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
;
66 arc_buf_contents_t dh_type
;
70 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
71 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, boolean_t fail_sparse
,
72 boolean_t fail_uncached
,
73 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
74 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
76 uint_t zfs_dbuf_evict_key
;
78 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
79 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
81 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
82 dmu_buf_evict_func_t
*evict_func_sync
,
83 dmu_buf_evict_func_t
*evict_func_async
,
84 dmu_buf_t
**clear_on_evict_dbufp
);
87 * Global data structures and functions for the dbuf cache.
89 static kmem_cache_t
*dbuf_kmem_cache
;
90 static taskq_t
*dbu_evict_taskq
;
92 static kthread_t
*dbuf_cache_evict_thread
;
93 static kmutex_t dbuf_evict_lock
;
94 static kcondvar_t dbuf_evict_cv
;
95 static boolean_t dbuf_evict_thread_exit
;
98 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
99 * are not currently held but have been recently released. These dbufs
100 * are not eligible for arc eviction until they are aged out of the cache.
101 * Dbufs are added to the dbuf cache once the last hold is released. If a
102 * dbuf is later accessed and still exists in the dbuf cache, then it will
103 * be removed from the cache and later re-added to the head of the cache.
104 * Dbufs that are aged out of the cache will be immediately destroyed and
105 * become eligible for arc eviction.
107 static multilist_t
*dbuf_cache
;
108 static refcount_t dbuf_cache_size
;
109 unsigned long dbuf_cache_max_bytes
= 100 * 1024 * 1024;
111 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
112 int dbuf_cache_max_shift
= 5;
115 * The dbuf cache uses a three-stage eviction policy:
116 * - A low water marker designates when the dbuf eviction thread
117 * should stop evicting from the dbuf cache.
118 * - When we reach the maximum size (aka mid water mark), we
119 * signal the eviction thread to run.
120 * - The high water mark indicates when the eviction thread
121 * is unable to keep up with the incoming load and eviction must
122 * happen in the context of the calling thread.
126 * low water mid water hi water
127 * +----------------------------------------+----------+----------+
132 * +----------------------------------------+----------+----------+
134 * evicting eviction directly
137 * The high and low water marks indicate the operating range for the eviction
138 * thread. The low water mark is, by default, 90% of the total size of the
139 * cache and the high water mark is at 110% (both of these percentages can be
140 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
141 * respectively). The eviction thread will try to ensure that the cache remains
142 * within this range by waking up every second and checking if the cache is
143 * above the low water mark. The thread can also be woken up by callers adding
144 * elements into the cache if the cache is larger than the mid water (i.e max
145 * cache size). Once the eviction thread is woken up and eviction is required,
146 * it will continue evicting buffers until it's able to reduce the cache size
147 * to the low water mark. If the cache size continues to grow and hits the high
148 * water mark, then callers adding elements to the cache will begin to evict
149 * directly from the cache until the cache is no longer above the high water
154 * The percentage above and below the maximum cache size.
156 uint_t dbuf_cache_hiwater_pct
= 10;
157 uint_t dbuf_cache_lowater_pct
= 10;
161 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
163 dmu_buf_impl_t
*db
= vdb
;
164 bzero(db
, sizeof (dmu_buf_impl_t
));
166 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
167 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
168 multilist_link_init(&db
->db_cache_link
);
169 refcount_create(&db
->db_holds
);
170 multilist_link_init(&db
->db_cache_link
);
177 dbuf_dest(void *vdb
, void *unused
)
179 dmu_buf_impl_t
*db
= vdb
;
180 mutex_destroy(&db
->db_mtx
);
181 cv_destroy(&db
->db_changed
);
182 ASSERT(!multilist_link_active(&db
->db_cache_link
));
183 refcount_destroy(&db
->db_holds
);
187 * dbuf hash table routines
189 static dbuf_hash_table_t dbuf_hash_table
;
191 static uint64_t dbuf_hash_count
;
194 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
196 uintptr_t osv
= (uintptr_t)os
;
197 uint64_t crc
= -1ULL;
199 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
200 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
201 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
202 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
203 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
204 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
205 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
207 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
212 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
213 ((dbuf)->db.db_object == (obj) && \
214 (dbuf)->db_objset == (os) && \
215 (dbuf)->db_level == (level) && \
216 (dbuf)->db_blkid == (blkid))
219 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
221 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
226 hv
= dbuf_hash(os
, obj
, level
, blkid
);
227 idx
= hv
& h
->hash_table_mask
;
229 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
230 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
231 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
232 mutex_enter(&db
->db_mtx
);
233 if (db
->db_state
!= DB_EVICTING
) {
234 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
237 mutex_exit(&db
->db_mtx
);
240 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
244 static dmu_buf_impl_t
*
245 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
248 dmu_buf_impl_t
*db
= NULL
;
250 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
251 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
252 if (dn
->dn_bonus
!= NULL
) {
254 mutex_enter(&db
->db_mtx
);
256 rw_exit(&dn
->dn_struct_rwlock
);
257 dnode_rele(dn
, FTAG
);
263 * Insert an entry into the hash table. If there is already an element
264 * equal to elem in the hash table, then the already existing element
265 * will be returned and the new element will not be inserted.
266 * Otherwise returns NULL.
268 static dmu_buf_impl_t
*
269 dbuf_hash_insert(dmu_buf_impl_t
*db
)
271 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
272 objset_t
*os
= db
->db_objset
;
273 uint64_t obj
= db
->db
.db_object
;
274 int level
= db
->db_level
;
275 uint64_t blkid
, hv
, idx
;
278 blkid
= db
->db_blkid
;
279 hv
= dbuf_hash(os
, obj
, level
, blkid
);
280 idx
= hv
& h
->hash_table_mask
;
282 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
283 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
284 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
285 mutex_enter(&dbf
->db_mtx
);
286 if (dbf
->db_state
!= DB_EVICTING
) {
287 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
290 mutex_exit(&dbf
->db_mtx
);
294 mutex_enter(&db
->db_mtx
);
295 db
->db_hash_next
= h
->hash_table
[idx
];
296 h
->hash_table
[idx
] = db
;
297 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
298 atomic_inc_64(&dbuf_hash_count
);
304 * Remove an entry from the hash table. It must be in the EVICTING state.
307 dbuf_hash_remove(dmu_buf_impl_t
*db
)
309 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
311 dmu_buf_impl_t
*dbf
, **dbp
;
313 hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
314 db
->db_level
, db
->db_blkid
);
315 idx
= hv
& h
->hash_table_mask
;
318 * We mustn't hold db_mtx to maintain lock ordering:
319 * DBUF_HASH_MUTEX > db_mtx.
321 ASSERT(refcount_is_zero(&db
->db_holds
));
322 ASSERT(db
->db_state
== DB_EVICTING
);
323 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
325 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
326 dbp
= &h
->hash_table
[idx
];
327 while ((dbf
= *dbp
) != db
) {
328 dbp
= &dbf
->db_hash_next
;
331 *dbp
= db
->db_hash_next
;
332 db
->db_hash_next
= NULL
;
333 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
334 atomic_dec_64(&dbuf_hash_count
);
340 } dbvu_verify_type_t
;
343 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
348 if (db
->db_user
== NULL
)
351 /* Only data blocks support the attachment of user data. */
352 ASSERT(db
->db_level
== 0);
354 /* Clients must resolve a dbuf before attaching user data. */
355 ASSERT(db
->db
.db_data
!= NULL
);
356 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
358 holds
= refcount_count(&db
->db_holds
);
359 if (verify_type
== DBVU_EVICTING
) {
361 * Immediate eviction occurs when holds == dirtycnt.
362 * For normal eviction buffers, holds is zero on
363 * eviction, except when dbuf_fix_old_data() calls
364 * dbuf_clear_data(). However, the hold count can grow
365 * during eviction even though db_mtx is held (see
366 * dmu_bonus_hold() for an example), so we can only
367 * test the generic invariant that holds >= dirtycnt.
369 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
371 if (db
->db_user_immediate_evict
== TRUE
)
372 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
374 ASSERT3U(holds
, >, 0);
380 dbuf_evict_user(dmu_buf_impl_t
*db
)
382 dmu_buf_user_t
*dbu
= db
->db_user
;
384 ASSERT(MUTEX_HELD(&db
->db_mtx
));
389 dbuf_verify_user(db
, DBVU_EVICTING
);
393 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
394 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
398 * There are two eviction callbacks - one that we call synchronously
399 * and one that we invoke via a taskq. The async one is useful for
400 * avoiding lock order reversals and limiting stack depth.
402 * Note that if we have a sync callback but no async callback,
403 * it's likely that the sync callback will free the structure
404 * containing the dbu. In that case we need to take care to not
405 * dereference dbu after calling the sync evict func.
407 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
409 if (dbu
->dbu_evict_func_sync
!= NULL
)
410 dbu
->dbu_evict_func_sync(dbu
);
413 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
414 dbu
, 0, &dbu
->dbu_tqent
);
419 dbuf_is_metadata(dmu_buf_impl_t
*db
)
422 * Consider indirect blocks and spill blocks to be meta data.
424 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
427 boolean_t is_metadata
;
430 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
433 return (is_metadata
);
439 * This function *must* return indices evenly distributed between all
440 * sublists of the multilist. This is needed due to how the dbuf eviction
441 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
442 * distributed between all sublists and uses this assumption when
443 * deciding which sublist to evict from and how much to evict from it.
446 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
448 dmu_buf_impl_t
*db
= obj
;
451 * The assumption here, is the hash value for a given
452 * dmu_buf_impl_t will remain constant throughout it's lifetime
453 * (i.e. it's objset, object, level and blkid fields don't change).
454 * Thus, we don't need to store the dbuf's sublist index
455 * on insertion, as this index can be recalculated on removal.
457 * Also, the low order bits of the hash value are thought to be
458 * distributed evenly. Otherwise, in the case that the multilist
459 * has a power of two number of sublists, each sublists' usage
460 * would not be evenly distributed.
462 return (dbuf_hash(db
->db_objset
, db
->db
.db_object
,
463 db
->db_level
, db
->db_blkid
) %
464 multilist_get_num_sublists(ml
));
467 static inline boolean_t
468 dbuf_cache_above_hiwater(void)
470 uint64_t dbuf_cache_hiwater_bytes
=
471 (dbuf_cache_max_bytes
* dbuf_cache_hiwater_pct
) / 100;
473 return (refcount_count(&dbuf_cache_size
) >
474 dbuf_cache_max_bytes
+ dbuf_cache_hiwater_bytes
);
477 static inline boolean_t
478 dbuf_cache_above_lowater(void)
480 uint64_t dbuf_cache_lowater_bytes
=
481 (dbuf_cache_max_bytes
* dbuf_cache_lowater_pct
) / 100;
483 return (refcount_count(&dbuf_cache_size
) >
484 dbuf_cache_max_bytes
- dbuf_cache_lowater_bytes
);
488 * Evict the oldest eligible dbuf from the dbuf cache.
493 int idx
= multilist_get_random_index(dbuf_cache
);
494 multilist_sublist_t
*mls
= multilist_sublist_lock(dbuf_cache
, idx
);
496 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
499 * Set the thread's tsd to indicate that it's processing evictions.
500 * Once a thread stops evicting from the dbuf cache it will
501 * reset its tsd to NULL.
503 ASSERT3P(tsd_get(zfs_dbuf_evict_key
), ==, NULL
);
504 (void) tsd_set(zfs_dbuf_evict_key
, (void *)B_TRUE
);
506 db
= multilist_sublist_tail(mls
);
507 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
508 db
= multilist_sublist_prev(mls
, db
);
511 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
512 multilist_sublist_t
*, mls
);
515 multilist_sublist_remove(mls
, db
);
516 multilist_sublist_unlock(mls
);
517 (void) refcount_remove_many(&dbuf_cache_size
,
521 multilist_sublist_unlock(mls
);
523 (void) tsd_set(zfs_dbuf_evict_key
, NULL
);
527 * The dbuf evict thread is responsible for aging out dbufs from the
528 * cache. Once the cache has reached it's maximum size, dbufs are removed
529 * and destroyed. The eviction thread will continue running until the size
530 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
531 * out of the cache it is destroyed and becomes eligible for arc eviction.
534 dbuf_evict_thread(void)
538 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
540 mutex_enter(&dbuf_evict_lock
);
541 while (!dbuf_evict_thread_exit
) {
542 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
543 CALLB_CPR_SAFE_BEGIN(&cpr
);
544 (void) cv_timedwait_sig_hires(&dbuf_evict_cv
,
545 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
546 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
548 mutex_exit(&dbuf_evict_lock
);
551 * Keep evicting as long as we're above the low water mark
552 * for the cache. We do this without holding the locks to
553 * minimize lock contention.
555 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
559 mutex_enter(&dbuf_evict_lock
);
562 dbuf_evict_thread_exit
= B_FALSE
;
563 cv_broadcast(&dbuf_evict_cv
);
564 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
569 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
570 * If the dbuf cache is at its high water mark, then evict a dbuf from the
571 * dbuf cache using the callers context.
574 dbuf_evict_notify(void)
578 * We use thread specific data to track when a thread has
579 * started processing evictions. This allows us to avoid deeply
580 * nested stacks that would have a call flow similar to this:
582 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
585 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
587 * The dbuf_eviction_thread will always have its tsd set until
588 * that thread exits. All other threads will only set their tsd
589 * if they are participating in the eviction process. This only
590 * happens if the eviction thread is unable to process evictions
591 * fast enough. To keep the dbuf cache size in check, other threads
592 * can evict from the dbuf cache directly. Those threads will set
593 * their tsd values so that we ensure that they only evict one dbuf
594 * from the dbuf cache.
596 if (tsd_get(zfs_dbuf_evict_key
) != NULL
)
600 * We check if we should evict without holding the dbuf_evict_lock,
601 * because it's OK to occasionally make the wrong decision here,
602 * and grabbing the lock results in massive lock contention.
604 if (refcount_count(&dbuf_cache_size
) > dbuf_cache_max_bytes
) {
605 if (dbuf_cache_above_hiwater())
607 cv_signal(&dbuf_evict_cv
);
616 uint64_t hsize
= 1ULL << 16;
617 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
621 * The hash table is big enough to fill all of physical memory
622 * with an average block size of zfs_arc_average_blocksize (default 8K).
623 * By default, the table will take up
624 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
626 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
630 h
->hash_table_mask
= hsize
- 1;
631 #if defined(_KERNEL) && defined(HAVE_SPL)
633 * Large allocations which do not require contiguous pages
634 * should be using vmem_alloc() in the linux kernel
636 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
638 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
640 if (h
->hash_table
== NULL
) {
641 /* XXX - we should really return an error instead of assert */
642 ASSERT(hsize
> (1ULL << 10));
647 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
648 sizeof (dmu_buf_impl_t
),
649 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
651 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
652 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
657 * Setup the parameters for the dbuf cache. We cap the size of the
658 * dbuf cache to 1/32nd (default) of the size of the ARC.
660 dbuf_cache_max_bytes
= MIN(dbuf_cache_max_bytes
,
661 arc_max_bytes() >> dbuf_cache_max_shift
);
664 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
665 * configuration is not required.
667 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
669 dbuf_cache
= multilist_create(sizeof (dmu_buf_impl_t
),
670 offsetof(dmu_buf_impl_t
, db_cache_link
),
671 dbuf_cache_multilist_index_func
);
672 refcount_create(&dbuf_cache_size
);
674 tsd_create(&zfs_dbuf_evict_key
, NULL
);
675 dbuf_evict_thread_exit
= B_FALSE
;
676 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
677 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
678 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
679 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
685 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
688 dbuf_stats_destroy();
690 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
691 mutex_destroy(&h
->hash_mutexes
[i
]);
692 #if defined(_KERNEL) && defined(HAVE_SPL)
694 * Large allocations which do not require contiguous pages
695 * should be using vmem_free() in the linux kernel
697 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
699 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
701 kmem_cache_destroy(dbuf_kmem_cache
);
702 taskq_destroy(dbu_evict_taskq
);
704 mutex_enter(&dbuf_evict_lock
);
705 dbuf_evict_thread_exit
= B_TRUE
;
706 while (dbuf_evict_thread_exit
) {
707 cv_signal(&dbuf_evict_cv
);
708 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
710 mutex_exit(&dbuf_evict_lock
);
711 tsd_destroy(&zfs_dbuf_evict_key
);
713 mutex_destroy(&dbuf_evict_lock
);
714 cv_destroy(&dbuf_evict_cv
);
716 refcount_destroy(&dbuf_cache_size
);
717 multilist_destroy(dbuf_cache
);
726 dbuf_verify(dmu_buf_impl_t
*db
)
729 dbuf_dirty_record_t
*dr
;
731 ASSERT(MUTEX_HELD(&db
->db_mtx
));
733 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
736 ASSERT(db
->db_objset
!= NULL
);
740 ASSERT(db
->db_parent
== NULL
);
741 ASSERT(db
->db_blkptr
== NULL
);
743 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
744 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
745 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
746 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
747 db
->db_blkid
== DMU_SPILL_BLKID
||
748 !avl_is_empty(&dn
->dn_dbufs
));
750 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
752 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
753 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
754 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
756 ASSERT0(db
->db
.db_offset
);
758 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
761 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
762 ASSERT(dr
->dr_dbuf
== db
);
764 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
765 ASSERT(dr
->dr_dbuf
== db
);
768 * We can't assert that db_size matches dn_datablksz because it
769 * can be momentarily different when another thread is doing
772 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
773 dr
= db
->db_data_pending
;
775 * It should only be modified in syncing context, so
776 * make sure we only have one copy of the data.
778 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
781 /* verify db->db_blkptr */
783 if (db
->db_parent
== dn
->dn_dbuf
) {
784 /* db is pointed to by the dnode */
785 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
786 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
787 ASSERT(db
->db_parent
== NULL
);
789 ASSERT(db
->db_parent
!= NULL
);
790 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
791 ASSERT3P(db
->db_blkptr
, ==,
792 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
794 /* db is pointed to by an indirect block */
795 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
797 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
798 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
801 * dnode_grow_indblksz() can make this fail if we don't
802 * have the struct_rwlock. XXX indblksz no longer
803 * grows. safe to do this now?
805 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
806 ASSERT3P(db
->db_blkptr
, ==,
807 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
808 db
->db_blkid
% epb
));
812 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
813 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
814 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
815 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
817 * If the blkptr isn't set but they have nonzero data,
818 * it had better be dirty, otherwise we'll lose that
819 * data when we evict this buffer.
821 * There is an exception to this rule for indirect blocks; in
822 * this case, if the indirect block is a hole, we fill in a few
823 * fields on each of the child blocks (importantly, birth time)
824 * to prevent hole birth times from being lost when you
825 * partially fill in a hole.
827 if (db
->db_dirtycnt
== 0) {
828 if (db
->db_level
== 0) {
829 uint64_t *buf
= db
->db
.db_data
;
832 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
837 blkptr_t
*bps
= db
->db
.db_data
;
838 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
841 * We want to verify that all the blkptrs in the
842 * indirect block are holes, but we may have
843 * automatically set up a few fields for them.
844 * We iterate through each blkptr and verify
845 * they only have those fields set.
848 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
850 blkptr_t
*bp
= &bps
[i
];
851 ASSERT(ZIO_CHECKSUM_IS_ZERO(
854 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
855 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
856 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
857 ASSERT0(bp
->blk_fill
);
858 ASSERT0(bp
->blk_pad
[0]);
859 ASSERT0(bp
->blk_pad
[1]);
860 ASSERT(!BP_IS_EMBEDDED(bp
));
861 ASSERT(BP_IS_HOLE(bp
));
862 ASSERT0(bp
->blk_phys_birth
);
872 dbuf_clear_data(dmu_buf_impl_t
*db
)
874 ASSERT(MUTEX_HELD(&db
->db_mtx
));
876 ASSERT3P(db
->db_buf
, ==, NULL
);
877 db
->db
.db_data
= NULL
;
878 if (db
->db_state
!= DB_NOFILL
)
879 db
->db_state
= DB_UNCACHED
;
883 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
885 ASSERT(MUTEX_HELD(&db
->db_mtx
));
889 ASSERT(buf
->b_data
!= NULL
);
890 db
->db
.db_data
= buf
->b_data
;
894 * Loan out an arc_buf for read. Return the loaned arc_buf.
897 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
901 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
902 mutex_enter(&db
->db_mtx
);
903 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
904 int blksz
= db
->db
.db_size
;
905 spa_t
*spa
= db
->db_objset
->os_spa
;
907 mutex_exit(&db
->db_mtx
);
908 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
909 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
912 arc_loan_inuse_buf(abuf
, db
);
915 mutex_exit(&db
->db_mtx
);
921 * Calculate which level n block references the data at the level 0 offset
925 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
927 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
929 * The level n blkid is equal to the level 0 blkid divided by
930 * the number of level 0s in a level n block.
932 * The level 0 blkid is offset >> datablkshift =
933 * offset / 2^datablkshift.
935 * The number of level 0s in a level n is the number of block
936 * pointers in an indirect block, raised to the power of level.
937 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
938 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
940 * Thus, the level n blkid is: offset /
941 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
942 * = offset / 2^(datablkshift + level *
943 * (indblkshift - SPA_BLKPTRSHIFT))
944 * = offset >> (datablkshift + level *
945 * (indblkshift - SPA_BLKPTRSHIFT))
948 const unsigned exp
= dn
->dn_datablkshift
+
949 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
951 if (exp
>= 8 * sizeof (offset
)) {
952 /* This only happens on the highest indirection level */
953 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
957 ASSERT3U(exp
, <, 8 * sizeof (offset
));
959 return (offset
>> exp
);
961 ASSERT3U(offset
, <, dn
->dn_datablksz
);
967 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
969 dmu_buf_impl_t
*db
= vdb
;
971 mutex_enter(&db
->db_mtx
);
972 ASSERT3U(db
->db_state
, ==, DB_READ
);
974 * All reads are synchronous, so we must have a hold on the dbuf
976 ASSERT(refcount_count(&db
->db_holds
) > 0);
977 ASSERT(db
->db_buf
== NULL
);
978 ASSERT(db
->db
.db_data
== NULL
);
979 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
980 /* we were freed in flight; disregard any error */
981 arc_release(buf
, db
);
982 bzero(buf
->b_data
, db
->db
.db_size
);
984 db
->db_freed_in_flight
= FALSE
;
985 dbuf_set_data(db
, buf
);
986 db
->db_state
= DB_CACHED
;
987 } else if (zio
== NULL
|| zio
->io_error
== 0) {
988 dbuf_set_data(db
, buf
);
989 db
->db_state
= DB_CACHED
;
991 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
992 ASSERT3P(db
->db_buf
, ==, NULL
);
993 arc_buf_destroy(buf
, db
);
994 db
->db_state
= DB_UNCACHED
;
996 cv_broadcast(&db
->db_changed
);
997 dbuf_rele_and_unlock(db
, NULL
);
1001 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1004 zbookmark_phys_t zb
;
1005 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1010 ASSERT(!refcount_is_zero(&db
->db_holds
));
1011 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1012 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1013 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1014 ASSERT(db
->db_state
== DB_UNCACHED
);
1015 ASSERT(db
->db_buf
== NULL
);
1017 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1019 * The bonus length stored in the dnode may be less than
1020 * the maximum available space in the bonus buffer.
1022 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1023 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1025 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1026 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1027 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1028 if (bonuslen
< max_bonuslen
)
1029 bzero(db
->db
.db_data
, max_bonuslen
);
1031 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1033 db
->db_state
= DB_CACHED
;
1034 mutex_exit(&db
->db_mtx
);
1039 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1040 * processes the delete record and clears the bp while we are waiting
1041 * for the dn_mtx (resulting in a "no" from block_freed).
1043 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1044 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1045 BP_IS_HOLE(db
->db_blkptr
)))) {
1046 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1048 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1050 bzero(db
->db
.db_data
, db
->db
.db_size
);
1052 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1053 BP_IS_HOLE(db
->db_blkptr
) &&
1054 db
->db_blkptr
->blk_birth
!= 0) {
1055 blkptr_t
*bps
= db
->db
.db_data
;
1057 for (i
= 0; i
< ((1 <<
1058 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1060 blkptr_t
*bp
= &bps
[i
];
1061 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1062 1 << dn
->dn_indblkshift
);
1064 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1066 BP_GET_LSIZE(db
->db_blkptr
));
1067 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1069 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1070 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1074 db
->db_state
= DB_CACHED
;
1075 mutex_exit(&db
->db_mtx
);
1081 db
->db_state
= DB_READ
;
1082 mutex_exit(&db
->db_mtx
);
1084 if (DBUF_IS_L2CACHEABLE(db
))
1085 aflags
|= ARC_FLAG_L2CACHE
;
1087 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
1088 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
1089 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1091 dbuf_add_ref(db
, NULL
);
1093 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1094 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
1095 (flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
1102 * This is our just-in-time copy function. It makes a copy of buffers that
1103 * have been modified in a previous transaction group before we access them in
1104 * the current active group.
1106 * This function is used in three places: when we are dirtying a buffer for the
1107 * first time in a txg, when we are freeing a range in a dnode that includes
1108 * this buffer, and when we are accessing a buffer which was received compressed
1109 * and later referenced in a WRITE_BYREF record.
1111 * Note that when we are called from dbuf_free_range() we do not put a hold on
1112 * the buffer, we just traverse the active dbuf list for the dnode.
1115 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1117 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1119 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1120 ASSERT(db
->db
.db_data
!= NULL
);
1121 ASSERT(db
->db_level
== 0);
1122 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1125 (dr
->dt
.dl
.dr_data
!=
1126 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1130 * If the last dirty record for this dbuf has not yet synced
1131 * and its referencing the dbuf data, either:
1132 * reset the reference to point to a new copy,
1133 * or (if there a no active holders)
1134 * just null out the current db_data pointer.
1136 ASSERT(dr
->dr_txg
>= txg
- 2);
1137 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1138 dnode_t
*dn
= DB_DNODE(db
);
1139 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1140 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1141 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1142 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1143 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1144 int size
= arc_buf_size(db
->db_buf
);
1145 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1146 spa_t
*spa
= db
->db_objset
->os_spa
;
1147 enum zio_compress compress_type
=
1148 arc_get_compression(db
->db_buf
);
1150 if (compress_type
== ZIO_COMPRESS_OFF
) {
1151 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1153 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1154 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1155 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1157 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1160 dbuf_clear_data(db
);
1165 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1172 * We don't have to hold the mutex to check db_state because it
1173 * can't be freed while we have a hold on the buffer.
1175 ASSERT(!refcount_is_zero(&db
->db_holds
));
1177 if (db
->db_state
== DB_NOFILL
)
1178 return (SET_ERROR(EIO
));
1182 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1183 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1185 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1186 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1187 DBUF_IS_CACHEABLE(db
);
1189 mutex_enter(&db
->db_mtx
);
1190 if (db
->db_state
== DB_CACHED
) {
1192 * If the arc buf is compressed, we need to decompress it to
1193 * read the data. This could happen during the "zfs receive" of
1194 * a stream which is compressed and deduplicated.
1196 if (db
->db_buf
!= NULL
&&
1197 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
) {
1198 dbuf_fix_old_data(db
,
1199 spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1200 err
= arc_decompress(db
->db_buf
);
1201 dbuf_set_data(db
, db
->db_buf
);
1203 mutex_exit(&db
->db_mtx
);
1205 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1206 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1207 rw_exit(&dn
->dn_struct_rwlock
);
1209 } else if (db
->db_state
== DB_UNCACHED
) {
1210 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1211 boolean_t need_wait
= B_FALSE
;
1214 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1215 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1218 err
= dbuf_read_impl(db
, zio
, flags
);
1220 /* dbuf_read_impl has dropped db_mtx for us */
1222 if (!err
&& prefetch
)
1223 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1225 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1226 rw_exit(&dn
->dn_struct_rwlock
);
1229 if (!err
&& need_wait
)
1230 err
= zio_wait(zio
);
1233 * Another reader came in while the dbuf was in flight
1234 * between UNCACHED and CACHED. Either a writer will finish
1235 * writing the buffer (sending the dbuf to CACHED) or the
1236 * first reader's request will reach the read_done callback
1237 * and send the dbuf to CACHED. Otherwise, a failure
1238 * occurred and the dbuf went to UNCACHED.
1240 mutex_exit(&db
->db_mtx
);
1242 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1243 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1244 rw_exit(&dn
->dn_struct_rwlock
);
1247 /* Skip the wait per the caller's request. */
1248 mutex_enter(&db
->db_mtx
);
1249 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1250 while (db
->db_state
== DB_READ
||
1251 db
->db_state
== DB_FILL
) {
1252 ASSERT(db
->db_state
== DB_READ
||
1253 (flags
& DB_RF_HAVESTRUCT
) == 0);
1254 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1256 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1258 if (db
->db_state
== DB_UNCACHED
)
1259 err
= SET_ERROR(EIO
);
1261 mutex_exit(&db
->db_mtx
);
1268 dbuf_noread(dmu_buf_impl_t
*db
)
1270 ASSERT(!refcount_is_zero(&db
->db_holds
));
1271 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1272 mutex_enter(&db
->db_mtx
);
1273 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1274 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1275 if (db
->db_state
== DB_UNCACHED
) {
1276 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1277 spa_t
*spa
= db
->db_objset
->os_spa
;
1279 ASSERT(db
->db_buf
== NULL
);
1280 ASSERT(db
->db
.db_data
== NULL
);
1281 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1282 db
->db_state
= DB_FILL
;
1283 } else if (db
->db_state
== DB_NOFILL
) {
1284 dbuf_clear_data(db
);
1286 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1288 mutex_exit(&db
->db_mtx
);
1292 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1294 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1295 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1296 uint64_t txg
= dr
->dr_txg
;
1298 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1300 * This assert is valid because dmu_sync() expects to be called by
1301 * a zilog's get_data while holding a range lock. This call only
1302 * comes from dbuf_dirty() callers who must also hold a range lock.
1304 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1305 ASSERT(db
->db_level
== 0);
1307 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1308 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1311 ASSERT(db
->db_data_pending
!= dr
);
1313 /* free this block */
1314 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1315 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1317 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1318 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1321 * Release the already-written buffer, so we leave it in
1322 * a consistent dirty state. Note that all callers are
1323 * modifying the buffer, so they will immediately do
1324 * another (redundant) arc_release(). Therefore, leave
1325 * the buf thawed to save the effort of freezing &
1326 * immediately re-thawing it.
1328 arc_release(dr
->dt
.dl
.dr_data
, db
);
1332 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1333 * data blocks in the free range, so that any future readers will find
1337 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1340 dmu_buf_impl_t
*db_search
;
1341 dmu_buf_impl_t
*db
, *db_next
;
1342 uint64_t txg
= tx
->tx_txg
;
1345 if (end_blkid
> dn
->dn_maxblkid
&&
1346 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1347 end_blkid
= dn
->dn_maxblkid
;
1348 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1350 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1351 db_search
->db_level
= 0;
1352 db_search
->db_blkid
= start_blkid
;
1353 db_search
->db_state
= DB_SEARCH
;
1355 mutex_enter(&dn
->dn_dbufs_mtx
);
1356 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1357 ASSERT3P(db
, ==, NULL
);
1359 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1361 for (; db
!= NULL
; db
= db_next
) {
1362 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1363 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1365 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1368 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1370 /* found a level 0 buffer in the range */
1371 mutex_enter(&db
->db_mtx
);
1372 if (dbuf_undirty(db
, tx
)) {
1373 /* mutex has been dropped and dbuf destroyed */
1377 if (db
->db_state
== DB_UNCACHED
||
1378 db
->db_state
== DB_NOFILL
||
1379 db
->db_state
== DB_EVICTING
) {
1380 ASSERT(db
->db
.db_data
== NULL
);
1381 mutex_exit(&db
->db_mtx
);
1384 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1385 /* will be handled in dbuf_read_done or dbuf_rele */
1386 db
->db_freed_in_flight
= TRUE
;
1387 mutex_exit(&db
->db_mtx
);
1390 if (refcount_count(&db
->db_holds
) == 0) {
1395 /* The dbuf is referenced */
1397 if (db
->db_last_dirty
!= NULL
) {
1398 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1400 if (dr
->dr_txg
== txg
) {
1402 * This buffer is "in-use", re-adjust the file
1403 * size to reflect that this buffer may
1404 * contain new data when we sync.
1406 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1407 db
->db_blkid
> dn
->dn_maxblkid
)
1408 dn
->dn_maxblkid
= db
->db_blkid
;
1409 dbuf_unoverride(dr
);
1412 * This dbuf is not dirty in the open context.
1413 * Either uncache it (if its not referenced in
1414 * the open context) or reset its contents to
1417 dbuf_fix_old_data(db
, txg
);
1420 /* clear the contents if its cached */
1421 if (db
->db_state
== DB_CACHED
) {
1422 ASSERT(db
->db
.db_data
!= NULL
);
1423 arc_release(db
->db_buf
, db
);
1424 bzero(db
->db
.db_data
, db
->db
.db_size
);
1425 arc_buf_freeze(db
->db_buf
);
1428 mutex_exit(&db
->db_mtx
);
1431 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1432 mutex_exit(&dn
->dn_dbufs_mtx
);
1436 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1438 arc_buf_t
*buf
, *obuf
;
1439 int osize
= db
->db
.db_size
;
1440 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1443 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1448 /* XXX does *this* func really need the lock? */
1449 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1452 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1453 * is OK, because there can be no other references to the db
1454 * when we are changing its size, so no concurrent DB_FILL can
1458 * XXX we should be doing a dbuf_read, checking the return
1459 * value and returning that up to our callers
1461 dmu_buf_will_dirty(&db
->db
, tx
);
1463 /* create the data buffer for the new block */
1464 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1466 /* copy old block data to the new block */
1468 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1469 /* zero the remainder */
1471 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1473 mutex_enter(&db
->db_mtx
);
1474 dbuf_set_data(db
, buf
);
1475 arc_buf_destroy(obuf
, db
);
1476 db
->db
.db_size
= size
;
1478 if (db
->db_level
== 0) {
1479 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1480 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1482 mutex_exit(&db
->db_mtx
);
1484 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1489 dbuf_release_bp(dmu_buf_impl_t
*db
)
1491 ASSERTV(objset_t
*os
= db
->db_objset
);
1493 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1494 ASSERT(arc_released(os
->os_phys_buf
) ||
1495 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1496 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1498 (void) arc_release(db
->db_buf
, db
);
1502 * We already have a dirty record for this TXG, and we are being
1506 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1508 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1510 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1512 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1514 * If this buffer has already been written out,
1515 * we now need to reset its state.
1517 dbuf_unoverride(dr
);
1518 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1519 db
->db_state
!= DB_NOFILL
) {
1520 /* Already released on initial dirty, so just thaw. */
1521 ASSERT(arc_released(db
->db_buf
));
1522 arc_buf_thaw(db
->db_buf
);
1527 dbuf_dirty_record_t
*
1528 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1532 dbuf_dirty_record_t
**drp
, *dr
;
1533 int drop_struct_lock
= FALSE
;
1534 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1536 ASSERT(tx
->tx_txg
!= 0);
1537 ASSERT(!refcount_is_zero(&db
->db_holds
));
1538 DMU_TX_DIRTY_BUF(tx
, db
);
1543 * Shouldn't dirty a regular buffer in syncing context. Private
1544 * objects may be dirtied in syncing context, but only if they
1545 * were already pre-dirtied in open context.
1548 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1549 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1552 ASSERT(!dmu_tx_is_syncing(tx
) ||
1553 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1554 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1555 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1556 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1557 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1560 * We make this assert for private objects as well, but after we
1561 * check if we're already dirty. They are allowed to re-dirty
1562 * in syncing context.
1564 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1565 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1566 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1568 mutex_enter(&db
->db_mtx
);
1570 * XXX make this true for indirects too? The problem is that
1571 * transactions created with dmu_tx_create_assigned() from
1572 * syncing context don't bother holding ahead.
1574 ASSERT(db
->db_level
!= 0 ||
1575 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1576 db
->db_state
== DB_NOFILL
);
1578 mutex_enter(&dn
->dn_mtx
);
1580 * Don't set dirtyctx to SYNC if we're just modifying this as we
1581 * initialize the objset.
1583 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1584 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1585 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1588 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1589 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1590 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1591 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1592 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1594 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1595 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1599 mutex_exit(&dn
->dn_mtx
);
1601 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1602 dn
->dn_have_spill
= B_TRUE
;
1605 * If this buffer is already dirty, we're done.
1607 drp
= &db
->db_last_dirty
;
1608 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1609 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1610 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1612 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1616 mutex_exit(&db
->db_mtx
);
1621 * Only valid if not already dirty.
1623 ASSERT(dn
->dn_object
== 0 ||
1624 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1625 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1627 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1630 * We should only be dirtying in syncing context if it's the
1631 * mos or we're initializing the os or it's a special object.
1632 * However, we are allowed to dirty in syncing context provided
1633 * we already dirtied it in open context. Hence we must make
1634 * this assertion only if we're not already dirty.
1637 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
1639 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1640 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
1641 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1642 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1643 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1644 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1646 ASSERT(db
->db
.db_size
!= 0);
1648 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1650 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1651 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
1655 * If this buffer is dirty in an old transaction group we need
1656 * to make a copy of it so that the changes we make in this
1657 * transaction group won't leak out when we sync the older txg.
1659 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1660 list_link_init(&dr
->dr_dirty_node
);
1661 if (db
->db_level
== 0) {
1662 void *data_old
= db
->db_buf
;
1664 if (db
->db_state
!= DB_NOFILL
) {
1665 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1666 dbuf_fix_old_data(db
, tx
->tx_txg
);
1667 data_old
= db
->db
.db_data
;
1668 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1670 * Release the data buffer from the cache so
1671 * that we can modify it without impacting
1672 * possible other users of this cached data
1673 * block. Note that indirect blocks and
1674 * private objects are not released until the
1675 * syncing state (since they are only modified
1678 arc_release(db
->db_buf
, db
);
1679 dbuf_fix_old_data(db
, tx
->tx_txg
);
1680 data_old
= db
->db_buf
;
1682 ASSERT(data_old
!= NULL
);
1684 dr
->dt
.dl
.dr_data
= data_old
;
1686 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1687 list_create(&dr
->dt
.di
.dr_children
,
1688 sizeof (dbuf_dirty_record_t
),
1689 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1691 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1692 dr
->dr_accounted
= db
->db
.db_size
;
1694 dr
->dr_txg
= tx
->tx_txg
;
1699 * We could have been freed_in_flight between the dbuf_noread
1700 * and dbuf_dirty. We win, as though the dbuf_noread() had
1701 * happened after the free.
1703 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1704 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1705 mutex_enter(&dn
->dn_mtx
);
1706 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1707 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1710 mutex_exit(&dn
->dn_mtx
);
1711 db
->db_freed_in_flight
= FALSE
;
1715 * This buffer is now part of this txg
1717 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1718 db
->db_dirtycnt
+= 1;
1719 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1721 mutex_exit(&db
->db_mtx
);
1723 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1724 db
->db_blkid
== DMU_SPILL_BLKID
) {
1725 mutex_enter(&dn
->dn_mtx
);
1726 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1727 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1728 mutex_exit(&dn
->dn_mtx
);
1729 dnode_setdirty(dn
, tx
);
1735 * The dn_struct_rwlock prevents db_blkptr from changing
1736 * due to a write from syncing context completing
1737 * while we are running, so we want to acquire it before
1738 * looking at db_blkptr.
1740 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1741 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1742 drop_struct_lock
= TRUE
;
1746 * We need to hold the dn_struct_rwlock to make this assertion,
1747 * because it protects dn_phys / dn_next_nlevels from changing.
1749 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1750 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1751 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1752 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1753 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1756 * If we are overwriting a dedup BP, then unless it is snapshotted,
1757 * when we get to syncing context we will need to decrement its
1758 * refcount in the DDT. Prefetch the relevant DDT block so that
1759 * syncing context won't have to wait for the i/o.
1761 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
1763 if (db
->db_level
== 0) {
1764 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1765 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1768 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1769 dmu_buf_impl_t
*parent
= db
->db_parent
;
1770 dbuf_dirty_record_t
*di
;
1771 int parent_held
= FALSE
;
1773 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1774 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1776 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1777 db
->db_blkid
>> epbs
, FTAG
);
1778 ASSERT(parent
!= NULL
);
1781 if (drop_struct_lock
)
1782 rw_exit(&dn
->dn_struct_rwlock
);
1783 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1784 di
= dbuf_dirty(parent
, tx
);
1786 dbuf_rele(parent
, FTAG
);
1788 mutex_enter(&db
->db_mtx
);
1790 * Since we've dropped the mutex, it's possible that
1791 * dbuf_undirty() might have changed this out from under us.
1793 if (db
->db_last_dirty
== dr
||
1794 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1795 mutex_enter(&di
->dt
.di
.dr_mtx
);
1796 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1797 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1798 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1799 mutex_exit(&di
->dt
.di
.dr_mtx
);
1802 mutex_exit(&db
->db_mtx
);
1804 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1805 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1806 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1807 mutex_enter(&dn
->dn_mtx
);
1808 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1809 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1810 mutex_exit(&dn
->dn_mtx
);
1811 if (drop_struct_lock
)
1812 rw_exit(&dn
->dn_struct_rwlock
);
1815 dnode_setdirty(dn
, tx
);
1821 * Undirty a buffer in the transaction group referenced by the given
1822 * transaction. Return whether this evicted the dbuf.
1825 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1828 uint64_t txg
= tx
->tx_txg
;
1829 dbuf_dirty_record_t
*dr
, **drp
;
1834 * Due to our use of dn_nlevels below, this can only be called
1835 * in open context, unless we are operating on the MOS.
1836 * From syncing context, dn_nlevels may be different from the
1837 * dn_nlevels used when dbuf was dirtied.
1839 ASSERT(db
->db_objset
==
1840 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1841 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1842 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1843 ASSERT0(db
->db_level
);
1844 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1847 * If this buffer is not dirty, we're done.
1849 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1850 if (dr
->dr_txg
<= txg
)
1852 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1854 ASSERT(dr
->dr_txg
== txg
);
1855 ASSERT(dr
->dr_dbuf
== db
);
1860 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1862 ASSERT(db
->db
.db_size
!= 0);
1864 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1865 dr
->dr_accounted
, txg
);
1870 * Note that there are three places in dbuf_dirty()
1871 * where this dirty record may be put on a list.
1872 * Make sure to do a list_remove corresponding to
1873 * every one of those list_insert calls.
1875 if (dr
->dr_parent
) {
1876 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1877 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1878 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1879 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1880 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1881 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1882 mutex_enter(&dn
->dn_mtx
);
1883 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1884 mutex_exit(&dn
->dn_mtx
);
1888 if (db
->db_state
!= DB_NOFILL
) {
1889 dbuf_unoverride(dr
);
1891 ASSERT(db
->db_buf
!= NULL
);
1892 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1893 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1894 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
1897 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1899 ASSERT(db
->db_dirtycnt
> 0);
1900 db
->db_dirtycnt
-= 1;
1902 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1903 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
1912 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1914 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1915 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1916 dbuf_dirty_record_t
*dr
;
1918 ASSERT(tx
->tx_txg
!= 0);
1919 ASSERT(!refcount_is_zero(&db
->db_holds
));
1922 * Quick check for dirtyness. For already dirty blocks, this
1923 * reduces runtime of this function by >90%, and overall performance
1924 * by 50% for some workloads (e.g. file deletion with indirect blocks
1927 mutex_enter(&db
->db_mtx
);
1929 for (dr
= db
->db_last_dirty
;
1930 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1932 * It's possible that it is already dirty but not cached,
1933 * because there are some calls to dbuf_dirty() that don't
1934 * go through dmu_buf_will_dirty().
1936 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1937 /* This dbuf is already dirty and cached. */
1939 mutex_exit(&db
->db_mtx
);
1943 mutex_exit(&db
->db_mtx
);
1946 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1947 rf
|= DB_RF_HAVESTRUCT
;
1949 (void) dbuf_read(db
, NULL
, rf
);
1950 (void) dbuf_dirty(db
, tx
);
1954 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1956 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1958 db
->db_state
= DB_NOFILL
;
1960 dmu_buf_will_fill(db_fake
, tx
);
1964 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1966 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1968 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1969 ASSERT(tx
->tx_txg
!= 0);
1970 ASSERT(db
->db_level
== 0);
1971 ASSERT(!refcount_is_zero(&db
->db_holds
));
1973 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1974 dmu_tx_private_ok(tx
));
1977 (void) dbuf_dirty(db
, tx
);
1980 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1983 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1985 mutex_enter(&db
->db_mtx
);
1988 if (db
->db_state
== DB_FILL
) {
1989 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1990 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1991 /* we were freed while filling */
1992 /* XXX dbuf_undirty? */
1993 bzero(db
->db
.db_data
, db
->db
.db_size
);
1994 db
->db_freed_in_flight
= FALSE
;
1996 db
->db_state
= DB_CACHED
;
1997 cv_broadcast(&db
->db_changed
);
1999 mutex_exit(&db
->db_mtx
);
2003 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2004 bp_embedded_type_t etype
, enum zio_compress comp
,
2005 int uncompressed_size
, int compressed_size
, int byteorder
,
2008 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2009 struct dirty_leaf
*dl
;
2010 dmu_object_type_t type
;
2012 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2013 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2014 SPA_FEATURE_EMBEDDED_DATA
));
2018 type
= DB_DNODE(db
)->dn_type
;
2021 ASSERT0(db
->db_level
);
2022 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2024 dmu_buf_will_not_fill(dbuf
, tx
);
2026 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2027 dl
= &db
->db_last_dirty
->dt
.dl
;
2028 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2029 data
, comp
, uncompressed_size
, compressed_size
);
2030 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2031 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2032 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2033 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2035 dl
->dr_override_state
= DR_OVERRIDDEN
;
2036 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2040 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2041 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2044 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2046 ASSERT(!refcount_is_zero(&db
->db_holds
));
2047 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2048 ASSERT(db
->db_level
== 0);
2049 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2050 ASSERT(buf
!= NULL
);
2051 ASSERT(arc_buf_lsize(buf
) == db
->db
.db_size
);
2052 ASSERT(tx
->tx_txg
!= 0);
2054 arc_return_buf(buf
, db
);
2055 ASSERT(arc_released(buf
));
2057 mutex_enter(&db
->db_mtx
);
2059 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2060 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2062 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2064 if (db
->db_state
== DB_CACHED
&&
2065 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2066 mutex_exit(&db
->db_mtx
);
2067 (void) dbuf_dirty(db
, tx
);
2068 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2069 arc_buf_destroy(buf
, db
);
2070 xuio_stat_wbuf_copied();
2074 xuio_stat_wbuf_nocopy();
2075 if (db
->db_state
== DB_CACHED
) {
2076 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2078 ASSERT(db
->db_buf
!= NULL
);
2079 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2080 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2081 if (!arc_released(db
->db_buf
)) {
2082 ASSERT(dr
->dt
.dl
.dr_override_state
==
2084 arc_release(db
->db_buf
, db
);
2086 dr
->dt
.dl
.dr_data
= buf
;
2087 arc_buf_destroy(db
->db_buf
, db
);
2088 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2089 arc_release(db
->db_buf
, db
);
2090 arc_buf_destroy(db
->db_buf
, db
);
2094 ASSERT(db
->db_buf
== NULL
);
2095 dbuf_set_data(db
, buf
);
2096 db
->db_state
= DB_FILL
;
2097 mutex_exit(&db
->db_mtx
);
2098 (void) dbuf_dirty(db
, tx
);
2099 dmu_buf_fill_done(&db
->db
, tx
);
2103 dbuf_destroy(dmu_buf_impl_t
*db
)
2106 dmu_buf_impl_t
*parent
= db
->db_parent
;
2107 dmu_buf_impl_t
*dndb
;
2109 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2110 ASSERT(refcount_is_zero(&db
->db_holds
));
2112 if (db
->db_buf
!= NULL
) {
2113 arc_buf_destroy(db
->db_buf
, db
);
2117 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2118 int slots
= DB_DNODE(db
)->dn_num_slots
;
2119 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2120 ASSERT(db
->db
.db_data
!= NULL
);
2121 kmem_free(db
->db
.db_data
, bonuslen
);
2122 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2123 db
->db_state
= DB_UNCACHED
;
2126 dbuf_clear_data(db
);
2128 if (multilist_link_active(&db
->db_cache_link
)) {
2129 multilist_remove(dbuf_cache
, db
);
2130 (void) refcount_remove_many(&dbuf_cache_size
,
2131 db
->db
.db_size
, db
);
2134 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2135 ASSERT(db
->db_data_pending
== NULL
);
2137 db
->db_state
= DB_EVICTING
;
2138 db
->db_blkptr
= NULL
;
2141 * Now that db_state is DB_EVICTING, nobody else can find this via
2142 * the hash table. We can now drop db_mtx, which allows us to
2143 * acquire the dn_dbufs_mtx.
2145 mutex_exit(&db
->db_mtx
);
2150 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2151 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2153 mutex_enter(&dn
->dn_dbufs_mtx
);
2154 avl_remove(&dn
->dn_dbufs
, db
);
2155 atomic_dec_32(&dn
->dn_dbufs_count
);
2159 mutex_exit(&dn
->dn_dbufs_mtx
);
2161 * Decrementing the dbuf count means that the hold corresponding
2162 * to the removed dbuf is no longer discounted in dnode_move(),
2163 * so the dnode cannot be moved until after we release the hold.
2164 * The membar_producer() ensures visibility of the decremented
2165 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2169 db
->db_dnode_handle
= NULL
;
2171 dbuf_hash_remove(db
);
2176 ASSERT(refcount_is_zero(&db
->db_holds
));
2178 db
->db_parent
= NULL
;
2180 ASSERT(db
->db_buf
== NULL
);
2181 ASSERT(db
->db
.db_data
== NULL
);
2182 ASSERT(db
->db_hash_next
== NULL
);
2183 ASSERT(db
->db_blkptr
== NULL
);
2184 ASSERT(db
->db_data_pending
== NULL
);
2185 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2187 kmem_cache_free(dbuf_kmem_cache
, db
);
2188 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2191 * If this dbuf is referenced from an indirect dbuf,
2192 * decrement the ref count on the indirect dbuf.
2194 if (parent
&& parent
!= dndb
)
2195 dbuf_rele(parent
, db
);
2199 * Note: While bpp will always be updated if the function returns success,
2200 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2201 * this happens when the dnode is the meta-dnode, or a userused or groupused
2204 __attribute__((always_inline
))
2206 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2207 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
2214 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2216 if (blkid
== DMU_SPILL_BLKID
) {
2217 mutex_enter(&dn
->dn_mtx
);
2218 if (dn
->dn_have_spill
&&
2219 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2220 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2223 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2224 *parentp
= dn
->dn_dbuf
;
2225 mutex_exit(&dn
->dn_mtx
);
2230 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2231 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2233 ASSERT3U(level
* epbs
, <, 64);
2234 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2236 * This assertion shouldn't trip as long as the max indirect block size
2237 * is less than 1M. The reason for this is that up to that point,
2238 * the number of levels required to address an entire object with blocks
2239 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2240 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2241 * (i.e. we can address the entire object), objects will all use at most
2242 * N-1 levels and the assertion won't overflow. However, once epbs is
2243 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2244 * enough to address an entire object, so objects will have 5 levels,
2245 * but then this assertion will overflow.
2247 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2248 * need to redo this logic to handle overflows.
2250 ASSERT(level
>= nlevels
||
2251 ((nlevels
- level
- 1) * epbs
) +
2252 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2253 if (level
>= nlevels
||
2254 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2255 ((nlevels
- level
- 1) * epbs
)) ||
2257 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2258 /* the buffer has no parent yet */
2259 return (SET_ERROR(ENOENT
));
2260 } else if (level
< nlevels
-1) {
2261 /* this block is referenced from an indirect block */
2264 err
= dbuf_hold_impl(dn
, level
+1,
2265 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2267 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
2268 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
2269 parentp
, dh
->dh_depth
+ 1);
2270 err
= __dbuf_hold_impl(dh
+ 1);
2274 err
= dbuf_read(*parentp
, NULL
,
2275 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2277 dbuf_rele(*parentp
, NULL
);
2281 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2282 (blkid
& ((1ULL << epbs
) - 1));
2283 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2284 ASSERT(BP_IS_HOLE(*bpp
));
2287 /* the block is referenced from the dnode */
2288 ASSERT3U(level
, ==, nlevels
-1);
2289 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2290 blkid
< dn
->dn_phys
->dn_nblkptr
);
2292 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2293 *parentp
= dn
->dn_dbuf
;
2295 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2300 static dmu_buf_impl_t
*
2301 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2302 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2304 objset_t
*os
= dn
->dn_objset
;
2305 dmu_buf_impl_t
*db
, *odb
;
2307 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2308 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2310 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2313 db
->db
.db_object
= dn
->dn_object
;
2314 db
->db_level
= level
;
2315 db
->db_blkid
= blkid
;
2316 db
->db_last_dirty
= NULL
;
2317 db
->db_dirtycnt
= 0;
2318 db
->db_dnode_handle
= dn
->dn_handle
;
2319 db
->db_parent
= parent
;
2320 db
->db_blkptr
= blkptr
;
2323 db
->db_user_immediate_evict
= FALSE
;
2324 db
->db_freed_in_flight
= FALSE
;
2325 db
->db_pending_evict
= FALSE
;
2327 if (blkid
== DMU_BONUS_BLKID
) {
2328 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2329 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2330 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2331 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2332 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2333 db
->db_state
= DB_UNCACHED
;
2334 /* the bonus dbuf is not placed in the hash table */
2335 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2337 } else if (blkid
== DMU_SPILL_BLKID
) {
2338 db
->db
.db_size
= (blkptr
!= NULL
) ?
2339 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2340 db
->db
.db_offset
= 0;
2343 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2344 db
->db
.db_size
= blocksize
;
2345 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2349 * Hold the dn_dbufs_mtx while we get the new dbuf
2350 * in the hash table *and* added to the dbufs list.
2351 * This prevents a possible deadlock with someone
2352 * trying to look up this dbuf before its added to the
2355 mutex_enter(&dn
->dn_dbufs_mtx
);
2356 db
->db_state
= DB_EVICTING
;
2357 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2358 /* someone else inserted it first */
2359 kmem_cache_free(dbuf_kmem_cache
, db
);
2360 mutex_exit(&dn
->dn_dbufs_mtx
);
2363 avl_add(&dn
->dn_dbufs
, db
);
2365 db
->db_state
= DB_UNCACHED
;
2366 mutex_exit(&dn
->dn_dbufs_mtx
);
2367 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2369 if (parent
&& parent
!= dn
->dn_dbuf
)
2370 dbuf_add_ref(parent
, db
);
2372 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2373 refcount_count(&dn
->dn_holds
) > 0);
2374 (void) refcount_add(&dn
->dn_holds
, db
);
2375 atomic_inc_32(&dn
->dn_dbufs_count
);
2377 dprintf_dbuf(db
, "db=%p\n", db
);
2382 typedef struct dbuf_prefetch_arg
{
2383 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2384 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2385 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2386 int dpa_curlevel
; /* The current level that we're reading */
2387 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2388 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2389 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2390 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2391 } dbuf_prefetch_arg_t
;
2394 * Actually issue the prefetch read for the block given.
2397 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2400 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2403 aflags
= dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2405 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2406 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2407 ASSERT(dpa
->dpa_zio
!= NULL
);
2408 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2409 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2410 &aflags
, &dpa
->dpa_zb
);
2414 * Called when an indirect block above our prefetch target is read in. This
2415 * will either read in the next indirect block down the tree or issue the actual
2416 * prefetch if the next block down is our target.
2419 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2421 dbuf_prefetch_arg_t
*dpa
= private;
2425 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2426 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2429 * The dpa_dnode is only valid if we are called with a NULL
2430 * zio. This indicates that the arc_read() returned without
2431 * first calling zio_read() to issue a physical read. Once
2432 * a physical read is made the dpa_dnode must be invalidated
2433 * as the locks guarding it may have been dropped. If the
2434 * dpa_dnode is still valid, then we want to add it to the dbuf
2435 * cache. To do so, we must hold the dbuf associated with the block
2436 * we just prefetched, read its contents so that we associate it
2437 * with an arc_buf_t, and then release it.
2440 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2441 if (zio
->io_flags
& ZIO_FLAG_RAW
) {
2442 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2444 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2446 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2448 dpa
->dpa_dnode
= NULL
;
2449 } else if (dpa
->dpa_dnode
!= NULL
) {
2450 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2451 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2452 dpa
->dpa_zb
.zb_level
));
2453 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2454 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2455 (void) dbuf_read(db
, NULL
,
2456 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2457 dbuf_rele(db
, FTAG
);
2460 dpa
->dpa_curlevel
--;
2462 nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2463 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2464 bp
= ((blkptr_t
*)abuf
->b_data
) +
2465 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2466 if (BP_IS_HOLE(bp
) || (zio
!= NULL
&& zio
->io_error
!= 0)) {
2467 kmem_free(dpa
, sizeof (*dpa
));
2468 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2469 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2470 dbuf_issue_final_prefetch(dpa
, bp
);
2471 kmem_free(dpa
, sizeof (*dpa
));
2473 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2474 zbookmark_phys_t zb
;
2476 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2478 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2479 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2481 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2482 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2483 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2487 arc_buf_destroy(abuf
, private);
2491 * Issue prefetch reads for the given block on the given level. If the indirect
2492 * blocks above that block are not in memory, we will read them in
2493 * asynchronously. As a result, this call never blocks waiting for a read to
2497 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2501 int epbs
, nlevels
, curlevel
;
2505 dbuf_prefetch_arg_t
*dpa
;
2508 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2509 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2511 if (blkid
> dn
->dn_maxblkid
)
2514 if (dnode_block_freed(dn
, blkid
))
2518 * This dnode hasn't been written to disk yet, so there's nothing to
2521 nlevels
= dn
->dn_phys
->dn_nlevels
;
2522 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2525 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2526 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2529 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2532 mutex_exit(&db
->db_mtx
);
2534 * This dbuf already exists. It is either CACHED, or
2535 * (we assume) about to be read or filled.
2541 * Find the closest ancestor (indirect block) of the target block
2542 * that is present in the cache. In this indirect block, we will
2543 * find the bp that is at curlevel, curblkid.
2547 while (curlevel
< nlevels
- 1) {
2548 int parent_level
= curlevel
+ 1;
2549 uint64_t parent_blkid
= curblkid
>> epbs
;
2552 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2553 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2554 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2555 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2556 dbuf_rele(db
, FTAG
);
2560 curlevel
= parent_level
;
2561 curblkid
= parent_blkid
;
2564 if (curlevel
== nlevels
- 1) {
2565 /* No cached indirect blocks found. */
2566 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2567 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2569 if (BP_IS_HOLE(&bp
))
2572 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2574 pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2577 dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2578 ds
= dn
->dn_objset
->os_dsl_dataset
;
2579 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2580 dn
->dn_object
, level
, blkid
);
2581 dpa
->dpa_curlevel
= curlevel
;
2582 dpa
->dpa_prio
= prio
;
2583 dpa
->dpa_aflags
= aflags
;
2584 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2585 dpa
->dpa_dnode
= dn
;
2586 dpa
->dpa_epbs
= epbs
;
2590 * If we have the indirect just above us, no need to do the asynchronous
2591 * prefetch chain; we'll just run the last step ourselves. If we're at
2592 * a higher level, though, we want to issue the prefetches for all the
2593 * indirect blocks asynchronously, so we can go on with whatever we were
2596 if (curlevel
== level
) {
2597 ASSERT3U(curblkid
, ==, blkid
);
2598 dbuf_issue_final_prefetch(dpa
, &bp
);
2599 kmem_free(dpa
, sizeof (*dpa
));
2601 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2602 zbookmark_phys_t zb
;
2604 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2605 dn
->dn_object
, curlevel
, curblkid
);
2606 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2607 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2608 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2612 * We use pio here instead of dpa_zio since it's possible that
2613 * dpa may have already been freed.
2618 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2621 * Returns with db_holds incremented, and db_mtx not held.
2622 * Note: dn_struct_rwlock must be held.
2625 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2627 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2628 dh
->dh_parent
= NULL
;
2630 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2631 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2632 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2634 *(dh
->dh_dbp
) = NULL
;
2636 /* dbuf_find() returns with db_mtx held */
2637 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2638 dh
->dh_level
, dh
->dh_blkid
);
2640 if (dh
->dh_db
== NULL
) {
2643 if (dh
->dh_fail_uncached
)
2644 return (SET_ERROR(ENOENT
));
2646 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2647 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2648 dh
->dh_fail_sparse
, &dh
->dh_parent
, &dh
->dh_bp
, dh
);
2649 if (dh
->dh_fail_sparse
) {
2650 if (dh
->dh_err
== 0 &&
2651 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2652 dh
->dh_err
= SET_ERROR(ENOENT
);
2655 dbuf_rele(dh
->dh_parent
, NULL
);
2656 return (dh
->dh_err
);
2659 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2660 return (dh
->dh_err
);
2661 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2662 dh
->dh_parent
, dh
->dh_bp
);
2665 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2666 mutex_exit(&dh
->dh_db
->db_mtx
);
2667 return (SET_ERROR(ENOENT
));
2670 if (dh
->dh_db
->db_buf
!= NULL
)
2671 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2673 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2676 * If this buffer is currently syncing out, and we are are
2677 * still referencing it from db_data, we need to make a copy
2678 * of it in case we decide we want to dirty it again in this txg.
2680 if (dh
->dh_db
->db_level
== 0 &&
2681 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2682 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2683 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2684 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2686 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2687 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2689 dbuf_set_data(dh
->dh_db
,
2690 arc_alloc_buf(dh
->dh_dn
->dn_objset
->os_spa
,
2691 dh
->dh_db
, dh
->dh_type
, dh
->dh_db
->db
.db_size
));
2692 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2693 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2697 if (multilist_link_active(&dh
->dh_db
->db_cache_link
)) {
2698 ASSERT(refcount_is_zero(&dh
->dh_db
->db_holds
));
2699 multilist_remove(dbuf_cache
, dh
->dh_db
);
2700 (void) refcount_remove_many(&dbuf_cache_size
,
2701 dh
->dh_db
->db
.db_size
, dh
->dh_db
);
2703 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2704 DBUF_VERIFY(dh
->dh_db
);
2705 mutex_exit(&dh
->dh_db
->db_mtx
);
2707 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2709 dbuf_rele(dh
->dh_parent
, NULL
);
2711 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2712 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2713 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2714 *(dh
->dh_dbp
) = dh
->dh_db
;
2720 * The following code preserves the recursive function dbuf_hold_impl()
2721 * but moves the local variables AND function arguments to the heap to
2722 * minimize the stack frame size. Enough space is initially allocated
2723 * on the stack for 20 levels of recursion.
2726 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2727 boolean_t fail_sparse
, boolean_t fail_uncached
,
2728 void *tag
, dmu_buf_impl_t
**dbp
)
2730 struct dbuf_hold_impl_data
*dh
;
2733 dh
= kmem_alloc(sizeof (struct dbuf_hold_impl_data
) *
2734 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2735 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2736 fail_uncached
, tag
, dbp
, 0);
2738 error
= __dbuf_hold_impl(dh
);
2740 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2741 DBUF_HOLD_IMPL_MAX_DEPTH
);
2747 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2748 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2749 boolean_t fail_sparse
, boolean_t fail_uncached
,
2750 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2753 dh
->dh_level
= level
;
2754 dh
->dh_blkid
= blkid
;
2756 dh
->dh_fail_sparse
= fail_sparse
;
2757 dh
->dh_fail_uncached
= fail_uncached
;
2763 dh
->dh_parent
= NULL
;
2769 dh
->dh_depth
= depth
;
2773 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2775 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2779 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2782 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2783 return (err
? NULL
: db
);
2787 dbuf_create_bonus(dnode_t
*dn
)
2789 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2791 ASSERT(dn
->dn_bonus
== NULL
);
2792 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2796 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2798 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2801 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2802 return (SET_ERROR(ENOTSUP
));
2804 blksz
= SPA_MINBLOCKSIZE
;
2805 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2806 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2810 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2811 dbuf_new_size(db
, blksz
, tx
);
2812 rw_exit(&dn
->dn_struct_rwlock
);
2819 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2821 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2824 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2826 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2828 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
2829 VERIFY3S(holds
, >, 1);
2832 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2834 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2837 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2838 dmu_buf_impl_t
*found_db
;
2839 boolean_t result
= B_FALSE
;
2841 if (blkid
== DMU_BONUS_BLKID
)
2842 found_db
= dbuf_find_bonus(os
, obj
);
2844 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2846 if (found_db
!= NULL
) {
2847 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2848 (void) refcount_add(&db
->db_holds
, tag
);
2851 mutex_exit(&found_db
->db_mtx
);
2857 * If you call dbuf_rele() you had better not be referencing the dnode handle
2858 * unless you have some other direct or indirect hold on the dnode. (An indirect
2859 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2860 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2861 * dnode's parent dbuf evicting its dnode handles.
2864 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2866 mutex_enter(&db
->db_mtx
);
2867 dbuf_rele_and_unlock(db
, tag
);
2871 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2873 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2877 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2878 * db_dirtycnt and db_holds to be updated atomically.
2881 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2885 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2889 * Remove the reference to the dbuf before removing its hold on the
2890 * dnode so we can guarantee in dnode_move() that a referenced bonus
2891 * buffer has a corresponding dnode hold.
2893 holds
= refcount_remove(&db
->db_holds
, tag
);
2897 * We can't freeze indirects if there is a possibility that they
2898 * may be modified in the current syncing context.
2900 if (db
->db_buf
!= NULL
&&
2901 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
2902 arc_buf_freeze(db
->db_buf
);
2905 if (holds
== db
->db_dirtycnt
&&
2906 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2907 dbuf_evict_user(db
);
2910 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2912 boolean_t evict_dbuf
= db
->db_pending_evict
;
2915 * If the dnode moves here, we cannot cross this
2916 * barrier until the move completes.
2921 atomic_dec_32(&dn
->dn_dbufs_count
);
2924 * Decrementing the dbuf count means that the bonus
2925 * buffer's dnode hold is no longer discounted in
2926 * dnode_move(). The dnode cannot move until after
2927 * the dnode_rele() below.
2932 * Do not reference db after its lock is dropped.
2933 * Another thread may evict it.
2935 mutex_exit(&db
->db_mtx
);
2938 dnode_evict_bonus(dn
);
2941 } else if (db
->db_buf
== NULL
) {
2943 * This is a special case: we never associated this
2944 * dbuf with any data allocated from the ARC.
2946 ASSERT(db
->db_state
== DB_UNCACHED
||
2947 db
->db_state
== DB_NOFILL
);
2949 } else if (arc_released(db
->db_buf
)) {
2951 * This dbuf has anonymous data associated with it.
2955 boolean_t do_arc_evict
= B_FALSE
;
2957 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
2959 if (!DBUF_IS_CACHEABLE(db
) &&
2960 db
->db_blkptr
!= NULL
&&
2961 !BP_IS_HOLE(db
->db_blkptr
) &&
2962 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2963 do_arc_evict
= B_TRUE
;
2964 bp
= *db
->db_blkptr
;
2967 if (!DBUF_IS_CACHEABLE(db
) ||
2968 db
->db_pending_evict
) {
2970 } else if (!multilist_link_active(&db
->db_cache_link
)) {
2971 multilist_insert(dbuf_cache
, db
);
2972 (void) refcount_add_many(&dbuf_cache_size
,
2973 db
->db
.db_size
, db
);
2974 mutex_exit(&db
->db_mtx
);
2976 dbuf_evict_notify();
2980 arc_freed(spa
, &bp
);
2983 mutex_exit(&db
->db_mtx
);
2988 #pragma weak dmu_buf_refcount = dbuf_refcount
2990 dbuf_refcount(dmu_buf_impl_t
*db
)
2992 return (refcount_count(&db
->db_holds
));
2996 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
2997 dmu_buf_user_t
*new_user
)
2999 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3001 mutex_enter(&db
->db_mtx
);
3002 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3003 if (db
->db_user
== old_user
)
3004 db
->db_user
= new_user
;
3006 old_user
= db
->db_user
;
3007 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3008 mutex_exit(&db
->db_mtx
);
3014 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3016 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3020 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3022 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3024 db
->db_user_immediate_evict
= TRUE
;
3025 return (dmu_buf_set_user(db_fake
, user
));
3029 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3031 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3035 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3037 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3039 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3040 return (db
->db_user
);
3044 dmu_buf_user_evict_wait()
3046 taskq_wait(dbu_evict_taskq
);
3050 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3052 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3053 return (dbi
->db_blkptr
);
3057 dmu_buf_get_objset(dmu_buf_t
*db
)
3059 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3060 return (dbi
->db_objset
);
3064 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3066 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3067 DB_DNODE_ENTER(dbi
);
3068 return (DB_DNODE(dbi
));
3072 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3074 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3079 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3081 /* ASSERT(dmu_tx_is_syncing(tx) */
3082 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3084 if (db
->db_blkptr
!= NULL
)
3087 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3088 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3089 BP_ZERO(db
->db_blkptr
);
3092 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3094 * This buffer was allocated at a time when there was
3095 * no available blkptrs from the dnode, or it was
3096 * inappropriate to hook it in (i.e., nlevels mis-match).
3098 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3099 ASSERT(db
->db_parent
== NULL
);
3100 db
->db_parent
= dn
->dn_dbuf
;
3101 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3104 dmu_buf_impl_t
*parent
= db
->db_parent
;
3105 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3107 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3108 if (parent
== NULL
) {
3109 mutex_exit(&db
->db_mtx
);
3110 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3111 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3112 db
->db_blkid
>> epbs
, db
);
3113 rw_exit(&dn
->dn_struct_rwlock
);
3114 mutex_enter(&db
->db_mtx
);
3115 db
->db_parent
= parent
;
3117 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3118 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3124 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3125 * is critical the we not allow the compiler to inline this function in to
3126 * dbuf_sync_list() thereby drastically bloating the stack usage.
3128 noinline
static void
3129 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3131 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3135 ASSERT(dmu_tx_is_syncing(tx
));
3137 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3139 mutex_enter(&db
->db_mtx
);
3141 ASSERT(db
->db_level
> 0);
3144 /* Read the block if it hasn't been read yet. */
3145 if (db
->db_buf
== NULL
) {
3146 mutex_exit(&db
->db_mtx
);
3147 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3148 mutex_enter(&db
->db_mtx
);
3150 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3151 ASSERT(db
->db_buf
!= NULL
);
3155 /* Indirect block size must match what the dnode thinks it is. */
3156 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3157 dbuf_check_blkptr(dn
, db
);
3160 /* Provide the pending dirty record to child dbufs */
3161 db
->db_data_pending
= dr
;
3163 mutex_exit(&db
->db_mtx
);
3164 dbuf_write(dr
, db
->db_buf
, tx
);
3167 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3168 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3169 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3170 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3175 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3176 * critical the we not allow the compiler to inline this function in to
3177 * dbuf_sync_list() thereby drastically bloating the stack usage.
3179 noinline
static void
3180 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3182 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3183 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3186 uint64_t txg
= tx
->tx_txg
;
3188 ASSERT(dmu_tx_is_syncing(tx
));
3190 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3192 mutex_enter(&db
->db_mtx
);
3194 * To be synced, we must be dirtied. But we
3195 * might have been freed after the dirty.
3197 if (db
->db_state
== DB_UNCACHED
) {
3198 /* This buffer has been freed since it was dirtied */
3199 ASSERT(db
->db
.db_data
== NULL
);
3200 } else if (db
->db_state
== DB_FILL
) {
3201 /* This buffer was freed and is now being re-filled */
3202 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3204 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3211 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3212 mutex_enter(&dn
->dn_mtx
);
3213 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3215 * In the previous transaction group, the bonus buffer
3216 * was entirely used to store the attributes for the
3217 * dnode which overrode the dn_spill field. However,
3218 * when adding more attributes to the file a spill
3219 * block was required to hold the extra attributes.
3221 * Make sure to clear the garbage left in the dn_spill
3222 * field from the previous attributes in the bonus
3223 * buffer. Otherwise, after writing out the spill
3224 * block to the new allocated dva, it will free
3225 * the old block pointed to by the invalid dn_spill.
3227 db
->db_blkptr
= NULL
;
3229 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3230 mutex_exit(&dn
->dn_mtx
);
3234 * If this is a bonus buffer, simply copy the bonus data into the
3235 * dnode. It will be written out when the dnode is synced (and it
3236 * will be synced, since it must have been dirty for dbuf_sync to
3239 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3240 dbuf_dirty_record_t
**drp
;
3242 ASSERT(*datap
!= NULL
);
3243 ASSERT0(db
->db_level
);
3244 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=,
3245 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3246 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
3249 if (*datap
!= db
->db
.db_data
) {
3250 int slots
= DB_DNODE(db
)->dn_num_slots
;
3251 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3252 kmem_free(*datap
, bonuslen
);
3253 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3255 db
->db_data_pending
= NULL
;
3256 drp
= &db
->db_last_dirty
;
3258 drp
= &(*drp
)->dr_next
;
3259 ASSERT(dr
->dr_next
== NULL
);
3260 ASSERT(dr
->dr_dbuf
== db
);
3262 if (dr
->dr_dbuf
->db_level
!= 0) {
3263 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3264 list_destroy(&dr
->dt
.di
.dr_children
);
3266 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3267 ASSERT(db
->db_dirtycnt
> 0);
3268 db
->db_dirtycnt
-= 1;
3269 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3276 * This function may have dropped the db_mtx lock allowing a dmu_sync
3277 * operation to sneak in. As a result, we need to ensure that we
3278 * don't check the dr_override_state until we have returned from
3279 * dbuf_check_blkptr.
3281 dbuf_check_blkptr(dn
, db
);
3284 * If this buffer is in the middle of an immediate write,
3285 * wait for the synchronous IO to complete.
3287 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3288 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3289 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3290 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3293 if (db
->db_state
!= DB_NOFILL
&&
3294 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3295 refcount_count(&db
->db_holds
) > 1 &&
3296 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3297 *datap
== db
->db_buf
) {
3299 * If this buffer is currently "in use" (i.e., there
3300 * are active holds and db_data still references it),
3301 * then make a copy before we start the write so that
3302 * any modifications from the open txg will not leak
3305 * NOTE: this copy does not need to be made for
3306 * objects only modified in the syncing context (e.g.
3307 * DNONE_DNODE blocks).
3309 int psize
= arc_buf_size(*datap
);
3310 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3311 enum zio_compress compress_type
= arc_get_compression(*datap
);
3313 if (compress_type
== ZIO_COMPRESS_OFF
) {
3314 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
3316 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
3317 int lsize
= arc_buf_lsize(*datap
);
3318 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
3319 psize
, lsize
, compress_type
);
3321 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
3323 db
->db_data_pending
= dr
;
3325 mutex_exit(&db
->db_mtx
);
3327 dbuf_write(dr
, *datap
, tx
);
3329 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3330 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3331 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3335 * Although zio_nowait() does not "wait for an IO", it does
3336 * initiate the IO. If this is an empty write it seems plausible
3337 * that the IO could actually be completed before the nowait
3338 * returns. We need to DB_DNODE_EXIT() first in case
3339 * zio_nowait() invalidates the dbuf.
3342 zio_nowait(dr
->dr_zio
);
3347 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3349 dbuf_dirty_record_t
*dr
;
3351 while ((dr
= list_head(list
))) {
3352 if (dr
->dr_zio
!= NULL
) {
3354 * If we find an already initialized zio then we
3355 * are processing the meta-dnode, and we have finished.
3356 * The dbufs for all dnodes are put back on the list
3357 * during processing, so that we can zio_wait()
3358 * these IOs after initiating all child IOs.
3360 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3361 DMU_META_DNODE_OBJECT
);
3364 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3365 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3366 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3368 list_remove(list
, dr
);
3369 if (dr
->dr_dbuf
->db_level
> 0)
3370 dbuf_sync_indirect(dr
, tx
);
3372 dbuf_sync_leaf(dr
, tx
);
3378 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3380 dmu_buf_impl_t
*db
= vdb
;
3382 blkptr_t
*bp
= zio
->io_bp
;
3383 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3384 spa_t
*spa
= zio
->io_spa
;
3389 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3390 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3394 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3395 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3396 zio
->io_prev_space_delta
= delta
;
3398 if (bp
->blk_birth
!= 0) {
3399 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3400 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3401 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3402 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3403 BP_IS_EMBEDDED(bp
));
3404 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3407 mutex_enter(&db
->db_mtx
);
3410 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3411 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3412 ASSERT(!(BP_IS_HOLE(bp
)) &&
3413 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3417 if (db
->db_level
== 0) {
3418 mutex_enter(&dn
->dn_mtx
);
3419 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3420 db
->db_blkid
!= DMU_SPILL_BLKID
)
3421 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3422 mutex_exit(&dn
->dn_mtx
);
3424 if (dn
->dn_type
== DMU_OT_DNODE
) {
3426 while (i
< db
->db
.db_size
) {
3428 (void *)(((char *)db
->db
.db_data
) + i
);
3430 i
+= DNODE_MIN_SIZE
;
3431 if (dnp
->dn_type
!= DMU_OT_NONE
) {
3433 i
+= dnp
->dn_extra_slots
*
3438 if (BP_IS_HOLE(bp
)) {
3445 blkptr_t
*ibp
= db
->db
.db_data
;
3446 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3447 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3448 if (BP_IS_HOLE(ibp
))
3450 fill
+= BP_GET_FILL(ibp
);
3455 if (!BP_IS_EMBEDDED(bp
))
3456 bp
->blk_fill
= fill
;
3458 mutex_exit(&db
->db_mtx
);
3460 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3461 *db
->db_blkptr
= *bp
;
3462 rw_exit(&dn
->dn_struct_rwlock
);
3467 * This function gets called just prior to running through the compression
3468 * stage of the zio pipeline. If we're an indirect block comprised of only
3469 * holes, then we want this indirect to be compressed away to a hole. In
3470 * order to do that we must zero out any information about the holes that
3471 * this indirect points to prior to before we try to compress it.
3474 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3476 dmu_buf_impl_t
*db
= vdb
;
3479 unsigned int epbs
, i
;
3481 ASSERT3U(db
->db_level
, >, 0);
3484 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3485 ASSERT3U(epbs
, <, 31);
3487 /* Determine if all our children are holes */
3488 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
3489 if (!BP_IS_HOLE(bp
))
3494 * If all the children are holes, then zero them all out so that
3495 * we may get compressed away.
3497 if (i
== 1ULL << epbs
) {
3499 * We only found holes. Grab the rwlock to prevent
3500 * anybody from reading the blocks we're about to
3503 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3504 bzero(db
->db
.db_data
, db
->db
.db_size
);
3505 rw_exit(&dn
->dn_struct_rwlock
);
3511 * The SPA will call this callback several times for each zio - once
3512 * for every physical child i/o (zio->io_phys_children times). This
3513 * allows the DMU to monitor the progress of each logical i/o. For example,
3514 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3515 * block. There may be a long delay before all copies/fragments are completed,
3516 * so this callback allows us to retire dirty space gradually, as the physical
3521 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3523 dmu_buf_impl_t
*db
= arg
;
3524 objset_t
*os
= db
->db_objset
;
3525 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3526 dbuf_dirty_record_t
*dr
;
3529 dr
= db
->db_data_pending
;
3530 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3533 * The callback will be called io_phys_children times. Retire one
3534 * portion of our dirty space each time we are called. Any rounding
3535 * error will be cleaned up by dsl_pool_sync()'s call to
3536 * dsl_pool_undirty_space().
3538 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3539 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3544 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3546 dmu_buf_impl_t
*db
= vdb
;
3547 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3548 blkptr_t
*bp
= db
->db_blkptr
;
3549 objset_t
*os
= db
->db_objset
;
3550 dmu_tx_t
*tx
= os
->os_synctx
;
3551 dbuf_dirty_record_t
**drp
, *dr
;
3553 ASSERT0(zio
->io_error
);
3554 ASSERT(db
->db_blkptr
== bp
);
3557 * For nopwrites and rewrites we ensure that the bp matches our
3558 * original and bypass all the accounting.
3560 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3561 ASSERT(BP_EQUAL(bp
, bp_orig
));
3563 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3564 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3565 dsl_dataset_block_born(ds
, bp
, tx
);
3568 mutex_enter(&db
->db_mtx
);
3572 drp
= &db
->db_last_dirty
;
3573 while ((dr
= *drp
) != db
->db_data_pending
)
3575 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3576 ASSERT(dr
->dr_dbuf
== db
);
3577 ASSERT(dr
->dr_next
== NULL
);
3581 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3586 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3587 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3588 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3593 if (db
->db_level
== 0) {
3594 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3595 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3596 if (db
->db_state
!= DB_NOFILL
) {
3597 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3598 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
3605 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3606 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3607 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3608 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3610 ASSERT3U(db
->db_blkid
, <=,
3611 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3612 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3616 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3617 list_destroy(&dr
->dt
.di
.dr_children
);
3619 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3621 cv_broadcast(&db
->db_changed
);
3622 ASSERT(db
->db_dirtycnt
> 0);
3623 db
->db_dirtycnt
-= 1;
3624 db
->db_data_pending
= NULL
;
3625 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3629 dbuf_write_nofill_ready(zio_t
*zio
)
3631 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3635 dbuf_write_nofill_done(zio_t
*zio
)
3637 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3641 dbuf_write_override_ready(zio_t
*zio
)
3643 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3644 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3646 dbuf_write_ready(zio
, NULL
, db
);
3650 dbuf_write_override_done(zio_t
*zio
)
3652 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3653 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3654 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3656 mutex_enter(&db
->db_mtx
);
3657 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3658 if (!BP_IS_HOLE(obp
))
3659 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3660 arc_release(dr
->dt
.dl
.dr_data
, db
);
3662 mutex_exit(&db
->db_mtx
);
3664 dbuf_write_done(zio
, NULL
, db
);
3666 if (zio
->io_abd
!= NULL
)
3667 abd_put(zio
->io_abd
);
3670 /* Issue I/O to commit a dirty buffer to disk. */
3672 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3674 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3677 dmu_buf_impl_t
*parent
= db
->db_parent
;
3678 uint64_t txg
= tx
->tx_txg
;
3679 zbookmark_phys_t zb
;
3684 ASSERT(dmu_tx_is_syncing(tx
));
3690 if (db
->db_state
!= DB_NOFILL
) {
3691 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3693 * Private object buffers are released here rather
3694 * than in dbuf_dirty() since they are only modified
3695 * in the syncing context and we don't want the
3696 * overhead of making multiple copies of the data.
3698 if (BP_IS_HOLE(db
->db_blkptr
)) {
3701 dbuf_release_bp(db
);
3706 if (parent
!= dn
->dn_dbuf
) {
3707 /* Our parent is an indirect block. */
3708 /* We have a dirty parent that has been scheduled for write. */
3709 ASSERT(parent
&& parent
->db_data_pending
);
3710 /* Our parent's buffer is one level closer to the dnode. */
3711 ASSERT(db
->db_level
== parent
->db_level
-1);
3713 * We're about to modify our parent's db_data by modifying
3714 * our block pointer, so the parent must be released.
3716 ASSERT(arc_released(parent
->db_buf
));
3717 zio
= parent
->db_data_pending
->dr_zio
;
3719 /* Our parent is the dnode itself. */
3720 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3721 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3722 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3723 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3724 ASSERT3P(db
->db_blkptr
, ==,
3725 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3729 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3730 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3733 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3734 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3735 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3737 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3739 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3741 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3745 * We copy the blkptr now (rather than when we instantiate the dirty
3746 * record), because its value can change between open context and
3747 * syncing context. We do not need to hold dn_struct_rwlock to read
3748 * db_blkptr because we are in syncing context.
3750 dr
->dr_bp_copy
= *db
->db_blkptr
;
3752 if (db
->db_level
== 0 &&
3753 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3755 * The BP for this block has been provided by open context
3756 * (by dmu_sync() or dmu_buf_write_embedded()).
3758 abd_t
*contents
= (data
!= NULL
) ?
3759 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
3761 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3762 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
3763 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
3764 dbuf_write_override_done
,
3765 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3766 mutex_enter(&db
->db_mtx
);
3767 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3768 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3769 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3770 mutex_exit(&db
->db_mtx
);
3771 } else if (db
->db_state
== DB_NOFILL
) {
3772 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
3773 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
3774 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3775 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
3776 dbuf_write_nofill_ready
, NULL
, NULL
,
3777 dbuf_write_nofill_done
, db
,
3778 ZIO_PRIORITY_ASYNC_WRITE
,
3779 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3781 arc_done_func_t
*children_ready_cb
= NULL
;
3782 ASSERT(arc_released(data
));
3785 * For indirect blocks, we want to setup the children
3786 * ready callback so that we can properly handle an indirect
3787 * block that only contains holes.
3789 if (db
->db_level
!= 0)
3790 children_ready_cb
= dbuf_write_children_ready
;
3792 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3793 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3794 &zp
, dbuf_write_ready
,
3795 children_ready_cb
, dbuf_write_physdone
,
3796 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
3797 ZIO_FLAG_MUSTSUCCEED
, &zb
);
3801 #if defined(_KERNEL) && defined(HAVE_SPL)
3802 EXPORT_SYMBOL(dbuf_find
);
3803 EXPORT_SYMBOL(dbuf_is_metadata
);
3804 EXPORT_SYMBOL(dbuf_destroy
);
3805 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
3806 EXPORT_SYMBOL(dbuf_whichblock
);
3807 EXPORT_SYMBOL(dbuf_read
);
3808 EXPORT_SYMBOL(dbuf_unoverride
);
3809 EXPORT_SYMBOL(dbuf_free_range
);
3810 EXPORT_SYMBOL(dbuf_new_size
);
3811 EXPORT_SYMBOL(dbuf_release_bp
);
3812 EXPORT_SYMBOL(dbuf_dirty
);
3813 EXPORT_SYMBOL(dmu_buf_will_dirty
);
3814 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
3815 EXPORT_SYMBOL(dmu_buf_will_fill
);
3816 EXPORT_SYMBOL(dmu_buf_fill_done
);
3817 EXPORT_SYMBOL(dmu_buf_rele
);
3818 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
3819 EXPORT_SYMBOL(dbuf_prefetch
);
3820 EXPORT_SYMBOL(dbuf_hold_impl
);
3821 EXPORT_SYMBOL(dbuf_hold
);
3822 EXPORT_SYMBOL(dbuf_hold_level
);
3823 EXPORT_SYMBOL(dbuf_create_bonus
);
3824 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
3825 EXPORT_SYMBOL(dbuf_rm_spill
);
3826 EXPORT_SYMBOL(dbuf_add_ref
);
3827 EXPORT_SYMBOL(dbuf_rele
);
3828 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
3829 EXPORT_SYMBOL(dbuf_refcount
);
3830 EXPORT_SYMBOL(dbuf_sync_list
);
3831 EXPORT_SYMBOL(dmu_buf_set_user
);
3832 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3833 EXPORT_SYMBOL(dmu_buf_get_user
);
3834 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
3837 module_param(dbuf_cache_max_bytes
, ulong
, 0644);
3838 MODULE_PARM_DESC(dbuf_cache_max_bytes
,
3839 "Maximum size in bytes of the dbuf cache.");
3841 module_param(dbuf_cache_hiwater_pct
, uint
, 0644);
3842 MODULE_PARM_DESC(dbuf_cache_hiwater_pct
,
3843 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
3846 module_param(dbuf_cache_lowater_pct
, uint
, 0644);
3847 MODULE_PARM_DESC(dbuf_cache_lowater_pct
,
3848 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
3851 module_param(dbuf_cache_max_shift
, int, 0644);
3852 MODULE_PARM_DESC(dbuf_cache_max_shift
,
3853 "Cap the size of the dbuf cache to a log2 fraction of arc size.");