4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_dbuf.h>
48 #include <sys/callb.h>
51 #include <sys/cityhash.h>
52 #include <sys/spa_impl.h>
56 typedef struct dbuf_stats
{
58 * Various statistics about the size of the dbuf cache.
60 kstat_named_t cache_count
;
61 kstat_named_t cache_size_bytes
;
62 kstat_named_t cache_size_bytes_max
;
64 * Statistics regarding the bounds on the dbuf cache size.
66 kstat_named_t cache_target_bytes
;
67 kstat_named_t cache_lowater_bytes
;
68 kstat_named_t cache_hiwater_bytes
;
70 * Total number of dbuf cache evictions that have occurred.
72 kstat_named_t cache_total_evicts
;
74 * The distribution of dbuf levels in the dbuf cache and
75 * the total size of all dbufs at each level.
77 kstat_named_t cache_levels
[DN_MAX_LEVELS
];
78 kstat_named_t cache_levels_bytes
[DN_MAX_LEVELS
];
80 * Statistics about the dbuf hash table.
82 kstat_named_t hash_hits
;
83 kstat_named_t hash_misses
;
84 kstat_named_t hash_collisions
;
85 kstat_named_t hash_elements
;
86 kstat_named_t hash_elements_max
;
88 * Number of sublists containing more than one dbuf in the dbuf
89 * hash table. Keep track of the longest hash chain.
91 kstat_named_t hash_chains
;
92 kstat_named_t hash_chain_max
;
94 * Number of times a dbuf_create() discovers that a dbuf was
95 * already created and in the dbuf hash table.
97 kstat_named_t hash_insert_race
;
99 * Statistics about the size of the metadata dbuf cache.
101 kstat_named_t metadata_cache_count
;
102 kstat_named_t metadata_cache_size_bytes
;
103 kstat_named_t metadata_cache_size_bytes_max
;
105 * For diagnostic purposes, this is incremented whenever we can't add
106 * something to the metadata cache because it's full, and instead put
107 * the data in the regular dbuf cache.
109 kstat_named_t metadata_cache_overflow
;
112 dbuf_stats_t dbuf_stats
= {
113 { "cache_count", KSTAT_DATA_UINT64
},
114 { "cache_size_bytes", KSTAT_DATA_UINT64
},
115 { "cache_size_bytes_max", KSTAT_DATA_UINT64
},
116 { "cache_target_bytes", KSTAT_DATA_UINT64
},
117 { "cache_lowater_bytes", KSTAT_DATA_UINT64
},
118 { "cache_hiwater_bytes", KSTAT_DATA_UINT64
},
119 { "cache_total_evicts", KSTAT_DATA_UINT64
},
120 { { "cache_levels_N", KSTAT_DATA_UINT64
} },
121 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64
} },
122 { "hash_hits", KSTAT_DATA_UINT64
},
123 { "hash_misses", KSTAT_DATA_UINT64
},
124 { "hash_collisions", KSTAT_DATA_UINT64
},
125 { "hash_elements", KSTAT_DATA_UINT64
},
126 { "hash_elements_max", KSTAT_DATA_UINT64
},
127 { "hash_chains", KSTAT_DATA_UINT64
},
128 { "hash_chain_max", KSTAT_DATA_UINT64
},
129 { "hash_insert_race", KSTAT_DATA_UINT64
},
130 { "metadata_cache_count", KSTAT_DATA_UINT64
},
131 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64
},
132 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64
},
133 { "metadata_cache_overflow", KSTAT_DATA_UINT64
}
136 #define DBUF_STAT_INCR(stat, val) \
137 atomic_add_64(&dbuf_stats.stat.value.ui64, (val));
138 #define DBUF_STAT_DECR(stat, val) \
139 DBUF_STAT_INCR(stat, -(val));
140 #define DBUF_STAT_BUMP(stat) \
141 DBUF_STAT_INCR(stat, 1);
142 #define DBUF_STAT_BUMPDOWN(stat) \
143 DBUF_STAT_INCR(stat, -1);
144 #define DBUF_STAT_MAX(stat, v) { \
146 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
147 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
151 typedef struct dbuf_hold_arg
{
152 /* Function arguments */
156 boolean_t dh_fail_sparse
;
157 boolean_t dh_fail_uncached
;
159 dmu_buf_impl_t
**dh_dbp
;
160 /* Local variables */
161 dmu_buf_impl_t
*dh_db
;
162 dmu_buf_impl_t
*dh_parent
;
165 dbuf_dirty_record_t
*dh_dr
;
168 static dbuf_hold_arg_t
*dbuf_hold_arg_create(dnode_t
*dn
, uint8_t level
,
169 uint64_t blkid
, boolean_t fail_sparse
, boolean_t fail_uncached
,
170 void *tag
, dmu_buf_impl_t
**dbp
);
171 static int dbuf_hold_impl_arg(dbuf_hold_arg_t
*dh
);
172 static void dbuf_hold_arg_destroy(dbuf_hold_arg_t
*dh
);
174 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
175 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
177 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
178 dmu_buf_evict_func_t
*evict_func_sync
,
179 dmu_buf_evict_func_t
*evict_func_async
,
180 dmu_buf_t
**clear_on_evict_dbufp
);
183 * Global data structures and functions for the dbuf cache.
185 static kmem_cache_t
*dbuf_kmem_cache
;
186 static taskq_t
*dbu_evict_taskq
;
188 static kthread_t
*dbuf_cache_evict_thread
;
189 static kmutex_t dbuf_evict_lock
;
190 static kcondvar_t dbuf_evict_cv
;
191 static boolean_t dbuf_evict_thread_exit
;
194 * There are two dbuf caches; each dbuf can only be in one of them at a time.
196 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
197 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
198 * that represent the metadata that describes filesystems/snapshots/
199 * bookmarks/properties/etc. We only evict from this cache when we export a
200 * pool, to short-circuit as much I/O as possible for all administrative
201 * commands that need the metadata. There is no eviction policy for this
202 * cache, because we try to only include types in it which would occupy a
203 * very small amount of space per object but create a large impact on the
204 * performance of these commands. Instead, after it reaches a maximum size
205 * (which should only happen on very small memory systems with a very large
206 * number of filesystem objects), we stop taking new dbufs into the
207 * metadata cache, instead putting them in the normal dbuf cache.
209 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
210 * are not currently held but have been recently released. These dbufs
211 * are not eligible for arc eviction until they are aged out of the cache.
212 * Dbufs that are aged out of the cache will be immediately destroyed and
213 * become eligible for arc eviction.
215 * Dbufs are added to these caches once the last hold is released. If a dbuf is
216 * later accessed and still exists in the dbuf cache, then it will be removed
217 * from the cache and later re-added to the head of the cache.
219 * If a given dbuf meets the requirements for the metadata cache, it will go
220 * there, otherwise it will be considered for the generic LRU dbuf cache. The
221 * caches and the refcounts tracking their sizes are stored in an array indexed
222 * by those caches' matching enum values (from dbuf_cached_state_t).
224 typedef struct dbuf_cache
{
228 dbuf_cache_t dbuf_caches
[DB_CACHE_MAX
];
230 /* Size limits for the caches */
231 unsigned long dbuf_cache_max_bytes
= 0;
232 unsigned long dbuf_metadata_cache_max_bytes
= 0;
233 /* Set the default sizes of the caches to log2 fraction of arc size */
234 int dbuf_cache_shift
= 5;
235 int dbuf_metadata_cache_shift
= 6;
238 * The LRU dbuf cache uses a three-stage eviction policy:
239 * - A low water marker designates when the dbuf eviction thread
240 * should stop evicting from the dbuf cache.
241 * - When we reach the maximum size (aka mid water mark), we
242 * signal the eviction thread to run.
243 * - The high water mark indicates when the eviction thread
244 * is unable to keep up with the incoming load and eviction must
245 * happen in the context of the calling thread.
249 * low water mid water hi water
250 * +----------------------------------------+----------+----------+
255 * +----------------------------------------+----------+----------+
257 * evicting eviction directly
260 * The high and low water marks indicate the operating range for the eviction
261 * thread. The low water mark is, by default, 90% of the total size of the
262 * cache and the high water mark is at 110% (both of these percentages can be
263 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
264 * respectively). The eviction thread will try to ensure that the cache remains
265 * within this range by waking up every second and checking if the cache is
266 * above the low water mark. The thread can also be woken up by callers adding
267 * elements into the cache if the cache is larger than the mid water (i.e max
268 * cache size). Once the eviction thread is woken up and eviction is required,
269 * it will continue evicting buffers until it's able to reduce the cache size
270 * to the low water mark. If the cache size continues to grow and hits the high
271 * water mark, then callers adding elements to the cache will begin to evict
272 * directly from the cache until the cache is no longer above the high water
277 * The percentage above and below the maximum cache size.
279 uint_t dbuf_cache_hiwater_pct
= 10;
280 uint_t dbuf_cache_lowater_pct
= 10;
284 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
286 dmu_buf_impl_t
*db
= vdb
;
287 bzero(db
, sizeof (dmu_buf_impl_t
));
289 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
290 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
291 multilist_link_init(&db
->db_cache_link
);
292 refcount_create(&db
->db_holds
);
299 dbuf_dest(void *vdb
, void *unused
)
301 dmu_buf_impl_t
*db
= vdb
;
302 mutex_destroy(&db
->db_mtx
);
303 cv_destroy(&db
->db_changed
);
304 ASSERT(!multilist_link_active(&db
->db_cache_link
));
305 refcount_destroy(&db
->db_holds
);
309 * dbuf hash table routines
311 static dbuf_hash_table_t dbuf_hash_table
;
313 static uint64_t dbuf_hash_count
;
316 * We use Cityhash for this. It's fast, and has good hash properties without
317 * requiring any large static buffers.
320 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
322 return (cityhash4((uintptr_t)os
, obj
, (uint64_t)lvl
, blkid
));
325 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
326 ((dbuf)->db.db_object == (obj) && \
327 (dbuf)->db_objset == (os) && \
328 (dbuf)->db_level == (level) && \
329 (dbuf)->db_blkid == (blkid))
332 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
334 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
339 hv
= dbuf_hash(os
, obj
, level
, blkid
);
340 idx
= hv
& h
->hash_table_mask
;
342 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
343 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
344 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
345 mutex_enter(&db
->db_mtx
);
346 if (db
->db_state
!= DB_EVICTING
) {
347 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
350 mutex_exit(&db
->db_mtx
);
353 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
357 static dmu_buf_impl_t
*
358 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
361 dmu_buf_impl_t
*db
= NULL
;
363 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
364 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
365 if (dn
->dn_bonus
!= NULL
) {
367 mutex_enter(&db
->db_mtx
);
369 rw_exit(&dn
->dn_struct_rwlock
);
370 dnode_rele(dn
, FTAG
);
376 * Insert an entry into the hash table. If there is already an element
377 * equal to elem in the hash table, then the already existing element
378 * will be returned and the new element will not be inserted.
379 * Otherwise returns NULL.
381 static dmu_buf_impl_t
*
382 dbuf_hash_insert(dmu_buf_impl_t
*db
)
384 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
385 objset_t
*os
= db
->db_objset
;
386 uint64_t obj
= db
->db
.db_object
;
387 int level
= db
->db_level
;
388 uint64_t blkid
, hv
, idx
;
392 blkid
= db
->db_blkid
;
393 hv
= dbuf_hash(os
, obj
, level
, blkid
);
394 idx
= hv
& h
->hash_table_mask
;
396 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
397 for (dbf
= h
->hash_table
[idx
], i
= 0; dbf
!= NULL
;
398 dbf
= dbf
->db_hash_next
, i
++) {
399 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
400 mutex_enter(&dbf
->db_mtx
);
401 if (dbf
->db_state
!= DB_EVICTING
) {
402 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
405 mutex_exit(&dbf
->db_mtx
);
410 DBUF_STAT_BUMP(hash_collisions
);
412 DBUF_STAT_BUMP(hash_chains
);
414 DBUF_STAT_MAX(hash_chain_max
, i
);
417 mutex_enter(&db
->db_mtx
);
418 db
->db_hash_next
= h
->hash_table
[idx
];
419 h
->hash_table
[idx
] = db
;
420 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
421 atomic_inc_64(&dbuf_hash_count
);
422 DBUF_STAT_MAX(hash_elements_max
, dbuf_hash_count
);
428 * This returns whether this dbuf should be stored in the metadata cache, which
429 * is based on whether it's from one of the dnode types that store data related
430 * to traversing dataset hierarchies.
433 dbuf_include_in_metadata_cache(dmu_buf_impl_t
*db
)
436 dmu_object_type_t type
= DB_DNODE(db
)->dn_type
;
439 /* Check if this dbuf is one of the types we care about */
440 if (DMU_OT_IS_METADATA_CACHED(type
)) {
441 /* If we hit this, then we set something up wrong in dmu_ot */
442 ASSERT(DMU_OT_IS_METADATA(type
));
445 * Sanity check for small-memory systems: don't allocate too
446 * much memory for this purpose.
448 if (refcount_count(&dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
) >
449 dbuf_metadata_cache_max_bytes
) {
450 DBUF_STAT_BUMP(metadata_cache_overflow
);
461 * Remove an entry from the hash table. It must be in the EVICTING state.
464 dbuf_hash_remove(dmu_buf_impl_t
*db
)
466 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
468 dmu_buf_impl_t
*dbf
, **dbp
;
470 hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
471 db
->db_level
, db
->db_blkid
);
472 idx
= hv
& h
->hash_table_mask
;
475 * We mustn't hold db_mtx to maintain lock ordering:
476 * DBUF_HASH_MUTEX > db_mtx.
478 ASSERT(refcount_is_zero(&db
->db_holds
));
479 ASSERT(db
->db_state
== DB_EVICTING
);
480 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
482 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
483 dbp
= &h
->hash_table
[idx
];
484 while ((dbf
= *dbp
) != db
) {
485 dbp
= &dbf
->db_hash_next
;
488 *dbp
= db
->db_hash_next
;
489 db
->db_hash_next
= NULL
;
490 if (h
->hash_table
[idx
] &&
491 h
->hash_table
[idx
]->db_hash_next
== NULL
)
492 DBUF_STAT_BUMPDOWN(hash_chains
);
493 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
494 atomic_dec_64(&dbuf_hash_count
);
500 } dbvu_verify_type_t
;
503 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
508 if (db
->db_user
== NULL
)
511 /* Only data blocks support the attachment of user data. */
512 ASSERT(db
->db_level
== 0);
514 /* Clients must resolve a dbuf before attaching user data. */
515 ASSERT(db
->db
.db_data
!= NULL
);
516 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
518 holds
= refcount_count(&db
->db_holds
);
519 if (verify_type
== DBVU_EVICTING
) {
521 * Immediate eviction occurs when holds == dirtycnt.
522 * For normal eviction buffers, holds is zero on
523 * eviction, except when dbuf_fix_old_data() calls
524 * dbuf_clear_data(). However, the hold count can grow
525 * during eviction even though db_mtx is held (see
526 * dmu_bonus_hold() for an example), so we can only
527 * test the generic invariant that holds >= dirtycnt.
529 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
531 if (db
->db_user_immediate_evict
== TRUE
)
532 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
534 ASSERT3U(holds
, >, 0);
540 dbuf_evict_user(dmu_buf_impl_t
*db
)
542 dmu_buf_user_t
*dbu
= db
->db_user
;
544 ASSERT(MUTEX_HELD(&db
->db_mtx
));
549 dbuf_verify_user(db
, DBVU_EVICTING
);
553 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
554 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
558 * There are two eviction callbacks - one that we call synchronously
559 * and one that we invoke via a taskq. The async one is useful for
560 * avoiding lock order reversals and limiting stack depth.
562 * Note that if we have a sync callback but no async callback,
563 * it's likely that the sync callback will free the structure
564 * containing the dbu. In that case we need to take care to not
565 * dereference dbu after calling the sync evict func.
567 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
569 if (dbu
->dbu_evict_func_sync
!= NULL
)
570 dbu
->dbu_evict_func_sync(dbu
);
573 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
574 dbu
, 0, &dbu
->dbu_tqent
);
579 dbuf_is_metadata(dmu_buf_impl_t
*db
)
582 * Consider indirect blocks and spill blocks to be meta data.
584 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
587 boolean_t is_metadata
;
590 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
593 return (is_metadata
);
599 * This function *must* return indices evenly distributed between all
600 * sublists of the multilist. This is needed due to how the dbuf eviction
601 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
602 * distributed between all sublists and uses this assumption when
603 * deciding which sublist to evict from and how much to evict from it.
606 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
608 dmu_buf_impl_t
*db
= obj
;
611 * The assumption here, is the hash value for a given
612 * dmu_buf_impl_t will remain constant throughout it's lifetime
613 * (i.e. it's objset, object, level and blkid fields don't change).
614 * Thus, we don't need to store the dbuf's sublist index
615 * on insertion, as this index can be recalculated on removal.
617 * Also, the low order bits of the hash value are thought to be
618 * distributed evenly. Otherwise, in the case that the multilist
619 * has a power of two number of sublists, each sublists' usage
620 * would not be evenly distributed.
622 return (dbuf_hash(db
->db_objset
, db
->db
.db_object
,
623 db
->db_level
, db
->db_blkid
) %
624 multilist_get_num_sublists(ml
));
627 static inline unsigned long
628 dbuf_cache_target_bytes(void)
630 return MIN(dbuf_cache_max_bytes
,
631 arc_target_bytes() >> dbuf_cache_shift
);
634 static inline uint64_t
635 dbuf_cache_hiwater_bytes(void)
637 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
638 return (dbuf_cache_target
+
639 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100);
642 static inline uint64_t
643 dbuf_cache_lowater_bytes(void)
645 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
646 return (dbuf_cache_target
-
647 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100);
650 static inline boolean_t
651 dbuf_cache_above_hiwater(void)
653 return (refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
654 dbuf_cache_hiwater_bytes());
657 static inline boolean_t
658 dbuf_cache_above_lowater(void)
660 return (refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
661 dbuf_cache_lowater_bytes());
665 * Evict the oldest eligible dbuf from the dbuf cache.
670 int idx
= multilist_get_random_index(dbuf_caches
[DB_DBUF_CACHE
].cache
);
671 multilist_sublist_t
*mls
= multilist_sublist_lock(
672 dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
674 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
676 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
677 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
678 db
= multilist_sublist_prev(mls
, db
);
681 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
682 multilist_sublist_t
*, mls
);
685 multilist_sublist_remove(mls
, db
);
686 multilist_sublist_unlock(mls
);
687 (void) refcount_remove_many(&dbuf_caches
[DB_DBUF_CACHE
].size
,
689 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
690 DBUF_STAT_BUMPDOWN(cache_count
);
691 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
693 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
694 db
->db_caching_status
= DB_NO_CACHE
;
696 DBUF_STAT_MAX(cache_size_bytes_max
,
697 refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
));
698 DBUF_STAT_BUMP(cache_total_evicts
);
700 multilist_sublist_unlock(mls
);
705 * The dbuf evict thread is responsible for aging out dbufs from the
706 * cache. Once the cache has reached it's maximum size, dbufs are removed
707 * and destroyed. The eviction thread will continue running until the size
708 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
709 * out of the cache it is destroyed and becomes eligible for arc eviction.
713 dbuf_evict_thread(void *unused
)
717 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
719 mutex_enter(&dbuf_evict_lock
);
720 while (!dbuf_evict_thread_exit
) {
721 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
722 CALLB_CPR_SAFE_BEGIN(&cpr
);
723 (void) cv_timedwait_sig_hires(&dbuf_evict_cv
,
724 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
725 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
727 mutex_exit(&dbuf_evict_lock
);
730 * Keep evicting as long as we're above the low water mark
731 * for the cache. We do this without holding the locks to
732 * minimize lock contention.
734 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
738 mutex_enter(&dbuf_evict_lock
);
741 dbuf_evict_thread_exit
= B_FALSE
;
742 cv_broadcast(&dbuf_evict_cv
);
743 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
748 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
749 * If the dbuf cache is at its high water mark, then evict a dbuf from the
750 * dbuf cache using the callers context.
753 dbuf_evict_notify(void)
756 * We check if we should evict without holding the dbuf_evict_lock,
757 * because it's OK to occasionally make the wrong decision here,
758 * and grabbing the lock results in massive lock contention.
760 if (refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
761 dbuf_cache_target_bytes()) {
762 if (dbuf_cache_above_hiwater())
764 cv_signal(&dbuf_evict_cv
);
769 dbuf_kstat_update(kstat_t
*ksp
, int rw
)
771 dbuf_stats_t
*ds
= ksp
->ks_data
;
773 if (rw
== KSTAT_WRITE
) {
774 return (SET_ERROR(EACCES
));
776 ds
->metadata_cache_size_bytes
.value
.ui64
=
777 refcount_count(&dbuf_caches
[DB_DBUF_METADATA_CACHE
].size
);
778 ds
->cache_size_bytes
.value
.ui64
=
779 refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
);
780 ds
->cache_target_bytes
.value
.ui64
= dbuf_cache_target_bytes();
781 ds
->cache_hiwater_bytes
.value
.ui64
= dbuf_cache_hiwater_bytes();
782 ds
->cache_lowater_bytes
.value
.ui64
= dbuf_cache_lowater_bytes();
783 ds
->hash_elements
.value
.ui64
= dbuf_hash_count
;
792 uint64_t hsize
= 1ULL << 16;
793 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
797 * The hash table is big enough to fill all of physical memory
798 * with an average block size of zfs_arc_average_blocksize (default 8K).
799 * By default, the table will take up
800 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
802 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
806 h
->hash_table_mask
= hsize
- 1;
809 * Large allocations which do not require contiguous pages
810 * should be using vmem_alloc() in the linux kernel
812 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
814 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
816 if (h
->hash_table
== NULL
) {
817 /* XXX - we should really return an error instead of assert */
818 ASSERT(hsize
> (1ULL << 10));
823 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
824 sizeof (dmu_buf_impl_t
),
825 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
827 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
828 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
833 * Setup the parameters for the dbuf caches. We set the sizes of the
834 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
835 * of the target size of the ARC. If the values has been specified as
836 * a module option and they're not greater than the target size of the
837 * ARC, then we honor that value.
839 if (dbuf_cache_max_bytes
== 0 ||
840 dbuf_cache_max_bytes
>= arc_target_bytes()) {
841 dbuf_cache_max_bytes
= arc_target_bytes() >> dbuf_cache_shift
;
843 if (dbuf_metadata_cache_max_bytes
== 0 ||
844 dbuf_metadata_cache_max_bytes
>= arc_target_bytes()) {
845 dbuf_metadata_cache_max_bytes
=
846 arc_target_bytes() >> dbuf_metadata_cache_shift
;
850 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
851 * configuration is not required.
853 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
855 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
856 dbuf_caches
[dcs
].cache
=
857 multilist_create(sizeof (dmu_buf_impl_t
),
858 offsetof(dmu_buf_impl_t
, db_cache_link
),
859 dbuf_cache_multilist_index_func
);
860 refcount_create(&dbuf_caches
[dcs
].size
);
863 dbuf_evict_thread_exit
= B_FALSE
;
864 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
865 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
866 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
867 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
869 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
870 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
872 if (dbuf_ksp
!= NULL
) {
873 dbuf_ksp
->ks_data
= &dbuf_stats
;
874 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
875 kstat_install(dbuf_ksp
);
877 for (i
= 0; i
< DN_MAX_LEVELS
; i
++) {
878 snprintf(dbuf_stats
.cache_levels
[i
].name
,
879 KSTAT_STRLEN
, "cache_level_%d", i
);
880 dbuf_stats
.cache_levels
[i
].data_type
=
882 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
883 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
884 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
893 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
896 dbuf_stats_destroy();
898 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
899 mutex_destroy(&h
->hash_mutexes
[i
]);
902 * Large allocations which do not require contiguous pages
903 * should be using vmem_free() in the linux kernel
905 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
907 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
909 kmem_cache_destroy(dbuf_kmem_cache
);
910 taskq_destroy(dbu_evict_taskq
);
912 mutex_enter(&dbuf_evict_lock
);
913 dbuf_evict_thread_exit
= B_TRUE
;
914 while (dbuf_evict_thread_exit
) {
915 cv_signal(&dbuf_evict_cv
);
916 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
918 mutex_exit(&dbuf_evict_lock
);
920 mutex_destroy(&dbuf_evict_lock
);
921 cv_destroy(&dbuf_evict_cv
);
923 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
924 refcount_destroy(&dbuf_caches
[dcs
].size
);
925 multilist_destroy(dbuf_caches
[dcs
].cache
);
928 if (dbuf_ksp
!= NULL
) {
929 kstat_delete(dbuf_ksp
);
940 dbuf_verify(dmu_buf_impl_t
*db
)
943 dbuf_dirty_record_t
*dr
;
945 ASSERT(MUTEX_HELD(&db
->db_mtx
));
947 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
950 ASSERT(db
->db_objset
!= NULL
);
954 ASSERT(db
->db_parent
== NULL
);
955 ASSERT(db
->db_blkptr
== NULL
);
957 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
958 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
959 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
960 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
961 db
->db_blkid
== DMU_SPILL_BLKID
||
962 !avl_is_empty(&dn
->dn_dbufs
));
964 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
966 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
967 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
968 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
970 ASSERT0(db
->db
.db_offset
);
972 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
975 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
976 ASSERT(dr
->dr_dbuf
== db
);
978 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
979 ASSERT(dr
->dr_dbuf
== db
);
982 * We can't assert that db_size matches dn_datablksz because it
983 * can be momentarily different when another thread is doing
986 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
987 dr
= db
->db_data_pending
;
989 * It should only be modified in syncing context, so
990 * make sure we only have one copy of the data.
992 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
995 /* verify db->db_blkptr */
997 if (db
->db_parent
== dn
->dn_dbuf
) {
998 /* db is pointed to by the dnode */
999 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1000 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
1001 ASSERT(db
->db_parent
== NULL
);
1003 ASSERT(db
->db_parent
!= NULL
);
1004 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
1005 ASSERT3P(db
->db_blkptr
, ==,
1006 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
1008 /* db is pointed to by an indirect block */
1009 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
1011 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
1012 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
1015 * dnode_grow_indblksz() can make this fail if we don't
1016 * have the struct_rwlock. XXX indblksz no longer
1017 * grows. safe to do this now?
1019 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1020 ASSERT3P(db
->db_blkptr
, ==,
1021 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
1022 db
->db_blkid
% epb
));
1026 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
1027 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
1028 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
1029 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
1031 * If the blkptr isn't set but they have nonzero data,
1032 * it had better be dirty, otherwise we'll lose that
1033 * data when we evict this buffer.
1035 * There is an exception to this rule for indirect blocks; in
1036 * this case, if the indirect block is a hole, we fill in a few
1037 * fields on each of the child blocks (importantly, birth time)
1038 * to prevent hole birth times from being lost when you
1039 * partially fill in a hole.
1041 if (db
->db_dirtycnt
== 0) {
1042 if (db
->db_level
== 0) {
1043 uint64_t *buf
= db
->db
.db_data
;
1046 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
1047 ASSERT(buf
[i
] == 0);
1050 blkptr_t
*bps
= db
->db
.db_data
;
1051 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1054 * We want to verify that all the blkptrs in the
1055 * indirect block are holes, but we may have
1056 * automatically set up a few fields for them.
1057 * We iterate through each blkptr and verify
1058 * they only have those fields set.
1061 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1063 blkptr_t
*bp
= &bps
[i
];
1064 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1067 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1068 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1069 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1070 ASSERT0(bp
->blk_fill
);
1071 ASSERT0(bp
->blk_pad
[0]);
1072 ASSERT0(bp
->blk_pad
[1]);
1073 ASSERT(!BP_IS_EMBEDDED(bp
));
1074 ASSERT(BP_IS_HOLE(bp
));
1075 ASSERT0(bp
->blk_phys_birth
);
1085 dbuf_clear_data(dmu_buf_impl_t
*db
)
1087 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1088 dbuf_evict_user(db
);
1089 ASSERT3P(db
->db_buf
, ==, NULL
);
1090 db
->db
.db_data
= NULL
;
1091 if (db
->db_state
!= DB_NOFILL
)
1092 db
->db_state
= DB_UNCACHED
;
1096 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1098 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1099 ASSERT(buf
!= NULL
);
1102 ASSERT(buf
->b_data
!= NULL
);
1103 db
->db
.db_data
= buf
->b_data
;
1107 * Loan out an arc_buf for read. Return the loaned arc_buf.
1110 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1114 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1115 mutex_enter(&db
->db_mtx
);
1116 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
1117 int blksz
= db
->db
.db_size
;
1118 spa_t
*spa
= db
->db_objset
->os_spa
;
1120 mutex_exit(&db
->db_mtx
);
1121 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1122 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
1125 arc_loan_inuse_buf(abuf
, db
);
1127 dbuf_clear_data(db
);
1128 mutex_exit(&db
->db_mtx
);
1134 * Calculate which level n block references the data at the level 0 offset
1138 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1140 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1142 * The level n blkid is equal to the level 0 blkid divided by
1143 * the number of level 0s in a level n block.
1145 * The level 0 blkid is offset >> datablkshift =
1146 * offset / 2^datablkshift.
1148 * The number of level 0s in a level n is the number of block
1149 * pointers in an indirect block, raised to the power of level.
1150 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1151 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1153 * Thus, the level n blkid is: offset /
1154 * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT))))
1155 * = offset / 2^(datablkshift + level *
1156 * (indblkshift - SPA_BLKPTRSHIFT))
1157 * = offset >> (datablkshift + level *
1158 * (indblkshift - SPA_BLKPTRSHIFT))
1161 const unsigned exp
= dn
->dn_datablkshift
+
1162 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1164 if (exp
>= 8 * sizeof (offset
)) {
1165 /* This only happens on the highest indirection level */
1166 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1170 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1172 return (offset
>> exp
);
1174 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1180 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1181 arc_buf_t
*buf
, void *vdb
)
1183 dmu_buf_impl_t
*db
= vdb
;
1185 mutex_enter(&db
->db_mtx
);
1186 ASSERT3U(db
->db_state
, ==, DB_READ
);
1188 * All reads are synchronous, so we must have a hold on the dbuf
1190 ASSERT(refcount_count(&db
->db_holds
) > 0);
1191 ASSERT(db
->db_buf
== NULL
);
1192 ASSERT(db
->db
.db_data
== NULL
);
1195 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
1196 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1197 ASSERT3P(db
->db_buf
, ==, NULL
);
1198 db
->db_state
= DB_UNCACHED
;
1199 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1200 /* freed in flight */
1201 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1202 arc_release(buf
, db
);
1203 bzero(buf
->b_data
, db
->db
.db_size
);
1204 arc_buf_freeze(buf
);
1205 db
->db_freed_in_flight
= FALSE
;
1206 dbuf_set_data(db
, buf
);
1207 db
->db_state
= DB_CACHED
;
1210 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
1211 dbuf_set_data(db
, buf
);
1212 db
->db_state
= DB_CACHED
;
1214 cv_broadcast(&db
->db_changed
);
1215 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
1220 * This function ensures that, when doing a decrypting read of a block,
1221 * we make sure we have decrypted the dnode associated with it. We must do
1222 * this so that we ensure we are fully authenticating the checksum-of-MACs
1223 * tree from the root of the objset down to this block. Indirect blocks are
1224 * always verified against their secure checksum-of-MACs assuming that the
1225 * dnode containing them is correct. Now that we are doing a decrypting read,
1226 * we can be sure that the key is loaded and verify that assumption. This is
1227 * especially important considering that we always read encrypted dnode
1228 * blocks as raw data (without verifying their MACs) to start, and
1229 * decrypt / authenticate them when we need to read an encrypted bonus buffer.
1232 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t
*db
, uint32_t flags
)
1235 objset_t
*os
= db
->db_objset
;
1236 arc_buf_t
*dnode_abuf
;
1238 zbookmark_phys_t zb
;
1240 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1242 if (!os
->os_encrypted
|| os
->os_raw_receive
||
1243 (flags
& DB_RF_NO_DECRYPT
) != 0)
1248 dnode_abuf
= (dn
->dn_dbuf
!= NULL
) ? dn
->dn_dbuf
->db_buf
: NULL
;
1250 if (dnode_abuf
== NULL
|| !arc_is_encrypted(dnode_abuf
)) {
1255 SET_BOOKMARK(&zb
, dmu_objset_id(os
),
1256 DMU_META_DNODE_OBJECT
, 0, dn
->dn_dbuf
->db_blkid
);
1257 err
= arc_untransform(dnode_abuf
, os
->os_spa
, &zb
, B_TRUE
);
1260 * An error code of EACCES tells us that the key is still not
1261 * available. This is ok if we are only reading authenticated
1262 * (and therefore non-encrypted) blocks.
1264 if (err
== EACCES
&& ((db
->db_blkid
!= DMU_BONUS_BLKID
&&
1265 !DMU_OT_IS_ENCRYPTED(dn
->dn_type
)) ||
1266 (db
->db_blkid
== DMU_BONUS_BLKID
&&
1267 !DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
))))
1276 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1279 zbookmark_phys_t zb
;
1280 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1281 int err
, zio_flags
= 0;
1285 ASSERT(!refcount_is_zero(&db
->db_holds
));
1286 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1287 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1288 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1289 ASSERT(db
->db_state
== DB_UNCACHED
);
1290 ASSERT(db
->db_buf
== NULL
);
1292 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1294 * The bonus length stored in the dnode may be less than
1295 * the maximum available space in the bonus buffer.
1297 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1298 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1300 /* if the underlying dnode block is encrypted, decrypt it */
1301 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1304 mutex_exit(&db
->db_mtx
);
1308 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1309 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1310 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1311 if (bonuslen
< max_bonuslen
)
1312 bzero(db
->db
.db_data
, max_bonuslen
);
1314 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1316 db
->db_state
= DB_CACHED
;
1317 mutex_exit(&db
->db_mtx
);
1322 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1323 * processes the delete record and clears the bp while we are waiting
1324 * for the dn_mtx (resulting in a "no" from block_freed).
1326 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1327 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1328 BP_IS_HOLE(db
->db_blkptr
)))) {
1329 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1331 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1333 bzero(db
->db
.db_data
, db
->db
.db_size
);
1335 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1336 BP_IS_HOLE(db
->db_blkptr
) &&
1337 db
->db_blkptr
->blk_birth
!= 0) {
1338 blkptr_t
*bps
= db
->db
.db_data
;
1339 for (int i
= 0; i
< ((1 <<
1340 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1342 blkptr_t
*bp
= &bps
[i
];
1343 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1344 1 << dn
->dn_indblkshift
);
1346 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1348 BP_GET_LSIZE(db
->db_blkptr
));
1349 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1351 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1352 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1356 db
->db_state
= DB_CACHED
;
1357 mutex_exit(&db
->db_mtx
);
1362 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1363 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1366 * All bps of an encrypted os should have the encryption bit set.
1367 * If this is not true it indicates tampering and we report an error.
1369 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(db
->db_blkptr
)) {
1370 spa_log_error(db
->db_objset
->os_spa
, &zb
);
1371 zfs_panic_recover("unencrypted block in encrypted "
1372 "object set %llu", dmu_objset_id(db
->db_objset
));
1374 mutex_exit(&db
->db_mtx
);
1375 return (SET_ERROR(EIO
));
1378 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1381 mutex_exit(&db
->db_mtx
);
1387 db
->db_state
= DB_READ
;
1388 mutex_exit(&db
->db_mtx
);
1390 if (DBUF_IS_L2CACHEABLE(db
))
1391 aflags
|= ARC_FLAG_L2CACHE
;
1393 dbuf_add_ref(db
, NULL
);
1395 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1396 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1398 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1399 zio_flags
|= ZIO_FLAG_RAW
;
1401 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1402 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1409 * This is our just-in-time copy function. It makes a copy of buffers that
1410 * have been modified in a previous transaction group before we access them in
1411 * the current active group.
1413 * This function is used in three places: when we are dirtying a buffer for the
1414 * first time in a txg, when we are freeing a range in a dnode that includes
1415 * this buffer, and when we are accessing a buffer which was received compressed
1416 * and later referenced in a WRITE_BYREF record.
1418 * Note that when we are called from dbuf_free_range() we do not put a hold on
1419 * the buffer, we just traverse the active dbuf list for the dnode.
1422 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1424 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1426 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1427 ASSERT(db
->db
.db_data
!= NULL
);
1428 ASSERT(db
->db_level
== 0);
1429 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1432 (dr
->dt
.dl
.dr_data
!=
1433 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1437 * If the last dirty record for this dbuf has not yet synced
1438 * and its referencing the dbuf data, either:
1439 * reset the reference to point to a new copy,
1440 * or (if there a no active holders)
1441 * just null out the current db_data pointer.
1443 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1444 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1445 dnode_t
*dn
= DB_DNODE(db
);
1446 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1447 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1448 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1449 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1450 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1451 dnode_t
*dn
= DB_DNODE(db
);
1452 int size
= arc_buf_size(db
->db_buf
);
1453 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1454 spa_t
*spa
= db
->db_objset
->os_spa
;
1455 enum zio_compress compress_type
=
1456 arc_get_compression(db
->db_buf
);
1458 if (arc_is_encrypted(db
->db_buf
)) {
1459 boolean_t byteorder
;
1460 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1461 uint8_t iv
[ZIO_DATA_IV_LEN
];
1462 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1464 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1466 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1467 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1468 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1470 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1471 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1472 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1473 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1475 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1477 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1480 dbuf_clear_data(db
);
1485 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1492 * We don't have to hold the mutex to check db_state because it
1493 * can't be freed while we have a hold on the buffer.
1495 ASSERT(!refcount_is_zero(&db
->db_holds
));
1497 if (db
->db_state
== DB_NOFILL
)
1498 return (SET_ERROR(EIO
));
1502 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1503 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1505 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1506 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1507 DBUF_IS_CACHEABLE(db
);
1509 mutex_enter(&db
->db_mtx
);
1510 if (db
->db_state
== DB_CACHED
) {
1511 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1514 * Ensure that this block's dnode has been decrypted if
1515 * the caller has requested decrypted data.
1517 err
= dbuf_read_verify_dnode_crypt(db
, flags
);
1520 * If the arc buf is compressed or encrypted and the caller
1521 * requested uncompressed data, we need to untransform it
1522 * before returning. We also call arc_untransform() on any
1523 * unauthenticated blocks, which will verify their MAC if
1524 * the key is now available.
1526 if (err
== 0 && db
->db_buf
!= NULL
&&
1527 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1528 (arc_is_encrypted(db
->db_buf
) ||
1529 arc_is_unauthenticated(db
->db_buf
) ||
1530 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1531 zbookmark_phys_t zb
;
1533 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1534 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1535 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1536 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1537 dbuf_set_data(db
, db
->db_buf
);
1539 mutex_exit(&db
->db_mtx
);
1540 if (err
== 0 && prefetch
)
1541 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1542 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1543 rw_exit(&dn
->dn_struct_rwlock
);
1545 DBUF_STAT_BUMP(hash_hits
);
1546 } else if (db
->db_state
== DB_UNCACHED
) {
1547 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1548 boolean_t need_wait
= B_FALSE
;
1551 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1552 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1555 err
= dbuf_read_impl(db
, zio
, flags
);
1557 /* dbuf_read_impl has dropped db_mtx for us */
1559 if (!err
&& prefetch
)
1560 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1562 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1563 rw_exit(&dn
->dn_struct_rwlock
);
1565 DBUF_STAT_BUMP(hash_misses
);
1567 if (!err
&& need_wait
)
1568 err
= zio_wait(zio
);
1571 * Another reader came in while the dbuf was in flight
1572 * between UNCACHED and CACHED. Either a writer will finish
1573 * writing the buffer (sending the dbuf to CACHED) or the
1574 * first reader's request will reach the read_done callback
1575 * and send the dbuf to CACHED. Otherwise, a failure
1576 * occurred and the dbuf went to UNCACHED.
1578 mutex_exit(&db
->db_mtx
);
1580 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1581 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1582 rw_exit(&dn
->dn_struct_rwlock
);
1584 DBUF_STAT_BUMP(hash_misses
);
1586 /* Skip the wait per the caller's request. */
1587 mutex_enter(&db
->db_mtx
);
1588 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1589 while (db
->db_state
== DB_READ
||
1590 db
->db_state
== DB_FILL
) {
1591 ASSERT(db
->db_state
== DB_READ
||
1592 (flags
& DB_RF_HAVESTRUCT
) == 0);
1593 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1595 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1597 if (db
->db_state
== DB_UNCACHED
)
1598 err
= SET_ERROR(EIO
);
1600 mutex_exit(&db
->db_mtx
);
1607 dbuf_noread(dmu_buf_impl_t
*db
)
1609 ASSERT(!refcount_is_zero(&db
->db_holds
));
1610 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1611 mutex_enter(&db
->db_mtx
);
1612 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1613 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1614 if (db
->db_state
== DB_UNCACHED
) {
1615 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1616 spa_t
*spa
= db
->db_objset
->os_spa
;
1618 ASSERT(db
->db_buf
== NULL
);
1619 ASSERT(db
->db
.db_data
== NULL
);
1620 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1621 db
->db_state
= DB_FILL
;
1622 } else if (db
->db_state
== DB_NOFILL
) {
1623 dbuf_clear_data(db
);
1625 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1627 mutex_exit(&db
->db_mtx
);
1631 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1633 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1634 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1635 uint64_t txg
= dr
->dr_txg
;
1637 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1639 * This assert is valid because dmu_sync() expects to be called by
1640 * a zilog's get_data while holding a range lock. This call only
1641 * comes from dbuf_dirty() callers who must also hold a range lock.
1643 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1644 ASSERT(db
->db_level
== 0);
1646 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1647 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1650 ASSERT(db
->db_data_pending
!= dr
);
1652 /* free this block */
1653 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1654 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1656 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1657 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1658 dr
->dt
.dl
.dr_has_raw_params
= B_FALSE
;
1661 * Release the already-written buffer, so we leave it in
1662 * a consistent dirty state. Note that all callers are
1663 * modifying the buffer, so they will immediately do
1664 * another (redundant) arc_release(). Therefore, leave
1665 * the buf thawed to save the effort of freezing &
1666 * immediately re-thawing it.
1668 arc_release(dr
->dt
.dl
.dr_data
, db
);
1672 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1673 * data blocks in the free range, so that any future readers will find
1677 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1680 dmu_buf_impl_t
*db_search
;
1681 dmu_buf_impl_t
*db
, *db_next
;
1682 uint64_t txg
= tx
->tx_txg
;
1685 if (end_blkid
> dn
->dn_maxblkid
&&
1686 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1687 end_blkid
= dn
->dn_maxblkid
;
1688 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1690 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1691 db_search
->db_level
= 0;
1692 db_search
->db_blkid
= start_blkid
;
1693 db_search
->db_state
= DB_SEARCH
;
1695 mutex_enter(&dn
->dn_dbufs_mtx
);
1696 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1697 ASSERT3P(db
, ==, NULL
);
1699 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1701 for (; db
!= NULL
; db
= db_next
) {
1702 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1703 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1705 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1708 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1710 /* found a level 0 buffer in the range */
1711 mutex_enter(&db
->db_mtx
);
1712 if (dbuf_undirty(db
, tx
)) {
1713 /* mutex has been dropped and dbuf destroyed */
1717 if (db
->db_state
== DB_UNCACHED
||
1718 db
->db_state
== DB_NOFILL
||
1719 db
->db_state
== DB_EVICTING
) {
1720 ASSERT(db
->db
.db_data
== NULL
);
1721 mutex_exit(&db
->db_mtx
);
1724 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1725 /* will be handled in dbuf_read_done or dbuf_rele */
1726 db
->db_freed_in_flight
= TRUE
;
1727 mutex_exit(&db
->db_mtx
);
1730 if (refcount_count(&db
->db_holds
) == 0) {
1735 /* The dbuf is referenced */
1737 if (db
->db_last_dirty
!= NULL
) {
1738 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1740 if (dr
->dr_txg
== txg
) {
1742 * This buffer is "in-use", re-adjust the file
1743 * size to reflect that this buffer may
1744 * contain new data when we sync.
1746 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1747 db
->db_blkid
> dn
->dn_maxblkid
)
1748 dn
->dn_maxblkid
= db
->db_blkid
;
1749 dbuf_unoverride(dr
);
1752 * This dbuf is not dirty in the open context.
1753 * Either uncache it (if its not referenced in
1754 * the open context) or reset its contents to
1757 dbuf_fix_old_data(db
, txg
);
1760 /* clear the contents if its cached */
1761 if (db
->db_state
== DB_CACHED
) {
1762 ASSERT(db
->db
.db_data
!= NULL
);
1763 arc_release(db
->db_buf
, db
);
1764 bzero(db
->db
.db_data
, db
->db
.db_size
);
1765 arc_buf_freeze(db
->db_buf
);
1768 mutex_exit(&db
->db_mtx
);
1771 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1772 mutex_exit(&dn
->dn_dbufs_mtx
);
1776 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1778 arc_buf_t
*buf
, *obuf
;
1779 int osize
= db
->db
.db_size
;
1780 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1783 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1788 /* XXX does *this* func really need the lock? */
1789 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1792 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1793 * is OK, because there can be no other references to the db
1794 * when we are changing its size, so no concurrent DB_FILL can
1798 * XXX we should be doing a dbuf_read, checking the return
1799 * value and returning that up to our callers
1801 dmu_buf_will_dirty(&db
->db
, tx
);
1803 /* create the data buffer for the new block */
1804 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1806 /* copy old block data to the new block */
1808 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1809 /* zero the remainder */
1811 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1813 mutex_enter(&db
->db_mtx
);
1814 dbuf_set_data(db
, buf
);
1815 arc_buf_destroy(obuf
, db
);
1816 db
->db
.db_size
= size
;
1818 if (db
->db_level
== 0) {
1819 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1820 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1822 mutex_exit(&db
->db_mtx
);
1824 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1829 dbuf_release_bp(dmu_buf_impl_t
*db
)
1831 ASSERTV(objset_t
*os
= db
->db_objset
);
1833 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1834 ASSERT(arc_released(os
->os_phys_buf
) ||
1835 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1836 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1838 (void) arc_release(db
->db_buf
, db
);
1842 * We already have a dirty record for this TXG, and we are being
1846 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1848 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1850 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1852 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1854 * If this buffer has already been written out,
1855 * we now need to reset its state.
1857 dbuf_unoverride(dr
);
1858 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1859 db
->db_state
!= DB_NOFILL
) {
1860 /* Already released on initial dirty, so just thaw. */
1861 ASSERT(arc_released(db
->db_buf
));
1862 arc_buf_thaw(db
->db_buf
);
1867 dbuf_dirty_record_t
*
1868 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1872 dbuf_dirty_record_t
**drp
, *dr
;
1873 int drop_struct_lock
= FALSE
;
1874 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1876 ASSERT(tx
->tx_txg
!= 0);
1877 ASSERT(!refcount_is_zero(&db
->db_holds
));
1878 DMU_TX_DIRTY_BUF(tx
, db
);
1883 * Shouldn't dirty a regular buffer in syncing context. Private
1884 * objects may be dirtied in syncing context, but only if they
1885 * were already pre-dirtied in open context.
1888 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1889 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1892 ASSERT(!dmu_tx_is_syncing(tx
) ||
1893 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1894 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1895 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1896 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1897 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1900 * We make this assert for private objects as well, but after we
1901 * check if we're already dirty. They are allowed to re-dirty
1902 * in syncing context.
1904 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1905 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1906 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1908 mutex_enter(&db
->db_mtx
);
1910 * XXX make this true for indirects too? The problem is that
1911 * transactions created with dmu_tx_create_assigned() from
1912 * syncing context don't bother holding ahead.
1914 ASSERT(db
->db_level
!= 0 ||
1915 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1916 db
->db_state
== DB_NOFILL
);
1918 mutex_enter(&dn
->dn_mtx
);
1920 * Don't set dirtyctx to SYNC if we're just modifying this as we
1921 * initialize the objset.
1923 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1924 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1925 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1928 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1929 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1930 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1931 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1932 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1934 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1935 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1940 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
1941 dn
->dn_dirty_txg
= tx
->tx_txg
;
1942 mutex_exit(&dn
->dn_mtx
);
1944 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1945 dn
->dn_have_spill
= B_TRUE
;
1948 * If this buffer is already dirty, we're done.
1950 drp
= &db
->db_last_dirty
;
1951 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1952 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1953 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1955 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1959 mutex_exit(&db
->db_mtx
);
1964 * Only valid if not already dirty.
1966 ASSERT(dn
->dn_object
== 0 ||
1967 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1968 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1970 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1973 * We should only be dirtying in syncing context if it's the
1974 * mos or we're initializing the os or it's a special object.
1975 * However, we are allowed to dirty in syncing context provided
1976 * we already dirtied it in open context. Hence we must make
1977 * this assertion only if we're not already dirty.
1980 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
1982 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1983 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
1984 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1985 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1986 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1987 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1989 ASSERT(db
->db
.db_size
!= 0);
1991 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1993 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1994 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
1998 * If this buffer is dirty in an old transaction group we need
1999 * to make a copy of it so that the changes we make in this
2000 * transaction group won't leak out when we sync the older txg.
2002 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
2003 list_link_init(&dr
->dr_dirty_node
);
2004 if (db
->db_level
== 0) {
2005 void *data_old
= db
->db_buf
;
2007 if (db
->db_state
!= DB_NOFILL
) {
2008 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2009 dbuf_fix_old_data(db
, tx
->tx_txg
);
2010 data_old
= db
->db
.db_data
;
2011 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
2013 * Release the data buffer from the cache so
2014 * that we can modify it without impacting
2015 * possible other users of this cached data
2016 * block. Note that indirect blocks and
2017 * private objects are not released until the
2018 * syncing state (since they are only modified
2021 arc_release(db
->db_buf
, db
);
2022 dbuf_fix_old_data(db
, tx
->tx_txg
);
2023 data_old
= db
->db_buf
;
2025 ASSERT(data_old
!= NULL
);
2027 dr
->dt
.dl
.dr_data
= data_old
;
2029 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
2030 list_create(&dr
->dt
.di
.dr_children
,
2031 sizeof (dbuf_dirty_record_t
),
2032 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
2034 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
2035 dr
->dr_accounted
= db
->db
.db_size
;
2037 dr
->dr_txg
= tx
->tx_txg
;
2042 * We could have been freed_in_flight between the dbuf_noread
2043 * and dbuf_dirty. We win, as though the dbuf_noread() had
2044 * happened after the free.
2046 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2047 db
->db_blkid
!= DMU_SPILL_BLKID
) {
2048 mutex_enter(&dn
->dn_mtx
);
2049 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
2050 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
2053 mutex_exit(&dn
->dn_mtx
);
2054 db
->db_freed_in_flight
= FALSE
;
2058 * This buffer is now part of this txg
2060 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
2061 db
->db_dirtycnt
+= 1;
2062 ASSERT3U(db
->db_dirtycnt
, <=, 3);
2064 mutex_exit(&db
->db_mtx
);
2066 if (db
->db_blkid
== DMU_BONUS_BLKID
||
2067 db
->db_blkid
== DMU_SPILL_BLKID
) {
2068 mutex_enter(&dn
->dn_mtx
);
2069 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2070 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2071 mutex_exit(&dn
->dn_mtx
);
2072 dnode_setdirty(dn
, tx
);
2078 * The dn_struct_rwlock prevents db_blkptr from changing
2079 * due to a write from syncing context completing
2080 * while we are running, so we want to acquire it before
2081 * looking at db_blkptr.
2083 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
2084 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2085 drop_struct_lock
= TRUE
;
2089 * We need to hold the dn_struct_rwlock to make this assertion,
2090 * because it protects dn_phys / dn_next_nlevels from changing.
2092 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
2093 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
2094 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
2095 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
2096 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
2099 * If we are overwriting a dedup BP, then unless it is snapshotted,
2100 * when we get to syncing context we will need to decrement its
2101 * refcount in the DDT. Prefetch the relevant DDT block so that
2102 * syncing context won't have to wait for the i/o.
2104 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
2106 if (db
->db_level
== 0) {
2107 ASSERT(!db
->db_objset
->os_raw_receive
||
2108 dn
->dn_maxblkid
>= db
->db_blkid
);
2109 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
2110 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
2113 if (db
->db_level
+1 < dn
->dn_nlevels
) {
2114 dmu_buf_impl_t
*parent
= db
->db_parent
;
2115 dbuf_dirty_record_t
*di
;
2116 int parent_held
= FALSE
;
2118 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
2119 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2121 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
2122 db
->db_blkid
>> epbs
, FTAG
);
2123 ASSERT(parent
!= NULL
);
2126 if (drop_struct_lock
)
2127 rw_exit(&dn
->dn_struct_rwlock
);
2128 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
2129 di
= dbuf_dirty(parent
, tx
);
2131 dbuf_rele(parent
, FTAG
);
2133 mutex_enter(&db
->db_mtx
);
2135 * Since we've dropped the mutex, it's possible that
2136 * dbuf_undirty() might have changed this out from under us.
2138 if (db
->db_last_dirty
== dr
||
2139 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2140 mutex_enter(&di
->dt
.di
.dr_mtx
);
2141 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2142 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2143 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2144 mutex_exit(&di
->dt
.di
.dr_mtx
);
2147 mutex_exit(&db
->db_mtx
);
2149 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
2150 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2151 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2152 mutex_enter(&dn
->dn_mtx
);
2153 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2154 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2155 mutex_exit(&dn
->dn_mtx
);
2156 if (drop_struct_lock
)
2157 rw_exit(&dn
->dn_struct_rwlock
);
2160 dnode_setdirty(dn
, tx
);
2166 * Undirty a buffer in the transaction group referenced by the given
2167 * transaction. Return whether this evicted the dbuf.
2170 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2173 uint64_t txg
= tx
->tx_txg
;
2174 dbuf_dirty_record_t
*dr
, **drp
;
2179 * Due to our use of dn_nlevels below, this can only be called
2180 * in open context, unless we are operating on the MOS.
2181 * From syncing context, dn_nlevels may be different from the
2182 * dn_nlevels used when dbuf was dirtied.
2184 ASSERT(db
->db_objset
==
2185 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2186 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2187 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2188 ASSERT0(db
->db_level
);
2189 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2192 * If this buffer is not dirty, we're done.
2194 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
2195 if (dr
->dr_txg
<= txg
)
2197 if (dr
== NULL
|| dr
->dr_txg
< txg
)
2199 ASSERT(dr
->dr_txg
== txg
);
2200 ASSERT(dr
->dr_dbuf
== db
);
2205 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2207 ASSERT(db
->db
.db_size
!= 0);
2209 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2210 dr
->dr_accounted
, txg
);
2215 * Note that there are three places in dbuf_dirty()
2216 * where this dirty record may be put on a list.
2217 * Make sure to do a list_remove corresponding to
2218 * every one of those list_insert calls.
2220 if (dr
->dr_parent
) {
2221 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2222 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2223 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2224 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2225 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2226 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2227 mutex_enter(&dn
->dn_mtx
);
2228 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2229 mutex_exit(&dn
->dn_mtx
);
2233 if (db
->db_state
!= DB_NOFILL
) {
2234 dbuf_unoverride(dr
);
2236 ASSERT(db
->db_buf
!= NULL
);
2237 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2238 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2239 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2242 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2244 ASSERT(db
->db_dirtycnt
> 0);
2245 db
->db_dirtycnt
-= 1;
2247 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2248 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
2257 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2259 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2261 ASSERT(tx
->tx_txg
!= 0);
2262 ASSERT(!refcount_is_zero(&db
->db_holds
));
2265 * Quick check for dirtyness. For already dirty blocks, this
2266 * reduces runtime of this function by >90%, and overall performance
2267 * by 50% for some workloads (e.g. file deletion with indirect blocks
2270 mutex_enter(&db
->db_mtx
);
2272 dbuf_dirty_record_t
*dr
;
2273 for (dr
= db
->db_last_dirty
;
2274 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
2276 * It's possible that it is already dirty but not cached,
2277 * because there are some calls to dbuf_dirty() that don't
2278 * go through dmu_buf_will_dirty().
2280 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
2281 /* This dbuf is already dirty and cached. */
2283 mutex_exit(&db
->db_mtx
);
2287 mutex_exit(&db
->db_mtx
);
2290 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2291 flags
|= DB_RF_HAVESTRUCT
;
2293 (void) dbuf_read(db
, NULL
, flags
);
2294 (void) dbuf_dirty(db
, tx
);
2298 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2300 dmu_buf_will_dirty_impl(db_fake
,
2301 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2305 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2307 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2309 db
->db_state
= DB_NOFILL
;
2311 dmu_buf_will_fill(db_fake
, tx
);
2315 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2317 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2319 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2320 ASSERT(tx
->tx_txg
!= 0);
2321 ASSERT(db
->db_level
== 0);
2322 ASSERT(!refcount_is_zero(&db
->db_holds
));
2324 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2325 dmu_tx_private_ok(tx
));
2328 (void) dbuf_dirty(db
, tx
);
2332 * This function is effectively the same as dmu_buf_will_dirty(), but
2333 * indicates the caller expects raw encrypted data in the db, and provides
2334 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2335 * blkptr_t when this dbuf is written. This is only used for blocks of
2336 * dnodes, during raw receive.
2339 dmu_buf_set_crypt_params(dmu_buf_t
*db_fake
, boolean_t byteorder
,
2340 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
, dmu_tx_t
*tx
)
2342 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2343 dbuf_dirty_record_t
*dr
;
2346 * dr_has_raw_params is only processed for blocks of dnodes
2347 * (see dbuf_sync_dnode_leaf_crypt()).
2349 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
2350 ASSERT3U(db
->db_level
, ==, 0);
2351 ASSERT(db
->db_objset
->os_raw_receive
);
2353 dmu_buf_will_dirty_impl(db_fake
,
2354 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2356 dr
= db
->db_last_dirty
;
2357 while (dr
!= NULL
&& dr
->dr_txg
> tx
->tx_txg
)
2360 ASSERT3P(dr
, !=, NULL
);
2361 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2363 dr
->dt
.dl
.dr_has_raw_params
= B_TRUE
;
2364 dr
->dt
.dl
.dr_byteorder
= byteorder
;
2365 bcopy(salt
, dr
->dt
.dl
.dr_salt
, ZIO_DATA_SALT_LEN
);
2366 bcopy(iv
, dr
->dt
.dl
.dr_iv
, ZIO_DATA_IV_LEN
);
2367 bcopy(mac
, dr
->dt
.dl
.dr_mac
, ZIO_DATA_MAC_LEN
);
2370 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2373 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2375 mutex_enter(&db
->db_mtx
);
2378 if (db
->db_state
== DB_FILL
) {
2379 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2380 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2381 /* we were freed while filling */
2382 /* XXX dbuf_undirty? */
2383 bzero(db
->db
.db_data
, db
->db
.db_size
);
2384 db
->db_freed_in_flight
= FALSE
;
2386 db
->db_state
= DB_CACHED
;
2387 cv_broadcast(&db
->db_changed
);
2389 mutex_exit(&db
->db_mtx
);
2393 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2394 bp_embedded_type_t etype
, enum zio_compress comp
,
2395 int uncompressed_size
, int compressed_size
, int byteorder
,
2398 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2399 struct dirty_leaf
*dl
;
2400 dmu_object_type_t type
;
2402 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2403 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2404 SPA_FEATURE_EMBEDDED_DATA
));
2408 type
= DB_DNODE(db
)->dn_type
;
2411 ASSERT0(db
->db_level
);
2412 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2414 dmu_buf_will_not_fill(dbuf
, tx
);
2416 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2417 dl
= &db
->db_last_dirty
->dt
.dl
;
2418 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2419 data
, comp
, uncompressed_size
, compressed_size
);
2420 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2421 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2422 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2423 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2425 dl
->dr_override_state
= DR_OVERRIDDEN
;
2426 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2430 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2431 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2434 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2436 ASSERT(!refcount_is_zero(&db
->db_holds
));
2437 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2438 ASSERT(db
->db_level
== 0);
2439 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2440 ASSERT(buf
!= NULL
);
2441 ASSERT(arc_buf_lsize(buf
) == db
->db
.db_size
);
2442 ASSERT(tx
->tx_txg
!= 0);
2444 arc_return_buf(buf
, db
);
2445 ASSERT(arc_released(buf
));
2447 mutex_enter(&db
->db_mtx
);
2449 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2450 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2452 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2454 if (db
->db_state
== DB_CACHED
&&
2455 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2457 * In practice, we will never have a case where we have an
2458 * encrypted arc buffer while additional holds exist on the
2459 * dbuf. We don't handle this here so we simply assert that
2462 ASSERT(!arc_is_encrypted(buf
));
2463 mutex_exit(&db
->db_mtx
);
2464 (void) dbuf_dirty(db
, tx
);
2465 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2466 arc_buf_destroy(buf
, db
);
2467 xuio_stat_wbuf_copied();
2471 xuio_stat_wbuf_nocopy();
2472 if (db
->db_state
== DB_CACHED
) {
2473 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2475 ASSERT(db
->db_buf
!= NULL
);
2476 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2477 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2479 if (!arc_released(db
->db_buf
)) {
2480 ASSERT(dr
->dt
.dl
.dr_override_state
==
2482 arc_release(db
->db_buf
, db
);
2484 dr
->dt
.dl
.dr_data
= buf
;
2485 arc_buf_destroy(db
->db_buf
, db
);
2486 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2487 arc_release(db
->db_buf
, db
);
2488 arc_buf_destroy(db
->db_buf
, db
);
2492 ASSERT(db
->db_buf
== NULL
);
2493 dbuf_set_data(db
, buf
);
2494 db
->db_state
= DB_FILL
;
2495 mutex_exit(&db
->db_mtx
);
2496 (void) dbuf_dirty(db
, tx
);
2497 dmu_buf_fill_done(&db
->db
, tx
);
2501 dbuf_destroy(dmu_buf_impl_t
*db
)
2504 dmu_buf_impl_t
*parent
= db
->db_parent
;
2505 dmu_buf_impl_t
*dndb
;
2507 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2508 ASSERT(refcount_is_zero(&db
->db_holds
));
2510 if (db
->db_buf
!= NULL
) {
2511 arc_buf_destroy(db
->db_buf
, db
);
2515 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2516 int slots
= DB_DNODE(db
)->dn_num_slots
;
2517 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2518 if (db
->db
.db_data
!= NULL
) {
2519 kmem_free(db
->db
.db_data
, bonuslen
);
2520 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2521 db
->db_state
= DB_UNCACHED
;
2525 dbuf_clear_data(db
);
2527 if (multilist_link_active(&db
->db_cache_link
)) {
2528 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
2529 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
2531 multilist_remove(dbuf_caches
[db
->db_caching_status
].cache
, db
);
2532 (void) refcount_remove_many(
2533 &dbuf_caches
[db
->db_caching_status
].size
,
2534 db
->db
.db_size
, db
);
2536 if (db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
2537 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
2539 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
2540 DBUF_STAT_BUMPDOWN(cache_count
);
2541 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
2544 db
->db_caching_status
= DB_NO_CACHE
;
2547 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2548 ASSERT(db
->db_data_pending
== NULL
);
2550 db
->db_state
= DB_EVICTING
;
2551 db
->db_blkptr
= NULL
;
2554 * Now that db_state is DB_EVICTING, nobody else can find this via
2555 * the hash table. We can now drop db_mtx, which allows us to
2556 * acquire the dn_dbufs_mtx.
2558 mutex_exit(&db
->db_mtx
);
2563 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2564 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2566 mutex_enter(&dn
->dn_dbufs_mtx
);
2567 avl_remove(&dn
->dn_dbufs
, db
);
2568 atomic_dec_32(&dn
->dn_dbufs_count
);
2572 mutex_exit(&dn
->dn_dbufs_mtx
);
2574 * Decrementing the dbuf count means that the hold corresponding
2575 * to the removed dbuf is no longer discounted in dnode_move(),
2576 * so the dnode cannot be moved until after we release the hold.
2577 * The membar_producer() ensures visibility of the decremented
2578 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2581 mutex_enter(&dn
->dn_mtx
);
2582 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
2583 db
->db_dnode_handle
= NULL
;
2585 dbuf_hash_remove(db
);
2590 ASSERT(refcount_is_zero(&db
->db_holds
));
2592 db
->db_parent
= NULL
;
2594 ASSERT(db
->db_buf
== NULL
);
2595 ASSERT(db
->db
.db_data
== NULL
);
2596 ASSERT(db
->db_hash_next
== NULL
);
2597 ASSERT(db
->db_blkptr
== NULL
);
2598 ASSERT(db
->db_data_pending
== NULL
);
2599 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
2600 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2602 kmem_cache_free(dbuf_kmem_cache
, db
);
2603 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2606 * If this dbuf is referenced from an indirect dbuf,
2607 * decrement the ref count on the indirect dbuf.
2609 if (parent
&& parent
!= dndb
) {
2610 mutex_enter(&parent
->db_mtx
);
2611 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
2616 * Note: While bpp will always be updated if the function returns success,
2617 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2618 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2621 __attribute__((always_inline
))
2623 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2624 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
2629 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2631 if (blkid
== DMU_SPILL_BLKID
) {
2632 mutex_enter(&dn
->dn_mtx
);
2633 if (dn
->dn_have_spill
&&
2634 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2635 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2638 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2639 *parentp
= dn
->dn_dbuf
;
2640 mutex_exit(&dn
->dn_mtx
);
2645 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2646 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2648 ASSERT3U(level
* epbs
, <, 64);
2649 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2651 * This assertion shouldn't trip as long as the max indirect block size
2652 * is less than 1M. The reason for this is that up to that point,
2653 * the number of levels required to address an entire object with blocks
2654 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2655 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2656 * (i.e. we can address the entire object), objects will all use at most
2657 * N-1 levels and the assertion won't overflow. However, once epbs is
2658 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2659 * enough to address an entire object, so objects will have 5 levels,
2660 * but then this assertion will overflow.
2662 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2663 * need to redo this logic to handle overflows.
2665 ASSERT(level
>= nlevels
||
2666 ((nlevels
- level
- 1) * epbs
) +
2667 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2668 if (level
>= nlevels
||
2669 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2670 ((nlevels
- level
- 1) * epbs
)) ||
2672 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2673 /* the buffer has no parent yet */
2674 return (SET_ERROR(ENOENT
));
2675 } else if (level
< nlevels
-1) {
2676 /* this block is referenced from an indirect block */
2678 dbuf_hold_arg_t
*dh
= dbuf_hold_arg_create(dn
, level
+ 1,
2679 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2680 err
= dbuf_hold_impl_arg(dh
);
2681 dbuf_hold_arg_destroy(dh
);
2684 err
= dbuf_read(*parentp
, NULL
,
2685 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2687 dbuf_rele(*parentp
, NULL
);
2691 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2692 (blkid
& ((1ULL << epbs
) - 1));
2693 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2694 ASSERT(BP_IS_HOLE(*bpp
));
2697 /* the block is referenced from the dnode */
2698 ASSERT3U(level
, ==, nlevels
-1);
2699 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2700 blkid
< dn
->dn_phys
->dn_nblkptr
);
2702 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2703 *parentp
= dn
->dn_dbuf
;
2705 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2710 static dmu_buf_impl_t
*
2711 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2712 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2714 objset_t
*os
= dn
->dn_objset
;
2715 dmu_buf_impl_t
*db
, *odb
;
2717 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2718 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2720 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2723 db
->db
.db_object
= dn
->dn_object
;
2724 db
->db_level
= level
;
2725 db
->db_blkid
= blkid
;
2726 db
->db_last_dirty
= NULL
;
2727 db
->db_dirtycnt
= 0;
2728 db
->db_dnode_handle
= dn
->dn_handle
;
2729 db
->db_parent
= parent
;
2730 db
->db_blkptr
= blkptr
;
2733 db
->db_user_immediate_evict
= FALSE
;
2734 db
->db_freed_in_flight
= FALSE
;
2735 db
->db_pending_evict
= FALSE
;
2737 if (blkid
== DMU_BONUS_BLKID
) {
2738 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2739 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2740 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2741 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2742 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2743 db
->db_state
= DB_UNCACHED
;
2744 db
->db_caching_status
= DB_NO_CACHE
;
2745 /* the bonus dbuf is not placed in the hash table */
2746 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2748 } else if (blkid
== DMU_SPILL_BLKID
) {
2749 db
->db
.db_size
= (blkptr
!= NULL
) ?
2750 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2751 db
->db
.db_offset
= 0;
2754 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2755 db
->db
.db_size
= blocksize
;
2756 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2760 * Hold the dn_dbufs_mtx while we get the new dbuf
2761 * in the hash table *and* added to the dbufs list.
2762 * This prevents a possible deadlock with someone
2763 * trying to look up this dbuf before its added to the
2766 mutex_enter(&dn
->dn_dbufs_mtx
);
2767 db
->db_state
= DB_EVICTING
;
2768 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2769 /* someone else inserted it first */
2770 kmem_cache_free(dbuf_kmem_cache
, db
);
2771 mutex_exit(&dn
->dn_dbufs_mtx
);
2772 DBUF_STAT_BUMP(hash_insert_race
);
2775 avl_add(&dn
->dn_dbufs
, db
);
2777 db
->db_state
= DB_UNCACHED
;
2778 db
->db_caching_status
= DB_NO_CACHE
;
2779 mutex_exit(&dn
->dn_dbufs_mtx
);
2780 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2782 if (parent
&& parent
!= dn
->dn_dbuf
)
2783 dbuf_add_ref(parent
, db
);
2785 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2786 refcount_count(&dn
->dn_holds
) > 0);
2787 (void) refcount_add(&dn
->dn_holds
, db
);
2788 atomic_inc_32(&dn
->dn_dbufs_count
);
2790 dprintf_dbuf(db
, "db=%p\n", db
);
2795 typedef struct dbuf_prefetch_arg
{
2796 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2797 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2798 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2799 int dpa_curlevel
; /* The current level that we're reading */
2800 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2801 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2802 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2803 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2804 } dbuf_prefetch_arg_t
;
2807 * Actually issue the prefetch read for the block given.
2810 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2812 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2815 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2816 arc_flags_t aflags
=
2817 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2819 /* dnodes are always read as raw and then converted later */
2820 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
2821 dpa
->dpa_curlevel
== 0)
2822 zio_flags
|= ZIO_FLAG_RAW
;
2824 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2825 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2826 ASSERT(dpa
->dpa_zio
!= NULL
);
2827 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2828 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
2832 * Called when an indirect block above our prefetch target is read in. This
2833 * will either read in the next indirect block down the tree or issue the actual
2834 * prefetch if the next block down is our target.
2837 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
2838 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
2840 dbuf_prefetch_arg_t
*dpa
= private;
2842 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2843 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2846 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
2847 kmem_free(dpa
, sizeof (*dpa
));
2850 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
2853 * The dpa_dnode is only valid if we are called with a NULL
2854 * zio. This indicates that the arc_read() returned without
2855 * first calling zio_read() to issue a physical read. Once
2856 * a physical read is made the dpa_dnode must be invalidated
2857 * as the locks guarding it may have been dropped. If the
2858 * dpa_dnode is still valid, then we want to add it to the dbuf
2859 * cache. To do so, we must hold the dbuf associated with the block
2860 * we just prefetched, read its contents so that we associate it
2861 * with an arc_buf_t, and then release it.
2864 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2865 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
2866 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2868 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2870 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2872 dpa
->dpa_dnode
= NULL
;
2873 } else if (dpa
->dpa_dnode
!= NULL
) {
2874 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2875 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2876 dpa
->dpa_zb
.zb_level
));
2877 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2878 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2879 (void) dbuf_read(db
, NULL
,
2880 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2881 dbuf_rele(db
, FTAG
);
2884 dpa
->dpa_curlevel
--;
2885 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2886 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2887 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
2888 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2890 if (BP_IS_HOLE(bp
)) {
2891 kmem_free(dpa
, sizeof (*dpa
));
2892 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2893 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2894 dbuf_issue_final_prefetch(dpa
, bp
);
2895 kmem_free(dpa
, sizeof (*dpa
));
2897 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2898 zbookmark_phys_t zb
;
2900 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2901 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
2902 iter_aflags
|= ARC_FLAG_L2CACHE
;
2904 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2906 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2907 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2909 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2910 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2911 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2915 arc_buf_destroy(abuf
, private);
2919 * Issue prefetch reads for the given block on the given level. If the indirect
2920 * blocks above that block are not in memory, we will read them in
2921 * asynchronously. As a result, this call never blocks waiting for a read to
2922 * complete. Note that the prefetch might fail if the dataset is encrypted and
2923 * the encryption key is unmapped before the IO completes.
2926 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2930 int epbs
, nlevels
, curlevel
;
2933 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2934 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2936 if (blkid
> dn
->dn_maxblkid
)
2939 if (dnode_block_freed(dn
, blkid
))
2943 * This dnode hasn't been written to disk yet, so there's nothing to
2946 nlevels
= dn
->dn_phys
->dn_nlevels
;
2947 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2950 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2951 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2954 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2957 mutex_exit(&db
->db_mtx
);
2959 * This dbuf already exists. It is either CACHED, or
2960 * (we assume) about to be read or filled.
2966 * Find the closest ancestor (indirect block) of the target block
2967 * that is present in the cache. In this indirect block, we will
2968 * find the bp that is at curlevel, curblkid.
2972 while (curlevel
< nlevels
- 1) {
2973 int parent_level
= curlevel
+ 1;
2974 uint64_t parent_blkid
= curblkid
>> epbs
;
2977 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2978 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2979 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2980 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2981 dbuf_rele(db
, FTAG
);
2985 curlevel
= parent_level
;
2986 curblkid
= parent_blkid
;
2989 if (curlevel
== nlevels
- 1) {
2990 /* No cached indirect blocks found. */
2991 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2992 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2994 if (BP_IS_HOLE(&bp
))
2997 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2999 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
3002 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
3003 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
3004 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3005 dn
->dn_object
, level
, blkid
);
3006 dpa
->dpa_curlevel
= curlevel
;
3007 dpa
->dpa_prio
= prio
;
3008 dpa
->dpa_aflags
= aflags
;
3009 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
3010 dpa
->dpa_dnode
= dn
;
3011 dpa
->dpa_epbs
= epbs
;
3014 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3015 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
3016 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
3019 * If we have the indirect just above us, no need to do the asynchronous
3020 * prefetch chain; we'll just run the last step ourselves. If we're at
3021 * a higher level, though, we want to issue the prefetches for all the
3022 * indirect blocks asynchronously, so we can go on with whatever we were
3025 if (curlevel
== level
) {
3026 ASSERT3U(curblkid
, ==, blkid
);
3027 dbuf_issue_final_prefetch(dpa
, &bp
);
3028 kmem_free(dpa
, sizeof (*dpa
));
3030 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
3031 zbookmark_phys_t zb
;
3033 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
3034 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
3035 iter_aflags
|= ARC_FLAG_L2CACHE
;
3037 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
3038 dn
->dn_object
, curlevel
, curblkid
);
3039 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
3040 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
3041 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3045 * We use pio here instead of dpa_zio since it's possible that
3046 * dpa may have already been freed.
3051 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
3054 * Helper function for dbuf_hold_impl_arg() to copy a buffer. Handles
3055 * the case of encrypted, compressed and uncompressed buffers by
3056 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
3057 * arc_alloc_compressed_buf() or arc_alloc_buf().*
3059 * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl_arg().
3061 noinline
static void
3062 dbuf_hold_copy(struct dbuf_hold_arg
*dh
)
3064 dnode_t
*dn
= dh
->dh_dn
;
3065 dmu_buf_impl_t
*db
= dh
->dh_db
;
3066 dbuf_dirty_record_t
*dr
= dh
->dh_dr
;
3067 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
3069 enum zio_compress compress_type
= arc_get_compression(data
);
3071 if (arc_is_encrypted(data
)) {
3072 boolean_t byteorder
;
3073 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3074 uint8_t iv
[ZIO_DATA_IV_LEN
];
3075 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3077 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
3078 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
3079 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
3080 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
3082 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3083 dbuf_set_data(db
, arc_alloc_compressed_buf(
3084 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
3085 arc_buf_lsize(data
), compress_type
));
3087 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
3088 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
3091 bcopy(data
->b_data
, db
->db
.db_data
, arc_buf_size(data
));
3095 * Returns with db_holds incremented, and db_mtx not held.
3096 * Note: dn_struct_rwlock must be held.
3099 dbuf_hold_impl_arg(struct dbuf_hold_arg
*dh
)
3101 dh
->dh_parent
= NULL
;
3103 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
3104 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
3105 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
3107 *(dh
->dh_dbp
) = NULL
;
3109 /* dbuf_find() returns with db_mtx held */
3110 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
3111 dh
->dh_level
, dh
->dh_blkid
);
3113 if (dh
->dh_db
== NULL
) {
3116 if (dh
->dh_fail_uncached
)
3117 return (SET_ERROR(ENOENT
));
3119 ASSERT3P(dh
->dh_parent
, ==, NULL
);
3120 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
3121 dh
->dh_fail_sparse
, &dh
->dh_parent
, &dh
->dh_bp
);
3122 if (dh
->dh_fail_sparse
) {
3123 if (dh
->dh_err
== 0 &&
3124 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
3125 dh
->dh_err
= SET_ERROR(ENOENT
);
3128 dbuf_rele(dh
->dh_parent
, NULL
);
3129 return (dh
->dh_err
);
3132 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
3133 return (dh
->dh_err
);
3134 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
3135 dh
->dh_parent
, dh
->dh_bp
);
3138 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
3139 mutex_exit(&dh
->dh_db
->db_mtx
);
3140 return (SET_ERROR(ENOENT
));
3143 if (dh
->dh_db
->db_buf
!= NULL
) {
3144 arc_buf_access(dh
->dh_db
->db_buf
);
3145 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
3148 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
3151 * If this buffer is currently syncing out, and we are are
3152 * still referencing it from db_data, we need to make a copy
3153 * of it in case we decide we want to dirty it again in this txg.
3155 if (dh
->dh_db
->db_level
== 0 &&
3156 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
3157 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3158 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
3159 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
3160 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
)
3164 if (multilist_link_active(&dh
->dh_db
->db_cache_link
)) {
3165 ASSERT(refcount_is_zero(&dh
->dh_db
->db_holds
));
3166 ASSERT(dh
->dh_db
->db_caching_status
== DB_DBUF_CACHE
||
3167 dh
->dh_db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
3170 dbuf_caches
[dh
->dh_db
->db_caching_status
].cache
,
3172 (void) refcount_remove_many(
3173 &dbuf_caches
[dh
->dh_db
->db_caching_status
].size
,
3174 dh
->dh_db
->db
.db_size
, dh
->dh_db
);
3176 if (dh
->dh_db
->db_caching_status
== DB_DBUF_METADATA_CACHE
) {
3177 DBUF_STAT_BUMPDOWN(metadata_cache_count
);
3179 DBUF_STAT_BUMPDOWN(cache_levels
[dh
->dh_db
->db_level
]);
3180 DBUF_STAT_BUMPDOWN(cache_count
);
3181 DBUF_STAT_DECR(cache_levels_bytes
[dh
->dh_db
->db_level
],
3182 dh
->dh_db
->db
.db_size
);
3184 dh
->dh_db
->db_caching_status
= DB_NO_CACHE
;
3186 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
3187 DBUF_VERIFY(dh
->dh_db
);
3188 mutex_exit(&dh
->dh_db
->db_mtx
);
3190 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3192 dbuf_rele(dh
->dh_parent
, NULL
);
3194 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
3195 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
3196 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
3197 *(dh
->dh_dbp
) = dh
->dh_db
;
3203 * dbuf_hold_impl_arg() is called recursively, via dbuf_findbp(). There can
3204 * be as many recursive calls as there are levels of on-disk indirect blocks,
3205 * but typically only 0-2 recursive calls. To minimize the stack frame size,
3206 * the recursive function's arguments and "local variables" are allocated on
3207 * the heap as the dbuf_hold_arg_t.
3210 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3211 boolean_t fail_sparse
, boolean_t fail_uncached
,
3212 void *tag
, dmu_buf_impl_t
**dbp
)
3214 dbuf_hold_arg_t
*dh
= dbuf_hold_arg_create(dn
, level
, blkid
,
3215 fail_sparse
, fail_uncached
, tag
, dbp
);
3217 int error
= dbuf_hold_impl_arg(dh
);
3219 dbuf_hold_arg_destroy(dh
);
3224 static dbuf_hold_arg_t
*
3225 dbuf_hold_arg_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3226 boolean_t fail_sparse
, boolean_t fail_uncached
,
3227 void *tag
, dmu_buf_impl_t
**dbp
)
3229 dbuf_hold_arg_t
*dh
= kmem_alloc(sizeof (*dh
), KM_SLEEP
);
3231 dh
->dh_level
= level
;
3232 dh
->dh_blkid
= blkid
;
3234 dh
->dh_fail_sparse
= fail_sparse
;
3235 dh
->dh_fail_uncached
= fail_uncached
;
3241 dh
->dh_parent
= NULL
;
3250 dbuf_hold_arg_destroy(dbuf_hold_arg_t
*dh
)
3252 kmem_free(dh
, sizeof (*dh
));
3256 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
3258 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3262 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
3265 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3266 return (err
? NULL
: db
);
3270 dbuf_create_bonus(dnode_t
*dn
)
3272 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3274 ASSERT(dn
->dn_bonus
== NULL
);
3275 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
3279 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3281 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3284 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3285 return (SET_ERROR(ENOTSUP
));
3287 blksz
= SPA_MINBLOCKSIZE
;
3288 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3289 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3293 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3294 dbuf_new_size(db
, blksz
, tx
);
3295 rw_exit(&dn
->dn_struct_rwlock
);
3302 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3304 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3307 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3309 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
3311 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
3312 VERIFY3S(holds
, >, 1);
3315 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3317 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3320 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3321 dmu_buf_impl_t
*found_db
;
3322 boolean_t result
= B_FALSE
;
3324 if (blkid
== DMU_BONUS_BLKID
)
3325 found_db
= dbuf_find_bonus(os
, obj
);
3327 found_db
= dbuf_find(os
, obj
, 0, blkid
);
3329 if (found_db
!= NULL
) {
3330 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3331 (void) refcount_add(&db
->db_holds
, tag
);
3334 mutex_exit(&found_db
->db_mtx
);
3340 * If you call dbuf_rele() you had better not be referencing the dnode handle
3341 * unless you have some other direct or indirect hold on the dnode. (An indirect
3342 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3343 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3344 * dnode's parent dbuf evicting its dnode handles.
3347 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
3349 mutex_enter(&db
->db_mtx
);
3350 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
3354 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
3356 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3360 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3361 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3362 * argument should be set if we are already in the dbuf-evicting code
3363 * path, in which case we don't want to recursively evict. This allows us to
3364 * avoid deeply nested stacks that would have a call flow similar to this:
3366 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3369 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3373 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
, boolean_t evicting
)
3377 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3381 * Remove the reference to the dbuf before removing its hold on the
3382 * dnode so we can guarantee in dnode_move() that a referenced bonus
3383 * buffer has a corresponding dnode hold.
3385 holds
= refcount_remove(&db
->db_holds
, tag
);
3389 * We can't freeze indirects if there is a possibility that they
3390 * may be modified in the current syncing context.
3392 if (db
->db_buf
!= NULL
&&
3393 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3394 arc_buf_freeze(db
->db_buf
);
3397 if (holds
== db
->db_dirtycnt
&&
3398 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3399 dbuf_evict_user(db
);
3402 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3404 boolean_t evict_dbuf
= db
->db_pending_evict
;
3407 * If the dnode moves here, we cannot cross this
3408 * barrier until the move completes.
3413 atomic_dec_32(&dn
->dn_dbufs_count
);
3416 * Decrementing the dbuf count means that the bonus
3417 * buffer's dnode hold is no longer discounted in
3418 * dnode_move(). The dnode cannot move until after
3419 * the dnode_rele() below.
3424 * Do not reference db after its lock is dropped.
3425 * Another thread may evict it.
3427 mutex_exit(&db
->db_mtx
);
3430 dnode_evict_bonus(dn
);
3433 } else if (db
->db_buf
== NULL
) {
3435 * This is a special case: we never associated this
3436 * dbuf with any data allocated from the ARC.
3438 ASSERT(db
->db_state
== DB_UNCACHED
||
3439 db
->db_state
== DB_NOFILL
);
3441 } else if (arc_released(db
->db_buf
)) {
3443 * This dbuf has anonymous data associated with it.
3447 boolean_t do_arc_evict
= B_FALSE
;
3449 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3451 if (!DBUF_IS_CACHEABLE(db
) &&
3452 db
->db_blkptr
!= NULL
&&
3453 !BP_IS_HOLE(db
->db_blkptr
) &&
3454 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
3455 do_arc_evict
= B_TRUE
;
3456 bp
= *db
->db_blkptr
;
3459 if (!DBUF_IS_CACHEABLE(db
) ||
3460 db
->db_pending_evict
) {
3462 } else if (!multilist_link_active(&db
->db_cache_link
)) {
3463 ASSERT3U(db
->db_caching_status
, ==,
3466 dbuf_cached_state_t dcs
=
3467 dbuf_include_in_metadata_cache(db
) ?
3468 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
3469 db
->db_caching_status
= dcs
;
3471 multilist_insert(dbuf_caches
[dcs
].cache
, db
);
3472 (void) refcount_add_many(&dbuf_caches
[dcs
].size
,
3473 db
->db
.db_size
, db
);
3475 if (dcs
== DB_DBUF_METADATA_CACHE
) {
3476 DBUF_STAT_BUMP(metadata_cache_count
);
3478 metadata_cache_size_bytes_max
,
3480 &dbuf_caches
[dcs
].size
));
3483 cache_levels
[db
->db_level
]);
3484 DBUF_STAT_BUMP(cache_count
);
3486 cache_levels_bytes
[db
->db_level
],
3488 DBUF_STAT_MAX(cache_size_bytes_max
,
3490 &dbuf_caches
[dcs
].size
));
3492 mutex_exit(&db
->db_mtx
);
3494 if (db
->db_caching_status
== DB_DBUF_CACHE
&&
3496 dbuf_evict_notify();
3501 arc_freed(spa
, &bp
);
3504 mutex_exit(&db
->db_mtx
);
3509 #pragma weak dmu_buf_refcount = dbuf_refcount
3511 dbuf_refcount(dmu_buf_impl_t
*db
)
3513 return (refcount_count(&db
->db_holds
));
3517 dmu_buf_user_refcount(dmu_buf_t
*db_fake
)
3520 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3522 mutex_enter(&db
->db_mtx
);
3523 ASSERT3U(refcount_count(&db
->db_holds
), >=, db
->db_dirtycnt
);
3524 holds
= refcount_count(&db
->db_holds
) - db
->db_dirtycnt
;
3525 mutex_exit(&db
->db_mtx
);
3531 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3532 dmu_buf_user_t
*new_user
)
3534 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3536 mutex_enter(&db
->db_mtx
);
3537 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3538 if (db
->db_user
== old_user
)
3539 db
->db_user
= new_user
;
3541 old_user
= db
->db_user
;
3542 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3543 mutex_exit(&db
->db_mtx
);
3549 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3551 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3555 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3557 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3559 db
->db_user_immediate_evict
= TRUE
;
3560 return (dmu_buf_set_user(db_fake
, user
));
3564 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3566 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3570 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3572 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3574 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3575 return (db
->db_user
);
3579 dmu_buf_user_evict_wait()
3581 taskq_wait(dbu_evict_taskq
);
3585 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3587 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3588 return (dbi
->db_blkptr
);
3592 dmu_buf_get_objset(dmu_buf_t
*db
)
3594 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3595 return (dbi
->db_objset
);
3599 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3601 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3602 DB_DNODE_ENTER(dbi
);
3603 return (DB_DNODE(dbi
));
3607 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3609 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3614 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3616 /* ASSERT(dmu_tx_is_syncing(tx) */
3617 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3619 if (db
->db_blkptr
!= NULL
)
3622 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3623 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3624 BP_ZERO(db
->db_blkptr
);
3627 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3629 * This buffer was allocated at a time when there was
3630 * no available blkptrs from the dnode, or it was
3631 * inappropriate to hook it in (i.e., nlevels mis-match).
3633 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3634 ASSERT(db
->db_parent
== NULL
);
3635 db
->db_parent
= dn
->dn_dbuf
;
3636 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3639 dmu_buf_impl_t
*parent
= db
->db_parent
;
3640 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3642 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3643 if (parent
== NULL
) {
3644 mutex_exit(&db
->db_mtx
);
3645 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3646 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3647 db
->db_blkid
>> epbs
, db
);
3648 rw_exit(&dn
->dn_struct_rwlock
);
3649 mutex_enter(&db
->db_mtx
);
3650 db
->db_parent
= parent
;
3652 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3653 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3659 * When syncing out a blocks of dnodes, adjust the block to deal with
3660 * encryption. Normally, we make sure the block is decrypted before writing
3661 * it. If we have crypt params, then we are writing a raw (encrypted) block,
3662 * from a raw receive. In this case, set the ARC buf's crypt params so
3663 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
3666 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t
*dr
)
3669 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3671 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3672 ASSERT3U(db
->db
.db_object
, ==, DMU_META_DNODE_OBJECT
);
3673 ASSERT3U(db
->db_level
, ==, 0);
3675 if (!db
->db_objset
->os_raw_receive
&& arc_is_encrypted(db
->db_buf
)) {
3676 zbookmark_phys_t zb
;
3679 * Unfortunately, there is currently no mechanism for
3680 * syncing context to handle decryption errors. An error
3681 * here is only possible if an attacker maliciously
3682 * changed a dnode block and updated the associated
3683 * checksums going up the block tree.
3685 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
3686 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3687 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
3690 panic("Invalid dnode block MAC");
3691 } else if (dr
->dt
.dl
.dr_has_raw_params
) {
3692 (void) arc_release(dr
->dt
.dl
.dr_data
, db
);
3693 arc_convert_to_raw(dr
->dt
.dl
.dr_data
,
3694 dmu_objset_id(db
->db_objset
),
3695 dr
->dt
.dl
.dr_byteorder
, DMU_OT_DNODE
,
3696 dr
->dt
.dl
.dr_salt
, dr
->dt
.dl
.dr_iv
, dr
->dt
.dl
.dr_mac
);
3701 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3702 * is critical the we not allow the compiler to inline this function in to
3703 * dbuf_sync_list() thereby drastically bloating the stack usage.
3705 noinline
static void
3706 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3708 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3712 ASSERT(dmu_tx_is_syncing(tx
));
3714 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3716 mutex_enter(&db
->db_mtx
);
3718 ASSERT(db
->db_level
> 0);
3721 /* Read the block if it hasn't been read yet. */
3722 if (db
->db_buf
== NULL
) {
3723 mutex_exit(&db
->db_mtx
);
3724 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3725 mutex_enter(&db
->db_mtx
);
3727 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3728 ASSERT(db
->db_buf
!= NULL
);
3732 /* Indirect block size must match what the dnode thinks it is. */
3733 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3734 dbuf_check_blkptr(dn
, db
);
3737 /* Provide the pending dirty record to child dbufs */
3738 db
->db_data_pending
= dr
;
3740 mutex_exit(&db
->db_mtx
);
3742 dbuf_write(dr
, db
->db_buf
, tx
);
3745 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3746 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3747 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3748 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3753 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3754 * critical the we not allow the compiler to inline this function in to
3755 * dbuf_sync_list() thereby drastically bloating the stack usage.
3757 noinline
static void
3758 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3760 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3761 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3764 uint64_t txg
= tx
->tx_txg
;
3766 ASSERT(dmu_tx_is_syncing(tx
));
3768 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3770 mutex_enter(&db
->db_mtx
);
3772 * To be synced, we must be dirtied. But we
3773 * might have been freed after the dirty.
3775 if (db
->db_state
== DB_UNCACHED
) {
3776 /* This buffer has been freed since it was dirtied */
3777 ASSERT(db
->db
.db_data
== NULL
);
3778 } else if (db
->db_state
== DB_FILL
) {
3779 /* This buffer was freed and is now being re-filled */
3780 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3782 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3789 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3790 mutex_enter(&dn
->dn_mtx
);
3791 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3793 * In the previous transaction group, the bonus buffer
3794 * was entirely used to store the attributes for the
3795 * dnode which overrode the dn_spill field. However,
3796 * when adding more attributes to the file a spill
3797 * block was required to hold the extra attributes.
3799 * Make sure to clear the garbage left in the dn_spill
3800 * field from the previous attributes in the bonus
3801 * buffer. Otherwise, after writing out the spill
3802 * block to the new allocated dva, it will free
3803 * the old block pointed to by the invalid dn_spill.
3805 db
->db_blkptr
= NULL
;
3807 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3808 mutex_exit(&dn
->dn_mtx
);
3812 * If this is a bonus buffer, simply copy the bonus data into the
3813 * dnode. It will be written out when the dnode is synced (and it
3814 * will be synced, since it must have been dirty for dbuf_sync to
3817 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3818 dbuf_dirty_record_t
**drp
;
3820 ASSERT(*datap
!= NULL
);
3821 ASSERT0(db
->db_level
);
3822 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
3823 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3824 bcopy(*datap
, DN_BONUS(dn
->dn_phys
),
3825 DN_MAX_BONUS_LEN(dn
->dn_phys
));
3828 if (*datap
!= db
->db
.db_data
) {
3829 int slots
= DB_DNODE(db
)->dn_num_slots
;
3830 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3831 kmem_free(*datap
, bonuslen
);
3832 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3834 db
->db_data_pending
= NULL
;
3835 drp
= &db
->db_last_dirty
;
3837 drp
= &(*drp
)->dr_next
;
3838 ASSERT(dr
->dr_next
== NULL
);
3839 ASSERT(dr
->dr_dbuf
== db
);
3841 if (dr
->dr_dbuf
->db_level
!= 0) {
3842 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3843 list_destroy(&dr
->dt
.di
.dr_children
);
3845 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3846 ASSERT(db
->db_dirtycnt
> 0);
3847 db
->db_dirtycnt
-= 1;
3848 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
, B_FALSE
);
3855 * This function may have dropped the db_mtx lock allowing a dmu_sync
3856 * operation to sneak in. As a result, we need to ensure that we
3857 * don't check the dr_override_state until we have returned from
3858 * dbuf_check_blkptr.
3860 dbuf_check_blkptr(dn
, db
);
3863 * If this buffer is in the middle of an immediate write,
3864 * wait for the synchronous IO to complete.
3866 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3867 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3868 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3869 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3873 * If this is a dnode block, ensure it is appropriately encrypted
3874 * or decrypted, depending on what we are writing to it this txg.
3876 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
3877 dbuf_prepare_encrypted_dnode_leaf(dr
);
3879 if (db
->db_state
!= DB_NOFILL
&&
3880 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3881 refcount_count(&db
->db_holds
) > 1 &&
3882 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3883 *datap
== db
->db_buf
) {
3885 * If this buffer is currently "in use" (i.e., there
3886 * are active holds and db_data still references it),
3887 * then make a copy before we start the write so that
3888 * any modifications from the open txg will not leak
3891 * NOTE: this copy does not need to be made for
3892 * objects only modified in the syncing context (e.g.
3893 * DNONE_DNODE blocks).
3895 int psize
= arc_buf_size(*datap
);
3896 int lsize
= arc_buf_lsize(*datap
);
3897 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3898 enum zio_compress compress_type
= arc_get_compression(*datap
);
3900 if (arc_is_encrypted(*datap
)) {
3901 boolean_t byteorder
;
3902 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3903 uint8_t iv
[ZIO_DATA_IV_LEN
];
3904 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3906 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
3907 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
3908 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
3909 dn
->dn_type
, psize
, lsize
, compress_type
);
3910 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3911 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
3912 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
3913 psize
, lsize
, compress_type
);
3915 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
3917 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
3919 db
->db_data_pending
= dr
;
3921 mutex_exit(&db
->db_mtx
);
3923 dbuf_write(dr
, *datap
, tx
);
3925 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3926 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3927 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3931 * Although zio_nowait() does not "wait for an IO", it does
3932 * initiate the IO. If this is an empty write it seems plausible
3933 * that the IO could actually be completed before the nowait
3934 * returns. We need to DB_DNODE_EXIT() first in case
3935 * zio_nowait() invalidates the dbuf.
3938 zio_nowait(dr
->dr_zio
);
3943 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3945 dbuf_dirty_record_t
*dr
;
3947 while ((dr
= list_head(list
))) {
3948 if (dr
->dr_zio
!= NULL
) {
3950 * If we find an already initialized zio then we
3951 * are processing the meta-dnode, and we have finished.
3952 * The dbufs for all dnodes are put back on the list
3953 * during processing, so that we can zio_wait()
3954 * these IOs after initiating all child IOs.
3956 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3957 DMU_META_DNODE_OBJECT
);
3960 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3961 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3962 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3964 list_remove(list
, dr
);
3965 if (dr
->dr_dbuf
->db_level
> 0)
3966 dbuf_sync_indirect(dr
, tx
);
3968 dbuf_sync_leaf(dr
, tx
);
3974 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3976 dmu_buf_impl_t
*db
= vdb
;
3978 blkptr_t
*bp
= zio
->io_bp
;
3979 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3980 spa_t
*spa
= zio
->io_spa
;
3985 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3986 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3990 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3991 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3992 zio
->io_prev_space_delta
= delta
;
3994 if (bp
->blk_birth
!= 0) {
3995 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3996 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3997 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3998 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3999 BP_IS_EMBEDDED(bp
));
4000 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
4003 mutex_enter(&db
->db_mtx
);
4006 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4007 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4008 ASSERT(!(BP_IS_HOLE(bp
)) &&
4009 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4013 if (db
->db_level
== 0) {
4014 mutex_enter(&dn
->dn_mtx
);
4015 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
4016 db
->db_blkid
!= DMU_SPILL_BLKID
) {
4017 ASSERT0(db
->db_objset
->os_raw_receive
);
4018 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
4020 mutex_exit(&dn
->dn_mtx
);
4022 if (dn
->dn_type
== DMU_OT_DNODE
) {
4024 while (i
< db
->db
.db_size
) {
4026 (void *)(((char *)db
->db
.db_data
) + i
);
4028 i
+= DNODE_MIN_SIZE
;
4029 if (dnp
->dn_type
!= DMU_OT_NONE
) {
4031 i
+= dnp
->dn_extra_slots
*
4036 if (BP_IS_HOLE(bp
)) {
4043 blkptr_t
*ibp
= db
->db
.db_data
;
4044 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
4045 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
4046 if (BP_IS_HOLE(ibp
))
4048 fill
+= BP_GET_FILL(ibp
);
4053 if (!BP_IS_EMBEDDED(bp
))
4054 BP_SET_FILL(bp
, fill
);
4056 mutex_exit(&db
->db_mtx
);
4058 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
4059 *db
->db_blkptr
= *bp
;
4060 rw_exit(&dn
->dn_struct_rwlock
);
4065 * This function gets called just prior to running through the compression
4066 * stage of the zio pipeline. If we're an indirect block comprised of only
4067 * holes, then we want this indirect to be compressed away to a hole. In
4068 * order to do that we must zero out any information about the holes that
4069 * this indirect points to prior to before we try to compress it.
4072 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4074 dmu_buf_impl_t
*db
= vdb
;
4077 unsigned int epbs
, i
;
4079 ASSERT3U(db
->db_level
, >, 0);
4082 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
4083 ASSERT3U(epbs
, <, 31);
4085 /* Determine if all our children are holes */
4086 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
4087 if (!BP_IS_HOLE(bp
))
4092 * If all the children are holes, then zero them all out so that
4093 * we may get compressed away.
4095 if (i
== 1ULL << epbs
) {
4097 * We only found holes. Grab the rwlock to prevent
4098 * anybody from reading the blocks we're about to
4101 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
4102 bzero(db
->db
.db_data
, db
->db
.db_size
);
4103 rw_exit(&dn
->dn_struct_rwlock
);
4109 * The SPA will call this callback several times for each zio - once
4110 * for every physical child i/o (zio->io_phys_children times). This
4111 * allows the DMU to monitor the progress of each logical i/o. For example,
4112 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
4113 * block. There may be a long delay before all copies/fragments are completed,
4114 * so this callback allows us to retire dirty space gradually, as the physical
4119 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4121 dmu_buf_impl_t
*db
= arg
;
4122 objset_t
*os
= db
->db_objset
;
4123 dsl_pool_t
*dp
= dmu_objset_pool(os
);
4124 dbuf_dirty_record_t
*dr
;
4127 dr
= db
->db_data_pending
;
4128 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
4131 * The callback will be called io_phys_children times. Retire one
4132 * portion of our dirty space each time we are called. Any rounding
4133 * error will be cleaned up by dsl_pool_sync()'s call to
4134 * dsl_pool_undirty_space().
4136 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
4137 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
4142 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
4144 dmu_buf_impl_t
*db
= vdb
;
4145 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
4146 blkptr_t
*bp
= db
->db_blkptr
;
4147 objset_t
*os
= db
->db_objset
;
4148 dmu_tx_t
*tx
= os
->os_synctx
;
4149 dbuf_dirty_record_t
**drp
, *dr
;
4151 ASSERT0(zio
->io_error
);
4152 ASSERT(db
->db_blkptr
== bp
);
4155 * For nopwrites and rewrites we ensure that the bp matches our
4156 * original and bypass all the accounting.
4158 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
4159 ASSERT(BP_EQUAL(bp
, bp_orig
));
4161 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
4162 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
4163 dsl_dataset_block_born(ds
, bp
, tx
);
4166 mutex_enter(&db
->db_mtx
);
4170 drp
= &db
->db_last_dirty
;
4171 while ((dr
= *drp
) != db
->db_data_pending
)
4173 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
4174 ASSERT(dr
->dr_dbuf
== db
);
4175 ASSERT(dr
->dr_next
== NULL
);
4179 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
4184 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
4185 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
4186 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
4191 if (db
->db_level
== 0) {
4192 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
4193 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
4194 if (db
->db_state
!= DB_NOFILL
) {
4195 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
4196 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
4203 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
4204 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
4205 if (!BP_IS_HOLE(db
->db_blkptr
)) {
4206 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
4208 ASSERT3U(db
->db_blkid
, <=,
4209 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
4210 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4214 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4215 list_destroy(&dr
->dt
.di
.dr_children
);
4217 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4219 cv_broadcast(&db
->db_changed
);
4220 ASSERT(db
->db_dirtycnt
> 0);
4221 db
->db_dirtycnt
-= 1;
4222 db
->db_data_pending
= NULL
;
4223 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
4227 dbuf_write_nofill_ready(zio_t
*zio
)
4229 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4233 dbuf_write_nofill_done(zio_t
*zio
)
4235 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4239 dbuf_write_override_ready(zio_t
*zio
)
4241 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4242 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4244 dbuf_write_ready(zio
, NULL
, db
);
4248 dbuf_write_override_done(zio_t
*zio
)
4250 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4251 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4252 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4254 mutex_enter(&db
->db_mtx
);
4255 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4256 if (!BP_IS_HOLE(obp
))
4257 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4258 arc_release(dr
->dt
.dl
.dr_data
, db
);
4260 mutex_exit(&db
->db_mtx
);
4262 dbuf_write_done(zio
, NULL
, db
);
4264 if (zio
->io_abd
!= NULL
)
4265 abd_put(zio
->io_abd
);
4268 typedef struct dbuf_remap_impl_callback_arg
{
4270 uint64_t drica_blk_birth
;
4272 } dbuf_remap_impl_callback_arg_t
;
4275 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
4278 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
4279 objset_t
*os
= drica
->drica_os
;
4280 spa_t
*spa
= dmu_objset_spa(os
);
4281 dmu_tx_t
*tx
= drica
->drica_tx
;
4283 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4285 if (os
== spa_meta_objset(spa
)) {
4286 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
4288 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
4289 size
, drica
->drica_blk_birth
, tx
);
4294 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, dmu_tx_t
*tx
)
4296 blkptr_t bp_copy
= *bp
;
4297 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4298 dbuf_remap_impl_callback_arg_t drica
;
4300 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4302 drica
.drica_os
= dn
->dn_objset
;
4303 drica
.drica_blk_birth
= bp
->blk_birth
;
4304 drica
.drica_tx
= tx
;
4305 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
4308 * The struct_rwlock prevents dbuf_read_impl() from
4309 * dereferencing the BP while we are changing it. To
4310 * avoid lock contention, only grab it when we are actually
4313 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
4315 rw_exit(&dn
->dn_struct_rwlock
);
4320 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
4321 * to remap a copy of every bp in the dbuf.
4324 dbuf_can_remap(const dmu_buf_impl_t
*db
)
4326 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
4327 blkptr_t
*bp
= db
->db
.db_data
;
4328 boolean_t ret
= B_FALSE
;
4330 ASSERT3U(db
->db_level
, >, 0);
4331 ASSERT3S(db
->db_state
, ==, DB_CACHED
);
4333 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
));
4335 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4336 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
4337 blkptr_t bp_copy
= bp
[i
];
4338 if (spa_remap_blkptr(spa
, &bp_copy
, NULL
, NULL
)) {
4343 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4349 dnode_needs_remap(const dnode_t
*dn
)
4351 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
4352 boolean_t ret
= B_FALSE
;
4354 if (dn
->dn_phys
->dn_nlevels
== 0) {
4358 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
));
4360 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4361 for (int j
= 0; j
< dn
->dn_phys
->dn_nblkptr
; j
++) {
4362 blkptr_t bp_copy
= dn
->dn_phys
->dn_blkptr
[j
];
4363 if (spa_remap_blkptr(spa
, &bp_copy
, NULL
, NULL
)) {
4368 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4374 * Remap any existing BP's to concrete vdevs, if possible.
4377 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
4379 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
4380 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
4382 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
4385 if (db
->db_level
> 0) {
4386 blkptr_t
*bp
= db
->db
.db_data
;
4387 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
4388 dbuf_remap_impl(dn
, &bp
[i
], tx
);
4390 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
4391 dnode_phys_t
*dnp
= db
->db
.db_data
;
4392 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
4394 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
;
4395 i
+= dnp
[i
].dn_extra_slots
+ 1) {
4396 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
4397 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], tx
);
4404 /* Issue I/O to commit a dirty buffer to disk. */
4406 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
4408 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4411 dmu_buf_impl_t
*parent
= db
->db_parent
;
4412 uint64_t txg
= tx
->tx_txg
;
4413 zbookmark_phys_t zb
;
4418 ASSERT(dmu_tx_is_syncing(tx
));
4424 if (db
->db_state
!= DB_NOFILL
) {
4425 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
4427 * Private object buffers are released here rather
4428 * than in dbuf_dirty() since they are only modified
4429 * in the syncing context and we don't want the
4430 * overhead of making multiple copies of the data.
4432 if (BP_IS_HOLE(db
->db_blkptr
)) {
4435 dbuf_release_bp(db
);
4437 dbuf_remap(dn
, db
, tx
);
4441 if (parent
!= dn
->dn_dbuf
) {
4442 /* Our parent is an indirect block. */
4443 /* We have a dirty parent that has been scheduled for write. */
4444 ASSERT(parent
&& parent
->db_data_pending
);
4445 /* Our parent's buffer is one level closer to the dnode. */
4446 ASSERT(db
->db_level
== parent
->db_level
-1);
4448 * We're about to modify our parent's db_data by modifying
4449 * our block pointer, so the parent must be released.
4451 ASSERT(arc_released(parent
->db_buf
));
4452 zio
= parent
->db_data_pending
->dr_zio
;
4454 /* Our parent is the dnode itself. */
4455 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
4456 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
4457 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
4458 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
4459 ASSERT3P(db
->db_blkptr
, ==,
4460 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
4464 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
4465 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
4468 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
4469 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
4470 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4472 if (db
->db_blkid
== DMU_SPILL_BLKID
)
4474 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
4476 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
4480 * We copy the blkptr now (rather than when we instantiate the dirty
4481 * record), because its value can change between open context and
4482 * syncing context. We do not need to hold dn_struct_rwlock to read
4483 * db_blkptr because we are in syncing context.
4485 dr
->dr_bp_copy
= *db
->db_blkptr
;
4487 if (db
->db_level
== 0 &&
4488 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
4490 * The BP for this block has been provided by open context
4491 * (by dmu_sync() or dmu_buf_write_embedded()).
4493 abd_t
*contents
= (data
!= NULL
) ?
4494 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
4496 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4497 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
4498 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
4499 dbuf_write_override_done
,
4500 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
4501 mutex_enter(&db
->db_mtx
);
4502 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
4503 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
4504 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
4505 mutex_exit(&db
->db_mtx
);
4506 } else if (db
->db_state
== DB_NOFILL
) {
4507 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
4508 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
4509 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4510 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
4511 dbuf_write_nofill_ready
, NULL
, NULL
,
4512 dbuf_write_nofill_done
, db
,
4513 ZIO_PRIORITY_ASYNC_WRITE
,
4514 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
4516 ASSERT(arc_released(data
));
4519 * For indirect blocks, we want to setup the children
4520 * ready callback so that we can properly handle an indirect
4521 * block that only contains holes.
4523 arc_write_done_func_t
*children_ready_cb
= NULL
;
4524 if (db
->db_level
!= 0)
4525 children_ready_cb
= dbuf_write_children_ready
;
4527 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
4528 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
4529 &zp
, dbuf_write_ready
,
4530 children_ready_cb
, dbuf_write_physdone
,
4531 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
4532 ZIO_FLAG_MUSTSUCCEED
, &zb
);
4536 #if defined(_KERNEL)
4537 EXPORT_SYMBOL(dbuf_find
);
4538 EXPORT_SYMBOL(dbuf_is_metadata
);
4539 EXPORT_SYMBOL(dbuf_destroy
);
4540 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
4541 EXPORT_SYMBOL(dbuf_whichblock
);
4542 EXPORT_SYMBOL(dbuf_read
);
4543 EXPORT_SYMBOL(dbuf_unoverride
);
4544 EXPORT_SYMBOL(dbuf_free_range
);
4545 EXPORT_SYMBOL(dbuf_new_size
);
4546 EXPORT_SYMBOL(dbuf_release_bp
);
4547 EXPORT_SYMBOL(dbuf_dirty
);
4548 EXPORT_SYMBOL(dmu_buf_set_crypt_params
);
4549 EXPORT_SYMBOL(dmu_buf_will_dirty
);
4550 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
4551 EXPORT_SYMBOL(dmu_buf_will_fill
);
4552 EXPORT_SYMBOL(dmu_buf_fill_done
);
4553 EXPORT_SYMBOL(dmu_buf_rele
);
4554 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
4555 EXPORT_SYMBOL(dbuf_prefetch
);
4556 EXPORT_SYMBOL(dbuf_hold_impl
);
4557 EXPORT_SYMBOL(dbuf_hold
);
4558 EXPORT_SYMBOL(dbuf_hold_level
);
4559 EXPORT_SYMBOL(dbuf_create_bonus
);
4560 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
4561 EXPORT_SYMBOL(dbuf_rm_spill
);
4562 EXPORT_SYMBOL(dbuf_add_ref
);
4563 EXPORT_SYMBOL(dbuf_rele
);
4564 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
4565 EXPORT_SYMBOL(dbuf_refcount
);
4566 EXPORT_SYMBOL(dbuf_sync_list
);
4567 EXPORT_SYMBOL(dmu_buf_set_user
);
4568 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
4569 EXPORT_SYMBOL(dmu_buf_get_user
);
4570 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
4573 module_param(dbuf_cache_max_bytes
, ulong
, 0644);
4574 MODULE_PARM_DESC(dbuf_cache_max_bytes
,
4575 "Maximum size in bytes of the dbuf cache.");
4577 module_param(dbuf_cache_hiwater_pct
, uint
, 0644);
4578 MODULE_PARM_DESC(dbuf_cache_hiwater_pct
,
4579 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4582 module_param(dbuf_cache_lowater_pct
, uint
, 0644);
4583 MODULE_PARM_DESC(dbuf_cache_lowater_pct
,
4584 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4587 module_param(dbuf_metadata_cache_max_bytes
, ulong
, 0644);
4588 MODULE_PARM_DESC(dbuf_metadata_cache_max_bytes
,
4589 "Maximum size in bytes of the dbuf metadata cache.");
4591 module_param(dbuf_cache_shift
, int, 0644);
4592 MODULE_PARM_DESC(dbuf_cache_shift
,
4593 "Set the size of the dbuf cache to a log2 fraction of arc size.");
4595 module_param(dbuf_metadata_cache_shift
, int, 0644);
4596 MODULE_PARM_DESC(dbuf_cache_shift
,
4597 "Set the size of the dbuf metadata cache to a log2 fraction of "