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, 2016 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>
50 uint_t zfs_dbuf_evict_key
;
53 * Number of times that zfs_free_range() took the slow path while doing
54 * a zfs receive. A nonzero value indicates a potential performance problem.
56 uint64_t zfs_free_range_recv_miss
;
58 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
59 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
62 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
63 dmu_buf_evict_func_t
*evict_func
, dmu_buf_t
**clear_on_evict_dbufp
);
67 * Global data structures and functions for the dbuf cache.
69 static kmem_cache_t
*dbuf_kmem_cache
;
70 static taskq_t
*dbu_evict_taskq
;
72 static kthread_t
*dbuf_cache_evict_thread
;
73 static kmutex_t dbuf_evict_lock
;
74 static kcondvar_t dbuf_evict_cv
;
75 static boolean_t dbuf_evict_thread_exit
;
78 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
79 * are not currently held but have been recently released. These dbufs
80 * are not eligible for arc eviction until they are aged out of the cache.
81 * Dbufs are added to the dbuf cache once the last hold is released. If a
82 * dbuf is later accessed and still exists in the dbuf cache, then it will
83 * be removed from the cache and later re-added to the head of the cache.
84 * Dbufs that are aged out of the cache will be immediately destroyed and
85 * become eligible for arc eviction.
87 static multilist_t dbuf_cache
;
88 static refcount_t dbuf_cache_size
;
89 uint64_t dbuf_cache_max_bytes
= 100 * 1024 * 1024;
91 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
92 int dbuf_cache_max_shift
= 5;
95 * The dbuf cache uses a three-stage eviction policy:
96 * - A low water marker designates when the dbuf eviction thread
97 * should stop evicting from the dbuf cache.
98 * - When we reach the maximum size (aka mid water mark), we
99 * signal the eviction thread to run.
100 * - The high water mark indicates when the eviction thread
101 * is unable to keep up with the incoming load and eviction must
102 * happen in the context of the calling thread.
106 * low water mid water hi water
107 * +----------------------------------------+----------+----------+
112 * +----------------------------------------+----------+----------+
114 * evicting eviction directly
117 * The high and low water marks indicate the operating range for the eviction
118 * thread. The low water mark is, by default, 90% of the total size of the
119 * cache and the high water mark is at 110% (both of these percentages can be
120 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
121 * respectively). The eviction thread will try to ensure that the cache remains
122 * within this range by waking up every second and checking if the cache is
123 * above the low water mark. The thread can also be woken up by callers adding
124 * elements into the cache if the cache is larger than the mid water (i.e max
125 * cache size). Once the eviction thread is woken up and eviction is required,
126 * it will continue evicting buffers until it's able to reduce the cache size
127 * to the low water mark. If the cache size continues to grow and hits the high
128 * water mark, then callers adding elments to the cache will begin to evict
129 * directly from the cache until the cache is no longer above the high water
134 * The percentage above and below the maximum cache size.
136 uint_t dbuf_cache_hiwater_pct
= 10;
137 uint_t dbuf_cache_lowater_pct
= 10;
141 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
143 dmu_buf_impl_t
*db
= vdb
;
144 bzero(db
, sizeof (dmu_buf_impl_t
));
146 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
147 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
148 multilist_link_init(&db
->db_cache_link
);
149 refcount_create(&db
->db_holds
);
156 dbuf_dest(void *vdb
, void *unused
)
158 dmu_buf_impl_t
*db
= vdb
;
159 mutex_destroy(&db
->db_mtx
);
160 cv_destroy(&db
->db_changed
);
161 ASSERT(!multilist_link_active(&db
->db_cache_link
));
162 refcount_destroy(&db
->db_holds
);
166 * dbuf hash table routines
168 static dbuf_hash_table_t dbuf_hash_table
;
170 static uint64_t dbuf_hash_count
;
173 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
175 uintptr_t osv
= (uintptr_t)os
;
176 uint64_t crc
= -1ULL;
178 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
179 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
180 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
181 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
182 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
183 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
184 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
186 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
191 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
192 ((dbuf)->db.db_object == (obj) && \
193 (dbuf)->db_objset == (os) && \
194 (dbuf)->db_level == (level) && \
195 (dbuf)->db_blkid == (blkid))
198 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
200 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
201 uint64_t hv
= dbuf_hash(os
, obj
, level
, blkid
);
202 uint64_t idx
= hv
& h
->hash_table_mask
;
205 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
206 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
207 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
208 mutex_enter(&db
->db_mtx
);
209 if (db
->db_state
!= DB_EVICTING
) {
210 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
213 mutex_exit(&db
->db_mtx
);
216 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
220 static dmu_buf_impl_t
*
221 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
224 dmu_buf_impl_t
*db
= NULL
;
226 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
227 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
228 if (dn
->dn_bonus
!= NULL
) {
230 mutex_enter(&db
->db_mtx
);
232 rw_exit(&dn
->dn_struct_rwlock
);
233 dnode_rele(dn
, FTAG
);
239 * Insert an entry into the hash table. If there is already an element
240 * equal to elem in the hash table, then the already existing element
241 * will be returned and the new element will not be inserted.
242 * Otherwise returns NULL.
244 static dmu_buf_impl_t
*
245 dbuf_hash_insert(dmu_buf_impl_t
*db
)
247 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
248 objset_t
*os
= db
->db_objset
;
249 uint64_t obj
= db
->db
.db_object
;
250 int level
= db
->db_level
;
251 uint64_t blkid
= db
->db_blkid
;
252 uint64_t hv
= dbuf_hash(os
, obj
, level
, blkid
);
253 uint64_t idx
= hv
& h
->hash_table_mask
;
256 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
257 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
258 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
259 mutex_enter(&dbf
->db_mtx
);
260 if (dbf
->db_state
!= DB_EVICTING
) {
261 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
264 mutex_exit(&dbf
->db_mtx
);
268 mutex_enter(&db
->db_mtx
);
269 db
->db_hash_next
= h
->hash_table
[idx
];
270 h
->hash_table
[idx
] = db
;
271 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
272 atomic_inc_64(&dbuf_hash_count
);
278 * Remove an entry from the hash table. It must be in the EVICTING state.
281 dbuf_hash_remove(dmu_buf_impl_t
*db
)
283 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
284 uint64_t hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
285 db
->db_level
, db
->db_blkid
);
286 uint64_t idx
= hv
& h
->hash_table_mask
;
287 dmu_buf_impl_t
*dbf
, **dbp
;
290 * We musn't hold db_mtx to maintain lock ordering:
291 * DBUF_HASH_MUTEX > db_mtx.
293 ASSERT(refcount_is_zero(&db
->db_holds
));
294 ASSERT(db
->db_state
== DB_EVICTING
);
295 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
297 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
298 dbp
= &h
->hash_table
[idx
];
299 while ((dbf
= *dbp
) != db
) {
300 dbp
= &dbf
->db_hash_next
;
303 *dbp
= db
->db_hash_next
;
304 db
->db_hash_next
= NULL
;
305 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
306 atomic_dec_64(&dbuf_hash_count
);
312 } dbvu_verify_type_t
;
315 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
320 if (db
->db_user
== NULL
)
323 /* Only data blocks support the attachment of user data. */
324 ASSERT(db
->db_level
== 0);
326 /* Clients must resolve a dbuf before attaching user data. */
327 ASSERT(db
->db
.db_data
!= NULL
);
328 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
330 holds
= refcount_count(&db
->db_holds
);
331 if (verify_type
== DBVU_EVICTING
) {
333 * Immediate eviction occurs when holds == dirtycnt.
334 * For normal eviction buffers, holds is zero on
335 * eviction, except when dbuf_fix_old_data() calls
336 * dbuf_clear_data(). However, the hold count can grow
337 * during eviction even though db_mtx is held (see
338 * dmu_bonus_hold() for an example), so we can only
339 * test the generic invariant that holds >= dirtycnt.
341 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
343 if (db
->db_user_immediate_evict
== TRUE
)
344 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
346 ASSERT3U(holds
, >, 0);
352 dbuf_evict_user(dmu_buf_impl_t
*db
)
354 dmu_buf_user_t
*dbu
= db
->db_user
;
356 ASSERT(MUTEX_HELD(&db
->db_mtx
));
361 dbuf_verify_user(db
, DBVU_EVICTING
);
365 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
366 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
370 * Invoke the callback from a taskq to avoid lock order reversals
371 * and limit stack depth.
373 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func
, dbu
, 0,
378 dbuf_is_metadata(dmu_buf_impl_t
*db
)
380 if (db
->db_level
> 0) {
383 boolean_t is_metadata
;
386 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
389 return (is_metadata
);
394 * This function *must* return indices evenly distributed between all
395 * sublists of the multilist. This is needed due to how the dbuf eviction
396 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
397 * distributed between all sublists and uses this assumption when
398 * deciding which sublist to evict from and how much to evict from it.
401 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
403 dmu_buf_impl_t
*db
= obj
;
406 * The assumption here, is the hash value for a given
407 * dmu_buf_impl_t will remain constant throughout it's lifetime
408 * (i.e. it's objset, object, level and blkid fields don't change).
409 * Thus, we don't need to store the dbuf's sublist index
410 * on insertion, as this index can be recalculated on removal.
412 * Also, the low order bits of the hash value are thought to be
413 * distributed evenly. Otherwise, in the case that the multilist
414 * has a power of two number of sublists, each sublists' usage
415 * would not be evenly distributed.
417 return (dbuf_hash(db
->db_objset
, db
->db
.db_object
,
418 db
->db_level
, db
->db_blkid
) %
419 multilist_get_num_sublists(ml
));
422 static inline boolean_t
423 dbuf_cache_above_hiwater(void)
425 uint64_t dbuf_cache_hiwater_bytes
=
426 (dbuf_cache_max_bytes
* dbuf_cache_hiwater_pct
) / 100;
428 return (refcount_count(&dbuf_cache_size
) >
429 dbuf_cache_max_bytes
+ dbuf_cache_hiwater_bytes
);
432 static inline boolean_t
433 dbuf_cache_above_lowater(void)
435 uint64_t dbuf_cache_lowater_bytes
=
436 (dbuf_cache_max_bytes
* dbuf_cache_lowater_pct
) / 100;
438 return (refcount_count(&dbuf_cache_size
) >
439 dbuf_cache_max_bytes
- dbuf_cache_lowater_bytes
);
443 * Evict the oldest eligible dbuf from the dbuf cache.
448 int idx
= multilist_get_random_index(&dbuf_cache
);
449 multilist_sublist_t
*mls
= multilist_sublist_lock(&dbuf_cache
, idx
);
451 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
454 * Set the thread's tsd to indicate that it's processing evictions.
455 * Once a thread stops evicting from the dbuf cache it will
456 * reset its tsd to NULL.
458 ASSERT3P(tsd_get(zfs_dbuf_evict_key
), ==, NULL
);
459 (void) tsd_set(zfs_dbuf_evict_key
, (void *)B_TRUE
);
461 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
462 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
463 db
= multilist_sublist_prev(mls
, db
);
466 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
467 multilist_sublist_t
*, mls
);
470 multilist_sublist_remove(mls
, db
);
471 multilist_sublist_unlock(mls
);
472 (void) refcount_remove_many(&dbuf_cache_size
,
476 multilist_sublist_unlock(mls
);
478 (void) tsd_set(zfs_dbuf_evict_key
, NULL
);
482 * The dbuf evict thread is responsible for aging out dbufs from the
483 * cache. Once the cache has reached it's maximum size, dbufs are removed
484 * and destroyed. The eviction thread will continue running until the size
485 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
486 * out of the cache it is destroyed and becomes eligible for arc eviction.
489 dbuf_evict_thread(void)
493 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
495 mutex_enter(&dbuf_evict_lock
);
496 while (!dbuf_evict_thread_exit
) {
497 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
498 CALLB_CPR_SAFE_BEGIN(&cpr
);
499 (void) cv_timedwait_hires(&dbuf_evict_cv
,
500 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
501 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
503 mutex_exit(&dbuf_evict_lock
);
506 * Keep evicting as long as we're above the low water mark
507 * for the cache. We do this without holding the locks to
508 * minimize lock contention.
510 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
514 mutex_enter(&dbuf_evict_lock
);
517 dbuf_evict_thread_exit
= B_FALSE
;
518 cv_broadcast(&dbuf_evict_cv
);
519 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
524 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
525 * If the dbuf cache is at its high water mark, then evict a dbuf from the
526 * dbuf cache using the callers context.
529 dbuf_evict_notify(void)
533 * We use thread specific data to track when a thread has
534 * started processing evictions. This allows us to avoid deeply
535 * nested stacks that would have a call flow similar to this:
537 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
540 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
542 * The dbuf_eviction_thread will always have its tsd set until
543 * that thread exits. All other threads will only set their tsd
544 * if they are participating in the eviction process. This only
545 * happens if the eviction thread is unable to process evictions
546 * fast enough. To keep the dbuf cache size in check, other threads
547 * can evict from the dbuf cache directly. Those threads will set
548 * their tsd values so that we ensure that they only evict one dbuf
549 * from the dbuf cache.
551 if (tsd_get(zfs_dbuf_evict_key
) != NULL
)
554 if (refcount_count(&dbuf_cache_size
) > dbuf_cache_max_bytes
) {
555 boolean_t evict_now
= B_FALSE
;
557 mutex_enter(&dbuf_evict_lock
);
558 if (refcount_count(&dbuf_cache_size
) > dbuf_cache_max_bytes
) {
559 evict_now
= dbuf_cache_above_hiwater();
560 cv_signal(&dbuf_evict_cv
);
562 mutex_exit(&dbuf_evict_lock
);
573 uint64_t hsize
= 1ULL << 16;
574 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
578 * The hash table is big enough to fill all of physical memory
579 * with an average 4K block size. The table will take up
580 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
582 while (hsize
* 4096 < physmem
* PAGESIZE
)
586 h
->hash_table_mask
= hsize
- 1;
587 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
588 if (h
->hash_table
== NULL
) {
589 /* XXX - we should really return an error instead of assert */
590 ASSERT(hsize
> (1ULL << 10));
595 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
596 sizeof (dmu_buf_impl_t
),
597 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
599 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
600 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
603 * Setup the parameters for the dbuf cache. We cap the size of the
604 * dbuf cache to 1/32nd (default) of the size of the ARC.
606 dbuf_cache_max_bytes
= MIN(dbuf_cache_max_bytes
,
607 arc_max_bytes() >> dbuf_cache_max_shift
);
610 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
611 * configuration is not required.
613 dbu_evict_taskq
= taskq_create("dbu_evict", 1, minclsyspri
, 0, 0, 0);
615 multilist_create(&dbuf_cache
, sizeof (dmu_buf_impl_t
),
616 offsetof(dmu_buf_impl_t
, db_cache_link
),
617 zfs_arc_num_sublists_per_state
,
618 dbuf_cache_multilist_index_func
);
619 refcount_create(&dbuf_cache_size
);
621 tsd_create(&zfs_dbuf_evict_key
, NULL
);
622 dbuf_evict_thread_exit
= B_FALSE
;
623 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
624 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
625 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
626 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
632 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
635 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
636 mutex_destroy(&h
->hash_mutexes
[i
]);
637 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
638 kmem_cache_destroy(dbuf_kmem_cache
);
639 taskq_destroy(dbu_evict_taskq
);
641 mutex_enter(&dbuf_evict_lock
);
642 dbuf_evict_thread_exit
= B_TRUE
;
643 while (dbuf_evict_thread_exit
) {
644 cv_signal(&dbuf_evict_cv
);
645 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
647 mutex_exit(&dbuf_evict_lock
);
648 tsd_destroy(&zfs_dbuf_evict_key
);
650 mutex_destroy(&dbuf_evict_lock
);
651 cv_destroy(&dbuf_evict_cv
);
653 refcount_destroy(&dbuf_cache_size
);
654 multilist_destroy(&dbuf_cache
);
663 dbuf_verify(dmu_buf_impl_t
*db
)
666 dbuf_dirty_record_t
*dr
;
668 ASSERT(MUTEX_HELD(&db
->db_mtx
));
670 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
673 ASSERT(db
->db_objset
!= NULL
);
677 ASSERT(db
->db_parent
== NULL
);
678 ASSERT(db
->db_blkptr
== NULL
);
680 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
681 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
682 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
683 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
684 db
->db_blkid
== DMU_SPILL_BLKID
||
685 !avl_is_empty(&dn
->dn_dbufs
));
687 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
689 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
690 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
691 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
693 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
694 ASSERT0(db
->db
.db_offset
);
696 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
699 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
700 ASSERT(dr
->dr_dbuf
== db
);
702 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
703 ASSERT(dr
->dr_dbuf
== db
);
706 * We can't assert that db_size matches dn_datablksz because it
707 * can be momentarily different when another thread is doing
710 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
711 dr
= db
->db_data_pending
;
713 * It should only be modified in syncing context, so
714 * make sure we only have one copy of the data.
716 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
719 /* verify db->db_blkptr */
721 if (db
->db_parent
== dn
->dn_dbuf
) {
722 /* db is pointed to by the dnode */
723 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
724 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
725 ASSERT(db
->db_parent
== NULL
);
727 ASSERT(db
->db_parent
!= NULL
);
728 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
729 ASSERT3P(db
->db_blkptr
, ==,
730 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
732 /* db is pointed to by an indirect block */
733 int epb
= db
->db_parent
->db
.db_size
>> SPA_BLKPTRSHIFT
;
734 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
735 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
738 * dnode_grow_indblksz() can make this fail if we don't
739 * have the struct_rwlock. XXX indblksz no longer
740 * grows. safe to do this now?
742 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
743 ASSERT3P(db
->db_blkptr
, ==,
744 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
745 db
->db_blkid
% epb
));
749 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
750 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
751 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
752 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
754 * If the blkptr isn't set but they have nonzero data,
755 * it had better be dirty, otherwise we'll lose that
756 * data when we evict this buffer.
758 * There is an exception to this rule for indirect blocks; in
759 * this case, if the indirect block is a hole, we fill in a few
760 * fields on each of the child blocks (importantly, birth time)
761 * to prevent hole birth times from being lost when you
762 * partially fill in a hole.
764 if (db
->db_dirtycnt
== 0) {
765 if (db
->db_level
== 0) {
766 uint64_t *buf
= db
->db
.db_data
;
769 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
773 blkptr_t
*bps
= db
->db
.db_data
;
774 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
777 * We want to verify that all the blkptrs in the
778 * indirect block are holes, but we may have
779 * automatically set up a few fields for them.
780 * We iterate through each blkptr and verify
781 * they only have those fields set.
784 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
786 blkptr_t
*bp
= &bps
[i
];
787 ASSERT(ZIO_CHECKSUM_IS_ZERO(
790 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
791 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
792 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
793 ASSERT0(bp
->blk_fill
);
794 ASSERT0(bp
->blk_pad
[0]);
795 ASSERT0(bp
->blk_pad
[1]);
796 ASSERT(!BP_IS_EMBEDDED(bp
));
797 ASSERT(BP_IS_HOLE(bp
));
798 ASSERT0(bp
->blk_phys_birth
);
808 dbuf_clear_data(dmu_buf_impl_t
*db
)
810 ASSERT(MUTEX_HELD(&db
->db_mtx
));
812 ASSERT3P(db
->db_buf
, ==, NULL
);
813 db
->db
.db_data
= NULL
;
814 if (db
->db_state
!= DB_NOFILL
)
815 db
->db_state
= DB_UNCACHED
;
819 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
821 ASSERT(MUTEX_HELD(&db
->db_mtx
));
825 ASSERT(buf
->b_data
!= NULL
);
826 db
->db
.db_data
= buf
->b_data
;
830 * Loan out an arc_buf for read. Return the loaned arc_buf.
833 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
837 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
838 mutex_enter(&db
->db_mtx
);
839 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
840 int blksz
= db
->db
.db_size
;
841 spa_t
*spa
= db
->db_objset
->os_spa
;
843 mutex_exit(&db
->db_mtx
);
844 abuf
= arc_loan_buf(spa
, blksz
);
845 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
848 arc_loan_inuse_buf(abuf
, db
);
851 mutex_exit(&db
->db_mtx
);
857 * Calculate which level n block references the data at the level 0 offset
861 dbuf_whichblock(dnode_t
*dn
, int64_t level
, uint64_t offset
)
863 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
865 * The level n blkid is equal to the level 0 blkid divided by
866 * the number of level 0s in a level n block.
868 * The level 0 blkid is offset >> datablkshift =
869 * offset / 2^datablkshift.
871 * The number of level 0s in a level n is the number of block
872 * pointers in an indirect block, raised to the power of level.
873 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
874 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
876 * Thus, the level n blkid is: offset /
877 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
878 * = offset / 2^(datablkshift + level *
879 * (indblkshift - SPA_BLKPTRSHIFT))
880 * = offset >> (datablkshift + level *
881 * (indblkshift - SPA_BLKPTRSHIFT))
883 return (offset
>> (dn
->dn_datablkshift
+ level
*
884 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
)));
886 ASSERT3U(offset
, <, dn
->dn_datablksz
);
892 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
894 dmu_buf_impl_t
*db
= vdb
;
896 mutex_enter(&db
->db_mtx
);
897 ASSERT3U(db
->db_state
, ==, DB_READ
);
899 * All reads are synchronous, so we must have a hold on the dbuf
901 ASSERT(refcount_count(&db
->db_holds
) > 0);
902 ASSERT(db
->db_buf
== NULL
);
903 ASSERT(db
->db
.db_data
== NULL
);
904 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
905 /* we were freed in flight; disregard any error */
906 arc_release(buf
, db
);
907 bzero(buf
->b_data
, db
->db
.db_size
);
909 db
->db_freed_in_flight
= FALSE
;
910 dbuf_set_data(db
, buf
);
911 db
->db_state
= DB_CACHED
;
912 } else if (zio
== NULL
|| zio
->io_error
== 0) {
913 dbuf_set_data(db
, buf
);
914 db
->db_state
= DB_CACHED
;
916 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
917 ASSERT3P(db
->db_buf
, ==, NULL
);
918 arc_buf_destroy(buf
, db
);
919 db
->db_state
= DB_UNCACHED
;
921 cv_broadcast(&db
->db_changed
);
922 dbuf_rele_and_unlock(db
, NULL
);
926 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
930 arc_flags_t aflags
= ARC_FLAG_NOWAIT
;
934 ASSERT(!refcount_is_zero(&db
->db_holds
));
935 /* We need the struct_rwlock to prevent db_blkptr from changing. */
936 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
937 ASSERT(MUTEX_HELD(&db
->db_mtx
));
938 ASSERT(db
->db_state
== DB_UNCACHED
);
939 ASSERT(db
->db_buf
== NULL
);
941 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
942 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
944 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
945 db
->db
.db_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
946 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
947 if (bonuslen
< DN_MAX_BONUSLEN
)
948 bzero(db
->db
.db_data
, DN_MAX_BONUSLEN
);
950 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
952 db
->db_state
= DB_CACHED
;
953 mutex_exit(&db
->db_mtx
);
958 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
959 * processes the delete record and clears the bp while we are waiting
960 * for the dn_mtx (resulting in a "no" from block_freed).
962 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
963 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
964 BP_IS_HOLE(db
->db_blkptr
)))) {
965 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
967 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
,
968 db
->db
.db_size
, db
, type
));
969 bzero(db
->db
.db_data
, db
->db
.db_size
);
971 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
972 BP_IS_HOLE(db
->db_blkptr
) &&
973 db
->db_blkptr
->blk_birth
!= 0) {
974 blkptr_t
*bps
= db
->db
.db_data
;
975 for (int i
= 0; i
< ((1 <<
976 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
978 blkptr_t
*bp
= &bps
[i
];
979 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
980 1 << dn
->dn_indblkshift
);
982 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
984 BP_GET_LSIZE(db
->db_blkptr
));
985 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
987 BP_GET_LEVEL(db
->db_blkptr
) - 1);
988 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
992 db
->db_state
= DB_CACHED
;
993 mutex_exit(&db
->db_mtx
);
999 db
->db_state
= DB_READ
;
1000 mutex_exit(&db
->db_mtx
);
1002 if (DBUF_IS_L2CACHEABLE(db
))
1003 aflags
|= ARC_FLAG_L2CACHE
;
1005 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
1006 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
1007 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1009 dbuf_add_ref(db
, NULL
);
1011 (void) arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1012 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
1013 (flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
1018 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1021 boolean_t havepzio
= (zio
!= NULL
);
1026 * We don't have to hold the mutex to check db_state because it
1027 * can't be freed while we have a hold on the buffer.
1029 ASSERT(!refcount_is_zero(&db
->db_holds
));
1031 if (db
->db_state
== DB_NOFILL
)
1032 return (SET_ERROR(EIO
));
1036 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1037 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1039 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1040 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1041 DBUF_IS_CACHEABLE(db
);
1043 mutex_enter(&db
->db_mtx
);
1044 if (db
->db_state
== DB_CACHED
) {
1045 mutex_exit(&db
->db_mtx
);
1047 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1048 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1049 rw_exit(&dn
->dn_struct_rwlock
);
1051 } else if (db
->db_state
== DB_UNCACHED
) {
1052 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1055 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1056 dbuf_read_impl(db
, zio
, flags
);
1058 /* dbuf_read_impl has dropped db_mtx for us */
1061 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1063 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1064 rw_exit(&dn
->dn_struct_rwlock
);
1068 err
= zio_wait(zio
);
1071 * Another reader came in while the dbuf was in flight
1072 * between UNCACHED and CACHED. Either a writer will finish
1073 * writing the buffer (sending the dbuf to CACHED) or the
1074 * first reader's request will reach the read_done callback
1075 * and send the dbuf to CACHED. Otherwise, a failure
1076 * occurred and the dbuf went to UNCACHED.
1078 mutex_exit(&db
->db_mtx
);
1080 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1081 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1082 rw_exit(&dn
->dn_struct_rwlock
);
1085 /* Skip the wait per the caller's request. */
1086 mutex_enter(&db
->db_mtx
);
1087 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1088 while (db
->db_state
== DB_READ
||
1089 db
->db_state
== DB_FILL
) {
1090 ASSERT(db
->db_state
== DB_READ
||
1091 (flags
& DB_RF_HAVESTRUCT
) == 0);
1092 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1094 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1096 if (db
->db_state
== DB_UNCACHED
)
1097 err
= SET_ERROR(EIO
);
1099 mutex_exit(&db
->db_mtx
);
1102 ASSERT(err
|| havepzio
|| db
->db_state
== DB_CACHED
);
1107 dbuf_noread(dmu_buf_impl_t
*db
)
1109 ASSERT(!refcount_is_zero(&db
->db_holds
));
1110 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1111 mutex_enter(&db
->db_mtx
);
1112 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1113 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1114 if (db
->db_state
== DB_UNCACHED
) {
1115 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1116 spa_t
*spa
= db
->db_objset
->os_spa
;
1118 ASSERT(db
->db_buf
== NULL
);
1119 ASSERT(db
->db
.db_data
== NULL
);
1120 dbuf_set_data(db
, arc_alloc_buf(spa
, db
->db
.db_size
, db
, type
));
1121 db
->db_state
= DB_FILL
;
1122 } else if (db
->db_state
== DB_NOFILL
) {
1123 dbuf_clear_data(db
);
1125 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1127 mutex_exit(&db
->db_mtx
);
1131 * This is our just-in-time copy function. It makes a copy of
1132 * buffers, that have been modified in a previous transaction
1133 * group, before we modify them in the current active group.
1135 * This function is used in two places: when we are dirtying a
1136 * buffer for the first time in a txg, and when we are freeing
1137 * a range in a dnode that includes this buffer.
1139 * Note that when we are called from dbuf_free_range() we do
1140 * not put a hold on the buffer, we just traverse the active
1141 * dbuf list for the dnode.
1144 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1146 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1148 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1149 ASSERT(db
->db
.db_data
!= NULL
);
1150 ASSERT(db
->db_level
== 0);
1151 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1154 (dr
->dt
.dl
.dr_data
!=
1155 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1159 * If the last dirty record for this dbuf has not yet synced
1160 * and its referencing the dbuf data, either:
1161 * reset the reference to point to a new copy,
1162 * or (if there a no active holders)
1163 * just null out the current db_data pointer.
1165 ASSERT(dr
->dr_txg
>= txg
- 2);
1166 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1167 /* Note that the data bufs here are zio_bufs */
1168 dr
->dt
.dl
.dr_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
1169 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1170 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, DN_MAX_BONUSLEN
);
1171 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1172 int size
= db
->db
.db_size
;
1173 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1174 spa_t
*spa
= db
->db_objset
->os_spa
;
1176 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, size
, db
, type
);
1177 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1180 dbuf_clear_data(db
);
1185 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1187 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1188 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1189 uint64_t txg
= dr
->dr_txg
;
1191 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1192 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1193 ASSERT(db
->db_level
== 0);
1195 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1196 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1199 ASSERT(db
->db_data_pending
!= dr
);
1201 /* free this block */
1202 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1203 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1205 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1206 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1209 * Release the already-written buffer, so we leave it in
1210 * a consistent dirty state. Note that all callers are
1211 * modifying the buffer, so they will immediately do
1212 * another (redundant) arc_release(). Therefore, leave
1213 * the buf thawed to save the effort of freezing &
1214 * immediately re-thawing it.
1216 arc_release(dr
->dt
.dl
.dr_data
, db
);
1220 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1221 * data blocks in the free range, so that any future readers will find
1224 * This is a no-op if the dataset is in the middle of an incremental
1225 * receive; see comment below for details.
1228 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1231 dmu_buf_impl_t db_search
;
1232 dmu_buf_impl_t
*db
, *db_next
;
1233 uint64_t txg
= tx
->tx_txg
;
1235 boolean_t freespill
=
1236 (start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
);
1238 if (end_blkid
> dn
->dn_maxblkid
&& !freespill
)
1239 end_blkid
= dn
->dn_maxblkid
;
1240 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1242 db_search
.db_level
= 0;
1243 db_search
.db_blkid
= start_blkid
;
1244 db_search
.db_state
= DB_SEARCH
;
1246 mutex_enter(&dn
->dn_dbufs_mtx
);
1247 if (start_blkid
>= dn
->dn_unlisted_l0_blkid
&& !freespill
) {
1248 /* There can't be any dbufs in this range; no need to search. */
1250 db
= avl_find(&dn
->dn_dbufs
, &db_search
, &where
);
1251 ASSERT3P(db
, ==, NULL
);
1252 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1253 ASSERT(db
== NULL
|| db
->db_level
> 0);
1255 mutex_exit(&dn
->dn_dbufs_mtx
);
1257 } else if (dmu_objset_is_receiving(dn
->dn_objset
)) {
1259 * If we are receiving, we expect there to be no dbufs in
1260 * the range to be freed, because receive modifies each
1261 * block at most once, and in offset order. If this is
1262 * not the case, it can lead to performance problems,
1263 * so note that we unexpectedly took the slow path.
1265 atomic_inc_64(&zfs_free_range_recv_miss
);
1268 db
= avl_find(&dn
->dn_dbufs
, &db_search
, &where
);
1269 ASSERT3P(db
, ==, NULL
);
1270 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1272 for (; db
!= NULL
; db
= db_next
) {
1273 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1274 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1276 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1279 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1281 /* found a level 0 buffer in the range */
1282 mutex_enter(&db
->db_mtx
);
1283 if (dbuf_undirty(db
, tx
)) {
1284 /* mutex has been dropped and dbuf destroyed */
1288 if (db
->db_state
== DB_UNCACHED
||
1289 db
->db_state
== DB_NOFILL
||
1290 db
->db_state
== DB_EVICTING
) {
1291 ASSERT(db
->db
.db_data
== NULL
);
1292 mutex_exit(&db
->db_mtx
);
1295 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1296 /* will be handled in dbuf_read_done or dbuf_rele */
1297 db
->db_freed_in_flight
= TRUE
;
1298 mutex_exit(&db
->db_mtx
);
1301 if (refcount_count(&db
->db_holds
) == 0) {
1306 /* The dbuf is referenced */
1308 if (db
->db_last_dirty
!= NULL
) {
1309 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1311 if (dr
->dr_txg
== txg
) {
1313 * This buffer is "in-use", re-adjust the file
1314 * size to reflect that this buffer may
1315 * contain new data when we sync.
1317 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1318 db
->db_blkid
> dn
->dn_maxblkid
)
1319 dn
->dn_maxblkid
= db
->db_blkid
;
1320 dbuf_unoverride(dr
);
1323 * This dbuf is not dirty in the open context.
1324 * Either uncache it (if its not referenced in
1325 * the open context) or reset its contents to
1328 dbuf_fix_old_data(db
, txg
);
1331 /* clear the contents if its cached */
1332 if (db
->db_state
== DB_CACHED
) {
1333 ASSERT(db
->db
.db_data
!= NULL
);
1334 arc_release(db
->db_buf
, db
);
1335 bzero(db
->db
.db_data
, db
->db
.db_size
);
1336 arc_buf_freeze(db
->db_buf
);
1339 mutex_exit(&db
->db_mtx
);
1341 mutex_exit(&dn
->dn_dbufs_mtx
);
1345 dbuf_block_freeable(dmu_buf_impl_t
*db
)
1347 dsl_dataset_t
*ds
= db
->db_objset
->os_dsl_dataset
;
1348 uint64_t birth_txg
= 0;
1351 * We don't need any locking to protect db_blkptr:
1352 * If it's syncing, then db_last_dirty will be set
1353 * so we'll ignore db_blkptr.
1355 * This logic ensures that only block births for
1356 * filled blocks are considered.
1358 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1359 if (db
->db_last_dirty
&& (db
->db_blkptr
== NULL
||
1360 !BP_IS_HOLE(db
->db_blkptr
))) {
1361 birth_txg
= db
->db_last_dirty
->dr_txg
;
1362 } else if (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1363 birth_txg
= db
->db_blkptr
->blk_birth
;
1367 * If this block don't exist or is in a snapshot, it can't be freed.
1368 * Don't pass the bp to dsl_dataset_block_freeable() since we
1369 * are holding the db_mtx lock and might deadlock if we are
1370 * prefetching a dedup-ed block.
1373 return (ds
== NULL
||
1374 dsl_dataset_block_freeable(ds
, NULL
, birth_txg
));
1380 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1382 arc_buf_t
*buf
, *obuf
;
1383 int osize
= db
->db
.db_size
;
1384 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1387 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1392 /* XXX does *this* func really need the lock? */
1393 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1396 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1397 * is OK, because there can be no other references to the db
1398 * when we are changing its size, so no concurrent DB_FILL can
1402 * XXX we should be doing a dbuf_read, checking the return
1403 * value and returning that up to our callers
1405 dmu_buf_will_dirty(&db
->db
, tx
);
1407 /* create the data buffer for the new block */
1408 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, size
, db
, type
);
1410 /* copy old block data to the new block */
1412 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1413 /* zero the remainder */
1415 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1417 mutex_enter(&db
->db_mtx
);
1418 dbuf_set_data(db
, buf
);
1419 arc_buf_destroy(obuf
, db
);
1420 db
->db
.db_size
= size
;
1422 if (db
->db_level
== 0) {
1423 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1424 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1426 mutex_exit(&db
->db_mtx
);
1428 dnode_willuse_space(dn
, size
-osize
, tx
);
1433 dbuf_release_bp(dmu_buf_impl_t
*db
)
1435 objset_t
*os
= db
->db_objset
;
1437 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1438 ASSERT(arc_released(os
->os_phys_buf
) ||
1439 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1440 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1442 (void) arc_release(db
->db_buf
, db
);
1446 * We already have a dirty record for this TXG, and we are being
1450 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1452 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1454 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1456 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1458 * If this buffer has already been written out,
1459 * we now need to reset its state.
1461 dbuf_unoverride(dr
);
1462 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1463 db
->db_state
!= DB_NOFILL
) {
1464 /* Already released on initial dirty, so just thaw. */
1465 ASSERT(arc_released(db
->db_buf
));
1466 arc_buf_thaw(db
->db_buf
);
1471 dbuf_dirty_record_t
*
1472 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1476 dbuf_dirty_record_t
**drp
, *dr
;
1477 int drop_struct_lock
= FALSE
;
1478 boolean_t do_free_accounting
= B_FALSE
;
1479 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1481 ASSERT(tx
->tx_txg
!= 0);
1482 ASSERT(!refcount_is_zero(&db
->db_holds
));
1483 DMU_TX_DIRTY_BUF(tx
, db
);
1488 * Shouldn't dirty a regular buffer in syncing context. Private
1489 * objects may be dirtied in syncing context, but only if they
1490 * were already pre-dirtied in open context.
1492 ASSERT(!dmu_tx_is_syncing(tx
) ||
1493 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1494 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1495 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1497 * We make this assert for private objects as well, but after we
1498 * check if we're already dirty. They are allowed to re-dirty
1499 * in syncing context.
1501 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1502 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1503 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1505 mutex_enter(&db
->db_mtx
);
1507 * XXX make this true for indirects too? The problem is that
1508 * transactions created with dmu_tx_create_assigned() from
1509 * syncing context don't bother holding ahead.
1511 ASSERT(db
->db_level
!= 0 ||
1512 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1513 db
->db_state
== DB_NOFILL
);
1515 mutex_enter(&dn
->dn_mtx
);
1517 * Don't set dirtyctx to SYNC if we're just modifying this as we
1518 * initialize the objset.
1520 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
&&
1521 !BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1523 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1524 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1525 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1527 mutex_exit(&dn
->dn_mtx
);
1529 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1530 dn
->dn_have_spill
= B_TRUE
;
1533 * If this buffer is already dirty, we're done.
1535 drp
= &db
->db_last_dirty
;
1536 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1537 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1538 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1540 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1544 mutex_exit(&db
->db_mtx
);
1549 * Only valid if not already dirty.
1551 ASSERT(dn
->dn_object
== 0 ||
1552 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1553 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1555 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1556 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1557 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1558 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1559 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1560 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1563 * We should only be dirtying in syncing context if it's the
1564 * mos or we're initializing the os or it's a special object.
1565 * However, we are allowed to dirty in syncing context provided
1566 * we already dirtied it in open context. Hence we must make
1567 * this assertion only if we're not already dirty.
1570 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1571 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1572 ASSERT(db
->db
.db_size
!= 0);
1574 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1576 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1578 * Update the accounting.
1579 * Note: we delay "free accounting" until after we drop
1580 * the db_mtx. This keeps us from grabbing other locks
1581 * (and possibly deadlocking) in bp_get_dsize() while
1582 * also holding the db_mtx.
1584 dnode_willuse_space(dn
, db
->db
.db_size
, tx
);
1585 do_free_accounting
= dbuf_block_freeable(db
);
1589 * If this buffer is dirty in an old transaction group we need
1590 * to make a copy of it so that the changes we make in this
1591 * transaction group won't leak out when we sync the older txg.
1593 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1594 if (db
->db_level
== 0) {
1595 void *data_old
= db
->db_buf
;
1597 if (db
->db_state
!= DB_NOFILL
) {
1598 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1599 dbuf_fix_old_data(db
, tx
->tx_txg
);
1600 data_old
= db
->db
.db_data
;
1601 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1603 * Release the data buffer from the cache so
1604 * that we can modify it without impacting
1605 * possible other users of this cached data
1606 * block. Note that indirect blocks and
1607 * private objects are not released until the
1608 * syncing state (since they are only modified
1611 arc_release(db
->db_buf
, db
);
1612 dbuf_fix_old_data(db
, tx
->tx_txg
);
1613 data_old
= db
->db_buf
;
1615 ASSERT(data_old
!= NULL
);
1617 dr
->dt
.dl
.dr_data
= data_old
;
1619 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
1620 list_create(&dr
->dt
.di
.dr_children
,
1621 sizeof (dbuf_dirty_record_t
),
1622 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1624 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1625 dr
->dr_accounted
= db
->db
.db_size
;
1627 dr
->dr_txg
= tx
->tx_txg
;
1632 * We could have been freed_in_flight between the dbuf_noread
1633 * and dbuf_dirty. We win, as though the dbuf_noread() had
1634 * happened after the free.
1636 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1637 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1638 mutex_enter(&dn
->dn_mtx
);
1639 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1640 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1643 mutex_exit(&dn
->dn_mtx
);
1644 db
->db_freed_in_flight
= FALSE
;
1648 * This buffer is now part of this txg
1650 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1651 db
->db_dirtycnt
+= 1;
1652 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1654 mutex_exit(&db
->db_mtx
);
1656 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1657 db
->db_blkid
== DMU_SPILL_BLKID
) {
1658 mutex_enter(&dn
->dn_mtx
);
1659 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1660 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1661 mutex_exit(&dn
->dn_mtx
);
1662 dnode_setdirty(dn
, tx
);
1668 * The dn_struct_rwlock prevents db_blkptr from changing
1669 * due to a write from syncing context completing
1670 * while we are running, so we want to acquire it before
1671 * looking at db_blkptr.
1673 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1674 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1675 drop_struct_lock
= TRUE
;
1678 if (do_free_accounting
) {
1679 blkptr_t
*bp
= db
->db_blkptr
;
1680 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1681 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1683 * This is only a guess -- if the dbuf is dirty
1684 * in a previous txg, we don't know how much
1685 * space it will use on disk yet. We should
1686 * really have the struct_rwlock to access
1687 * db_blkptr, but since this is just a guess,
1688 * it's OK if we get an odd answer.
1690 ddt_prefetch(os
->os_spa
, bp
);
1691 dnode_willuse_space(dn
, -willfree
, tx
);
1694 if (db
->db_level
== 0) {
1695 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1696 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1699 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1700 dmu_buf_impl_t
*parent
= db
->db_parent
;
1701 dbuf_dirty_record_t
*di
;
1702 int parent_held
= FALSE
;
1704 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1705 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1707 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1708 db
->db_blkid
>> epbs
, FTAG
);
1709 ASSERT(parent
!= NULL
);
1712 if (drop_struct_lock
)
1713 rw_exit(&dn
->dn_struct_rwlock
);
1714 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1715 di
= dbuf_dirty(parent
, tx
);
1717 dbuf_rele(parent
, FTAG
);
1719 mutex_enter(&db
->db_mtx
);
1721 * Since we've dropped the mutex, it's possible that
1722 * dbuf_undirty() might have changed this out from under us.
1724 if (db
->db_last_dirty
== dr
||
1725 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1726 mutex_enter(&di
->dt
.di
.dr_mtx
);
1727 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1728 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1729 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1730 mutex_exit(&di
->dt
.di
.dr_mtx
);
1733 mutex_exit(&db
->db_mtx
);
1735 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1736 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1737 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1738 mutex_enter(&dn
->dn_mtx
);
1739 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1740 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1741 mutex_exit(&dn
->dn_mtx
);
1742 if (drop_struct_lock
)
1743 rw_exit(&dn
->dn_struct_rwlock
);
1746 dnode_setdirty(dn
, tx
);
1752 * Undirty a buffer in the transaction group referenced by the given
1753 * transaction. Return whether this evicted the dbuf.
1756 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1759 uint64_t txg
= tx
->tx_txg
;
1760 dbuf_dirty_record_t
*dr
, **drp
;
1765 * Due to our use of dn_nlevels below, this can only be called
1766 * in open context, unless we are operating on the MOS.
1767 * From syncing context, dn_nlevels may be different from the
1768 * dn_nlevels used when dbuf was dirtied.
1770 ASSERT(db
->db_objset
==
1771 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1772 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1773 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1774 ASSERT0(db
->db_level
);
1775 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1778 * If this buffer is not dirty, we're done.
1780 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1781 if (dr
->dr_txg
<= txg
)
1783 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1785 ASSERT(dr
->dr_txg
== txg
);
1786 ASSERT(dr
->dr_dbuf
== db
);
1791 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1793 ASSERT(db
->db
.db_size
!= 0);
1795 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1796 dr
->dr_accounted
, txg
);
1801 * Note that there are three places in dbuf_dirty()
1802 * where this dirty record may be put on a list.
1803 * Make sure to do a list_remove corresponding to
1804 * every one of those list_insert calls.
1806 if (dr
->dr_parent
) {
1807 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1808 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1809 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1810 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1811 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1812 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1813 mutex_enter(&dn
->dn_mtx
);
1814 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1815 mutex_exit(&dn
->dn_mtx
);
1819 if (db
->db_state
!= DB_NOFILL
) {
1820 dbuf_unoverride(dr
);
1822 ASSERT(db
->db_buf
!= NULL
);
1823 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1824 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1825 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
1828 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1830 ASSERT(db
->db_dirtycnt
> 0);
1831 db
->db_dirtycnt
-= 1;
1833 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1834 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
1843 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1845 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1846 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1848 ASSERT(tx
->tx_txg
!= 0);
1849 ASSERT(!refcount_is_zero(&db
->db_holds
));
1852 * Quick check for dirtyness. For already dirty blocks, this
1853 * reduces runtime of this function by >90%, and overall performance
1854 * by 50% for some workloads (e.g. file deletion with indirect blocks
1857 mutex_enter(&db
->db_mtx
);
1858 dbuf_dirty_record_t
*dr
;
1859 for (dr
= db
->db_last_dirty
;
1860 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1862 * It's possible that it is already dirty but not cached,
1863 * because there are some calls to dbuf_dirty() that don't
1864 * go through dmu_buf_will_dirty().
1866 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1867 /* This dbuf is already dirty and cached. */
1869 mutex_exit(&db
->db_mtx
);
1873 mutex_exit(&db
->db_mtx
);
1876 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1877 rf
|= DB_RF_HAVESTRUCT
;
1879 (void) dbuf_read(db
, NULL
, rf
);
1880 (void) dbuf_dirty(db
, tx
);
1884 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1886 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1888 db
->db_state
= DB_NOFILL
;
1890 dmu_buf_will_fill(db_fake
, tx
);
1894 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1896 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1898 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1899 ASSERT(tx
->tx_txg
!= 0);
1900 ASSERT(db
->db_level
== 0);
1901 ASSERT(!refcount_is_zero(&db
->db_holds
));
1903 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1904 dmu_tx_private_ok(tx
));
1907 (void) dbuf_dirty(db
, tx
);
1910 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1913 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1915 mutex_enter(&db
->db_mtx
);
1918 if (db
->db_state
== DB_FILL
) {
1919 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1920 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1921 /* we were freed while filling */
1922 /* XXX dbuf_undirty? */
1923 bzero(db
->db
.db_data
, db
->db
.db_size
);
1924 db
->db_freed_in_flight
= FALSE
;
1926 db
->db_state
= DB_CACHED
;
1927 cv_broadcast(&db
->db_changed
);
1929 mutex_exit(&db
->db_mtx
);
1933 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
1934 bp_embedded_type_t etype
, enum zio_compress comp
,
1935 int uncompressed_size
, int compressed_size
, int byteorder
,
1938 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
1939 struct dirty_leaf
*dl
;
1940 dmu_object_type_t type
;
1942 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
1943 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
1944 SPA_FEATURE_EMBEDDED_DATA
));
1948 type
= DB_DNODE(db
)->dn_type
;
1951 ASSERT0(db
->db_level
);
1952 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1954 dmu_buf_will_not_fill(dbuf
, tx
);
1956 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1957 dl
= &db
->db_last_dirty
->dt
.dl
;
1958 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
1959 data
, comp
, uncompressed_size
, compressed_size
);
1960 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
1961 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
1962 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
1963 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
1965 dl
->dr_override_state
= DR_OVERRIDDEN
;
1966 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
1970 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1971 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1974 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1976 ASSERT(!refcount_is_zero(&db
->db_holds
));
1977 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1978 ASSERT(db
->db_level
== 0);
1979 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1980 ASSERT(buf
!= NULL
);
1981 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1982 ASSERT(tx
->tx_txg
!= 0);
1984 arc_return_buf(buf
, db
);
1985 ASSERT(arc_released(buf
));
1987 mutex_enter(&db
->db_mtx
);
1989 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1990 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1992 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1994 if (db
->db_state
== DB_CACHED
&&
1995 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1996 mutex_exit(&db
->db_mtx
);
1997 (void) dbuf_dirty(db
, tx
);
1998 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1999 arc_buf_destroy(buf
, db
);
2000 xuio_stat_wbuf_copied();
2004 xuio_stat_wbuf_nocopy();
2005 if (db
->db_state
== DB_CACHED
) {
2006 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2008 ASSERT(db
->db_buf
!= NULL
);
2009 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2010 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2011 if (!arc_released(db
->db_buf
)) {
2012 ASSERT(dr
->dt
.dl
.dr_override_state
==
2014 arc_release(db
->db_buf
, db
);
2016 dr
->dt
.dl
.dr_data
= buf
;
2017 arc_buf_destroy(db
->db_buf
, db
);
2018 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2019 arc_release(db
->db_buf
, db
);
2020 arc_buf_destroy(db
->db_buf
, db
);
2024 ASSERT(db
->db_buf
== NULL
);
2025 dbuf_set_data(db
, buf
);
2026 db
->db_state
= DB_FILL
;
2027 mutex_exit(&db
->db_mtx
);
2028 (void) dbuf_dirty(db
, tx
);
2029 dmu_buf_fill_done(&db
->db
, tx
);
2033 dbuf_destroy(dmu_buf_impl_t
*db
)
2036 dmu_buf_impl_t
*parent
= db
->db_parent
;
2037 dmu_buf_impl_t
*dndb
;
2039 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2040 ASSERT(refcount_is_zero(&db
->db_holds
));
2042 if (db
->db_buf
!= NULL
) {
2043 arc_buf_destroy(db
->db_buf
, db
);
2047 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2048 ASSERT(db
->db
.db_data
!= NULL
);
2049 zio_buf_free(db
->db
.db_data
, DN_MAX_BONUSLEN
);
2050 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
2051 db
->db_state
= DB_UNCACHED
;
2054 dbuf_clear_data(db
);
2056 if (multilist_link_active(&db
->db_cache_link
)) {
2057 multilist_remove(&dbuf_cache
, db
);
2058 (void) refcount_remove_many(&dbuf_cache_size
,
2059 db
->db
.db_size
, db
);
2062 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2063 ASSERT(db
->db_data_pending
== NULL
);
2065 db
->db_state
= DB_EVICTING
;
2066 db
->db_blkptr
= NULL
;
2069 * Now that db_state is DB_EVICTING, nobody else can find this via
2070 * the hash table. We can now drop db_mtx, which allows us to
2071 * acquire the dn_dbufs_mtx.
2073 mutex_exit(&db
->db_mtx
);
2078 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2079 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2081 mutex_enter(&dn
->dn_dbufs_mtx
);
2082 avl_remove(&dn
->dn_dbufs
, db
);
2083 atomic_dec_32(&dn
->dn_dbufs_count
);
2087 mutex_exit(&dn
->dn_dbufs_mtx
);
2089 * Decrementing the dbuf count means that the hold corresponding
2090 * to the removed dbuf is no longer discounted in dnode_move(),
2091 * so the dnode cannot be moved until after we release the hold.
2092 * The membar_producer() ensures visibility of the decremented
2093 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2097 db
->db_dnode_handle
= NULL
;
2099 dbuf_hash_remove(db
);
2104 ASSERT(refcount_is_zero(&db
->db_holds
));
2106 db
->db_parent
= NULL
;
2108 ASSERT(db
->db_buf
== NULL
);
2109 ASSERT(db
->db
.db_data
== NULL
);
2110 ASSERT(db
->db_hash_next
== NULL
);
2111 ASSERT(db
->db_blkptr
== NULL
);
2112 ASSERT(db
->db_data_pending
== NULL
);
2113 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2115 kmem_cache_free(dbuf_kmem_cache
, db
);
2116 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2119 * If this dbuf is referenced from an indirect dbuf,
2120 * decrement the ref count on the indirect dbuf.
2122 if (parent
&& parent
!= dndb
)
2123 dbuf_rele(parent
, db
);
2127 * Note: While bpp will always be updated if the function returns success,
2128 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2129 * this happens when the dnode is the meta-dnode, or a userused or groupused
2133 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2134 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
2141 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2143 if (blkid
== DMU_SPILL_BLKID
) {
2144 mutex_enter(&dn
->dn_mtx
);
2145 if (dn
->dn_have_spill
&&
2146 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2147 *bpp
= &dn
->dn_phys
->dn_spill
;
2150 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2151 *parentp
= dn
->dn_dbuf
;
2152 mutex_exit(&dn
->dn_mtx
);
2156 if (dn
->dn_phys
->dn_nlevels
== 0)
2159 nlevels
= dn
->dn_phys
->dn_nlevels
;
2161 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2163 ASSERT3U(level
* epbs
, <, 64);
2164 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2165 if (level
>= nlevels
||
2166 (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2167 /* the buffer has no parent yet */
2168 return (SET_ERROR(ENOENT
));
2169 } else if (level
< nlevels
-1) {
2170 /* this block is referenced from an indirect block */
2171 int err
= dbuf_hold_impl(dn
, level
+1,
2172 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2175 err
= dbuf_read(*parentp
, NULL
,
2176 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2178 dbuf_rele(*parentp
, NULL
);
2182 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2183 (blkid
& ((1ULL << epbs
) - 1));
2186 /* the block is referenced from the dnode */
2187 ASSERT3U(level
, ==, nlevels
-1);
2188 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2189 blkid
< dn
->dn_phys
->dn_nblkptr
);
2191 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2192 *parentp
= dn
->dn_dbuf
;
2194 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2199 static dmu_buf_impl_t
*
2200 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2201 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2203 objset_t
*os
= dn
->dn_objset
;
2204 dmu_buf_impl_t
*db
, *odb
;
2206 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2207 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2209 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2212 db
->db
.db_object
= dn
->dn_object
;
2213 db
->db_level
= level
;
2214 db
->db_blkid
= blkid
;
2215 db
->db_last_dirty
= NULL
;
2216 db
->db_dirtycnt
= 0;
2217 db
->db_dnode_handle
= dn
->dn_handle
;
2218 db
->db_parent
= parent
;
2219 db
->db_blkptr
= blkptr
;
2222 db
->db_user_immediate_evict
= FALSE
;
2223 db
->db_freed_in_flight
= FALSE
;
2224 db
->db_pending_evict
= FALSE
;
2226 if (blkid
== DMU_BONUS_BLKID
) {
2227 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2228 db
->db
.db_size
= DN_MAX_BONUSLEN
-
2229 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2230 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2231 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2232 db
->db_state
= DB_UNCACHED
;
2233 /* the bonus dbuf is not placed in the hash table */
2234 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2236 } else if (blkid
== DMU_SPILL_BLKID
) {
2237 db
->db
.db_size
= (blkptr
!= NULL
) ?
2238 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2239 db
->db
.db_offset
= 0;
2242 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2243 db
->db
.db_size
= blocksize
;
2244 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2248 * Hold the dn_dbufs_mtx while we get the new dbuf
2249 * in the hash table *and* added to the dbufs list.
2250 * This prevents a possible deadlock with someone
2251 * trying to look up this dbuf before its added to the
2254 mutex_enter(&dn
->dn_dbufs_mtx
);
2255 db
->db_state
= DB_EVICTING
;
2256 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2257 /* someone else inserted it first */
2258 kmem_cache_free(dbuf_kmem_cache
, db
);
2259 mutex_exit(&dn
->dn_dbufs_mtx
);
2262 avl_add(&dn
->dn_dbufs
, db
);
2263 if (db
->db_level
== 0 && db
->db_blkid
>=
2264 dn
->dn_unlisted_l0_blkid
)
2265 dn
->dn_unlisted_l0_blkid
= db
->db_blkid
+ 1;
2266 db
->db_state
= DB_UNCACHED
;
2267 mutex_exit(&dn
->dn_dbufs_mtx
);
2268 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2270 if (parent
&& parent
!= dn
->dn_dbuf
)
2271 dbuf_add_ref(parent
, db
);
2273 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2274 refcount_count(&dn
->dn_holds
) > 0);
2275 (void) refcount_add(&dn
->dn_holds
, db
);
2276 atomic_inc_32(&dn
->dn_dbufs_count
);
2278 dprintf_dbuf(db
, "db=%p\n", db
);
2283 typedef struct dbuf_prefetch_arg
{
2284 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2285 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2286 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2287 int dpa_curlevel
; /* The current level that we're reading */
2288 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2289 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2290 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2291 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2292 } dbuf_prefetch_arg_t
;
2295 * Actually issue the prefetch read for the block given.
2298 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2300 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2303 arc_flags_t aflags
=
2304 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2306 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2307 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2308 ASSERT(dpa
->dpa_zio
!= NULL
);
2309 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2310 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2311 &aflags
, &dpa
->dpa_zb
);
2315 * Called when an indirect block above our prefetch target is read in. This
2316 * will either read in the next indirect block down the tree or issue the actual
2317 * prefetch if the next block down is our target.
2320 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2322 dbuf_prefetch_arg_t
*dpa
= private;
2324 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2325 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2328 * The dpa_dnode is only valid if we are called with a NULL
2329 * zio. This indicates that the arc_read() returned without
2330 * first calling zio_read() to issue a physical read. Once
2331 * a physical read is made the dpa_dnode must be invalidated
2332 * as the locks guarding it may have been dropped. If the
2333 * dpa_dnode is still valid, then we want to add it to the dbuf
2334 * cache. To do so, we must hold the dbuf associated with the block
2335 * we just prefetched, read its contents so that we associate it
2336 * with an arc_buf_t, and then release it.
2339 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2340 if (zio
->io_flags
& ZIO_FLAG_RAW
) {
2341 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2343 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2345 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2347 dpa
->dpa_dnode
= NULL
;
2348 } else if (dpa
->dpa_dnode
!= NULL
) {
2349 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2350 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2351 dpa
->dpa_zb
.zb_level
));
2352 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2353 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2354 (void) dbuf_read(db
, NULL
,
2355 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2356 dbuf_rele(db
, FTAG
);
2359 dpa
->dpa_curlevel
--;
2361 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2362 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2363 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
2364 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2365 if (BP_IS_HOLE(bp
) || (zio
!= NULL
&& zio
->io_error
!= 0)) {
2366 kmem_free(dpa
, sizeof (*dpa
));
2367 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2368 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2369 dbuf_issue_final_prefetch(dpa
, bp
);
2370 kmem_free(dpa
, sizeof (*dpa
));
2372 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2373 zbookmark_phys_t zb
;
2375 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2377 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2378 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2380 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2381 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2382 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2386 arc_buf_destroy(abuf
, private);
2390 * Issue prefetch reads for the given block on the given level. If the indirect
2391 * blocks above that block are not in memory, we will read them in
2392 * asynchronously. As a result, this call never blocks waiting for a read to
2396 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2400 int epbs
, nlevels
, curlevel
;
2403 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2404 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2406 if (blkid
> dn
->dn_maxblkid
)
2409 if (dnode_block_freed(dn
, blkid
))
2413 * This dnode hasn't been written to disk yet, so there's nothing to
2416 nlevels
= dn
->dn_phys
->dn_nlevels
;
2417 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2420 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2421 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2424 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2427 mutex_exit(&db
->db_mtx
);
2429 * This dbuf already exists. It is either CACHED, or
2430 * (we assume) about to be read or filled.
2436 * Find the closest ancestor (indirect block) of the target block
2437 * that is present in the cache. In this indirect block, we will
2438 * find the bp that is at curlevel, curblkid.
2442 while (curlevel
< nlevels
- 1) {
2443 int parent_level
= curlevel
+ 1;
2444 uint64_t parent_blkid
= curblkid
>> epbs
;
2447 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2448 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2449 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2450 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2451 dbuf_rele(db
, FTAG
);
2455 curlevel
= parent_level
;
2456 curblkid
= parent_blkid
;
2459 if (curlevel
== nlevels
- 1) {
2460 /* No cached indirect blocks found. */
2461 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2462 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2464 if (BP_IS_HOLE(&bp
))
2467 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2469 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2472 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2473 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
2474 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2475 dn
->dn_object
, level
, blkid
);
2476 dpa
->dpa_curlevel
= curlevel
;
2477 dpa
->dpa_prio
= prio
;
2478 dpa
->dpa_aflags
= aflags
;
2479 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2480 dpa
->dpa_dnode
= dn
;
2481 dpa
->dpa_epbs
= epbs
;
2485 * If we have the indirect just above us, no need to do the asynchronous
2486 * prefetch chain; we'll just run the last step ourselves. If we're at
2487 * a higher level, though, we want to issue the prefetches for all the
2488 * indirect blocks asynchronously, so we can go on with whatever we were
2491 if (curlevel
== level
) {
2492 ASSERT3U(curblkid
, ==, blkid
);
2493 dbuf_issue_final_prefetch(dpa
, &bp
);
2494 kmem_free(dpa
, sizeof (*dpa
));
2496 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2497 zbookmark_phys_t zb
;
2499 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2500 dn
->dn_object
, curlevel
, curblkid
);
2501 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2502 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2503 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2507 * We use pio here instead of dpa_zio since it's possible that
2508 * dpa may have already been freed.
2514 * Returns with db_holds incremented, and db_mtx not held.
2515 * Note: dn_struct_rwlock must be held.
2518 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2519 boolean_t fail_sparse
, boolean_t fail_uncached
,
2520 void *tag
, dmu_buf_impl_t
**dbp
)
2522 dmu_buf_impl_t
*db
, *parent
= NULL
;
2524 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2525 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2526 ASSERT3U(dn
->dn_nlevels
, >, level
);
2530 /* dbuf_find() returns with db_mtx held */
2531 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
, level
, blkid
);
2534 blkptr_t
*bp
= NULL
;
2538 return (SET_ERROR(ENOENT
));
2540 ASSERT3P(parent
, ==, NULL
);
2541 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
2543 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
2544 err
= SET_ERROR(ENOENT
);
2547 dbuf_rele(parent
, NULL
);
2551 if (err
&& err
!= ENOENT
)
2553 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
);
2556 if (fail_uncached
&& db
->db_state
!= DB_CACHED
) {
2557 mutex_exit(&db
->db_mtx
);
2558 return (SET_ERROR(ENOENT
));
2561 if (db
->db_buf
!= NULL
)
2562 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
2564 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
2567 * If this buffer is currently syncing out, and we are are
2568 * still referencing it from db_data, we need to make a copy
2569 * of it in case we decide we want to dirty it again in this txg.
2571 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
2572 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2573 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
2574 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
2576 if (dr
->dt
.dl
.dr_data
== db
->db_buf
) {
2577 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2580 arc_alloc_buf(dn
->dn_objset
->os_spa
,
2581 db
->db
.db_size
, db
, type
));
2582 bcopy(dr
->dt
.dl
.dr_data
->b_data
, db
->db
.db_data
,
2587 if (multilist_link_active(&db
->db_cache_link
)) {
2588 ASSERT(refcount_is_zero(&db
->db_holds
));
2589 multilist_remove(&dbuf_cache
, db
);
2590 (void) refcount_remove_many(&dbuf_cache_size
,
2591 db
->db
.db_size
, db
);
2593 (void) refcount_add(&db
->db_holds
, tag
);
2595 mutex_exit(&db
->db_mtx
);
2597 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2599 dbuf_rele(parent
, NULL
);
2601 ASSERT3P(DB_DNODE(db
), ==, dn
);
2602 ASSERT3U(db
->db_blkid
, ==, blkid
);
2603 ASSERT3U(db
->db_level
, ==, level
);
2610 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2612 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2616 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2619 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2620 return (err
? NULL
: db
);
2624 dbuf_create_bonus(dnode_t
*dn
)
2626 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2628 ASSERT(dn
->dn_bonus
== NULL
);
2629 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2633 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2635 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2638 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2639 return (SET_ERROR(ENOTSUP
));
2641 blksz
= SPA_MINBLOCKSIZE
;
2642 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2643 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2647 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2648 dbuf_new_size(db
, blksz
, tx
);
2649 rw_exit(&dn
->dn_struct_rwlock
);
2656 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2658 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2661 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2663 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2665 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
2666 ASSERT3S(holds
, >, 1);
2669 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2671 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2674 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2675 dmu_buf_impl_t
*found_db
;
2676 boolean_t result
= B_FALSE
;
2678 if (db
->db_blkid
== DMU_BONUS_BLKID
)
2679 found_db
= dbuf_find_bonus(os
, obj
);
2681 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2683 if (found_db
!= NULL
) {
2684 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2685 (void) refcount_add(&db
->db_holds
, tag
);
2688 mutex_exit(&db
->db_mtx
);
2694 * If you call dbuf_rele() you had better not be referencing the dnode handle
2695 * unless you have some other direct or indirect hold on the dnode. (An indirect
2696 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2697 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2698 * dnode's parent dbuf evicting its dnode handles.
2701 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2703 mutex_enter(&db
->db_mtx
);
2704 dbuf_rele_and_unlock(db
, tag
);
2708 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2710 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2714 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2715 * db_dirtycnt and db_holds to be updated atomically.
2718 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2722 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2726 * Remove the reference to the dbuf before removing its hold on the
2727 * dnode so we can guarantee in dnode_move() that a referenced bonus
2728 * buffer has a corresponding dnode hold.
2730 holds
= refcount_remove(&db
->db_holds
, tag
);
2734 * We can't freeze indirects if there is a possibility that they
2735 * may be modified in the current syncing context.
2737 if (db
->db_buf
!= NULL
&&
2738 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
2739 arc_buf_freeze(db
->db_buf
);
2742 if (holds
== db
->db_dirtycnt
&&
2743 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2744 dbuf_evict_user(db
);
2747 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2749 boolean_t evict_dbuf
= db
->db_pending_evict
;
2752 * If the dnode moves here, we cannot cross this
2753 * barrier until the move completes.
2758 atomic_dec_32(&dn
->dn_dbufs_count
);
2761 * Decrementing the dbuf count means that the bonus
2762 * buffer's dnode hold is no longer discounted in
2763 * dnode_move(). The dnode cannot move until after
2764 * the dnode_rele() below.
2769 * Do not reference db after its lock is dropped.
2770 * Another thread may evict it.
2772 mutex_exit(&db
->db_mtx
);
2775 dnode_evict_bonus(dn
);
2778 } else if (db
->db_buf
== NULL
) {
2780 * This is a special case: we never associated this
2781 * dbuf with any data allocated from the ARC.
2783 ASSERT(db
->db_state
== DB_UNCACHED
||
2784 db
->db_state
== DB_NOFILL
);
2786 } else if (arc_released(db
->db_buf
)) {
2788 * This dbuf has anonymous data associated with it.
2792 boolean_t do_arc_evict
= B_FALSE
;
2794 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
2796 if (!DBUF_IS_CACHEABLE(db
) &&
2797 db
->db_blkptr
!= NULL
&&
2798 !BP_IS_HOLE(db
->db_blkptr
) &&
2799 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2800 do_arc_evict
= B_TRUE
;
2801 bp
= *db
->db_blkptr
;
2804 if (!DBUF_IS_CACHEABLE(db
) ||
2805 db
->db_pending_evict
) {
2807 } else if (!multilist_link_active(&db
->db_cache_link
)) {
2808 multilist_insert(&dbuf_cache
, db
);
2809 (void) refcount_add_many(&dbuf_cache_size
,
2810 db
->db
.db_size
, db
);
2811 mutex_exit(&db
->db_mtx
);
2813 dbuf_evict_notify();
2817 arc_freed(spa
, &bp
);
2820 mutex_exit(&db
->db_mtx
);
2825 #pragma weak dmu_buf_refcount = dbuf_refcount
2827 dbuf_refcount(dmu_buf_impl_t
*db
)
2829 return (refcount_count(&db
->db_holds
));
2833 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
2834 dmu_buf_user_t
*new_user
)
2836 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2838 mutex_enter(&db
->db_mtx
);
2839 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2840 if (db
->db_user
== old_user
)
2841 db
->db_user
= new_user
;
2843 old_user
= db
->db_user
;
2844 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2845 mutex_exit(&db
->db_mtx
);
2851 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2853 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
2857 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2859 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2861 db
->db_user_immediate_evict
= TRUE
;
2862 return (dmu_buf_set_user(db_fake
, user
));
2866 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2868 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
2872 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2874 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2876 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2877 return (db
->db_user
);
2881 dmu_buf_user_evict_wait()
2883 taskq_wait(dbu_evict_taskq
);
2887 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2889 boolean_t res
= B_FALSE
;
2890 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2893 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2894 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2900 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2902 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2903 return (dbi
->db_blkptr
);
2907 dmu_buf_get_objset(dmu_buf_t
*db
)
2909 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2910 return (dbi
->db_objset
);
2914 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2916 /* ASSERT(dmu_tx_is_syncing(tx) */
2917 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2919 if (db
->db_blkptr
!= NULL
)
2922 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2923 db
->db_blkptr
= &dn
->dn_phys
->dn_spill
;
2924 BP_ZERO(db
->db_blkptr
);
2927 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2929 * This buffer was allocated at a time when there was
2930 * no available blkptrs from the dnode, or it was
2931 * inappropriate to hook it in (i.e., nlevels mis-match).
2933 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2934 ASSERT(db
->db_parent
== NULL
);
2935 db
->db_parent
= dn
->dn_dbuf
;
2936 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2939 dmu_buf_impl_t
*parent
= db
->db_parent
;
2940 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2942 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2943 if (parent
== NULL
) {
2944 mutex_exit(&db
->db_mtx
);
2945 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2946 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2947 db
->db_blkid
>> epbs
, db
);
2948 rw_exit(&dn
->dn_struct_rwlock
);
2949 mutex_enter(&db
->db_mtx
);
2950 db
->db_parent
= parent
;
2952 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2953 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2959 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2961 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2965 ASSERT(dmu_tx_is_syncing(tx
));
2967 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2969 mutex_enter(&db
->db_mtx
);
2971 ASSERT(db
->db_level
> 0);
2974 /* Read the block if it hasn't been read yet. */
2975 if (db
->db_buf
== NULL
) {
2976 mutex_exit(&db
->db_mtx
);
2977 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2978 mutex_enter(&db
->db_mtx
);
2980 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2981 ASSERT(db
->db_buf
!= NULL
);
2985 /* Indirect block size must match what the dnode thinks it is. */
2986 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2987 dbuf_check_blkptr(dn
, db
);
2990 /* Provide the pending dirty record to child dbufs */
2991 db
->db_data_pending
= dr
;
2993 mutex_exit(&db
->db_mtx
);
2994 dbuf_write(dr
, db
->db_buf
, tx
);
2997 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2998 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
2999 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3000 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3005 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3007 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3008 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3011 uint64_t txg
= tx
->tx_txg
;
3013 ASSERT(dmu_tx_is_syncing(tx
));
3015 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3017 mutex_enter(&db
->db_mtx
);
3019 * To be synced, we must be dirtied. But we
3020 * might have been freed after the dirty.
3022 if (db
->db_state
== DB_UNCACHED
) {
3023 /* This buffer has been freed since it was dirtied */
3024 ASSERT(db
->db
.db_data
== NULL
);
3025 } else if (db
->db_state
== DB_FILL
) {
3026 /* This buffer was freed and is now being re-filled */
3027 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3029 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3036 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3037 mutex_enter(&dn
->dn_mtx
);
3038 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3039 mutex_exit(&dn
->dn_mtx
);
3043 * If this is a bonus buffer, simply copy the bonus data into the
3044 * dnode. It will be written out when the dnode is synced (and it
3045 * will be synced, since it must have been dirty for dbuf_sync to
3048 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3049 dbuf_dirty_record_t
**drp
;
3051 ASSERT(*datap
!= NULL
);
3052 ASSERT0(db
->db_level
);
3053 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
3054 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
3057 if (*datap
!= db
->db
.db_data
) {
3058 zio_buf_free(*datap
, DN_MAX_BONUSLEN
);
3059 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
3061 db
->db_data_pending
= NULL
;
3062 drp
= &db
->db_last_dirty
;
3064 drp
= &(*drp
)->dr_next
;
3065 ASSERT(dr
->dr_next
== NULL
);
3066 ASSERT(dr
->dr_dbuf
== db
);
3068 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3069 ASSERT(db
->db_dirtycnt
> 0);
3070 db
->db_dirtycnt
-= 1;
3071 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3078 * This function may have dropped the db_mtx lock allowing a dmu_sync
3079 * operation to sneak in. As a result, we need to ensure that we
3080 * don't check the dr_override_state until we have returned from
3081 * dbuf_check_blkptr.
3083 dbuf_check_blkptr(dn
, db
);
3086 * If this buffer is in the middle of an immediate write,
3087 * wait for the synchronous IO to complete.
3089 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3090 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3091 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3092 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3095 if (db
->db_state
!= DB_NOFILL
&&
3096 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3097 refcount_count(&db
->db_holds
) > 1 &&
3098 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3099 *datap
== db
->db_buf
) {
3101 * If this buffer is currently "in use" (i.e., there
3102 * are active holds and db_data still references it),
3103 * then make a copy before we start the write so that
3104 * any modifications from the open txg will not leak
3107 * NOTE: this copy does not need to be made for
3108 * objects only modified in the syncing context (e.g.
3109 * DNONE_DNODE blocks).
3111 int blksz
= arc_buf_size(*datap
);
3112 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3113 *datap
= arc_alloc_buf(os
->os_spa
, blksz
, db
, type
);
3114 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
3116 db
->db_data_pending
= dr
;
3118 mutex_exit(&db
->db_mtx
);
3120 dbuf_write(dr
, *datap
, tx
);
3122 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3123 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3124 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3128 * Although zio_nowait() does not "wait for an IO", it does
3129 * initiate the IO. If this is an empty write it seems plausible
3130 * that the IO could actually be completed before the nowait
3131 * returns. We need to DB_DNODE_EXIT() first in case
3132 * zio_nowait() invalidates the dbuf.
3135 zio_nowait(dr
->dr_zio
);
3140 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3142 dbuf_dirty_record_t
*dr
;
3144 while (dr
= list_head(list
)) {
3145 if (dr
->dr_zio
!= NULL
) {
3147 * If we find an already initialized zio then we
3148 * are processing the meta-dnode, and we have finished.
3149 * The dbufs for all dnodes are put back on the list
3150 * during processing, so that we can zio_wait()
3151 * these IOs after initiating all child IOs.
3153 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3154 DMU_META_DNODE_OBJECT
);
3157 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3158 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3159 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3161 list_remove(list
, dr
);
3162 if (dr
->dr_dbuf
->db_level
> 0)
3163 dbuf_sync_indirect(dr
, tx
);
3165 dbuf_sync_leaf(dr
, tx
);
3171 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3173 dmu_buf_impl_t
*db
= vdb
;
3175 blkptr_t
*bp
= zio
->io_bp
;
3176 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3177 spa_t
*spa
= zio
->io_spa
;
3182 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3183 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3187 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3188 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3189 zio
->io_prev_space_delta
= delta
;
3191 if (bp
->blk_birth
!= 0) {
3192 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3193 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3194 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3195 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3196 BP_IS_EMBEDDED(bp
));
3197 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3200 mutex_enter(&db
->db_mtx
);
3203 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3204 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3205 ASSERT(!(BP_IS_HOLE(bp
)) &&
3206 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3210 if (db
->db_level
== 0) {
3211 mutex_enter(&dn
->dn_mtx
);
3212 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3213 db
->db_blkid
!= DMU_SPILL_BLKID
)
3214 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3215 mutex_exit(&dn
->dn_mtx
);
3217 if (dn
->dn_type
== DMU_OT_DNODE
) {
3218 dnode_phys_t
*dnp
= db
->db
.db_data
;
3219 for (i
= db
->db
.db_size
>> DNODE_SHIFT
; i
> 0;
3221 if (dnp
->dn_type
!= DMU_OT_NONE
)
3225 if (BP_IS_HOLE(bp
)) {
3232 blkptr_t
*ibp
= db
->db
.db_data
;
3233 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3234 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3235 if (BP_IS_HOLE(ibp
))
3237 fill
+= BP_GET_FILL(ibp
);
3242 if (!BP_IS_EMBEDDED(bp
))
3243 bp
->blk_fill
= fill
;
3245 mutex_exit(&db
->db_mtx
);
3247 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3248 *db
->db_blkptr
= *bp
;
3249 rw_exit(&dn
->dn_struct_rwlock
);
3254 * This function gets called just prior to running through the compression
3255 * stage of the zio pipeline. If we're an indirect block comprised of only
3256 * holes, then we want this indirect to be compressed away to a hole. In
3257 * order to do that we must zero out any information about the holes that
3258 * this indirect points to prior to before we try to compress it.
3261 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3263 dmu_buf_impl_t
*db
= vdb
;
3269 ASSERT3U(db
->db_level
, >, 0);
3272 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3274 /* Determine if all our children are holes */
3275 for (i
= 0, bp
= db
->db
.db_data
; i
< 1 << epbs
; i
++, bp
++) {
3276 if (!BP_IS_HOLE(bp
))
3281 * If all the children are holes, then zero them all out so that
3282 * we may get compressed away.
3284 if (i
== 1 << epbs
) {
3285 /* didn't find any non-holes */
3286 bzero(db
->db
.db_data
, db
->db
.db_size
);
3292 * The SPA will call this callback several times for each zio - once
3293 * for every physical child i/o (zio->io_phys_children times). This
3294 * allows the DMU to monitor the progress of each logical i/o. For example,
3295 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3296 * block. There may be a long delay before all copies/fragments are completed,
3297 * so this callback allows us to retire dirty space gradually, as the physical
3302 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3304 dmu_buf_impl_t
*db
= arg
;
3305 objset_t
*os
= db
->db_objset
;
3306 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3307 dbuf_dirty_record_t
*dr
;
3310 dr
= db
->db_data_pending
;
3311 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3314 * The callback will be called io_phys_children times. Retire one
3315 * portion of our dirty space each time we are called. Any rounding
3316 * error will be cleaned up by dsl_pool_sync()'s call to
3317 * dsl_pool_undirty_space().
3319 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3320 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3325 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3327 dmu_buf_impl_t
*db
= vdb
;
3328 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3329 blkptr_t
*bp
= db
->db_blkptr
;
3330 objset_t
*os
= db
->db_objset
;
3331 dmu_tx_t
*tx
= os
->os_synctx
;
3332 dbuf_dirty_record_t
**drp
, *dr
;
3334 ASSERT0(zio
->io_error
);
3335 ASSERT(db
->db_blkptr
== bp
);
3338 * For nopwrites and rewrites we ensure that the bp matches our
3339 * original and bypass all the accounting.
3341 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3342 ASSERT(BP_EQUAL(bp
, bp_orig
));
3344 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3345 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3346 dsl_dataset_block_born(ds
, bp
, tx
);
3349 mutex_enter(&db
->db_mtx
);
3353 drp
= &db
->db_last_dirty
;
3354 while ((dr
= *drp
) != db
->db_data_pending
)
3356 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3357 ASSERT(dr
->dr_dbuf
== db
);
3358 ASSERT(dr
->dr_next
== NULL
);
3362 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3367 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3368 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3369 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3374 if (db
->db_level
== 0) {
3375 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3376 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3377 if (db
->db_state
!= DB_NOFILL
) {
3378 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3379 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
3386 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3387 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3388 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3390 dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3391 ASSERT3U(db
->db_blkid
, <=,
3392 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3393 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3397 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3398 list_destroy(&dr
->dt
.di
.dr_children
);
3400 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3402 cv_broadcast(&db
->db_changed
);
3403 ASSERT(db
->db_dirtycnt
> 0);
3404 db
->db_dirtycnt
-= 1;
3405 db
->db_data_pending
= NULL
;
3406 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3410 dbuf_write_nofill_ready(zio_t
*zio
)
3412 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3416 dbuf_write_nofill_done(zio_t
*zio
)
3418 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3422 dbuf_write_override_ready(zio_t
*zio
)
3424 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3425 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3427 dbuf_write_ready(zio
, NULL
, db
);
3431 dbuf_write_override_done(zio_t
*zio
)
3433 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3434 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3435 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3437 mutex_enter(&db
->db_mtx
);
3438 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3439 if (!BP_IS_HOLE(obp
))
3440 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3441 arc_release(dr
->dt
.dl
.dr_data
, db
);
3443 mutex_exit(&db
->db_mtx
);
3445 dbuf_write_done(zio
, NULL
, db
);
3448 /* Issue I/O to commit a dirty buffer to disk. */
3450 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3452 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3455 dmu_buf_impl_t
*parent
= db
->db_parent
;
3456 uint64_t txg
= tx
->tx_txg
;
3457 zbookmark_phys_t zb
;
3462 ASSERT(dmu_tx_is_syncing(tx
));
3468 if (db
->db_state
!= DB_NOFILL
) {
3469 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3471 * Private object buffers are released here rather
3472 * than in dbuf_dirty() since they are only modified
3473 * in the syncing context and we don't want the
3474 * overhead of making multiple copies of the data.
3476 if (BP_IS_HOLE(db
->db_blkptr
)) {
3479 dbuf_release_bp(db
);
3484 if (parent
!= dn
->dn_dbuf
) {
3485 /* Our parent is an indirect block. */
3486 /* We have a dirty parent that has been scheduled for write. */
3487 ASSERT(parent
&& parent
->db_data_pending
);
3488 /* Our parent's buffer is one level closer to the dnode. */
3489 ASSERT(db
->db_level
== parent
->db_level
-1);
3491 * We're about to modify our parent's db_data by modifying
3492 * our block pointer, so the parent must be released.
3494 ASSERT(arc_released(parent
->db_buf
));
3495 zio
= parent
->db_data_pending
->dr_zio
;
3497 /* Our parent is the dnode itself. */
3498 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3499 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3500 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3501 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3502 ASSERT3P(db
->db_blkptr
, ==,
3503 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3507 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3508 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3511 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3512 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3513 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3515 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3517 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3519 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3523 * We copy the blkptr now (rather than when we instantiate the dirty
3524 * record), because its value can change between open context and
3525 * syncing context. We do not need to hold dn_struct_rwlock to read
3526 * db_blkptr because we are in syncing context.
3528 dr
->dr_bp_copy
= *db
->db_blkptr
;
3530 if (db
->db_level
== 0 &&
3531 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3533 * The BP for this block has been provided by open context
3534 * (by dmu_sync() or dmu_buf_write_embedded()).
3536 void *contents
= (data
!= NULL
) ? data
->b_data
: NULL
;
3538 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3539 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, &zp
,
3540 dbuf_write_override_ready
, NULL
, NULL
,
3541 dbuf_write_override_done
,
3542 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3543 mutex_enter(&db
->db_mtx
);
3544 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3545 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3546 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3547 mutex_exit(&db
->db_mtx
);
3548 } else if (db
->db_state
== DB_NOFILL
) {
3549 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
3550 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
3551 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3552 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, &zp
,
3553 dbuf_write_nofill_ready
, NULL
, NULL
,
3554 dbuf_write_nofill_done
, db
,
3555 ZIO_PRIORITY_ASYNC_WRITE
,
3556 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3558 ASSERT(arc_released(data
));
3561 * For indirect blocks, we want to setup the children
3562 * ready callback so that we can properly handle an indirect
3563 * block that only contains holes.
3565 arc_done_func_t
*children_ready_cb
= NULL
;
3566 if (db
->db_level
!= 0)
3567 children_ready_cb
= dbuf_write_children_ready
;
3569 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3570 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3571 &zp
, dbuf_write_ready
, children_ready_cb
,
3572 dbuf_write_physdone
, dbuf_write_done
, db
,
3573 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);