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 unsigned long
468 dbuf_cache_target_bytes(void)
470 return MIN(dbuf_cache_max_bytes
,
471 arc_target_bytes() >> dbuf_cache_max_shift
);
474 static inline boolean_t
475 dbuf_cache_above_hiwater(void)
477 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
479 uint64_t dbuf_cache_hiwater_bytes
=
480 (dbuf_cache_target
* dbuf_cache_hiwater_pct
) / 100;
482 return (refcount_count(&dbuf_cache_size
) >
483 dbuf_cache_target
+ dbuf_cache_hiwater_bytes
);
486 static inline boolean_t
487 dbuf_cache_above_lowater(void)
489 uint64_t dbuf_cache_target
= dbuf_cache_target_bytes();
491 uint64_t dbuf_cache_lowater_bytes
=
492 (dbuf_cache_target
* dbuf_cache_lowater_pct
) / 100;
494 return (refcount_count(&dbuf_cache_size
) >
495 dbuf_cache_target
- dbuf_cache_lowater_bytes
);
499 * Evict the oldest eligible dbuf from the dbuf cache.
504 int idx
= multilist_get_random_index(dbuf_cache
);
505 multilist_sublist_t
*mls
= multilist_sublist_lock(dbuf_cache
, idx
);
507 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
510 * Set the thread's tsd to indicate that it's processing evictions.
511 * Once a thread stops evicting from the dbuf cache it will
512 * reset its tsd to NULL.
514 ASSERT3P(tsd_get(zfs_dbuf_evict_key
), ==, NULL
);
515 (void) tsd_set(zfs_dbuf_evict_key
, (void *)B_TRUE
);
517 db
= multilist_sublist_tail(mls
);
518 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
519 db
= multilist_sublist_prev(mls
, db
);
522 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
523 multilist_sublist_t
*, mls
);
526 multilist_sublist_remove(mls
, db
);
527 multilist_sublist_unlock(mls
);
528 (void) refcount_remove_many(&dbuf_cache_size
,
532 multilist_sublist_unlock(mls
);
534 (void) tsd_set(zfs_dbuf_evict_key
, NULL
);
538 * The dbuf evict thread is responsible for aging out dbufs from the
539 * cache. Once the cache has reached it's maximum size, dbufs are removed
540 * and destroyed. The eviction thread will continue running until the size
541 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
542 * out of the cache it is destroyed and becomes eligible for arc eviction.
545 dbuf_evict_thread(void)
549 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
551 mutex_enter(&dbuf_evict_lock
);
552 while (!dbuf_evict_thread_exit
) {
553 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
554 CALLB_CPR_SAFE_BEGIN(&cpr
);
555 (void) cv_timedwait_sig_hires(&dbuf_evict_cv
,
556 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
557 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
559 mutex_exit(&dbuf_evict_lock
);
562 * Keep evicting as long as we're above the low water mark
563 * for the cache. We do this without holding the locks to
564 * minimize lock contention.
566 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
570 mutex_enter(&dbuf_evict_lock
);
573 dbuf_evict_thread_exit
= B_FALSE
;
574 cv_broadcast(&dbuf_evict_cv
);
575 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
580 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
581 * If the dbuf cache is at its high water mark, then evict a dbuf from the
582 * dbuf cache using the callers context.
585 dbuf_evict_notify(void)
589 * We use thread specific data to track when a thread has
590 * started processing evictions. This allows us to avoid deeply
591 * nested stacks that would have a call flow similar to this:
593 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
596 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
598 * The dbuf_eviction_thread will always have its tsd set until
599 * that thread exits. All other threads will only set their tsd
600 * if they are participating in the eviction process. This only
601 * happens if the eviction thread is unable to process evictions
602 * fast enough. To keep the dbuf cache size in check, other threads
603 * can evict from the dbuf cache directly. Those threads will set
604 * their tsd values so that we ensure that they only evict one dbuf
605 * from the dbuf cache.
607 if (tsd_get(zfs_dbuf_evict_key
) != NULL
)
611 * We check if we should evict without holding the dbuf_evict_lock,
612 * because it's OK to occasionally make the wrong decision here,
613 * and grabbing the lock results in massive lock contention.
615 if (refcount_count(&dbuf_cache_size
) > dbuf_cache_target_bytes()) {
616 if (dbuf_cache_above_hiwater())
618 cv_signal(&dbuf_evict_cv
);
627 uint64_t hsize
= 1ULL << 16;
628 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
632 * The hash table is big enough to fill all of physical memory
633 * with an average block size of zfs_arc_average_blocksize (default 8K).
634 * By default, the table will take up
635 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
637 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
641 h
->hash_table_mask
= hsize
- 1;
642 #if defined(_KERNEL) && defined(HAVE_SPL)
644 * Large allocations which do not require contiguous pages
645 * should be using vmem_alloc() in the linux kernel
647 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
649 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
651 if (h
->hash_table
== NULL
) {
652 /* XXX - we should really return an error instead of assert */
653 ASSERT(hsize
> (1ULL << 10));
658 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
659 sizeof (dmu_buf_impl_t
),
660 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
662 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
663 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
668 * Setup the parameters for the dbuf cache. We cap the size of the
669 * dbuf cache to 1/32nd (default) of the size of the ARC.
671 dbuf_cache_max_bytes
= MIN(dbuf_cache_max_bytes
,
672 arc_target_bytes() >> dbuf_cache_max_shift
);
675 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
676 * configuration is not required.
678 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
680 dbuf_cache
= multilist_create(sizeof (dmu_buf_impl_t
),
681 offsetof(dmu_buf_impl_t
, db_cache_link
),
682 dbuf_cache_multilist_index_func
);
683 refcount_create(&dbuf_cache_size
);
685 tsd_create(&zfs_dbuf_evict_key
, NULL
);
686 dbuf_evict_thread_exit
= B_FALSE
;
687 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
688 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
689 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
690 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
696 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
699 dbuf_stats_destroy();
701 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
702 mutex_destroy(&h
->hash_mutexes
[i
]);
703 #if defined(_KERNEL) && defined(HAVE_SPL)
705 * Large allocations which do not require contiguous pages
706 * should be using vmem_free() in the linux kernel
708 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
710 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
712 kmem_cache_destroy(dbuf_kmem_cache
);
713 taskq_destroy(dbu_evict_taskq
);
715 mutex_enter(&dbuf_evict_lock
);
716 dbuf_evict_thread_exit
= B_TRUE
;
717 while (dbuf_evict_thread_exit
) {
718 cv_signal(&dbuf_evict_cv
);
719 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
721 mutex_exit(&dbuf_evict_lock
);
722 tsd_destroy(&zfs_dbuf_evict_key
);
724 mutex_destroy(&dbuf_evict_lock
);
725 cv_destroy(&dbuf_evict_cv
);
727 refcount_destroy(&dbuf_cache_size
);
728 multilist_destroy(dbuf_cache
);
737 dbuf_verify(dmu_buf_impl_t
*db
)
740 dbuf_dirty_record_t
*dr
;
742 ASSERT(MUTEX_HELD(&db
->db_mtx
));
744 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
747 ASSERT(db
->db_objset
!= NULL
);
751 ASSERT(db
->db_parent
== NULL
);
752 ASSERT(db
->db_blkptr
== NULL
);
754 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
755 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
756 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
757 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
758 db
->db_blkid
== DMU_SPILL_BLKID
||
759 !avl_is_empty(&dn
->dn_dbufs
));
761 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
763 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
764 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
765 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
767 ASSERT0(db
->db
.db_offset
);
769 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
772 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
773 ASSERT(dr
->dr_dbuf
== db
);
775 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
776 ASSERT(dr
->dr_dbuf
== db
);
779 * We can't assert that db_size matches dn_datablksz because it
780 * can be momentarily different when another thread is doing
783 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
784 dr
= db
->db_data_pending
;
786 * It should only be modified in syncing context, so
787 * make sure we only have one copy of the data.
789 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
792 /* verify db->db_blkptr */
794 if (db
->db_parent
== dn
->dn_dbuf
) {
795 /* db is pointed to by the dnode */
796 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
797 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
798 ASSERT(db
->db_parent
== NULL
);
800 ASSERT(db
->db_parent
!= NULL
);
801 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
802 ASSERT3P(db
->db_blkptr
, ==,
803 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
805 /* db is pointed to by an indirect block */
806 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
808 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
809 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
812 * dnode_grow_indblksz() can make this fail if we don't
813 * have the struct_rwlock. XXX indblksz no longer
814 * grows. safe to do this now?
816 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
817 ASSERT3P(db
->db_blkptr
, ==,
818 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
819 db
->db_blkid
% epb
));
823 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
824 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
825 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
826 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
828 * If the blkptr isn't set but they have nonzero data,
829 * it had better be dirty, otherwise we'll lose that
830 * data when we evict this buffer.
832 * There is an exception to this rule for indirect blocks; in
833 * this case, if the indirect block is a hole, we fill in a few
834 * fields on each of the child blocks (importantly, birth time)
835 * to prevent hole birth times from being lost when you
836 * partially fill in a hole.
838 if (db
->db_dirtycnt
== 0) {
839 if (db
->db_level
== 0) {
840 uint64_t *buf
= db
->db
.db_data
;
843 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
848 blkptr_t
*bps
= db
->db
.db_data
;
849 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
852 * We want to verify that all the blkptrs in the
853 * indirect block are holes, but we may have
854 * automatically set up a few fields for them.
855 * We iterate through each blkptr and verify
856 * they only have those fields set.
859 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
861 blkptr_t
*bp
= &bps
[i
];
862 ASSERT(ZIO_CHECKSUM_IS_ZERO(
865 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
866 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
867 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
868 ASSERT0(bp
->blk_fill
);
869 ASSERT0(bp
->blk_pad
[0]);
870 ASSERT0(bp
->blk_pad
[1]);
871 ASSERT(!BP_IS_EMBEDDED(bp
));
872 ASSERT(BP_IS_HOLE(bp
));
873 ASSERT0(bp
->blk_phys_birth
);
883 dbuf_clear_data(dmu_buf_impl_t
*db
)
885 ASSERT(MUTEX_HELD(&db
->db_mtx
));
887 ASSERT3P(db
->db_buf
, ==, NULL
);
888 db
->db
.db_data
= NULL
;
889 if (db
->db_state
!= DB_NOFILL
)
890 db
->db_state
= DB_UNCACHED
;
894 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
896 ASSERT(MUTEX_HELD(&db
->db_mtx
));
900 ASSERT(buf
->b_data
!= NULL
);
901 db
->db
.db_data
= buf
->b_data
;
905 * Loan out an arc_buf for read. Return the loaned arc_buf.
908 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
912 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
913 mutex_enter(&db
->db_mtx
);
914 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
915 int blksz
= db
->db
.db_size
;
916 spa_t
*spa
= db
->db_objset
->os_spa
;
918 mutex_exit(&db
->db_mtx
);
919 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
920 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
923 arc_loan_inuse_buf(abuf
, db
);
926 mutex_exit(&db
->db_mtx
);
932 * Calculate which level n block references the data at the level 0 offset
936 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
938 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
940 * The level n blkid is equal to the level 0 blkid divided by
941 * the number of level 0s in a level n block.
943 * The level 0 blkid is offset >> datablkshift =
944 * offset / 2^datablkshift.
946 * The number of level 0s in a level n is the number of block
947 * pointers in an indirect block, raised to the power of level.
948 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
949 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
951 * Thus, the level n blkid is: offset /
952 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
953 * = offset / 2^(datablkshift + level *
954 * (indblkshift - SPA_BLKPTRSHIFT))
955 * = offset >> (datablkshift + level *
956 * (indblkshift - SPA_BLKPTRSHIFT))
959 const unsigned exp
= dn
->dn_datablkshift
+
960 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
962 if (exp
>= 8 * sizeof (offset
)) {
963 /* This only happens on the highest indirection level */
964 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
968 ASSERT3U(exp
, <, 8 * sizeof (offset
));
970 return (offset
>> exp
);
972 ASSERT3U(offset
, <, dn
->dn_datablksz
);
978 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
980 dmu_buf_impl_t
*db
= vdb
;
982 mutex_enter(&db
->db_mtx
);
983 ASSERT3U(db
->db_state
, ==, DB_READ
);
985 * All reads are synchronous, so we must have a hold on the dbuf
987 ASSERT(refcount_count(&db
->db_holds
) > 0);
988 ASSERT(db
->db_buf
== NULL
);
989 ASSERT(db
->db
.db_data
== NULL
);
990 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
991 /* we were freed in flight; disregard any error */
992 arc_release(buf
, db
);
993 bzero(buf
->b_data
, db
->db
.db_size
);
995 db
->db_freed_in_flight
= FALSE
;
996 dbuf_set_data(db
, buf
);
997 db
->db_state
= DB_CACHED
;
998 } else if (zio
== NULL
|| zio
->io_error
== 0) {
999 dbuf_set_data(db
, buf
);
1000 db
->db_state
= DB_CACHED
;
1002 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1003 ASSERT3P(db
->db_buf
, ==, NULL
);
1004 arc_buf_destroy(buf
, db
);
1005 db
->db_state
= DB_UNCACHED
;
1007 cv_broadcast(&db
->db_changed
);
1008 dbuf_rele_and_unlock(db
, NULL
);
1012 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1015 zbookmark_phys_t zb
;
1016 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1021 ASSERT(!refcount_is_zero(&db
->db_holds
));
1022 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1023 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1024 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1025 ASSERT(db
->db_state
== DB_UNCACHED
);
1026 ASSERT(db
->db_buf
== NULL
);
1028 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1030 * The bonus length stored in the dnode may be less than
1031 * the maximum available space in the bonus buffer.
1033 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1034 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1036 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1037 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1038 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1039 if (bonuslen
< max_bonuslen
)
1040 bzero(db
->db
.db_data
, max_bonuslen
);
1042 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1044 db
->db_state
= DB_CACHED
;
1045 mutex_exit(&db
->db_mtx
);
1050 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1051 * processes the delete record and clears the bp while we are waiting
1052 * for the dn_mtx (resulting in a "no" from block_freed).
1054 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1055 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1056 BP_IS_HOLE(db
->db_blkptr
)))) {
1057 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1059 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1061 bzero(db
->db
.db_data
, db
->db
.db_size
);
1063 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1064 BP_IS_HOLE(db
->db_blkptr
) &&
1065 db
->db_blkptr
->blk_birth
!= 0) {
1066 blkptr_t
*bps
= db
->db
.db_data
;
1068 for (i
= 0; i
< ((1 <<
1069 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1071 blkptr_t
*bp
= &bps
[i
];
1072 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1073 1 << dn
->dn_indblkshift
);
1075 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1077 BP_GET_LSIZE(db
->db_blkptr
));
1078 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1080 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1081 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1085 db
->db_state
= DB_CACHED
;
1086 mutex_exit(&db
->db_mtx
);
1092 db
->db_state
= DB_READ
;
1093 mutex_exit(&db
->db_mtx
);
1095 if (DBUF_IS_L2CACHEABLE(db
))
1096 aflags
|= ARC_FLAG_L2CACHE
;
1098 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
1099 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
1100 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1102 dbuf_add_ref(db
, NULL
);
1104 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1105 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
1106 (flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
1113 * This is our just-in-time copy function. It makes a copy of buffers that
1114 * have been modified in a previous transaction group before we access them in
1115 * the current active group.
1117 * This function is used in three places: when we are dirtying a buffer for the
1118 * first time in a txg, when we are freeing a range in a dnode that includes
1119 * this buffer, and when we are accessing a buffer which was received compressed
1120 * and later referenced in a WRITE_BYREF record.
1122 * Note that when we are called from dbuf_free_range() we do not put a hold on
1123 * the buffer, we just traverse the active dbuf list for the dnode.
1126 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1128 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1130 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1131 ASSERT(db
->db
.db_data
!= NULL
);
1132 ASSERT(db
->db_level
== 0);
1133 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1136 (dr
->dt
.dl
.dr_data
!=
1137 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1141 * If the last dirty record for this dbuf has not yet synced
1142 * and its referencing the dbuf data, either:
1143 * reset the reference to point to a new copy,
1144 * or (if there a no active holders)
1145 * just null out the current db_data pointer.
1147 ASSERT(dr
->dr_txg
>= txg
- 2);
1148 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1149 dnode_t
*dn
= DB_DNODE(db
);
1150 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1151 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1152 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1153 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1154 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1155 int size
= arc_buf_size(db
->db_buf
);
1156 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1157 spa_t
*spa
= db
->db_objset
->os_spa
;
1158 enum zio_compress compress_type
=
1159 arc_get_compression(db
->db_buf
);
1161 if (compress_type
== ZIO_COMPRESS_OFF
) {
1162 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1164 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1165 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1166 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1168 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1171 dbuf_clear_data(db
);
1176 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1183 * We don't have to hold the mutex to check db_state because it
1184 * can't be freed while we have a hold on the buffer.
1186 ASSERT(!refcount_is_zero(&db
->db_holds
));
1188 if (db
->db_state
== DB_NOFILL
)
1189 return (SET_ERROR(EIO
));
1193 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1194 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1196 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1197 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1198 DBUF_IS_CACHEABLE(db
);
1200 mutex_enter(&db
->db_mtx
);
1201 if (db
->db_state
== DB_CACHED
) {
1203 * If the arc buf is compressed, we need to decompress it to
1204 * read the data. This could happen during the "zfs receive" of
1205 * a stream which is compressed and deduplicated.
1207 if (db
->db_buf
!= NULL
&&
1208 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
) {
1209 dbuf_fix_old_data(db
,
1210 spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1211 err
= arc_decompress(db
->db_buf
);
1212 dbuf_set_data(db
, db
->db_buf
);
1214 mutex_exit(&db
->db_mtx
);
1216 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1217 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1218 rw_exit(&dn
->dn_struct_rwlock
);
1220 } else if (db
->db_state
== DB_UNCACHED
) {
1221 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1222 boolean_t need_wait
= B_FALSE
;
1225 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1226 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1229 err
= dbuf_read_impl(db
, zio
, flags
);
1231 /* dbuf_read_impl has dropped db_mtx for us */
1233 if (!err
&& prefetch
)
1234 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1236 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1237 rw_exit(&dn
->dn_struct_rwlock
);
1240 if (!err
&& need_wait
)
1241 err
= zio_wait(zio
);
1244 * Another reader came in while the dbuf was in flight
1245 * between UNCACHED and CACHED. Either a writer will finish
1246 * writing the buffer (sending the dbuf to CACHED) or the
1247 * first reader's request will reach the read_done callback
1248 * and send the dbuf to CACHED. Otherwise, a failure
1249 * occurred and the dbuf went to UNCACHED.
1251 mutex_exit(&db
->db_mtx
);
1253 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1254 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1255 rw_exit(&dn
->dn_struct_rwlock
);
1258 /* Skip the wait per the caller's request. */
1259 mutex_enter(&db
->db_mtx
);
1260 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1261 while (db
->db_state
== DB_READ
||
1262 db
->db_state
== DB_FILL
) {
1263 ASSERT(db
->db_state
== DB_READ
||
1264 (flags
& DB_RF_HAVESTRUCT
) == 0);
1265 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1267 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1269 if (db
->db_state
== DB_UNCACHED
)
1270 err
= SET_ERROR(EIO
);
1272 mutex_exit(&db
->db_mtx
);
1279 dbuf_noread(dmu_buf_impl_t
*db
)
1281 ASSERT(!refcount_is_zero(&db
->db_holds
));
1282 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1283 mutex_enter(&db
->db_mtx
);
1284 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1285 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1286 if (db
->db_state
== DB_UNCACHED
) {
1287 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1288 spa_t
*spa
= db
->db_objset
->os_spa
;
1290 ASSERT(db
->db_buf
== NULL
);
1291 ASSERT(db
->db
.db_data
== NULL
);
1292 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1293 db
->db_state
= DB_FILL
;
1294 } else if (db
->db_state
== DB_NOFILL
) {
1295 dbuf_clear_data(db
);
1297 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1299 mutex_exit(&db
->db_mtx
);
1303 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1305 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1306 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1307 uint64_t txg
= dr
->dr_txg
;
1309 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1311 * This assert is valid because dmu_sync() expects to be called by
1312 * a zilog's get_data while holding a range lock. This call only
1313 * comes from dbuf_dirty() callers who must also hold a range lock.
1315 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1316 ASSERT(db
->db_level
== 0);
1318 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1319 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1322 ASSERT(db
->db_data_pending
!= dr
);
1324 /* free this block */
1325 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1326 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1328 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1329 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1332 * Release the already-written buffer, so we leave it in
1333 * a consistent dirty state. Note that all callers are
1334 * modifying the buffer, so they will immediately do
1335 * another (redundant) arc_release(). Therefore, leave
1336 * the buf thawed to save the effort of freezing &
1337 * immediately re-thawing it.
1339 arc_release(dr
->dt
.dl
.dr_data
, db
);
1343 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1344 * data blocks in the free range, so that any future readers will find
1348 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1351 dmu_buf_impl_t
*db_search
;
1352 dmu_buf_impl_t
*db
, *db_next
;
1353 uint64_t txg
= tx
->tx_txg
;
1356 if (end_blkid
> dn
->dn_maxblkid
&&
1357 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1358 end_blkid
= dn
->dn_maxblkid
;
1359 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1361 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1362 db_search
->db_level
= 0;
1363 db_search
->db_blkid
= start_blkid
;
1364 db_search
->db_state
= DB_SEARCH
;
1366 mutex_enter(&dn
->dn_dbufs_mtx
);
1367 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1368 ASSERT3P(db
, ==, NULL
);
1370 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1372 for (; db
!= NULL
; db
= db_next
) {
1373 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1374 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1376 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1379 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1381 /* found a level 0 buffer in the range */
1382 mutex_enter(&db
->db_mtx
);
1383 if (dbuf_undirty(db
, tx
)) {
1384 /* mutex has been dropped and dbuf destroyed */
1388 if (db
->db_state
== DB_UNCACHED
||
1389 db
->db_state
== DB_NOFILL
||
1390 db
->db_state
== DB_EVICTING
) {
1391 ASSERT(db
->db
.db_data
== NULL
);
1392 mutex_exit(&db
->db_mtx
);
1395 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1396 /* will be handled in dbuf_read_done or dbuf_rele */
1397 db
->db_freed_in_flight
= TRUE
;
1398 mutex_exit(&db
->db_mtx
);
1401 if (refcount_count(&db
->db_holds
) == 0) {
1406 /* The dbuf is referenced */
1408 if (db
->db_last_dirty
!= NULL
) {
1409 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1411 if (dr
->dr_txg
== txg
) {
1413 * This buffer is "in-use", re-adjust the file
1414 * size to reflect that this buffer may
1415 * contain new data when we sync.
1417 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1418 db
->db_blkid
> dn
->dn_maxblkid
)
1419 dn
->dn_maxblkid
= db
->db_blkid
;
1420 dbuf_unoverride(dr
);
1423 * This dbuf is not dirty in the open context.
1424 * Either uncache it (if its not referenced in
1425 * the open context) or reset its contents to
1428 dbuf_fix_old_data(db
, txg
);
1431 /* clear the contents if its cached */
1432 if (db
->db_state
== DB_CACHED
) {
1433 ASSERT(db
->db
.db_data
!= NULL
);
1434 arc_release(db
->db_buf
, db
);
1435 bzero(db
->db
.db_data
, db
->db
.db_size
);
1436 arc_buf_freeze(db
->db_buf
);
1439 mutex_exit(&db
->db_mtx
);
1442 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1443 mutex_exit(&dn
->dn_dbufs_mtx
);
1447 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1449 arc_buf_t
*buf
, *obuf
;
1450 int osize
= db
->db
.db_size
;
1451 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1454 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1459 /* XXX does *this* func really need the lock? */
1460 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1463 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1464 * is OK, because there can be no other references to the db
1465 * when we are changing its size, so no concurrent DB_FILL can
1469 * XXX we should be doing a dbuf_read, checking the return
1470 * value and returning that up to our callers
1472 dmu_buf_will_dirty(&db
->db
, tx
);
1474 /* create the data buffer for the new block */
1475 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1477 /* copy old block data to the new block */
1479 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1480 /* zero the remainder */
1482 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1484 mutex_enter(&db
->db_mtx
);
1485 dbuf_set_data(db
, buf
);
1486 arc_buf_destroy(obuf
, db
);
1487 db
->db
.db_size
= size
;
1489 if (db
->db_level
== 0) {
1490 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1491 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1493 mutex_exit(&db
->db_mtx
);
1495 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1500 dbuf_release_bp(dmu_buf_impl_t
*db
)
1502 ASSERTV(objset_t
*os
= db
->db_objset
);
1504 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1505 ASSERT(arc_released(os
->os_phys_buf
) ||
1506 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1507 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1509 (void) arc_release(db
->db_buf
, db
);
1513 * We already have a dirty record for this TXG, and we are being
1517 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1519 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1521 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1523 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1525 * If this buffer has already been written out,
1526 * we now need to reset its state.
1528 dbuf_unoverride(dr
);
1529 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1530 db
->db_state
!= DB_NOFILL
) {
1531 /* Already released on initial dirty, so just thaw. */
1532 ASSERT(arc_released(db
->db_buf
));
1533 arc_buf_thaw(db
->db_buf
);
1538 dbuf_dirty_record_t
*
1539 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1543 dbuf_dirty_record_t
**drp
, *dr
;
1544 int drop_struct_lock
= FALSE
;
1545 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1547 ASSERT(tx
->tx_txg
!= 0);
1548 ASSERT(!refcount_is_zero(&db
->db_holds
));
1549 DMU_TX_DIRTY_BUF(tx
, db
);
1554 * Shouldn't dirty a regular buffer in syncing context. Private
1555 * objects may be dirtied in syncing context, but only if they
1556 * were already pre-dirtied in open context.
1559 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1560 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1563 ASSERT(!dmu_tx_is_syncing(tx
) ||
1564 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1565 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1566 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1567 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1568 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1571 * We make this assert for private objects as well, but after we
1572 * check if we're already dirty. They are allowed to re-dirty
1573 * in syncing context.
1575 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1576 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1577 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1579 mutex_enter(&db
->db_mtx
);
1581 * XXX make this true for indirects too? The problem is that
1582 * transactions created with dmu_tx_create_assigned() from
1583 * syncing context don't bother holding ahead.
1585 ASSERT(db
->db_level
!= 0 ||
1586 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1587 db
->db_state
== DB_NOFILL
);
1589 mutex_enter(&dn
->dn_mtx
);
1591 * Don't set dirtyctx to SYNC if we're just modifying this as we
1592 * initialize the objset.
1594 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1595 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1596 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1599 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1600 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1601 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1602 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1603 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1605 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1606 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1610 mutex_exit(&dn
->dn_mtx
);
1612 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1613 dn
->dn_have_spill
= B_TRUE
;
1616 * If this buffer is already dirty, we're done.
1618 drp
= &db
->db_last_dirty
;
1619 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1620 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1621 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1623 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1627 mutex_exit(&db
->db_mtx
);
1632 * Only valid if not already dirty.
1634 ASSERT(dn
->dn_object
== 0 ||
1635 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1636 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1638 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1641 * We should only be dirtying in syncing context if it's the
1642 * mos or we're initializing the os or it's a special object.
1643 * However, we are allowed to dirty in syncing context provided
1644 * we already dirtied it in open context. Hence we must make
1645 * this assertion only if we're not already dirty.
1648 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
1650 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1651 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
1652 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1653 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1654 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1655 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1657 ASSERT(db
->db
.db_size
!= 0);
1659 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1661 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1662 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
1666 * If this buffer is dirty in an old transaction group we need
1667 * to make a copy of it so that the changes we make in this
1668 * transaction group won't leak out when we sync the older txg.
1670 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1671 list_link_init(&dr
->dr_dirty_node
);
1672 if (db
->db_level
== 0) {
1673 void *data_old
= db
->db_buf
;
1675 if (db
->db_state
!= DB_NOFILL
) {
1676 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1677 dbuf_fix_old_data(db
, tx
->tx_txg
);
1678 data_old
= db
->db
.db_data
;
1679 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1681 * Release the data buffer from the cache so
1682 * that we can modify it without impacting
1683 * possible other users of this cached data
1684 * block. Note that indirect blocks and
1685 * private objects are not released until the
1686 * syncing state (since they are only modified
1689 arc_release(db
->db_buf
, db
);
1690 dbuf_fix_old_data(db
, tx
->tx_txg
);
1691 data_old
= db
->db_buf
;
1693 ASSERT(data_old
!= NULL
);
1695 dr
->dt
.dl
.dr_data
= data_old
;
1697 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1698 list_create(&dr
->dt
.di
.dr_children
,
1699 sizeof (dbuf_dirty_record_t
),
1700 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1702 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1703 dr
->dr_accounted
= db
->db
.db_size
;
1705 dr
->dr_txg
= tx
->tx_txg
;
1710 * We could have been freed_in_flight between the dbuf_noread
1711 * and dbuf_dirty. We win, as though the dbuf_noread() had
1712 * happened after the free.
1714 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1715 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1716 mutex_enter(&dn
->dn_mtx
);
1717 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1718 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1721 mutex_exit(&dn
->dn_mtx
);
1722 db
->db_freed_in_flight
= FALSE
;
1726 * This buffer is now part of this txg
1728 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1729 db
->db_dirtycnt
+= 1;
1730 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1732 mutex_exit(&db
->db_mtx
);
1734 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1735 db
->db_blkid
== DMU_SPILL_BLKID
) {
1736 mutex_enter(&dn
->dn_mtx
);
1737 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1738 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1739 mutex_exit(&dn
->dn_mtx
);
1740 dnode_setdirty(dn
, tx
);
1746 * The dn_struct_rwlock prevents db_blkptr from changing
1747 * due to a write from syncing context completing
1748 * while we are running, so we want to acquire it before
1749 * looking at db_blkptr.
1751 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1752 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1753 drop_struct_lock
= TRUE
;
1757 * We need to hold the dn_struct_rwlock to make this assertion,
1758 * because it protects dn_phys / dn_next_nlevels from changing.
1760 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1761 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1762 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1763 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1764 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1767 * If we are overwriting a dedup BP, then unless it is snapshotted,
1768 * when we get to syncing context we will need to decrement its
1769 * refcount in the DDT. Prefetch the relevant DDT block so that
1770 * syncing context won't have to wait for the i/o.
1772 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
1774 if (db
->db_level
== 0) {
1775 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1776 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1779 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1780 dmu_buf_impl_t
*parent
= db
->db_parent
;
1781 dbuf_dirty_record_t
*di
;
1782 int parent_held
= FALSE
;
1784 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1785 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1787 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1788 db
->db_blkid
>> epbs
, FTAG
);
1789 ASSERT(parent
!= NULL
);
1792 if (drop_struct_lock
)
1793 rw_exit(&dn
->dn_struct_rwlock
);
1794 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1795 di
= dbuf_dirty(parent
, tx
);
1797 dbuf_rele(parent
, FTAG
);
1799 mutex_enter(&db
->db_mtx
);
1801 * Since we've dropped the mutex, it's possible that
1802 * dbuf_undirty() might have changed this out from under us.
1804 if (db
->db_last_dirty
== dr
||
1805 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1806 mutex_enter(&di
->dt
.di
.dr_mtx
);
1807 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1808 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1809 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1810 mutex_exit(&di
->dt
.di
.dr_mtx
);
1813 mutex_exit(&db
->db_mtx
);
1815 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1816 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1817 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1818 mutex_enter(&dn
->dn_mtx
);
1819 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1820 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1821 mutex_exit(&dn
->dn_mtx
);
1822 if (drop_struct_lock
)
1823 rw_exit(&dn
->dn_struct_rwlock
);
1826 dnode_setdirty(dn
, tx
);
1832 * Undirty a buffer in the transaction group referenced by the given
1833 * transaction. Return whether this evicted the dbuf.
1836 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1839 uint64_t txg
= tx
->tx_txg
;
1840 dbuf_dirty_record_t
*dr
, **drp
;
1845 * Due to our use of dn_nlevels below, this can only be called
1846 * in open context, unless we are operating on the MOS.
1847 * From syncing context, dn_nlevels may be different from the
1848 * dn_nlevels used when dbuf was dirtied.
1850 ASSERT(db
->db_objset
==
1851 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1852 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1853 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1854 ASSERT0(db
->db_level
);
1855 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1858 * If this buffer is not dirty, we're done.
1860 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1861 if (dr
->dr_txg
<= txg
)
1863 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1865 ASSERT(dr
->dr_txg
== txg
);
1866 ASSERT(dr
->dr_dbuf
== db
);
1871 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1873 ASSERT(db
->db
.db_size
!= 0);
1875 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1876 dr
->dr_accounted
, txg
);
1881 * Note that there are three places in dbuf_dirty()
1882 * where this dirty record may be put on a list.
1883 * Make sure to do a list_remove corresponding to
1884 * every one of those list_insert calls.
1886 if (dr
->dr_parent
) {
1887 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1888 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1889 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1890 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1891 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1892 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1893 mutex_enter(&dn
->dn_mtx
);
1894 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1895 mutex_exit(&dn
->dn_mtx
);
1899 if (db
->db_state
!= DB_NOFILL
) {
1900 dbuf_unoverride(dr
);
1902 ASSERT(db
->db_buf
!= NULL
);
1903 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1904 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1905 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
1908 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1910 ASSERT(db
->db_dirtycnt
> 0);
1911 db
->db_dirtycnt
-= 1;
1913 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1914 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
1923 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1925 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1926 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1927 dbuf_dirty_record_t
*dr
;
1929 ASSERT(tx
->tx_txg
!= 0);
1930 ASSERT(!refcount_is_zero(&db
->db_holds
));
1933 * Quick check for dirtyness. For already dirty blocks, this
1934 * reduces runtime of this function by >90%, and overall performance
1935 * by 50% for some workloads (e.g. file deletion with indirect blocks
1938 mutex_enter(&db
->db_mtx
);
1940 for (dr
= db
->db_last_dirty
;
1941 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1943 * It's possible that it is already dirty but not cached,
1944 * because there are some calls to dbuf_dirty() that don't
1945 * go through dmu_buf_will_dirty().
1947 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1948 /* This dbuf is already dirty and cached. */
1950 mutex_exit(&db
->db_mtx
);
1954 mutex_exit(&db
->db_mtx
);
1957 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1958 rf
|= DB_RF_HAVESTRUCT
;
1960 (void) dbuf_read(db
, NULL
, rf
);
1961 (void) dbuf_dirty(db
, tx
);
1965 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1967 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1969 db
->db_state
= DB_NOFILL
;
1971 dmu_buf_will_fill(db_fake
, tx
);
1975 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1977 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1979 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1980 ASSERT(tx
->tx_txg
!= 0);
1981 ASSERT(db
->db_level
== 0);
1982 ASSERT(!refcount_is_zero(&db
->db_holds
));
1984 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1985 dmu_tx_private_ok(tx
));
1988 (void) dbuf_dirty(db
, tx
);
1991 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1994 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1996 mutex_enter(&db
->db_mtx
);
1999 if (db
->db_state
== DB_FILL
) {
2000 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2001 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2002 /* we were freed while filling */
2003 /* XXX dbuf_undirty? */
2004 bzero(db
->db
.db_data
, db
->db
.db_size
);
2005 db
->db_freed_in_flight
= FALSE
;
2007 db
->db_state
= DB_CACHED
;
2008 cv_broadcast(&db
->db_changed
);
2010 mutex_exit(&db
->db_mtx
);
2014 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2015 bp_embedded_type_t etype
, enum zio_compress comp
,
2016 int uncompressed_size
, int compressed_size
, int byteorder
,
2019 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2020 struct dirty_leaf
*dl
;
2021 dmu_object_type_t type
;
2023 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2024 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2025 SPA_FEATURE_EMBEDDED_DATA
));
2029 type
= DB_DNODE(db
)->dn_type
;
2032 ASSERT0(db
->db_level
);
2033 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2035 dmu_buf_will_not_fill(dbuf
, tx
);
2037 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2038 dl
= &db
->db_last_dirty
->dt
.dl
;
2039 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2040 data
, comp
, uncompressed_size
, compressed_size
);
2041 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2042 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2043 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2044 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2046 dl
->dr_override_state
= DR_OVERRIDDEN
;
2047 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2051 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2052 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2055 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2057 ASSERT(!refcount_is_zero(&db
->db_holds
));
2058 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2059 ASSERT(db
->db_level
== 0);
2060 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2061 ASSERT(buf
!= NULL
);
2062 ASSERT(arc_buf_lsize(buf
) == db
->db
.db_size
);
2063 ASSERT(tx
->tx_txg
!= 0);
2065 arc_return_buf(buf
, db
);
2066 ASSERT(arc_released(buf
));
2068 mutex_enter(&db
->db_mtx
);
2070 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2071 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2073 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2075 if (db
->db_state
== DB_CACHED
&&
2076 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2077 mutex_exit(&db
->db_mtx
);
2078 (void) dbuf_dirty(db
, tx
);
2079 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2080 arc_buf_destroy(buf
, db
);
2081 xuio_stat_wbuf_copied();
2085 xuio_stat_wbuf_nocopy();
2086 if (db
->db_state
== DB_CACHED
) {
2087 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2089 ASSERT(db
->db_buf
!= NULL
);
2090 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2091 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2092 if (!arc_released(db
->db_buf
)) {
2093 ASSERT(dr
->dt
.dl
.dr_override_state
==
2095 arc_release(db
->db_buf
, db
);
2097 dr
->dt
.dl
.dr_data
= buf
;
2098 arc_buf_destroy(db
->db_buf
, db
);
2099 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2100 arc_release(db
->db_buf
, db
);
2101 arc_buf_destroy(db
->db_buf
, db
);
2105 ASSERT(db
->db_buf
== NULL
);
2106 dbuf_set_data(db
, buf
);
2107 db
->db_state
= DB_FILL
;
2108 mutex_exit(&db
->db_mtx
);
2109 (void) dbuf_dirty(db
, tx
);
2110 dmu_buf_fill_done(&db
->db
, tx
);
2114 dbuf_destroy(dmu_buf_impl_t
*db
)
2117 dmu_buf_impl_t
*parent
= db
->db_parent
;
2118 dmu_buf_impl_t
*dndb
;
2120 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2121 ASSERT(refcount_is_zero(&db
->db_holds
));
2123 if (db
->db_buf
!= NULL
) {
2124 arc_buf_destroy(db
->db_buf
, db
);
2128 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2129 int slots
= DB_DNODE(db
)->dn_num_slots
;
2130 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2131 ASSERT(db
->db
.db_data
!= NULL
);
2132 kmem_free(db
->db
.db_data
, bonuslen
);
2133 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2134 db
->db_state
= DB_UNCACHED
;
2137 dbuf_clear_data(db
);
2139 if (multilist_link_active(&db
->db_cache_link
)) {
2140 multilist_remove(dbuf_cache
, db
);
2141 (void) refcount_remove_many(&dbuf_cache_size
,
2142 db
->db
.db_size
, db
);
2145 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2146 ASSERT(db
->db_data_pending
== NULL
);
2148 db
->db_state
= DB_EVICTING
;
2149 db
->db_blkptr
= NULL
;
2152 * Now that db_state is DB_EVICTING, nobody else can find this via
2153 * the hash table. We can now drop db_mtx, which allows us to
2154 * acquire the dn_dbufs_mtx.
2156 mutex_exit(&db
->db_mtx
);
2161 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2162 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2164 mutex_enter(&dn
->dn_dbufs_mtx
);
2165 avl_remove(&dn
->dn_dbufs
, db
);
2166 atomic_dec_32(&dn
->dn_dbufs_count
);
2170 mutex_exit(&dn
->dn_dbufs_mtx
);
2172 * Decrementing the dbuf count means that the hold corresponding
2173 * to the removed dbuf is no longer discounted in dnode_move(),
2174 * so the dnode cannot be moved until after we release the hold.
2175 * The membar_producer() ensures visibility of the decremented
2176 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2180 db
->db_dnode_handle
= NULL
;
2182 dbuf_hash_remove(db
);
2187 ASSERT(refcount_is_zero(&db
->db_holds
));
2189 db
->db_parent
= NULL
;
2191 ASSERT(db
->db_buf
== NULL
);
2192 ASSERT(db
->db
.db_data
== NULL
);
2193 ASSERT(db
->db_hash_next
== NULL
);
2194 ASSERT(db
->db_blkptr
== NULL
);
2195 ASSERT(db
->db_data_pending
== NULL
);
2196 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2198 kmem_cache_free(dbuf_kmem_cache
, db
);
2199 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2202 * If this dbuf is referenced from an indirect dbuf,
2203 * decrement the ref count on the indirect dbuf.
2205 if (parent
&& parent
!= dndb
)
2206 dbuf_rele(parent
, db
);
2210 * Note: While bpp will always be updated if the function returns success,
2211 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2212 * this happens when the dnode is the meta-dnode, or a userused or groupused
2215 __attribute__((always_inline
))
2217 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2218 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
2225 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2227 if (blkid
== DMU_SPILL_BLKID
) {
2228 mutex_enter(&dn
->dn_mtx
);
2229 if (dn
->dn_have_spill
&&
2230 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2231 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2234 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2235 *parentp
= dn
->dn_dbuf
;
2236 mutex_exit(&dn
->dn_mtx
);
2241 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2242 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2244 ASSERT3U(level
* epbs
, <, 64);
2245 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2247 * This assertion shouldn't trip as long as the max indirect block size
2248 * is less than 1M. The reason for this is that up to that point,
2249 * the number of levels required to address an entire object with blocks
2250 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2251 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2252 * (i.e. we can address the entire object), objects will all use at most
2253 * N-1 levels and the assertion won't overflow. However, once epbs is
2254 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2255 * enough to address an entire object, so objects will have 5 levels,
2256 * but then this assertion will overflow.
2258 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2259 * need to redo this logic to handle overflows.
2261 ASSERT(level
>= nlevels
||
2262 ((nlevels
- level
- 1) * epbs
) +
2263 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2264 if (level
>= nlevels
||
2265 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2266 ((nlevels
- level
- 1) * epbs
)) ||
2268 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2269 /* the buffer has no parent yet */
2270 return (SET_ERROR(ENOENT
));
2271 } else if (level
< nlevels
-1) {
2272 /* this block is referenced from an indirect block */
2275 err
= dbuf_hold_impl(dn
, level
+1,
2276 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2278 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
2279 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
2280 parentp
, dh
->dh_depth
+ 1);
2281 err
= __dbuf_hold_impl(dh
+ 1);
2285 err
= dbuf_read(*parentp
, NULL
,
2286 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2288 dbuf_rele(*parentp
, NULL
);
2292 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2293 (blkid
& ((1ULL << epbs
) - 1));
2294 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2295 ASSERT(BP_IS_HOLE(*bpp
));
2298 /* the block is referenced from the dnode */
2299 ASSERT3U(level
, ==, nlevels
-1);
2300 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2301 blkid
< dn
->dn_phys
->dn_nblkptr
);
2303 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2304 *parentp
= dn
->dn_dbuf
;
2306 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2311 static dmu_buf_impl_t
*
2312 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2313 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2315 objset_t
*os
= dn
->dn_objset
;
2316 dmu_buf_impl_t
*db
, *odb
;
2318 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2319 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2321 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2324 db
->db
.db_object
= dn
->dn_object
;
2325 db
->db_level
= level
;
2326 db
->db_blkid
= blkid
;
2327 db
->db_last_dirty
= NULL
;
2328 db
->db_dirtycnt
= 0;
2329 db
->db_dnode_handle
= dn
->dn_handle
;
2330 db
->db_parent
= parent
;
2331 db
->db_blkptr
= blkptr
;
2334 db
->db_user_immediate_evict
= FALSE
;
2335 db
->db_freed_in_flight
= FALSE
;
2336 db
->db_pending_evict
= FALSE
;
2338 if (blkid
== DMU_BONUS_BLKID
) {
2339 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2340 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2341 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2342 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2343 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2344 db
->db_state
= DB_UNCACHED
;
2345 /* the bonus dbuf is not placed in the hash table */
2346 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2348 } else if (blkid
== DMU_SPILL_BLKID
) {
2349 db
->db
.db_size
= (blkptr
!= NULL
) ?
2350 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2351 db
->db
.db_offset
= 0;
2354 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2355 db
->db
.db_size
= blocksize
;
2356 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2360 * Hold the dn_dbufs_mtx while we get the new dbuf
2361 * in the hash table *and* added to the dbufs list.
2362 * This prevents a possible deadlock with someone
2363 * trying to look up this dbuf before its added to the
2366 mutex_enter(&dn
->dn_dbufs_mtx
);
2367 db
->db_state
= DB_EVICTING
;
2368 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2369 /* someone else inserted it first */
2370 kmem_cache_free(dbuf_kmem_cache
, db
);
2371 mutex_exit(&dn
->dn_dbufs_mtx
);
2374 avl_add(&dn
->dn_dbufs
, db
);
2376 db
->db_state
= DB_UNCACHED
;
2377 mutex_exit(&dn
->dn_dbufs_mtx
);
2378 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2380 if (parent
&& parent
!= dn
->dn_dbuf
)
2381 dbuf_add_ref(parent
, db
);
2383 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2384 refcount_count(&dn
->dn_holds
) > 0);
2385 (void) refcount_add(&dn
->dn_holds
, db
);
2386 atomic_inc_32(&dn
->dn_dbufs_count
);
2388 dprintf_dbuf(db
, "db=%p\n", db
);
2393 typedef struct dbuf_prefetch_arg
{
2394 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2395 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2396 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2397 int dpa_curlevel
; /* The current level that we're reading */
2398 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2399 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2400 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2401 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2402 } dbuf_prefetch_arg_t
;
2405 * Actually issue the prefetch read for the block given.
2408 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2411 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2414 aflags
= dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2416 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2417 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2418 ASSERT(dpa
->dpa_zio
!= NULL
);
2419 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2420 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2421 &aflags
, &dpa
->dpa_zb
);
2425 * Called when an indirect block above our prefetch target is read in. This
2426 * will either read in the next indirect block down the tree or issue the actual
2427 * prefetch if the next block down is our target.
2430 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2432 dbuf_prefetch_arg_t
*dpa
= private;
2436 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2437 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2440 * The dpa_dnode is only valid if we are called with a NULL
2441 * zio. This indicates that the arc_read() returned without
2442 * first calling zio_read() to issue a physical read. Once
2443 * a physical read is made the dpa_dnode must be invalidated
2444 * as the locks guarding it may have been dropped. If the
2445 * dpa_dnode is still valid, then we want to add it to the dbuf
2446 * cache. To do so, we must hold the dbuf associated with the block
2447 * we just prefetched, read its contents so that we associate it
2448 * with an arc_buf_t, and then release it.
2451 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2452 if (zio
->io_flags
& ZIO_FLAG_RAW
) {
2453 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2455 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2457 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2459 dpa
->dpa_dnode
= NULL
;
2460 } else if (dpa
->dpa_dnode
!= NULL
) {
2461 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2462 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2463 dpa
->dpa_zb
.zb_level
));
2464 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2465 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2466 (void) dbuf_read(db
, NULL
,
2467 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2468 dbuf_rele(db
, FTAG
);
2471 dpa
->dpa_curlevel
--;
2473 nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2474 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2475 bp
= ((blkptr_t
*)abuf
->b_data
) +
2476 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2477 if (BP_IS_HOLE(bp
) || (zio
!= NULL
&& zio
->io_error
!= 0)) {
2478 kmem_free(dpa
, sizeof (*dpa
));
2479 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2480 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2481 dbuf_issue_final_prefetch(dpa
, bp
);
2482 kmem_free(dpa
, sizeof (*dpa
));
2484 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2485 zbookmark_phys_t zb
;
2487 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2489 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2490 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2492 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2493 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2494 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2498 arc_buf_destroy(abuf
, private);
2502 * Issue prefetch reads for the given block on the given level. If the indirect
2503 * blocks above that block are not in memory, we will read them in
2504 * asynchronously. As a result, this call never blocks waiting for a read to
2508 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2512 int epbs
, nlevels
, curlevel
;
2516 dbuf_prefetch_arg_t
*dpa
;
2519 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2520 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2522 if (blkid
> dn
->dn_maxblkid
)
2525 if (dnode_block_freed(dn
, blkid
))
2529 * This dnode hasn't been written to disk yet, so there's nothing to
2532 nlevels
= dn
->dn_phys
->dn_nlevels
;
2533 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2536 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2537 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2540 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2543 mutex_exit(&db
->db_mtx
);
2545 * This dbuf already exists. It is either CACHED, or
2546 * (we assume) about to be read or filled.
2552 * Find the closest ancestor (indirect block) of the target block
2553 * that is present in the cache. In this indirect block, we will
2554 * find the bp that is at curlevel, curblkid.
2558 while (curlevel
< nlevels
- 1) {
2559 int parent_level
= curlevel
+ 1;
2560 uint64_t parent_blkid
= curblkid
>> epbs
;
2563 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2564 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2565 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2566 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2567 dbuf_rele(db
, FTAG
);
2571 curlevel
= parent_level
;
2572 curblkid
= parent_blkid
;
2575 if (curlevel
== nlevels
- 1) {
2576 /* No cached indirect blocks found. */
2577 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2578 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2580 if (BP_IS_HOLE(&bp
))
2583 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2585 pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2588 dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2589 ds
= dn
->dn_objset
->os_dsl_dataset
;
2590 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2591 dn
->dn_object
, level
, blkid
);
2592 dpa
->dpa_curlevel
= curlevel
;
2593 dpa
->dpa_prio
= prio
;
2594 dpa
->dpa_aflags
= aflags
;
2595 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2596 dpa
->dpa_dnode
= dn
;
2597 dpa
->dpa_epbs
= epbs
;
2601 * If we have the indirect just above us, no need to do the asynchronous
2602 * prefetch chain; we'll just run the last step ourselves. If we're at
2603 * a higher level, though, we want to issue the prefetches for all the
2604 * indirect blocks asynchronously, so we can go on with whatever we were
2607 if (curlevel
== level
) {
2608 ASSERT3U(curblkid
, ==, blkid
);
2609 dbuf_issue_final_prefetch(dpa
, &bp
);
2610 kmem_free(dpa
, sizeof (*dpa
));
2612 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2613 zbookmark_phys_t zb
;
2615 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2616 dn
->dn_object
, curlevel
, curblkid
);
2617 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2618 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2619 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2623 * We use pio here instead of dpa_zio since it's possible that
2624 * dpa may have already been freed.
2629 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2632 * Returns with db_holds incremented, and db_mtx not held.
2633 * Note: dn_struct_rwlock must be held.
2636 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2638 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2639 dh
->dh_parent
= NULL
;
2641 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2642 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2643 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2645 *(dh
->dh_dbp
) = NULL
;
2647 /* dbuf_find() returns with db_mtx held */
2648 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2649 dh
->dh_level
, dh
->dh_blkid
);
2651 if (dh
->dh_db
== NULL
) {
2654 if (dh
->dh_fail_uncached
)
2655 return (SET_ERROR(ENOENT
));
2657 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2658 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2659 dh
->dh_fail_sparse
, &dh
->dh_parent
, &dh
->dh_bp
, dh
);
2660 if (dh
->dh_fail_sparse
) {
2661 if (dh
->dh_err
== 0 &&
2662 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2663 dh
->dh_err
= SET_ERROR(ENOENT
);
2666 dbuf_rele(dh
->dh_parent
, NULL
);
2667 return (dh
->dh_err
);
2670 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2671 return (dh
->dh_err
);
2672 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2673 dh
->dh_parent
, dh
->dh_bp
);
2676 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2677 mutex_exit(&dh
->dh_db
->db_mtx
);
2678 return (SET_ERROR(ENOENT
));
2681 if (dh
->dh_db
->db_buf
!= NULL
)
2682 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2684 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2687 * If this buffer is currently syncing out, and we are are
2688 * still referencing it from db_data, we need to make a copy
2689 * of it in case we decide we want to dirty it again in this txg.
2691 if (dh
->dh_db
->db_level
== 0 &&
2692 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2693 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2694 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2695 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2697 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2698 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2700 dbuf_set_data(dh
->dh_db
,
2701 arc_alloc_buf(dh
->dh_dn
->dn_objset
->os_spa
,
2702 dh
->dh_db
, dh
->dh_type
, dh
->dh_db
->db
.db_size
));
2703 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2704 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2708 if (multilist_link_active(&dh
->dh_db
->db_cache_link
)) {
2709 ASSERT(refcount_is_zero(&dh
->dh_db
->db_holds
));
2710 multilist_remove(dbuf_cache
, dh
->dh_db
);
2711 (void) refcount_remove_many(&dbuf_cache_size
,
2712 dh
->dh_db
->db
.db_size
, dh
->dh_db
);
2714 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2715 DBUF_VERIFY(dh
->dh_db
);
2716 mutex_exit(&dh
->dh_db
->db_mtx
);
2718 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2720 dbuf_rele(dh
->dh_parent
, NULL
);
2722 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2723 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2724 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2725 *(dh
->dh_dbp
) = dh
->dh_db
;
2731 * The following code preserves the recursive function dbuf_hold_impl()
2732 * but moves the local variables AND function arguments to the heap to
2733 * minimize the stack frame size. Enough space is initially allocated
2734 * on the stack for 20 levels of recursion.
2737 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2738 boolean_t fail_sparse
, boolean_t fail_uncached
,
2739 void *tag
, dmu_buf_impl_t
**dbp
)
2741 struct dbuf_hold_impl_data
*dh
;
2744 dh
= kmem_alloc(sizeof (struct dbuf_hold_impl_data
) *
2745 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2746 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2747 fail_uncached
, tag
, dbp
, 0);
2749 error
= __dbuf_hold_impl(dh
);
2751 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2752 DBUF_HOLD_IMPL_MAX_DEPTH
);
2758 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2759 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2760 boolean_t fail_sparse
, boolean_t fail_uncached
,
2761 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2764 dh
->dh_level
= level
;
2765 dh
->dh_blkid
= blkid
;
2767 dh
->dh_fail_sparse
= fail_sparse
;
2768 dh
->dh_fail_uncached
= fail_uncached
;
2774 dh
->dh_parent
= NULL
;
2780 dh
->dh_depth
= depth
;
2784 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2786 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2790 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2793 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2794 return (err
? NULL
: db
);
2798 dbuf_create_bonus(dnode_t
*dn
)
2800 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2802 ASSERT(dn
->dn_bonus
== NULL
);
2803 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2807 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2809 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2812 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2813 return (SET_ERROR(ENOTSUP
));
2815 blksz
= SPA_MINBLOCKSIZE
;
2816 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2817 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2821 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2822 dbuf_new_size(db
, blksz
, tx
);
2823 rw_exit(&dn
->dn_struct_rwlock
);
2830 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2832 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2835 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2837 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2839 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
2840 VERIFY3S(holds
, >, 1);
2843 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2845 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2848 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2849 dmu_buf_impl_t
*found_db
;
2850 boolean_t result
= B_FALSE
;
2852 if (blkid
== DMU_BONUS_BLKID
)
2853 found_db
= dbuf_find_bonus(os
, obj
);
2855 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2857 if (found_db
!= NULL
) {
2858 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2859 (void) refcount_add(&db
->db_holds
, tag
);
2862 mutex_exit(&found_db
->db_mtx
);
2868 * If you call dbuf_rele() you had better not be referencing the dnode handle
2869 * unless you have some other direct or indirect hold on the dnode. (An indirect
2870 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2871 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2872 * dnode's parent dbuf evicting its dnode handles.
2875 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2877 mutex_enter(&db
->db_mtx
);
2878 dbuf_rele_and_unlock(db
, tag
);
2882 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2884 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2888 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2889 * db_dirtycnt and db_holds to be updated atomically.
2892 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2896 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2900 * Remove the reference to the dbuf before removing its hold on the
2901 * dnode so we can guarantee in dnode_move() that a referenced bonus
2902 * buffer has a corresponding dnode hold.
2904 holds
= refcount_remove(&db
->db_holds
, tag
);
2908 * We can't freeze indirects if there is a possibility that they
2909 * may be modified in the current syncing context.
2911 if (db
->db_buf
!= NULL
&&
2912 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
2913 arc_buf_freeze(db
->db_buf
);
2916 if (holds
== db
->db_dirtycnt
&&
2917 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2918 dbuf_evict_user(db
);
2921 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2923 boolean_t evict_dbuf
= db
->db_pending_evict
;
2926 * If the dnode moves here, we cannot cross this
2927 * barrier until the move completes.
2932 atomic_dec_32(&dn
->dn_dbufs_count
);
2935 * Decrementing the dbuf count means that the bonus
2936 * buffer's dnode hold is no longer discounted in
2937 * dnode_move(). The dnode cannot move until after
2938 * the dnode_rele() below.
2943 * Do not reference db after its lock is dropped.
2944 * Another thread may evict it.
2946 mutex_exit(&db
->db_mtx
);
2949 dnode_evict_bonus(dn
);
2952 } else if (db
->db_buf
== NULL
) {
2954 * This is a special case: we never associated this
2955 * dbuf with any data allocated from the ARC.
2957 ASSERT(db
->db_state
== DB_UNCACHED
||
2958 db
->db_state
== DB_NOFILL
);
2960 } else if (arc_released(db
->db_buf
)) {
2962 * This dbuf has anonymous data associated with it.
2966 boolean_t do_arc_evict
= B_FALSE
;
2968 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
2970 if (!DBUF_IS_CACHEABLE(db
) &&
2971 db
->db_blkptr
!= NULL
&&
2972 !BP_IS_HOLE(db
->db_blkptr
) &&
2973 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2974 do_arc_evict
= B_TRUE
;
2975 bp
= *db
->db_blkptr
;
2978 if (!DBUF_IS_CACHEABLE(db
) ||
2979 db
->db_pending_evict
) {
2981 } else if (!multilist_link_active(&db
->db_cache_link
)) {
2982 multilist_insert(dbuf_cache
, db
);
2983 (void) refcount_add_many(&dbuf_cache_size
,
2984 db
->db
.db_size
, db
);
2985 mutex_exit(&db
->db_mtx
);
2987 dbuf_evict_notify();
2991 arc_freed(spa
, &bp
);
2994 mutex_exit(&db
->db_mtx
);
2999 #pragma weak dmu_buf_refcount = dbuf_refcount
3001 dbuf_refcount(dmu_buf_impl_t
*db
)
3003 return (refcount_count(&db
->db_holds
));
3007 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3008 dmu_buf_user_t
*new_user
)
3010 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3012 mutex_enter(&db
->db_mtx
);
3013 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3014 if (db
->db_user
== old_user
)
3015 db
->db_user
= new_user
;
3017 old_user
= db
->db_user
;
3018 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3019 mutex_exit(&db
->db_mtx
);
3025 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3027 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3031 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3033 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3035 db
->db_user_immediate_evict
= TRUE
;
3036 return (dmu_buf_set_user(db_fake
, user
));
3040 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3042 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3046 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3048 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3050 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3051 return (db
->db_user
);
3055 dmu_buf_user_evict_wait()
3057 taskq_wait(dbu_evict_taskq
);
3061 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3063 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3064 return (dbi
->db_blkptr
);
3068 dmu_buf_get_objset(dmu_buf_t
*db
)
3070 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3071 return (dbi
->db_objset
);
3075 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3077 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3078 DB_DNODE_ENTER(dbi
);
3079 return (DB_DNODE(dbi
));
3083 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3085 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3090 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3092 /* ASSERT(dmu_tx_is_syncing(tx) */
3093 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3095 if (db
->db_blkptr
!= NULL
)
3098 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3099 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3100 BP_ZERO(db
->db_blkptr
);
3103 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3105 * This buffer was allocated at a time when there was
3106 * no available blkptrs from the dnode, or it was
3107 * inappropriate to hook it in (i.e., nlevels mis-match).
3109 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3110 ASSERT(db
->db_parent
== NULL
);
3111 db
->db_parent
= dn
->dn_dbuf
;
3112 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3115 dmu_buf_impl_t
*parent
= db
->db_parent
;
3116 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3118 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3119 if (parent
== NULL
) {
3120 mutex_exit(&db
->db_mtx
);
3121 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3122 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3123 db
->db_blkid
>> epbs
, db
);
3124 rw_exit(&dn
->dn_struct_rwlock
);
3125 mutex_enter(&db
->db_mtx
);
3126 db
->db_parent
= parent
;
3128 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3129 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3135 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3136 * is critical the we not allow the compiler to inline this function in to
3137 * dbuf_sync_list() thereby drastically bloating the stack usage.
3139 noinline
static void
3140 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3142 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3146 ASSERT(dmu_tx_is_syncing(tx
));
3148 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3150 mutex_enter(&db
->db_mtx
);
3152 ASSERT(db
->db_level
> 0);
3155 /* Read the block if it hasn't been read yet. */
3156 if (db
->db_buf
== NULL
) {
3157 mutex_exit(&db
->db_mtx
);
3158 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3159 mutex_enter(&db
->db_mtx
);
3161 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3162 ASSERT(db
->db_buf
!= NULL
);
3166 /* Indirect block size must match what the dnode thinks it is. */
3167 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3168 dbuf_check_blkptr(dn
, db
);
3171 /* Provide the pending dirty record to child dbufs */
3172 db
->db_data_pending
= dr
;
3174 mutex_exit(&db
->db_mtx
);
3175 dbuf_write(dr
, db
->db_buf
, tx
);
3178 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3179 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3180 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3181 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3186 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3187 * critical the we not allow the compiler to inline this function in to
3188 * dbuf_sync_list() thereby drastically bloating the stack usage.
3190 noinline
static void
3191 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3193 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3194 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3197 uint64_t txg
= tx
->tx_txg
;
3199 ASSERT(dmu_tx_is_syncing(tx
));
3201 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3203 mutex_enter(&db
->db_mtx
);
3205 * To be synced, we must be dirtied. But we
3206 * might have been freed after the dirty.
3208 if (db
->db_state
== DB_UNCACHED
) {
3209 /* This buffer has been freed since it was dirtied */
3210 ASSERT(db
->db
.db_data
== NULL
);
3211 } else if (db
->db_state
== DB_FILL
) {
3212 /* This buffer was freed and is now being re-filled */
3213 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3215 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3222 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3223 mutex_enter(&dn
->dn_mtx
);
3224 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3226 * In the previous transaction group, the bonus buffer
3227 * was entirely used to store the attributes for the
3228 * dnode which overrode the dn_spill field. However,
3229 * when adding more attributes to the file a spill
3230 * block was required to hold the extra attributes.
3232 * Make sure to clear the garbage left in the dn_spill
3233 * field from the previous attributes in the bonus
3234 * buffer. Otherwise, after writing out the spill
3235 * block to the new allocated dva, it will free
3236 * the old block pointed to by the invalid dn_spill.
3238 db
->db_blkptr
= NULL
;
3240 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3241 mutex_exit(&dn
->dn_mtx
);
3245 * If this is a bonus buffer, simply copy the bonus data into the
3246 * dnode. It will be written out when the dnode is synced (and it
3247 * will be synced, since it must have been dirty for dbuf_sync to
3250 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3251 dbuf_dirty_record_t
**drp
;
3253 ASSERT(*datap
!= NULL
);
3254 ASSERT0(db
->db_level
);
3255 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=,
3256 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3257 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
3260 if (*datap
!= db
->db
.db_data
) {
3261 int slots
= DB_DNODE(db
)->dn_num_slots
;
3262 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3263 kmem_free(*datap
, bonuslen
);
3264 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3266 db
->db_data_pending
= NULL
;
3267 drp
= &db
->db_last_dirty
;
3269 drp
= &(*drp
)->dr_next
;
3270 ASSERT(dr
->dr_next
== NULL
);
3271 ASSERT(dr
->dr_dbuf
== db
);
3273 if (dr
->dr_dbuf
->db_level
!= 0) {
3274 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3275 list_destroy(&dr
->dt
.di
.dr_children
);
3277 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3278 ASSERT(db
->db_dirtycnt
> 0);
3279 db
->db_dirtycnt
-= 1;
3280 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3287 * This function may have dropped the db_mtx lock allowing a dmu_sync
3288 * operation to sneak in. As a result, we need to ensure that we
3289 * don't check the dr_override_state until we have returned from
3290 * dbuf_check_blkptr.
3292 dbuf_check_blkptr(dn
, db
);
3295 * If this buffer is in the middle of an immediate write,
3296 * wait for the synchronous IO to complete.
3298 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3299 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3300 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3301 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3304 if (db
->db_state
!= DB_NOFILL
&&
3305 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3306 refcount_count(&db
->db_holds
) > 1 &&
3307 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3308 *datap
== db
->db_buf
) {
3310 * If this buffer is currently "in use" (i.e., there
3311 * are active holds and db_data still references it),
3312 * then make a copy before we start the write so that
3313 * any modifications from the open txg will not leak
3316 * NOTE: this copy does not need to be made for
3317 * objects only modified in the syncing context (e.g.
3318 * DNONE_DNODE blocks).
3320 int psize
= arc_buf_size(*datap
);
3321 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3322 enum zio_compress compress_type
= arc_get_compression(*datap
);
3324 if (compress_type
== ZIO_COMPRESS_OFF
) {
3325 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
3327 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
3328 int lsize
= arc_buf_lsize(*datap
);
3329 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
3330 psize
, lsize
, compress_type
);
3332 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
3334 db
->db_data_pending
= dr
;
3336 mutex_exit(&db
->db_mtx
);
3338 dbuf_write(dr
, *datap
, tx
);
3340 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3341 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3342 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3346 * Although zio_nowait() does not "wait for an IO", it does
3347 * initiate the IO. If this is an empty write it seems plausible
3348 * that the IO could actually be completed before the nowait
3349 * returns. We need to DB_DNODE_EXIT() first in case
3350 * zio_nowait() invalidates the dbuf.
3353 zio_nowait(dr
->dr_zio
);
3358 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3360 dbuf_dirty_record_t
*dr
;
3362 while ((dr
= list_head(list
))) {
3363 if (dr
->dr_zio
!= NULL
) {
3365 * If we find an already initialized zio then we
3366 * are processing the meta-dnode, and we have finished.
3367 * The dbufs for all dnodes are put back on the list
3368 * during processing, so that we can zio_wait()
3369 * these IOs after initiating all child IOs.
3371 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3372 DMU_META_DNODE_OBJECT
);
3375 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3376 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3377 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3379 list_remove(list
, dr
);
3380 if (dr
->dr_dbuf
->db_level
> 0)
3381 dbuf_sync_indirect(dr
, tx
);
3383 dbuf_sync_leaf(dr
, tx
);
3389 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3391 dmu_buf_impl_t
*db
= vdb
;
3393 blkptr_t
*bp
= zio
->io_bp
;
3394 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3395 spa_t
*spa
= zio
->io_spa
;
3400 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3401 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3405 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3406 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3407 zio
->io_prev_space_delta
= delta
;
3409 if (bp
->blk_birth
!= 0) {
3410 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3411 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3412 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3413 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3414 BP_IS_EMBEDDED(bp
));
3415 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3418 mutex_enter(&db
->db_mtx
);
3421 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3422 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3423 ASSERT(!(BP_IS_HOLE(bp
)) &&
3424 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3428 if (db
->db_level
== 0) {
3429 mutex_enter(&dn
->dn_mtx
);
3430 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3431 db
->db_blkid
!= DMU_SPILL_BLKID
)
3432 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3433 mutex_exit(&dn
->dn_mtx
);
3435 if (dn
->dn_type
== DMU_OT_DNODE
) {
3437 while (i
< db
->db
.db_size
) {
3439 (void *)(((char *)db
->db
.db_data
) + i
);
3441 i
+= DNODE_MIN_SIZE
;
3442 if (dnp
->dn_type
!= DMU_OT_NONE
) {
3444 i
+= dnp
->dn_extra_slots
*
3449 if (BP_IS_HOLE(bp
)) {
3456 blkptr_t
*ibp
= db
->db
.db_data
;
3457 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3458 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3459 if (BP_IS_HOLE(ibp
))
3461 fill
+= BP_GET_FILL(ibp
);
3466 if (!BP_IS_EMBEDDED(bp
))
3467 bp
->blk_fill
= fill
;
3469 mutex_exit(&db
->db_mtx
);
3471 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3472 *db
->db_blkptr
= *bp
;
3473 rw_exit(&dn
->dn_struct_rwlock
);
3478 * This function gets called just prior to running through the compression
3479 * stage of the zio pipeline. If we're an indirect block comprised of only
3480 * holes, then we want this indirect to be compressed away to a hole. In
3481 * order to do that we must zero out any information about the holes that
3482 * this indirect points to prior to before we try to compress it.
3485 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3487 dmu_buf_impl_t
*db
= vdb
;
3490 unsigned int epbs
, i
;
3492 ASSERT3U(db
->db_level
, >, 0);
3495 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3496 ASSERT3U(epbs
, <, 31);
3498 /* Determine if all our children are holes */
3499 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
3500 if (!BP_IS_HOLE(bp
))
3505 * If all the children are holes, then zero them all out so that
3506 * we may get compressed away.
3508 if (i
== 1ULL << epbs
) {
3510 * We only found holes. Grab the rwlock to prevent
3511 * anybody from reading the blocks we're about to
3514 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3515 bzero(db
->db
.db_data
, db
->db
.db_size
);
3516 rw_exit(&dn
->dn_struct_rwlock
);
3522 * The SPA will call this callback several times for each zio - once
3523 * for every physical child i/o (zio->io_phys_children times). This
3524 * allows the DMU to monitor the progress of each logical i/o. For example,
3525 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3526 * block. There may be a long delay before all copies/fragments are completed,
3527 * so this callback allows us to retire dirty space gradually, as the physical
3532 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3534 dmu_buf_impl_t
*db
= arg
;
3535 objset_t
*os
= db
->db_objset
;
3536 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3537 dbuf_dirty_record_t
*dr
;
3540 dr
= db
->db_data_pending
;
3541 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3544 * The callback will be called io_phys_children times. Retire one
3545 * portion of our dirty space each time we are called. Any rounding
3546 * error will be cleaned up by dsl_pool_sync()'s call to
3547 * dsl_pool_undirty_space().
3549 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3550 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3555 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3557 dmu_buf_impl_t
*db
= vdb
;
3558 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3559 blkptr_t
*bp
= db
->db_blkptr
;
3560 objset_t
*os
= db
->db_objset
;
3561 dmu_tx_t
*tx
= os
->os_synctx
;
3562 dbuf_dirty_record_t
**drp
, *dr
;
3564 ASSERT0(zio
->io_error
);
3565 ASSERT(db
->db_blkptr
== bp
);
3568 * For nopwrites and rewrites we ensure that the bp matches our
3569 * original and bypass all the accounting.
3571 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3572 ASSERT(BP_EQUAL(bp
, bp_orig
));
3574 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3575 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3576 dsl_dataset_block_born(ds
, bp
, tx
);
3579 mutex_enter(&db
->db_mtx
);
3583 drp
= &db
->db_last_dirty
;
3584 while ((dr
= *drp
) != db
->db_data_pending
)
3586 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3587 ASSERT(dr
->dr_dbuf
== db
);
3588 ASSERT(dr
->dr_next
== NULL
);
3592 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3597 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3598 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3599 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3604 if (db
->db_level
== 0) {
3605 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3606 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3607 if (db
->db_state
!= DB_NOFILL
) {
3608 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3609 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
3616 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3617 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3618 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3619 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3621 ASSERT3U(db
->db_blkid
, <=,
3622 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3623 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3627 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3628 list_destroy(&dr
->dt
.di
.dr_children
);
3630 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3632 cv_broadcast(&db
->db_changed
);
3633 ASSERT(db
->db_dirtycnt
> 0);
3634 db
->db_dirtycnt
-= 1;
3635 db
->db_data_pending
= NULL
;
3636 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3640 dbuf_write_nofill_ready(zio_t
*zio
)
3642 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3646 dbuf_write_nofill_done(zio_t
*zio
)
3648 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3652 dbuf_write_override_ready(zio_t
*zio
)
3654 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3655 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3657 dbuf_write_ready(zio
, NULL
, db
);
3661 dbuf_write_override_done(zio_t
*zio
)
3663 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3664 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3665 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3667 mutex_enter(&db
->db_mtx
);
3668 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3669 if (!BP_IS_HOLE(obp
))
3670 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3671 arc_release(dr
->dt
.dl
.dr_data
, db
);
3673 mutex_exit(&db
->db_mtx
);
3675 dbuf_write_done(zio
, NULL
, db
);
3677 if (zio
->io_abd
!= NULL
)
3678 abd_put(zio
->io_abd
);
3681 /* Issue I/O to commit a dirty buffer to disk. */
3683 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3685 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3688 dmu_buf_impl_t
*parent
= db
->db_parent
;
3689 uint64_t txg
= tx
->tx_txg
;
3690 zbookmark_phys_t zb
;
3695 ASSERT(dmu_tx_is_syncing(tx
));
3701 if (db
->db_state
!= DB_NOFILL
) {
3702 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3704 * Private object buffers are released here rather
3705 * than in dbuf_dirty() since they are only modified
3706 * in the syncing context and we don't want the
3707 * overhead of making multiple copies of the data.
3709 if (BP_IS_HOLE(db
->db_blkptr
)) {
3712 dbuf_release_bp(db
);
3717 if (parent
!= dn
->dn_dbuf
) {
3718 /* Our parent is an indirect block. */
3719 /* We have a dirty parent that has been scheduled for write. */
3720 ASSERT(parent
&& parent
->db_data_pending
);
3721 /* Our parent's buffer is one level closer to the dnode. */
3722 ASSERT(db
->db_level
== parent
->db_level
-1);
3724 * We're about to modify our parent's db_data by modifying
3725 * our block pointer, so the parent must be released.
3727 ASSERT(arc_released(parent
->db_buf
));
3728 zio
= parent
->db_data_pending
->dr_zio
;
3730 /* Our parent is the dnode itself. */
3731 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3732 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3733 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3734 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3735 ASSERT3P(db
->db_blkptr
, ==,
3736 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3740 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3741 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3744 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3745 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3746 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3748 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3750 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3752 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3756 * We copy the blkptr now (rather than when we instantiate the dirty
3757 * record), because its value can change between open context and
3758 * syncing context. We do not need to hold dn_struct_rwlock to read
3759 * db_blkptr because we are in syncing context.
3761 dr
->dr_bp_copy
= *db
->db_blkptr
;
3763 if (db
->db_level
== 0 &&
3764 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3766 * The BP for this block has been provided by open context
3767 * (by dmu_sync() or dmu_buf_write_embedded()).
3769 abd_t
*contents
= (data
!= NULL
) ?
3770 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
3772 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3773 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
3774 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
3775 dbuf_write_override_done
,
3776 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3777 mutex_enter(&db
->db_mtx
);
3778 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3779 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3780 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3781 mutex_exit(&db
->db_mtx
);
3782 } else if (db
->db_state
== DB_NOFILL
) {
3783 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
3784 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
3785 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3786 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
3787 dbuf_write_nofill_ready
, NULL
, NULL
,
3788 dbuf_write_nofill_done
, db
,
3789 ZIO_PRIORITY_ASYNC_WRITE
,
3790 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3792 arc_done_func_t
*children_ready_cb
= NULL
;
3793 ASSERT(arc_released(data
));
3796 * For indirect blocks, we want to setup the children
3797 * ready callback so that we can properly handle an indirect
3798 * block that only contains holes.
3800 if (db
->db_level
!= 0)
3801 children_ready_cb
= dbuf_write_children_ready
;
3803 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3804 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3805 &zp
, dbuf_write_ready
,
3806 children_ready_cb
, dbuf_write_physdone
,
3807 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
3808 ZIO_FLAG_MUSTSUCCEED
, &zb
);
3812 #if defined(_KERNEL) && defined(HAVE_SPL)
3813 EXPORT_SYMBOL(dbuf_find
);
3814 EXPORT_SYMBOL(dbuf_is_metadata
);
3815 EXPORT_SYMBOL(dbuf_destroy
);
3816 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
3817 EXPORT_SYMBOL(dbuf_whichblock
);
3818 EXPORT_SYMBOL(dbuf_read
);
3819 EXPORT_SYMBOL(dbuf_unoverride
);
3820 EXPORT_SYMBOL(dbuf_free_range
);
3821 EXPORT_SYMBOL(dbuf_new_size
);
3822 EXPORT_SYMBOL(dbuf_release_bp
);
3823 EXPORT_SYMBOL(dbuf_dirty
);
3824 EXPORT_SYMBOL(dmu_buf_will_dirty
);
3825 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
3826 EXPORT_SYMBOL(dmu_buf_will_fill
);
3827 EXPORT_SYMBOL(dmu_buf_fill_done
);
3828 EXPORT_SYMBOL(dmu_buf_rele
);
3829 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
3830 EXPORT_SYMBOL(dbuf_prefetch
);
3831 EXPORT_SYMBOL(dbuf_hold_impl
);
3832 EXPORT_SYMBOL(dbuf_hold
);
3833 EXPORT_SYMBOL(dbuf_hold_level
);
3834 EXPORT_SYMBOL(dbuf_create_bonus
);
3835 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
3836 EXPORT_SYMBOL(dbuf_rm_spill
);
3837 EXPORT_SYMBOL(dbuf_add_ref
);
3838 EXPORT_SYMBOL(dbuf_rele
);
3839 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
3840 EXPORT_SYMBOL(dbuf_refcount
);
3841 EXPORT_SYMBOL(dbuf_sync_list
);
3842 EXPORT_SYMBOL(dmu_buf_set_user
);
3843 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3844 EXPORT_SYMBOL(dmu_buf_get_user
);
3845 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
3848 module_param(dbuf_cache_max_bytes
, ulong
, 0644);
3849 MODULE_PARM_DESC(dbuf_cache_max_bytes
,
3850 "Maximum size in bytes of the dbuf cache.");
3852 module_param(dbuf_cache_hiwater_pct
, uint
, 0644);
3853 MODULE_PARM_DESC(dbuf_cache_hiwater_pct
,
3854 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
3857 module_param(dbuf_cache_lowater_pct
, uint
, 0644);
3858 MODULE_PARM_DESC(dbuf_cache_lowater_pct
,
3859 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
3862 module_param(dbuf_cache_max_shift
, int, 0644);
3863 MODULE_PARM_DESC(dbuf_cache_max_shift
,
3864 "Cap the size of the dbuf cache to a log2 fraction of arc size.");