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) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
42 #include <sys/dmu_zfetch.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
51 #include <sys/cityhash.h>
52 #include <sys/spa_impl.h>
54 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
55 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
57 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
58 dmu_buf_evict_func_t
*evict_func_sync
,
59 dmu_buf_evict_func_t
*evict_func_async
,
60 dmu_buf_t
**clear_on_evict_dbufp
);
63 * Global data structures and functions for the dbuf cache.
65 static kmem_cache_t
*dbuf_kmem_cache
;
66 static taskq_t
*dbu_evict_taskq
;
68 static kthread_t
*dbuf_cache_evict_thread
;
69 static kmutex_t dbuf_evict_lock
;
70 static kcondvar_t dbuf_evict_cv
;
71 static boolean_t dbuf_evict_thread_exit
;
74 * There are two dbuf caches; each dbuf can only be in one of them at a time.
76 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
77 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
78 * that represent the metadata that describes filesystems/snapshots/
79 * bookmarks/properties/etc. We only evict from this cache when we export a
80 * pool, to short-circuit as much I/O as possible for all administrative
81 * commands that need the metadata. There is no eviction policy for this
82 * cache, because we try to only include types in it which would occupy a
83 * very small amount of space per object but create a large impact on the
84 * performance of these commands. Instead, after it reaches a maximum size
85 * (which should only happen on very small memory systems with a very large
86 * number of filesystem objects), we stop taking new dbufs into the
87 * metadata cache, instead putting them in the normal dbuf cache.
89 * 2. LRU cache of dbufs. The "dbuf cache" maintains a list of dbufs that
90 * are not currently held but have been recently released. These dbufs
91 * are not eligible for arc eviction until they are aged out of the cache.
92 * Dbufs that are aged out of the cache will be immediately destroyed and
93 * become eligible for arc eviction.
95 * Dbufs are added to these caches once the last hold is released. If a dbuf is
96 * later accessed and still exists in the dbuf cache, then it will be removed
97 * from the cache and later re-added to the head of the cache.
99 * If a given dbuf meets the requirements for the metadata cache, it will go
100 * there, otherwise it will be considered for the generic LRU dbuf cache. The
101 * caches and the refcounts tracking their sizes are stored in an array indexed
102 * by those caches' matching enum values (from dbuf_cached_state_t).
104 typedef struct dbuf_cache
{
108 dbuf_cache_t dbuf_caches
[DB_CACHE_MAX
];
110 /* Size limits for the caches */
111 uint64_t dbuf_cache_max_bytes
= 0;
112 uint64_t dbuf_metadata_cache_max_bytes
= 0;
113 /* Set the default sizes of the caches to log2 fraction of arc size */
114 int dbuf_cache_shift
= 5;
115 int dbuf_metadata_cache_shift
= 6;
118 * For diagnostic purposes, this is incremented whenever we can't add
119 * something to the metadata cache because it's full, and instead put
120 * the data in the regular dbuf cache.
122 uint64_t dbuf_metadata_cache_overflow
;
125 * The LRU dbuf cache uses a three-stage eviction policy:
126 * - A low water marker designates when the dbuf eviction thread
127 * should stop evicting from the dbuf cache.
128 * - When we reach the maximum size (aka mid water mark), we
129 * signal the eviction thread to run.
130 * - The high water mark indicates when the eviction thread
131 * is unable to keep up with the incoming load and eviction must
132 * happen in the context of the calling thread.
136 * low water mid water hi water
137 * +----------------------------------------+----------+----------+
142 * +----------------------------------------+----------+----------+
144 * evicting eviction directly
147 * The high and low water marks indicate the operating range for the eviction
148 * thread. The low water mark is, by default, 90% of the total size of the
149 * cache and the high water mark is at 110% (both of these percentages can be
150 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
151 * respectively). The eviction thread will try to ensure that the cache remains
152 * within this range by waking up every second and checking if the cache is
153 * above the low water mark. The thread can also be woken up by callers adding
154 * elements into the cache if the cache is larger than the mid water (i.e max
155 * cache size). Once the eviction thread is woken up and eviction is required,
156 * it will continue evicting buffers until it's able to reduce the cache size
157 * to the low water mark. If the cache size continues to grow and hits the high
158 * water mark, then callers adding elments to the cache will begin to evict
159 * directly from the cache until the cache is no longer above the high water
164 * The percentage above and below the maximum cache size.
166 uint_t dbuf_cache_hiwater_pct
= 10;
167 uint_t dbuf_cache_lowater_pct
= 10;
171 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
173 dmu_buf_impl_t
*db
= vdb
;
174 bzero(db
, sizeof (dmu_buf_impl_t
));
176 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
177 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
178 multilist_link_init(&db
->db_cache_link
);
179 refcount_create(&db
->db_holds
);
186 dbuf_dest(void *vdb
, void *unused
)
188 dmu_buf_impl_t
*db
= vdb
;
189 mutex_destroy(&db
->db_mtx
);
190 cv_destroy(&db
->db_changed
);
191 ASSERT(!multilist_link_active(&db
->db_cache_link
));
192 refcount_destroy(&db
->db_holds
);
196 * dbuf hash table routines
198 static dbuf_hash_table_t dbuf_hash_table
;
200 static uint64_t dbuf_hash_count
;
203 * We use Cityhash for this. It's fast, and has good hash properties without
204 * requiring any large static buffers.
207 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
209 return (cityhash4((uintptr_t)os
, obj
, (uint64_t)lvl
, blkid
));
212 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
213 ((dbuf)->db.db_object == (obj) && \
214 (dbuf)->db_objset == (os) && \
215 (dbuf)->db_level == (level) && \
216 (dbuf)->db_blkid == (blkid))
219 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
221 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
222 uint64_t hv
= dbuf_hash(os
, obj
, level
, blkid
);
223 uint64_t idx
= hv
& h
->hash_table_mask
;
226 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
227 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
228 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
229 mutex_enter(&db
->db_mtx
);
230 if (db
->db_state
!= DB_EVICTING
) {
231 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
234 mutex_exit(&db
->db_mtx
);
237 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
241 static dmu_buf_impl_t
*
242 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
245 dmu_buf_impl_t
*db
= NULL
;
247 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
248 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
249 if (dn
->dn_bonus
!= NULL
) {
251 mutex_enter(&db
->db_mtx
);
253 rw_exit(&dn
->dn_struct_rwlock
);
254 dnode_rele(dn
, FTAG
);
260 * Insert an entry into the hash table. If there is already an element
261 * equal to elem in the hash table, then the already existing element
262 * will be returned and the new element will not be inserted.
263 * Otherwise returns NULL.
265 static dmu_buf_impl_t
*
266 dbuf_hash_insert(dmu_buf_impl_t
*db
)
268 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
269 objset_t
*os
= db
->db_objset
;
270 uint64_t obj
= db
->db
.db_object
;
271 int level
= db
->db_level
;
272 uint64_t blkid
= db
->db_blkid
;
273 uint64_t hv
= dbuf_hash(os
, obj
, level
, blkid
);
274 uint64_t idx
= hv
& h
->hash_table_mask
;
277 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
278 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
279 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
280 mutex_enter(&dbf
->db_mtx
);
281 if (dbf
->db_state
!= DB_EVICTING
) {
282 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
285 mutex_exit(&dbf
->db_mtx
);
289 mutex_enter(&db
->db_mtx
);
290 db
->db_hash_next
= h
->hash_table
[idx
];
291 h
->hash_table
[idx
] = db
;
292 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
293 atomic_inc_64(&dbuf_hash_count
);
299 * Remove an entry from the hash table. It must be in the EVICTING state.
302 dbuf_hash_remove(dmu_buf_impl_t
*db
)
304 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
305 uint64_t hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
306 db
->db_level
, db
->db_blkid
);
307 uint64_t idx
= hv
& h
->hash_table_mask
;
308 dmu_buf_impl_t
*dbf
, **dbp
;
311 * We musn't hold db_mtx to maintain lock ordering:
312 * DBUF_HASH_MUTEX > db_mtx.
314 ASSERT(refcount_is_zero(&db
->db_holds
));
315 ASSERT(db
->db_state
== DB_EVICTING
);
316 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
318 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
319 dbp
= &h
->hash_table
[idx
];
320 while ((dbf
= *dbp
) != db
) {
321 dbp
= &dbf
->db_hash_next
;
324 *dbp
= db
->db_hash_next
;
325 db
->db_hash_next
= NULL
;
326 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
327 atomic_dec_64(&dbuf_hash_count
);
333 } dbvu_verify_type_t
;
336 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
341 if (db
->db_user
== NULL
)
344 /* Only data blocks support the attachment of user data. */
345 ASSERT(db
->db_level
== 0);
347 /* Clients must resolve a dbuf before attaching user data. */
348 ASSERT(db
->db
.db_data
!= NULL
);
349 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
351 holds
= refcount_count(&db
->db_holds
);
352 if (verify_type
== DBVU_EVICTING
) {
354 * Immediate eviction occurs when holds == dirtycnt.
355 * For normal eviction buffers, holds is zero on
356 * eviction, except when dbuf_fix_old_data() calls
357 * dbuf_clear_data(). However, the hold count can grow
358 * during eviction even though db_mtx is held (see
359 * dmu_bonus_hold() for an example), so we can only
360 * test the generic invariant that holds >= dirtycnt.
362 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
364 if (db
->db_user_immediate_evict
== TRUE
)
365 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
367 ASSERT3U(holds
, >, 0);
373 dbuf_evict_user(dmu_buf_impl_t
*db
)
375 dmu_buf_user_t
*dbu
= db
->db_user
;
377 ASSERT(MUTEX_HELD(&db
->db_mtx
));
382 dbuf_verify_user(db
, DBVU_EVICTING
);
386 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
387 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
391 * There are two eviction callbacks - one that we call synchronously
392 * and one that we invoke via a taskq. The async one is useful for
393 * avoiding lock order reversals and limiting stack depth.
395 * Note that if we have a sync callback but no async callback,
396 * it's likely that the sync callback will free the structure
397 * containing the dbu. In that case we need to take care to not
398 * dereference dbu after calling the sync evict func.
400 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
402 if (dbu
->dbu_evict_func_sync
!= NULL
)
403 dbu
->dbu_evict_func_sync(dbu
);
406 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
407 dbu
, 0, &dbu
->dbu_tqent
);
412 dbuf_is_metadata(dmu_buf_impl_t
*db
)
414 if (db
->db_level
> 0) {
417 boolean_t is_metadata
;
420 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
423 return (is_metadata
);
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_metadata_cache_overflow
++;
451 DTRACE_PROBE1(dbuf__metadata__cache__overflow
,
452 dmu_buf_impl_t
*, db
);
463 * This function *must* return indices evenly distributed between all
464 * sublists of the multilist. This is needed due to how the dbuf eviction
465 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
466 * distributed between all sublists and uses this assumption when
467 * deciding which sublist to evict from and how much to evict from it.
470 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
472 dmu_buf_impl_t
*db
= obj
;
475 * The assumption here, is the hash value for a given
476 * dmu_buf_impl_t will remain constant throughout it's lifetime
477 * (i.e. it's objset, object, level and blkid fields don't change).
478 * Thus, we don't need to store the dbuf's sublist index
479 * on insertion, as this index can be recalculated on removal.
481 * Also, the low order bits of the hash value are thought to be
482 * distributed evenly. Otherwise, in the case that the multilist
483 * has a power of two number of sublists, each sublists' usage
484 * would not be evenly distributed.
486 return (dbuf_hash(db
->db_objset
, db
->db
.db_object
,
487 db
->db_level
, db
->db_blkid
) %
488 multilist_get_num_sublists(ml
));
491 static inline boolean_t
492 dbuf_cache_above_hiwater(void)
494 uint64_t dbuf_cache_hiwater_bytes
=
495 (dbuf_cache_max_bytes
* dbuf_cache_hiwater_pct
) / 100;
497 return (refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
498 dbuf_cache_max_bytes
+ dbuf_cache_hiwater_bytes
);
501 static inline boolean_t
502 dbuf_cache_above_lowater(void)
504 uint64_t dbuf_cache_lowater_bytes
=
505 (dbuf_cache_max_bytes
* dbuf_cache_lowater_pct
) / 100;
507 return (refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
508 dbuf_cache_max_bytes
- dbuf_cache_lowater_bytes
);
512 * Evict the oldest eligible dbuf from the dbuf cache.
517 int idx
= multilist_get_random_index(dbuf_caches
[DB_DBUF_CACHE
].cache
);
518 multilist_sublist_t
*mls
= multilist_sublist_lock(
519 dbuf_caches
[DB_DBUF_CACHE
].cache
, idx
);
521 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
523 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
524 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
525 db
= multilist_sublist_prev(mls
, db
);
528 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
529 multilist_sublist_t
*, mls
);
532 multilist_sublist_remove(mls
, db
);
533 multilist_sublist_unlock(mls
);
534 (void) refcount_remove_many(&dbuf_caches
[DB_DBUF_CACHE
].size
,
536 ASSERT3U(db
->db_caching_status
, ==, DB_DBUF_CACHE
);
537 db
->db_caching_status
= DB_NO_CACHE
;
540 multilist_sublist_unlock(mls
);
545 * The dbuf evict thread is responsible for aging out dbufs from the
546 * cache. Once the cache has reached it's maximum size, dbufs are removed
547 * and destroyed. The eviction thread will continue running until the size
548 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
549 * out of the cache it is destroyed and becomes eligible for arc eviction.
553 dbuf_evict_thread(void *unused
)
557 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
559 mutex_enter(&dbuf_evict_lock
);
560 while (!dbuf_evict_thread_exit
) {
561 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
562 CALLB_CPR_SAFE_BEGIN(&cpr
);
563 (void) cv_timedwait_hires(&dbuf_evict_cv
,
564 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
565 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
567 mutex_exit(&dbuf_evict_lock
);
570 * Keep evicting as long as we're above the low water mark
571 * for the cache. We do this without holding the locks to
572 * minimize lock contention.
574 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
578 mutex_enter(&dbuf_evict_lock
);
581 dbuf_evict_thread_exit
= B_FALSE
;
582 cv_broadcast(&dbuf_evict_cv
);
583 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
588 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
589 * If the dbuf cache is at its high water mark, then evict a dbuf from the
590 * dbuf cache using the callers context.
593 dbuf_evict_notify(void)
596 * We check if we should evict without holding the dbuf_evict_lock,
597 * because it's OK to occasionally make the wrong decision here,
598 * and grabbing the lock results in massive lock contention.
600 if (refcount_count(&dbuf_caches
[DB_DBUF_CACHE
].size
) >
601 dbuf_cache_max_bytes
) {
602 if (dbuf_cache_above_hiwater())
604 cv_signal(&dbuf_evict_cv
);
611 uint64_t hsize
= 1ULL << 16;
612 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
616 * The hash table is big enough to fill all of physical memory
617 * with an average 4K block size. The table will take up
618 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
620 while (hsize
* 4096 < physmem
* PAGESIZE
)
624 h
->hash_table_mask
= hsize
- 1;
625 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
626 if (h
->hash_table
== NULL
) {
627 /* XXX - we should really return an error instead of assert */
628 ASSERT(hsize
> (1ULL << 10));
633 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
634 sizeof (dmu_buf_impl_t
),
635 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
637 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
638 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
641 * Setup the parameters for the dbuf caches. We set the sizes of the
642 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
643 * of the size of the ARC, respectively. If the values are set in
644 * /etc/system and they're not greater than the size of the ARC, then
645 * we honor that value.
647 if (dbuf_cache_max_bytes
== 0 ||
648 dbuf_cache_max_bytes
>= arc_max_bytes()) {
649 dbuf_cache_max_bytes
= arc_max_bytes() >> dbuf_cache_shift
;
651 if (dbuf_metadata_cache_max_bytes
== 0 ||
652 dbuf_metadata_cache_max_bytes
>= arc_max_bytes()) {
653 dbuf_metadata_cache_max_bytes
=
654 arc_max_bytes() >> dbuf_metadata_cache_shift
;
658 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
659 * configuration is not required.
661 dbu_evict_taskq
= taskq_create("dbu_evict", 1, minclsyspri
, 0, 0, 0);
663 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
664 dbuf_caches
[dcs
].cache
=
665 multilist_create(sizeof (dmu_buf_impl_t
),
666 offsetof(dmu_buf_impl_t
, db_cache_link
),
667 dbuf_cache_multilist_index_func
);
668 refcount_create(&dbuf_caches
[dcs
].size
);
671 dbuf_evict_thread_exit
= B_FALSE
;
672 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
673 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
674 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
675 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
681 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
684 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
685 mutex_destroy(&h
->hash_mutexes
[i
]);
686 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
687 kmem_cache_destroy(dbuf_kmem_cache
);
688 taskq_destroy(dbu_evict_taskq
);
690 mutex_enter(&dbuf_evict_lock
);
691 dbuf_evict_thread_exit
= B_TRUE
;
692 while (dbuf_evict_thread_exit
) {
693 cv_signal(&dbuf_evict_cv
);
694 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
696 mutex_exit(&dbuf_evict_lock
);
698 mutex_destroy(&dbuf_evict_lock
);
699 cv_destroy(&dbuf_evict_cv
);
701 for (dbuf_cached_state_t dcs
= 0; dcs
< DB_CACHE_MAX
; dcs
++) {
702 refcount_destroy(&dbuf_caches
[dcs
].size
);
703 multilist_destroy(dbuf_caches
[dcs
].cache
);
713 dbuf_verify(dmu_buf_impl_t
*db
)
716 dbuf_dirty_record_t
*dr
;
718 ASSERT(MUTEX_HELD(&db
->db_mtx
));
720 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
723 ASSERT(db
->db_objset
!= NULL
);
727 ASSERT(db
->db_parent
== NULL
);
728 ASSERT(db
->db_blkptr
== NULL
);
730 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
731 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
732 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
733 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
734 db
->db_blkid
== DMU_SPILL_BLKID
||
735 !avl_is_empty(&dn
->dn_dbufs
));
737 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
739 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
740 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
741 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
743 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
744 ASSERT0(db
->db
.db_offset
);
746 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
749 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
750 ASSERT(dr
->dr_dbuf
== db
);
752 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
753 ASSERT(dr
->dr_dbuf
== db
);
756 * We can't assert that db_size matches dn_datablksz because it
757 * can be momentarily different when another thread is doing
760 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
761 dr
= db
->db_data_pending
;
763 * It should only be modified in syncing context, so
764 * make sure we only have one copy of the data.
766 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
769 /* verify db->db_blkptr */
771 if (db
->db_parent
== dn
->dn_dbuf
) {
772 /* db is pointed to by the dnode */
773 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
774 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
775 ASSERT(db
->db_parent
== NULL
);
777 ASSERT(db
->db_parent
!= NULL
);
778 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
779 ASSERT3P(db
->db_blkptr
, ==,
780 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
782 /* db is pointed to by an indirect block */
783 int epb
= db
->db_parent
->db
.db_size
>> SPA_BLKPTRSHIFT
;
784 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
785 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
788 * dnode_grow_indblksz() can make this fail if we don't
789 * have the struct_rwlock. XXX indblksz no longer
790 * grows. safe to do this now?
792 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
793 ASSERT3P(db
->db_blkptr
, ==,
794 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
795 db
->db_blkid
% epb
));
799 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
800 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
801 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
802 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
804 * If the blkptr isn't set but they have nonzero data,
805 * it had better be dirty, otherwise we'll lose that
806 * data when we evict this buffer.
808 * There is an exception to this rule for indirect blocks; in
809 * this case, if the indirect block is a hole, we fill in a few
810 * fields on each of the child blocks (importantly, birth time)
811 * to prevent hole birth times from being lost when you
812 * partially fill in a hole.
814 if (db
->db_dirtycnt
== 0) {
815 if (db
->db_level
== 0) {
816 uint64_t *buf
= db
->db
.db_data
;
819 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
823 blkptr_t
*bps
= db
->db
.db_data
;
824 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
827 * We want to verify that all the blkptrs in the
828 * indirect block are holes, but we may have
829 * automatically set up a few fields for them.
830 * We iterate through each blkptr and verify
831 * they only have those fields set.
834 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
836 blkptr_t
*bp
= &bps
[i
];
837 ASSERT(ZIO_CHECKSUM_IS_ZERO(
840 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
841 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
842 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
843 ASSERT0(bp
->blk_fill
);
844 ASSERT0(bp
->blk_pad
[0]);
845 ASSERT0(bp
->blk_pad
[1]);
846 ASSERT(!BP_IS_EMBEDDED(bp
));
847 ASSERT(BP_IS_HOLE(bp
));
848 ASSERT0(bp
->blk_phys_birth
);
858 dbuf_clear_data(dmu_buf_impl_t
*db
)
860 ASSERT(MUTEX_HELD(&db
->db_mtx
));
862 ASSERT3P(db
->db_buf
, ==, NULL
);
863 db
->db
.db_data
= NULL
;
864 if (db
->db_state
!= DB_NOFILL
)
865 db
->db_state
= DB_UNCACHED
;
869 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
871 ASSERT(MUTEX_HELD(&db
->db_mtx
));
875 ASSERT(buf
->b_data
!= NULL
);
876 db
->db
.db_data
= buf
->b_data
;
880 * Loan out an arc_buf for read. Return the loaned arc_buf.
883 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
887 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
888 mutex_enter(&db
->db_mtx
);
889 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
890 int blksz
= db
->db
.db_size
;
891 spa_t
*spa
= db
->db_objset
->os_spa
;
893 mutex_exit(&db
->db_mtx
);
894 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
895 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
898 arc_loan_inuse_buf(abuf
, db
);
901 mutex_exit(&db
->db_mtx
);
907 * Calculate which level n block references the data at the level 0 offset
911 dbuf_whichblock(dnode_t
*dn
, int64_t level
, uint64_t offset
)
913 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
915 * The level n blkid is equal to the level 0 blkid divided by
916 * the number of level 0s in a level n block.
918 * The level 0 blkid is offset >> datablkshift =
919 * offset / 2^datablkshift.
921 * The number of level 0s in a level n is the number of block
922 * pointers in an indirect block, raised to the power of level.
923 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
924 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
926 * Thus, the level n blkid is: offset /
927 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
928 * = offset / 2^(datablkshift + level *
929 * (indblkshift - SPA_BLKPTRSHIFT))
930 * = offset >> (datablkshift + level *
931 * (indblkshift - SPA_BLKPTRSHIFT))
933 return (offset
>> (dn
->dn_datablkshift
+ level
*
934 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
)));
936 ASSERT3U(offset
, <, dn
->dn_datablksz
);
942 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
944 dmu_buf_impl_t
*db
= vdb
;
946 mutex_enter(&db
->db_mtx
);
947 ASSERT3U(db
->db_state
, ==, DB_READ
);
949 * All reads are synchronous, so we must have a hold on the dbuf
951 ASSERT(refcount_count(&db
->db_holds
) > 0);
952 ASSERT(db
->db_buf
== NULL
);
953 ASSERT(db
->db
.db_data
== NULL
);
956 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
957 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
958 ASSERT3P(db
->db_buf
, ==, NULL
);
959 db
->db_state
= DB_UNCACHED
;
960 } else if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
961 /* freed in flight */
962 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
963 arc_release(buf
, db
);
964 bzero(buf
->b_data
, db
->db
.db_size
);
966 db
->db_freed_in_flight
= FALSE
;
967 dbuf_set_data(db
, buf
);
968 db
->db_state
= DB_CACHED
;
971 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
972 dbuf_set_data(db
, buf
);
973 db
->db_state
= DB_CACHED
;
975 cv_broadcast(&db
->db_changed
);
976 dbuf_rele_and_unlock(db
, NULL
, B_FALSE
);
980 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
984 arc_flags_t aflags
= ARC_FLAG_NOWAIT
;
988 ASSERT(!refcount_is_zero(&db
->db_holds
));
989 /* We need the struct_rwlock to prevent db_blkptr from changing. */
990 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
991 ASSERT(MUTEX_HELD(&db
->db_mtx
));
992 ASSERT(db
->db_state
== DB_UNCACHED
);
993 ASSERT(db
->db_buf
== NULL
);
995 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
996 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
998 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
999 db
->db
.db_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
1000 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1001 if (bonuslen
< DN_MAX_BONUSLEN
)
1002 bzero(db
->db
.db_data
, DN_MAX_BONUSLEN
);
1004 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1006 db
->db_state
= DB_CACHED
;
1007 mutex_exit(&db
->db_mtx
);
1012 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1013 * processes the delete record and clears the bp while we are waiting
1014 * for the dn_mtx (resulting in a "no" from block_freed).
1016 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1017 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1018 BP_IS_HOLE(db
->db_blkptr
)))) {
1019 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1021 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1023 bzero(db
->db
.db_data
, db
->db
.db_size
);
1025 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1026 BP_IS_HOLE(db
->db_blkptr
) &&
1027 db
->db_blkptr
->blk_birth
!= 0) {
1028 blkptr_t
*bps
= db
->db
.db_data
;
1029 for (int i
= 0; i
< ((1 <<
1030 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1032 blkptr_t
*bp
= &bps
[i
];
1033 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1034 1 << dn
->dn_indblkshift
);
1036 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1038 BP_GET_LSIZE(db
->db_blkptr
));
1039 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1041 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1042 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1046 db
->db_state
= DB_CACHED
;
1047 mutex_exit(&db
->db_mtx
);
1053 db
->db_state
= DB_READ
;
1054 mutex_exit(&db
->db_mtx
);
1056 if (DBUF_IS_L2CACHEABLE(db
))
1057 aflags
|= ARC_FLAG_L2CACHE
;
1059 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
1060 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
1061 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1063 dbuf_add_ref(db
, NULL
);
1065 (void) arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1066 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
1067 (flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
1072 * This is our just-in-time copy function. It makes a copy of buffers that
1073 * have been modified in a previous transaction group before we access them in
1074 * the current active group.
1076 * This function is used in three places: when we are dirtying a buffer for the
1077 * first time in a txg, when we are freeing a range in a dnode that includes
1078 * this buffer, and when we are accessing a buffer which was received compressed
1079 * and later referenced in a WRITE_BYREF record.
1081 * Note that when we are called from dbuf_free_range() we do not put a hold on
1082 * the buffer, we just traverse the active dbuf list for the dnode.
1085 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1087 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1089 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1090 ASSERT(db
->db
.db_data
!= NULL
);
1091 ASSERT(db
->db_level
== 0);
1092 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1095 (dr
->dt
.dl
.dr_data
!=
1096 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1100 * If the last dirty record for this dbuf has not yet synced
1101 * and its referencing the dbuf data, either:
1102 * reset the reference to point to a new copy,
1103 * or (if there a no active holders)
1104 * just null out the current db_data pointer.
1106 ASSERT(dr
->dr_txg
>= txg
- 2);
1107 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1108 /* Note that the data bufs here are zio_bufs */
1109 dr
->dt
.dl
.dr_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
1110 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1111 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, DN_MAX_BONUSLEN
);
1112 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1113 int size
= arc_buf_size(db
->db_buf
);
1114 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1115 spa_t
*spa
= db
->db_objset
->os_spa
;
1116 enum zio_compress compress_type
=
1117 arc_get_compression(db
->db_buf
);
1119 if (compress_type
== ZIO_COMPRESS_OFF
) {
1120 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1122 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1123 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1124 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1126 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1129 dbuf_clear_data(db
);
1134 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1141 * We don't have to hold the mutex to check db_state because it
1142 * can't be freed while we have a hold on the buffer.
1144 ASSERT(!refcount_is_zero(&db
->db_holds
));
1146 if (db
->db_state
== DB_NOFILL
)
1147 return (SET_ERROR(EIO
));
1151 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1152 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1154 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1155 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1156 DBUF_IS_CACHEABLE(db
);
1158 mutex_enter(&db
->db_mtx
);
1159 if (db
->db_state
== DB_CACHED
) {
1161 * If the arc buf is compressed, we need to decompress it to
1162 * read the data. This could happen during the "zfs receive" of
1163 * a stream which is compressed and deduplicated.
1165 if (db
->db_buf
!= NULL
&&
1166 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
) {
1167 dbuf_fix_old_data(db
,
1168 spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1169 err
= arc_decompress(db
->db_buf
);
1170 dbuf_set_data(db
, db
->db_buf
);
1172 mutex_exit(&db
->db_mtx
);
1174 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1175 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1176 rw_exit(&dn
->dn_struct_rwlock
);
1178 } else if (db
->db_state
== DB_UNCACHED
) {
1179 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1180 boolean_t need_wait
= B_FALSE
;
1183 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1184 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1187 dbuf_read_impl(db
, zio
, flags
);
1189 /* dbuf_read_impl has dropped db_mtx for us */
1192 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1194 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1195 rw_exit(&dn
->dn_struct_rwlock
);
1199 err
= zio_wait(zio
);
1202 * Another reader came in while the dbuf was in flight
1203 * between UNCACHED and CACHED. Either a writer will finish
1204 * writing the buffer (sending the dbuf to CACHED) or the
1205 * first reader's request will reach the read_done callback
1206 * and send the dbuf to CACHED. Otherwise, a failure
1207 * occurred and the dbuf went to UNCACHED.
1209 mutex_exit(&db
->db_mtx
);
1211 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1212 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1213 rw_exit(&dn
->dn_struct_rwlock
);
1216 /* Skip the wait per the caller's request. */
1217 mutex_enter(&db
->db_mtx
);
1218 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1219 while (db
->db_state
== DB_READ
||
1220 db
->db_state
== DB_FILL
) {
1221 ASSERT(db
->db_state
== DB_READ
||
1222 (flags
& DB_RF_HAVESTRUCT
) == 0);
1223 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1225 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1227 if (db
->db_state
== DB_UNCACHED
)
1228 err
= SET_ERROR(EIO
);
1230 mutex_exit(&db
->db_mtx
);
1237 dbuf_noread(dmu_buf_impl_t
*db
)
1239 ASSERT(!refcount_is_zero(&db
->db_holds
));
1240 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1241 mutex_enter(&db
->db_mtx
);
1242 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1243 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1244 if (db
->db_state
== DB_UNCACHED
) {
1245 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1246 spa_t
*spa
= db
->db_objset
->os_spa
;
1248 ASSERT(db
->db_buf
== NULL
);
1249 ASSERT(db
->db
.db_data
== NULL
);
1250 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1251 db
->db_state
= DB_FILL
;
1252 } else if (db
->db_state
== DB_NOFILL
) {
1253 dbuf_clear_data(db
);
1255 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1257 mutex_exit(&db
->db_mtx
);
1261 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1263 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1264 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1265 uint64_t txg
= dr
->dr_txg
;
1267 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1269 * This assert is valid because dmu_sync() expects to be called by
1270 * a zilog's get_data while holding a range lock. This call only
1271 * comes from dbuf_dirty() callers who must also hold a range lock.
1273 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1274 ASSERT(db
->db_level
== 0);
1276 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1277 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1280 ASSERT(db
->db_data_pending
!= dr
);
1282 /* free this block */
1283 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1284 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1286 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1287 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1290 * Release the already-written buffer, so we leave it in
1291 * a consistent dirty state. Note that all callers are
1292 * modifying the buffer, so they will immediately do
1293 * another (redundant) arc_release(). Therefore, leave
1294 * the buf thawed to save the effort of freezing &
1295 * immediately re-thawing it.
1297 arc_release(dr
->dt
.dl
.dr_data
, db
);
1301 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1302 * data blocks in the free range, so that any future readers will find
1306 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1309 dmu_buf_impl_t db_search
;
1310 dmu_buf_impl_t
*db
, *db_next
;
1311 uint64_t txg
= tx
->tx_txg
;
1314 if (end_blkid
> dn
->dn_maxblkid
&&
1315 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1316 end_blkid
= dn
->dn_maxblkid
;
1317 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1319 db_search
.db_level
= 0;
1320 db_search
.db_blkid
= start_blkid
;
1321 db_search
.db_state
= DB_SEARCH
;
1323 mutex_enter(&dn
->dn_dbufs_mtx
);
1324 db
= avl_find(&dn
->dn_dbufs
, &db_search
, &where
);
1325 ASSERT3P(db
, ==, NULL
);
1327 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1329 for (; db
!= NULL
; db
= db_next
) {
1330 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1331 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1333 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1336 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1338 /* found a level 0 buffer in the range */
1339 mutex_enter(&db
->db_mtx
);
1340 if (dbuf_undirty(db
, tx
)) {
1341 /* mutex has been dropped and dbuf destroyed */
1345 if (db
->db_state
== DB_UNCACHED
||
1346 db
->db_state
== DB_NOFILL
||
1347 db
->db_state
== DB_EVICTING
) {
1348 ASSERT(db
->db
.db_data
== NULL
);
1349 mutex_exit(&db
->db_mtx
);
1352 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1353 /* will be handled in dbuf_read_done or dbuf_rele */
1354 db
->db_freed_in_flight
= TRUE
;
1355 mutex_exit(&db
->db_mtx
);
1358 if (refcount_count(&db
->db_holds
) == 0) {
1363 /* The dbuf is referenced */
1365 if (db
->db_last_dirty
!= NULL
) {
1366 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1368 if (dr
->dr_txg
== txg
) {
1370 * This buffer is "in-use", re-adjust the file
1371 * size to reflect that this buffer may
1372 * contain new data when we sync.
1374 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1375 db
->db_blkid
> dn
->dn_maxblkid
)
1376 dn
->dn_maxblkid
= db
->db_blkid
;
1377 dbuf_unoverride(dr
);
1380 * This dbuf is not dirty in the open context.
1381 * Either uncache it (if its not referenced in
1382 * the open context) or reset its contents to
1385 dbuf_fix_old_data(db
, txg
);
1388 /* clear the contents if its cached */
1389 if (db
->db_state
== DB_CACHED
) {
1390 ASSERT(db
->db
.db_data
!= NULL
);
1391 arc_release(db
->db_buf
, db
);
1392 bzero(db
->db
.db_data
, db
->db
.db_size
);
1393 arc_buf_freeze(db
->db_buf
);
1396 mutex_exit(&db
->db_mtx
);
1398 mutex_exit(&dn
->dn_dbufs_mtx
);
1402 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1404 arc_buf_t
*buf
, *obuf
;
1405 int osize
= db
->db
.db_size
;
1406 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1409 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1414 /* XXX does *this* func really need the lock? */
1415 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1418 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1419 * is OK, because there can be no other references to the db
1420 * when we are changing its size, so no concurrent DB_FILL can
1424 * XXX we should be doing a dbuf_read, checking the return
1425 * value and returning that up to our callers
1427 dmu_buf_will_dirty(&db
->db
, tx
);
1429 /* create the data buffer for the new block */
1430 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1432 /* copy old block data to the new block */
1434 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1435 /* zero the remainder */
1437 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1439 mutex_enter(&db
->db_mtx
);
1440 dbuf_set_data(db
, buf
);
1441 arc_buf_destroy(obuf
, db
);
1442 db
->db
.db_size
= size
;
1444 if (db
->db_level
== 0) {
1445 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1446 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1448 mutex_exit(&db
->db_mtx
);
1450 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1455 dbuf_release_bp(dmu_buf_impl_t
*db
)
1457 objset_t
*os
= db
->db_objset
;
1459 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1460 ASSERT(arc_released(os
->os_phys_buf
) ||
1461 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1462 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1464 (void) arc_release(db
->db_buf
, db
);
1468 * We already have a dirty record for this TXG, and we are being
1472 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1474 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1476 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1478 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1480 * If this buffer has already been written out,
1481 * we now need to reset its state.
1483 dbuf_unoverride(dr
);
1484 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1485 db
->db_state
!= DB_NOFILL
) {
1486 /* Already released on initial dirty, so just thaw. */
1487 ASSERT(arc_released(db
->db_buf
));
1488 arc_buf_thaw(db
->db_buf
);
1493 dbuf_dirty_record_t
*
1494 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1498 dbuf_dirty_record_t
**drp
, *dr
;
1499 int drop_struct_lock
= FALSE
;
1500 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1502 ASSERT(tx
->tx_txg
!= 0);
1503 ASSERT(!refcount_is_zero(&db
->db_holds
));
1504 DMU_TX_DIRTY_BUF(tx
, db
);
1509 * Shouldn't dirty a regular buffer in syncing context. Private
1510 * objects may be dirtied in syncing context, but only if they
1511 * were already pre-dirtied in open context.
1514 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1515 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1518 ASSERT(!dmu_tx_is_syncing(tx
) ||
1519 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1520 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1521 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1522 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1523 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1526 * We make this assert for private objects as well, but after we
1527 * check if we're already dirty. They are allowed to re-dirty
1528 * in syncing context.
1530 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1531 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1532 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1534 mutex_enter(&db
->db_mtx
);
1536 * XXX make this true for indirects too? The problem is that
1537 * transactions created with dmu_tx_create_assigned() from
1538 * syncing context don't bother holding ahead.
1540 ASSERT(db
->db_level
!= 0 ||
1541 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1542 db
->db_state
== DB_NOFILL
);
1544 mutex_enter(&dn
->dn_mtx
);
1546 * Don't set dirtyctx to SYNC if we're just modifying this as we
1547 * initialize the objset.
1549 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1550 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1551 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1554 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1555 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1556 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1557 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1558 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1560 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1561 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1565 mutex_exit(&dn
->dn_mtx
);
1567 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1568 dn
->dn_have_spill
= B_TRUE
;
1571 * If this buffer is already dirty, we're done.
1573 drp
= &db
->db_last_dirty
;
1574 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1575 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1576 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1578 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1582 mutex_exit(&db
->db_mtx
);
1587 * Only valid if not already dirty.
1589 ASSERT(dn
->dn_object
== 0 ||
1590 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1591 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1593 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1596 * We should only be dirtying in syncing context if it's the
1597 * mos or we're initializing the os or it's a special object.
1598 * However, we are allowed to dirty in syncing context provided
1599 * we already dirtied it in open context. Hence we must make
1600 * this assertion only if we're not already dirty.
1603 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
1605 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1606 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
1607 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1608 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1609 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1610 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1612 ASSERT(db
->db
.db_size
!= 0);
1614 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1616 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1617 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
1621 * If this buffer is dirty in an old transaction group we need
1622 * to make a copy of it so that the changes we make in this
1623 * transaction group won't leak out when we sync the older txg.
1625 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1626 if (db
->db_level
== 0) {
1627 void *data_old
= db
->db_buf
;
1629 if (db
->db_state
!= DB_NOFILL
) {
1630 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1631 dbuf_fix_old_data(db
, tx
->tx_txg
);
1632 data_old
= db
->db
.db_data
;
1633 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1635 * Release the data buffer from the cache so
1636 * that we can modify it without impacting
1637 * possible other users of this cached data
1638 * block. Note that indirect blocks and
1639 * private objects are not released until the
1640 * syncing state (since they are only modified
1643 arc_release(db
->db_buf
, db
);
1644 dbuf_fix_old_data(db
, tx
->tx_txg
);
1645 data_old
= db
->db_buf
;
1647 ASSERT(data_old
!= NULL
);
1649 dr
->dt
.dl
.dr_data
= data_old
;
1651 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
1652 list_create(&dr
->dt
.di
.dr_children
,
1653 sizeof (dbuf_dirty_record_t
),
1654 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1656 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1657 dr
->dr_accounted
= db
->db
.db_size
;
1659 dr
->dr_txg
= tx
->tx_txg
;
1664 * We could have been freed_in_flight between the dbuf_noread
1665 * and dbuf_dirty. We win, as though the dbuf_noread() had
1666 * happened after the free.
1668 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1669 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1670 mutex_enter(&dn
->dn_mtx
);
1671 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1672 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1675 mutex_exit(&dn
->dn_mtx
);
1676 db
->db_freed_in_flight
= FALSE
;
1680 * This buffer is now part of this txg
1682 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1683 db
->db_dirtycnt
+= 1;
1684 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1686 mutex_exit(&db
->db_mtx
);
1688 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1689 db
->db_blkid
== DMU_SPILL_BLKID
) {
1690 mutex_enter(&dn
->dn_mtx
);
1691 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1692 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1693 mutex_exit(&dn
->dn_mtx
);
1694 dnode_setdirty(dn
, tx
);
1700 * The dn_struct_rwlock prevents db_blkptr from changing
1701 * due to a write from syncing context completing
1702 * while we are running, so we want to acquire it before
1703 * looking at db_blkptr.
1705 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1706 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1707 drop_struct_lock
= TRUE
;
1711 * We need to hold the dn_struct_rwlock to make this assertion,
1712 * because it protects dn_phys / dn_next_nlevels from changing.
1714 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1715 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1716 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1717 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1718 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1721 * If we are overwriting a dedup BP, then unless it is snapshotted,
1722 * when we get to syncing context we will need to decrement its
1723 * refcount in the DDT. Prefetch the relevant DDT block so that
1724 * syncing context won't have to wait for the i/o.
1726 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
1728 if (db
->db_level
== 0) {
1729 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1730 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1733 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1734 dmu_buf_impl_t
*parent
= db
->db_parent
;
1735 dbuf_dirty_record_t
*di
;
1736 int parent_held
= FALSE
;
1738 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1739 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1741 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1742 db
->db_blkid
>> epbs
, FTAG
);
1743 ASSERT(parent
!= NULL
);
1746 if (drop_struct_lock
)
1747 rw_exit(&dn
->dn_struct_rwlock
);
1748 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1749 di
= dbuf_dirty(parent
, tx
);
1751 dbuf_rele(parent
, FTAG
);
1753 mutex_enter(&db
->db_mtx
);
1755 * Since we've dropped the mutex, it's possible that
1756 * dbuf_undirty() might have changed this out from under us.
1758 if (db
->db_last_dirty
== dr
||
1759 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1760 mutex_enter(&di
->dt
.di
.dr_mtx
);
1761 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1762 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1763 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1764 mutex_exit(&di
->dt
.di
.dr_mtx
);
1767 mutex_exit(&db
->db_mtx
);
1769 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1770 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1771 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1772 mutex_enter(&dn
->dn_mtx
);
1773 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1774 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1775 mutex_exit(&dn
->dn_mtx
);
1776 if (drop_struct_lock
)
1777 rw_exit(&dn
->dn_struct_rwlock
);
1780 dnode_setdirty(dn
, tx
);
1786 * Undirty a buffer in the transaction group referenced by the given
1787 * transaction. Return whether this evicted the dbuf.
1790 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1793 uint64_t txg
= tx
->tx_txg
;
1794 dbuf_dirty_record_t
*dr
, **drp
;
1799 * Due to our use of dn_nlevels below, this can only be called
1800 * in open context, unless we are operating on the MOS.
1801 * From syncing context, dn_nlevels may be different from the
1802 * dn_nlevels used when dbuf was dirtied.
1804 ASSERT(db
->db_objset
==
1805 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1806 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1807 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1808 ASSERT0(db
->db_level
);
1809 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1812 * If this buffer is not dirty, we're done.
1814 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1815 if (dr
->dr_txg
<= txg
)
1817 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1819 ASSERT(dr
->dr_txg
== txg
);
1820 ASSERT(dr
->dr_dbuf
== db
);
1825 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1827 ASSERT(db
->db
.db_size
!= 0);
1829 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1830 dr
->dr_accounted
, txg
);
1835 * Note that there are three places in dbuf_dirty()
1836 * where this dirty record may be put on a list.
1837 * Make sure to do a list_remove corresponding to
1838 * every one of those list_insert calls.
1840 if (dr
->dr_parent
) {
1841 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1842 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1843 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1844 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1845 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1846 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1847 mutex_enter(&dn
->dn_mtx
);
1848 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1849 mutex_exit(&dn
->dn_mtx
);
1853 if (db
->db_state
!= DB_NOFILL
) {
1854 dbuf_unoverride(dr
);
1856 ASSERT(db
->db_buf
!= NULL
);
1857 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1858 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1859 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
1862 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1864 ASSERT(db
->db_dirtycnt
> 0);
1865 db
->db_dirtycnt
-= 1;
1867 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1868 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
1877 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1879 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1880 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1882 ASSERT(tx
->tx_txg
!= 0);
1883 ASSERT(!refcount_is_zero(&db
->db_holds
));
1886 * Quick check for dirtyness. For already dirty blocks, this
1887 * reduces runtime of this function by >90%, and overall performance
1888 * by 50% for some workloads (e.g. file deletion with indirect blocks
1891 mutex_enter(&db
->db_mtx
);
1892 dbuf_dirty_record_t
*dr
;
1893 for (dr
= db
->db_last_dirty
;
1894 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1896 * It's possible that it is already dirty but not cached,
1897 * because there are some calls to dbuf_dirty() that don't
1898 * go through dmu_buf_will_dirty().
1900 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1901 /* This dbuf is already dirty and cached. */
1903 mutex_exit(&db
->db_mtx
);
1907 mutex_exit(&db
->db_mtx
);
1910 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1911 rf
|= DB_RF_HAVESTRUCT
;
1913 (void) dbuf_read(db
, NULL
, rf
);
1914 (void) dbuf_dirty(db
, tx
);
1918 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1920 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1922 db
->db_state
= DB_NOFILL
;
1924 dmu_buf_will_fill(db_fake
, tx
);
1928 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1930 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1932 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1933 ASSERT(tx
->tx_txg
!= 0);
1934 ASSERT(db
->db_level
== 0);
1935 ASSERT(!refcount_is_zero(&db
->db_holds
));
1937 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1938 dmu_tx_private_ok(tx
));
1941 (void) dbuf_dirty(db
, tx
);
1944 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1947 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1949 mutex_enter(&db
->db_mtx
);
1952 if (db
->db_state
== DB_FILL
) {
1953 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1954 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1955 /* we were freed while filling */
1956 /* XXX dbuf_undirty? */
1957 bzero(db
->db
.db_data
, db
->db
.db_size
);
1958 db
->db_freed_in_flight
= FALSE
;
1960 db
->db_state
= DB_CACHED
;
1961 cv_broadcast(&db
->db_changed
);
1963 mutex_exit(&db
->db_mtx
);
1967 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
1968 bp_embedded_type_t etype
, enum zio_compress comp
,
1969 int uncompressed_size
, int compressed_size
, int byteorder
,
1972 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
1973 struct dirty_leaf
*dl
;
1974 dmu_object_type_t type
;
1976 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
1977 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
1978 SPA_FEATURE_EMBEDDED_DATA
));
1982 type
= DB_DNODE(db
)->dn_type
;
1985 ASSERT0(db
->db_level
);
1986 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1988 dmu_buf_will_not_fill(dbuf
, tx
);
1990 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1991 dl
= &db
->db_last_dirty
->dt
.dl
;
1992 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
1993 data
, comp
, uncompressed_size
, compressed_size
);
1994 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
1995 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
1996 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
1997 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
1999 dl
->dr_override_state
= DR_OVERRIDDEN
;
2000 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2004 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2005 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2008 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2010 ASSERT(!refcount_is_zero(&db
->db_holds
));
2011 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2012 ASSERT(db
->db_level
== 0);
2013 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2014 ASSERT(buf
!= NULL
);
2015 ASSERT(arc_buf_lsize(buf
) == db
->db
.db_size
);
2016 ASSERT(tx
->tx_txg
!= 0);
2018 arc_return_buf(buf
, db
);
2019 ASSERT(arc_released(buf
));
2021 mutex_enter(&db
->db_mtx
);
2023 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2024 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2026 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2028 if (db
->db_state
== DB_CACHED
&&
2029 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2030 mutex_exit(&db
->db_mtx
);
2031 (void) dbuf_dirty(db
, tx
);
2032 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2033 arc_buf_destroy(buf
, db
);
2034 xuio_stat_wbuf_copied();
2038 xuio_stat_wbuf_nocopy();
2039 if (db
->db_state
== DB_CACHED
) {
2040 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2042 ASSERT(db
->db_buf
!= NULL
);
2043 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2044 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2045 if (!arc_released(db
->db_buf
)) {
2046 ASSERT(dr
->dt
.dl
.dr_override_state
==
2048 arc_release(db
->db_buf
, db
);
2050 dr
->dt
.dl
.dr_data
= buf
;
2051 arc_buf_destroy(db
->db_buf
, db
);
2052 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2053 arc_release(db
->db_buf
, db
);
2054 arc_buf_destroy(db
->db_buf
, db
);
2058 ASSERT(db
->db_buf
== NULL
);
2059 dbuf_set_data(db
, buf
);
2060 db
->db_state
= DB_FILL
;
2061 mutex_exit(&db
->db_mtx
);
2062 (void) dbuf_dirty(db
, tx
);
2063 dmu_buf_fill_done(&db
->db
, tx
);
2067 dbuf_destroy(dmu_buf_impl_t
*db
)
2070 dmu_buf_impl_t
*parent
= db
->db_parent
;
2071 dmu_buf_impl_t
*dndb
;
2073 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2074 ASSERT(refcount_is_zero(&db
->db_holds
));
2076 if (db
->db_buf
!= NULL
) {
2077 arc_buf_destroy(db
->db_buf
, db
);
2081 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2082 ASSERT(db
->db
.db_data
!= NULL
);
2083 zio_buf_free(db
->db
.db_data
, DN_MAX_BONUSLEN
);
2084 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
2085 db
->db_state
= DB_UNCACHED
;
2088 dbuf_clear_data(db
);
2090 if (multilist_link_active(&db
->db_cache_link
)) {
2091 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
2092 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
2094 multilist_remove(dbuf_caches
[db
->db_caching_status
].cache
, db
);
2095 (void) refcount_remove_many(
2096 &dbuf_caches
[db
->db_caching_status
].size
,
2097 db
->db
.db_size
, db
);
2099 db
->db_caching_status
= DB_NO_CACHE
;
2102 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2103 ASSERT(db
->db_data_pending
== NULL
);
2105 db
->db_state
= DB_EVICTING
;
2106 db
->db_blkptr
= NULL
;
2109 * Now that db_state is DB_EVICTING, nobody else can find this via
2110 * the hash table. We can now drop db_mtx, which allows us to
2111 * acquire the dn_dbufs_mtx.
2113 mutex_exit(&db
->db_mtx
);
2118 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2119 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2121 mutex_enter(&dn
->dn_dbufs_mtx
);
2122 avl_remove(&dn
->dn_dbufs
, db
);
2123 atomic_dec_32(&dn
->dn_dbufs_count
);
2127 mutex_exit(&dn
->dn_dbufs_mtx
);
2129 * Decrementing the dbuf count means that the hold corresponding
2130 * to the removed dbuf is no longer discounted in dnode_move(),
2131 * so the dnode cannot be moved until after we release the hold.
2132 * The membar_producer() ensures visibility of the decremented
2133 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2136 mutex_enter(&dn
->dn_mtx
);
2137 dnode_rele_and_unlock(dn
, db
, B_TRUE
);
2138 db
->db_dnode_handle
= NULL
;
2140 dbuf_hash_remove(db
);
2145 ASSERT(refcount_is_zero(&db
->db_holds
));
2147 db
->db_parent
= NULL
;
2149 ASSERT(db
->db_buf
== NULL
);
2150 ASSERT(db
->db
.db_data
== NULL
);
2151 ASSERT(db
->db_hash_next
== NULL
);
2152 ASSERT(db
->db_blkptr
== NULL
);
2153 ASSERT(db
->db_data_pending
== NULL
);
2154 ASSERT3U(db
->db_caching_status
, ==, DB_NO_CACHE
);
2155 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2157 kmem_cache_free(dbuf_kmem_cache
, db
);
2158 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2161 * If this dbuf is referenced from an indirect dbuf,
2162 * decrement the ref count on the indirect dbuf.
2164 if (parent
&& parent
!= dndb
) {
2165 mutex_enter(&parent
->db_mtx
);
2166 dbuf_rele_and_unlock(parent
, db
, B_TRUE
);
2171 * Note: While bpp will always be updated if the function returns success,
2172 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2173 * this happens when the dnode is the meta-dnode, or a userused or groupused
2177 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2178 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
2183 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2185 if (blkid
== DMU_SPILL_BLKID
) {
2186 mutex_enter(&dn
->dn_mtx
);
2187 if (dn
->dn_have_spill
&&
2188 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2189 *bpp
= &dn
->dn_phys
->dn_spill
;
2192 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2193 *parentp
= dn
->dn_dbuf
;
2194 mutex_exit(&dn
->dn_mtx
);
2199 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2200 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2202 ASSERT3U(level
* epbs
, <, 64);
2203 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2205 * This assertion shouldn't trip as long as the max indirect block size
2206 * is less than 1M. The reason for this is that up to that point,
2207 * the number of levels required to address an entire object with blocks
2208 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2209 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2210 * (i.e. we can address the entire object), objects will all use at most
2211 * N-1 levels and the assertion won't overflow. However, once epbs is
2212 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2213 * enough to address an entire object, so objects will have 5 levels,
2214 * but then this assertion will overflow.
2216 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2217 * need to redo this logic to handle overflows.
2219 ASSERT(level
>= nlevels
||
2220 ((nlevels
- level
- 1) * epbs
) +
2221 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2222 if (level
>= nlevels
||
2223 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2224 ((nlevels
- level
- 1) * epbs
)) ||
2226 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2227 /* the buffer has no parent yet */
2228 return (SET_ERROR(ENOENT
));
2229 } else if (level
< nlevels
-1) {
2230 /* this block is referenced from an indirect block */
2231 int err
= dbuf_hold_impl(dn
, level
+1,
2232 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2235 err
= dbuf_read(*parentp
, NULL
,
2236 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2238 dbuf_rele(*parentp
, NULL
);
2242 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2243 (blkid
& ((1ULL << epbs
) - 1));
2244 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2245 ASSERT(BP_IS_HOLE(*bpp
));
2248 /* the block is referenced from the dnode */
2249 ASSERT3U(level
, ==, nlevels
-1);
2250 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2251 blkid
< dn
->dn_phys
->dn_nblkptr
);
2253 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2254 *parentp
= dn
->dn_dbuf
;
2256 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2261 static dmu_buf_impl_t
*
2262 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2263 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2265 objset_t
*os
= dn
->dn_objset
;
2266 dmu_buf_impl_t
*db
, *odb
;
2268 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2269 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2271 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2274 db
->db
.db_object
= dn
->dn_object
;
2275 db
->db_level
= level
;
2276 db
->db_blkid
= blkid
;
2277 db
->db_last_dirty
= NULL
;
2278 db
->db_dirtycnt
= 0;
2279 db
->db_dnode_handle
= dn
->dn_handle
;
2280 db
->db_parent
= parent
;
2281 db
->db_blkptr
= blkptr
;
2284 db
->db_user_immediate_evict
= FALSE
;
2285 db
->db_freed_in_flight
= FALSE
;
2286 db
->db_pending_evict
= FALSE
;
2288 if (blkid
== DMU_BONUS_BLKID
) {
2289 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2290 db
->db
.db_size
= DN_MAX_BONUSLEN
-
2291 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2292 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2293 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2294 db
->db_state
= DB_UNCACHED
;
2295 db
->db_caching_status
= DB_NO_CACHE
;
2296 /* the bonus dbuf is not placed in the hash table */
2297 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2299 } else if (blkid
== DMU_SPILL_BLKID
) {
2300 db
->db
.db_size
= (blkptr
!= NULL
) ?
2301 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2302 db
->db
.db_offset
= 0;
2305 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2306 db
->db
.db_size
= blocksize
;
2307 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2311 * Hold the dn_dbufs_mtx while we get the new dbuf
2312 * in the hash table *and* added to the dbufs list.
2313 * This prevents a possible deadlock with someone
2314 * trying to look up this dbuf before its added to the
2317 mutex_enter(&dn
->dn_dbufs_mtx
);
2318 db
->db_state
= DB_EVICTING
;
2319 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2320 /* someone else inserted it first */
2321 kmem_cache_free(dbuf_kmem_cache
, db
);
2322 mutex_exit(&dn
->dn_dbufs_mtx
);
2325 avl_add(&dn
->dn_dbufs
, db
);
2327 db
->db_state
= DB_UNCACHED
;
2328 db
->db_caching_status
= DB_NO_CACHE
;
2329 mutex_exit(&dn
->dn_dbufs_mtx
);
2330 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2332 if (parent
&& parent
!= dn
->dn_dbuf
)
2333 dbuf_add_ref(parent
, db
);
2335 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2336 refcount_count(&dn
->dn_holds
) > 0);
2337 (void) refcount_add(&dn
->dn_holds
, db
);
2338 atomic_inc_32(&dn
->dn_dbufs_count
);
2340 dprintf_dbuf(db
, "db=%p\n", db
);
2345 typedef struct dbuf_prefetch_arg
{
2346 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2347 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2348 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2349 int dpa_curlevel
; /* The current level that we're reading */
2350 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2351 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2352 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2353 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2354 } dbuf_prefetch_arg_t
;
2357 * Actually issue the prefetch read for the block given.
2360 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2362 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2365 arc_flags_t aflags
=
2366 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2368 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2369 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2370 ASSERT(dpa
->dpa_zio
!= NULL
);
2371 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2372 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2373 &aflags
, &dpa
->dpa_zb
);
2377 * Called when an indirect block above our prefetch target is read in. This
2378 * will either read in the next indirect block down the tree or issue the actual
2379 * prefetch if the next block down is our target.
2382 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2384 dbuf_prefetch_arg_t
*dpa
= private;
2386 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2387 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2390 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
2391 kmem_free(dpa
, sizeof (*dpa
));
2394 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
2397 * The dpa_dnode is only valid if we are called with a NULL
2398 * zio. This indicates that the arc_read() returned without
2399 * first calling zio_read() to issue a physical read. Once
2400 * a physical read is made the dpa_dnode must be invalidated
2401 * as the locks guarding it may have been dropped. If the
2402 * dpa_dnode is still valid, then we want to add it to the dbuf
2403 * cache. To do so, we must hold the dbuf associated with the block
2404 * we just prefetched, read its contents so that we associate it
2405 * with an arc_buf_t, and then release it.
2408 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2409 if (zio
->io_flags
& ZIO_FLAG_RAW
) {
2410 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2412 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2414 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2416 dpa
->dpa_dnode
= NULL
;
2417 } else if (dpa
->dpa_dnode
!= NULL
) {
2418 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2419 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2420 dpa
->dpa_zb
.zb_level
));
2421 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2422 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2423 (void) dbuf_read(db
, NULL
,
2424 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2425 dbuf_rele(db
, FTAG
);
2428 dpa
->dpa_curlevel
--;
2430 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2431 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2432 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
2433 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2434 if (BP_IS_HOLE(bp
)) {
2435 kmem_free(dpa
, sizeof (*dpa
));
2436 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2437 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2438 dbuf_issue_final_prefetch(dpa
, bp
);
2439 kmem_free(dpa
, sizeof (*dpa
));
2441 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2442 zbookmark_phys_t zb
;
2444 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2445 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
2446 iter_aflags
|= ARC_FLAG_L2CACHE
;
2448 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2450 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2451 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2453 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2454 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2455 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2459 arc_buf_destroy(abuf
, private);
2463 * Issue prefetch reads for the given block on the given level. If the indirect
2464 * blocks above that block are not in memory, we will read them in
2465 * asynchronously. As a result, this call never blocks waiting for a read to
2469 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2473 int epbs
, nlevels
, curlevel
;
2476 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2477 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2479 if (blkid
> dn
->dn_maxblkid
)
2482 if (dnode_block_freed(dn
, blkid
))
2486 * This dnode hasn't been written to disk yet, so there's nothing to
2489 nlevels
= dn
->dn_phys
->dn_nlevels
;
2490 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2493 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2494 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2497 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2500 mutex_exit(&db
->db_mtx
);
2502 * This dbuf already exists. It is either CACHED, or
2503 * (we assume) about to be read or filled.
2509 * Find the closest ancestor (indirect block) of the target block
2510 * that is present in the cache. In this indirect block, we will
2511 * find the bp that is at curlevel, curblkid.
2515 while (curlevel
< nlevels
- 1) {
2516 int parent_level
= curlevel
+ 1;
2517 uint64_t parent_blkid
= curblkid
>> epbs
;
2520 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2521 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2522 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2523 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2524 dbuf_rele(db
, FTAG
);
2528 curlevel
= parent_level
;
2529 curblkid
= parent_blkid
;
2532 if (curlevel
== nlevels
- 1) {
2533 /* No cached indirect blocks found. */
2534 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2535 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2537 if (BP_IS_HOLE(&bp
))
2540 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2542 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2545 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2546 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
2547 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2548 dn
->dn_object
, level
, blkid
);
2549 dpa
->dpa_curlevel
= curlevel
;
2550 dpa
->dpa_prio
= prio
;
2551 dpa
->dpa_aflags
= aflags
;
2552 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2553 dpa
->dpa_dnode
= dn
;
2554 dpa
->dpa_epbs
= epbs
;
2557 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2558 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
2559 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
2562 * If we have the indirect just above us, no need to do the asynchronous
2563 * prefetch chain; we'll just run the last step ourselves. If we're at
2564 * a higher level, though, we want to issue the prefetches for all the
2565 * indirect blocks asynchronously, so we can go on with whatever we were
2568 if (curlevel
== level
) {
2569 ASSERT3U(curblkid
, ==, blkid
);
2570 dbuf_issue_final_prefetch(dpa
, &bp
);
2571 kmem_free(dpa
, sizeof (*dpa
));
2573 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2574 zbookmark_phys_t zb
;
2576 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2577 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
2578 iter_aflags
|= ARC_FLAG_L2CACHE
;
2580 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2581 dn
->dn_object
, curlevel
, curblkid
);
2582 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2583 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2584 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2588 * We use pio here instead of dpa_zio since it's possible that
2589 * dpa may have already been freed.
2595 * Returns with db_holds incremented, and db_mtx not held.
2596 * Note: dn_struct_rwlock must be held.
2599 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2600 boolean_t fail_sparse
, boolean_t fail_uncached
,
2601 void *tag
, dmu_buf_impl_t
**dbp
)
2603 dmu_buf_impl_t
*db
, *parent
= NULL
;
2605 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2606 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2607 ASSERT3U(dn
->dn_nlevels
, >, level
);
2611 /* dbuf_find() returns with db_mtx held */
2612 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
);
2615 blkptr_t
*bp
= NULL
;
2619 return (SET_ERROR(ENOENT
));
2621 ASSERT3P(parent
, ==, NULL
);
2622 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
2624 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
2625 err
= SET_ERROR(ENOENT
);
2628 dbuf_rele(parent
, NULL
);
2632 if (err
&& err
!= ENOENT
)
2634 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
);
2637 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
2638 mutex_exit(&db
->db_mtx
);
2639 return (SET_ERROR(ENOENT
));
2642 if (db
->db_buf
!= NULL
)
2643 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
2645 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
2648 * If this buffer is currently syncing out, and we are are
2649 * still referencing it from db_data, we need to make a copy
2650 * of it in case we decide we want to dirty it again in this txg.
2652 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2653 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2654 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
2655 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
2657 if (dr
->dt
.dl
.dr_data
== db
->db_buf
) {
2658 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2661 arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
,
2663 bcopy(dr
->dt
.dl
.dr_data
->b_data
, db
->db
.db_data
,
2668 if (multilist_link_active(&db
->db_cache_link
)) {
2669 ASSERT(refcount_is_zero(&db
->db_holds
));
2670 ASSERT(db
->db_caching_status
== DB_DBUF_CACHE
||
2671 db
->db_caching_status
== DB_DBUF_METADATA_CACHE
);
2673 multilist_remove(dbuf_caches
[db
->db_caching_status
].cache
, db
);
2674 (void) refcount_remove_many(
2675 &dbuf_caches
[db
->db_caching_status
].size
,
2676 db
->db
.db_size
, db
);
2678 db
->db_caching_status
= DB_NO_CACHE
;
2680 (void) refcount_add(&db
->db_holds
, tag
);
2682 mutex_exit(&db
->db_mtx
);
2684 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2686 dbuf_rele(parent
, NULL
);
2688 ASSERT3P(DB_DNODE(db
), ==, dn
);
2689 ASSERT3U(db
->db_blkid
, ==, blkid
);
2690 ASSERT3U(db
->db_level
, ==, level
);
2697 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2699 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2703 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2706 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2707 return (err
? NULL
: db
);
2711 dbuf_create_bonus(dnode_t
*dn
)
2713 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2715 ASSERT(dn
->dn_bonus
== NULL
);
2716 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2720 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2722 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2725 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2726 return (SET_ERROR(ENOTSUP
));
2728 blksz
= SPA_MINBLOCKSIZE
;
2729 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2730 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2734 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2735 dbuf_new_size(db
, blksz
, tx
);
2736 rw_exit(&dn
->dn_struct_rwlock
);
2743 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2745 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2748 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2750 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2752 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
2753 ASSERT3S(holds
, >, 1);
2756 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2758 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2761 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2762 dmu_buf_impl_t
*found_db
;
2763 boolean_t result
= B_FALSE
;
2765 if (db
->db_blkid
== DMU_BONUS_BLKID
)
2766 found_db
= dbuf_find_bonus(os
, obj
);
2768 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2770 if (found_db
!= NULL
) {
2771 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2772 (void) refcount_add(&db
->db_holds
, tag
);
2775 mutex_exit(&db
->db_mtx
);
2781 * If you call dbuf_rele() you had better not be referencing the dnode handle
2782 * unless you have some other direct or indirect hold on the dnode. (An indirect
2783 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2784 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2785 * dnode's parent dbuf evicting its dnode handles.
2788 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2790 mutex_enter(&db
->db_mtx
);
2791 dbuf_rele_and_unlock(db
, tag
, B_FALSE
);
2795 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2797 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2801 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2802 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
2803 * argument should be set if we are already in the dbuf-evicting code
2804 * path, in which case we don't want to recursively evict. This allows us to
2805 * avoid deeply nested stacks that would have a call flow similar to this:
2807 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
2810 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
2814 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
, boolean_t evicting
)
2818 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2822 * Remove the reference to the dbuf before removing its hold on the
2823 * dnode so we can guarantee in dnode_move() that a referenced bonus
2824 * buffer has a corresponding dnode hold.
2826 holds
= refcount_remove(&db
->db_holds
, tag
);
2830 * We can't freeze indirects if there is a possibility that they
2831 * may be modified in the current syncing context.
2833 if (db
->db_buf
!= NULL
&&
2834 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
2835 arc_buf_freeze(db
->db_buf
);
2838 if (holds
== db
->db_dirtycnt
&&
2839 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2840 dbuf_evict_user(db
);
2843 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2845 boolean_t evict_dbuf
= db
->db_pending_evict
;
2848 * If the dnode moves here, we cannot cross this
2849 * barrier until the move completes.
2854 atomic_dec_32(&dn
->dn_dbufs_count
);
2857 * Decrementing the dbuf count means that the bonus
2858 * buffer's dnode hold is no longer discounted in
2859 * dnode_move(). The dnode cannot move until after
2860 * the dnode_rele() below.
2865 * Do not reference db after its lock is dropped.
2866 * Another thread may evict it.
2868 mutex_exit(&db
->db_mtx
);
2871 dnode_evict_bonus(dn
);
2874 } else if (db
->db_buf
== NULL
) {
2876 * This is a special case: we never associated this
2877 * dbuf with any data allocated from the ARC.
2879 ASSERT(db
->db_state
== DB_UNCACHED
||
2880 db
->db_state
== DB_NOFILL
);
2882 } else if (arc_released(db
->db_buf
)) {
2884 * This dbuf has anonymous data associated with it.
2888 boolean_t do_arc_evict
= B_FALSE
;
2890 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
2892 if (!DBUF_IS_CACHEABLE(db
) &&
2893 db
->db_blkptr
!= NULL
&&
2894 !BP_IS_HOLE(db
->db_blkptr
) &&
2895 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2896 do_arc_evict
= B_TRUE
;
2897 bp
= *db
->db_blkptr
;
2900 if (!DBUF_IS_CACHEABLE(db
) ||
2901 db
->db_pending_evict
) {
2903 } else if (!multilist_link_active(&db
->db_cache_link
)) {
2904 ASSERT3U(db
->db_caching_status
, ==,
2907 dbuf_cached_state_t dcs
=
2908 dbuf_include_in_metadata_cache(db
) ?
2909 DB_DBUF_METADATA_CACHE
: DB_DBUF_CACHE
;
2910 db
->db_caching_status
= dcs
;
2912 multilist_insert(dbuf_caches
[dcs
].cache
, db
);
2913 (void) refcount_add_many(&dbuf_caches
[dcs
].size
,
2914 db
->db
.db_size
, db
);
2915 mutex_exit(&db
->db_mtx
);
2917 if (db
->db_caching_status
== DB_DBUF_CACHE
&&
2919 dbuf_evict_notify();
2924 arc_freed(spa
, &bp
);
2927 mutex_exit(&db
->db_mtx
);
2932 #pragma weak dmu_buf_refcount = dbuf_refcount
2934 dbuf_refcount(dmu_buf_impl_t
*db
)
2936 return (refcount_count(&db
->db_holds
));
2940 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
2941 dmu_buf_user_t
*new_user
)
2943 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2945 mutex_enter(&db
->db_mtx
);
2946 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2947 if (db
->db_user
== old_user
)
2948 db
->db_user
= new_user
;
2950 old_user
= db
->db_user
;
2951 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2952 mutex_exit(&db
->db_mtx
);
2958 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2960 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
2964 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2966 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2968 db
->db_user_immediate_evict
= TRUE
;
2969 return (dmu_buf_set_user(db_fake
, user
));
2973 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2975 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
2979 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2981 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2983 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2984 return (db
->db_user
);
2988 dmu_buf_user_evict_wait()
2990 taskq_wait(dbu_evict_taskq
);
2994 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2996 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2997 return (dbi
->db_blkptr
);
3001 dmu_buf_get_objset(dmu_buf_t
*db
)
3003 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3004 return (dbi
->db_objset
);
3008 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3010 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3011 DB_DNODE_ENTER(dbi
);
3012 return (DB_DNODE(dbi
));
3016 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3018 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3023 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3025 /* ASSERT(dmu_tx_is_syncing(tx) */
3026 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3028 if (db
->db_blkptr
!= NULL
)
3031 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3032 db
->db_blkptr
= &dn
->dn_phys
->dn_spill
;
3033 BP_ZERO(db
->db_blkptr
);
3036 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3038 * This buffer was allocated at a time when there was
3039 * no available blkptrs from the dnode, or it was
3040 * inappropriate to hook it in (i.e., nlevels mis-match).
3042 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3043 ASSERT(db
->db_parent
== NULL
);
3044 db
->db_parent
= dn
->dn_dbuf
;
3045 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3048 dmu_buf_impl_t
*parent
= db
->db_parent
;
3049 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3051 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3052 if (parent
== NULL
) {
3053 mutex_exit(&db
->db_mtx
);
3054 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3055 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3056 db
->db_blkid
>> epbs
, db
);
3057 rw_exit(&dn
->dn_struct_rwlock
);
3058 mutex_enter(&db
->db_mtx
);
3059 db
->db_parent
= parent
;
3061 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3062 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3068 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3070 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3074 ASSERT(dmu_tx_is_syncing(tx
));
3076 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3078 mutex_enter(&db
->db_mtx
);
3080 ASSERT(db
->db_level
> 0);
3083 /* Read the block if it hasn't been read yet. */
3084 if (db
->db_buf
== NULL
) {
3085 mutex_exit(&db
->db_mtx
);
3086 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3087 mutex_enter(&db
->db_mtx
);
3089 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3090 ASSERT(db
->db_buf
!= NULL
);
3094 /* Indirect block size must match what the dnode thinks it is. */
3095 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3096 dbuf_check_blkptr(dn
, db
);
3099 /* Provide the pending dirty record to child dbufs */
3100 db
->db_data_pending
= dr
;
3102 mutex_exit(&db
->db_mtx
);
3104 dbuf_write(dr
, db
->db_buf
, tx
);
3107 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3108 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3109 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3110 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3115 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3117 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3118 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3121 uint64_t txg
= tx
->tx_txg
;
3123 ASSERT(dmu_tx_is_syncing(tx
));
3125 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3127 mutex_enter(&db
->db_mtx
);
3129 * To be synced, we must be dirtied. But we
3130 * might have been freed after the dirty.
3132 if (db
->db_state
== DB_UNCACHED
) {
3133 /* This buffer has been freed since it was dirtied */
3134 ASSERT(db
->db
.db_data
== NULL
);
3135 } else if (db
->db_state
== DB_FILL
) {
3136 /* This buffer was freed and is now being re-filled */
3137 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3139 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3146 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3147 mutex_enter(&dn
->dn_mtx
);
3148 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3149 mutex_exit(&dn
->dn_mtx
);
3153 * If this is a bonus buffer, simply copy the bonus data into the
3154 * dnode. It will be written out when the dnode is synced (and it
3155 * will be synced, since it must have been dirty for dbuf_sync to
3158 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3159 dbuf_dirty_record_t
**drp
;
3161 ASSERT(*datap
!= NULL
);
3162 ASSERT0(db
->db_level
);
3163 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
3164 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
3167 if (*datap
!= db
->db
.db_data
) {
3168 zio_buf_free(*datap
, DN_MAX_BONUSLEN
);
3169 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
3171 db
->db_data_pending
= NULL
;
3172 drp
= &db
->db_last_dirty
;
3174 drp
= &(*drp
)->dr_next
;
3175 ASSERT(dr
->dr_next
== NULL
);
3176 ASSERT(dr
->dr_dbuf
== db
);
3178 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3179 ASSERT(db
->db_dirtycnt
> 0);
3180 db
->db_dirtycnt
-= 1;
3181 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
, B_FALSE
);
3188 * This function may have dropped the db_mtx lock allowing a dmu_sync
3189 * operation to sneak in. As a result, we need to ensure that we
3190 * don't check the dr_override_state until we have returned from
3191 * dbuf_check_blkptr.
3193 dbuf_check_blkptr(dn
, db
);
3196 * If this buffer is in the middle of an immediate write,
3197 * wait for the synchronous IO to complete.
3199 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3200 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3201 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3202 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3205 if (db
->db_state
!= DB_NOFILL
&&
3206 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3207 refcount_count(&db
->db_holds
) > 1 &&
3208 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3209 *datap
== db
->db_buf
) {
3211 * If this buffer is currently "in use" (i.e., there
3212 * are active holds and db_data still references it),
3213 * then make a copy before we start the write so that
3214 * any modifications from the open txg will not leak
3217 * NOTE: this copy does not need to be made for
3218 * objects only modified in the syncing context (e.g.
3219 * DNONE_DNODE blocks).
3221 int psize
= arc_buf_size(*datap
);
3222 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3223 enum zio_compress compress_type
= arc_get_compression(*datap
);
3225 if (compress_type
== ZIO_COMPRESS_OFF
) {
3226 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
3228 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
3229 int lsize
= arc_buf_lsize(*datap
);
3230 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
3231 psize
, lsize
, compress_type
);
3233 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
3235 db
->db_data_pending
= dr
;
3237 mutex_exit(&db
->db_mtx
);
3239 dbuf_write(dr
, *datap
, tx
);
3241 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3242 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3243 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3247 * Although zio_nowait() does not "wait for an IO", it does
3248 * initiate the IO. If this is an empty write it seems plausible
3249 * that the IO could actually be completed before the nowait
3250 * returns. We need to DB_DNODE_EXIT() first in case
3251 * zio_nowait() invalidates the dbuf.
3254 zio_nowait(dr
->dr_zio
);
3259 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3261 dbuf_dirty_record_t
*dr
;
3263 while (dr
= list_head(list
)) {
3264 if (dr
->dr_zio
!= NULL
) {
3266 * If we find an already initialized zio then we
3267 * are processing the meta-dnode, and we have finished.
3268 * The dbufs for all dnodes are put back on the list
3269 * during processing, so that we can zio_wait()
3270 * these IOs after initiating all child IOs.
3272 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3273 DMU_META_DNODE_OBJECT
);
3276 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3277 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3278 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3280 list_remove(list
, dr
);
3281 if (dr
->dr_dbuf
->db_level
> 0)
3282 dbuf_sync_indirect(dr
, tx
);
3284 dbuf_sync_leaf(dr
, tx
);
3290 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3292 dmu_buf_impl_t
*db
= vdb
;
3294 blkptr_t
*bp
= zio
->io_bp
;
3295 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3296 spa_t
*spa
= zio
->io_spa
;
3301 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3302 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3306 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3307 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3308 zio
->io_prev_space_delta
= delta
;
3310 if (bp
->blk_birth
!= 0) {
3311 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3312 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3313 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3314 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3315 BP_IS_EMBEDDED(bp
));
3316 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3319 mutex_enter(&db
->db_mtx
);
3322 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3323 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3324 ASSERT(!(BP_IS_HOLE(bp
)) &&
3325 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3329 if (db
->db_level
== 0) {
3330 mutex_enter(&dn
->dn_mtx
);
3331 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3332 db
->db_blkid
!= DMU_SPILL_BLKID
)
3333 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3334 mutex_exit(&dn
->dn_mtx
);
3336 if (dn
->dn_type
== DMU_OT_DNODE
) {
3337 dnode_phys_t
*dnp
= db
->db
.db_data
;
3338 for (i
= db
->db
.db_size
>> DNODE_SHIFT
; i
> 0;
3340 if (dnp
->dn_type
!= DMU_OT_NONE
)
3344 if (BP_IS_HOLE(bp
)) {
3351 blkptr_t
*ibp
= db
->db
.db_data
;
3352 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3353 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3354 if (BP_IS_HOLE(ibp
))
3356 fill
+= BP_GET_FILL(ibp
);
3361 if (!BP_IS_EMBEDDED(bp
))
3362 bp
->blk_fill
= fill
;
3364 mutex_exit(&db
->db_mtx
);
3366 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3367 *db
->db_blkptr
= *bp
;
3368 rw_exit(&dn
->dn_struct_rwlock
);
3373 * This function gets called just prior to running through the compression
3374 * stage of the zio pipeline. If we're an indirect block comprised of only
3375 * holes, then we want this indirect to be compressed away to a hole. In
3376 * order to do that we must zero out any information about the holes that
3377 * this indirect points to prior to before we try to compress it.
3380 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3382 dmu_buf_impl_t
*db
= vdb
;
3385 unsigned int epbs
, i
;
3387 ASSERT3U(db
->db_level
, >, 0);
3390 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3391 ASSERT3U(epbs
, <, 31);
3393 /* Determine if all our children are holes */
3394 for (i
= 0, bp
= db
->db
.db_data
; i
< 1 << epbs
; i
++, bp
++) {
3395 if (!BP_IS_HOLE(bp
))
3400 * If all the children are holes, then zero them all out so that
3401 * we may get compressed away.
3403 if (i
== 1 << epbs
) {
3405 * We only found holes. Grab the rwlock to prevent
3406 * anybody from reading the blocks we're about to
3409 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3410 bzero(db
->db
.db_data
, db
->db
.db_size
);
3411 rw_exit(&dn
->dn_struct_rwlock
);
3417 * The SPA will call this callback several times for each zio - once
3418 * for every physical child i/o (zio->io_phys_children times). This
3419 * allows the DMU to monitor the progress of each logical i/o. For example,
3420 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3421 * block. There may be a long delay before all copies/fragments are completed,
3422 * so this callback allows us to retire dirty space gradually, as the physical
3427 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3429 dmu_buf_impl_t
*db
= arg
;
3430 objset_t
*os
= db
->db_objset
;
3431 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3432 dbuf_dirty_record_t
*dr
;
3435 dr
= db
->db_data_pending
;
3436 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3439 * The callback will be called io_phys_children times. Retire one
3440 * portion of our dirty space each time we are called. Any rounding
3441 * error will be cleaned up by dsl_pool_sync()'s call to
3442 * dsl_pool_undirty_space().
3444 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3445 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3450 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3452 dmu_buf_impl_t
*db
= vdb
;
3453 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3454 blkptr_t
*bp
= db
->db_blkptr
;
3455 objset_t
*os
= db
->db_objset
;
3456 dmu_tx_t
*tx
= os
->os_synctx
;
3457 dbuf_dirty_record_t
**drp
, *dr
;
3459 ASSERT0(zio
->io_error
);
3460 ASSERT(db
->db_blkptr
== bp
);
3463 * For nopwrites and rewrites we ensure that the bp matches our
3464 * original and bypass all the accounting.
3466 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3467 ASSERT(BP_EQUAL(bp
, bp_orig
));
3469 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3470 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3471 dsl_dataset_block_born(ds
, bp
, tx
);
3474 mutex_enter(&db
->db_mtx
);
3478 drp
= &db
->db_last_dirty
;
3479 while ((dr
= *drp
) != db
->db_data_pending
)
3481 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3482 ASSERT(dr
->dr_dbuf
== db
);
3483 ASSERT(dr
->dr_next
== NULL
);
3487 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3492 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3493 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3494 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3499 if (db
->db_level
== 0) {
3500 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3501 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3502 if (db
->db_state
!= DB_NOFILL
) {
3503 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3504 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
3511 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3512 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3513 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3515 dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3516 ASSERT3U(db
->db_blkid
, <=,
3517 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3518 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3522 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3523 list_destroy(&dr
->dt
.di
.dr_children
);
3525 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3527 cv_broadcast(&db
->db_changed
);
3528 ASSERT(db
->db_dirtycnt
> 0);
3529 db
->db_dirtycnt
-= 1;
3530 db
->db_data_pending
= NULL
;
3531 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
, B_FALSE
);
3535 dbuf_write_nofill_ready(zio_t
*zio
)
3537 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3541 dbuf_write_nofill_done(zio_t
*zio
)
3543 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3547 dbuf_write_override_ready(zio_t
*zio
)
3549 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3550 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3552 dbuf_write_ready(zio
, NULL
, db
);
3556 dbuf_write_override_done(zio_t
*zio
)
3558 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3559 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3560 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3562 mutex_enter(&db
->db_mtx
);
3563 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3564 if (!BP_IS_HOLE(obp
))
3565 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3566 arc_release(dr
->dt
.dl
.dr_data
, db
);
3568 mutex_exit(&db
->db_mtx
);
3569 dbuf_write_done(zio
, NULL
, db
);
3571 if (zio
->io_abd
!= NULL
)
3572 abd_put(zio
->io_abd
);
3575 typedef struct dbuf_remap_impl_callback_arg
{
3577 uint64_t drica_blk_birth
;
3579 } dbuf_remap_impl_callback_arg_t
;
3582 dbuf_remap_impl_callback(uint64_t vdev
, uint64_t offset
, uint64_t size
,
3585 dbuf_remap_impl_callback_arg_t
*drica
= arg
;
3586 objset_t
*os
= drica
->drica_os
;
3587 spa_t
*spa
= dmu_objset_spa(os
);
3588 dmu_tx_t
*tx
= drica
->drica_tx
;
3590 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
3592 if (os
== spa_meta_objset(spa
)) {
3593 spa_vdev_indirect_mark_obsolete(spa
, vdev
, offset
, size
, tx
);
3595 dsl_dataset_block_remapped(dmu_objset_ds(os
), vdev
, offset
,
3596 size
, drica
->drica_blk_birth
, tx
);
3601 dbuf_remap_impl(dnode_t
*dn
, blkptr_t
*bp
, dmu_tx_t
*tx
)
3603 blkptr_t bp_copy
= *bp
;
3604 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
3605 dbuf_remap_impl_callback_arg_t drica
;
3607 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
3609 drica
.drica_os
= dn
->dn_objset
;
3610 drica
.drica_blk_birth
= bp
->blk_birth
;
3611 drica
.drica_tx
= tx
;
3612 if (spa_remap_blkptr(spa
, &bp_copy
, dbuf_remap_impl_callback
,
3615 * The struct_rwlock prevents dbuf_read_impl() from
3616 * dereferencing the BP while we are changing it. To
3617 * avoid lock contention, only grab it when we are actually
3620 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3622 rw_exit(&dn
->dn_struct_rwlock
);
3627 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
3628 * to remap a copy of every bp in the dbuf.
3631 dbuf_can_remap(const dmu_buf_impl_t
*db
)
3633 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3634 blkptr_t
*bp
= db
->db
.db_data
;
3635 boolean_t ret
= B_FALSE
;
3637 ASSERT3U(db
->db_level
, >, 0);
3638 ASSERT3S(db
->db_state
, ==, DB_CACHED
);
3640 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
));
3642 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
3643 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
3644 blkptr_t bp_copy
= bp
[i
];
3645 if (spa_remap_blkptr(spa
, &bp_copy
, NULL
, NULL
)) {
3650 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
3656 dnode_needs_remap(const dnode_t
*dn
)
3658 spa_t
*spa
= dmu_objset_spa(dn
->dn_objset
);
3659 boolean_t ret
= B_FALSE
;
3661 if (dn
->dn_phys
->dn_nlevels
== 0) {
3665 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
));
3667 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
3668 for (int j
= 0; j
< dn
->dn_phys
->dn_nblkptr
; j
++) {
3669 blkptr_t bp_copy
= dn
->dn_phys
->dn_blkptr
[j
];
3670 if (spa_remap_blkptr(spa
, &bp_copy
, NULL
, NULL
)) {
3675 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
3681 * Remap any existing BP's to concrete vdevs, if possible.
3684 dbuf_remap(dnode_t
*dn
, dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
3686 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3687 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa
)));
3689 if (!spa_feature_is_active(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
3692 if (db
->db_level
> 0) {
3693 blkptr_t
*bp
= db
->db
.db_data
;
3694 for (int i
= 0; i
< db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
++) {
3695 dbuf_remap_impl(dn
, &bp
[i
], tx
);
3697 } else if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
3698 dnode_phys_t
*dnp
= db
->db
.db_data
;
3699 ASSERT3U(db
->db_dnode_handle
->dnh_dnode
->dn_type
, ==,
3701 for (int i
= 0; i
< db
->db
.db_size
>> DNODE_SHIFT
; i
++) {
3702 for (int j
= 0; j
< dnp
[i
].dn_nblkptr
; j
++) {
3703 dbuf_remap_impl(dn
, &dnp
[i
].dn_blkptr
[j
], tx
);
3710 /* Issue I/O to commit a dirty buffer to disk. */
3712 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3714 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3717 dmu_buf_impl_t
*parent
= db
->db_parent
;
3718 uint64_t txg
= tx
->tx_txg
;
3719 zbookmark_phys_t zb
;
3724 ASSERT(dmu_tx_is_syncing(tx
));
3730 if (db
->db_state
!= DB_NOFILL
) {
3731 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3733 * Private object buffers are released here rather
3734 * than in dbuf_dirty() since they are only modified
3735 * in the syncing context and we don't want the
3736 * overhead of making multiple copies of the data.
3738 if (BP_IS_HOLE(db
->db_blkptr
)) {
3741 dbuf_release_bp(db
);
3743 dbuf_remap(dn
, db
, tx
);
3747 if (parent
!= dn
->dn_dbuf
) {
3748 /* Our parent is an indirect block. */
3749 /* We have a dirty parent that has been scheduled for write. */
3750 ASSERT(parent
&& parent
->db_data_pending
);
3751 /* Our parent's buffer is one level closer to the dnode. */
3752 ASSERT(db
->db_level
== parent
->db_level
-1);
3754 * We're about to modify our parent's db_data by modifying
3755 * our block pointer, so the parent must be released.
3757 ASSERT(arc_released(parent
->db_buf
));
3758 zio
= parent
->db_data_pending
->dr_zio
;
3760 /* Our parent is the dnode itself. */
3761 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3762 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3763 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3764 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3765 ASSERT3P(db
->db_blkptr
, ==,
3766 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3770 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3771 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3774 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3775 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3776 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3778 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3780 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3782 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3786 * We copy the blkptr now (rather than when we instantiate the dirty
3787 * record), because its value can change between open context and
3788 * syncing context. We do not need to hold dn_struct_rwlock to read
3789 * db_blkptr because we are in syncing context.
3791 dr
->dr_bp_copy
= *db
->db_blkptr
;
3793 if (db
->db_level
== 0 &&
3794 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3796 * The BP for this block has been provided by open context
3797 * (by dmu_sync() or dmu_buf_write_embedded()).
3799 abd_t
*contents
= (data
!= NULL
) ?
3800 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
3802 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
, &dr
->dr_bp_copy
,
3803 contents
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
3804 dbuf_write_override_ready
, NULL
, NULL
,
3805 dbuf_write_override_done
,
3806 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3807 mutex_enter(&db
->db_mtx
);
3808 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3809 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3810 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3811 mutex_exit(&db
->db_mtx
);
3812 } else if (db
->db_state
== DB_NOFILL
) {
3813 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
3814 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
3815 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3816 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
3817 dbuf_write_nofill_ready
, NULL
, NULL
,
3818 dbuf_write_nofill_done
, db
,
3819 ZIO_PRIORITY_ASYNC_WRITE
,
3820 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3822 ASSERT(arc_released(data
));
3825 * For indirect blocks, we want to setup the children
3826 * ready callback so that we can properly handle an indirect
3827 * block that only contains holes.
3829 arc_done_func_t
*children_ready_cb
= NULL
;
3830 if (db
->db_level
!= 0)
3831 children_ready_cb
= dbuf_write_children_ready
;
3833 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3834 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3835 &zp
, dbuf_write_ready
, children_ready_cb
,
3836 dbuf_write_physdone
, dbuf_write_done
, db
,
3837 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);