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, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_dbuf.h>
48 #include <sys/callb.h>
51 struct dbuf_hold_impl_data
{
52 /* Function arguments */
56 boolean_t dh_fail_sparse
;
57 boolean_t dh_fail_uncached
;
59 dmu_buf_impl_t
**dh_dbp
;
61 dmu_buf_impl_t
*dh_db
;
62 dmu_buf_impl_t
*dh_parent
;
65 dbuf_dirty_record_t
*dh_dr
;
69 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
70 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, boolean_t fail_sparse
,
71 boolean_t fail_uncached
,
72 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
73 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
75 uint_t zfs_dbuf_evict_key
;
77 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
78 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
80 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
81 dmu_buf_evict_func_t
*evict_func_sync
,
82 dmu_buf_evict_func_t
*evict_func_async
,
83 dmu_buf_t
**clear_on_evict_dbufp
);
86 * Global data structures and functions for the dbuf cache.
88 static kmem_cache_t
*dbuf_kmem_cache
;
89 static taskq_t
*dbu_evict_taskq
;
91 static kthread_t
*dbuf_cache_evict_thread
;
92 static kmutex_t dbuf_evict_lock
;
93 static kcondvar_t dbuf_evict_cv
;
94 static boolean_t dbuf_evict_thread_exit
;
97 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
98 * are not currently held but have been recently released. These dbufs
99 * are not eligible for arc eviction until they are aged out of the cache.
100 * Dbufs are added to the dbuf cache once the last hold is released. If a
101 * dbuf is later accessed and still exists in the dbuf cache, then it will
102 * be removed from the cache and later re-added to the head of the cache.
103 * Dbufs that are aged out of the cache will be immediately destroyed and
104 * become eligible for arc eviction.
106 static multilist_t
*dbuf_cache
;
107 static refcount_t dbuf_cache_size
;
108 unsigned long dbuf_cache_max_bytes
= 100 * 1024 * 1024;
110 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
111 int dbuf_cache_max_shift
= 5;
114 * The dbuf cache uses a three-stage eviction policy:
115 * - A low water marker designates when the dbuf eviction thread
116 * should stop evicting from the dbuf cache.
117 * - When we reach the maximum size (aka mid water mark), we
118 * signal the eviction thread to run.
119 * - The high water mark indicates when the eviction thread
120 * is unable to keep up with the incoming load and eviction must
121 * happen in the context of the calling thread.
125 * low water mid water hi water
126 * +----------------------------------------+----------+----------+
131 * +----------------------------------------+----------+----------+
133 * evicting eviction directly
136 * The high and low water marks indicate the operating range for the eviction
137 * thread. The low water mark is, by default, 90% of the total size of the
138 * cache and the high water mark is at 110% (both of these percentages can be
139 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
140 * respectively). The eviction thread will try to ensure that the cache remains
141 * within this range by waking up every second and checking if the cache is
142 * above the low water mark. The thread can also be woken up by callers adding
143 * elements into the cache if the cache is larger than the mid water (i.e max
144 * cache size). Once the eviction thread is woken up and eviction is required,
145 * it will continue evicting buffers until it's able to reduce the cache size
146 * to the low water mark. If the cache size continues to grow and hits the high
147 * water mark, then callers adding elements to the cache will begin to evict
148 * directly from the cache until the cache is no longer above the high water
153 * The percentage above and below the maximum cache size.
155 uint_t dbuf_cache_hiwater_pct
= 10;
156 uint_t dbuf_cache_lowater_pct
= 10;
160 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
162 dmu_buf_impl_t
*db
= vdb
;
163 bzero(db
, sizeof (dmu_buf_impl_t
));
165 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
166 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
167 multilist_link_init(&db
->db_cache_link
);
168 refcount_create(&db
->db_holds
);
175 dbuf_dest(void *vdb
, void *unused
)
177 dmu_buf_impl_t
*db
= vdb
;
178 mutex_destroy(&db
->db_mtx
);
179 cv_destroy(&db
->db_changed
);
180 ASSERT(!multilist_link_active(&db
->db_cache_link
));
181 refcount_destroy(&db
->db_holds
);
185 * dbuf hash table routines
187 static dbuf_hash_table_t dbuf_hash_table
;
189 static uint64_t dbuf_hash_count
;
192 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
194 uintptr_t osv
= (uintptr_t)os
;
195 uint64_t crc
= -1ULL;
197 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
198 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
199 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
200 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
201 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
202 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
203 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
205 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
210 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
211 ((dbuf)->db.db_object == (obj) && \
212 (dbuf)->db_objset == (os) && \
213 (dbuf)->db_level == (level) && \
214 (dbuf)->db_blkid == (blkid))
217 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
219 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
224 hv
= dbuf_hash(os
, obj
, level
, blkid
);
225 idx
= hv
& h
->hash_table_mask
;
227 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
228 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
229 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
230 mutex_enter(&db
->db_mtx
);
231 if (db
->db_state
!= DB_EVICTING
) {
232 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
235 mutex_exit(&db
->db_mtx
);
238 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
242 static dmu_buf_impl_t
*
243 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
246 dmu_buf_impl_t
*db
= NULL
;
248 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
249 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
250 if (dn
->dn_bonus
!= NULL
) {
252 mutex_enter(&db
->db_mtx
);
254 rw_exit(&dn
->dn_struct_rwlock
);
255 dnode_rele(dn
, FTAG
);
261 * Insert an entry into the hash table. If there is already an element
262 * equal to elem in the hash table, then the already existing element
263 * will be returned and the new element will not be inserted.
264 * Otherwise returns NULL.
266 static dmu_buf_impl_t
*
267 dbuf_hash_insert(dmu_buf_impl_t
*db
)
269 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
270 objset_t
*os
= db
->db_objset
;
271 uint64_t obj
= db
->db
.db_object
;
272 int level
= db
->db_level
;
273 uint64_t blkid
, hv
, idx
;
276 blkid
= db
->db_blkid
;
277 hv
= dbuf_hash(os
, obj
, level
, blkid
);
278 idx
= hv
& h
->hash_table_mask
;
280 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
281 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
282 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
283 mutex_enter(&dbf
->db_mtx
);
284 if (dbf
->db_state
!= DB_EVICTING
) {
285 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
288 mutex_exit(&dbf
->db_mtx
);
292 mutex_enter(&db
->db_mtx
);
293 db
->db_hash_next
= h
->hash_table
[idx
];
294 h
->hash_table
[idx
] = db
;
295 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
296 atomic_inc_64(&dbuf_hash_count
);
302 * Remove an entry from the hash table. It must be in the EVICTING state.
305 dbuf_hash_remove(dmu_buf_impl_t
*db
)
307 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
309 dmu_buf_impl_t
*dbf
, **dbp
;
311 hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
312 db
->db_level
, db
->db_blkid
);
313 idx
= hv
& h
->hash_table_mask
;
316 * We mustn't hold db_mtx to maintain lock ordering:
317 * DBUF_HASH_MUTEX > db_mtx.
319 ASSERT(refcount_is_zero(&db
->db_holds
));
320 ASSERT(db
->db_state
== DB_EVICTING
);
321 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
323 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
324 dbp
= &h
->hash_table
[idx
];
325 while ((dbf
= *dbp
) != db
) {
326 dbp
= &dbf
->db_hash_next
;
329 *dbp
= db
->db_hash_next
;
330 db
->db_hash_next
= NULL
;
331 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
332 atomic_dec_64(&dbuf_hash_count
);
338 } dbvu_verify_type_t
;
341 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
346 if (db
->db_user
== NULL
)
349 /* Only data blocks support the attachment of user data. */
350 ASSERT(db
->db_level
== 0);
352 /* Clients must resolve a dbuf before attaching user data. */
353 ASSERT(db
->db
.db_data
!= NULL
);
354 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
356 holds
= refcount_count(&db
->db_holds
);
357 if (verify_type
== DBVU_EVICTING
) {
359 * Immediate eviction occurs when holds == dirtycnt.
360 * For normal eviction buffers, holds is zero on
361 * eviction, except when dbuf_fix_old_data() calls
362 * dbuf_clear_data(). However, the hold count can grow
363 * during eviction even though db_mtx is held (see
364 * dmu_bonus_hold() for an example), so we can only
365 * test the generic invariant that holds >= dirtycnt.
367 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
369 if (db
->db_user_immediate_evict
== TRUE
)
370 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
372 ASSERT3U(holds
, >, 0);
378 dbuf_evict_user(dmu_buf_impl_t
*db
)
380 dmu_buf_user_t
*dbu
= db
->db_user
;
382 ASSERT(MUTEX_HELD(&db
->db_mtx
));
387 dbuf_verify_user(db
, DBVU_EVICTING
);
391 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
392 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
396 * There are two eviction callbacks - one that we call synchronously
397 * and one that we invoke via a taskq. The async one is useful for
398 * avoiding lock order reversals and limiting stack depth.
400 * Note that if we have a sync callback but no async callback,
401 * it's likely that the sync callback will free the structure
402 * containing the dbu. In that case we need to take care to not
403 * dereference dbu after calling the sync evict func.
405 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
407 if (dbu
->dbu_evict_func_sync
!= NULL
)
408 dbu
->dbu_evict_func_sync(dbu
);
411 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
412 dbu
, 0, &dbu
->dbu_tqent
);
417 dbuf_is_metadata(dmu_buf_impl_t
*db
)
420 * Consider indirect blocks and spill blocks to be meta data.
422 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
425 boolean_t is_metadata
;
428 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
431 return (is_metadata
);
437 * This function *must* return indices evenly distributed between all
438 * sublists of the multilist. This is needed due to how the dbuf eviction
439 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
440 * distributed between all sublists and uses this assumption when
441 * deciding which sublist to evict from and how much to evict from it.
444 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
446 dmu_buf_impl_t
*db
= obj
;
449 * The assumption here, is the hash value for a given
450 * dmu_buf_impl_t will remain constant throughout it's lifetime
451 * (i.e. it's objset, object, level and blkid fields don't change).
452 * Thus, we don't need to store the dbuf's sublist index
453 * on insertion, as this index can be recalculated on removal.
455 * Also, the low order bits of the hash value are thought to be
456 * distributed evenly. Otherwise, in the case that the multilist
457 * has a power of two number of sublists, each sublists' usage
458 * would not be evenly distributed.
460 return (dbuf_hash(db
->db_objset
, db
->db
.db_object
,
461 db
->db_level
, db
->db_blkid
) %
462 multilist_get_num_sublists(ml
));
465 static inline unsigned long
466 dbuf_cache_target_bytes(void)
468 return MIN(dbuf_cache_max_bytes
,
469 arc_target_bytes() >> dbuf_cache_max_shift
);
472 static inline boolean_t
473 dbuf_cache_above_hiwater(void)
475 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
477 uint64_t dbuf_cache_hiwater_bytes
=
478 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100;
480 return (refcount_count(&dbuf_cache_size
) >
481 dbuf_cache_target
+ dbuf_cache_hiwater_bytes
);
484 static inline boolean_t
485 dbuf_cache_above_lowater(void)
487 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
489 uint64_t dbuf_cache_lowater_bytes
=
490 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100;
492 return (refcount_count(&dbuf_cache_size
) >
493 dbuf_cache_target
- dbuf_cache_lowater_bytes
);
497 * Evict the oldest eligible dbuf from the dbuf cache.
502 int idx
= multilist_get_random_index(dbuf_cache
);
503 multilist_sublist_t
*mls
= multilist_sublist_lock(dbuf_cache
, idx
);
505 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
508 * Set the thread's tsd to indicate that it's processing evictions.
509 * Once a thread stops evicting from the dbuf cache it will
510 * reset its tsd to NULL.
512 ASSERT3P(tsd_get(zfs_dbuf_evict_key
), ==, NULL
);
513 (void) tsd_set(zfs_dbuf_evict_key
, (void *)B_TRUE
);
515 dmu_buf_impl_t
*db
= multilist_sublist_tail(mls
);
516 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
517 db
= multilist_sublist_prev(mls
, db
);
520 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
521 multilist_sublist_t
*, mls
);
524 multilist_sublist_remove(mls
, db
);
525 multilist_sublist_unlock(mls
);
526 (void) refcount_remove_many(&dbuf_cache_size
,
530 multilist_sublist_unlock(mls
);
532 (void) tsd_set(zfs_dbuf_evict_key
, NULL
);
536 * The dbuf evict thread is responsible for aging out dbufs from the
537 * cache. Once the cache has reached it's maximum size, dbufs are removed
538 * and destroyed. The eviction thread will continue running until the size
539 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
540 * out of the cache it is destroyed and becomes eligible for arc eviction.
544 dbuf_evict_thread(void *unused
)
548 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
550 mutex_enter(&dbuf_evict_lock
);
551 while (!dbuf_evict_thread_exit
) {
552 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
553 CALLB_CPR_SAFE_BEGIN(&cpr
);
554 (void) cv_timedwait_sig_hires(&dbuf_evict_cv
,
555 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
556 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
558 mutex_exit(&dbuf_evict_lock
);
561 * Keep evicting as long as we're above the low water mark
562 * for the cache. We do this without holding the locks to
563 * minimize lock contention.
565 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
569 mutex_enter(&dbuf_evict_lock
);
572 dbuf_evict_thread_exit
= B_FALSE
;
573 cv_broadcast(&dbuf_evict_cv
);
574 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
579 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
580 * If the dbuf cache is at its high water mark, then evict a dbuf from the
581 * dbuf cache using the callers context.
584 dbuf_evict_notify(void)
588 * We use thread specific data to track when a thread has
589 * started processing evictions. This allows us to avoid deeply
590 * nested stacks that would have a call flow similar to this:
592 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
595 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
597 * The dbuf_eviction_thread will always have its tsd set until
598 * that thread exits. All other threads will only set their tsd
599 * if they are participating in the eviction process. This only
600 * happens if the eviction thread is unable to process evictions
601 * fast enough. To keep the dbuf cache size in check, other threads
602 * can evict from the dbuf cache directly. Those threads will set
603 * their tsd values so that we ensure that they only evict one dbuf
604 * from the dbuf cache.
606 if (tsd_get(zfs_dbuf_evict_key
) != NULL
)
610 * We check if we should evict without holding the dbuf_evict_lock,
611 * because it's OK to occasionally make the wrong decision here,
612 * and grabbing the lock results in massive lock contention.
614 if (refcount_count(&dbuf_cache_size
) > dbuf_cache_target_bytes()) {
615 if (dbuf_cache_above_hiwater())
617 cv_signal(&dbuf_evict_cv
);
626 uint64_t hsize
= 1ULL << 16;
627 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
631 * The hash table is big enough to fill all of physical memory
632 * with an average block size of zfs_arc_average_blocksize (default 8K).
633 * By default, the table will take up
634 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
636 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
640 h
->hash_table_mask
= hsize
- 1;
641 #if defined(_KERNEL) && defined(HAVE_SPL)
643 * Large allocations which do not require contiguous pages
644 * should be using vmem_alloc() in the linux kernel
646 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
648 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
650 if (h
->hash_table
== NULL
) {
651 /* XXX - we should really return an error instead of assert */
652 ASSERT(hsize
> (1ULL << 10));
657 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
658 sizeof (dmu_buf_impl_t
),
659 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
661 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
662 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
667 * Setup the parameters for the dbuf cache. We cap the size of the
668 * dbuf cache to 1/32nd (default) of the size of the ARC.
670 dbuf_cache_max_bytes
= MIN(dbuf_cache_max_bytes
,
671 arc_target_bytes() >> dbuf_cache_max_shift
);
674 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
675 * configuration is not required.
677 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
679 dbuf_cache
= multilist_create(sizeof (dmu_buf_impl_t
),
680 offsetof(dmu_buf_impl_t
, db_cache_link
),
681 dbuf_cache_multilist_index_func
);
682 refcount_create(&dbuf_cache_size
);
684 tsd_create(&zfs_dbuf_evict_key
, NULL
);
685 dbuf_evict_thread_exit
= B_FALSE
;
686 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
687 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
688 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
689 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
695 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
698 dbuf_stats_destroy();
700 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
701 mutex_destroy(&h
->hash_mutexes
[i
]);
702 #if defined(_KERNEL) && defined(HAVE_SPL)
704 * Large allocations which do not require contiguous pages
705 * should be using vmem_free() in the linux kernel
707 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
709 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
711 kmem_cache_destroy(dbuf_kmem_cache
);
712 taskq_destroy(dbu_evict_taskq
);
714 mutex_enter(&dbuf_evict_lock
);
715 dbuf_evict_thread_exit
= B_TRUE
;
716 while (dbuf_evict_thread_exit
) {
717 cv_signal(&dbuf_evict_cv
);
718 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
720 mutex_exit(&dbuf_evict_lock
);
721 tsd_destroy(&zfs_dbuf_evict_key
);
723 mutex_destroy(&dbuf_evict_lock
);
724 cv_destroy(&dbuf_evict_cv
);
726 refcount_destroy(&dbuf_cache_size
);
727 multilist_destroy(dbuf_cache
);
736 dbuf_verify(dmu_buf_impl_t
*db
)
739 dbuf_dirty_record_t
*dr
;
741 ASSERT(MUTEX_HELD(&db
->db_mtx
));
743 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
746 ASSERT(db
->db_objset
!= NULL
);
750 ASSERT(db
->db_parent
== NULL
);
751 ASSERT(db
->db_blkptr
== NULL
);
753 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
754 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
755 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
756 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
757 db
->db_blkid
== DMU_SPILL_BLKID
||
758 !avl_is_empty(&dn
->dn_dbufs
));
760 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
762 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
763 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
764 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
766 ASSERT0(db
->db
.db_offset
);
768 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
771 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
772 ASSERT(dr
->dr_dbuf
== db
);
774 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
775 ASSERT(dr
->dr_dbuf
== db
);
778 * We can't assert that db_size matches dn_datablksz because it
779 * can be momentarily different when another thread is doing
782 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
783 dr
= db
->db_data_pending
;
785 * It should only be modified in syncing context, so
786 * make sure we only have one copy of the data.
788 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
791 /* verify db->db_blkptr */
793 if (db
->db_parent
== dn
->dn_dbuf
) {
794 /* db is pointed to by the dnode */
795 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
796 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
797 ASSERT(db
->db_parent
== NULL
);
799 ASSERT(db
->db_parent
!= NULL
);
800 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
801 ASSERT3P(db
->db_blkptr
, ==,
802 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
804 /* db is pointed to by an indirect block */
805 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
807 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
808 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
811 * dnode_grow_indblksz() can make this fail if we don't
812 * have the struct_rwlock. XXX indblksz no longer
813 * grows. safe to do this now?
815 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
816 ASSERT3P(db
->db_blkptr
, ==,
817 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
818 db
->db_blkid
% epb
));
822 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
823 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
824 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
825 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
827 * If the blkptr isn't set but they have nonzero data,
828 * it had better be dirty, otherwise we'll lose that
829 * data when we evict this buffer.
831 * There is an exception to this rule for indirect blocks; in
832 * this case, if the indirect block is a hole, we fill in a few
833 * fields on each of the child blocks (importantly, birth time)
834 * to prevent hole birth times from being lost when you
835 * partially fill in a hole.
837 if (db
->db_dirtycnt
== 0) {
838 if (db
->db_level
== 0) {
839 uint64_t *buf
= db
->db
.db_data
;
842 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
846 blkptr_t
*bps
= db
->db
.db_data
;
847 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
850 * We want to verify that all the blkptrs in the
851 * indirect block are holes, but we may have
852 * automatically set up a few fields for them.
853 * We iterate through each blkptr and verify
854 * they only have those fields set.
857 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
859 blkptr_t
*bp
= &bps
[i
];
860 ASSERT(ZIO_CHECKSUM_IS_ZERO(
863 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
864 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
865 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
866 ASSERT0(bp
->blk_fill
);
867 ASSERT0(bp
->blk_pad
[0]);
868 ASSERT0(bp
->blk_pad
[1]);
869 ASSERT(!BP_IS_EMBEDDED(bp
));
870 ASSERT(BP_IS_HOLE(bp
));
871 ASSERT0(bp
->blk_phys_birth
);
881 dbuf_clear_data(dmu_buf_impl_t
*db
)
883 ASSERT(MUTEX_HELD(&db
->db_mtx
));
885 ASSERT3P(db
->db_buf
, ==, NULL
);
886 db
->db
.db_data
= NULL
;
887 if (db
->db_state
!= DB_NOFILL
)
888 db
->db_state
= DB_UNCACHED
;
892 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
894 ASSERT(MUTEX_HELD(&db
->db_mtx
));
898 ASSERT(buf
->b_data
!= NULL
);
899 db
->db
.db_data
= buf
->b_data
;
903 * Loan out an arc_buf for read. Return the loaned arc_buf.
906 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
910 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
911 mutex_enter(&db
->db_mtx
);
912 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
913 int blksz
= db
->db
.db_size
;
914 spa_t
*spa
= db
->db_objset
->os_spa
;
916 mutex_exit(&db
->db_mtx
);
917 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
918 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
921 arc_loan_inuse_buf(abuf
, db
);
924 mutex_exit(&db
->db_mtx
);
930 * Calculate which level n block references the data at the level 0 offset
934 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
936 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
938 * The level n blkid is equal to the level 0 blkid divided by
939 * the number of level 0s in a level n block.
941 * The level 0 blkid is offset >> datablkshift =
942 * offset / 2^datablkshift.
944 * The number of level 0s in a level n is the number of block
945 * pointers in an indirect block, raised to the power of level.
946 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
947 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
949 * Thus, the level n blkid is: offset /
950 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
951 * = offset / 2^(datablkshift + level *
952 * (indblkshift - SPA_BLKPTRSHIFT))
953 * = offset >> (datablkshift + level *
954 * (indblkshift - SPA_BLKPTRSHIFT))
957 const unsigned exp
= dn
->dn_datablkshift
+
958 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
960 if (exp
>= 8 * sizeof (offset
)) {
961 /* This only happens on the highest indirection level */
962 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
966 ASSERT3U(exp
, <, 8 * sizeof (offset
));
968 return (offset
>> exp
);
970 ASSERT3U(offset
, <, dn
->dn_datablksz
);
976 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
977 arc_buf_t
*buf
, void *vdb
)
979 dmu_buf_impl_t
*db
= vdb
;
981 mutex_enter(&db
->db_mtx
);
982 ASSERT3U(db
->db_state
, ==, DB_READ
);
984 * All reads are synchronous, so we must have a hold on the dbuf
986 ASSERT(refcount_count(&db
->db_holds
) > 0);
987 ASSERT(db
->db_buf
== NULL
);
988 ASSERT(db
->db
.db_data
== NULL
);
989 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
990 /* we were freed in flight; disregard any error */
992 buf
= arc_alloc_buf(db
->db_objset
->os_spa
,
993 db
, DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
);
995 arc_release(buf
, db
);
996 bzero(buf
->b_data
, db
->db
.db_size
);
998 db
->db_freed_in_flight
= FALSE
;
999 dbuf_set_data(db
, buf
);
1000 db
->db_state
= DB_CACHED
;
1001 } else if (buf
!= NULL
) {
1002 dbuf_set_data(db
, buf
);
1003 db
->db_state
= DB_CACHED
;
1005 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1006 ASSERT3P(db
->db_buf
, ==, NULL
);
1007 db
->db_state
= DB_UNCACHED
;
1009 cv_broadcast(&db
->db_changed
);
1010 dbuf_rele_and_unlock(db
, NULL
);
1014 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1017 zbookmark_phys_t zb
;
1018 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1019 int err
, zio_flags
= 0;
1023 ASSERT(!refcount_is_zero(&db
->db_holds
));
1024 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1025 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1026 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1027 ASSERT(db
->db_state
== DB_UNCACHED
);
1028 ASSERT(db
->db_buf
== NULL
);
1030 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1032 * The bonus length stored in the dnode may be less than
1033 * the maximum available space in the bonus buffer.
1035 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1036 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1037 arc_buf_t
*dn_buf
= (dn
->dn_dbuf
!= NULL
) ?
1038 dn
->dn_dbuf
->db_buf
: NULL
;
1040 /* if the underlying dnode block is encrypted, decrypt it */
1041 if (dn_buf
!= NULL
&& dn
->dn_objset
->os_encrypted
&&
1042 DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
) &&
1043 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1044 arc_is_encrypted(dn_buf
)) {
1045 err
= arc_untransform(dn_buf
, dn
->dn_objset
->os_spa
,
1046 dmu_objset_id(dn
->dn_objset
), B_TRUE
);
1049 mutex_exit(&db
->db_mtx
);
1054 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1055 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1056 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1057 if (bonuslen
< max_bonuslen
)
1058 bzero(db
->db
.db_data
, max_bonuslen
);
1060 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1062 db
->db_state
= DB_CACHED
;
1063 mutex_exit(&db
->db_mtx
);
1068 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1069 * processes the delete record and clears the bp while we are waiting
1070 * for the dn_mtx (resulting in a "no" from block_freed).
1072 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1073 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1074 BP_IS_HOLE(db
->db_blkptr
)))) {
1075 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1077 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1079 bzero(db
->db
.db_data
, db
->db
.db_size
);
1081 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1082 BP_IS_HOLE(db
->db_blkptr
) &&
1083 db
->db_blkptr
->blk_birth
!= 0) {
1084 blkptr_t
*bps
= db
->db
.db_data
;
1085 for (int i
= 0; i
< ((1 <<
1086 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1088 blkptr_t
*bp
= &bps
[i
];
1089 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1090 1 << dn
->dn_indblkshift
);
1092 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1094 BP_GET_LSIZE(db
->db_blkptr
));
1095 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1097 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1098 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1102 db
->db_state
= DB_CACHED
;
1103 mutex_exit(&db
->db_mtx
);
1109 db
->db_state
= DB_READ
;
1110 mutex_exit(&db
->db_mtx
);
1112 if (DBUF_IS_L2CACHEABLE(db
))
1113 aflags
|= ARC_FLAG_L2CACHE
;
1115 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
1116 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
1117 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1120 * All bps of an encrypted os should have the encryption bit set.
1121 * If this is not true it indicates tampering and we report an error.
1123 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(db
->db_blkptr
)) {
1124 spa_log_error(db
->db_objset
->os_spa
, &zb
);
1125 zfs_panic_recover("unencrypted block in encrypted "
1126 "object set %llu", dmu_objset_id(db
->db_objset
));
1127 return (SET_ERROR(EIO
));
1130 dbuf_add_ref(db
, NULL
);
1132 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1133 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1135 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1136 zio_flags
|= ZIO_FLAG_RAW
;
1138 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1139 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1146 * This is our just-in-time copy function. It makes a copy of buffers that
1147 * have been modified in a previous transaction group before we access them in
1148 * the current active group.
1150 * This function is used in three places: when we are dirtying a buffer for the
1151 * first time in a txg, when we are freeing a range in a dnode that includes
1152 * this buffer, and when we are accessing a buffer which was received compressed
1153 * and later referenced in a WRITE_BYREF record.
1155 * Note that when we are called from dbuf_free_range() we do not put a hold on
1156 * the buffer, we just traverse the active dbuf list for the dnode.
1159 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1161 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1163 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1164 ASSERT(db
->db
.db_data
!= NULL
);
1165 ASSERT(db
->db_level
== 0);
1166 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1169 (dr
->dt
.dl
.dr_data
!=
1170 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1174 * If the last dirty record for this dbuf has not yet synced
1175 * and its referencing the dbuf data, either:
1176 * reset the reference to point to a new copy,
1177 * or (if there a no active holders)
1178 * just null out the current db_data pointer.
1180 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1181 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1182 dnode_t
*dn
= DB_DNODE(db
);
1183 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1184 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1185 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1186 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1187 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1188 dnode_t
*dn
= DB_DNODE(db
);
1189 int size
= arc_buf_size(db
->db_buf
);
1190 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1191 spa_t
*spa
= db
->db_objset
->os_spa
;
1192 enum zio_compress compress_type
=
1193 arc_get_compression(db
->db_buf
);
1195 if (arc_is_encrypted(db
->db_buf
)) {
1196 boolean_t byteorder
;
1197 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1198 uint8_t iv
[ZIO_DATA_IV_LEN
];
1199 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1201 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1203 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1204 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1205 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1207 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1208 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1209 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1210 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1212 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1214 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1217 dbuf_clear_data(db
);
1222 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1229 * We don't have to hold the mutex to check db_state because it
1230 * can't be freed while we have a hold on the buffer.
1232 ASSERT(!refcount_is_zero(&db
->db_holds
));
1234 if (db
->db_state
== DB_NOFILL
)
1235 return (SET_ERROR(EIO
));
1239 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1240 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1242 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1243 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1244 DBUF_IS_CACHEABLE(db
);
1246 mutex_enter(&db
->db_mtx
);
1247 if (db
->db_state
== DB_CACHED
) {
1248 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1251 * If the arc buf is compressed or encrypted, we need to
1252 * untransform it to read the data. This could happen during
1253 * the "zfs receive" of a stream which is deduplicated and
1254 * either raw or compressed. We do not need to do this if the
1255 * caller wants raw encrypted data.
1257 if (db
->db_buf
!= NULL
&& (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1258 (arc_is_encrypted(db
->db_buf
) ||
1259 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1260 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1261 err
= arc_untransform(db
->db_buf
, spa
,
1262 dmu_objset_id(db
->db_objset
), B_FALSE
);
1263 dbuf_set_data(db
, db
->db_buf
);
1265 mutex_exit(&db
->db_mtx
);
1267 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1268 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1269 rw_exit(&dn
->dn_struct_rwlock
);
1271 } else if (db
->db_state
== DB_UNCACHED
) {
1272 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1273 boolean_t need_wait
= B_FALSE
;
1276 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1277 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1280 err
= dbuf_read_impl(db
, zio
, flags
);
1282 /* dbuf_read_impl has dropped db_mtx for us */
1284 if (!err
&& prefetch
)
1285 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1287 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1288 rw_exit(&dn
->dn_struct_rwlock
);
1291 if (!err
&& need_wait
)
1292 err
= zio_wait(zio
);
1295 * Another reader came in while the dbuf was in flight
1296 * between UNCACHED and CACHED. Either a writer will finish
1297 * writing the buffer (sending the dbuf to CACHED) or the
1298 * first reader's request will reach the read_done callback
1299 * and send the dbuf to CACHED. Otherwise, a failure
1300 * occurred and the dbuf went to UNCACHED.
1302 mutex_exit(&db
->db_mtx
);
1304 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1305 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1306 rw_exit(&dn
->dn_struct_rwlock
);
1309 /* Skip the wait per the caller's request. */
1310 mutex_enter(&db
->db_mtx
);
1311 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1312 while (db
->db_state
== DB_READ
||
1313 db
->db_state
== DB_FILL
) {
1314 ASSERT(db
->db_state
== DB_READ
||
1315 (flags
& DB_RF_HAVESTRUCT
) == 0);
1316 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1318 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1320 if (db
->db_state
== DB_UNCACHED
)
1321 err
= SET_ERROR(EIO
);
1323 mutex_exit(&db
->db_mtx
);
1330 dbuf_noread(dmu_buf_impl_t
*db
)
1332 ASSERT(!refcount_is_zero(&db
->db_holds
));
1333 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1334 mutex_enter(&db
->db_mtx
);
1335 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1336 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1337 if (db
->db_state
== DB_UNCACHED
) {
1338 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1339 spa_t
*spa
= db
->db_objset
->os_spa
;
1341 ASSERT(db
->db_buf
== NULL
);
1342 ASSERT(db
->db
.db_data
== NULL
);
1343 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1344 db
->db_state
= DB_FILL
;
1345 } else if (db
->db_state
== DB_NOFILL
) {
1346 dbuf_clear_data(db
);
1348 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1350 mutex_exit(&db
->db_mtx
);
1354 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1356 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1357 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1358 uint64_t txg
= dr
->dr_txg
;
1360 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1362 * This assert is valid because dmu_sync() expects to be called by
1363 * a zilog's get_data while holding a range lock. This call only
1364 * comes from dbuf_dirty() callers who must also hold a range lock.
1366 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1367 ASSERT(db
->db_level
== 0);
1369 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1370 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1373 ASSERT(db
->db_data_pending
!= dr
);
1375 /* free this block */
1376 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1377 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1379 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1380 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1381 dr
->dt
.dl
.dr_raw
= B_FALSE
;
1384 * Release the already-written buffer, so we leave it in
1385 * a consistent dirty state. Note that all callers are
1386 * modifying the buffer, so they will immediately do
1387 * another (redundant) arc_release(). Therefore, leave
1388 * the buf thawed to save the effort of freezing &
1389 * immediately re-thawing it.
1391 arc_release(dr
->dt
.dl
.dr_data
, db
);
1395 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1396 * data blocks in the free range, so that any future readers will find
1400 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1403 dmu_buf_impl_t
*db_search
;
1404 dmu_buf_impl_t
*db
, *db_next
;
1405 uint64_t txg
= tx
->tx_txg
;
1408 if (end_blkid
> dn
->dn_maxblkid
&&
1409 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1410 end_blkid
= dn
->dn_maxblkid
;
1411 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1413 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1414 db_search
->db_level
= 0;
1415 db_search
->db_blkid
= start_blkid
;
1416 db_search
->db_state
= DB_SEARCH
;
1418 mutex_enter(&dn
->dn_dbufs_mtx
);
1419 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1420 ASSERT3P(db
, ==, NULL
);
1422 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1424 for (; db
!= NULL
; db
= db_next
) {
1425 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1426 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1428 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1431 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1433 /* found a level 0 buffer in the range */
1434 mutex_enter(&db
->db_mtx
);
1435 if (dbuf_undirty(db
, tx
)) {
1436 /* mutex has been dropped and dbuf destroyed */
1440 if (db
->db_state
== DB_UNCACHED
||
1441 db
->db_state
== DB_NOFILL
||
1442 db
->db_state
== DB_EVICTING
) {
1443 ASSERT(db
->db
.db_data
== NULL
);
1444 mutex_exit(&db
->db_mtx
);
1447 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1448 /* will be handled in dbuf_read_done or dbuf_rele */
1449 db
->db_freed_in_flight
= TRUE
;
1450 mutex_exit(&db
->db_mtx
);
1453 if (refcount_count(&db
->db_holds
) == 0) {
1458 /* The dbuf is referenced */
1460 if (db
->db_last_dirty
!= NULL
) {
1461 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1463 if (dr
->dr_txg
== txg
) {
1465 * This buffer is "in-use", re-adjust the file
1466 * size to reflect that this buffer may
1467 * contain new data when we sync.
1469 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1470 db
->db_blkid
> dn
->dn_maxblkid
)
1471 dn
->dn_maxblkid
= db
->db_blkid
;
1472 dbuf_unoverride(dr
);
1475 * This dbuf is not dirty in the open context.
1476 * Either uncache it (if its not referenced in
1477 * the open context) or reset its contents to
1480 dbuf_fix_old_data(db
, txg
);
1483 /* clear the contents if its cached */
1484 if (db
->db_state
== DB_CACHED
) {
1485 ASSERT(db
->db
.db_data
!= NULL
);
1486 arc_release(db
->db_buf
, db
);
1487 bzero(db
->db
.db_data
, db
->db
.db_size
);
1488 arc_buf_freeze(db
->db_buf
);
1491 mutex_exit(&db
->db_mtx
);
1494 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1495 mutex_exit(&dn
->dn_dbufs_mtx
);
1499 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1501 arc_buf_t
*buf
, *obuf
;
1502 int osize
= db
->db
.db_size
;
1503 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1506 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1511 /* XXX does *this* func really need the lock? */
1512 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1515 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1516 * is OK, because there can be no other references to the db
1517 * when we are changing its size, so no concurrent DB_FILL can
1521 * XXX we should be doing a dbuf_read, checking the return
1522 * value and returning that up to our callers
1524 dmu_buf_will_dirty(&db
->db
, tx
);
1526 /* create the data buffer for the new block */
1527 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1529 /* copy old block data to the new block */
1531 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1532 /* zero the remainder */
1534 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1536 mutex_enter(&db
->db_mtx
);
1537 dbuf_set_data(db
, buf
);
1538 arc_buf_destroy(obuf
, db
);
1539 db
->db
.db_size
= size
;
1541 if (db
->db_level
== 0) {
1542 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1543 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1545 mutex_exit(&db
->db_mtx
);
1547 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1552 dbuf_release_bp(dmu_buf_impl_t
*db
)
1554 ASSERTV(objset_t
*os
= db
->db_objset
);
1556 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1557 ASSERT(arc_released(os
->os_phys_buf
) ||
1558 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1559 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1561 (void) arc_release(db
->db_buf
, db
);
1565 * We already have a dirty record for this TXG, and we are being
1569 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1571 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1573 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1575 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1577 * If this buffer has already been written out,
1578 * we now need to reset its state.
1580 dbuf_unoverride(dr
);
1581 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1582 db
->db_state
!= DB_NOFILL
) {
1583 /* Already released on initial dirty, so just thaw. */
1584 ASSERT(arc_released(db
->db_buf
));
1585 arc_buf_thaw(db
->db_buf
);
1590 dbuf_dirty_record_t
*
1591 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1595 dbuf_dirty_record_t
**drp
, *dr
;
1596 int drop_struct_lock
= FALSE
;
1597 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1599 ASSERT(tx
->tx_txg
!= 0);
1600 ASSERT(!refcount_is_zero(&db
->db_holds
));
1601 DMU_TX_DIRTY_BUF(tx
, db
);
1606 * Shouldn't dirty a regular buffer in syncing context. Private
1607 * objects may be dirtied in syncing context, but only if they
1608 * were already pre-dirtied in open context.
1611 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1612 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1615 ASSERT(!dmu_tx_is_syncing(tx
) ||
1616 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1617 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1618 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1619 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1620 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1623 * We make this assert for private objects as well, but after we
1624 * check if we're already dirty. They are allowed to re-dirty
1625 * in syncing context.
1627 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1628 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1629 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1631 mutex_enter(&db
->db_mtx
);
1633 * XXX make this true for indirects too? The problem is that
1634 * transactions created with dmu_tx_create_assigned() from
1635 * syncing context don't bother holding ahead.
1637 ASSERT(db
->db_level
!= 0 ||
1638 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1639 db
->db_state
== DB_NOFILL
);
1641 mutex_enter(&dn
->dn_mtx
);
1643 * Don't set dirtyctx to SYNC if we're just modifying this as we
1644 * initialize the objset.
1646 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1647 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1648 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1651 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1652 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1653 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1654 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1655 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1657 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1658 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1662 mutex_exit(&dn
->dn_mtx
);
1664 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1665 dn
->dn_have_spill
= B_TRUE
;
1668 * If this buffer is already dirty, we're done.
1670 drp
= &db
->db_last_dirty
;
1671 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1672 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1673 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1675 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1679 mutex_exit(&db
->db_mtx
);
1684 * Only valid if not already dirty.
1686 ASSERT(dn
->dn_object
== 0 ||
1687 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1688 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1690 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1693 * We should only be dirtying in syncing context if it's the
1694 * mos or we're initializing the os or it's a special object.
1695 * However, we are allowed to dirty in syncing context provided
1696 * we already dirtied it in open context. Hence we must make
1697 * this assertion only if we're not already dirty.
1700 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
1702 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1703 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
1704 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1705 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1706 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1707 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1709 ASSERT(db
->db
.db_size
!= 0);
1711 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1713 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1714 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
1718 * If this buffer is dirty in an old transaction group we need
1719 * to make a copy of it so that the changes we make in this
1720 * transaction group won't leak out when we sync the older txg.
1722 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1723 list_link_init(&dr
->dr_dirty_node
);
1724 if (db
->db_level
== 0) {
1725 void *data_old
= db
->db_buf
;
1727 if (db
->db_state
!= DB_NOFILL
) {
1728 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1729 dbuf_fix_old_data(db
, tx
->tx_txg
);
1730 data_old
= db
->db
.db_data
;
1731 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1733 * Release the data buffer from the cache so
1734 * that we can modify it without impacting
1735 * possible other users of this cached data
1736 * block. Note that indirect blocks and
1737 * private objects are not released until the
1738 * syncing state (since they are only modified
1741 arc_release(db
->db_buf
, db
);
1742 dbuf_fix_old_data(db
, tx
->tx_txg
);
1743 data_old
= db
->db_buf
;
1745 ASSERT(data_old
!= NULL
);
1747 dr
->dt
.dl
.dr_data
= data_old
;
1749 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1750 list_create(&dr
->dt
.di
.dr_children
,
1751 sizeof (dbuf_dirty_record_t
),
1752 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1754 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1755 dr
->dr_accounted
= db
->db
.db_size
;
1757 dr
->dr_txg
= tx
->tx_txg
;
1762 * We could have been freed_in_flight between the dbuf_noread
1763 * and dbuf_dirty. We win, as though the dbuf_noread() had
1764 * happened after the free.
1766 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1767 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1768 mutex_enter(&dn
->dn_mtx
);
1769 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1770 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1773 mutex_exit(&dn
->dn_mtx
);
1774 db
->db_freed_in_flight
= FALSE
;
1778 * This buffer is now part of this txg
1780 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1781 db
->db_dirtycnt
+= 1;
1782 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1784 mutex_exit(&db
->db_mtx
);
1786 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1787 db
->db_blkid
== DMU_SPILL_BLKID
) {
1788 mutex_enter(&dn
->dn_mtx
);
1789 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1790 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1791 mutex_exit(&dn
->dn_mtx
);
1792 dnode_setdirty(dn
, tx
);
1798 * The dn_struct_rwlock prevents db_blkptr from changing
1799 * due to a write from syncing context completing
1800 * while we are running, so we want to acquire it before
1801 * looking at db_blkptr.
1803 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1804 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1805 drop_struct_lock
= TRUE
;
1809 * We need to hold the dn_struct_rwlock to make this assertion,
1810 * because it protects dn_phys / dn_next_nlevels from changing.
1812 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1813 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1814 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1815 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1816 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1819 * If we are overwriting a dedup BP, then unless it is snapshotted,
1820 * when we get to syncing context we will need to decrement its
1821 * refcount in the DDT. Prefetch the relevant DDT block so that
1822 * syncing context won't have to wait for the i/o.
1824 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
1826 if (db
->db_level
== 0) {
1827 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1828 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1831 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1832 dmu_buf_impl_t
*parent
= db
->db_parent
;
1833 dbuf_dirty_record_t
*di
;
1834 int parent_held
= FALSE
;
1836 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1837 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1839 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1840 db
->db_blkid
>> epbs
, FTAG
);
1841 ASSERT(parent
!= NULL
);
1844 if (drop_struct_lock
)
1845 rw_exit(&dn
->dn_struct_rwlock
);
1846 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1847 di
= dbuf_dirty(parent
, tx
);
1849 dbuf_rele(parent
, FTAG
);
1851 mutex_enter(&db
->db_mtx
);
1853 * Since we've dropped the mutex, it's possible that
1854 * dbuf_undirty() might have changed this out from under us.
1856 if (db
->db_last_dirty
== dr
||
1857 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1858 mutex_enter(&di
->dt
.di
.dr_mtx
);
1859 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1860 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1861 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1862 mutex_exit(&di
->dt
.di
.dr_mtx
);
1865 mutex_exit(&db
->db_mtx
);
1867 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1868 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1869 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1870 mutex_enter(&dn
->dn_mtx
);
1871 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1872 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1873 mutex_exit(&dn
->dn_mtx
);
1874 if (drop_struct_lock
)
1875 rw_exit(&dn
->dn_struct_rwlock
);
1878 dnode_setdirty(dn
, tx
);
1884 * Undirty a buffer in the transaction group referenced by the given
1885 * transaction. Return whether this evicted the dbuf.
1888 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1891 uint64_t txg
= tx
->tx_txg
;
1892 dbuf_dirty_record_t
*dr
, **drp
;
1897 * Due to our use of dn_nlevels below, this can only be called
1898 * in open context, unless we are operating on the MOS.
1899 * From syncing context, dn_nlevels may be different from the
1900 * dn_nlevels used when dbuf was dirtied.
1902 ASSERT(db
->db_objset
==
1903 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1904 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1905 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1906 ASSERT0(db
->db_level
);
1907 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1910 * If this buffer is not dirty, we're done.
1912 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1913 if (dr
->dr_txg
<= txg
)
1915 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1917 ASSERT(dr
->dr_txg
== txg
);
1918 ASSERT(dr
->dr_dbuf
== db
);
1923 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1925 ASSERT(db
->db
.db_size
!= 0);
1927 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1928 dr
->dr_accounted
, txg
);
1933 * Note that there are three places in dbuf_dirty()
1934 * where this dirty record may be put on a list.
1935 * Make sure to do a list_remove corresponding to
1936 * every one of those list_insert calls.
1938 if (dr
->dr_parent
) {
1939 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1940 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1941 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1942 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1943 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1944 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1945 mutex_enter(&dn
->dn_mtx
);
1946 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1947 mutex_exit(&dn
->dn_mtx
);
1951 if (db
->db_state
!= DB_NOFILL
) {
1952 dbuf_unoverride(dr
);
1954 ASSERT(db
->db_buf
!= NULL
);
1955 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1956 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1957 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
1960 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1962 ASSERT(db
->db_dirtycnt
> 0);
1963 db
->db_dirtycnt
-= 1;
1965 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1966 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
1975 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
1977 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1979 ASSERT(tx
->tx_txg
!= 0);
1980 ASSERT(!refcount_is_zero(&db
->db_holds
));
1983 * Quick check for dirtyness. For already dirty blocks, this
1984 * reduces runtime of this function by >90%, and overall performance
1985 * by 50% for some workloads (e.g. file deletion with indirect blocks
1988 mutex_enter(&db
->db_mtx
);
1990 dbuf_dirty_record_t
*dr
;
1991 for (dr
= db
->db_last_dirty
;
1992 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1994 * It's possible that it is already dirty but not cached,
1995 * because there are some calls to dbuf_dirty() that don't
1996 * go through dmu_buf_will_dirty().
1998 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1999 /* This dbuf is already dirty and cached. */
2001 mutex_exit(&db
->db_mtx
);
2005 mutex_exit(&db
->db_mtx
);
2008 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2009 flags
|= DB_RF_HAVESTRUCT
;
2011 (void) dbuf_read(db
, NULL
, flags
);
2012 (void) dbuf_dirty(db
, tx
);
2016 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2018 dmu_buf_will_dirty_impl(db_fake
,
2019 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2023 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2025 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2027 db
->db_state
= DB_NOFILL
;
2029 dmu_buf_will_fill(db_fake
, tx
);
2033 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2035 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2037 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2038 ASSERT(tx
->tx_txg
!= 0);
2039 ASSERT(db
->db_level
== 0);
2040 ASSERT(!refcount_is_zero(&db
->db_holds
));
2042 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2043 dmu_tx_private_ok(tx
));
2046 (void) dbuf_dirty(db
, tx
);
2050 * This function is effectively the same as dmu_buf_will_dirty(), but
2051 * indicates the caller expects raw encrypted data in the db. It will
2052 * also set the raw flag on the created dirty record.
2055 dmu_buf_will_change_crypt_params(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2057 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2058 dbuf_dirty_record_t
*dr
;
2060 dmu_buf_will_dirty_impl(db_fake
,
2061 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2063 dr
= db
->db_last_dirty
;
2064 while (dr
!= NULL
&& dr
->dr_txg
> tx
->tx_txg
)
2067 ASSERT3P(dr
, !=, NULL
);
2068 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2069 dr
->dt
.dl
.dr_raw
= B_TRUE
;
2072 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2075 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2077 mutex_enter(&db
->db_mtx
);
2080 if (db
->db_state
== DB_FILL
) {
2081 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2082 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2083 /* we were freed while filling */
2084 /* XXX dbuf_undirty? */
2085 bzero(db
->db
.db_data
, db
->db
.db_size
);
2086 db
->db_freed_in_flight
= FALSE
;
2088 db
->db_state
= DB_CACHED
;
2089 cv_broadcast(&db
->db_changed
);
2091 mutex_exit(&db
->db_mtx
);
2095 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2096 bp_embedded_type_t etype
, enum zio_compress comp
,
2097 int uncompressed_size
, int compressed_size
, int byteorder
,
2100 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2101 struct dirty_leaf
*dl
;
2102 dmu_object_type_t type
;
2104 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2105 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2106 SPA_FEATURE_EMBEDDED_DATA
));
2110 type
= DB_DNODE(db
)->dn_type
;
2113 ASSERT0(db
->db_level
);
2114 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2116 dmu_buf_will_not_fill(dbuf
, tx
);
2118 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2119 dl
= &db
->db_last_dirty
->dt
.dl
;
2120 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2121 data
, comp
, uncompressed_size
, compressed_size
);
2122 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2123 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2124 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2125 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2127 dl
->dr_override_state
= DR_OVERRIDDEN
;
2128 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2132 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2133 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2136 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2138 ASSERT(!refcount_is_zero(&db
->db_holds
));
2139 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2140 ASSERT(db
->db_level
== 0);
2141 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2142 ASSERT(buf
!= NULL
);
2143 ASSERT(arc_buf_lsize(buf
) == db
->db
.db_size
);
2144 ASSERT(tx
->tx_txg
!= 0);
2146 arc_return_buf(buf
, db
);
2147 ASSERT(arc_released(buf
));
2149 mutex_enter(&db
->db_mtx
);
2151 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2152 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2154 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2156 if (db
->db_state
== DB_CACHED
&&
2157 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2159 * In practice, we will never have a case where we have an
2160 * encrypted arc buffer while additional holds exist on the
2161 * dbuf. We don't handle this here so we simply assert that
2164 ASSERT(!arc_is_encrypted(buf
));
2165 mutex_exit(&db
->db_mtx
);
2166 (void) dbuf_dirty(db
, tx
);
2167 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2168 arc_buf_destroy(buf
, db
);
2169 xuio_stat_wbuf_copied();
2173 xuio_stat_wbuf_nocopy();
2174 if (db
->db_state
== DB_CACHED
) {
2175 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2177 ASSERT(db
->db_buf
!= NULL
);
2178 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2179 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2180 IMPLY(arc_is_encrypted(buf
), dr
->dt
.dl
.dr_raw
);
2182 if (!arc_released(db
->db_buf
)) {
2183 ASSERT(dr
->dt
.dl
.dr_override_state
==
2185 arc_release(db
->db_buf
, db
);
2187 dr
->dt
.dl
.dr_data
= buf
;
2188 arc_buf_destroy(db
->db_buf
, db
);
2189 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2190 arc_release(db
->db_buf
, db
);
2191 arc_buf_destroy(db
->db_buf
, db
);
2195 ASSERT(db
->db_buf
== NULL
);
2196 dbuf_set_data(db
, buf
);
2197 db
->db_state
= DB_FILL
;
2198 mutex_exit(&db
->db_mtx
);
2199 (void) dbuf_dirty(db
, tx
);
2200 dmu_buf_fill_done(&db
->db
, tx
);
2204 dbuf_destroy(dmu_buf_impl_t
*db
)
2207 dmu_buf_impl_t
*parent
= db
->db_parent
;
2208 dmu_buf_impl_t
*dndb
;
2210 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2211 ASSERT(refcount_is_zero(&db
->db_holds
));
2213 if (db
->db_buf
!= NULL
) {
2214 arc_buf_destroy(db
->db_buf
, db
);
2218 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2219 int slots
= DB_DNODE(db
)->dn_num_slots
;
2220 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2221 if (db
->db
.db_data
!= NULL
) {
2222 kmem_free(db
->db
.db_data
, bonuslen
);
2223 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2224 db
->db_state
= DB_UNCACHED
;
2228 dbuf_clear_data(db
);
2230 if (multilist_link_active(&db
->db_cache_link
)) {
2231 multilist_remove(dbuf_cache
, db
);
2232 (void) refcount_remove_many(&dbuf_cache_size
,
2233 db
->db
.db_size
, db
);
2236 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2237 ASSERT(db
->db_data_pending
== NULL
);
2239 db
->db_state
= DB_EVICTING
;
2240 db
->db_blkptr
= NULL
;
2243 * Now that db_state is DB_EVICTING, nobody else can find this via
2244 * the hash table. We can now drop db_mtx, which allows us to
2245 * acquire the dn_dbufs_mtx.
2247 mutex_exit(&db
->db_mtx
);
2252 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2253 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2255 mutex_enter(&dn
->dn_dbufs_mtx
);
2256 avl_remove(&dn
->dn_dbufs
, db
);
2257 atomic_dec_32(&dn
->dn_dbufs_count
);
2261 mutex_exit(&dn
->dn_dbufs_mtx
);
2263 * Decrementing the dbuf count means that the hold corresponding
2264 * to the removed dbuf is no longer discounted in dnode_move(),
2265 * so the dnode cannot be moved until after we release the hold.
2266 * The membar_producer() ensures visibility of the decremented
2267 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2271 db
->db_dnode_handle
= NULL
;
2273 dbuf_hash_remove(db
);
2278 ASSERT(refcount_is_zero(&db
->db_holds
));
2280 db
->db_parent
= NULL
;
2282 ASSERT(db
->db_buf
== NULL
);
2283 ASSERT(db
->db
.db_data
== NULL
);
2284 ASSERT(db
->db_hash_next
== NULL
);
2285 ASSERT(db
->db_blkptr
== NULL
);
2286 ASSERT(db
->db_data_pending
== NULL
);
2287 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2289 kmem_cache_free(dbuf_kmem_cache
, db
);
2290 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2293 * If this dbuf is referenced from an indirect dbuf,
2294 * decrement the ref count on the indirect dbuf.
2296 if (parent
&& parent
!= dndb
)
2297 dbuf_rele(parent
, db
);
2301 * Note: While bpp will always be updated if the function returns success,
2302 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2303 * this happens when the dnode is the meta-dnode, or a userused or groupused
2306 __attribute__((always_inline
))
2308 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2309 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
2314 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2316 if (blkid
== DMU_SPILL_BLKID
) {
2317 mutex_enter(&dn
->dn_mtx
);
2318 if (dn
->dn_have_spill
&&
2319 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2320 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2323 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2324 *parentp
= dn
->dn_dbuf
;
2325 mutex_exit(&dn
->dn_mtx
);
2330 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2331 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2333 ASSERT3U(level
* epbs
, <, 64);
2334 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2336 * This assertion shouldn't trip as long as the max indirect block size
2337 * is less than 1M. The reason for this is that up to that point,
2338 * the number of levels required to address an entire object with blocks
2339 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2340 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2341 * (i.e. we can address the entire object), objects will all use at most
2342 * N-1 levels and the assertion won't overflow. However, once epbs is
2343 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2344 * enough to address an entire object, so objects will have 5 levels,
2345 * but then this assertion will overflow.
2347 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2348 * need to redo this logic to handle overflows.
2350 ASSERT(level
>= nlevels
||
2351 ((nlevels
- level
- 1) * epbs
) +
2352 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2353 if (level
>= nlevels
||
2354 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2355 ((nlevels
- level
- 1) * epbs
)) ||
2357 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2358 /* the buffer has no parent yet */
2359 return (SET_ERROR(ENOENT
));
2360 } else if (level
< nlevels
-1) {
2361 /* this block is referenced from an indirect block */
2364 err
= dbuf_hold_impl(dn
, level
+1,
2365 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2367 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
2368 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
2369 parentp
, dh
->dh_depth
+ 1);
2370 err
= __dbuf_hold_impl(dh
+ 1);
2374 err
= dbuf_read(*parentp
, NULL
,
2375 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2377 dbuf_rele(*parentp
, NULL
);
2381 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2382 (blkid
& ((1ULL << epbs
) - 1));
2383 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2384 ASSERT(BP_IS_HOLE(*bpp
));
2387 /* the block is referenced from the dnode */
2388 ASSERT3U(level
, ==, nlevels
-1);
2389 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2390 blkid
< dn
->dn_phys
->dn_nblkptr
);
2392 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2393 *parentp
= dn
->dn_dbuf
;
2395 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2400 static dmu_buf_impl_t
*
2401 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2402 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2404 objset_t
*os
= dn
->dn_objset
;
2405 dmu_buf_impl_t
*db
, *odb
;
2407 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2408 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2410 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2413 db
->db
.db_object
= dn
->dn_object
;
2414 db
->db_level
= level
;
2415 db
->db_blkid
= blkid
;
2416 db
->db_last_dirty
= NULL
;
2417 db
->db_dirtycnt
= 0;
2418 db
->db_dnode_handle
= dn
->dn_handle
;
2419 db
->db_parent
= parent
;
2420 db
->db_blkptr
= blkptr
;
2423 db
->db_user_immediate_evict
= FALSE
;
2424 db
->db_freed_in_flight
= FALSE
;
2425 db
->db_pending_evict
= FALSE
;
2427 if (blkid
== DMU_BONUS_BLKID
) {
2428 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2429 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2430 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2431 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2432 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2433 db
->db_state
= DB_UNCACHED
;
2434 /* the bonus dbuf is not placed in the hash table */
2435 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2437 } else if (blkid
== DMU_SPILL_BLKID
) {
2438 db
->db
.db_size
= (blkptr
!= NULL
) ?
2439 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2440 db
->db
.db_offset
= 0;
2443 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2444 db
->db
.db_size
= blocksize
;
2445 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2449 * Hold the dn_dbufs_mtx while we get the new dbuf
2450 * in the hash table *and* added to the dbufs list.
2451 * This prevents a possible deadlock with someone
2452 * trying to look up this dbuf before its added to the
2455 mutex_enter(&dn
->dn_dbufs_mtx
);
2456 db
->db_state
= DB_EVICTING
;
2457 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2458 /* someone else inserted it first */
2459 kmem_cache_free(dbuf_kmem_cache
, db
);
2460 mutex_exit(&dn
->dn_dbufs_mtx
);
2463 avl_add(&dn
->dn_dbufs
, db
);
2465 db
->db_state
= DB_UNCACHED
;
2466 mutex_exit(&dn
->dn_dbufs_mtx
);
2467 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2469 if (parent
&& parent
!= dn
->dn_dbuf
)
2470 dbuf_add_ref(parent
, db
);
2472 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2473 refcount_count(&dn
->dn_holds
) > 0);
2474 (void) refcount_add(&dn
->dn_holds
, db
);
2475 atomic_inc_32(&dn
->dn_dbufs_count
);
2477 dprintf_dbuf(db
, "db=%p\n", db
);
2482 typedef struct dbuf_prefetch_arg
{
2483 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2484 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2485 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2486 int dpa_curlevel
; /* The current level that we're reading */
2487 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2488 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2489 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2490 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2491 } dbuf_prefetch_arg_t
;
2494 * Actually issue the prefetch read for the block given.
2497 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2499 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2502 arc_flags_t aflags
=
2503 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2505 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2506 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2507 ASSERT(dpa
->dpa_zio
!= NULL
);
2508 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2509 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2510 &aflags
, &dpa
->dpa_zb
);
2514 * Called when an indirect block above our prefetch target is read in. This
2515 * will either read in the next indirect block down the tree or issue the actual
2516 * prefetch if the next block down is our target.
2519 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
2520 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
2522 dbuf_prefetch_arg_t
*dpa
= private;
2524 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2525 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2528 * The dpa_dnode is only valid if we are called with a NULL
2529 * zio. This indicates that the arc_read() returned without
2530 * first calling zio_read() to issue a physical read. Once
2531 * a physical read is made the dpa_dnode must be invalidated
2532 * as the locks guarding it may have been dropped. If the
2533 * dpa_dnode is still valid, then we want to add it to the dbuf
2534 * cache. To do so, we must hold the dbuf associated with the block
2535 * we just prefetched, read its contents so that we associate it
2536 * with an arc_buf_t, and then release it.
2539 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2540 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
2541 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2543 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2545 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2547 dpa
->dpa_dnode
= NULL
;
2548 } else if (dpa
->dpa_dnode
!= NULL
) {
2549 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2550 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2551 dpa
->dpa_zb
.zb_level
));
2552 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2553 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2554 (void) dbuf_read(db
, NULL
,
2555 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2556 dbuf_rele(db
, FTAG
);
2560 kmem_free(dpa
, sizeof (*dpa
));
2564 dpa
->dpa_curlevel
--;
2565 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2566 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2567 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
2568 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2570 if (BP_IS_HOLE(bp
)) {
2571 kmem_free(dpa
, sizeof (*dpa
));
2572 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2573 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2574 dbuf_issue_final_prefetch(dpa
, bp
);
2575 kmem_free(dpa
, sizeof (*dpa
));
2577 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2578 zbookmark_phys_t zb
;
2580 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2581 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
2582 iter_aflags
|= ARC_FLAG_L2CACHE
;
2584 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2586 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2587 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2589 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2590 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2591 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2595 arc_buf_destroy(abuf
, private);
2599 * Issue prefetch reads for the given block on the given level. If the indirect
2600 * blocks above that block are not in memory, we will read them in
2601 * asynchronously. As a result, this call never blocks waiting for a read to
2602 * complete. Note that the prefetch might fail if the dataset is encrypted and
2603 * the encryption key is unmapped before the IO completes.
2606 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2610 int epbs
, nlevels
, curlevel
;
2613 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2614 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2616 if (blkid
> dn
->dn_maxblkid
)
2619 if (dnode_block_freed(dn
, blkid
))
2623 * This dnode hasn't been written to disk yet, so there's nothing to
2626 nlevels
= dn
->dn_phys
->dn_nlevels
;
2627 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2630 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2631 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2634 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2637 mutex_exit(&db
->db_mtx
);
2639 * This dbuf already exists. It is either CACHED, or
2640 * (we assume) about to be read or filled.
2646 * Find the closest ancestor (indirect block) of the target block
2647 * that is present in the cache. In this indirect block, we will
2648 * find the bp that is at curlevel, curblkid.
2652 while (curlevel
< nlevels
- 1) {
2653 int parent_level
= curlevel
+ 1;
2654 uint64_t parent_blkid
= curblkid
>> epbs
;
2657 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2658 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2659 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2660 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2661 dbuf_rele(db
, FTAG
);
2665 curlevel
= parent_level
;
2666 curblkid
= parent_blkid
;
2669 if (curlevel
== nlevels
- 1) {
2670 /* No cached indirect blocks found. */
2671 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2672 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2674 if (BP_IS_HOLE(&bp
))
2677 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2679 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2682 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2683 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
2684 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2685 dn
->dn_object
, level
, blkid
);
2686 dpa
->dpa_curlevel
= curlevel
;
2687 dpa
->dpa_prio
= prio
;
2688 dpa
->dpa_aflags
= aflags
;
2689 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2690 dpa
->dpa_dnode
= dn
;
2691 dpa
->dpa_epbs
= epbs
;
2694 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2695 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
2696 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
2699 * If we have the indirect just above us, no need to do the asynchronous
2700 * prefetch chain; we'll just run the last step ourselves. If we're at
2701 * a higher level, though, we want to issue the prefetches for all the
2702 * indirect blocks asynchronously, so we can go on with whatever we were
2705 if (curlevel
== level
) {
2706 ASSERT3U(curblkid
, ==, blkid
);
2707 dbuf_issue_final_prefetch(dpa
, &bp
);
2708 kmem_free(dpa
, sizeof (*dpa
));
2710 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2711 zbookmark_phys_t zb
;
2713 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2714 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
2715 iter_aflags
|= ARC_FLAG_L2CACHE
;
2717 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2718 dn
->dn_object
, curlevel
, curblkid
);
2719 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2720 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2721 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2725 * We use pio here instead of dpa_zio since it's possible that
2726 * dpa may have already been freed.
2731 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2734 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2735 * the case of encrypted, compressed and uncompressed buffers by
2736 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2737 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2739 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2741 noinline
static void
2742 dbuf_hold_copy(struct dbuf_hold_impl_data
*dh
)
2744 dnode_t
*dn
= dh
->dh_dn
;
2745 dmu_buf_impl_t
*db
= dh
->dh_db
;
2746 dbuf_dirty_record_t
*dr
= dh
->dh_dr
;
2747 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
2749 enum zio_compress compress_type
= arc_get_compression(data
);
2751 if (arc_is_encrypted(data
)) {
2752 boolean_t byteorder
;
2753 uint8_t salt
[ZIO_DATA_SALT_LEN
];
2754 uint8_t iv
[ZIO_DATA_IV_LEN
];
2755 uint8_t mac
[ZIO_DATA_MAC_LEN
];
2757 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
2758 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
2759 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
2760 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
2762 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
2763 dbuf_set_data(db
, arc_alloc_compressed_buf(
2764 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
2765 arc_buf_lsize(data
), compress_type
));
2767 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
2768 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
2771 bcopy(data
->b_data
, db
->db
.db_data
, arc_buf_size(data
));
2775 * Returns with db_holds incremented, and db_mtx not held.
2776 * Note: dn_struct_rwlock must be held.
2779 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2781 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2782 dh
->dh_parent
= NULL
;
2784 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2785 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2786 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2788 *(dh
->dh_dbp
) = NULL
;
2790 /* dbuf_find() returns with db_mtx held */
2791 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2792 dh
->dh_level
, dh
->dh_blkid
);
2794 if (dh
->dh_db
== NULL
) {
2797 if (dh
->dh_fail_uncached
)
2798 return (SET_ERROR(ENOENT
));
2800 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2801 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2802 dh
->dh_fail_sparse
, &dh
->dh_parent
, &dh
->dh_bp
, dh
);
2803 if (dh
->dh_fail_sparse
) {
2804 if (dh
->dh_err
== 0 &&
2805 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2806 dh
->dh_err
= SET_ERROR(ENOENT
);
2809 dbuf_rele(dh
->dh_parent
, NULL
);
2810 return (dh
->dh_err
);
2813 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2814 return (dh
->dh_err
);
2815 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2816 dh
->dh_parent
, dh
->dh_bp
);
2819 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2820 mutex_exit(&dh
->dh_db
->db_mtx
);
2821 return (SET_ERROR(ENOENT
));
2824 if (dh
->dh_db
->db_buf
!= NULL
)
2825 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2827 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2830 * If this buffer is currently syncing out, and we are are
2831 * still referencing it from db_data, we need to make a copy
2832 * of it in case we decide we want to dirty it again in this txg.
2834 if (dh
->dh_db
->db_level
== 0 &&
2835 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2836 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2837 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2838 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2839 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
)
2843 if (multilist_link_active(&dh
->dh_db
->db_cache_link
)) {
2844 ASSERT(refcount_is_zero(&dh
->dh_db
->db_holds
));
2845 multilist_remove(dbuf_cache
, dh
->dh_db
);
2846 (void) refcount_remove_many(&dbuf_cache_size
,
2847 dh
->dh_db
->db
.db_size
, dh
->dh_db
);
2849 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2850 DBUF_VERIFY(dh
->dh_db
);
2851 mutex_exit(&dh
->dh_db
->db_mtx
);
2853 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2855 dbuf_rele(dh
->dh_parent
, NULL
);
2857 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2858 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2859 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2860 *(dh
->dh_dbp
) = dh
->dh_db
;
2866 * The following code preserves the recursive function dbuf_hold_impl()
2867 * but moves the local variables AND function arguments to the heap to
2868 * minimize the stack frame size. Enough space is initially allocated
2869 * on the stack for 20 levels of recursion.
2872 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2873 boolean_t fail_sparse
, boolean_t fail_uncached
,
2874 void *tag
, dmu_buf_impl_t
**dbp
)
2876 struct dbuf_hold_impl_data
*dh
;
2879 dh
= kmem_alloc(sizeof (struct dbuf_hold_impl_data
) *
2880 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2881 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2882 fail_uncached
, tag
, dbp
, 0);
2884 error
= __dbuf_hold_impl(dh
);
2886 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2887 DBUF_HOLD_IMPL_MAX_DEPTH
);
2893 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2894 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2895 boolean_t fail_sparse
, boolean_t fail_uncached
,
2896 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2899 dh
->dh_level
= level
;
2900 dh
->dh_blkid
= blkid
;
2902 dh
->dh_fail_sparse
= fail_sparse
;
2903 dh
->dh_fail_uncached
= fail_uncached
;
2909 dh
->dh_parent
= NULL
;
2914 dh
->dh_depth
= depth
;
2918 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2920 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2924 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2927 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2928 return (err
? NULL
: db
);
2932 dbuf_create_bonus(dnode_t
*dn
)
2934 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2936 ASSERT(dn
->dn_bonus
== NULL
);
2937 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2941 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2943 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2946 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2947 return (SET_ERROR(ENOTSUP
));
2949 blksz
= SPA_MINBLOCKSIZE
;
2950 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2951 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2955 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2956 dbuf_new_size(db
, blksz
, tx
);
2957 rw_exit(&dn
->dn_struct_rwlock
);
2964 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2966 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2969 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2971 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2973 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
2974 VERIFY3S(holds
, >, 1);
2977 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2979 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2982 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2983 dmu_buf_impl_t
*found_db
;
2984 boolean_t result
= B_FALSE
;
2986 if (blkid
== DMU_BONUS_BLKID
)
2987 found_db
= dbuf_find_bonus(os
, obj
);
2989 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2991 if (found_db
!= NULL
) {
2992 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2993 (void) refcount_add(&db
->db_holds
, tag
);
2996 mutex_exit(&found_db
->db_mtx
);
3002 * If you call dbuf_rele() you had better not be referencing the dnode handle
3003 * unless you have some other direct or indirect hold on the dnode. (An indirect
3004 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3005 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3006 * dnode's parent dbuf evicting its dnode handles.
3009 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
3011 mutex_enter(&db
->db_mtx
);
3012 dbuf_rele_and_unlock(db
, tag
);
3016 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
3018 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3022 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3023 * db_dirtycnt and db_holds to be updated atomically.
3026 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
3030 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3034 * Remove the reference to the dbuf before removing its hold on the
3035 * dnode so we can guarantee in dnode_move() that a referenced bonus
3036 * buffer has a corresponding dnode hold.
3038 holds
= refcount_remove(&db
->db_holds
, tag
);
3042 * We can't freeze indirects if there is a possibility that they
3043 * may be modified in the current syncing context.
3045 if (db
->db_buf
!= NULL
&&
3046 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3047 arc_buf_freeze(db
->db_buf
);
3050 if (holds
== db
->db_dirtycnt
&&
3051 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3052 dbuf_evict_user(db
);
3055 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3057 boolean_t evict_dbuf
= db
->db_pending_evict
;
3060 * If the dnode moves here, we cannot cross this
3061 * barrier until the move completes.
3066 atomic_dec_32(&dn
->dn_dbufs_count
);
3069 * Decrementing the dbuf count means that the bonus
3070 * buffer's dnode hold is no longer discounted in
3071 * dnode_move(). The dnode cannot move until after
3072 * the dnode_rele() below.
3077 * Do not reference db after its lock is dropped.
3078 * Another thread may evict it.
3080 mutex_exit(&db
->db_mtx
);
3083 dnode_evict_bonus(dn
);
3086 } else if (db
->db_buf
== NULL
) {
3088 * This is a special case: we never associated this
3089 * dbuf with any data allocated from the ARC.
3091 ASSERT(db
->db_state
== DB_UNCACHED
||
3092 db
->db_state
== DB_NOFILL
);
3094 } else if (arc_released(db
->db_buf
)) {
3096 * This dbuf has anonymous data associated with it.
3100 boolean_t do_arc_evict
= B_FALSE
;
3102 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3104 if (!DBUF_IS_CACHEABLE(db
) &&
3105 db
->db_blkptr
!= NULL
&&
3106 !BP_IS_HOLE(db
->db_blkptr
) &&
3107 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
3108 do_arc_evict
= B_TRUE
;
3109 bp
= *db
->db_blkptr
;
3112 if (!DBUF_IS_CACHEABLE(db
) ||
3113 db
->db_pending_evict
) {
3115 } else if (!multilist_link_active(&db
->db_cache_link
)) {
3116 multilist_insert(dbuf_cache
, db
);
3117 (void) refcount_add_many(&dbuf_cache_size
,
3118 db
->db
.db_size
, db
);
3119 mutex_exit(&db
->db_mtx
);
3121 dbuf_evict_notify();
3125 arc_freed(spa
, &bp
);
3128 mutex_exit(&db
->db_mtx
);
3133 #pragma weak dmu_buf_refcount = dbuf_refcount
3135 dbuf_refcount(dmu_buf_impl_t
*db
)
3137 return (refcount_count(&db
->db_holds
));
3141 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3142 dmu_buf_user_t
*new_user
)
3144 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3146 mutex_enter(&db
->db_mtx
);
3147 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3148 if (db
->db_user
== old_user
)
3149 db
->db_user
= new_user
;
3151 old_user
= db
->db_user
;
3152 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3153 mutex_exit(&db
->db_mtx
);
3159 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3161 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3165 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3167 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3169 db
->db_user_immediate_evict
= TRUE
;
3170 return (dmu_buf_set_user(db_fake
, user
));
3174 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3176 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3180 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3182 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3184 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3185 return (db
->db_user
);
3189 dmu_buf_user_evict_wait()
3191 taskq_wait(dbu_evict_taskq
);
3195 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3197 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3198 return (dbi
->db_blkptr
);
3202 dmu_buf_get_objset(dmu_buf_t
*db
)
3204 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3205 return (dbi
->db_objset
);
3209 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3211 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3212 DB_DNODE_ENTER(dbi
);
3213 return (DB_DNODE(dbi
));
3217 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3219 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3224 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3226 /* ASSERT(dmu_tx_is_syncing(tx) */
3227 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3229 if (db
->db_blkptr
!= NULL
)
3232 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3233 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3234 BP_ZERO(db
->db_blkptr
);
3237 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3239 * This buffer was allocated at a time when there was
3240 * no available blkptrs from the dnode, or it was
3241 * inappropriate to hook it in (i.e., nlevels mis-match).
3243 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3244 ASSERT(db
->db_parent
== NULL
);
3245 db
->db_parent
= dn
->dn_dbuf
;
3246 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3249 dmu_buf_impl_t
*parent
= db
->db_parent
;
3250 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3252 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3253 if (parent
== NULL
) {
3254 mutex_exit(&db
->db_mtx
);
3255 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3256 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3257 db
->db_blkid
>> epbs
, db
);
3258 rw_exit(&dn
->dn_struct_rwlock
);
3259 mutex_enter(&db
->db_mtx
);
3260 db
->db_parent
= parent
;
3262 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3263 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3269 * Ensure the dbuf's data is untransformed if the associated dirty
3270 * record requires it. This is used by dbuf_sync_leaf() to ensure
3271 * that a dnode block is decrypted before we write new data to it.
3272 * For raw writes we assert that the buffer is already encrypted.
3275 dbuf_check_crypt(dbuf_dirty_record_t
*dr
)
3278 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3280 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3282 if (!dr
->dt
.dl
.dr_raw
&& arc_is_encrypted(db
->db_buf
)) {
3284 * Unfortunately, there is currently no mechanism for
3285 * syncing context to handle decryption errors. An error
3286 * here is only possible if an attacker maliciously
3287 * changed a dnode block and updated the associated
3288 * checksums going up the block tree.
3290 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
3291 dmu_objset_id(db
->db_objset
), B_TRUE
);
3293 panic("Invalid dnode block MAC");
3294 } else if (dr
->dt
.dl
.dr_raw
) {
3296 * Writing raw encrypted data requires the db's arc buffer
3297 * to be converted to raw by the caller.
3299 ASSERT(arc_is_encrypted(db
->db_buf
));
3304 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3305 * is critical the we not allow the compiler to inline this function in to
3306 * dbuf_sync_list() thereby drastically bloating the stack usage.
3308 noinline
static void
3309 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3311 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3315 ASSERT(dmu_tx_is_syncing(tx
));
3317 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3319 mutex_enter(&db
->db_mtx
);
3321 ASSERT(db
->db_level
> 0);
3324 /* Read the block if it hasn't been read yet. */
3325 if (db
->db_buf
== NULL
) {
3326 mutex_exit(&db
->db_mtx
);
3327 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3328 mutex_enter(&db
->db_mtx
);
3330 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3331 ASSERT(db
->db_buf
!= NULL
);
3335 /* Indirect block size must match what the dnode thinks it is. */
3336 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3337 dbuf_check_blkptr(dn
, db
);
3340 /* Provide the pending dirty record to child dbufs */
3341 db
->db_data_pending
= dr
;
3343 mutex_exit(&db
->db_mtx
);
3344 dbuf_write(dr
, db
->db_buf
, tx
);
3347 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3348 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3349 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3350 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3355 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3356 * critical the we not allow the compiler to inline this function in to
3357 * dbuf_sync_list() thereby drastically bloating the stack usage.
3359 noinline
static void
3360 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3362 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3363 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3366 uint64_t txg
= tx
->tx_txg
;
3368 ASSERT(dmu_tx_is_syncing(tx
));
3370 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3372 mutex_enter(&db
->db_mtx
);
3374 * To be synced, we must be dirtied. But we
3375 * might have been freed after the dirty.
3377 if (db
->db_state
== DB_UNCACHED
) {
3378 /* This buffer has been freed since it was dirtied */
3379 ASSERT(db
->db
.db_data
== NULL
);
3380 } else if (db
->db_state
== DB_FILL
) {
3381 /* This buffer was freed and is now being re-filled */
3382 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3384 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3391 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3392 mutex_enter(&dn
->dn_mtx
);
3393 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3395 * In the previous transaction group, the bonus buffer
3396 * was entirely used to store the attributes for the
3397 * dnode which overrode the dn_spill field. However,
3398 * when adding more attributes to the file a spill
3399 * block was required to hold the extra attributes.
3401 * Make sure to clear the garbage left in the dn_spill
3402 * field from the previous attributes in the bonus
3403 * buffer. Otherwise, after writing out the spill
3404 * block to the new allocated dva, it will free
3405 * the old block pointed to by the invalid dn_spill.
3407 db
->db_blkptr
= NULL
;
3409 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3410 mutex_exit(&dn
->dn_mtx
);
3414 * If this is a bonus buffer, simply copy the bonus data into the
3415 * dnode. It will be written out when the dnode is synced (and it
3416 * will be synced, since it must have been dirty for dbuf_sync to
3419 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3420 dbuf_dirty_record_t
**drp
;
3422 ASSERT(*datap
!= NULL
);
3423 ASSERT0(db
->db_level
);
3424 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
3425 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3426 bcopy(*datap
, DN_BONUS(dn
->dn_phys
),
3427 DN_MAX_BONUS_LEN(dn
->dn_phys
));
3430 if (*datap
!= db
->db
.db_data
) {
3431 int slots
= DB_DNODE(db
)->dn_num_slots
;
3432 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3433 kmem_free(*datap
, bonuslen
);
3434 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3436 db
->db_data_pending
= NULL
;
3437 drp
= &db
->db_last_dirty
;
3439 drp
= &(*drp
)->dr_next
;
3440 ASSERT(dr
->dr_next
== NULL
);
3441 ASSERT(dr
->dr_dbuf
== db
);
3443 if (dr
->dr_dbuf
->db_level
!= 0) {
3444 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3445 list_destroy(&dr
->dt
.di
.dr_children
);
3447 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3448 ASSERT(db
->db_dirtycnt
> 0);
3449 db
->db_dirtycnt
-= 1;
3450 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3457 * This function may have dropped the db_mtx lock allowing a dmu_sync
3458 * operation to sneak in. As a result, we need to ensure that we
3459 * don't check the dr_override_state until we have returned from
3460 * dbuf_check_blkptr.
3462 dbuf_check_blkptr(dn
, db
);
3465 * If this buffer is in the middle of an immediate write,
3466 * wait for the synchronous IO to complete.
3468 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3469 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3470 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3471 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3475 * If this is a dnode block, ensure it is appropriately encrypted
3476 * or decrypted, depending on what we are writing to it this txg.
3478 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
3479 dbuf_check_crypt(dr
);
3481 if (db
->db_state
!= DB_NOFILL
&&
3482 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3483 refcount_count(&db
->db_holds
) > 1 &&
3484 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3485 *datap
== db
->db_buf
) {
3487 * If this buffer is currently "in use" (i.e., there
3488 * are active holds and db_data still references it),
3489 * then make a copy before we start the write so that
3490 * any modifications from the open txg will not leak
3493 * NOTE: this copy does not need to be made for
3494 * objects only modified in the syncing context (e.g.
3495 * DNONE_DNODE blocks).
3497 int psize
= arc_buf_size(*datap
);
3498 int lsize
= arc_buf_lsize(*datap
);
3499 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3500 enum zio_compress compress_type
= arc_get_compression(*datap
);
3502 if (arc_is_encrypted(*datap
)) {
3503 boolean_t byteorder
;
3504 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3505 uint8_t iv
[ZIO_DATA_IV_LEN
];
3506 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3508 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
3509 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
3510 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
3511 dn
->dn_type
, psize
, lsize
, compress_type
);
3512 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3513 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
3514 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
3515 psize
, lsize
, compress_type
);
3517 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
3519 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
3521 db
->db_data_pending
= dr
;
3523 mutex_exit(&db
->db_mtx
);
3525 dbuf_write(dr
, *datap
, tx
);
3527 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3528 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3529 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3533 * Although zio_nowait() does not "wait for an IO", it does
3534 * initiate the IO. If this is an empty write it seems plausible
3535 * that the IO could actually be completed before the nowait
3536 * returns. We need to DB_DNODE_EXIT() first in case
3537 * zio_nowait() invalidates the dbuf.
3540 zio_nowait(dr
->dr_zio
);
3545 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3547 dbuf_dirty_record_t
*dr
;
3549 while ((dr
= list_head(list
))) {
3550 if (dr
->dr_zio
!= NULL
) {
3552 * If we find an already initialized zio then we
3553 * are processing the meta-dnode, and we have finished.
3554 * The dbufs for all dnodes are put back on the list
3555 * during processing, so that we can zio_wait()
3556 * these IOs after initiating all child IOs.
3558 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3559 DMU_META_DNODE_OBJECT
);
3562 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3563 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3564 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3566 list_remove(list
, dr
);
3567 if (dr
->dr_dbuf
->db_level
> 0)
3568 dbuf_sync_indirect(dr
, tx
);
3570 dbuf_sync_leaf(dr
, tx
);
3576 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3578 dmu_buf_impl_t
*db
= vdb
;
3580 blkptr_t
*bp
= zio
->io_bp
;
3581 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3582 spa_t
*spa
= zio
->io_spa
;
3587 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3588 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3592 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3593 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3594 zio
->io_prev_space_delta
= delta
;
3596 if (bp
->blk_birth
!= 0) {
3597 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3598 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3599 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3600 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3601 BP_IS_EMBEDDED(bp
));
3602 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3605 mutex_enter(&db
->db_mtx
);
3608 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3609 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3610 ASSERT(!(BP_IS_HOLE(bp
)) &&
3611 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3615 if (db
->db_level
== 0) {
3616 mutex_enter(&dn
->dn_mtx
);
3617 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3618 db
->db_blkid
!= DMU_SPILL_BLKID
)
3619 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3620 mutex_exit(&dn
->dn_mtx
);
3622 if (dn
->dn_type
== DMU_OT_DNODE
) {
3624 while (i
< db
->db
.db_size
) {
3626 (void *)(((char *)db
->db
.db_data
) + i
);
3628 i
+= DNODE_MIN_SIZE
;
3629 if (dnp
->dn_type
!= DMU_OT_NONE
) {
3631 i
+= dnp
->dn_extra_slots
*
3636 if (BP_IS_HOLE(bp
)) {
3643 blkptr_t
*ibp
= db
->db
.db_data
;
3644 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3645 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3646 if (BP_IS_HOLE(ibp
))
3648 fill
+= BP_GET_FILL(ibp
);
3653 if (!BP_IS_EMBEDDED(bp
))
3654 BP_SET_FILL(bp
, fill
);
3656 mutex_exit(&db
->db_mtx
);
3658 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3659 *db
->db_blkptr
= *bp
;
3660 rw_exit(&dn
->dn_struct_rwlock
);
3665 * This function gets called just prior to running through the compression
3666 * stage of the zio pipeline. If we're an indirect block comprised of only
3667 * holes, then we want this indirect to be compressed away to a hole. In
3668 * order to do that we must zero out any information about the holes that
3669 * this indirect points to prior to before we try to compress it.
3672 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3674 dmu_buf_impl_t
*db
= vdb
;
3677 unsigned int epbs
, i
;
3679 ASSERT3U(db
->db_level
, >, 0);
3682 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3683 ASSERT3U(epbs
, <, 31);
3685 /* Determine if all our children are holes */
3686 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
3687 if (!BP_IS_HOLE(bp
))
3692 * If all the children are holes, then zero them all out so that
3693 * we may get compressed away.
3695 if (i
== 1ULL << epbs
) {
3697 * We only found holes. Grab the rwlock to prevent
3698 * anybody from reading the blocks we're about to
3701 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3702 bzero(db
->db
.db_data
, db
->db
.db_size
);
3703 rw_exit(&dn
->dn_struct_rwlock
);
3709 * The SPA will call this callback several times for each zio - once
3710 * for every physical child i/o (zio->io_phys_children times). This
3711 * allows the DMU to monitor the progress of each logical i/o. For example,
3712 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3713 * block. There may be a long delay before all copies/fragments are completed,
3714 * so this callback allows us to retire dirty space gradually, as the physical
3719 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3721 dmu_buf_impl_t
*db
= arg
;
3722 objset_t
*os
= db
->db_objset
;
3723 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3724 dbuf_dirty_record_t
*dr
;
3727 dr
= db
->db_data_pending
;
3728 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3731 * The callback will be called io_phys_children times. Retire one
3732 * portion of our dirty space each time we are called. Any rounding
3733 * error will be cleaned up by dsl_pool_sync()'s call to
3734 * dsl_pool_undirty_space().
3736 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3737 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3742 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3744 dmu_buf_impl_t
*db
= vdb
;
3745 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3746 blkptr_t
*bp
= db
->db_blkptr
;
3747 objset_t
*os
= db
->db_objset
;
3748 dmu_tx_t
*tx
= os
->os_synctx
;
3749 dbuf_dirty_record_t
**drp
, *dr
;
3751 ASSERT0(zio
->io_error
);
3752 ASSERT(db
->db_blkptr
== bp
);
3755 * For nopwrites and rewrites we ensure that the bp matches our
3756 * original and bypass all the accounting.
3758 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3759 ASSERT(BP_EQUAL(bp
, bp_orig
));
3761 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3762 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3763 dsl_dataset_block_born(ds
, bp
, tx
);
3766 mutex_enter(&db
->db_mtx
);
3770 drp
= &db
->db_last_dirty
;
3771 while ((dr
= *drp
) != db
->db_data_pending
)
3773 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3774 ASSERT(dr
->dr_dbuf
== db
);
3775 ASSERT(dr
->dr_next
== NULL
);
3779 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3784 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3785 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3786 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3791 if (db
->db_level
== 0) {
3792 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3793 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3794 if (db
->db_state
!= DB_NOFILL
) {
3795 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3796 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
3803 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3804 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3805 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3806 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3808 ASSERT3U(db
->db_blkid
, <=,
3809 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3810 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3814 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3815 list_destroy(&dr
->dt
.di
.dr_children
);
3817 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3819 cv_broadcast(&db
->db_changed
);
3820 ASSERT(db
->db_dirtycnt
> 0);
3821 db
->db_dirtycnt
-= 1;
3822 db
->db_data_pending
= NULL
;
3823 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3827 dbuf_write_nofill_ready(zio_t
*zio
)
3829 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3833 dbuf_write_nofill_done(zio_t
*zio
)
3835 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3839 dbuf_write_override_ready(zio_t
*zio
)
3841 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3842 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3844 dbuf_write_ready(zio
, NULL
, db
);
3848 dbuf_write_override_done(zio_t
*zio
)
3850 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3851 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3852 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3854 mutex_enter(&db
->db_mtx
);
3855 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3856 if (!BP_IS_HOLE(obp
))
3857 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3858 arc_release(dr
->dt
.dl
.dr_data
, db
);
3860 mutex_exit(&db
->db_mtx
);
3862 dbuf_write_done(zio
, NULL
, db
);
3864 if (zio
->io_abd
!= NULL
)
3865 abd_put(zio
->io_abd
);
3868 /* Issue I/O to commit a dirty buffer to disk. */
3870 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3872 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3875 dmu_buf_impl_t
*parent
= db
->db_parent
;
3876 uint64_t txg
= tx
->tx_txg
;
3877 zbookmark_phys_t zb
;
3882 ASSERT(dmu_tx_is_syncing(tx
));
3888 if (db
->db_state
!= DB_NOFILL
) {
3889 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3891 * Private object buffers are released here rather
3892 * than in dbuf_dirty() since they are only modified
3893 * in the syncing context and we don't want the
3894 * overhead of making multiple copies of the data.
3896 if (BP_IS_HOLE(db
->db_blkptr
)) {
3899 dbuf_release_bp(db
);
3904 if (parent
!= dn
->dn_dbuf
) {
3905 /* Our parent is an indirect block. */
3906 /* We have a dirty parent that has been scheduled for write. */
3907 ASSERT(parent
&& parent
->db_data_pending
);
3908 /* Our parent's buffer is one level closer to the dnode. */
3909 ASSERT(db
->db_level
== parent
->db_level
-1);
3911 * We're about to modify our parent's db_data by modifying
3912 * our block pointer, so the parent must be released.
3914 ASSERT(arc_released(parent
->db_buf
));
3915 zio
= parent
->db_data_pending
->dr_zio
;
3917 /* Our parent is the dnode itself. */
3918 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3919 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3920 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3921 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3922 ASSERT3P(db
->db_blkptr
, ==,
3923 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3927 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3928 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3931 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3932 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3933 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3935 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3937 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3939 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3943 * We copy the blkptr now (rather than when we instantiate the dirty
3944 * record), because its value can change between open context and
3945 * syncing context. We do not need to hold dn_struct_rwlock to read
3946 * db_blkptr because we are in syncing context.
3948 dr
->dr_bp_copy
= *db
->db_blkptr
;
3950 if (db
->db_level
== 0 &&
3951 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3953 * The BP for this block has been provided by open context
3954 * (by dmu_sync() or dmu_buf_write_embedded()).
3956 abd_t
*contents
= (data
!= NULL
) ?
3957 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
3959 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3960 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
3961 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
3962 dbuf_write_override_done
,
3963 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3964 mutex_enter(&db
->db_mtx
);
3965 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3966 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3967 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3968 mutex_exit(&db
->db_mtx
);
3969 } else if (db
->db_state
== DB_NOFILL
) {
3970 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
3971 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
3972 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3973 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
3974 dbuf_write_nofill_ready
, NULL
, NULL
,
3975 dbuf_write_nofill_done
, db
,
3976 ZIO_PRIORITY_ASYNC_WRITE
,
3977 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3979 ASSERT(arc_released(data
));
3982 * For indirect blocks, we want to setup the children
3983 * ready callback so that we can properly handle an indirect
3984 * block that only contains holes.
3986 arc_write_done_func_t
*children_ready_cb
= NULL
;
3987 if (db
->db_level
!= 0)
3988 children_ready_cb
= dbuf_write_children_ready
;
3990 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3991 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3992 &zp
, dbuf_write_ready
,
3993 children_ready_cb
, dbuf_write_physdone
,
3994 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
3995 ZIO_FLAG_MUSTSUCCEED
, &zb
);
3999 #if defined(_KERNEL) && defined(HAVE_SPL)
4000 EXPORT_SYMBOL(dbuf_find
);
4001 EXPORT_SYMBOL(dbuf_is_metadata
);
4002 EXPORT_SYMBOL(dbuf_destroy
);
4003 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
4004 EXPORT_SYMBOL(dbuf_whichblock
);
4005 EXPORT_SYMBOL(dbuf_read
);
4006 EXPORT_SYMBOL(dbuf_unoverride
);
4007 EXPORT_SYMBOL(dbuf_free_range
);
4008 EXPORT_SYMBOL(dbuf_new_size
);
4009 EXPORT_SYMBOL(dbuf_release_bp
);
4010 EXPORT_SYMBOL(dbuf_dirty
);
4011 EXPORT_SYMBOL(dmu_buf_will_change_crypt_params
);
4012 EXPORT_SYMBOL(dmu_buf_will_dirty
);
4013 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
4014 EXPORT_SYMBOL(dmu_buf_will_fill
);
4015 EXPORT_SYMBOL(dmu_buf_fill_done
);
4016 EXPORT_SYMBOL(dmu_buf_rele
);
4017 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
4018 EXPORT_SYMBOL(dbuf_prefetch
);
4019 EXPORT_SYMBOL(dbuf_hold_impl
);
4020 EXPORT_SYMBOL(dbuf_hold
);
4021 EXPORT_SYMBOL(dbuf_hold_level
);
4022 EXPORT_SYMBOL(dbuf_create_bonus
);
4023 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
4024 EXPORT_SYMBOL(dbuf_rm_spill
);
4025 EXPORT_SYMBOL(dbuf_add_ref
);
4026 EXPORT_SYMBOL(dbuf_rele
);
4027 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
4028 EXPORT_SYMBOL(dbuf_refcount
);
4029 EXPORT_SYMBOL(dbuf_sync_list
);
4030 EXPORT_SYMBOL(dmu_buf_set_user
);
4031 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
4032 EXPORT_SYMBOL(dmu_buf_get_user
);
4033 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
4036 module_param(dbuf_cache_max_bytes
, ulong
, 0644);
4037 MODULE_PARM_DESC(dbuf_cache_max_bytes
,
4038 "Maximum size in bytes of the dbuf cache.");
4040 module_param(dbuf_cache_hiwater_pct
, uint
, 0644);
4041 MODULE_PARM_DESC(dbuf_cache_hiwater_pct
,
4042 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4045 module_param(dbuf_cache_lowater_pct
, uint
, 0644);
4046 MODULE_PARM_DESC(dbuf_cache_lowater_pct
,
4047 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4050 module_param(dbuf_cache_max_shift
, int, 0644);
4051 MODULE_PARM_DESC(dbuf_cache_max_shift
,
4052 "Cap the size of the dbuf cache to a log2 fraction of arc size.");