4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_dbuf.h>
48 #include <sys/callb.h>
51 struct dbuf_hold_impl_data
{
52 /* Function arguments */
56 boolean_t dh_fail_sparse
;
57 boolean_t dh_fail_uncached
;
59 dmu_buf_impl_t
**dh_dbp
;
61 dmu_buf_impl_t
*dh_db
;
62 dmu_buf_impl_t
*dh_parent
;
65 dbuf_dirty_record_t
*dh_dr
;
66 arc_buf_contents_t dh_type
;
70 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
71 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, boolean_t fail_sparse
,
72 boolean_t fail_uncached
,
73 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
74 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
76 uint_t zfs_dbuf_evict_key
;
78 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
79 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
81 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
82 dmu_buf_evict_func_t
*evict_func_sync
,
83 dmu_buf_evict_func_t
*evict_func_async
,
84 dmu_buf_t
**clear_on_evict_dbufp
);
87 * Global data structures and functions for the dbuf cache.
89 static kmem_cache_t
*dbuf_kmem_cache
;
90 static taskq_t
*dbu_evict_taskq
;
92 static kthread_t
*dbuf_cache_evict_thread
;
93 static kmutex_t dbuf_evict_lock
;
94 static kcondvar_t dbuf_evict_cv
;
95 static boolean_t dbuf_evict_thread_exit
;
98 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
99 * are not currently held but have been recently released. These dbufs
100 * are not eligible for arc eviction until they are aged out of the cache.
101 * Dbufs are added to the dbuf cache once the last hold is released. If a
102 * dbuf is later accessed and still exists in the dbuf cache, then it will
103 * be removed from the cache and later re-added to the head of the cache.
104 * Dbufs that are aged out of the cache will be immediately destroyed and
105 * become eligible for arc eviction.
107 static multilist_t
*dbuf_cache
;
108 static refcount_t dbuf_cache_size
;
109 unsigned long dbuf_cache_max_bytes
= 100 * 1024 * 1024;
111 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
112 int dbuf_cache_max_shift
= 5;
115 * The dbuf cache uses a three-stage eviction policy:
116 * - A low water marker designates when the dbuf eviction thread
117 * should stop evicting from the dbuf cache.
118 * - When we reach the maximum size (aka mid water mark), we
119 * signal the eviction thread to run.
120 * - The high water mark indicates when the eviction thread
121 * is unable to keep up with the incoming load and eviction must
122 * happen in the context of the calling thread.
126 * low water mid water hi water
127 * +----------------------------------------+----------+----------+
132 * +----------------------------------------+----------+----------+
134 * evicting eviction directly
137 * The high and low water marks indicate the operating range for the eviction
138 * thread. The low water mark is, by default, 90% of the total size of the
139 * cache and the high water mark is at 110% (both of these percentages can be
140 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
141 * respectively). The eviction thread will try to ensure that the cache remains
142 * within this range by waking up every second and checking if the cache is
143 * above the low water mark. The thread can also be woken up by callers adding
144 * elements into the cache if the cache is larger than the mid water (i.e max
145 * cache size). Once the eviction thread is woken up and eviction is required,
146 * it will continue evicting buffers until it's able to reduce the cache size
147 * to the low water mark. If the cache size continues to grow and hits the high
148 * water mark, then callers adding elements to the cache will begin to evict
149 * directly from the cache until the cache is no longer above the high water
154 * The percentage above and below the maximum cache size.
156 uint_t dbuf_cache_hiwater_pct
= 10;
157 uint_t dbuf_cache_lowater_pct
= 10;
161 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
163 dmu_buf_impl_t
*db
= vdb
;
164 bzero(db
, sizeof (dmu_buf_impl_t
));
166 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
167 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
168 multilist_link_init(&db
->db_cache_link
);
169 refcount_create(&db
->db_holds
);
170 multilist_link_init(&db
->db_cache_link
);
177 dbuf_dest(void *vdb
, void *unused
)
179 dmu_buf_impl_t
*db
= vdb
;
180 mutex_destroy(&db
->db_mtx
);
181 cv_destroy(&db
->db_changed
);
182 ASSERT(!multilist_link_active(&db
->db_cache_link
));
183 refcount_destroy(&db
->db_holds
);
187 * dbuf hash table routines
189 static dbuf_hash_table_t dbuf_hash_table
;
191 static uint64_t dbuf_hash_count
;
194 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
196 uintptr_t osv
= (uintptr_t)os
;
197 uint64_t crc
= -1ULL;
199 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
200 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
201 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
202 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
203 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
204 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
205 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
207 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
212 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
213 ((dbuf)->db.db_object == (obj) && \
214 (dbuf)->db_objset == (os) && \
215 (dbuf)->db_level == (level) && \
216 (dbuf)->db_blkid == (blkid))
219 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
221 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
226 hv
= dbuf_hash(os
, obj
, level
, blkid
);
227 idx
= hv
& h
->hash_table_mask
;
229 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
230 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
231 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
232 mutex_enter(&db
->db_mtx
);
233 if (db
->db_state
!= DB_EVICTING
) {
234 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
237 mutex_exit(&db
->db_mtx
);
240 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
244 static dmu_buf_impl_t
*
245 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
248 dmu_buf_impl_t
*db
= NULL
;
250 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
251 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
252 if (dn
->dn_bonus
!= NULL
) {
254 mutex_enter(&db
->db_mtx
);
256 rw_exit(&dn
->dn_struct_rwlock
);
257 dnode_rele(dn
, FTAG
);
263 * Insert an entry into the hash table. If there is already an element
264 * equal to elem in the hash table, then the already existing element
265 * will be returned and the new element will not be inserted.
266 * Otherwise returns NULL.
268 static dmu_buf_impl_t
*
269 dbuf_hash_insert(dmu_buf_impl_t
*db
)
271 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
272 objset_t
*os
= db
->db_objset
;
273 uint64_t obj
= db
->db
.db_object
;
274 int level
= db
->db_level
;
275 uint64_t blkid
, hv
, idx
;
278 blkid
= db
->db_blkid
;
279 hv
= dbuf_hash(os
, obj
, level
, blkid
);
280 idx
= hv
& h
->hash_table_mask
;
282 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
283 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
284 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
285 mutex_enter(&dbf
->db_mtx
);
286 if (dbf
->db_state
!= DB_EVICTING
) {
287 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
290 mutex_exit(&dbf
->db_mtx
);
294 mutex_enter(&db
->db_mtx
);
295 db
->db_hash_next
= h
->hash_table
[idx
];
296 h
->hash_table
[idx
] = db
;
297 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
298 atomic_inc_64(&dbuf_hash_count
);
304 * Remove an entry from the hash table. It must be in the EVICTING state.
307 dbuf_hash_remove(dmu_buf_impl_t
*db
)
309 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
311 dmu_buf_impl_t
*dbf
, **dbp
;
313 hv
= dbuf_hash(db
->db_objset
, db
->db
.db_object
,
314 db
->db_level
, db
->db_blkid
);
315 idx
= hv
& h
->hash_table_mask
;
318 * We mustn't hold db_mtx to maintain lock ordering:
319 * DBUF_HASH_MUTEX > db_mtx.
321 ASSERT(refcount_is_zero(&db
->db_holds
));
322 ASSERT(db
->db_state
== DB_EVICTING
);
323 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
325 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
326 dbp
= &h
->hash_table
[idx
];
327 while ((dbf
= *dbp
) != db
) {
328 dbp
= &dbf
->db_hash_next
;
331 *dbp
= db
->db_hash_next
;
332 db
->db_hash_next
= NULL
;
333 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
334 atomic_dec_64(&dbuf_hash_count
);
340 } dbvu_verify_type_t
;
343 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
348 if (db
->db_user
== NULL
)
351 /* Only data blocks support the attachment of user data. */
352 ASSERT(db
->db_level
== 0);
354 /* Clients must resolve a dbuf before attaching user data. */
355 ASSERT(db
->db
.db_data
!= NULL
);
356 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
358 holds
= refcount_count(&db
->db_holds
);
359 if (verify_type
== DBVU_EVICTING
) {
361 * Immediate eviction occurs when holds == dirtycnt.
362 * For normal eviction buffers, holds is zero on
363 * eviction, except when dbuf_fix_old_data() calls
364 * dbuf_clear_data(). However, the hold count can grow
365 * during eviction even though db_mtx is held (see
366 * dmu_bonus_hold() for an example), so we can only
367 * test the generic invariant that holds >= dirtycnt.
369 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
371 if (db
->db_user_immediate_evict
== TRUE
)
372 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
374 ASSERT3U(holds
, >, 0);
380 dbuf_evict_user(dmu_buf_impl_t
*db
)
382 dmu_buf_user_t
*dbu
= db
->db_user
;
384 ASSERT(MUTEX_HELD(&db
->db_mtx
));
389 dbuf_verify_user(db
, DBVU_EVICTING
);
393 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
394 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
398 * There are two eviction callbacks - one that we call synchronously
399 * and one that we invoke via a taskq. The async one is useful for
400 * avoiding lock order reversals and limiting stack depth.
402 * Note that if we have a sync callback but no async callback,
403 * it's likely that the sync callback will free the structure
404 * containing the dbu. In that case we need to take care to not
405 * dereference dbu after calling the sync evict func.
407 boolean_t has_async
= (dbu
->dbu_evict_func_async
!= NULL
);
409 if (dbu
->dbu_evict_func_sync
!= NULL
)
410 dbu
->dbu_evict_func_sync(dbu
);
413 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func_async
,
414 dbu
, 0, &dbu
->dbu_tqent
);
419 dbuf_is_metadata(dmu_buf_impl_t
*db
)
422 * Consider indirect blocks and spill blocks to be meta data.
424 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
427 boolean_t is_metadata
;
430 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
433 return (is_metadata
);
439 * This function *must* return indices evenly distributed between all
440 * sublists of the multilist. This is needed due to how the dbuf eviction
441 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
442 * distributed between all sublists and uses this assumption when
443 * deciding which sublist to evict from and how much to evict from it.
446 dbuf_cache_multilist_index_func(multilist_t
*ml
, void *obj
)
448 dmu_buf_impl_t
*db
= obj
;
451 * The assumption here, is the hash value for a given
452 * dmu_buf_impl_t will remain constant throughout it's lifetime
453 * (i.e. it's objset, object, level and blkid fields don't change).
454 * Thus, we don't need to store the dbuf's sublist index
455 * on insertion, as this index can be recalculated on removal.
457 * Also, the low order bits of the hash value are thought to be
458 * distributed evenly. Otherwise, in the case that the multilist
459 * has a power of two number of sublists, each sublists' usage
460 * would not be evenly distributed.
462 return (dbuf_hash(db
->db_objset
, db
->db
.db_object
,
463 db
->db_level
, db
->db_blkid
) %
464 multilist_get_num_sublists(ml
));
467 static inline boolean_t
468 dbuf_cache_above_hiwater(void)
470 uint64_t dbuf_cache_hiwater_bytes
=
471 (dbuf_cache_max_bytes
* dbuf_cache_hiwater_pct
) / 100;
473 return (refcount_count(&dbuf_cache_size
) >
474 dbuf_cache_max_bytes
+ dbuf_cache_hiwater_bytes
);
477 static inline boolean_t
478 dbuf_cache_above_lowater(void)
480 uint64_t dbuf_cache_lowater_bytes
=
481 (dbuf_cache_max_bytes
* dbuf_cache_lowater_pct
) / 100;
483 return (refcount_count(&dbuf_cache_size
) >
484 dbuf_cache_max_bytes
- dbuf_cache_lowater_bytes
);
488 * Evict the oldest eligible dbuf from the dbuf cache.
493 int idx
= multilist_get_random_index(dbuf_cache
);
494 multilist_sublist_t
*mls
= multilist_sublist_lock(dbuf_cache
, idx
);
496 ASSERT(!MUTEX_HELD(&dbuf_evict_lock
));
499 * Set the thread's tsd to indicate that it's processing evictions.
500 * Once a thread stops evicting from the dbuf cache it will
501 * reset its tsd to NULL.
503 ASSERT3P(tsd_get(zfs_dbuf_evict_key
), ==, NULL
);
504 (void) tsd_set(zfs_dbuf_evict_key
, (void *)B_TRUE
);
506 db
= multilist_sublist_tail(mls
);
507 while (db
!= NULL
&& mutex_tryenter(&db
->db_mtx
) == 0) {
508 db
= multilist_sublist_prev(mls
, db
);
511 DTRACE_PROBE2(dbuf__evict__one
, dmu_buf_impl_t
*, db
,
512 multilist_sublist_t
*, mls
);
515 multilist_sublist_remove(mls
, db
);
516 multilist_sublist_unlock(mls
);
517 (void) refcount_remove_many(&dbuf_cache_size
,
521 multilist_sublist_unlock(mls
);
523 (void) tsd_set(zfs_dbuf_evict_key
, NULL
);
527 * The dbuf evict thread is responsible for aging out dbufs from the
528 * cache. Once the cache has reached it's maximum size, dbufs are removed
529 * and destroyed. The eviction thread will continue running until the size
530 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
531 * out of the cache it is destroyed and becomes eligible for arc eviction.
534 dbuf_evict_thread(void *unused
)
538 CALLB_CPR_INIT(&cpr
, &dbuf_evict_lock
, callb_generic_cpr
, FTAG
);
540 mutex_enter(&dbuf_evict_lock
);
541 while (!dbuf_evict_thread_exit
) {
542 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
543 CALLB_CPR_SAFE_BEGIN(&cpr
);
544 (void) cv_timedwait_sig_hires(&dbuf_evict_cv
,
545 &dbuf_evict_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
546 CALLB_CPR_SAFE_END(&cpr
, &dbuf_evict_lock
);
548 mutex_exit(&dbuf_evict_lock
);
551 * Keep evicting as long as we're above the low water mark
552 * for the cache. We do this without holding the locks to
553 * minimize lock contention.
555 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit
) {
559 mutex_enter(&dbuf_evict_lock
);
562 dbuf_evict_thread_exit
= B_FALSE
;
563 cv_broadcast(&dbuf_evict_cv
);
564 CALLB_CPR_EXIT(&cpr
); /* drops dbuf_evict_lock */
569 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
570 * If the dbuf cache is at its high water mark, then evict a dbuf from the
571 * dbuf cache using the callers context.
574 dbuf_evict_notify(void)
578 * We use thread specific data to track when a thread has
579 * started processing evictions. This allows us to avoid deeply
580 * nested stacks that would have a call flow similar to this:
582 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
585 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
587 * The dbuf_eviction_thread will always have its tsd set until
588 * that thread exits. All other threads will only set their tsd
589 * if they are participating in the eviction process. This only
590 * happens if the eviction thread is unable to process evictions
591 * fast enough. To keep the dbuf cache size in check, other threads
592 * can evict from the dbuf cache directly. Those threads will set
593 * their tsd values so that we ensure that they only evict one dbuf
594 * from the dbuf cache.
596 if (tsd_get(zfs_dbuf_evict_key
) != NULL
)
600 * We check if we should evict without holding the dbuf_evict_lock,
601 * because it's OK to occasionally make the wrong decision here,
602 * and grabbing the lock results in massive lock contention.
604 if (refcount_count(&dbuf_cache_size
) > dbuf_cache_max_bytes
) {
605 if (dbuf_cache_above_hiwater())
607 cv_signal(&dbuf_evict_cv
);
616 uint64_t hsize
= 1ULL << 16;
617 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
621 * The hash table is big enough to fill all of physical memory
622 * with an average block size of zfs_arc_average_blocksize (default 8K).
623 * By default, the table will take up
624 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
626 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
630 h
->hash_table_mask
= hsize
- 1;
631 #if defined(_KERNEL) && defined(HAVE_SPL)
633 * Large allocations which do not require contiguous pages
634 * should be using vmem_alloc() in the linux kernel
636 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
638 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
640 if (h
->hash_table
== NULL
) {
641 /* XXX - we should really return an error instead of assert */
642 ASSERT(hsize
> (1ULL << 10));
647 dbuf_kmem_cache
= kmem_cache_create("dmu_buf_impl_t",
648 sizeof (dmu_buf_impl_t
),
649 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
651 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
652 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
657 * Setup the parameters for the dbuf cache. We cap the size of the
658 * dbuf cache to 1/32nd (default) of the size of the ARC.
660 dbuf_cache_max_bytes
= MIN(dbuf_cache_max_bytes
,
661 arc_max_bytes() >> dbuf_cache_max_shift
);
664 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
665 * configuration is not required.
667 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
669 dbuf_cache
= multilist_create(sizeof (dmu_buf_impl_t
),
670 offsetof(dmu_buf_impl_t
, db_cache_link
),
671 dbuf_cache_multilist_index_func
);
672 refcount_create(&dbuf_cache_size
);
674 tsd_create(&zfs_dbuf_evict_key
, NULL
);
675 dbuf_evict_thread_exit
= B_FALSE
;
676 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
677 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
678 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
679 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
685 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
688 dbuf_stats_destroy();
690 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
691 mutex_destroy(&h
->hash_mutexes
[i
]);
692 #if defined(_KERNEL) && defined(HAVE_SPL)
694 * Large allocations which do not require contiguous pages
695 * should be using vmem_free() in the linux kernel
697 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
699 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
701 kmem_cache_destroy(dbuf_kmem_cache
);
702 taskq_destroy(dbu_evict_taskq
);
704 mutex_enter(&dbuf_evict_lock
);
705 dbuf_evict_thread_exit
= B_TRUE
;
706 while (dbuf_evict_thread_exit
) {
707 cv_signal(&dbuf_evict_cv
);
708 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
710 mutex_exit(&dbuf_evict_lock
);
711 tsd_destroy(&zfs_dbuf_evict_key
);
713 mutex_destroy(&dbuf_evict_lock
);
714 cv_destroy(&dbuf_evict_cv
);
716 refcount_destroy(&dbuf_cache_size
);
717 multilist_destroy(dbuf_cache
);
726 dbuf_verify(dmu_buf_impl_t
*db
)
729 dbuf_dirty_record_t
*dr
;
731 ASSERT(MUTEX_HELD(&db
->db_mtx
));
733 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
736 ASSERT(db
->db_objset
!= NULL
);
740 ASSERT(db
->db_parent
== NULL
);
741 ASSERT(db
->db_blkptr
== NULL
);
743 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
744 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
745 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
746 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
747 db
->db_blkid
== DMU_SPILL_BLKID
||
748 !avl_is_empty(&dn
->dn_dbufs
));
750 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
752 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
753 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
754 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
756 ASSERT0(db
->db
.db_offset
);
758 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
761 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
762 ASSERT(dr
->dr_dbuf
== db
);
764 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
765 ASSERT(dr
->dr_dbuf
== db
);
768 * We can't assert that db_size matches dn_datablksz because it
769 * can be momentarily different when another thread is doing
772 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
773 dr
= db
->db_data_pending
;
775 * It should only be modified in syncing context, so
776 * make sure we only have one copy of the data.
778 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
781 /* verify db->db_blkptr */
783 if (db
->db_parent
== dn
->dn_dbuf
) {
784 /* db is pointed to by the dnode */
785 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
786 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
787 ASSERT(db
->db_parent
== NULL
);
789 ASSERT(db
->db_parent
!= NULL
);
790 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
791 ASSERT3P(db
->db_blkptr
, ==,
792 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
794 /* db is pointed to by an indirect block */
795 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
797 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
798 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
801 * dnode_grow_indblksz() can make this fail if we don't
802 * have the struct_rwlock. XXX indblksz no longer
803 * grows. safe to do this now?
805 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
806 ASSERT3P(db
->db_blkptr
, ==,
807 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
808 db
->db_blkid
% epb
));
812 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
813 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
814 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
815 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
817 * If the blkptr isn't set but they have nonzero data,
818 * it had better be dirty, otherwise we'll lose that
819 * data when we evict this buffer.
821 * There is an exception to this rule for indirect blocks; in
822 * this case, if the indirect block is a hole, we fill in a few
823 * fields on each of the child blocks (importantly, birth time)
824 * to prevent hole birth times from being lost when you
825 * partially fill in a hole.
827 if (db
->db_dirtycnt
== 0) {
828 if (db
->db_level
== 0) {
829 uint64_t *buf
= db
->db
.db_data
;
832 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
837 blkptr_t
*bps
= db
->db
.db_data
;
838 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
841 * We want to verify that all the blkptrs in the
842 * indirect block are holes, but we may have
843 * automatically set up a few fields for them.
844 * We iterate through each blkptr and verify
845 * they only have those fields set.
848 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
850 blkptr_t
*bp
= &bps
[i
];
851 ASSERT(ZIO_CHECKSUM_IS_ZERO(
854 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
855 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
856 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
857 ASSERT0(bp
->blk_fill
);
858 ASSERT0(bp
->blk_pad
[0]);
859 ASSERT0(bp
->blk_pad
[1]);
860 ASSERT(!BP_IS_EMBEDDED(bp
));
861 ASSERT(BP_IS_HOLE(bp
));
862 ASSERT0(bp
->blk_phys_birth
);
872 dbuf_clear_data(dmu_buf_impl_t
*db
)
874 ASSERT(MUTEX_HELD(&db
->db_mtx
));
876 ASSERT3P(db
->db_buf
, ==, NULL
);
877 db
->db
.db_data
= NULL
;
878 if (db
->db_state
!= DB_NOFILL
)
879 db
->db_state
= DB_UNCACHED
;
883 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
885 ASSERT(MUTEX_HELD(&db
->db_mtx
));
889 ASSERT(buf
->b_data
!= NULL
);
890 db
->db
.db_data
= buf
->b_data
;
894 * Loan out an arc_buf for read. Return the loaned arc_buf.
897 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
901 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
902 mutex_enter(&db
->db_mtx
);
903 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
904 int blksz
= db
->db
.db_size
;
905 spa_t
*spa
= db
->db_objset
->os_spa
;
907 mutex_exit(&db
->db_mtx
);
908 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
909 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
912 arc_loan_inuse_buf(abuf
, db
);
915 mutex_exit(&db
->db_mtx
);
921 * Calculate which level n block references the data at the level 0 offset
925 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
927 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
929 * The level n blkid is equal to the level 0 blkid divided by
930 * the number of level 0s in a level n block.
932 * The level 0 blkid is offset >> datablkshift =
933 * offset / 2^datablkshift.
935 * The number of level 0s in a level n is the number of block
936 * pointers in an indirect block, raised to the power of level.
937 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
938 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
940 * Thus, the level n blkid is: offset /
941 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
942 * = offset / 2^(datablkshift + level *
943 * (indblkshift - SPA_BLKPTRSHIFT))
944 * = offset >> (datablkshift + level *
945 * (indblkshift - SPA_BLKPTRSHIFT))
948 const unsigned exp
= dn
->dn_datablkshift
+
949 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
951 if (exp
>= 8 * sizeof (offset
)) {
952 /* This only happens on the highest indirection level */
953 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
957 ASSERT3U(exp
, <, 8 * sizeof (offset
));
959 return (offset
>> exp
);
961 ASSERT3U(offset
, <, dn
->dn_datablksz
);
967 dbuf_read_done(zio_t
*zio
, int err
, arc_buf_t
*buf
, void *vdb
)
969 dmu_buf_impl_t
*db
= vdb
;
971 mutex_enter(&db
->db_mtx
);
972 ASSERT3U(db
->db_state
, ==, DB_READ
);
974 * All reads are synchronous, so we must have a hold on the dbuf
976 ASSERT(refcount_count(&db
->db_holds
) > 0);
977 ASSERT(db
->db_buf
== NULL
);
978 ASSERT(db
->db
.db_data
== NULL
);
979 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
980 /* we were freed in flight; disregard any error */
981 arc_release(buf
, db
);
982 bzero(buf
->b_data
, db
->db
.db_size
);
984 db
->db_freed_in_flight
= FALSE
;
985 dbuf_set_data(db
, buf
);
986 db
->db_state
= DB_CACHED
;
987 } else if (err
== 0) {
988 dbuf_set_data(db
, buf
);
989 db
->db_state
= DB_CACHED
;
991 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
992 ASSERT3P(db
->db_buf
, ==, NULL
);
993 arc_buf_destroy(buf
, db
);
994 db
->db_state
= DB_UNCACHED
;
996 cv_broadcast(&db
->db_changed
);
997 dbuf_rele_and_unlock(db
, NULL
);
1001 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1004 zbookmark_phys_t zb
;
1005 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1006 int err
, zio_flags
= 0;
1010 ASSERT(!refcount_is_zero(&db
->db_holds
));
1011 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1012 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1013 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1014 ASSERT(db
->db_state
== DB_UNCACHED
);
1015 ASSERT(db
->db_buf
== NULL
);
1017 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1019 * The bonus length stored in the dnode may be less than
1020 * the maximum available space in the bonus buffer.
1022 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1023 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1024 arc_buf_t
*dn_buf
= (dn
->dn_dbuf
!= NULL
) ?
1025 dn
->dn_dbuf
->db_buf
: NULL
;
1027 /* if the underlying dnode block is encrypted, decrypt it */
1028 if (dn_buf
!= NULL
&& dn
->dn_objset
->os_encrypted
&&
1029 DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
) &&
1030 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1031 arc_is_encrypted(dn_buf
)) {
1032 err
= arc_untransform(dn_buf
, dn
->dn_objset
->os_spa
,
1033 dmu_objset_id(dn
->dn_objset
), B_TRUE
);
1036 mutex_exit(&db
->db_mtx
);
1041 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1042 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1043 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1044 if (bonuslen
< max_bonuslen
)
1045 bzero(db
->db
.db_data
, max_bonuslen
);
1047 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1049 db
->db_state
= DB_CACHED
;
1050 mutex_exit(&db
->db_mtx
);
1055 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1056 * processes the delete record and clears the bp while we are waiting
1057 * for the dn_mtx (resulting in a "no" from block_freed).
1059 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1060 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1061 BP_IS_HOLE(db
->db_blkptr
)))) {
1062 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1064 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1066 bzero(db
->db
.db_data
, db
->db
.db_size
);
1068 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1069 BP_IS_HOLE(db
->db_blkptr
) &&
1070 db
->db_blkptr
->blk_birth
!= 0) {
1071 blkptr_t
*bps
= db
->db
.db_data
;
1073 for (i
= 0; i
< ((1 <<
1074 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1076 blkptr_t
*bp
= &bps
[i
];
1077 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1078 1 << dn
->dn_indblkshift
);
1080 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1082 BP_GET_LSIZE(db
->db_blkptr
));
1083 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1085 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1086 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1090 db
->db_state
= DB_CACHED
;
1091 mutex_exit(&db
->db_mtx
);
1097 db
->db_state
= DB_READ
;
1098 mutex_exit(&db
->db_mtx
);
1100 if (DBUF_IS_L2CACHEABLE(db
))
1101 aflags
|= ARC_FLAG_L2CACHE
;
1103 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
1104 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
1105 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1108 * All bps of an encrypted os should have the encryption bit set.
1109 * If this is not true it indicates tampering and we report an error.
1111 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(db
->db_blkptr
)) {
1112 spa_log_error(db
->db_objset
->os_spa
, &zb
);
1113 zfs_panic_recover("unencrypted block in encrypted "
1114 "object set %llu", dmu_objset_id(db
->db_objset
));
1115 return (SET_ERROR(EIO
));
1118 dbuf_add_ref(db
, NULL
);
1120 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1121 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1123 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1124 zio_flags
|= ZIO_FLAG_RAW
;
1126 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1127 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1134 * This is our just-in-time copy function. It makes a copy of buffers that
1135 * have been modified in a previous transaction group before we access them in
1136 * the current active group.
1138 * This function is used in three places: when we are dirtying a buffer for the
1139 * first time in a txg, when we are freeing a range in a dnode that includes
1140 * this buffer, and when we are accessing a buffer which was received compressed
1141 * and later referenced in a WRITE_BYREF record.
1143 * Note that when we are called from dbuf_free_range() we do not put a hold on
1144 * the buffer, we just traverse the active dbuf list for the dnode.
1147 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1149 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1151 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1152 ASSERT(db
->db
.db_data
!= NULL
);
1153 ASSERT(db
->db_level
== 0);
1154 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1157 (dr
->dt
.dl
.dr_data
!=
1158 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1162 * If the last dirty record for this dbuf has not yet synced
1163 * and its referencing the dbuf data, either:
1164 * reset the reference to point to a new copy,
1165 * or (if there a no active holders)
1166 * just null out the current db_data pointer.
1168 ASSERT(dr
->dr_txg
>= txg
- 2);
1169 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1170 dnode_t
*dn
= DB_DNODE(db
);
1171 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1172 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1173 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1174 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1175 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1176 dnode_t
*dn
= DB_DNODE(db
);
1177 int size
= arc_buf_size(db
->db_buf
);
1178 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1179 spa_t
*spa
= db
->db_objset
->os_spa
;
1180 enum zio_compress compress_type
=
1181 arc_get_compression(db
->db_buf
);
1183 if (arc_is_encrypted(db
->db_buf
)) {
1184 boolean_t byteorder
;
1185 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1186 uint8_t iv
[ZIO_DATA_IV_LEN
];
1187 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1189 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1191 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1192 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1193 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1195 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1196 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1197 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1198 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1200 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1202 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1205 dbuf_clear_data(db
);
1210 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1217 * We don't have to hold the mutex to check db_state because it
1218 * can't be freed while we have a hold on the buffer.
1220 ASSERT(!refcount_is_zero(&db
->db_holds
));
1222 if (db
->db_state
== DB_NOFILL
)
1223 return (SET_ERROR(EIO
));
1227 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1228 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1230 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1231 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1232 DBUF_IS_CACHEABLE(db
);
1234 mutex_enter(&db
->db_mtx
);
1235 if (db
->db_state
== DB_CACHED
) {
1236 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1239 * If the arc buf is compressed or encrypted, we need to
1240 * untransform it to read the data. This could happen during
1241 * the "zfs receive" of a stream which is deduplicated and
1242 * either raw or compressed. We do not need to do this if the
1243 * caller wants raw encrypted data.
1245 if (db
->db_buf
!= NULL
&& (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1246 (arc_is_encrypted(db
->db_buf
) ||
1247 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1248 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1249 err
= arc_untransform(db
->db_buf
, spa
,
1250 dmu_objset_id(db
->db_objset
), B_FALSE
);
1251 dbuf_set_data(db
, db
->db_buf
);
1253 mutex_exit(&db
->db_mtx
);
1255 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1256 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1257 rw_exit(&dn
->dn_struct_rwlock
);
1259 } else if (db
->db_state
== DB_UNCACHED
) {
1260 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1261 boolean_t need_wait
= B_FALSE
;
1264 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1265 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1268 err
= dbuf_read_impl(db
, zio
, flags
);
1270 /* dbuf_read_impl has dropped db_mtx for us */
1272 if (!err
&& prefetch
)
1273 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1275 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1276 rw_exit(&dn
->dn_struct_rwlock
);
1279 if (!err
&& need_wait
)
1280 err
= zio_wait(zio
);
1283 * Another reader came in while the dbuf was in flight
1284 * between UNCACHED and CACHED. Either a writer will finish
1285 * writing the buffer (sending the dbuf to CACHED) or the
1286 * first reader's request will reach the read_done callback
1287 * and send the dbuf to CACHED. Otherwise, a failure
1288 * occurred and the dbuf went to UNCACHED.
1290 mutex_exit(&db
->db_mtx
);
1292 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1293 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1294 rw_exit(&dn
->dn_struct_rwlock
);
1297 /* Skip the wait per the caller's request. */
1298 mutex_enter(&db
->db_mtx
);
1299 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1300 while (db
->db_state
== DB_READ
||
1301 db
->db_state
== DB_FILL
) {
1302 ASSERT(db
->db_state
== DB_READ
||
1303 (flags
& DB_RF_HAVESTRUCT
) == 0);
1304 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1306 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1308 if (db
->db_state
== DB_UNCACHED
)
1309 err
= SET_ERROR(EIO
);
1311 mutex_exit(&db
->db_mtx
);
1318 dbuf_noread(dmu_buf_impl_t
*db
)
1320 ASSERT(!refcount_is_zero(&db
->db_holds
));
1321 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1322 mutex_enter(&db
->db_mtx
);
1323 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1324 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1325 if (db
->db_state
== DB_UNCACHED
) {
1326 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1327 spa_t
*spa
= db
->db_objset
->os_spa
;
1329 ASSERT(db
->db_buf
== NULL
);
1330 ASSERT(db
->db
.db_data
== NULL
);
1331 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1332 db
->db_state
= DB_FILL
;
1333 } else if (db
->db_state
== DB_NOFILL
) {
1334 dbuf_clear_data(db
);
1336 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1338 mutex_exit(&db
->db_mtx
);
1342 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1344 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1345 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1346 uint64_t txg
= dr
->dr_txg
;
1348 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1350 * This assert is valid because dmu_sync() expects to be called by
1351 * a zilog's get_data while holding a range lock. This call only
1352 * comes from dbuf_dirty() callers who must also hold a range lock.
1354 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1355 ASSERT(db
->db_level
== 0);
1357 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1358 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1361 ASSERT(db
->db_data_pending
!= dr
);
1363 /* free this block */
1364 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1365 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1367 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1368 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1369 dr
->dt
.dl
.dr_raw
= B_FALSE
;
1372 * Release the already-written buffer, so we leave it in
1373 * a consistent dirty state. Note that all callers are
1374 * modifying the buffer, so they will immediately do
1375 * another (redundant) arc_release(). Therefore, leave
1376 * the buf thawed to save the effort of freezing &
1377 * immediately re-thawing it.
1379 arc_release(dr
->dt
.dl
.dr_data
, db
);
1383 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1384 * data blocks in the free range, so that any future readers will find
1388 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1391 dmu_buf_impl_t
*db_search
;
1392 dmu_buf_impl_t
*db
, *db_next
;
1393 uint64_t txg
= tx
->tx_txg
;
1396 if (end_blkid
> dn
->dn_maxblkid
&&
1397 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1398 end_blkid
= dn
->dn_maxblkid
;
1399 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1401 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1402 db_search
->db_level
= 0;
1403 db_search
->db_blkid
= start_blkid
;
1404 db_search
->db_state
= DB_SEARCH
;
1406 mutex_enter(&dn
->dn_dbufs_mtx
);
1407 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1408 ASSERT3P(db
, ==, NULL
);
1410 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1412 for (; db
!= NULL
; db
= db_next
) {
1413 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1414 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1416 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1419 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1421 /* found a level 0 buffer in the range */
1422 mutex_enter(&db
->db_mtx
);
1423 if (dbuf_undirty(db
, tx
)) {
1424 /* mutex has been dropped and dbuf destroyed */
1428 if (db
->db_state
== DB_UNCACHED
||
1429 db
->db_state
== DB_NOFILL
||
1430 db
->db_state
== DB_EVICTING
) {
1431 ASSERT(db
->db
.db_data
== NULL
);
1432 mutex_exit(&db
->db_mtx
);
1435 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1436 /* will be handled in dbuf_read_done or dbuf_rele */
1437 db
->db_freed_in_flight
= TRUE
;
1438 mutex_exit(&db
->db_mtx
);
1441 if (refcount_count(&db
->db_holds
) == 0) {
1446 /* The dbuf is referenced */
1448 if (db
->db_last_dirty
!= NULL
) {
1449 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1451 if (dr
->dr_txg
== txg
) {
1453 * This buffer is "in-use", re-adjust the file
1454 * size to reflect that this buffer may
1455 * contain new data when we sync.
1457 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1458 db
->db_blkid
> dn
->dn_maxblkid
)
1459 dn
->dn_maxblkid
= db
->db_blkid
;
1460 dbuf_unoverride(dr
);
1463 * This dbuf is not dirty in the open context.
1464 * Either uncache it (if its not referenced in
1465 * the open context) or reset its contents to
1468 dbuf_fix_old_data(db
, txg
);
1471 /* clear the contents if its cached */
1472 if (db
->db_state
== DB_CACHED
) {
1473 ASSERT(db
->db
.db_data
!= NULL
);
1474 arc_release(db
->db_buf
, db
);
1475 bzero(db
->db
.db_data
, db
->db
.db_size
);
1476 arc_buf_freeze(db
->db_buf
);
1479 mutex_exit(&db
->db_mtx
);
1482 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1483 mutex_exit(&dn
->dn_dbufs_mtx
);
1487 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1489 arc_buf_t
*buf
, *obuf
;
1490 int osize
= db
->db
.db_size
;
1491 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1494 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1499 /* XXX does *this* func really need the lock? */
1500 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1503 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1504 * is OK, because there can be no other references to the db
1505 * when we are changing its size, so no concurrent DB_FILL can
1509 * XXX we should be doing a dbuf_read, checking the return
1510 * value and returning that up to our callers
1512 dmu_buf_will_dirty(&db
->db
, tx
);
1514 /* create the data buffer for the new block */
1515 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1517 /* copy old block data to the new block */
1519 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1520 /* zero the remainder */
1522 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1524 mutex_enter(&db
->db_mtx
);
1525 dbuf_set_data(db
, buf
);
1526 arc_buf_destroy(obuf
, db
);
1527 db
->db
.db_size
= size
;
1529 if (db
->db_level
== 0) {
1530 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1531 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1533 mutex_exit(&db
->db_mtx
);
1535 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1540 dbuf_release_bp(dmu_buf_impl_t
*db
)
1542 ASSERTV(objset_t
*os
= db
->db_objset
);
1544 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1545 ASSERT(arc_released(os
->os_phys_buf
) ||
1546 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1547 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1549 (void) arc_release(db
->db_buf
, db
);
1553 * We already have a dirty record for this TXG, and we are being
1557 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1559 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1561 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1563 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1565 * If this buffer has already been written out,
1566 * we now need to reset its state.
1568 dbuf_unoverride(dr
);
1569 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1570 db
->db_state
!= DB_NOFILL
) {
1571 /* Already released on initial dirty, so just thaw. */
1572 ASSERT(arc_released(db
->db_buf
));
1573 arc_buf_thaw(db
->db_buf
);
1578 dbuf_dirty_record_t
*
1579 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1583 dbuf_dirty_record_t
**drp
, *dr
;
1584 int drop_struct_lock
= FALSE
;
1585 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1587 ASSERT(tx
->tx_txg
!= 0);
1588 ASSERT(!refcount_is_zero(&db
->db_holds
));
1589 DMU_TX_DIRTY_BUF(tx
, db
);
1594 * Shouldn't dirty a regular buffer in syncing context. Private
1595 * objects may be dirtied in syncing context, but only if they
1596 * were already pre-dirtied in open context.
1599 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1600 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1603 ASSERT(!dmu_tx_is_syncing(tx
) ||
1604 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1605 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1606 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1607 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1608 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1611 * We make this assert for private objects as well, but after we
1612 * check if we're already dirty. They are allowed to re-dirty
1613 * in syncing context.
1615 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1616 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1617 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1619 mutex_enter(&db
->db_mtx
);
1621 * XXX make this true for indirects too? The problem is that
1622 * transactions created with dmu_tx_create_assigned() from
1623 * syncing context don't bother holding ahead.
1625 ASSERT(db
->db_level
!= 0 ||
1626 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1627 db
->db_state
== DB_NOFILL
);
1629 mutex_enter(&dn
->dn_mtx
);
1631 * Don't set dirtyctx to SYNC if we're just modifying this as we
1632 * initialize the objset.
1634 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1635 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1636 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1639 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1640 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1641 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1642 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1643 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1645 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1646 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1650 mutex_exit(&dn
->dn_mtx
);
1652 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1653 dn
->dn_have_spill
= B_TRUE
;
1656 * If this buffer is already dirty, we're done.
1658 drp
= &db
->db_last_dirty
;
1659 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1660 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1661 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1663 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1667 mutex_exit(&db
->db_mtx
);
1672 * Only valid if not already dirty.
1674 ASSERT(dn
->dn_object
== 0 ||
1675 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1676 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1678 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1681 * We should only be dirtying in syncing context if it's the
1682 * mos or we're initializing the os or it's a special object.
1683 * However, we are allowed to dirty in syncing context provided
1684 * we already dirtied it in open context. Hence we must make
1685 * this assertion only if we're not already dirty.
1688 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
1690 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1691 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
1692 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1693 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1694 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1695 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1697 ASSERT(db
->db
.db_size
!= 0);
1699 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1701 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1702 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
1706 * If this buffer is dirty in an old transaction group we need
1707 * to make a copy of it so that the changes we make in this
1708 * transaction group won't leak out when we sync the older txg.
1710 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1711 list_link_init(&dr
->dr_dirty_node
);
1712 if (db
->db_level
== 0) {
1713 void *data_old
= db
->db_buf
;
1715 if (db
->db_state
!= DB_NOFILL
) {
1716 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1717 dbuf_fix_old_data(db
, tx
->tx_txg
);
1718 data_old
= db
->db
.db_data
;
1719 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1721 * Release the data buffer from the cache so
1722 * that we can modify it without impacting
1723 * possible other users of this cached data
1724 * block. Note that indirect blocks and
1725 * private objects are not released until the
1726 * syncing state (since they are only modified
1729 arc_release(db
->db_buf
, db
);
1730 dbuf_fix_old_data(db
, tx
->tx_txg
);
1731 data_old
= db
->db_buf
;
1733 ASSERT(data_old
!= NULL
);
1735 dr
->dt
.dl
.dr_data
= data_old
;
1737 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1738 list_create(&dr
->dt
.di
.dr_children
,
1739 sizeof (dbuf_dirty_record_t
),
1740 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1742 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1743 dr
->dr_accounted
= db
->db
.db_size
;
1745 dr
->dr_txg
= tx
->tx_txg
;
1750 * We could have been freed_in_flight between the dbuf_noread
1751 * and dbuf_dirty. We win, as though the dbuf_noread() had
1752 * happened after the free.
1754 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1755 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1756 mutex_enter(&dn
->dn_mtx
);
1757 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1758 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1761 mutex_exit(&dn
->dn_mtx
);
1762 db
->db_freed_in_flight
= FALSE
;
1766 * This buffer is now part of this txg
1768 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1769 db
->db_dirtycnt
+= 1;
1770 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1772 mutex_exit(&db
->db_mtx
);
1774 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1775 db
->db_blkid
== DMU_SPILL_BLKID
) {
1776 mutex_enter(&dn
->dn_mtx
);
1777 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1778 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1779 mutex_exit(&dn
->dn_mtx
);
1780 dnode_setdirty(dn
, tx
);
1786 * The dn_struct_rwlock prevents db_blkptr from changing
1787 * due to a write from syncing context completing
1788 * while we are running, so we want to acquire it before
1789 * looking at db_blkptr.
1791 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1792 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1793 drop_struct_lock
= TRUE
;
1797 * We need to hold the dn_struct_rwlock to make this assertion,
1798 * because it protects dn_phys / dn_next_nlevels from changing.
1800 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1801 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1802 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1803 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1804 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1807 * If we are overwriting a dedup BP, then unless it is snapshotted,
1808 * when we get to syncing context we will need to decrement its
1809 * refcount in the DDT. Prefetch the relevant DDT block so that
1810 * syncing context won't have to wait for the i/o.
1812 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
1814 if (db
->db_level
== 0) {
1815 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1816 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1819 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1820 dmu_buf_impl_t
*parent
= db
->db_parent
;
1821 dbuf_dirty_record_t
*di
;
1822 int parent_held
= FALSE
;
1824 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1825 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1827 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1828 db
->db_blkid
>> epbs
, FTAG
);
1829 ASSERT(parent
!= NULL
);
1832 if (drop_struct_lock
)
1833 rw_exit(&dn
->dn_struct_rwlock
);
1834 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1835 di
= dbuf_dirty(parent
, tx
);
1837 dbuf_rele(parent
, FTAG
);
1839 mutex_enter(&db
->db_mtx
);
1841 * Since we've dropped the mutex, it's possible that
1842 * dbuf_undirty() might have changed this out from under us.
1844 if (db
->db_last_dirty
== dr
||
1845 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1846 mutex_enter(&di
->dt
.di
.dr_mtx
);
1847 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1848 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1849 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1850 mutex_exit(&di
->dt
.di
.dr_mtx
);
1853 mutex_exit(&db
->db_mtx
);
1855 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1856 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1857 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1858 mutex_enter(&dn
->dn_mtx
);
1859 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1860 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1861 mutex_exit(&dn
->dn_mtx
);
1862 if (drop_struct_lock
)
1863 rw_exit(&dn
->dn_struct_rwlock
);
1866 dnode_setdirty(dn
, tx
);
1872 * Undirty a buffer in the transaction group referenced by the given
1873 * transaction. Return whether this evicted the dbuf.
1876 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1879 uint64_t txg
= tx
->tx_txg
;
1880 dbuf_dirty_record_t
*dr
, **drp
;
1885 * Due to our use of dn_nlevels below, this can only be called
1886 * in open context, unless we are operating on the MOS.
1887 * From syncing context, dn_nlevels may be different from the
1888 * dn_nlevels used when dbuf was dirtied.
1890 ASSERT(db
->db_objset
==
1891 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1892 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1893 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1894 ASSERT0(db
->db_level
);
1895 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1898 * If this buffer is not dirty, we're done.
1900 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1901 if (dr
->dr_txg
<= txg
)
1903 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1905 ASSERT(dr
->dr_txg
== txg
);
1906 ASSERT(dr
->dr_dbuf
== db
);
1911 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1913 ASSERT(db
->db
.db_size
!= 0);
1915 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1916 dr
->dr_accounted
, txg
);
1921 * Note that there are three places in dbuf_dirty()
1922 * where this dirty record may be put on a list.
1923 * Make sure to do a list_remove corresponding to
1924 * every one of those list_insert calls.
1926 if (dr
->dr_parent
) {
1927 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1928 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1929 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1930 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1931 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1932 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1933 mutex_enter(&dn
->dn_mtx
);
1934 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1935 mutex_exit(&dn
->dn_mtx
);
1939 if (db
->db_state
!= DB_NOFILL
) {
1940 dbuf_unoverride(dr
);
1942 ASSERT(db
->db_buf
!= NULL
);
1943 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1944 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1945 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
1948 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1950 ASSERT(db
->db_dirtycnt
> 0);
1951 db
->db_dirtycnt
-= 1;
1953 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1954 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
1963 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
1965 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1966 dbuf_dirty_record_t
*dr
;
1968 ASSERT(tx
->tx_txg
!= 0);
1969 ASSERT(!refcount_is_zero(&db
->db_holds
));
1972 * Quick check for dirtyness. For already dirty blocks, this
1973 * reduces runtime of this function by >90%, and overall performance
1974 * by 50% for some workloads (e.g. file deletion with indirect blocks
1977 mutex_enter(&db
->db_mtx
);
1979 for (dr
= db
->db_last_dirty
;
1980 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1982 * It's possible that it is already dirty but not cached,
1983 * because there are some calls to dbuf_dirty() that don't
1984 * go through dmu_buf_will_dirty().
1986 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1987 /* This dbuf is already dirty and cached. */
1989 mutex_exit(&db
->db_mtx
);
1993 mutex_exit(&db
->db_mtx
);
1996 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1997 flags
|= DB_RF_HAVESTRUCT
;
1999 (void) dbuf_read(db
, NULL
, flags
);
2000 (void) dbuf_dirty(db
, tx
);
2004 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2006 dmu_buf_will_dirty_impl(db_fake
,
2007 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2011 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2013 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2015 db
->db_state
= DB_NOFILL
;
2017 dmu_buf_will_fill(db_fake
, tx
);
2021 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2023 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2025 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2026 ASSERT(tx
->tx_txg
!= 0);
2027 ASSERT(db
->db_level
== 0);
2028 ASSERT(!refcount_is_zero(&db
->db_holds
));
2030 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2031 dmu_tx_private_ok(tx
));
2034 (void) dbuf_dirty(db
, tx
);
2038 * This function is effectively the same as dmu_buf_will_dirty(), but
2039 * indicates the caller expects raw encrypted data in the db. It will
2040 * also set the raw flag on the created dirty record.
2043 dmu_buf_will_change_crypt_params(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2045 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2046 dbuf_dirty_record_t
*dr
;
2048 dmu_buf_will_dirty_impl(db_fake
,
2049 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2051 dr
= db
->db_last_dirty
;
2052 while (dr
!= NULL
&& dr
->dr_txg
> tx
->tx_txg
)
2055 ASSERT3P(dr
, !=, NULL
);
2056 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2057 dr
->dt
.dl
.dr_raw
= B_TRUE
;
2060 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2063 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2065 mutex_enter(&db
->db_mtx
);
2068 if (db
->db_state
== DB_FILL
) {
2069 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2070 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2071 /* we were freed while filling */
2072 /* XXX dbuf_undirty? */
2073 bzero(db
->db
.db_data
, db
->db
.db_size
);
2074 db
->db_freed_in_flight
= FALSE
;
2076 db
->db_state
= DB_CACHED
;
2077 cv_broadcast(&db
->db_changed
);
2079 mutex_exit(&db
->db_mtx
);
2083 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2084 bp_embedded_type_t etype
, enum zio_compress comp
,
2085 int uncompressed_size
, int compressed_size
, int byteorder
,
2088 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2089 struct dirty_leaf
*dl
;
2090 dmu_object_type_t type
;
2092 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2093 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2094 SPA_FEATURE_EMBEDDED_DATA
));
2098 type
= DB_DNODE(db
)->dn_type
;
2101 ASSERT0(db
->db_level
);
2102 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2104 dmu_buf_will_not_fill(dbuf
, tx
);
2106 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2107 dl
= &db
->db_last_dirty
->dt
.dl
;
2108 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2109 data
, comp
, uncompressed_size
, compressed_size
);
2110 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2111 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2112 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2113 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2115 dl
->dr_override_state
= DR_OVERRIDDEN
;
2116 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2120 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2121 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2124 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2126 ASSERT(!refcount_is_zero(&db
->db_holds
));
2127 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2128 ASSERT(db
->db_level
== 0);
2129 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2130 ASSERT(buf
!= NULL
);
2131 ASSERT(arc_buf_lsize(buf
) == db
->db
.db_size
);
2132 ASSERT(tx
->tx_txg
!= 0);
2134 arc_return_buf(buf
, db
);
2135 ASSERT(arc_released(buf
));
2137 mutex_enter(&db
->db_mtx
);
2139 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2140 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2142 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2144 if (db
->db_state
== DB_CACHED
&&
2145 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2146 mutex_exit(&db
->db_mtx
);
2147 (void) dbuf_dirty(db
, tx
);
2148 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2149 arc_buf_destroy(buf
, db
);
2150 xuio_stat_wbuf_copied();
2154 xuio_stat_wbuf_nocopy();
2155 if (db
->db_state
== DB_CACHED
) {
2156 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2158 ASSERT(db
->db_buf
!= NULL
);
2159 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2160 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2161 if (!arc_released(db
->db_buf
)) {
2162 ASSERT(dr
->dt
.dl
.dr_override_state
==
2164 arc_release(db
->db_buf
, db
);
2166 dr
->dt
.dl
.dr_data
= buf
;
2167 arc_buf_destroy(db
->db_buf
, db
);
2168 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2169 arc_release(db
->db_buf
, db
);
2170 arc_buf_destroy(db
->db_buf
, db
);
2174 ASSERT(db
->db_buf
== NULL
);
2175 dbuf_set_data(db
, buf
);
2176 db
->db_state
= DB_FILL
;
2177 mutex_exit(&db
->db_mtx
);
2178 (void) dbuf_dirty(db
, tx
);
2179 dmu_buf_fill_done(&db
->db
, tx
);
2183 dbuf_destroy(dmu_buf_impl_t
*db
)
2186 dmu_buf_impl_t
*parent
= db
->db_parent
;
2187 dmu_buf_impl_t
*dndb
;
2189 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2190 ASSERT(refcount_is_zero(&db
->db_holds
));
2192 if (db
->db_buf
!= NULL
) {
2193 arc_buf_destroy(db
->db_buf
, db
);
2197 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2198 int slots
= DB_DNODE(db
)->dn_num_slots
;
2199 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2200 if (db
->db
.db_data
!= NULL
) {
2201 kmem_free(db
->db
.db_data
, bonuslen
);
2202 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2203 db
->db_state
= DB_UNCACHED
;
2207 dbuf_clear_data(db
);
2209 if (multilist_link_active(&db
->db_cache_link
)) {
2210 multilist_remove(dbuf_cache
, db
);
2211 (void) refcount_remove_many(&dbuf_cache_size
,
2212 db
->db
.db_size
, db
);
2215 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2216 ASSERT(db
->db_data_pending
== NULL
);
2218 db
->db_state
= DB_EVICTING
;
2219 db
->db_blkptr
= NULL
;
2222 * Now that db_state is DB_EVICTING, nobody else can find this via
2223 * the hash table. We can now drop db_mtx, which allows us to
2224 * acquire the dn_dbufs_mtx.
2226 mutex_exit(&db
->db_mtx
);
2231 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2232 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2234 mutex_enter(&dn
->dn_dbufs_mtx
);
2235 avl_remove(&dn
->dn_dbufs
, db
);
2236 atomic_dec_32(&dn
->dn_dbufs_count
);
2240 mutex_exit(&dn
->dn_dbufs_mtx
);
2242 * Decrementing the dbuf count means that the hold corresponding
2243 * to the removed dbuf is no longer discounted in dnode_move(),
2244 * so the dnode cannot be moved until after we release the hold.
2245 * The membar_producer() ensures visibility of the decremented
2246 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2250 db
->db_dnode_handle
= NULL
;
2252 dbuf_hash_remove(db
);
2257 ASSERT(refcount_is_zero(&db
->db_holds
));
2259 db
->db_parent
= NULL
;
2261 ASSERT(db
->db_buf
== NULL
);
2262 ASSERT(db
->db
.db_data
== NULL
);
2263 ASSERT(db
->db_hash_next
== NULL
);
2264 ASSERT(db
->db_blkptr
== NULL
);
2265 ASSERT(db
->db_data_pending
== NULL
);
2266 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2268 kmem_cache_free(dbuf_kmem_cache
, db
);
2269 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2272 * If this dbuf is referenced from an indirect dbuf,
2273 * decrement the ref count on the indirect dbuf.
2275 if (parent
&& parent
!= dndb
)
2276 dbuf_rele(parent
, db
);
2280 * Note: While bpp will always be updated if the function returns success,
2281 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2282 * this happens when the dnode is the meta-dnode, or a userused or groupused
2285 __attribute__((always_inline
))
2287 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2288 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
2295 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2297 if (blkid
== DMU_SPILL_BLKID
) {
2298 mutex_enter(&dn
->dn_mtx
);
2299 if (dn
->dn_have_spill
&&
2300 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2301 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2304 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2305 *parentp
= dn
->dn_dbuf
;
2306 mutex_exit(&dn
->dn_mtx
);
2311 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2312 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2314 ASSERT3U(level
* epbs
, <, 64);
2315 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2317 * This assertion shouldn't trip as long as the max indirect block size
2318 * is less than 1M. The reason for this is that up to that point,
2319 * the number of levels required to address an entire object with blocks
2320 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2321 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2322 * (i.e. we can address the entire object), objects will all use at most
2323 * N-1 levels and the assertion won't overflow. However, once epbs is
2324 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2325 * enough to address an entire object, so objects will have 5 levels,
2326 * but then this assertion will overflow.
2328 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2329 * need to redo this logic to handle overflows.
2331 ASSERT(level
>= nlevels
||
2332 ((nlevels
- level
- 1) * epbs
) +
2333 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2334 if (level
>= nlevels
||
2335 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2336 ((nlevels
- level
- 1) * epbs
)) ||
2338 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2339 /* the buffer has no parent yet */
2340 return (SET_ERROR(ENOENT
));
2341 } else if (level
< nlevels
-1) {
2342 /* this block is referenced from an indirect block */
2345 err
= dbuf_hold_impl(dn
, level
+1,
2346 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2348 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
2349 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
2350 parentp
, dh
->dh_depth
+ 1);
2351 err
= __dbuf_hold_impl(dh
+ 1);
2355 err
= dbuf_read(*parentp
, NULL
,
2356 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2358 dbuf_rele(*parentp
, NULL
);
2362 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2363 (blkid
& ((1ULL << epbs
) - 1));
2364 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2365 ASSERT(BP_IS_HOLE(*bpp
));
2368 /* the block is referenced from the dnode */
2369 ASSERT3U(level
, ==, nlevels
-1);
2370 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2371 blkid
< dn
->dn_phys
->dn_nblkptr
);
2373 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2374 *parentp
= dn
->dn_dbuf
;
2376 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2381 static dmu_buf_impl_t
*
2382 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2383 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2385 objset_t
*os
= dn
->dn_objset
;
2386 dmu_buf_impl_t
*db
, *odb
;
2388 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2389 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2391 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2394 db
->db
.db_object
= dn
->dn_object
;
2395 db
->db_level
= level
;
2396 db
->db_blkid
= blkid
;
2397 db
->db_last_dirty
= NULL
;
2398 db
->db_dirtycnt
= 0;
2399 db
->db_dnode_handle
= dn
->dn_handle
;
2400 db
->db_parent
= parent
;
2401 db
->db_blkptr
= blkptr
;
2404 db
->db_user_immediate_evict
= FALSE
;
2405 db
->db_freed_in_flight
= FALSE
;
2406 db
->db_pending_evict
= FALSE
;
2408 if (blkid
== DMU_BONUS_BLKID
) {
2409 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2410 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2411 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2412 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2413 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2414 db
->db_state
= DB_UNCACHED
;
2415 /* the bonus dbuf is not placed in the hash table */
2416 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2418 } else if (blkid
== DMU_SPILL_BLKID
) {
2419 db
->db
.db_size
= (blkptr
!= NULL
) ?
2420 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2421 db
->db
.db_offset
= 0;
2424 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2425 db
->db
.db_size
= blocksize
;
2426 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2430 * Hold the dn_dbufs_mtx while we get the new dbuf
2431 * in the hash table *and* added to the dbufs list.
2432 * This prevents a possible deadlock with someone
2433 * trying to look up this dbuf before its added to the
2436 mutex_enter(&dn
->dn_dbufs_mtx
);
2437 db
->db_state
= DB_EVICTING
;
2438 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2439 /* someone else inserted it first */
2440 kmem_cache_free(dbuf_kmem_cache
, db
);
2441 mutex_exit(&dn
->dn_dbufs_mtx
);
2444 avl_add(&dn
->dn_dbufs
, db
);
2446 db
->db_state
= DB_UNCACHED
;
2447 mutex_exit(&dn
->dn_dbufs_mtx
);
2448 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2450 if (parent
&& parent
!= dn
->dn_dbuf
)
2451 dbuf_add_ref(parent
, db
);
2453 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2454 refcount_count(&dn
->dn_holds
) > 0);
2455 (void) refcount_add(&dn
->dn_holds
, db
);
2456 atomic_inc_32(&dn
->dn_dbufs_count
);
2458 dprintf_dbuf(db
, "db=%p\n", db
);
2463 typedef struct dbuf_prefetch_arg
{
2464 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2465 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2466 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2467 int dpa_curlevel
; /* The current level that we're reading */
2468 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2469 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2470 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2471 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2472 } dbuf_prefetch_arg_t
;
2475 * Actually issue the prefetch read for the block given.
2478 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2481 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2484 aflags
= dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2486 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2487 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2488 ASSERT(dpa
->dpa_zio
!= NULL
);
2489 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2490 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2491 &aflags
, &dpa
->dpa_zb
);
2495 * Called when an indirect block above our prefetch target is read in. This
2496 * will either read in the next indirect block down the tree or issue the actual
2497 * prefetch if the next block down is our target.
2500 dbuf_prefetch_indirect_done(zio_t
*zio
, int err
, arc_buf_t
*abuf
, void *private)
2502 dbuf_prefetch_arg_t
*dpa
= private;
2506 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2507 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2510 * The dpa_dnode is only valid if we are called with a NULL
2511 * zio. This indicates that the arc_read() returned without
2512 * first calling zio_read() to issue a physical read. Once
2513 * a physical read is made the dpa_dnode must be invalidated
2514 * as the locks guarding it may have been dropped. If the
2515 * dpa_dnode is still valid, then we want to add it to the dbuf
2516 * cache. To do so, we must hold the dbuf associated with the block
2517 * we just prefetched, read its contents so that we associate it
2518 * with an arc_buf_t, and then release it.
2521 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2522 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
2523 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2525 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2527 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2529 dpa
->dpa_dnode
= NULL
;
2530 } else if (dpa
->dpa_dnode
!= NULL
) {
2531 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2532 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2533 dpa
->dpa_zb
.zb_level
));
2534 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2535 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2536 (void) dbuf_read(db
, NULL
,
2537 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2538 dbuf_rele(db
, FTAG
);
2541 dpa
->dpa_curlevel
--;
2543 nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2544 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2545 bp
= ((blkptr_t
*)abuf
->b_data
) +
2546 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2547 if (BP_IS_HOLE(bp
) || err
!= 0) {
2548 kmem_free(dpa
, sizeof (*dpa
));
2549 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2550 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2551 dbuf_issue_final_prefetch(dpa
, bp
);
2552 kmem_free(dpa
, sizeof (*dpa
));
2554 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2555 zbookmark_phys_t zb
;
2557 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2559 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2560 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2562 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2563 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2564 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2568 arc_buf_destroy(abuf
, private);
2572 * Issue prefetch reads for the given block on the given level. If the indirect
2573 * blocks above that block are not in memory, we will read them in
2574 * asynchronously. As a result, this call never blocks waiting for a read to
2575 * complete. Note that the prefetch might fail if the dataset is encrypted and
2576 * the encryption key is unmapped before the IO completes.
2579 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2583 int epbs
, nlevels
, curlevel
;
2587 dbuf_prefetch_arg_t
*dpa
;
2590 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2591 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2593 if (blkid
> dn
->dn_maxblkid
)
2596 if (dnode_block_freed(dn
, blkid
))
2600 * This dnode hasn't been written to disk yet, so there's nothing to
2603 nlevels
= dn
->dn_phys
->dn_nlevels
;
2604 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2607 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2608 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2611 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2614 mutex_exit(&db
->db_mtx
);
2616 * This dbuf already exists. It is either CACHED, or
2617 * (we assume) about to be read or filled.
2623 * Find the closest ancestor (indirect block) of the target block
2624 * that is present in the cache. In this indirect block, we will
2625 * find the bp that is at curlevel, curblkid.
2629 while (curlevel
< nlevels
- 1) {
2630 int parent_level
= curlevel
+ 1;
2631 uint64_t parent_blkid
= curblkid
>> epbs
;
2634 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2635 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2636 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2637 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2638 dbuf_rele(db
, FTAG
);
2642 curlevel
= parent_level
;
2643 curblkid
= parent_blkid
;
2646 if (curlevel
== nlevels
- 1) {
2647 /* No cached indirect blocks found. */
2648 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2649 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2651 if (BP_IS_HOLE(&bp
))
2654 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2656 pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2659 dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2660 ds
= dn
->dn_objset
->os_dsl_dataset
;
2661 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2662 dn
->dn_object
, level
, blkid
);
2663 dpa
->dpa_curlevel
= curlevel
;
2664 dpa
->dpa_prio
= prio
;
2665 dpa
->dpa_aflags
= aflags
;
2666 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2667 dpa
->dpa_dnode
= dn
;
2668 dpa
->dpa_epbs
= epbs
;
2672 * If we have the indirect just above us, no need to do the asynchronous
2673 * prefetch chain; we'll just run the last step ourselves. If we're at
2674 * a higher level, though, we want to issue the prefetches for all the
2675 * indirect blocks asynchronously, so we can go on with whatever we were
2678 if (curlevel
== level
) {
2679 ASSERT3U(curblkid
, ==, blkid
);
2680 dbuf_issue_final_prefetch(dpa
, &bp
);
2681 kmem_free(dpa
, sizeof (*dpa
));
2683 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2684 zbookmark_phys_t zb
;
2686 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2687 dn
->dn_object
, curlevel
, curblkid
);
2688 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2689 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2690 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2694 * We use pio here instead of dpa_zio since it's possible that
2695 * dpa may have already been freed.
2700 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2703 * Returns with db_holds incremented, and db_mtx not held.
2704 * Note: dn_struct_rwlock must be held.
2707 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2709 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2710 dh
->dh_parent
= NULL
;
2712 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2713 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2714 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2716 *(dh
->dh_dbp
) = NULL
;
2718 /* dbuf_find() returns with db_mtx held */
2719 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2720 dh
->dh_level
, dh
->dh_blkid
);
2722 if (dh
->dh_db
== NULL
) {
2725 if (dh
->dh_fail_uncached
)
2726 return (SET_ERROR(ENOENT
));
2728 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2729 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2730 dh
->dh_fail_sparse
, &dh
->dh_parent
, &dh
->dh_bp
, dh
);
2731 if (dh
->dh_fail_sparse
) {
2732 if (dh
->dh_err
== 0 &&
2733 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2734 dh
->dh_err
= SET_ERROR(ENOENT
);
2737 dbuf_rele(dh
->dh_parent
, NULL
);
2738 return (dh
->dh_err
);
2741 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2742 return (dh
->dh_err
);
2743 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2744 dh
->dh_parent
, dh
->dh_bp
);
2747 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2748 mutex_exit(&dh
->dh_db
->db_mtx
);
2749 return (SET_ERROR(ENOENT
));
2752 if (dh
->dh_db
->db_buf
!= NULL
)
2753 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2755 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2758 * If this buffer is currently syncing out, and we are are
2759 * still referencing it from db_data, we need to make a copy
2760 * of it in case we decide we want to dirty it again in this txg.
2762 if (dh
->dh_db
->db_level
== 0 &&
2763 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2764 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2765 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2766 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2768 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2769 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2771 dbuf_set_data(dh
->dh_db
,
2772 arc_alloc_buf(dh
->dh_dn
->dn_objset
->os_spa
,
2773 dh
->dh_db
, dh
->dh_type
, dh
->dh_db
->db
.db_size
));
2774 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2775 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2779 if (multilist_link_active(&dh
->dh_db
->db_cache_link
)) {
2780 ASSERT(refcount_is_zero(&dh
->dh_db
->db_holds
));
2781 multilist_remove(dbuf_cache
, dh
->dh_db
);
2782 (void) refcount_remove_many(&dbuf_cache_size
,
2783 dh
->dh_db
->db
.db_size
, dh
->dh_db
);
2785 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2786 DBUF_VERIFY(dh
->dh_db
);
2787 mutex_exit(&dh
->dh_db
->db_mtx
);
2789 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2791 dbuf_rele(dh
->dh_parent
, NULL
);
2793 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2794 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2795 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2796 *(dh
->dh_dbp
) = dh
->dh_db
;
2802 * The following code preserves the recursive function dbuf_hold_impl()
2803 * but moves the local variables AND function arguments to the heap to
2804 * minimize the stack frame size. Enough space is initially allocated
2805 * on the stack for 20 levels of recursion.
2808 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2809 boolean_t fail_sparse
, boolean_t fail_uncached
,
2810 void *tag
, dmu_buf_impl_t
**dbp
)
2812 struct dbuf_hold_impl_data
*dh
;
2815 dh
= kmem_alloc(sizeof (struct dbuf_hold_impl_data
) *
2816 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2817 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2818 fail_uncached
, tag
, dbp
, 0);
2820 error
= __dbuf_hold_impl(dh
);
2822 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2823 DBUF_HOLD_IMPL_MAX_DEPTH
);
2829 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2830 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2831 boolean_t fail_sparse
, boolean_t fail_uncached
,
2832 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2835 dh
->dh_level
= level
;
2836 dh
->dh_blkid
= blkid
;
2838 dh
->dh_fail_sparse
= fail_sparse
;
2839 dh
->dh_fail_uncached
= fail_uncached
;
2845 dh
->dh_parent
= NULL
;
2851 dh
->dh_depth
= depth
;
2855 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2857 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2861 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2864 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2865 return (err
? NULL
: db
);
2869 dbuf_create_bonus(dnode_t
*dn
)
2871 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2873 ASSERT(dn
->dn_bonus
== NULL
);
2874 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2878 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2880 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2883 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2884 return (SET_ERROR(ENOTSUP
));
2886 blksz
= SPA_MINBLOCKSIZE
;
2887 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2888 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2892 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2893 dbuf_new_size(db
, blksz
, tx
);
2894 rw_exit(&dn
->dn_struct_rwlock
);
2901 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2903 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2906 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2908 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2910 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
2911 VERIFY3S(holds
, >, 1);
2914 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2916 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2919 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2920 dmu_buf_impl_t
*found_db
;
2921 boolean_t result
= B_FALSE
;
2923 if (blkid
== DMU_BONUS_BLKID
)
2924 found_db
= dbuf_find_bonus(os
, obj
);
2926 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2928 if (found_db
!= NULL
) {
2929 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2930 (void) refcount_add(&db
->db_holds
, tag
);
2933 mutex_exit(&found_db
->db_mtx
);
2939 * If you call dbuf_rele() you had better not be referencing the dnode handle
2940 * unless you have some other direct or indirect hold on the dnode. (An indirect
2941 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2942 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2943 * dnode's parent dbuf evicting its dnode handles.
2946 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2948 mutex_enter(&db
->db_mtx
);
2949 dbuf_rele_and_unlock(db
, tag
);
2953 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2955 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2959 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2960 * db_dirtycnt and db_holds to be updated atomically.
2963 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2967 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2971 * Remove the reference to the dbuf before removing its hold on the
2972 * dnode so we can guarantee in dnode_move() that a referenced bonus
2973 * buffer has a corresponding dnode hold.
2975 holds
= refcount_remove(&db
->db_holds
, tag
);
2979 * We can't freeze indirects if there is a possibility that they
2980 * may be modified in the current syncing context.
2982 if (db
->db_buf
!= NULL
&&
2983 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
2984 arc_buf_freeze(db
->db_buf
);
2987 if (holds
== db
->db_dirtycnt
&&
2988 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2989 dbuf_evict_user(db
);
2992 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2994 boolean_t evict_dbuf
= db
->db_pending_evict
;
2997 * If the dnode moves here, we cannot cross this
2998 * barrier until the move completes.
3003 atomic_dec_32(&dn
->dn_dbufs_count
);
3006 * Decrementing the dbuf count means that the bonus
3007 * buffer's dnode hold is no longer discounted in
3008 * dnode_move(). The dnode cannot move until after
3009 * the dnode_rele() below.
3014 * Do not reference db after its lock is dropped.
3015 * Another thread may evict it.
3017 mutex_exit(&db
->db_mtx
);
3020 dnode_evict_bonus(dn
);
3023 } else if (db
->db_buf
== NULL
) {
3025 * This is a special case: we never associated this
3026 * dbuf with any data allocated from the ARC.
3028 ASSERT(db
->db_state
== DB_UNCACHED
||
3029 db
->db_state
== DB_NOFILL
);
3031 } else if (arc_released(db
->db_buf
)) {
3033 * This dbuf has anonymous data associated with it.
3037 boolean_t do_arc_evict
= B_FALSE
;
3039 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3041 if (!DBUF_IS_CACHEABLE(db
) &&
3042 db
->db_blkptr
!= NULL
&&
3043 !BP_IS_HOLE(db
->db_blkptr
) &&
3044 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
3045 do_arc_evict
= B_TRUE
;
3046 bp
= *db
->db_blkptr
;
3049 if (!DBUF_IS_CACHEABLE(db
) ||
3050 db
->db_pending_evict
) {
3052 } else if (!multilist_link_active(&db
->db_cache_link
)) {
3053 multilist_insert(dbuf_cache
, db
);
3054 (void) refcount_add_many(&dbuf_cache_size
,
3055 db
->db
.db_size
, db
);
3056 mutex_exit(&db
->db_mtx
);
3058 dbuf_evict_notify();
3062 arc_freed(spa
, &bp
);
3065 mutex_exit(&db
->db_mtx
);
3070 #pragma weak dmu_buf_refcount = dbuf_refcount
3072 dbuf_refcount(dmu_buf_impl_t
*db
)
3074 return (refcount_count(&db
->db_holds
));
3078 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3079 dmu_buf_user_t
*new_user
)
3081 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3083 mutex_enter(&db
->db_mtx
);
3084 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3085 if (db
->db_user
== old_user
)
3086 db
->db_user
= new_user
;
3088 old_user
= db
->db_user
;
3089 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3090 mutex_exit(&db
->db_mtx
);
3096 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3098 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3102 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3104 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3106 db
->db_user_immediate_evict
= TRUE
;
3107 return (dmu_buf_set_user(db_fake
, user
));
3111 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3113 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3117 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3119 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3121 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3122 return (db
->db_user
);
3126 dmu_buf_user_evict_wait()
3128 taskq_wait(dbu_evict_taskq
);
3132 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3134 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3135 return (dbi
->db_blkptr
);
3139 dmu_buf_get_objset(dmu_buf_t
*db
)
3141 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3142 return (dbi
->db_objset
);
3146 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3148 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3149 DB_DNODE_ENTER(dbi
);
3150 return (DB_DNODE(dbi
));
3154 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3156 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3161 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3163 /* ASSERT(dmu_tx_is_syncing(tx) */
3164 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3166 if (db
->db_blkptr
!= NULL
)
3169 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3170 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3171 BP_ZERO(db
->db_blkptr
);
3174 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3176 * This buffer was allocated at a time when there was
3177 * no available blkptrs from the dnode, or it was
3178 * inappropriate to hook it in (i.e., nlevels mis-match).
3180 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3181 ASSERT(db
->db_parent
== NULL
);
3182 db
->db_parent
= dn
->dn_dbuf
;
3183 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3186 dmu_buf_impl_t
*parent
= db
->db_parent
;
3187 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3189 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3190 if (parent
== NULL
) {
3191 mutex_exit(&db
->db_mtx
);
3192 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3193 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3194 db
->db_blkid
>> epbs
, db
);
3195 rw_exit(&dn
->dn_struct_rwlock
);
3196 mutex_enter(&db
->db_mtx
);
3197 db
->db_parent
= parent
;
3199 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3200 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3206 * Ensure the dbuf's data is untransformed if the associated dirty
3207 * record requires it. This is used by dbuf_sync_leaf() to ensure
3208 * that a dnode block is decrypted before we write new data to it.
3209 * For raw writes we assert that the buffer is already encrypted.
3212 dbuf_check_crypt(dbuf_dirty_record_t
*dr
)
3215 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3217 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3219 if (!dr
->dt
.dl
.dr_raw
&& arc_is_encrypted(db
->db_buf
)) {
3221 * Unfortunately, there is currently no mechanism for
3222 * syncing context to handle decryption errors. An error
3223 * here is only possible if an attacker maliciously
3224 * changed a dnode block and updated the associated
3225 * checksums going up the block tree.
3227 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
3228 dmu_objset_id(db
->db_objset
), B_TRUE
);
3230 panic("Invalid dnode block MAC");
3231 } else if (dr
->dt
.dl
.dr_raw
) {
3233 * Writing raw encrypted data requires the db's arc buffer
3234 * to be converted to raw by the caller.
3236 ASSERT(arc_is_encrypted(db
->db_buf
));
3241 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3242 * is critical the we not allow the compiler to inline this function in to
3243 * dbuf_sync_list() thereby drastically bloating the stack usage.
3245 noinline
static void
3246 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3248 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3252 ASSERT(dmu_tx_is_syncing(tx
));
3254 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3256 mutex_enter(&db
->db_mtx
);
3258 ASSERT(db
->db_level
> 0);
3261 /* Read the block if it hasn't been read yet. */
3262 if (db
->db_buf
== NULL
) {
3263 mutex_exit(&db
->db_mtx
);
3264 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3265 mutex_enter(&db
->db_mtx
);
3267 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3268 ASSERT(db
->db_buf
!= NULL
);
3272 /* Indirect block size must match what the dnode thinks it is. */
3273 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3274 dbuf_check_blkptr(dn
, db
);
3277 /* Provide the pending dirty record to child dbufs */
3278 db
->db_data_pending
= dr
;
3280 mutex_exit(&db
->db_mtx
);
3281 dbuf_write(dr
, db
->db_buf
, tx
);
3284 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3285 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3286 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3287 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3292 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3293 * critical the we not allow the compiler to inline this function in to
3294 * dbuf_sync_list() thereby drastically bloating the stack usage.
3296 noinline
static void
3297 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3299 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3300 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3303 uint64_t txg
= tx
->tx_txg
;
3305 ASSERT(dmu_tx_is_syncing(tx
));
3307 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3309 mutex_enter(&db
->db_mtx
);
3311 * To be synced, we must be dirtied. But we
3312 * might have been freed after the dirty.
3314 if (db
->db_state
== DB_UNCACHED
) {
3315 /* This buffer has been freed since it was dirtied */
3316 ASSERT(db
->db
.db_data
== NULL
);
3317 } else if (db
->db_state
== DB_FILL
) {
3318 /* This buffer was freed and is now being re-filled */
3319 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3321 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3328 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3329 mutex_enter(&dn
->dn_mtx
);
3330 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3332 * In the previous transaction group, the bonus buffer
3333 * was entirely used to store the attributes for the
3334 * dnode which overrode the dn_spill field. However,
3335 * when adding more attributes to the file a spill
3336 * block was required to hold the extra attributes.
3338 * Make sure to clear the garbage left in the dn_spill
3339 * field from the previous attributes in the bonus
3340 * buffer. Otherwise, after writing out the spill
3341 * block to the new allocated dva, it will free
3342 * the old block pointed to by the invalid dn_spill.
3344 db
->db_blkptr
= NULL
;
3346 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3347 mutex_exit(&dn
->dn_mtx
);
3351 * If this is a bonus buffer, simply copy the bonus data into the
3352 * dnode. It will be written out when the dnode is synced (and it
3353 * will be synced, since it must have been dirty for dbuf_sync to
3356 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3357 dbuf_dirty_record_t
**drp
;
3359 ASSERT(*datap
!= NULL
);
3360 ASSERT0(db
->db_level
);
3361 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
3362 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3363 bcopy(*datap
, DN_BONUS(dn
->dn_phys
),
3364 DN_MAX_BONUS_LEN(dn
->dn_phys
));
3367 if (*datap
!= db
->db
.db_data
) {
3368 int slots
= DB_DNODE(db
)->dn_num_slots
;
3369 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3370 kmem_free(*datap
, bonuslen
);
3371 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3373 db
->db_data_pending
= NULL
;
3374 drp
= &db
->db_last_dirty
;
3376 drp
= &(*drp
)->dr_next
;
3377 ASSERT(dr
->dr_next
== NULL
);
3378 ASSERT(dr
->dr_dbuf
== db
);
3380 if (dr
->dr_dbuf
->db_level
!= 0) {
3381 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3382 list_destroy(&dr
->dt
.di
.dr_children
);
3384 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3385 ASSERT(db
->db_dirtycnt
> 0);
3386 db
->db_dirtycnt
-= 1;
3387 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3394 * This function may have dropped the db_mtx lock allowing a dmu_sync
3395 * operation to sneak in. As a result, we need to ensure that we
3396 * don't check the dr_override_state until we have returned from
3397 * dbuf_check_blkptr.
3399 dbuf_check_blkptr(dn
, db
);
3402 * If this buffer is in the middle of an immediate write,
3403 * wait for the synchronous IO to complete.
3405 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3406 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3407 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3408 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3412 * If this is a dnode block, ensure it is appropriately encrypted
3413 * or decrypted, depending on what we are writing to it this txg.
3415 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
3416 dbuf_check_crypt(dr
);
3418 if (db
->db_state
!= DB_NOFILL
&&
3419 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3420 refcount_count(&db
->db_holds
) > 1 &&
3421 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3422 *datap
== db
->db_buf
) {
3424 * If this buffer is currently "in use" (i.e., there
3425 * are active holds and db_data still references it),
3426 * then make a copy before we start the write so that
3427 * any modifications from the open txg will not leak
3430 * NOTE: this copy does not need to be made for
3431 * objects only modified in the syncing context (e.g.
3432 * DNONE_DNODE blocks).
3434 int psize
= arc_buf_size(*datap
);
3435 int lsize
= arc_buf_lsize(*datap
);
3436 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3437 enum zio_compress compress_type
= arc_get_compression(*datap
);
3439 if (arc_is_encrypted(*datap
)) {
3440 boolean_t byteorder
;
3441 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3442 uint8_t iv
[ZIO_DATA_IV_LEN
];
3443 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3445 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
3446 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
3447 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
3448 dn
->dn_type
, psize
, lsize
, compress_type
);
3449 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3450 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
3451 int lsize
= arc_buf_lsize(*datap
);
3452 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
3453 psize
, lsize
, compress_type
);
3455 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
3457 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
3459 db
->db_data_pending
= dr
;
3461 mutex_exit(&db
->db_mtx
);
3463 dbuf_write(dr
, *datap
, tx
);
3465 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3466 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3467 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3471 * Although zio_nowait() does not "wait for an IO", it does
3472 * initiate the IO. If this is an empty write it seems plausible
3473 * that the IO could actually be completed before the nowait
3474 * returns. We need to DB_DNODE_EXIT() first in case
3475 * zio_nowait() invalidates the dbuf.
3478 zio_nowait(dr
->dr_zio
);
3483 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3485 dbuf_dirty_record_t
*dr
;
3487 while ((dr
= list_head(list
))) {
3488 if (dr
->dr_zio
!= NULL
) {
3490 * If we find an already initialized zio then we
3491 * are processing the meta-dnode, and we have finished.
3492 * The dbufs for all dnodes are put back on the list
3493 * during processing, so that we can zio_wait()
3494 * these IOs after initiating all child IOs.
3496 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3497 DMU_META_DNODE_OBJECT
);
3500 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3501 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3502 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3504 list_remove(list
, dr
);
3505 if (dr
->dr_dbuf
->db_level
> 0)
3506 dbuf_sync_indirect(dr
, tx
);
3508 dbuf_sync_leaf(dr
, tx
);
3514 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3516 dmu_buf_impl_t
*db
= vdb
;
3518 blkptr_t
*bp
= zio
->io_bp
;
3519 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3520 spa_t
*spa
= zio
->io_spa
;
3525 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3526 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3530 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3531 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3532 zio
->io_prev_space_delta
= delta
;
3534 if (bp
->blk_birth
!= 0) {
3535 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3536 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3537 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3538 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3539 BP_IS_EMBEDDED(bp
));
3540 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3543 mutex_enter(&db
->db_mtx
);
3546 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3547 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3548 ASSERT(!(BP_IS_HOLE(bp
)) &&
3549 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3553 if (db
->db_level
== 0) {
3554 mutex_enter(&dn
->dn_mtx
);
3555 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3556 db
->db_blkid
!= DMU_SPILL_BLKID
)
3557 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3558 mutex_exit(&dn
->dn_mtx
);
3560 if (dn
->dn_type
== DMU_OT_DNODE
) {
3562 while (i
< db
->db
.db_size
) {
3564 (void *)(((char *)db
->db
.db_data
) + i
);
3566 i
+= DNODE_MIN_SIZE
;
3567 if (dnp
->dn_type
!= DMU_OT_NONE
) {
3569 i
+= dnp
->dn_extra_slots
*
3574 if (BP_IS_HOLE(bp
)) {
3581 blkptr_t
*ibp
= db
->db
.db_data
;
3582 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3583 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3584 if (BP_IS_HOLE(ibp
))
3586 fill
+= BP_GET_FILL(ibp
);
3591 if (!BP_IS_EMBEDDED(bp
))
3592 BP_SET_FILL(bp
, fill
);
3594 mutex_exit(&db
->db_mtx
);
3596 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3597 *db
->db_blkptr
= *bp
;
3598 rw_exit(&dn
->dn_struct_rwlock
);
3603 * This function gets called just prior to running through the compression
3604 * stage of the zio pipeline. If we're an indirect block comprised of only
3605 * holes, then we want this indirect to be compressed away to a hole. In
3606 * order to do that we must zero out any information about the holes that
3607 * this indirect points to prior to before we try to compress it.
3610 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3612 dmu_buf_impl_t
*db
= vdb
;
3615 unsigned int epbs
, i
;
3617 ASSERT3U(db
->db_level
, >, 0);
3620 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3621 ASSERT3U(epbs
, <, 31);
3623 /* Determine if all our children are holes */
3624 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
3625 if (!BP_IS_HOLE(bp
))
3630 * If all the children are holes, then zero them all out so that
3631 * we may get compressed away.
3633 if (i
== 1ULL << epbs
) {
3635 * We only found holes. Grab the rwlock to prevent
3636 * anybody from reading the blocks we're about to
3639 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3640 bzero(db
->db
.db_data
, db
->db
.db_size
);
3641 rw_exit(&dn
->dn_struct_rwlock
);
3647 * The SPA will call this callback several times for each zio - once
3648 * for every physical child i/o (zio->io_phys_children times). This
3649 * allows the DMU to monitor the progress of each logical i/o. For example,
3650 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3651 * block. There may be a long delay before all copies/fragments are completed,
3652 * so this callback allows us to retire dirty space gradually, as the physical
3657 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3659 dmu_buf_impl_t
*db
= arg
;
3660 objset_t
*os
= db
->db_objset
;
3661 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3662 dbuf_dirty_record_t
*dr
;
3665 dr
= db
->db_data_pending
;
3666 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3669 * The callback will be called io_phys_children times. Retire one
3670 * portion of our dirty space each time we are called. Any rounding
3671 * error will be cleaned up by dsl_pool_sync()'s call to
3672 * dsl_pool_undirty_space().
3674 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3675 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3680 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3682 dmu_buf_impl_t
*db
= vdb
;
3683 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3684 blkptr_t
*bp
= db
->db_blkptr
;
3685 objset_t
*os
= db
->db_objset
;
3686 dmu_tx_t
*tx
= os
->os_synctx
;
3687 dbuf_dirty_record_t
**drp
, *dr
;
3689 ASSERT0(zio
->io_error
);
3690 ASSERT(db
->db_blkptr
== bp
);
3693 * For nopwrites and rewrites we ensure that the bp matches our
3694 * original and bypass all the accounting.
3696 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3697 ASSERT(BP_EQUAL(bp
, bp_orig
));
3699 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3700 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3701 dsl_dataset_block_born(ds
, bp
, tx
);
3704 mutex_enter(&db
->db_mtx
);
3708 drp
= &db
->db_last_dirty
;
3709 while ((dr
= *drp
) != db
->db_data_pending
)
3711 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3712 ASSERT(dr
->dr_dbuf
== db
);
3713 ASSERT(dr
->dr_next
== NULL
);
3717 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3722 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3723 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3724 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3729 if (db
->db_level
== 0) {
3730 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3731 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3732 if (db
->db_state
!= DB_NOFILL
) {
3733 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3734 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
3741 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3742 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3743 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3744 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3746 ASSERT3U(db
->db_blkid
, <=,
3747 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3748 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3752 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3753 list_destroy(&dr
->dt
.di
.dr_children
);
3755 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3757 cv_broadcast(&db
->db_changed
);
3758 ASSERT(db
->db_dirtycnt
> 0);
3759 db
->db_dirtycnt
-= 1;
3760 db
->db_data_pending
= NULL
;
3761 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3765 dbuf_write_nofill_ready(zio_t
*zio
)
3767 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3771 dbuf_write_nofill_done(zio_t
*zio
)
3773 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3777 dbuf_write_override_ready(zio_t
*zio
)
3779 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3780 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3782 dbuf_write_ready(zio
, NULL
, db
);
3786 dbuf_write_override_done(zio_t
*zio
)
3788 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3789 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3790 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3792 mutex_enter(&db
->db_mtx
);
3793 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3794 if (!BP_IS_HOLE(obp
))
3795 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3796 arc_release(dr
->dt
.dl
.dr_data
, db
);
3798 mutex_exit(&db
->db_mtx
);
3800 dbuf_write_done(zio
, NULL
, db
);
3802 if (zio
->io_abd
!= NULL
)
3803 abd_put(zio
->io_abd
);
3806 /* Issue I/O to commit a dirty buffer to disk. */
3808 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3810 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3813 dmu_buf_impl_t
*parent
= db
->db_parent
;
3814 uint64_t txg
= tx
->tx_txg
;
3815 zbookmark_phys_t zb
;
3820 ASSERT(dmu_tx_is_syncing(tx
));
3826 if (db
->db_state
!= DB_NOFILL
) {
3827 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3829 * Private object buffers are released here rather
3830 * than in dbuf_dirty() since they are only modified
3831 * in the syncing context and we don't want the
3832 * overhead of making multiple copies of the data.
3834 if (BP_IS_HOLE(db
->db_blkptr
)) {
3837 dbuf_release_bp(db
);
3842 if (parent
!= dn
->dn_dbuf
) {
3843 /* Our parent is an indirect block. */
3844 /* We have a dirty parent that has been scheduled for write. */
3845 ASSERT(parent
&& parent
->db_data_pending
);
3846 /* Our parent's buffer is one level closer to the dnode. */
3847 ASSERT(db
->db_level
== parent
->db_level
-1);
3849 * We're about to modify our parent's db_data by modifying
3850 * our block pointer, so the parent must be released.
3852 ASSERT(arc_released(parent
->db_buf
));
3853 zio
= parent
->db_data_pending
->dr_zio
;
3855 /* Our parent is the dnode itself. */
3856 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3857 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3858 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3859 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3860 ASSERT3P(db
->db_blkptr
, ==,
3861 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3865 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3866 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3869 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3870 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3871 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3873 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3875 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3877 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3881 * We copy the blkptr now (rather than when we instantiate the dirty
3882 * record), because its value can change between open context and
3883 * syncing context. We do not need to hold dn_struct_rwlock to read
3884 * db_blkptr because we are in syncing context.
3886 dr
->dr_bp_copy
= *db
->db_blkptr
;
3888 if (db
->db_level
== 0 &&
3889 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3891 * The BP for this block has been provided by open context
3892 * (by dmu_sync() or dmu_buf_write_embedded()).
3894 abd_t
*contents
= (data
!= NULL
) ?
3895 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
3897 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3898 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
3899 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
3900 dbuf_write_override_done
,
3901 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3902 mutex_enter(&db
->db_mtx
);
3903 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3904 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3905 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3906 mutex_exit(&db
->db_mtx
);
3907 } else if (db
->db_state
== DB_NOFILL
) {
3908 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
3909 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
3910 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3911 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
3912 dbuf_write_nofill_ready
, NULL
, NULL
,
3913 dbuf_write_nofill_done
, db
,
3914 ZIO_PRIORITY_ASYNC_WRITE
,
3915 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3917 arc_write_done_func_t
*children_ready_cb
= NULL
;
3918 ASSERT(arc_released(data
));
3921 * For indirect blocks, we want to setup the children
3922 * ready callback so that we can properly handle an indirect
3923 * block that only contains holes.
3925 if (db
->db_level
!= 0)
3926 children_ready_cb
= dbuf_write_children_ready
;
3928 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3929 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3930 &zp
, dbuf_write_ready
,
3931 children_ready_cb
, dbuf_write_physdone
,
3932 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
3933 ZIO_FLAG_MUSTSUCCEED
, &zb
);
3937 #if defined(_KERNEL) && defined(HAVE_SPL)
3938 EXPORT_SYMBOL(dbuf_find
);
3939 EXPORT_SYMBOL(dbuf_is_metadata
);
3940 EXPORT_SYMBOL(dbuf_destroy
);
3941 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
3942 EXPORT_SYMBOL(dbuf_whichblock
);
3943 EXPORT_SYMBOL(dbuf_read
);
3944 EXPORT_SYMBOL(dbuf_unoverride
);
3945 EXPORT_SYMBOL(dbuf_free_range
);
3946 EXPORT_SYMBOL(dbuf_new_size
);
3947 EXPORT_SYMBOL(dbuf_release_bp
);
3948 EXPORT_SYMBOL(dbuf_dirty
);
3949 EXPORT_SYMBOL(dmu_buf_will_change_crypt_params
);
3950 EXPORT_SYMBOL(dmu_buf_will_dirty
);
3951 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
3952 EXPORT_SYMBOL(dmu_buf_will_fill
);
3953 EXPORT_SYMBOL(dmu_buf_fill_done
);
3954 EXPORT_SYMBOL(dmu_buf_rele
);
3955 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
3956 EXPORT_SYMBOL(dbuf_prefetch
);
3957 EXPORT_SYMBOL(dbuf_hold_impl
);
3958 EXPORT_SYMBOL(dbuf_hold
);
3959 EXPORT_SYMBOL(dbuf_hold_level
);
3960 EXPORT_SYMBOL(dbuf_create_bonus
);
3961 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
3962 EXPORT_SYMBOL(dbuf_rm_spill
);
3963 EXPORT_SYMBOL(dbuf_add_ref
);
3964 EXPORT_SYMBOL(dbuf_rele
);
3965 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
3966 EXPORT_SYMBOL(dbuf_refcount
);
3967 EXPORT_SYMBOL(dbuf_sync_list
);
3968 EXPORT_SYMBOL(dmu_buf_set_user
);
3969 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3970 EXPORT_SYMBOL(dmu_buf_get_user
);
3971 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
3974 module_param(dbuf_cache_max_bytes
, ulong
, 0644);
3975 MODULE_PARM_DESC(dbuf_cache_max_bytes
,
3976 "Maximum size in bytes of the dbuf cache.");
3978 module_param(dbuf_cache_hiwater_pct
, uint
, 0644);
3979 MODULE_PARM_DESC(dbuf_cache_hiwater_pct
,
3980 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
3983 module_param(dbuf_cache_lowater_pct
, uint
, 0644);
3984 MODULE_PARM_DESC(dbuf_cache_lowater_pct
,
3985 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
3988 module_param(dbuf_cache_max_shift
, int, 0644);
3989 MODULE_PARM_DESC(dbuf_cache_max_shift
,
3990 "Cap the size of the dbuf cache to a log2 fraction of arc size.");