122 dr
->dr_accounted
= dn
->dn_datablksz
;
2125 * There should not be any dbuf for the block that we're dirtying.
2126 * Otherwise the buffer contents could be inconsistent between the
2127 * dbuf and the lightweight dirty record.
2129 ASSERT3P(NULL
, ==, dbuf_find(dn
->dn_objset
, dn
->dn_object
, 0, blkid
));
2131 mutex_enter(&dn
->dn_mtx
);
2132 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2133 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2134 range_tree_clear(dn
->dn_free_ranges
[txgoff
], blkid
, 1);
2137 if (dn
->dn_nlevels
== 1) {
2138 ASSERT3U(blkid
, <, dn
->dn_nblkptr
);
2139 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2140 mutex_exit(&dn
->dn_mtx
);
2141 rw_exit(&dn
->dn_struct_rwlock
);
2142 dnode_setdirty(dn
, tx
);
2144 mutex_exit(&dn
->dn_mtx
);
2146 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2147 dmu_buf_impl_t
*parent_db
= dbuf_hold_level(dn
,
2148 1, blkid
>> epbs
, FTAG
);
2149 rw_exit(&dn
->dn_struct_rwlock
);
2150 if (parent_db
== NULL
) {
2151 kmem_free(dr
, sizeof (*dr
));
2154 int err
= dbuf_read(parent_db
, NULL
,
2155 (DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2157 dbuf_rele(parent_db
, FTAG
);
2158 kmem_free(dr
, sizeof (*dr
));
2162 dbuf_dirty_record_t
*parent_dr
= dbuf_dirty(parent_db
, tx
);
2163 dbuf_rele(parent_db
, FTAG
);
2164 mutex_enter(&parent_dr
->dt
.di
.dr_mtx
);
2165 ASSERT3U(parent_dr
->dr_txg
, ==, tx
->tx_txg
);
2166 list_insert_tail(&parent_dr
->dt
.di
.dr_children
, dr
);
2167 mutex_exit(&parent_dr
->dt
.di
.dr_mtx
);
2168 dr
->dr_parent
= parent_dr
;
2171 dmu_objset_willuse_space(dn
->dn_objset
, dr
->dr_accounted
, tx
);
2176 dbuf_dirty_record_t
*
2177 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2181 dbuf_dirty_record_t
*dr
, *dr_next
, *dr_head
;
2182 int txgoff
= tx
->tx_txg
& TXG_MASK
;
2183 boolean_t drop_struct_rwlock
= B_FALSE
;
2185 ASSERT(tx
->tx_txg
!= 0);
2186 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2187 DMU_TX_DIRTY_BUF(tx
, db
);
2192 * Shouldn't dirty a regular buffer in syncing context. Private
2193 * objects may be dirtied in syncing context, but only if they
2194 * were already pre-dirtied in open context.
2197 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
2198 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
2201 ASSERT(!dmu_tx_is_syncing(tx
) ||
2202 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
2203 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2204 dn
->dn_objset
->os_dsl_dataset
== NULL
);
2205 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2206 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2209 * We make this assert for private objects as well, but after we
2210 * check if we're already dirty. They are allowed to re-dirty
2211 * in syncing context.
2213 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2214 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2215 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2217 mutex_enter(&db
->db_mtx
);
2219 * XXX make this true for indirects too? The problem is that
2220 * transactions created with dmu_tx_create_assigned() from
2221 * syncing context don't bother holding ahead.
2223 ASSERT(db
->db_level
!= 0 ||
2224 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
2225 db
->db_state
== DB_NOFILL
);
2227 mutex_enter(&dn
->dn_mtx
);
2228 dnode_set_dirtyctx(dn
, tx
, db
);
2229 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
2230 dn
->dn_dirty_txg
= tx
->tx_txg
;
2231 mutex_exit(&dn
->dn_mtx
);
2233 if (db
->db_blkid
== DMU_SPILL_BLKID
)
2234 dn
->dn_have_spill
= B_TRUE
;
2237 * If this buffer is already dirty, we're done.
2239 dr_head
= list_head(&db
->db_dirty_records
);
2240 ASSERT(dr_head
== NULL
|| dr_head
->dr_txg
<= tx
->tx_txg
||
2241 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
2242 dr_next
= dbuf_find_dirty_lte(db
, tx
->tx_txg
);
2243 if (dr_next
&& dr_next
->dr_txg
== tx
->tx_txg
) {
2246 dbuf_redirty(dr_next
);
2247 mutex_exit(&db
->db_mtx
);
2252 * Only valid if not already dirty.
2254 ASSERT(dn
->dn_object
== 0 ||
2255 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
2256 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
2258 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
2261 * We should only be dirtying in syncing context if it's the
2262 * mos or we're initializing the os or it's a special object.
2263 * However, we are allowed to dirty in syncing context provided
2264 * we already dirtied it in open context. Hence we must make
2265 * this assertion only if we're not already dirty.
2268 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
2270 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2271 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
2272 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
2273 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
2274 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
2275 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
2277 ASSERT(db
->db
.db_size
!= 0);
2279 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2281 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2282 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
2286 * If this buffer is dirty in an old transaction group we need
2287 * to make a copy of it so that the changes we make in this
2288 * transaction group won't leak out when we sync the older txg.
2290 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
2291 list_link_init(&dr
->dr_dirty_node
);
2292 list_link_init(&dr
->dr_dbuf_node
);
2294 if (db
->db_level
== 0) {
2295 void *data_old
= db
->db_buf
;
2297 if (db
->db_state
!= DB_NOFILL
) {
2298 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2299 dbuf_fix_old_data(db
, tx
->tx_txg
);
2300 data_old
= db
->db
.db_data
;
2301 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
2303 * Release the data buffer from the cache so
2304 * that we can modify it without impacting
2305 * possible other users of this cached data
2306 * block. Note that indirect blocks and
2307 * private objects are not released until the
2308 * syncing state (since they are only modified
2311 arc_release(db
->db_buf
, db
);
2312 dbuf_fix_old_data(db
, tx
->tx_txg
);
2313 data_old
= db
->db_buf
;
2315 ASSERT(data_old
!= NULL
);
2317 dr
->dt
.dl
.dr_data
= data_old
;
2319 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
2320 list_create(&dr
->dt
.di
.dr_children
,
2321 sizeof (dbuf_dirty_record_t
),
2322 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
2324 if (db
->db_blkid
!= DMU_BONUS_BLKID
)
2325 dr
->dr_accounted
= db
->db
.db_size
;
2327 dr
->dr_txg
= tx
->tx_txg
;
2328 list_insert_before(&db
->db_dirty_records
, dr_next
, dr
);
2331 * We could have been freed_in_flight between the dbuf_noread
2332 * and dbuf_dirty. We win, as though the dbuf_noread() had
2333 * happened after the free.
2335 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2336 db
->db_blkid
!= DMU_SPILL_BLKID
) {
2337 mutex_enter(&dn
->dn_mtx
);
2338 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2339 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
2342 mutex_exit(&dn
->dn_mtx
);
2343 db
->db_freed_in_flight
= FALSE
;
2347 * This buffer is now part of this txg
2349 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
2350 db
->db_dirtycnt
+= 1;
2351 ASSERT3U(db
->db_dirtycnt
, <=, 3);
2353 mutex_exit(&db
->db_mtx
);
2355 if (db
->db_blkid
== DMU_BONUS_BLKID
||
2356 db
->db_blkid
== DMU_SPILL_BLKID
) {
2357 mutex_enter(&dn
->dn_mtx
);
2358 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2359 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2360 mutex_exit(&dn
->dn_mtx
);
2361 dnode_setdirty(dn
, tx
);
2366 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
2367 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2368 drop_struct_rwlock
= B_TRUE
;
2372 * If we are overwriting a dedup BP, then unless it is snapshotted,
2373 * when we get to syncing context we will need to decrement its
2374 * refcount in the DDT. Prefetch the relevant DDT block so that
2375 * syncing context won't have to wait for the i/o.
2377 if (db
->db_blkptr
!= NULL
) {
2378 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_READER
, FTAG
);
2379 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
2380 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
2384 * We need to hold the dn_struct_rwlock to make this assertion,
2385 * because it protects dn_phys / dn_next_nlevels from changing.
2387 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
2388 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
2389 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
2390 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
2391 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
2394 if (db
->db_level
== 0) {
2395 ASSERT(!db
->db_objset
->os_raw_receive
||
2396 dn
->dn_maxblkid
>= db
->db_blkid
);
2397 dnode_new_blkid(dn
, db
->db_blkid
, tx
,
2398 drop_struct_rwlock
, B_FALSE
);
2399 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
2402 if (db
->db_level
+1 < dn
->dn_nlevels
) {
2403 dmu_buf_impl_t
*parent
= db
->db_parent
;
2404 dbuf_dirty_record_t
*di
;
2405 int parent_held
= FALSE
;
2407 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
2408 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2409 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2410 db
->db_blkid
>> epbs
, FTAG
);
2411 ASSERT(parent
!= NULL
);
2414 if (drop_struct_rwlock
)
2415 rw_exit(&dn
->dn_struct_rwlock
);
2416 ASSERT3U(db
->db_level
+ 1, ==, parent
->db_level
);
2417 di
= dbuf_dirty(parent
, tx
);
2419 dbuf_rele(parent
, FTAG
);
2421 mutex_enter(&db
->db_mtx
);
2423 * Since we've dropped the mutex, it's possible that
2424 * dbuf_undirty() might have changed this out from under us.
2426 if (list_head(&db
->db_dirty_records
) == dr
||
2427 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2428 mutex_enter(&di
->dt
.di
.dr_mtx
);
2429 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2430 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2431 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2432 mutex_exit(&di
->dt
.di
.dr_mtx
);
2435 mutex_exit(&db
->db_mtx
);
2437 ASSERT(db
->db_level
+ 1 == dn
->dn_nlevels
);
2438 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2439 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2440 mutex_enter(&dn
->dn_mtx
);
2441 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2442 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2443 mutex_exit(&dn
->dn_mtx
);
2444 if (drop_struct_rwlock
)
2445 rw_exit(&dn
->dn_struct_rwlock
);
2448 dnode_setdirty(dn
, tx
);
2454 dbuf_undirty_bonus(dbuf_dirty_record_t
*dr
)
2456 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2458 if (dr
->dt
.dl
.dr_data
!= db
->db
.db_data
) {
2459 struct dnode
*dn
= dr
->dr_dnode
;
2460 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
2462 kmem_free(dr
->dt
.dl
.dr_data
, max_bonuslen
);
2463 arc_space_return(max_bonuslen
, ARC_SPACE_BONUS
);
2465 db
->db_data_pending
= NULL
;
2466 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
2467 list_remove(&db
->db_dirty_records
, dr
);
2468 if (dr
->dr_dbuf
->db_level
!= 0) {
2469 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2470 list_destroy(&dr
->dt
.di
.dr_children
);
2472 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2473 ASSERT3U(db
->db_dirtycnt
, >, 0);
2474 db
->db_dirtycnt
-= 1;
2478 * Undirty a buffer in the transaction group referenced by the given
2479 * transaction. Return whether this evicted the dbuf.
2482 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2484 uint64_t txg
= tx
->tx_txg
;
2489 * Due to our use of dn_nlevels below, this can only be called
2490 * in open context, unless we are operating on the MOS.
2491 * From syncing context, dn_nlevels may be different from the
2492 * dn_nlevels used when dbuf was dirtied.
2494 ASSERT(db
->db_objset
==
2495 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2496 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2497 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2498 ASSERT0(db
->db_level
);
2499 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2502 * If this buffer is not dirty, we're done.
2504 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, txg
);
2507 ASSERT(dr
->dr_dbuf
== db
);
2509 dnode_t
*dn
= dr
->dr_dnode
;
2511 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2513 ASSERT(db
->db
.db_size
!= 0);
2515 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2516 dr
->dr_accounted
, txg
);
2518 list_remove(&db
->db_dirty_records
, dr
);
2521 * Note that there are three places in dbuf_dirty()
2522 * where this dirty record may be put on a list.
2523 * Make sure to do a list_remove corresponding to
2524 * every one of those list_insert calls.
2526 if (dr
->dr_parent
) {
2527 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2528 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2529 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2530 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2531 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2532 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2533 mutex_enter(&dn
->dn_mtx
);
2534 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2535 mutex_exit(&dn
->dn_mtx
);
2538 if (db
->db_state
!= DB_NOFILL
) {
2539 dbuf_unoverride(dr
);
2541 ASSERT(db
->db_buf
!= NULL
);
2542 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2543 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2544 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2547 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2549 ASSERT(db
->db_dirtycnt
> 0);
2550 db
->db_dirtycnt
-= 1;
2552 if (zfs_refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2553 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
2562 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2564 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2566 ASSERT(tx
->tx_txg
!= 0);
2567 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2570 * Quick check for dirtiness. For already dirty blocks, this
2571 * reduces runtime of this function by >90%, and overall performance
2572 * by 50% for some workloads (e.g. file deletion with indirect blocks
2575 mutex_enter(&db
->db_mtx
);
2577 if (db
->db_state
== DB_CACHED
) {
2578 dbuf_dirty_record_t
*dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2580 * It's possible that it is already dirty but not cached,
2581 * because there are some calls to dbuf_dirty() that don't
2582 * go through dmu_buf_will_dirty().
2585 /* This dbuf is already dirty and cached. */
2587 mutex_exit(&db
->db_mtx
);
2591 mutex_exit(&db
->db_mtx
);
2594 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2595 flags
|= DB_RF_HAVESTRUCT
;
2597 (void) dbuf_read(db
, NULL
, flags
);
2598 (void) dbuf_dirty(db
, tx
);
2602 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2604 dmu_buf_will_dirty_impl(db_fake
,
2605 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2609 dmu_buf_is_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2611 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2612 dbuf_dirty_record_t
*dr
;
2614 mutex_enter(&db
->db_mtx
);
2615 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2616 mutex_exit(&db
->db_mtx
);
2617 return (dr
!= NULL
);
2621 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2623 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2625 db
->db_state
= DB_NOFILL
;
2626 DTRACE_SET_STATE(db
, "allocating NOFILL buffer");
2627 dmu_buf_will_fill(db_fake
, tx
);
2631 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2633 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2635 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2636 ASSERT(tx
->tx_txg
!= 0);
2637 ASSERT(db
->db_level
== 0);
2638 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2640 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2641 dmu_tx_private_ok(tx
));
2644 (void) dbuf_dirty(db
, tx
);
2648 * This function is effectively the same as dmu_buf_will_dirty(), but
2649 * indicates the caller expects raw encrypted data in the db, and provides
2650 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2651 * blkptr_t when this dbuf is written. This is only used for blocks of
2652 * dnodes, during raw receive.
2655 dmu_buf_set_crypt_params(dmu_buf_t
*db_fake
, boolean_t byteorder
,
2656 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
, dmu_tx_t
*tx
)
2658 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2659 dbuf_dirty_record_t
*dr
;
2662 * dr_has_raw_params is only processed for blocks of dnodes
2663 * (see dbuf_sync_dnode_leaf_crypt()).
2665 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
2666 ASSERT3U(db
->db_level
, ==, 0);
2667 ASSERT(db
->db_objset
->os_raw_receive
);
2669 dmu_buf_will_dirty_impl(db_fake
,
2670 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2672 dr
= dbuf_find_dirty_eq(db
, tx
->tx_txg
);
2674 ASSERT3P(dr
, !=, NULL
);
2676 dr
->dt
.dl
.dr_has_raw_params
= B_TRUE
;
2677 dr
->dt
.dl
.dr_byteorder
= byteorder
;
2678 memcpy(dr
->dt
.dl
.dr_salt
, salt
, ZIO_DATA_SALT_LEN
);
2679 memcpy(dr
->dt
.dl
.dr_iv
, iv
, ZIO_DATA_IV_LEN
);
2680 memcpy(dr
->dt
.dl
.dr_mac
, mac
, ZIO_DATA_MAC_LEN
);
2684 dbuf_override_impl(dmu_buf_impl_t
*db
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
2686 struct dirty_leaf
*dl
;
2687 dbuf_dirty_record_t
*dr
;
2689 dr
= list_head(&db
->db_dirty_records
);
2690 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2692 dl
->dr_overridden_by
= *bp
;
2693 dl
->dr_override_state
= DR_OVERRIDDEN
;
2694 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2698 dmu_buf_fill_done(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2701 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2702 dbuf_states_t old_state
;
2703 mutex_enter(&db
->db_mtx
);
2706 old_state
= db
->db_state
;
2707 db
->db_state
= DB_CACHED
;
2708 if (old_state
== DB_FILL
) {
2709 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2710 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2711 /* we were freed while filling */
2712 /* XXX dbuf_undirty? */
2713 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
2714 db
->db_freed_in_flight
= FALSE
;
2715 DTRACE_SET_STATE(db
,
2716 "fill done handling freed in flight");
2718 DTRACE_SET_STATE(db
, "fill done");
2720 cv_broadcast(&db
->db_changed
);
2722 mutex_exit(&db
->db_mtx
);
2726 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2727 bp_embedded_type_t etype
, enum zio_compress comp
,
2728 int uncompressed_size
, int compressed_size
, int byteorder
,
2731 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2732 struct dirty_leaf
*dl
;
2733 dmu_object_type_t type
;
2734 dbuf_dirty_record_t
*dr
;
2736 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2737 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2738 SPA_FEATURE_EMBEDDED_DATA
));
2742 type
= DB_DNODE(db
)->dn_type
;
2745 ASSERT0(db
->db_level
);
2746 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2748 dmu_buf_will_not_fill(dbuf
, tx
);
2750 dr
= list_head(&db
->db_dirty_records
);
2751 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2753 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2754 data
, comp
, uncompressed_size
, compressed_size
);
2755 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2756 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2757 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2758 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2760 dl
->dr_override_state
= DR_OVERRIDDEN
;
2761 dl
->dr_overridden_by
.blk_birth
= dr
->dr_txg
;
2765 dmu_buf_redact(dmu_buf_t
*dbuf
, dmu_tx_t
*tx
)
2767 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2768 dmu_object_type_t type
;
2769 ASSERT(dsl_dataset_feature_is_active(db
->db_objset
->os_dsl_dataset
,
2770 SPA_FEATURE_REDACTED_DATASETS
));
2773 type
= DB_DNODE(db
)->dn_type
;
2776 ASSERT0(db
->db_level
);
2777 dmu_buf_will_not_fill(dbuf
, tx
);
2779 blkptr_t bp
= { { { {0} } } };
2780 BP_SET_TYPE(&bp
, type
);
2781 BP_SET_LEVEL(&bp
, 0);
2782 BP_SET_BIRTH(&bp
, tx
->tx_txg
, 0);
2783 BP_SET_REDACTED(&bp
);
2784 BPE_SET_LSIZE(&bp
, dbuf
->db_size
);
2786 dbuf_override_impl(db
, &bp
, tx
);
2790 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2791 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2794 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2796 ASSERT(!zfs_refcount_is_zero(&db
->db_holds
));
2797 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2798 ASSERT(db
->db_level
== 0);
2799 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2800 ASSERT(buf
!= NULL
);
2801 ASSERT3U(arc_buf_lsize(buf
), ==, db
->db
.db_size
);
2802 ASSERT(tx
->tx_txg
!= 0);
2804 arc_return_buf(buf
, db
);
2805 ASSERT(arc_released(buf
));
2807 mutex_enter(&db
->db_mtx
);
2809 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2810 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2812 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2814 if (db
->db_state
== DB_CACHED
&&
2815 zfs_refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2817 * In practice, we will never have a case where we have an
2818 * encrypted arc buffer while additional holds exist on the
2819 * dbuf. We don't handle this here so we simply assert that
2822 ASSERT(!arc_is_encrypted(buf
));
2823 mutex_exit(&db
->db_mtx
);
2824 (void) dbuf_dirty(db
, tx
);
2825 memcpy(db
->db
.db_data
, buf
->b_data
, db
->db
.db_size
);
2826 arc_buf_destroy(buf
, db
);
2830 if (db
->db_state
== DB_CACHED
) {
2831 dbuf_dirty_record_t
*dr
= list_head(&db
->db_dirty_records
);
2833 ASSERT(db
->db_buf
!= NULL
);
2834 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2835 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2837 if (!arc_released(db
->db_buf
)) {
2838 ASSERT(dr
->dt
.dl
.dr_override_state
==
2840 arc_release(db
->db_buf
, db
);
2842 dr
->dt
.dl
.dr_data
= buf
;
2843 arc_buf_destroy(db
->db_buf
, db
);
2844 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2845 arc_release(db
->db_buf
, db
);
2846 arc_buf_destroy(db
->db_buf
, db
);
2850 ASSERT(db
->db_buf
== NULL
);
2851 dbuf_set_data(db
, buf
);
2852 db
->db_state
= DB_FILL
;
2853 DTRACE_SET_STATE(db
, "filling assigned arcbuf");
2854 mutex_exit(&db
->db_mtx
);
2855 (void) dbuf_dirty(db
, tx
);
2856 dmu_buf_fill_done(&db
->db
, tx
);
2860 dbuf_destroy(dmu_buf_impl_t
*db
)
2863 dmu_buf_impl_t
*parent
= db
->db_parent
;
2864 dmu_buf_impl_t
*dndb
;
2866 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2867 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
2869 if (db
->db_buf
!= NULL
) {
2870 arc_buf_destroy(db
->db_buf
, db
);
2874 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2875 int slots
= DB_DNODE(db
)->dn_num_slots
;
2876 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2877 if (db
->db
.db_data
!= NULL
) {
2878 kmem_free(db
->db
.db_data
, bonuslen
);
2879 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2880 db
->db_state
= DB_UNCACHED
;
2881 DTRACE_SET_STATE(db
, "buffer cleared");
2885 dbuf_clear_data(db
);
2887 if (multilist_link_active(&db
->db_cache_link
)) {
2888 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
2889 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
2891 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
2892 (void) zfs_refcount_remove_many(
2893 &dbuf_caches
[db
->db_caching_status
].size
,
2894 db
->db
.db_size
, db
);
2896 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
2897 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
2899 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
2900 DBUF_STAT_BUMPDOWN(cache_count
);
2901 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
2904 db
->db_caching_status
= DB_NO_CACHE
;
2907 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2908 ASSERT(db
->db_data_pending
== NULL
);
2909 ASSERT(list_is_empty(&db
->db_dirty_records
));
2911 db
->db_state
= DB_EVICTING
;
2912 DTRACE_SET_STATE(db
, "buffer eviction started");
2913 db
->db_blkptr
= NULL
;
2916 * Now that db_state is DB_EVICTING, nobody else can find this via
2917 * the hash table. We can now drop db_mtx, which allows us to
2918 * acquire the dn_dbufs_mtx.
2920 mutex_exit(&db
->db_mtx
);
2925 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2926 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2928 mutex_enter_nested(&dn
->dn_dbufs_mtx
,
2930 avl_remove(&dn
->dn_dbufs
, db
);
2934 mutex_exit(&dn
->dn_dbufs_mtx
);
2936 * Decrementing the dbuf count means that the hold corresponding
2937 * to the removed dbuf is no longer discounted in dnode_move(),
2938 * so the dnode cannot be moved until after we release the hold.
2939 * The membar_producer() ensures visibility of the decremented
2940 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2943 mutex_enter(&dn
->dn_mtx
);
2944 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
2945 db
->db_dnode_handle
= NULL
;
2947 dbuf_hash_remove(db
);
2952 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
2954 db
->db_parent
= NULL
;
2956 ASSERT(db
->db_buf
== NULL
);
2957 ASSERT(db
->db
.db_data
== NULL
);
2958 ASSERT(db
->db_hash_next
== NULL
);
2959 ASSERT(db
->db_blkptr
== NULL
);
2960 ASSERT(db
->db_data_pending
== NULL
);
2961 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
2962 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2965 * If this dbuf is referenced from an indirect dbuf,
2966 * decrement the ref count on the indirect dbuf.
2968 if (parent
&& parent
!= dndb
) {
2969 mutex_enter(&parent
->db_mtx
);
2970 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
2973 kmem_cache_free(dbuf_kmem_cache
, db
);
2974 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2978 * Note: While bpp will always be updated if the function returns success,
2979 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2980 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2983 __attribute__((always_inline
))
2985 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2986 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
2991 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2993 if (blkid
== DMU_SPILL_BLKID
) {
2994 mutex_enter(&dn
->dn_mtx
);
2995 if (dn
->dn_have_spill
&&
2996 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2997 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3000 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3001 *parentp
= dn
->dn_dbuf
;
3002 mutex_exit(&dn
->dn_mtx
);
3007 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
3008 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3010 ASSERT3U(level
* epbs
, <, 64);
3011 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3013 * This assertion shouldn't trip as long as the max indirect block size
3014 * is less than 1M. The reason for this is that up to that point,
3015 * the number of levels required to address an entire object with blocks
3016 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
3017 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
3018 * (i.e. we can address the entire object), objects will all use at most
3019 * N-1 levels and the assertion won't overflow. However, once epbs is
3020 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
3021 * enough to address an entire object, so objects will have 5 levels,
3022 * but then this assertion will overflow.
3024 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
3025 * need to redo this logic to handle overflows.
3027 ASSERT(level
>= nlevels
||
3028 ((nlevels
- level
- 1) * epbs
) +
3029 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
3030 if (level
>= nlevels
||
3031 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
3032 ((nlevels
- level
- 1) * epbs
)) ||
3034 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
3035 /* the buffer has no parent yet */
3036 return (SET_ERROR(ENOENT
));
3037 } else if (level
< nlevels
-1) {
3038 /* this block is referenced from an indirect block */
3041 err
= dbuf_hold_impl(dn
, level
+ 1,
3042 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
3046 err
= dbuf_read(*parentp
, NULL
,
3047 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
3049 dbuf_rele(*parentp
, NULL
);
3053 rw_enter(&(*parentp
)->db_rwlock
, RW_READER
);
3054 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
3055 (blkid
& ((1ULL << epbs
) - 1));
3056 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
3057 ASSERT(BP_IS_HOLE(*bpp
));
3058 rw_exit(&(*parentp
)->db_rwlock
);
3061 /* the block is referenced from the dnode */
3062 ASSERT3U(level
, ==, nlevels
-1);
3063 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
3064 blkid
< dn
->dn_phys
->dn_nblkptr
);
3066 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
3067 *parentp
= dn
->dn_dbuf
;
3069 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
3074 static dmu_buf_impl_t
*
3075 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3076 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
3078 objset_t
*os
= dn
->dn_objset
;
3079 dmu_buf_impl_t
*db
, *odb
;
3081 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3082 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
3084 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
3086 list_create(&db
->db_dirty_records
, sizeof (dbuf_dirty_record_t
),
3087 offsetof(dbuf_dirty_record_t
, dr_dbuf_node
));
3090 db
->db
.db_object
= dn
->dn_object
;
3091 db
->db_level
= level
;
3092 db
->db_blkid
= blkid
;
3093 db
->db_dirtycnt
= 0;
3094 db
->db_dnode_handle
= dn
->dn_handle
;
3095 db
->db_parent
= parent
;
3096 db
->db_blkptr
= blkptr
;
3099 db
->db_user_immediate_evict
= FALSE
;
3100 db
->db_freed_in_flight
= FALSE
;
3101 db
->db_pending_evict
= FALSE
;
3103 if (blkid
== DMU_BONUS_BLKID
) {
3104 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
3105 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
3106 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
3107 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
3108 db
->db
.db_offset
= DMU_BONUS_BLKID
;
3109 db
->db_state
= DB_UNCACHED
;
3110 DTRACE_SET_STATE(db
, "bonus buffer created");
3111 db
->db_caching_status
= DB_NO_CACHE
;
3112 /* the bonus dbuf is not placed in the hash table */
3113 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3115 } else if (blkid
== DMU_SPILL_BLKID
) {
3116 db
->db
.db_size
= (blkptr
!= NULL
) ?
3117 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
3118 db
->db
.db_offset
= 0;
3121 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
3122 db
->db
.db_size
= blocksize
;
3123 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
3127 * Hold the dn_dbufs_mtx while we get the new dbuf
3128 * in the hash table *and* added to the dbufs list.
3129 * This prevents a possible deadlock with someone
3130 * trying to look up this dbuf before it's added to the
3133 mutex_enter(&dn
->dn_dbufs_mtx
);
3134 db
->db_state
= DB_EVICTING
; /* not worth logging this state change */
3135 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
3136 /* someone else inserted it first */
3137 mutex_exit(&dn
->dn_dbufs_mtx
);
3138 kmem_cache_free(dbuf_kmem_cache
, db
);
3139 DBUF_STAT_BUMP(hash_insert_race
);
3142 avl_add(&dn
->dn_dbufs
, db
);
3144 db
->db_state
= DB_UNCACHED
;
3145 DTRACE_SET_STATE(db
, "regular buffer created");
3146 db
->db_caching_status
= DB_NO_CACHE
;
3147 mutex_exit(&dn
->dn_dbufs_mtx
);
3148 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
3150 if (parent
&& parent
!= dn
->dn_dbuf
)
3151 dbuf_add_ref(parent
, db
);
3153 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
3154 zfs_refcount_count(&dn
->dn_holds
) > 0);
3155 (void) zfs_refcount_add(&dn
->dn_holds
, db
);
3157 dprintf_dbuf(db
, "db=%p\n", db
);
3163 * This function returns a block pointer and information about the object,
3164 * given a dnode and a block. This is a publicly accessible version of
3165 * dbuf_findbp that only returns some information, rather than the
3166 * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock
3167 * should be locked as (at least) a reader.
3170 dbuf_dnode_findbp(dnode_t
*dn
, uint64_t level
, uint64_t blkid
,
3171 blkptr_t
*bp
, uint16_t *datablkszsec
, uint8_t *indblkshift
)
3173 dmu_buf_impl_t
*dbp
= NULL
;
3176 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3178 err
= dbuf_findbp(dn
, level
, blkid
, B_FALSE
, &dbp
, &bp2
);
3182 dbuf_rele(dbp
, NULL
);
3183 if (datablkszsec
!= NULL
)
3184 *datablkszsec
= dn
->dn_phys
->dn_datablkszsec
;
3185 if (indblkshift
!= NULL
)
3186 *indblkshift
= dn
->dn_phys
->dn_indblkshift
;
3192 typedef struct dbuf_prefetch_arg
{
3193 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
3194 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
3195 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
3196 int dpa_curlevel
; /* The current level that we're reading */
3197 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
3198 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
3199 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
3200 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
3201 dbuf_prefetch_fn dpa_cb
; /* prefetch completion callback */
3202 void *dpa_arg
; /* prefetch completion arg */
3203 } dbuf_prefetch_arg_t
;
3206 dbuf_prefetch_fini(dbuf_prefetch_arg_t
*dpa
, boolean_t io_done
)
3208 if (dpa
->dpa_cb
!= NULL
) {
3209 dpa
->dpa_cb(dpa
->dpa_arg
, dpa
->dpa_zb
.zb_level
,
3210 dpa
->dpa_zb
.zb_blkid
, io_done
);
3212 kmem_free(dpa
, sizeof (*dpa
));
3216 dbuf_issue_final_prefetch_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3217 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3219 (void) zio
, (void) zb
, (void) iobp
;
3220 dbuf_prefetch_arg_t
*dpa
= private;
3223 arc_buf_destroy(abuf
, private);
3225 dbuf_prefetch_fini(dpa
, B_TRUE
);
3229 * Actually issue the prefetch read for the block given.
3232 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
3234 ASSERT(!BP_IS_REDACTED(bp
) ||
3235 dsl_dataset_feature_is_active(
3236 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3237 SPA_FEATURE_REDACTED_DATASETS
));
3239 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
) || BP_IS_REDACTED(bp
))
3240 return (dbuf_prefetch_fini(dpa
, B_FALSE
));
3242 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
3243 arc_flags_t aflags
=
3244 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
|
3247 /* dnodes are always read as raw and then converted later */
3248 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
3249 dpa
->dpa_curlevel
== 0)
3250 zio_flags
|= ZIO_FLAG_RAW
;
3252 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3253 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
3254 ASSERT(dpa
->dpa_zio
!= NULL
);
3255 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
,
3256 dbuf_issue_final_prefetch_done
, dpa
,
3257 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
3261 * Called when an indirect block above our prefetch target is read in. This
3262 * will either read in the next indirect block down the tree or issue the actual
3263 * prefetch if the next block down is our target.
3266 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
3267 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
3269 (void) zb
, (void) iobp
;
3270 dbuf_prefetch_arg_t
*dpa
= private;
3272 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
3273 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
3276 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
3277 dbuf_prefetch_fini(dpa
, B_TRUE
);
3280 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
3283 * The dpa_dnode is only valid if we are called with a NULL
3284 * zio. This indicates that the arc_read() returned without
3285 * first calling zio_read() to issue a physical read. Once
3286 * a physical read is made the dpa_dnode must be invalidated
3287 * as the locks guarding it may have been dropped. If the
3288 * dpa_dnode is still valid, then we want to add it to the dbuf
3289 * cache. To do so, we must hold the dbuf associated with the block
3290 * we just prefetched, read its contents so that we associate it
3291 * with an arc_buf_t, and then release it.
3294 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
3295 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
3296 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
3298 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
3300 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
3302 dpa
->dpa_dnode
= NULL
;
3303 } else if (dpa
->dpa_dnode
!= NULL
) {
3304 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
3305 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
3306 dpa
->dpa_zb
.zb_level
));
3307 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
3308 dpa
->dpa_curlevel
, curblkid
, FTAG
);
3310 arc_buf_destroy(abuf
, private);
3311 dbuf_prefetch_fini(dpa
, B_TRUE
);
3314 (void) dbuf_read(db
, NULL
,
3315 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
3316 dbuf_rele(db
, FTAG
);
3319 dpa
->dpa_curlevel
--;
3320 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
3321 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
3322 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
3323 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
3325 ASSERT(!BP_IS_REDACTED(bp
) ||
3326 dsl_dataset_feature_is_active(
3327 dpa
->dpa_dnode
->dn_objset
->os_dsl_dataset
,
3328 SPA_FEATURE_REDACTED_DATASETS
));
3329 if (BP_IS_HOLE(bp
) || BP_IS_REDACTED(bp
)) {
3330 arc_buf_destroy(abuf
, private);
3331 dbuf_prefetch_fini(dpa
, B_TRUE
);
3333 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
3334 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
3335 dbuf_issue_final_prefetch(dpa
, bp
);
3337 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3338 zbookmark_phys_t zb
;
3340 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3341 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
3342 iter_aflags
|= ARC_FLAG_L2CACHE
;
3344 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
3346 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
3347 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
3349 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3350 bp
, dbuf_prefetch_indirect_done
, dpa
,
3351 ZIO_PRIORITY_SYNC_READ
,
3352 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3356 arc_buf_destroy(abuf
, private);
3360 * Issue prefetch reads for the given block on the given level. If the indirect
3361 * blocks above that block are not in memory, we will read them in
3362 * asynchronously. As a result, this call never blocks waiting for a read to
3363 * complete. Note that the prefetch might fail if the dataset is encrypted and
3364 * the encryption key is unmapped before the IO completes.
3367 dbuf_prefetch_impl(dnode_t
*dn
, int64_t level
, uint64_t blkid
,
3368 zio_priority_t prio
, arc_flags_t aflags
, dbuf_prefetch_fn cb
,
3372 int epbs
, nlevels
, curlevel
;
3375 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3376 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3378 if (blkid
> dn
->dn_maxblkid
)
3381 if (level
== 0 && dnode_block_freed(dn
, blkid
))
3385 * This dnode hasn't been written to disk yet, so there's nothing to
3388 nlevels
= dn
->dn_phys
->dn_nlevels
;
3389 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
3392 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3393 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
3396 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
3399 mutex_exit(&db
->db_mtx
);
3401 * This dbuf already exists. It is either CACHED, or
3402 * (we assume) about to be read or filled.
3408 * Find the closest ancestor (indirect block) of the target block
3409 * that is present in the cache. In this indirect block, we will
3410 * find the bp that is at curlevel, curblkid.
3414 while (curlevel
< nlevels
- 1) {
3415 int parent_level
= curlevel
+ 1;
3416 uint64_t parent_blkid
= curblkid
>> epbs
;
3419 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
3420 FALSE
, TRUE
, FTAG
, &db
) == 0) {
3421 blkptr_t
*bpp
= db
->db_buf
->b_data
;
3422 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
3423 dbuf_rele(db
, FTAG
);
3427 curlevel
= parent_level
;
3428 curblkid
= parent_blkid
;
3431 if (curlevel
== nlevels
- 1) {
3432 /* No cached indirect blocks found. */
3433 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
3434 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
3436 ASSERT(!BP_IS_REDACTED(&bp
) ||
3437 dsl_dataset_feature_is_active(dn
->dn_objset
->os_dsl_dataset
,
3438 SPA_FEATURE_REDACTED_DATASETS
));
3439 if (BP_IS_HOLE(&bp
) || BP_IS_REDACTED(&bp
))
3442 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
3444 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
3447 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
3448 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
3449 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3450 dn
->dn_object
, level
, blkid
);
3451 dpa
->dpa_curlevel
= curlevel
;
3452 dpa
->dpa_prio
= prio
;
3453 dpa
->dpa_aflags
= aflags
;
3454 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
3455 dpa
->dpa_dnode
= dn
;
3456 dpa
->dpa_epbs
= epbs
;
3461 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3462 if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3463 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
3466 * If we have the indirect just above us, no need to do the asynchronous
3467 * prefetch chain; we'll just run the last step ourselves. If we're at
3468 * a higher level, though, we want to issue the prefetches for all the
3469 * indirect blocks asynchronously, so we can go on with whatever we were
3472 if (curlevel
== level
) {
3473 ASSERT3U(curblkid
, ==, blkid
);
3474 dbuf_issue_final_prefetch(dpa
, &bp
);
3476 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3477 zbookmark_phys_t zb
;
3479 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3480 if (dnode_level_is_l2cacheable(&bp
, dn
, level
))
3481 iter_aflags
|= ARC_FLAG_L2CACHE
;
3483 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3484 dn
->dn_object
, curlevel
, curblkid
);
3485 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3486 &bp
, dbuf_prefetch_indirect_done
, dpa
,
3487 ZIO_PRIORITY_SYNC_READ
,
3488 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3492 * We use pio here instead of dpa_zio since it's possible that
3493 * dpa may have already been freed.
3499 cb(arg
, level
, blkid
, B_FALSE
);
3504 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
3508 return (dbuf_prefetch_impl(dn
, level
, blkid
, prio
, aflags
, NULL
, NULL
));
3512 * Helper function for dbuf_hold_impl() to copy a buffer. Handles
3513 * the case of encrypted, compressed and uncompressed buffers by
3514 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3515 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3517 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl().
3519 noinline
static void
3520 dbuf_hold_copy(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3522 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3523 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
3524 enum zio_compress compress_type
= arc_get_compression(data
);
3525 uint8_t complevel
= arc_get_complevel(data
);
3527 if (arc_is_encrypted(data
)) {
3528 boolean_t byteorder
;
3529 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3530 uint8_t iv
[ZIO_DATA_IV_LEN
];
3531 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3533 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
3534 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
3535 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
3536 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
3537 compress_type
, complevel
));
3538 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3539 dbuf_set_data(db
, arc_alloc_compressed_buf(
3540 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
3541 arc_buf_lsize(data
), compress_type
, complevel
));
3543 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
3544 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
3547 rw_enter(&db
->db_rwlock
, RW_WRITER
);
3548 memcpy(db
->db
.db_data
, data
->b_data
, arc_buf_size(data
));
3549 rw_exit(&db
->db_rwlock
);
3553 * Returns with db_holds incremented, and db_mtx not held.
3554 * Note: dn_struct_rwlock must be held.
3557 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3558 boolean_t fail_sparse
, boolean_t fail_uncached
,
3559 const void *tag
, dmu_buf_impl_t
**dbp
)
3561 dmu_buf_impl_t
*db
, *parent
= NULL
;
3563 /* If the pool has been created, verify the tx_sync_lock is not held */
3564 spa_t
*spa
= dn
->dn_objset
->os_spa
;
3565 dsl_pool_t
*dp
= spa
->spa_dsl_pool
;
3567 ASSERT(!MUTEX_HELD(&dp
->dp_tx
.tx_sync_lock
));
3570 ASSERT(blkid
!= DMU_BONUS_BLKID
);
3571 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
3572 ASSERT3U(dn
->dn_nlevels
, >, level
);
3576 /* dbuf_find() returns with db_mtx held */
3577 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
);
3580 blkptr_t
*bp
= NULL
;
3584 return (SET_ERROR(ENOENT
));
3586 ASSERT3P(parent
, ==, NULL
);
3587 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
3589 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
3590 err
= SET_ERROR(ENOENT
);
3593 dbuf_rele(parent
, NULL
);
3597 if (err
&& err
!= ENOENT
)
3599 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
);
3602 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
3603 mutex_exit(&db
->db_mtx
);
3604 return (SET_ERROR(ENOENT
));
3607 if (db
->db_buf
!= NULL
) {
3608 arc_buf_access(db
->db_buf
);
3609 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
3612 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
3615 * If this buffer is currently syncing out, and we are
3616 * still referencing it from db_data, we need to make a copy
3617 * of it in case we decide we want to dirty it again in this txg.
3619 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
3620 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3621 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
3622 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
3623 if (dr
->dt
.dl
.dr_data
== db
->db_buf
)
3624 dbuf_hold_copy(dn
, db
);
3627 if (multilist_link_active(&db
->db_cache_link
)) {
3628 ASSERT(zfs_refcount_is_zero(&db
->db_holds
));
3629 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
3630 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3632 multilist_remove(&dbuf_caches
[db
->db_caching_status
].cache
, db
);
3633 (void) zfs_refcount_remove_many(
3634 &dbuf_caches
[db
->db_caching_status
].size
,
3635 db
->db
.db_size
, db
);
3637 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3638 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3640 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
3641 DBUF_STAT_BUMPDOWN(cache_count
);
3642 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
3645 db
->db_caching_status
= DB_NO_CACHE
;
3647 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3649 mutex_exit(&db
->db_mtx
);
3651 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3653 dbuf_rele(parent
, NULL
);
3655 ASSERT3P(DB_DNODE(db
), ==, dn
);
3656 ASSERT3U(db
->db_blkid
, ==, blkid
);
3657 ASSERT3U(db
->db_level
, ==, level
);
3664 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, const void *tag
)
3666 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3670 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, const void *tag
)
3673 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3674 return (err
? NULL
: db
);
3678 dbuf_create_bonus(dnode_t
*dn
)
3680 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3682 ASSERT(dn
->dn_bonus
== NULL
);
3683 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
3687 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3689 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3691 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3692 return (SET_ERROR(ENOTSUP
));
3694 blksz
= SPA_MINBLOCKSIZE
;
3695 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3696 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3698 dbuf_new_size(db
, blksz
, tx
);
3704 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3706 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3709 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3711 dbuf_add_ref(dmu_buf_impl_t
*db
, const void *tag
)
3713 int64_t holds
= zfs_refcount_add(&db
->db_holds
, tag
);
3714 VERIFY3S(holds
, >, 1);
3717 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3719 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3722 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3723 dmu_buf_impl_t
*found_db
;
3724 boolean_t result
= B_FALSE
;
3726 if (blkid
== DMU_BONUS_BLKID
)
3727 found_db
= dbuf_find_bonus(os
, obj
);
3729 found_db
= dbuf_find(os
, obj
, 0, blkid
);
3731 if (found_db
!= NULL
) {
3732 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3733 (void) zfs_refcount_add(&db
->db_holds
, tag
);
3736 mutex_exit(&found_db
->db_mtx
);
3742 * If you call dbuf_rele() you had better not be referencing the dnode handle
3743 * unless you have some other direct or indirect hold on the dnode. (An indirect
3744 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3745 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3746 * dnode's parent dbuf evicting its dnode handles.
3749 dbuf_rele(dmu_buf_impl_t
*db
, const void *tag
)
3751 mutex_enter(&db
->db_mtx
);
3752 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
3756 dmu_buf_rele(dmu_buf_t
*db
, const void *tag
)
3758 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3762 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3763 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3764 * argument should be set if we are already in the dbuf-evicting code
3765 * path, in which case we don't want to recursively evict. This allows us to
3766 * avoid deeply nested stacks that would have a call flow similar to this:
3768 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3771 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3775 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, const void *tag
, boolean_t evicting
)
3780 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3784 * Remove the reference to the dbuf before removing its hold on the
3785 * dnode so we can guarantee in dnode_move() that a referenced bonus
3786 * buffer has a corresponding dnode hold.
3788 holds
= zfs_refcount_remove(&db
->db_holds
, tag
);
3792 * We can't freeze indirects if there is a possibility that they
3793 * may be modified in the current syncing context.
3795 if (db
->db_buf
!= NULL
&&
3796 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3797 arc_buf_freeze(db
->db_buf
);
3800 if (holds
== db
->db_dirtycnt
&&
3801 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3802 dbuf_evict_user(db
);
3805 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3807 boolean_t evict_dbuf
= db
->db_pending_evict
;
3810 * If the dnode moves here, we cannot cross this
3811 * barrier until the move completes.
3816 atomic_dec_32(&dn
->dn_dbufs_count
);
3819 * Decrementing the dbuf count means that the bonus
3820 * buffer's dnode hold is no longer discounted in
3821 * dnode_move(). The dnode cannot move until after
3822 * the dnode_rele() below.
3827 * Do not reference db after its lock is dropped.
3828 * Another thread may evict it.
3830 mutex_exit(&db
->db_mtx
);
3833 dnode_evict_bonus(dn
);
3836 } else if (db
->db_buf
== NULL
) {
3838 * This is a special case: we never associated this
3839 * dbuf with any data allocated from the ARC.
3841 ASSERT(db
->db_state
== DB_UNCACHED
||
3842 db
->db_state
== DB_NOFILL
);
3844 } else if (arc_released(db
->db_buf
)) {
3846 * This dbuf has anonymous data associated with it.
3850 boolean_t do_arc_evict
= B_FALSE
;
3852 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3854 if (!DBUF_IS_CACHEABLE(db
) &&
3855 db
->db_blkptr
!= NULL
&&
3856 !BP_IS_HOLE(db
->db_blkptr
) &&
3857 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
3858 do_arc_evict
= B_TRUE
;
3859 bp
= *db
->db_blkptr
;
3862 if (!DBUF_IS_CACHEABLE(db
) ||
3863 db
->db_pending_evict
) {
3865 } else if (!multilist_link_active(&db
->db_cache_link
)) {
3866 ASSERT3U(db
->db_caching_status
, ==,
3869 dbuf_cached_state_t dcs
=
3870 dbuf_include_in_metadata_cache(db
) ?
3871 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
3872 db
->db_caching_status
= dcs
;
3874 multilist_insert(&dbuf_caches
[dcs
].cache
, db
);
3875 uint64_t db_size
= db
->db
.db_size
;
3876 size
= zfs_refcount_add_many(
3877 &dbuf_caches
[dcs
].size
, db_size
, db
);
3878 uint8_t db_level
= db
->db_level
;
3879 mutex_exit(&db
->db_mtx
);
3881 if (dcs
== DB_DBUF_METADATA_CACHE
) {
3882 DBUF_STAT_BUMP(metadata_cache_count
);
3884 metadata_cache_size_bytes_max
,
3887 DBUF_STAT_BUMP(cache_count
);
3888 DBUF_STAT_MAX(cache_size_bytes_max
,
3890 DBUF_STAT_BUMP(cache_levels
[db_level
]);
3892 cache_levels_bytes
[db_level
],
3896 if (dcs
== DB_DBUF_CACHE
&& !evicting
)
3897 dbuf_evict_notify(size
);
3901 arc_freed(spa
, &bp
);
3904 mutex_exit(&db
->db_mtx
);
3909 #pragma weak dmu_buf_refcount = dbuf_refcount
3911 dbuf_refcount(dmu_buf_impl_t
*db
)
3913 return (zfs_refcount_count(&db
->db_holds
));
3917 dmu_buf_user_refcount(dmu_buf_t
*db_fake
)
3920 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3922 mutex_enter(&db
->db_mtx
);
3923 ASSERT3U(zfs_refcount_count(&db
->db_holds
), >=, db
->db_dirtycnt
);
3924 holds
= zfs_refcount_count(&db
->db_holds
) - db
->db_dirtycnt
;
3925 mutex_exit(&db
->db_mtx
);
3931 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3932 dmu_buf_user_t
*new_user
)
3934 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3936 mutex_enter(&db
->db_mtx
);
3937 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3938 if (db
->db_user
== old_user
)
3939 db
->db_user
= new_user
;
3941 old_user
= db
->db_user
;
3942 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3943 mutex_exit(&db
->db_mtx
);
3949 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3951 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3955 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3957 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3959 db
->db_user_immediate_evict
= TRUE
;
3960 return (dmu_buf_set_user(db_fake
, user
));
3964 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3966 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3970 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3972 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3974 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3975 return (db
->db_user
);
3979 dmu_buf_user_evict_wait(void)
3981 taskq_wait(dbu_evict_taskq
);
3985 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3987 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3988 return (dbi
->db_blkptr
);
3992 dmu_buf_get_objset(dmu_buf_t
*db
)
3994 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3995 return (dbi
->db_objset
);
3999 dmu_buf_dnode_enter(dmu_buf_t
*db
)
4001 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4002 DB_DNODE_ENTER(dbi
);
4003 return (DB_DNODE(dbi
));
4007 dmu_buf_dnode_exit(dmu_buf_t
*db
)
4009 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
4014 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
4016 /* ASSERT(dmu_tx_is_syncing(tx) */
4017 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4019 if (db
->db_blkptr
!= NULL
)
4022 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4023 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
4024 BP_ZERO(db
->db_blkptr
);
4027 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
4029 * This buffer was allocated at a time when there was
4030 * no available blkptrs from the dnode, or it was
4031 * inappropriate to hook it in (i.e., nlevels mismatch).
4033 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
4034 ASSERT(db
->db_parent
== NULL
);
4035 db
->db_parent
= dn
->dn_dbuf
;
4036 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
4039 dmu_buf_impl_t
*parent
= db
->db_parent
;
4040 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4042 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
4043 if (parent
== NULL
) {
4044 mutex_exit(&db
->db_mtx
);
4045 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
4046 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
4047 db
->db_blkid
>> epbs
, db
);
4048 rw_exit(&dn
->dn_struct_rwlock
);
4049 mutex_enter(&db
->db_mtx
);
4050 db
->db_parent
= parent
;
4052 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
4053 (db
->db_blkid
& ((1ULL << epbs
) - 1));
4059 dbuf_sync_bonus(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4061 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4062 void *data
= dr
->dt
.dl
.dr_data
;
4064 ASSERT0(db
->db_level
);
4065 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4066 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
);
4067 ASSERT(data
!= NULL
);
4069 dnode_t
*dn
= dr
->dr_dnode
;
4070 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
4071 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
4072 memcpy(DN_BONUS(dn
->dn_phys
), data
, DN_MAX_BONUS_LEN(dn
->dn_phys
));
4074 dbuf_sync_leaf_verify_bonus_dnode(dr
);
4076 dbuf_undirty_bonus(dr
);
4077 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4081 * When syncing out a blocks of dnodes, adjust the block to deal with
4082 * encryption. Normally, we make sure the block is decrypted before writing
4083 * it. If we have crypt params, then we are writing a raw (encrypted) block,
4084 * from a raw receive. In this case, set the ARC buf's crypt params so
4085 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
4088 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t
*dr
)
4091 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4093 ASSERT(MUTEX_HELD(&db
->db_mtx
));
4094 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
4095 ASSERT3U(db
->db_level
, ==, 0);
4097 if (!db
->db_objset
->os_raw_receive
&& arc_is_encrypted(db
->db_buf
)) {
4098 zbookmark_phys_t zb
;
4101 * Unfortunately, there is currently no mechanism for
4102 * syncing context to handle decryption errors. An error
4103 * here is only possible if an attacker maliciously
4104 * changed a dnode block and updated the associated
4105 * checksums going up the block tree.
4107 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
4108 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4109 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
4112 panic("Invalid dnode block MAC");
4113 } else if (dr
->dt
.dl
.dr_has_raw_params
) {
4114 (void) arc_release(dr
->dt
.dl
.dr_data
, db
);
4115 arc_convert_to_raw(dr
->dt
.dl
.dr_data
,
4116 dmu_objset_id(db
->db_objset
),
4117 dr
->dt
.dl
.dr_byteorder
, DMU_OT_DNODE
,
4118 dr
->dt
.dl
.dr_salt
, dr
->dt
.dl
.dr_iv
, dr
->dt
.dl
.dr_mac
);
4123 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
4124 * is critical the we not allow the compiler to inline this function in to
4125 * dbuf_sync_list() thereby drastically bloating the stack usage.
4127 noinline
static void
4128 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4130 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4131 dnode_t
*dn
= dr
->dr_dnode
;
4133 ASSERT(dmu_tx_is_syncing(tx
));
4135 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4137 mutex_enter(&db
->db_mtx
);
4139 ASSERT(db
->db_level
> 0);
4142 /* Read the block if it hasn't been read yet. */
4143 if (db
->db_buf
== NULL
) {
4144 mutex_exit(&db
->db_mtx
);
4145 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
4146 mutex_enter(&db
->db_mtx
);
4148 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
4149 ASSERT(db
->db_buf
!= NULL
);
4151 /* Indirect block size must match what the dnode thinks it is. */
4152 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4153 dbuf_check_blkptr(dn
, db
);
4155 /* Provide the pending dirty record to child dbufs */
4156 db
->db_data_pending
= dr
;
4158 mutex_exit(&db
->db_mtx
);
4160 dbuf_write(dr
, db
->db_buf
, tx
);
4162 zio_t
*zio
= dr
->dr_zio
;
4163 mutex_enter(&dr
->dt
.di
.dr_mtx
);
4164 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
4165 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4166 mutex_exit(&dr
->dt
.di
.dr_mtx
);
4171 * Verify that the size of the data in our bonus buffer does not exceed
4172 * its recorded size.
4174 * The purpose of this verification is to catch any cases in development
4175 * where the size of a phys structure (i.e space_map_phys_t) grows and,
4176 * due to incorrect feature management, older pools expect to read more
4177 * data even though they didn't actually write it to begin with.
4179 * For a example, this would catch an error in the feature logic where we
4180 * open an older pool and we expect to write the space map histogram of
4181 * a space map with size SPACE_MAP_SIZE_V0.
4184 dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t
*dr
)
4187 dnode_t
*dn
= dr
->dr_dnode
;
4190 * Encrypted bonus buffers can have data past their bonuslen.
4191 * Skip the verification of these blocks.
4193 if (DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))
4196 uint16_t bonuslen
= dn
->dn_phys
->dn_bonuslen
;
4197 uint16_t maxbonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
4198 ASSERT3U(bonuslen
, <=, maxbonuslen
);
4200 arc_buf_t
*datap
= dr
->dt
.dl
.dr_data
;
4201 char *datap_end
= ((char *)datap
) + bonuslen
;
4202 char *datap_max
= ((char *)datap
) + maxbonuslen
;
4204 /* ensure that everything is zero after our data */
4205 for (; datap_end
< datap_max
; datap_end
++)
4206 ASSERT(*datap_end
== 0);
4211 dbuf_lightweight_bp(dbuf_dirty_record_t
*dr
)
4213 /* This must be a lightweight dirty record. */
4214 ASSERT3P(dr
->dr_dbuf
, ==, NULL
);
4215 dnode_t
*dn
= dr
->dr_dnode
;
4217 if (dn
->dn_phys
->dn_nlevels
== 1) {
4218 VERIFY3U(dr
->dt
.dll
.dr_blkid
, <, dn
->dn_phys
->dn_nblkptr
);
4219 return (&dn
->dn_phys
->dn_blkptr
[dr
->dt
.dll
.dr_blkid
]);
4221 dmu_buf_impl_t
*parent_db
= dr
->dr_parent
->dr_dbuf
;
4222 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4223 VERIFY3U(parent_db
->db_level
, ==, 1);
4224 VERIFY3P(parent_db
->db_dnode_handle
->dnh_dnode
, ==, dn
);
4225 VERIFY3U(dr
->dt
.dll
.dr_blkid
>> epbs
, ==, parent_db
->db_blkid
);
4226 blkptr_t
*bp
= parent_db
->db
.db_data
;
4227 return (&bp
[dr
->dt
.dll
.dr_blkid
& ((1 << epbs
) - 1)]);
4232 dbuf_lightweight_ready(zio_t
*zio
)
4234 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4235 blkptr_t
*bp
= zio
->io_bp
;
4237 if (zio
->io_error
!= 0)
4240 dnode_t
*dn
= dr
->dr_dnode
;
4242 blkptr_t
*bp_orig
= dbuf_lightweight_bp(dr
);
4243 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4244 int64_t delta
= bp_get_dsize_sync(spa
, bp
) -
4245 bp_get_dsize_sync(spa
, bp_orig
);
4246 dnode_diduse_space(dn
, delta
);
4248 uint64_t blkid
= dr
->dt
.dll
.dr_blkid
;
4249 mutex_enter(&dn
->dn_mtx
);
4250 if (blkid
> dn
->dn_phys
->dn_maxblkid
) {
4251 ASSERT0(dn
->dn_objset
->os_raw_receive
);
4252 dn
->dn_phys
->dn_maxblkid
= blkid
;
4254 mutex_exit(&dn
->dn_mtx
);
4256 if (!BP_IS_EMBEDDED(bp
)) {
4257 uint64_t fill
= BP_IS_HOLE(bp
) ? 0 : 1;
4258 BP_SET_FILL(bp
, fill
);
4261 dmu_buf_impl_t
*parent_db
;
4262 EQUIV(dr
->dr_parent
== NULL
, dn
->dn_phys
->dn_nlevels
== 1);
4263 if (dr
->dr_parent
== NULL
) {
4264 parent_db
= dn
->dn_dbuf
;
4266 parent_db
= dr
->dr_parent
->dr_dbuf
;
4268 rw_enter(&parent_db
->db_rwlock
, RW_WRITER
);
4270 rw_exit(&parent_db
->db_rwlock
);
4274 dbuf_lightweight_physdone(zio_t
*zio
)
4276 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4277 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
4278 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
4281 * The callback will be called io_phys_children times. Retire one
4282 * portion of our dirty space each time we are called. Any rounding
4283 * error will be cleaned up by dbuf_lightweight_done().
4285 int delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
4286 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
4290 dbuf_lightweight_done(zio_t
*zio
)
4292 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4294 VERIFY0(zio
->io_error
);
4296 objset_t
*os
= dr
->dr_dnode
->dn_objset
;
4297 dmu_tx_t
*tx
= os
->os_synctx
;
4299 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4300 ASSERT(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4302 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4303 (void) dsl_dataset_block_kill(ds
, &zio
->io_bp_orig
, tx
, B_TRUE
);
4304 dsl_dataset_block_born(ds
, zio
->io_bp
, tx
);
4308 * See comment in dbuf_write_done().
4310 if (zio
->io_phys_children
== 0) {
4311 dsl_pool_undirty_space(dmu_objset_pool(os
),
4312 dr
->dr_accounted
, zio
->io_txg
);
4314 dsl_pool_undirty_space(dmu_objset_pool(os
),
4315 dr
->dr_accounted
% zio
->io_phys_children
, zio
->io_txg
);
4318 abd_free(dr
->dt
.dll
.dr_abd
);
4319 kmem_free(dr
, sizeof (*dr
));
4322 noinline
static void
4323 dbuf_sync_lightweight(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4325 dnode_t
*dn
= dr
->dr_dnode
;
4327 if (dn
->dn_phys
->dn_nlevels
== 1) {
4330 pio
= dr
->dr_parent
->dr_zio
;
4333 zbookmark_phys_t zb
= {
4334 .zb_objset
= dmu_objset_id(dn
->dn_objset
),
4335 .zb_object
= dn
->dn_object
,
4337 .zb_blkid
= dr
->dt
.dll
.dr_blkid
,
4341 * See comment in dbuf_write(). This is so that zio->io_bp_orig
4342 * will have the old BP in dbuf_lightweight_done().
4344 dr
->dr_bp_copy
= *dbuf_lightweight_bp(dr
);
4346 dr
->dr_zio
= zio_write(pio
, dmu_objset_spa(dn
->dn_objset
),
4347 dmu_tx_get_txg(tx
), &dr
->dr_bp_copy
, dr
->dt
.dll
.dr_abd
,
4348 dn
->dn_datablksz
, abd_get_size(dr
->dt
.dll
.dr_abd
),
4349 &dr
->dt
.dll
.dr_props
, dbuf_lightweight_ready
, NULL
,
4350 dbuf_lightweight_physdone
, dbuf_lightweight_done
, dr
,
4351 ZIO_PRIORITY_ASYNC_WRITE
,
4352 ZIO_FLAG_MUSTSUCCEED
| dr
->dt
.dll
.dr_flags
, &zb
);
4354 zio_nowait(dr
->dr_zio
);
4358 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
4359 * critical the we not allow the compiler to inline this function in to
4360 * dbuf_sync_list() thereby drastically bloating the stack usage.
4362 noinline
static void
4363 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
4365 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
4366 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4367 dnode_t
*dn
= dr
->dr_dnode
;
4369 uint64_t txg
= tx
->tx_txg
;
4371 ASSERT(dmu_tx_is_syncing(tx
));
4373 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
4375 mutex_enter(&db
->db_mtx
);
4377 * To be synced, we must be dirtied. But we
4378 * might have been freed after the dirty.
4380 if (db
->db_state
== DB_UNCACHED
) {
4381 /* This buffer has been freed since it was dirtied */
4382 ASSERT(db
->db
.db_data
== NULL
);
4383 } else if (db
->db_state
== DB_FILL
) {
4384 /* This buffer was freed and is now being re-filled */
4385 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
4387 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
4391 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4392 mutex_enter(&dn
->dn_mtx
);
4393 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
4395 * In the previous transaction group, the bonus buffer
4396 * was entirely used to store the attributes for the
4397 * dnode which overrode the dn_spill field. However,
4398 * when adding more attributes to the file a spill
4399 * block was required to hold the extra attributes.
4401 * Make sure to clear the garbage left in the dn_spill
4402 * field from the previous attributes in the bonus
4403 * buffer. Otherwise, after writing out the spill
4404 * block to the new allocated dva, it will free
4405 * the old block pointed to by the invalid dn_spill.
4407 db
->db_blkptr
= NULL
;
4409 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
4410 mutex_exit(&dn
->dn_mtx
);
4414 * If this is a bonus buffer, simply copy the bonus data into the
4415 * dnode. It will be written out when the dnode is synced (and it
4416 * will be synced, since it must have been dirty for dbuf_sync to
4419 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
4420 ASSERT(dr
->dr_dbuf
== db
);
4421 dbuf_sync_bonus(dr
, tx
);
4428 * This function may have dropped the db_mtx lock allowing a dmu_sync
4429 * operation to sneak in. As a result, we need to ensure that we
4430 * don't check the dr_override_state until we have returned from
4431 * dbuf_check_blkptr.
4433 dbuf_check_blkptr(dn
, db
);
4436 * If this buffer is in the middle of an immediate write,
4437 * wait for the synchronous IO to complete.
4439 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
4440 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
4441 cv_wait(&db
->db_changed
, &db
->db_mtx
);
4442 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
4446 * If this is a dnode block, ensure it is appropriately encrypted
4447 * or decrypted, depending on what we are writing to it this txg.
4449 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
4450 dbuf_prepare_encrypted_dnode_leaf(dr
);
4452 if (db
->db_state
!= DB_NOFILL
&&
4453 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
4454 zfs_refcount_count(&db
->db_holds
) > 1 &&
4455 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
4456 *datap
== db
->db_buf
) {
4458 * If this buffer is currently "in use" (i.e., there
4459 * are active holds and db_data still references it),
4460 * then make a copy before we start the write so that
4461 * any modifications from the open txg will not leak
4464 * NOTE: this copy does not need to be made for
4465 * objects only modified in the syncing context (e.g.
4466 * DNONE_DNODE blocks).
4468 int psize
= arc_buf_size(*datap
);
4469 int lsize
= arc_buf_lsize(*datap
);
4470 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
4471 enum zio_compress compress_type
= arc_get_compression(*datap
);
4472 uint8_t complevel
= arc_get_complevel(*datap
);
4474 if (arc_is_encrypted(*datap
)) {
4475 boolean_t byteorder
;
4476 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4477 uint8_t iv
[ZIO_DATA_IV_LEN
];
4478 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4480 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
4481 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
4482 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
4483 dn
->dn_type
, psize
, lsize
, compress_type
,
4485 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
4486 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
4487 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
4488 psize
, lsize
, compress_type
, complevel
);
4490 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
4492 memcpy((*datap
)->b_data
, db
->db
.db_data
, psize
);
4494 db
->db_data_pending
= dr
;
4496 mutex_exit(&db
->db_mtx
);
4498 dbuf_write(dr
, *datap
, tx
);
4500 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4501 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
4502 list_insert_tail(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
4504 zio_nowait(dr
->dr_zio
);
4509 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
4511 dbuf_dirty_record_t
*dr
;
4513 while ((dr
= list_head(list
))) {
4514 if (dr
->dr_zio
!= NULL
) {
4516 * If we find an already initialized zio then we
4517 * are processing the meta-dnode, and we have finished.
4518 * The dbufs for all dnodes are put back on the list
4519 * during processing, so that we can zio_wait()
4520 * these IOs after initiating all child IOs.
4522 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
4523 DMU_META_DNODE_OBJECT
);
4526 list_remove(list
, dr
);
4527 if (dr
->dr_dbuf
== NULL
) {
4528 dbuf_sync_lightweight(dr
, tx
);
4530 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
4531 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
4532 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
4534 if (dr
->dr_dbuf
->db_level
> 0)
4535 dbuf_sync_indirect(dr
, tx
);
4537 dbuf_sync_leaf(dr
, tx
);
4543 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4546 dmu_buf_impl_t
*db
= vdb
;
4548 blkptr_t
*bp
= zio
->io_bp
;
4549 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4550 spa_t
*spa
= zio
->io_spa
;
4555 ASSERT3P(db
->db_blkptr
, !=, NULL
);
4556 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
4560 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
4561 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
4562 zio
->io_prev_space_delta
= delta
;
4564 if (bp
->blk_birth
!= 0) {
4565 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
4566 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
4567 (db
->db_blkid
== DMU_SPILL_BLKID
&&
4568 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
4569 BP_IS_EMBEDDED(bp
));
4570 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
4573 mutex_enter(&db
->db_mtx
);
4576 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4577 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4578 ASSERT(!(BP_IS_HOLE(bp
)) &&
4579 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4583 if (db
->db_level
== 0) {
4584 mutex_enter(&dn
->dn_mtx
);
4585 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
4586 db
->db_blkid
!= DMU_SPILL_BLKID
) {
4587 ASSERT0(db
->db_objset
->os_raw_receive
);
4588 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
4590 mutex_exit(&dn
->dn_mtx
);
4592 if (dn
->dn_type
== DMU_OT_DNODE
) {
4594 while (i
< db
->db
.db_size
) {
4596 (void *)(((char *)db
->db
.db_data
) + i
);
4598 i
+= DNODE_MIN_SIZE
;
4599 if (dnp
->dn_type
!= DMU_OT_NONE
) {
4601 i
+= dnp
->dn_extra_slots
*
4606 if (BP_IS_HOLE(bp
)) {
4613 blkptr_t
*ibp
= db
->db
.db_data
;
4614 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4615 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
4616 if (BP_IS_HOLE(ibp
))
4618 fill
+= BP_GET_FILL(ibp
);
4623 if (!BP_IS_EMBEDDED(bp
))
4624 BP_SET_FILL(bp
, fill
);
4626 mutex_exit(&db
->db_mtx
);
4628 db_lock_type_t dblt
= dmu_buf_lock_parent(db
, RW_WRITER
, FTAG
);
4629 *db
->db_blkptr
= *bp
;
4630 dmu_buf_unlock_parent(db
, dblt
, FTAG
);
4634 * This function gets called just prior to running through the compression
4635 * stage of the zio pipeline. If we're an indirect block comprised of only
4636 * holes, then we want this indirect to be compressed away to a hole. In
4637 * order to do that we must zero out any information about the holes that
4638 * this indirect points to prior to before we try to compress it.
4641 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4643 (void) zio
, (void) buf
;
4644 dmu_buf_impl_t
*db
= vdb
;
4647 unsigned int epbs
, i
;
4649 ASSERT3U(db
->db_level
, >, 0);
4652 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4653 ASSERT3U(epbs
, <, 31);
4655 /* Determine if all our children are holes */
4656 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
4657 if (!BP_IS_HOLE(bp
))
4662 * If all the children are holes, then zero them all out so that
4663 * we may get compressed away.
4665 if (i
== 1ULL << epbs
) {
4667 * We only found holes. Grab the rwlock to prevent
4668 * anybody from reading the blocks we're about to
4671 rw_enter(&db
->db_rwlock
, RW_WRITER
);
4672 memset(db
->db
.db_data
, 0, db
->db
.db_size
);
4673 rw_exit(&db
->db_rwlock
);
4679 * The SPA will call this callback several times for each zio - once
4680 * for every physical child i/o (zio->io_phys_children times). This
4681 * allows the DMU to monitor the progress of each logical i/o. For example,
4682 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
4683 * block. There may be a long delay before all copies/fragments are completed,
4684 * so this callback allows us to retire dirty space gradually, as the physical
4688 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4691 dmu_buf_impl_t
*db
= arg
;
4692 objset_t
*os
= db
->db_objset
;
4693 dsl_pool_t
*dp
= dmu_objset_pool(os
);
4694 dbuf_dirty_record_t
*dr
;
4697 dr
= db
->db_data_pending
;
4698 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
4701 * The callback will be called io_phys_children times. Retire one
4702 * portion of our dirty space each time we are called. Any rounding
4703 * error will be cleaned up by dbuf_write_done().
4705 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
4706 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
4710 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4713 dmu_buf_impl_t
*db
= vdb
;
4714 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4715 blkptr_t
*bp
= db
->db_blkptr
;
4716 objset_t
*os
= db
->db_objset
;
4717 dmu_tx_t
*tx
= os
->os_synctx
;
4719 ASSERT0(zio
->io_error
);
4720 ASSERT(db
->db_blkptr
== bp
);
4723 * For nopwrites and rewrites we ensure that the bp matches our
4724 * original and bypass all the accounting.
4726 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4727 ASSERT(BP_EQUAL(bp
, bp_orig
));
4729 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4730 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
4731 dsl_dataset_block_born(ds
, bp
, tx
);
4734 mutex_enter(&db
->db_mtx
);
4738 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
4739 dnode_t
*dn
= dr
->dr_dnode
;
4740 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4741 ASSERT(dr
->dr_dbuf
== db
);
4742 ASSERT(list_next(&db
->db_dirty_records
, dr
) == NULL
);
4743 list_remove(&db
->db_dirty_records
, dr
);
4746 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4747 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4748 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
4749 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4753 if (db
->db_level
== 0) {
4754 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
4755 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
4756 if (db
->db_state
!= DB_NOFILL
) {
4757 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
4758 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
4761 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4762 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
4763 if (!BP_IS_HOLE(db
->db_blkptr
)) {
4764 int epbs __maybe_unused
= dn
->dn_phys
->dn_indblkshift
-
4766 ASSERT3U(db
->db_blkid
, <=,
4767 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
4768 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4771 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4772 list_destroy(&dr
->dt
.di
.dr_children
);
4775 cv_broadcast(&db
->db_changed
);
4776 ASSERT(db
->db_dirtycnt
> 0);
4777 db
->db_dirtycnt
-= 1;
4778 db
->db_data_pending
= NULL
;
4779 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4782 * If we didn't do a physical write in this ZIO and we
4783 * still ended up here, it means that the space of the
4784 * dbuf that we just released (and undirtied) above hasn't
4785 * been marked as undirtied in the pool's accounting.
4787 * Thus, we undirty that space in the pool's view of the
4788 * world here. For physical writes this type of update
4789 * happens in dbuf_write_physdone().
4791 * If we did a physical write, cleanup any rounding errors
4792 * that came up due to writing multiple copies of a block
4793 * on disk [see dbuf_write_physdone()].
4795 if (zio
->io_phys_children
== 0) {
4796 dsl_pool_undirty_space(dmu_objset_pool(os
),
4797 dr
->dr_accounted
, zio
->io_txg
);
4799 dsl_pool_undirty_space(dmu_objset_pool(os
),
4800 dr
->dr_accounted
% zio
->io_phys_children
, zio
->io_txg
);
4803 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4807 dbuf_write_nofill_ready(zio_t
*zio
)
4809 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4813 dbuf_write_nofill_done(zio_t
*zio
)
4815 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4819 dbuf_write_override_ready(zio_t
*zio
)
4821 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4822 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4824 dbuf_write_ready(zio
, NULL
, db
);
4828 dbuf_write_override_done(zio_t
*zio
)
4830 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4831 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4832 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4834 mutex_enter(&db
->db_mtx
);
4835 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4836 if (!BP_IS_HOLE(obp
))
4837 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4838 arc_release(dr
->dt
.dl
.dr_data
, db
);
4840 mutex_exit(&db
->db_mtx
);
4842 dbuf_write_done(zio
, NULL
, db
);
4844 if (zio
->io_abd
!= NULL
)
4845 abd_free(zio
->io_abd
);
4848 typedef struct dbuf_remap_impl_callback_arg
{
4850 uint64_t drica_blk_birth
;
4852 } dbuf_remap_impl_callback_arg_t
;
4855 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
4858 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
4859 objset_t
*os
= drica
->drica_os
;
4860 spa_t
*spa
= dmu_objset_spa(os
);
4861 dmu_tx_t
*tx
= drica
->drica_tx
;
4863 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4865 if (os
== spa_meta_objset(spa
)) {
4866 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
4868 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
4869 size
, drica
->drica_blk_birth
, tx
);
4874 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, krwlock_t
*rw
, dmu_tx_t
*tx
)
4876 blkptr_t bp_copy
= *bp
;
4877 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4878 dbuf_remap_impl_callback_arg_t drica
;
4880 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4882 drica
.drica_os
= dn
->dn_objset
;
4883 drica
.drica_blk_birth
= bp
->blk_birth
;
4884 drica
.drica_tx
= tx
;
4885 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
4888 * If the blkptr being remapped is tracked by a livelist,
4889 * then we need to make sure the livelist reflects the update.
4890 * First, cancel out the old blkptr by appending a 'FREE'
4891 * entry. Next, add an 'ALLOC' to track the new version. This
4892 * way we avoid trying to free an inaccurate blkptr at delete.
4893 * Note that embedded blkptrs are not tracked in livelists.
4895 if (dn
->dn_objset
!= spa_meta_objset(spa
)) {
4896 dsl_dataset_t
*ds
= dmu_objset_ds(dn
->dn_objset
);
4897 if (dsl_deadlist_is_open(&ds
->ds_dir
->dd_livelist
) &&
4898 bp
->blk_birth
> ds
->ds_dir
->dd_origin_txg
) {
4899 ASSERT(!BP_IS_EMBEDDED(bp
));
4900 ASSERT(dsl_dir_is_clone(ds
->ds_dir
));
4901 ASSERT(spa_feature_is_enabled(spa
,
4902 SPA_FEATURE_LIVELIST
));
4903 bplist_append(&ds
->ds_dir
->dd_pending_frees
,
4905 bplist_append(&ds
->ds_dir
->dd_pending_allocs
,
4911 * The db_rwlock prevents dbuf_read_impl() from
4912 * dereferencing the BP while we are changing it. To
4913 * avoid lock contention, only grab it when we are actually
4917 rw_enter(rw
, RW_WRITER
);
4925 * Remap any existing BP's to concrete vdevs, if possible.
4928 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
4930 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
4931 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4933 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
4936 if (db
->db_level
> 0) {
4937 blkptr_t
*bp
= db
->db
.db_data
;
4938 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
4939 dbuf_remap_impl(dn
, &bp
[i
], &db
->db_rwlock
, tx
);
4941 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
4942 dnode_phys_t
*dnp
= db
->db
.db_data
;
4943 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
4945 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
;
4946 i
+= dnp
[i
].dn_extra_slots
+ 1) {
4947 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
4948 krwlock_t
*lock
= (dn
->dn_dbuf
== NULL
? NULL
:
4949 &dn
->dn_dbuf
->db_rwlock
);
4950 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], lock
,
4958 /* Issue I/O to commit a dirty buffer to disk. */
4960 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
4962 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4963 dnode_t
*dn
= dr
->dr_dnode
;
4965 dmu_buf_impl_t
*parent
= db
->db_parent
;
4966 uint64_t txg
= tx
->tx_txg
;
4967 zbookmark_phys_t zb
;
4969 zio_t
*pio
; /* parent I/O */
4972 ASSERT(dmu_tx_is_syncing(tx
));
4976 if (db
->db_state
!= DB_NOFILL
) {
4977 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
4979 * Private object buffers are released here rather
4980 * than in dbuf_dirty() since they are only modified
4981 * in the syncing context and we don't want the
4982 * overhead of making multiple copies of the data.
4984 if (BP_IS_HOLE(db
->db_blkptr
)) {
4987 dbuf_release_bp(db
);
4989 dbuf_remap(dn
, db
, tx
);
4993 if (parent
!= dn
->dn_dbuf
) {
4994 /* Our parent is an indirect block. */
4995 /* We have a dirty parent that has been scheduled for write. */
4996 ASSERT(parent
&& parent
->db_data_pending
);
4997 /* Our parent's buffer is one level closer to the dnode. */
4998 ASSERT(db
->db_level
== parent
->db_level
-1);
5000 * We're about to modify our parent's db_data by modifying
5001 * our block pointer, so the parent must be released.
5003 ASSERT(arc_released(parent
->db_buf
));
5004 pio
= parent
->db_data_pending
->dr_zio
;
5006 /* Our parent is the dnode itself. */
5007 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
5008 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
5009 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
5010 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
5011 ASSERT3P(db
->db_blkptr
, ==,
5012 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
5016 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
5017 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
5020 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
5021 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
5022 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
5024 if (db
->db_blkid
== DMU_SPILL_BLKID
)
5026 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
5028 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
5031 * We copy the blkptr now (rather than when we instantiate the dirty
5032 * record), because its value can change between open context and
5033 * syncing context. We do not need to hold dn_struct_rwlock to read
5034 * db_blkptr because we are in syncing context.
5036 dr
->dr_bp_copy
= *db
->db_blkptr
;
5038 if (db
->db_level
== 0 &&
5039 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
5041 * The BP for this block has been provided by open context
5042 * (by dmu_sync() or dmu_buf_write_embedded()).
5044 abd_t
*contents
= (data
!= NULL
) ?
5045 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
5047 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
, &dr
->dr_bp_copy
,
5048 contents
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5049 dbuf_write_override_ready
, NULL
, NULL
,
5050 dbuf_write_override_done
,
5051 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
5052 mutex_enter(&db
->db_mtx
);
5053 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
5054 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
5055 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
5056 mutex_exit(&db
->db_mtx
);
5057 } else if (db
->db_state
== DB_NOFILL
) {
5058 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
5059 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
5060 dr
->dr_zio
= zio_write(pio
, os
->os_spa
, txg
,
5061 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
5062 dbuf_write_nofill_ready
, NULL
, NULL
,
5063 dbuf_write_nofill_done
, db
,
5064 ZIO_PRIORITY_ASYNC_WRITE
,
5065 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
5067 ASSERT(arc_released(data
));
5070 * For indirect blocks, we want to setup the children
5071 * ready callback so that we can properly handle an indirect
5072 * block that only contains holes.
5074 arc_write_done_func_t
*children_ready_cb
= NULL
;
5075 if (db
->db_level
!= 0)
5076 children_ready_cb
= dbuf_write_children_ready
;
5078 dr
->dr_zio
= arc_write(pio
, os
->os_spa
, txg
,
5079 &dr
->dr_bp_copy
, data
, dbuf_is_l2cacheable(db
),
5080 &zp
, dbuf_write_ready
,
5081 children_ready_cb
, dbuf_write_physdone
,
5082 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
5083 ZIO_FLAG_MUSTSUCCEED
, &zb
);
5087 EXPORT_SYMBOL(dbuf_find
);
5088 EXPORT_SYMBOL(dbuf_is_metadata
);
5089 EXPORT_SYMBOL(dbuf_destroy
);
5090 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
5091 EXPORT_SYMBOL(dbuf_whichblock
);
5092 EXPORT_SYMBOL(dbuf_read
);
5093 EXPORT_SYMBOL(dbuf_unoverride
);
5094 EXPORT_SYMBOL(dbuf_free_range
);
5095 EXPORT_SYMBOL(dbuf_new_size
);
5096 EXPORT_SYMBOL(dbuf_release_bp
);
5097 EXPORT_SYMBOL(dbuf_dirty
);
5098 EXPORT_SYMBOL(dmu_buf_set_crypt_params
);
5099 EXPORT_SYMBOL(dmu_buf_will_dirty
);
5100 EXPORT_SYMBOL(dmu_buf_is_dirty
);
5101 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
5102 EXPORT_SYMBOL(dmu_buf_will_fill
);
5103 EXPORT_SYMBOL(dmu_buf_fill_done
);
5104 EXPORT_SYMBOL(dmu_buf_rele
);
5105 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
5106 EXPORT_SYMBOL(dbuf_prefetch
);
5107 EXPORT_SYMBOL(dbuf_hold_impl
);
5108 EXPORT_SYMBOL(dbuf_hold
);
5109 EXPORT_SYMBOL(dbuf_hold_level
);
5110 EXPORT_SYMBOL(dbuf_create_bonus
);
5111 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
5112 EXPORT_SYMBOL(dbuf_rm_spill
);
5113 EXPORT_SYMBOL(dbuf_add_ref
);
5114 EXPORT_SYMBOL(dbuf_rele
);
5115 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
5116 EXPORT_SYMBOL(dbuf_refcount
);
5117 EXPORT_SYMBOL(dbuf_sync_list
);
5118 EXPORT_SYMBOL(dmu_buf_set_user
);
5119 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
5120 EXPORT_SYMBOL(dmu_buf_get_user
);
5121 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
5123 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, max_bytes
, ULONG
, ZMOD_RW
,
5124 "Maximum size in bytes of the dbuf cache.");
5126 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, hiwater_pct
, UINT
, ZMOD_RW
,
5127 "Percentage over dbuf_cache_max_bytes for direct dbuf eviction.");
5129 ZFS_MODULE_PARAM(zfs_dbuf_cache
, dbuf_cache_
, lowater_pct
, UINT
, ZMOD_RW
,
5130 "Percentage below dbuf_cache_max_bytes when dbuf eviction stops.");
5132 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_max_bytes
, ULONG
, ZMOD_RW
,
5133 "Maximum size in bytes of dbuf metadata cache.");
5135 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, cache_shift
, UINT
, ZMOD_RW
,
5136 "Set size of dbuf cache to log2 fraction of arc size.");
5138 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, metadata_cache_shift
, UINT
, ZMOD_RW
,
5139 "Set size of dbuf metadata cache to log2 fraction of arc size.");
5141 ZFS_MODULE_PARAM(zfs_dbuf
, dbuf_
, mutex_cache_shift
, UINT
, ZMOD_RD
,
5142 "Set size of dbuf cache mutex array as log2 shift.");