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 by Delphix. All rights reserved.
27 #include <sys/zfs_context.h>
29 #include <sys/dmu_impl.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dmu_tx.h>
37 #include <sys/dmu_zfetch.h>
39 #include <sys/sa_impl.h>
41 static void dbuf_destroy(dmu_buf_impl_t
*db
);
42 static int dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
43 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
46 * Global data structures and functions for the dbuf cache.
48 static kmem_cache_t
*dbuf_cache
;
52 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
54 dmu_buf_impl_t
*db
= vdb
;
55 bzero(db
, sizeof (dmu_buf_impl_t
));
57 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
58 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
59 refcount_create(&db
->db_holds
);
65 dbuf_dest(void *vdb
, void *unused
)
67 dmu_buf_impl_t
*db
= vdb
;
68 mutex_destroy(&db
->db_mtx
);
69 cv_destroy(&db
->db_changed
);
70 refcount_destroy(&db
->db_holds
);
74 * dbuf hash table routines
76 static dbuf_hash_table_t dbuf_hash_table
;
78 static uint64_t dbuf_hash_count
;
81 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
83 uintptr_t osv
= (uintptr_t)os
;
86 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
87 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
88 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
89 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
90 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
91 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
92 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
94 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
99 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
101 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
102 ((dbuf)->db.db_object == (obj) && \
103 (dbuf)->db_objset == (os) && \
104 (dbuf)->db_level == (level) && \
105 (dbuf)->db_blkid == (blkid))
108 dbuf_find(dnode_t
*dn
, uint8_t level
, uint64_t blkid
)
110 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
111 objset_t
*os
= dn
->dn_objset
;
112 uint64_t obj
= dn
->dn_object
;
113 uint64_t hv
= DBUF_HASH(os
, obj
, level
, blkid
);
114 uint64_t idx
= hv
& h
->hash_table_mask
;
117 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
118 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
119 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
120 mutex_enter(&db
->db_mtx
);
121 if (db
->db_state
!= DB_EVICTING
) {
122 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
125 mutex_exit(&db
->db_mtx
);
128 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
133 * Insert an entry into the hash table. If there is already an element
134 * equal to elem in the hash table, then the already existing element
135 * will be returned and the new element will not be inserted.
136 * Otherwise returns NULL.
138 static dmu_buf_impl_t
*
139 dbuf_hash_insert(dmu_buf_impl_t
*db
)
141 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
142 objset_t
*os
= db
->db_objset
;
143 uint64_t obj
= db
->db
.db_object
;
144 int level
= db
->db_level
;
145 uint64_t blkid
= db
->db_blkid
;
146 uint64_t hv
= DBUF_HASH(os
, obj
, level
, blkid
);
147 uint64_t idx
= hv
& h
->hash_table_mask
;
150 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
151 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
152 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
153 mutex_enter(&dbf
->db_mtx
);
154 if (dbf
->db_state
!= DB_EVICTING
) {
155 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
158 mutex_exit(&dbf
->db_mtx
);
162 mutex_enter(&db
->db_mtx
);
163 db
->db_hash_next
= h
->hash_table
[idx
];
164 h
->hash_table
[idx
] = db
;
165 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
166 atomic_add_64(&dbuf_hash_count
, 1);
172 * Remove an entry from the hash table. This operation will
173 * fail if there are any existing holds on the db.
176 dbuf_hash_remove(dmu_buf_impl_t
*db
)
178 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
179 uint64_t hv
= DBUF_HASH(db
->db_objset
, db
->db
.db_object
,
180 db
->db_level
, db
->db_blkid
);
181 uint64_t idx
= hv
& h
->hash_table_mask
;
182 dmu_buf_impl_t
*dbf
, **dbp
;
185 * We musn't hold db_mtx to maintin lock ordering:
186 * DBUF_HASH_MUTEX > db_mtx.
188 ASSERT(refcount_is_zero(&db
->db_holds
));
189 ASSERT(db
->db_state
== DB_EVICTING
);
190 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
192 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
193 dbp
= &h
->hash_table
[idx
];
194 while ((dbf
= *dbp
) != db
) {
195 dbp
= &dbf
->db_hash_next
;
198 *dbp
= db
->db_hash_next
;
199 db
->db_hash_next
= NULL
;
200 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
201 atomic_add_64(&dbuf_hash_count
, -1);
204 static arc_evict_func_t dbuf_do_evict
;
207 dbuf_evict_user(dmu_buf_impl_t
*db
)
209 ASSERT(MUTEX_HELD(&db
->db_mtx
));
211 if (db
->db_level
!= 0 || db
->db_evict_func
== NULL
)
214 if (db
->db_user_data_ptr_ptr
)
215 *db
->db_user_data_ptr_ptr
= db
->db
.db_data
;
216 db
->db_evict_func(&db
->db
, db
->db_user_ptr
);
217 db
->db_user_ptr
= NULL
;
218 db
->db_user_data_ptr_ptr
= NULL
;
219 db
->db_evict_func
= NULL
;
223 dbuf_is_metadata(dmu_buf_impl_t
*db
)
225 if (db
->db_level
> 0) {
228 boolean_t is_metadata
;
231 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
234 return (is_metadata
);
239 dbuf_evict(dmu_buf_impl_t
*db
)
241 ASSERT(MUTEX_HELD(&db
->db_mtx
));
242 ASSERT(db
->db_buf
== NULL
);
243 ASSERT(db
->db_data_pending
== NULL
);
252 uint64_t hsize
= 1ULL << 16;
253 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
257 * The hash table is big enough to fill all of physical memory
258 * with an average 4K block size. The table will take up
259 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
261 while (hsize
* 4096 < physmem
* PAGESIZE
)
265 h
->hash_table_mask
= hsize
- 1;
266 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
267 if (h
->hash_table
== NULL
) {
268 /* XXX - we should really return an error instead of assert */
269 ASSERT(hsize
> (1ULL << 10));
274 dbuf_cache
= kmem_cache_create("dmu_buf_impl_t",
275 sizeof (dmu_buf_impl_t
),
276 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
278 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
279 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
285 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
288 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
289 mutex_destroy(&h
->hash_mutexes
[i
]);
290 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
291 kmem_cache_destroy(dbuf_cache
);
300 dbuf_verify(dmu_buf_impl_t
*db
)
303 dbuf_dirty_record_t
*dr
;
305 ASSERT(MUTEX_HELD(&db
->db_mtx
));
307 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
310 ASSERT(db
->db_objset
!= NULL
);
314 ASSERT(db
->db_parent
== NULL
);
315 ASSERT(db
->db_blkptr
== NULL
);
317 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
318 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
319 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
320 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
321 db
->db_blkid
== DMU_SPILL_BLKID
||
322 !list_is_empty(&dn
->dn_dbufs
));
324 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
326 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
327 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
328 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
330 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
331 ASSERT0(db
->db
.db_offset
);
333 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
336 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
337 ASSERT(dr
->dr_dbuf
== db
);
339 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
340 ASSERT(dr
->dr_dbuf
== db
);
343 * We can't assert that db_size matches dn_datablksz because it
344 * can be momentarily different when another thread is doing
347 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
348 dr
= db
->db_data_pending
;
350 * It should only be modified in syncing context, so
351 * make sure we only have one copy of the data.
353 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
356 /* verify db->db_blkptr */
358 if (db
->db_parent
== dn
->dn_dbuf
) {
359 /* db is pointed to by the dnode */
360 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
361 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
362 ASSERT(db
->db_parent
== NULL
);
364 ASSERT(db
->db_parent
!= NULL
);
365 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
366 ASSERT3P(db
->db_blkptr
, ==,
367 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
369 /* db is pointed to by an indirect block */
370 int epb
= db
->db_parent
->db
.db_size
>> SPA_BLKPTRSHIFT
;
371 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
372 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
375 * dnode_grow_indblksz() can make this fail if we don't
376 * have the struct_rwlock. XXX indblksz no longer
377 * grows. safe to do this now?
379 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
380 ASSERT3P(db
->db_blkptr
, ==,
381 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
382 db
->db_blkid
% epb
));
386 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
387 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
388 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
389 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
391 * If the blkptr isn't set but they have nonzero data,
392 * it had better be dirty, otherwise we'll lose that
393 * data when we evict this buffer.
395 if (db
->db_dirtycnt
== 0) {
396 uint64_t *buf
= db
->db
.db_data
;
399 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
409 dbuf_update_data(dmu_buf_impl_t
*db
)
411 ASSERT(MUTEX_HELD(&db
->db_mtx
));
412 if (db
->db_level
== 0 && db
->db_user_data_ptr_ptr
) {
413 ASSERT(!refcount_is_zero(&db
->db_holds
));
414 *db
->db_user_data_ptr_ptr
= db
->db
.db_data
;
419 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
421 ASSERT(MUTEX_HELD(&db
->db_mtx
));
422 ASSERT(db
->db_buf
== NULL
|| !arc_has_callback(db
->db_buf
));
425 ASSERT(buf
->b_data
!= NULL
);
426 db
->db
.db_data
= buf
->b_data
;
427 if (!arc_released(buf
))
428 arc_set_callback(buf
, dbuf_do_evict
, db
);
429 dbuf_update_data(db
);
432 db
->db
.db_data
= NULL
;
433 if (db
->db_state
!= DB_NOFILL
)
434 db
->db_state
= DB_UNCACHED
;
439 * Loan out an arc_buf for read. Return the loaned arc_buf.
442 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
446 mutex_enter(&db
->db_mtx
);
447 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
448 int blksz
= db
->db
.db_size
;
451 mutex_exit(&db
->db_mtx
);
452 DB_GET_SPA(&spa
, db
);
453 abuf
= arc_loan_buf(spa
, blksz
);
454 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
457 arc_loan_inuse_buf(abuf
, db
);
458 dbuf_set_data(db
, NULL
);
459 mutex_exit(&db
->db_mtx
);
465 dbuf_whichblock(dnode_t
*dn
, uint64_t offset
)
467 if (dn
->dn_datablkshift
) {
468 return (offset
>> dn
->dn_datablkshift
);
470 ASSERT3U(offset
, <, dn
->dn_datablksz
);
476 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
478 dmu_buf_impl_t
*db
= vdb
;
480 mutex_enter(&db
->db_mtx
);
481 ASSERT3U(db
->db_state
, ==, DB_READ
);
483 * All reads are synchronous, so we must have a hold on the dbuf
485 ASSERT(refcount_count(&db
->db_holds
) > 0);
486 ASSERT(db
->db_buf
== NULL
);
487 ASSERT(db
->db
.db_data
== NULL
);
488 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
489 /* we were freed in flight; disregard any error */
490 arc_release(buf
, db
);
491 bzero(buf
->b_data
, db
->db
.db_size
);
493 db
->db_freed_in_flight
= FALSE
;
494 dbuf_set_data(db
, buf
);
495 db
->db_state
= DB_CACHED
;
496 } else if (zio
== NULL
|| zio
->io_error
== 0) {
497 dbuf_set_data(db
, buf
);
498 db
->db_state
= DB_CACHED
;
500 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
501 ASSERT3P(db
->db_buf
, ==, NULL
);
502 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
503 db
->db_state
= DB_UNCACHED
;
505 cv_broadcast(&db
->db_changed
);
506 dbuf_rele_and_unlock(db
, NULL
);
510 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t *flags
)
515 uint32_t aflags
= ARC_NOWAIT
;
520 ASSERT(!refcount_is_zero(&db
->db_holds
));
521 /* We need the struct_rwlock to prevent db_blkptr from changing. */
522 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
523 ASSERT(MUTEX_HELD(&db
->db_mtx
));
524 ASSERT(db
->db_state
== DB_UNCACHED
);
525 ASSERT(db
->db_buf
== NULL
);
527 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
528 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
530 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
531 db
->db
.db_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
532 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
533 if (bonuslen
< DN_MAX_BONUSLEN
)
534 bzero(db
->db
.db_data
, DN_MAX_BONUSLEN
);
536 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
538 dbuf_update_data(db
);
539 db
->db_state
= DB_CACHED
;
540 mutex_exit(&db
->db_mtx
);
545 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
546 * processes the delete record and clears the bp while we are waiting
547 * for the dn_mtx (resulting in a "no" from block_freed).
549 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
550 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
551 BP_IS_HOLE(db
->db_blkptr
)))) {
552 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
554 dbuf_set_data(db
, arc_buf_alloc(dn
->dn_objset
->os_spa
,
555 db
->db
.db_size
, db
, type
));
557 bzero(db
->db
.db_data
, db
->db
.db_size
);
558 db
->db_state
= DB_CACHED
;
559 *flags
|= DB_RF_CACHED
;
560 mutex_exit(&db
->db_mtx
);
564 spa
= dn
->dn_objset
->os_spa
;
567 db
->db_state
= DB_READ
;
568 mutex_exit(&db
->db_mtx
);
570 if (DBUF_IS_L2CACHEABLE(db
))
571 aflags
|= ARC_L2CACHE
;
573 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
574 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
575 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
577 dbuf_add_ref(db
, NULL
);
578 /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */
581 pbuf
= db
->db_parent
->db_buf
;
583 pbuf
= db
->db_objset
->os_phys_buf
;
585 (void) dsl_read(zio
, spa
, db
->db_blkptr
, pbuf
,
586 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
587 (*flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
589 if (aflags
& ARC_CACHED
)
590 *flags
|= DB_RF_CACHED
;
594 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
597 int havepzio
= (zio
!= NULL
);
602 * We don't have to hold the mutex to check db_state because it
603 * can't be freed while we have a hold on the buffer.
605 ASSERT(!refcount_is_zero(&db
->db_holds
));
607 if (db
->db_state
== DB_NOFILL
)
612 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
613 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
615 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
616 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
617 DBUF_IS_CACHEABLE(db
);
619 mutex_enter(&db
->db_mtx
);
620 if (db
->db_state
== DB_CACHED
) {
621 mutex_exit(&db
->db_mtx
);
623 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
624 db
->db
.db_size
, TRUE
);
625 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
626 rw_exit(&dn
->dn_struct_rwlock
);
628 } else if (db
->db_state
== DB_UNCACHED
) {
629 spa_t
*spa
= dn
->dn_objset
->os_spa
;
632 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
633 dbuf_read_impl(db
, zio
, &flags
);
635 /* dbuf_read_impl has dropped db_mtx for us */
638 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
639 db
->db
.db_size
, flags
& DB_RF_CACHED
);
641 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
642 rw_exit(&dn
->dn_struct_rwlock
);
648 mutex_exit(&db
->db_mtx
);
650 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
651 db
->db
.db_size
, TRUE
);
652 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
653 rw_exit(&dn
->dn_struct_rwlock
);
656 mutex_enter(&db
->db_mtx
);
657 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
658 while (db
->db_state
== DB_READ
||
659 db
->db_state
== DB_FILL
) {
660 ASSERT(db
->db_state
== DB_READ
||
661 (flags
& DB_RF_HAVESTRUCT
) == 0);
662 cv_wait(&db
->db_changed
, &db
->db_mtx
);
664 if (db
->db_state
== DB_UNCACHED
)
667 mutex_exit(&db
->db_mtx
);
670 ASSERT(err
|| havepzio
|| db
->db_state
== DB_CACHED
);
675 dbuf_noread(dmu_buf_impl_t
*db
)
677 ASSERT(!refcount_is_zero(&db
->db_holds
));
678 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
679 mutex_enter(&db
->db_mtx
);
680 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
681 cv_wait(&db
->db_changed
, &db
->db_mtx
);
682 if (db
->db_state
== DB_UNCACHED
) {
683 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
686 ASSERT(db
->db_buf
== NULL
);
687 ASSERT(db
->db
.db_data
== NULL
);
688 DB_GET_SPA(&spa
, db
);
689 dbuf_set_data(db
, arc_buf_alloc(spa
, db
->db
.db_size
, db
, type
));
690 db
->db_state
= DB_FILL
;
691 } else if (db
->db_state
== DB_NOFILL
) {
692 dbuf_set_data(db
, NULL
);
694 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
696 mutex_exit(&db
->db_mtx
);
700 * This is our just-in-time copy function. It makes a copy of
701 * buffers, that have been modified in a previous transaction
702 * group, before we modify them in the current active group.
704 * This function is used in two places: when we are dirtying a
705 * buffer for the first time in a txg, and when we are freeing
706 * a range in a dnode that includes this buffer.
708 * Note that when we are called from dbuf_free_range() we do
709 * not put a hold on the buffer, we just traverse the active
710 * dbuf list for the dnode.
713 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
715 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
717 ASSERT(MUTEX_HELD(&db
->db_mtx
));
718 ASSERT(db
->db
.db_data
!= NULL
);
719 ASSERT(db
->db_level
== 0);
720 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
723 (dr
->dt
.dl
.dr_data
!=
724 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
728 * If the last dirty record for this dbuf has not yet synced
729 * and its referencing the dbuf data, either:
730 * reset the reference to point to a new copy,
731 * or (if there a no active holders)
732 * just null out the current db_data pointer.
734 ASSERT(dr
->dr_txg
>= txg
- 2);
735 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
736 /* Note that the data bufs here are zio_bufs */
737 dr
->dt
.dl
.dr_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
738 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
739 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, DN_MAX_BONUSLEN
);
740 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
741 int size
= db
->db
.db_size
;
742 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
745 DB_GET_SPA(&spa
, db
);
746 dr
->dt
.dl
.dr_data
= arc_buf_alloc(spa
, size
, db
, type
);
747 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
749 dbuf_set_data(db
, NULL
);
754 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
756 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
757 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
758 uint64_t txg
= dr
->dr_txg
;
760 ASSERT(MUTEX_HELD(&db
->db_mtx
));
761 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
762 ASSERT(db
->db_level
== 0);
764 if (db
->db_blkid
== DMU_BONUS_BLKID
||
765 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
768 ASSERT(db
->db_data_pending
!= dr
);
770 /* free this block */
771 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
) {
774 DB_GET_SPA(&spa
, db
);
775 zio_free(spa
, txg
, bp
);
777 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
778 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
781 * Release the already-written buffer, so we leave it in
782 * a consistent dirty state. Note that all callers are
783 * modifying the buffer, so they will immediately do
784 * another (redundant) arc_release(). Therefore, leave
785 * the buf thawed to save the effort of freezing &
786 * immediately re-thawing it.
788 arc_release(dr
->dt
.dl
.dr_data
, db
);
792 * Evict (if its unreferenced) or clear (if its referenced) any level-0
793 * data blocks in the free range, so that any future readers will find
794 * empty blocks. Also, if we happen accross any level-1 dbufs in the
795 * range that have not already been marked dirty, mark them dirty so
796 * they stay in memory.
799 dbuf_free_range(dnode_t
*dn
, uint64_t start
, uint64_t end
, dmu_tx_t
*tx
)
801 dmu_buf_impl_t
*db
, *db_next
;
802 uint64_t txg
= tx
->tx_txg
;
803 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
804 uint64_t first_l1
= start
>> epbs
;
805 uint64_t last_l1
= end
>> epbs
;
807 if (end
> dn
->dn_maxblkid
&& (end
!= DMU_SPILL_BLKID
)) {
808 end
= dn
->dn_maxblkid
;
809 last_l1
= end
>> epbs
;
811 dprintf_dnode(dn
, "start=%llu end=%llu\n", start
, end
);
812 mutex_enter(&dn
->dn_dbufs_mtx
);
813 for (db
= list_head(&dn
->dn_dbufs
); db
; db
= db_next
) {
814 db_next
= list_next(&dn
->dn_dbufs
, db
);
815 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
817 if (db
->db_level
== 1 &&
818 db
->db_blkid
>= first_l1
&& db
->db_blkid
<= last_l1
) {
819 mutex_enter(&db
->db_mtx
);
820 if (db
->db_last_dirty
&&
821 db
->db_last_dirty
->dr_txg
< txg
) {
822 dbuf_add_ref(db
, FTAG
);
823 mutex_exit(&db
->db_mtx
);
824 dbuf_will_dirty(db
, tx
);
827 mutex_exit(&db
->db_mtx
);
831 if (db
->db_level
!= 0)
833 dprintf_dbuf(db
, "found buf %s\n", "");
834 if (db
->db_blkid
< start
|| db
->db_blkid
> end
)
837 /* found a level 0 buffer in the range */
838 if (dbuf_undirty(db
, tx
))
841 mutex_enter(&db
->db_mtx
);
842 if (db
->db_state
== DB_UNCACHED
||
843 db
->db_state
== DB_NOFILL
||
844 db
->db_state
== DB_EVICTING
) {
845 ASSERT(db
->db
.db_data
== NULL
);
846 mutex_exit(&db
->db_mtx
);
849 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
850 /* will be handled in dbuf_read_done or dbuf_rele */
851 db
->db_freed_in_flight
= TRUE
;
852 mutex_exit(&db
->db_mtx
);
855 if (refcount_count(&db
->db_holds
) == 0) {
860 /* The dbuf is referenced */
862 if (db
->db_last_dirty
!= NULL
) {
863 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
865 if (dr
->dr_txg
== txg
) {
867 * This buffer is "in-use", re-adjust the file
868 * size to reflect that this buffer may
869 * contain new data when we sync.
871 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
872 db
->db_blkid
> dn
->dn_maxblkid
)
873 dn
->dn_maxblkid
= db
->db_blkid
;
877 * This dbuf is not dirty in the open context.
878 * Either uncache it (if its not referenced in
879 * the open context) or reset its contents to
882 dbuf_fix_old_data(db
, txg
);
885 /* clear the contents if its cached */
886 if (db
->db_state
== DB_CACHED
) {
887 ASSERT(db
->db
.db_data
!= NULL
);
888 arc_release(db
->db_buf
, db
);
889 bzero(db
->db
.db_data
, db
->db
.db_size
);
890 arc_buf_freeze(db
->db_buf
);
893 mutex_exit(&db
->db_mtx
);
895 mutex_exit(&dn
->dn_dbufs_mtx
);
899 dbuf_block_freeable(dmu_buf_impl_t
*db
)
901 dsl_dataset_t
*ds
= db
->db_objset
->os_dsl_dataset
;
902 uint64_t birth_txg
= 0;
905 * We don't need any locking to protect db_blkptr:
906 * If it's syncing, then db_last_dirty will be set
907 * so we'll ignore db_blkptr.
909 ASSERT(MUTEX_HELD(&db
->db_mtx
));
910 if (db
->db_last_dirty
)
911 birth_txg
= db
->db_last_dirty
->dr_txg
;
912 else if (db
->db_blkptr
)
913 birth_txg
= db
->db_blkptr
->blk_birth
;
916 * If we don't exist or are in a snapshot, we can't be freed.
917 * Don't pass the bp to dsl_dataset_block_freeable() since we
918 * are holding the db_mtx lock and might deadlock if we are
919 * prefetching a dedup-ed block.
922 return (ds
== NULL
||
923 dsl_dataset_block_freeable(ds
, NULL
, birth_txg
));
929 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
931 arc_buf_t
*buf
, *obuf
;
932 int osize
= db
->db
.db_size
;
933 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
936 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
941 /* XXX does *this* func really need the lock? */
942 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
945 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
946 * is OK, because there can be no other references to the db
947 * when we are changing its size, so no concurrent DB_FILL can
951 * XXX we should be doing a dbuf_read, checking the return
952 * value and returning that up to our callers
954 dbuf_will_dirty(db
, tx
);
956 /* create the data buffer for the new block */
957 buf
= arc_buf_alloc(dn
->dn_objset
->os_spa
, size
, db
, type
);
959 /* copy old block data to the new block */
961 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
962 /* zero the remainder */
964 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
966 mutex_enter(&db
->db_mtx
);
967 dbuf_set_data(db
, buf
);
968 VERIFY(arc_buf_remove_ref(obuf
, db
) == 1);
969 db
->db
.db_size
= size
;
971 if (db
->db_level
== 0) {
972 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
973 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
975 mutex_exit(&db
->db_mtx
);
977 dnode_willuse_space(dn
, size
-osize
, tx
);
982 dbuf_release_bp(dmu_buf_impl_t
*db
)
987 DB_GET_OBJSET(&os
, db
);
988 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
989 ASSERT(arc_released(os
->os_phys_buf
) ||
990 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
991 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
993 zb
.zb_objset
= os
->os_dsl_dataset
?
994 os
->os_dsl_dataset
->ds_object
: 0;
995 zb
.zb_object
= db
->db
.db_object
;
996 zb
.zb_level
= db
->db_level
;
997 zb
.zb_blkid
= db
->db_blkid
;
998 (void) arc_release_bp(db
->db_buf
, db
,
999 db
->db_blkptr
, os
->os_spa
, &zb
);
1002 dbuf_dirty_record_t
*
1003 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1007 dbuf_dirty_record_t
**drp
, *dr
;
1008 int drop_struct_lock
= FALSE
;
1009 boolean_t do_free_accounting
= B_FALSE
;
1010 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1012 ASSERT(tx
->tx_txg
!= 0);
1013 ASSERT(!refcount_is_zero(&db
->db_holds
));
1014 DMU_TX_DIRTY_BUF(tx
, db
);
1019 * Shouldn't dirty a regular buffer in syncing context. Private
1020 * objects may be dirtied in syncing context, but only if they
1021 * were already pre-dirtied in open context.
1023 ASSERT(!dmu_tx_is_syncing(tx
) ||
1024 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1025 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1026 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1028 * We make this assert for private objects as well, but after we
1029 * check if we're already dirty. They are allowed to re-dirty
1030 * in syncing context.
1032 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1033 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1034 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1036 mutex_enter(&db
->db_mtx
);
1038 * XXX make this true for indirects too? The problem is that
1039 * transactions created with dmu_tx_create_assigned() from
1040 * syncing context don't bother holding ahead.
1042 ASSERT(db
->db_level
!= 0 ||
1043 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1044 db
->db_state
== DB_NOFILL
);
1046 mutex_enter(&dn
->dn_mtx
);
1048 * Don't set dirtyctx to SYNC if we're just modifying this as we
1049 * initialize the objset.
1051 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
&&
1052 !BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1054 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1055 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1056 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1058 mutex_exit(&dn
->dn_mtx
);
1060 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1061 dn
->dn_have_spill
= B_TRUE
;
1064 * If this buffer is already dirty, we're done.
1066 drp
= &db
->db_last_dirty
;
1067 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1068 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1069 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1071 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1074 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1076 * If this buffer has already been written out,
1077 * we now need to reset its state.
1079 dbuf_unoverride(dr
);
1080 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1081 db
->db_state
!= DB_NOFILL
)
1082 arc_buf_thaw(db
->db_buf
);
1084 mutex_exit(&db
->db_mtx
);
1089 * Only valid if not already dirty.
1091 ASSERT(dn
->dn_object
== 0 ||
1092 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1093 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1095 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1096 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1097 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1098 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1099 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1100 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1103 * We should only be dirtying in syncing context if it's the
1104 * mos or we're initializing the os or it's a special object.
1105 * However, we are allowed to dirty in syncing context provided
1106 * we already dirtied it in open context. Hence we must make
1107 * this assertion only if we're not already dirty.
1110 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1111 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1112 ASSERT(db
->db
.db_size
!= 0);
1114 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1116 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1118 * Update the accounting.
1119 * Note: we delay "free accounting" until after we drop
1120 * the db_mtx. This keeps us from grabbing other locks
1121 * (and possibly deadlocking) in bp_get_dsize() while
1122 * also holding the db_mtx.
1124 dnode_willuse_space(dn
, db
->db
.db_size
, tx
);
1125 do_free_accounting
= dbuf_block_freeable(db
);
1129 * If this buffer is dirty in an old transaction group we need
1130 * to make a copy of it so that the changes we make in this
1131 * transaction group won't leak out when we sync the older txg.
1133 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1134 if (db
->db_level
== 0) {
1135 void *data_old
= db
->db_buf
;
1137 if (db
->db_state
!= DB_NOFILL
) {
1138 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1139 dbuf_fix_old_data(db
, tx
->tx_txg
);
1140 data_old
= db
->db
.db_data
;
1141 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1143 * Release the data buffer from the cache so
1144 * that we can modify it without impacting
1145 * possible other users of this cached data
1146 * block. Note that indirect blocks and
1147 * private objects are not released until the
1148 * syncing state (since they are only modified
1151 arc_release(db
->db_buf
, db
);
1152 dbuf_fix_old_data(db
, tx
->tx_txg
);
1153 data_old
= db
->db_buf
;
1155 ASSERT(data_old
!= NULL
);
1157 dr
->dt
.dl
.dr_data
= data_old
;
1159 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
1160 list_create(&dr
->dt
.di
.dr_children
,
1161 sizeof (dbuf_dirty_record_t
),
1162 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1165 dr
->dr_txg
= tx
->tx_txg
;
1170 * We could have been freed_in_flight between the dbuf_noread
1171 * and dbuf_dirty. We win, as though the dbuf_noread() had
1172 * happened after the free.
1174 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1175 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1176 mutex_enter(&dn
->dn_mtx
);
1177 dnode_clear_range(dn
, db
->db_blkid
, 1, tx
);
1178 mutex_exit(&dn
->dn_mtx
);
1179 db
->db_freed_in_flight
= FALSE
;
1183 * This buffer is now part of this txg
1185 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1186 db
->db_dirtycnt
+= 1;
1187 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1189 mutex_exit(&db
->db_mtx
);
1191 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1192 db
->db_blkid
== DMU_SPILL_BLKID
) {
1193 mutex_enter(&dn
->dn_mtx
);
1194 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1195 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1196 mutex_exit(&dn
->dn_mtx
);
1197 dnode_setdirty(dn
, tx
);
1200 } else if (do_free_accounting
) {
1201 blkptr_t
*bp
= db
->db_blkptr
;
1202 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1203 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1205 * This is only a guess -- if the dbuf is dirty
1206 * in a previous txg, we don't know how much
1207 * space it will use on disk yet. We should
1208 * really have the struct_rwlock to access
1209 * db_blkptr, but since this is just a guess,
1210 * it's OK if we get an odd answer.
1212 ddt_prefetch(os
->os_spa
, bp
);
1213 dnode_willuse_space(dn
, -willfree
, tx
);
1216 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1217 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1218 drop_struct_lock
= TRUE
;
1221 if (db
->db_level
== 0) {
1222 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1223 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1226 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1227 dmu_buf_impl_t
*parent
= db
->db_parent
;
1228 dbuf_dirty_record_t
*di
;
1229 int parent_held
= FALSE
;
1231 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1232 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1234 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1235 db
->db_blkid
>> epbs
, FTAG
);
1236 ASSERT(parent
!= NULL
);
1239 if (drop_struct_lock
)
1240 rw_exit(&dn
->dn_struct_rwlock
);
1241 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1242 di
= dbuf_dirty(parent
, tx
);
1244 dbuf_rele(parent
, FTAG
);
1246 mutex_enter(&db
->db_mtx
);
1247 /* possible race with dbuf_undirty() */
1248 if (db
->db_last_dirty
== dr
||
1249 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1250 mutex_enter(&di
->dt
.di
.dr_mtx
);
1251 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1252 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1253 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1254 mutex_exit(&di
->dt
.di
.dr_mtx
);
1257 mutex_exit(&db
->db_mtx
);
1259 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1260 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1261 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1262 mutex_enter(&dn
->dn_mtx
);
1263 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1264 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1265 mutex_exit(&dn
->dn_mtx
);
1266 if (drop_struct_lock
)
1267 rw_exit(&dn
->dn_struct_rwlock
);
1270 dnode_setdirty(dn
, tx
);
1276 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1279 uint64_t txg
= tx
->tx_txg
;
1280 dbuf_dirty_record_t
*dr
, **drp
;
1283 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1285 mutex_enter(&db
->db_mtx
);
1287 * If this buffer is not dirty, we're done.
1289 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1290 if (dr
->dr_txg
<= txg
)
1292 if (dr
== NULL
|| dr
->dr_txg
< txg
) {
1293 mutex_exit(&db
->db_mtx
);
1296 ASSERT(dr
->dr_txg
== txg
);
1297 ASSERT(dr
->dr_dbuf
== db
);
1303 * If this buffer is currently held, we cannot undirty
1304 * it, since one of the current holders may be in the
1305 * middle of an update. Note that users of dbuf_undirty()
1306 * should not place a hold on the dbuf before the call.
1307 * Also note: we can get here with a spill block, so
1308 * test for that similar to how dbuf_dirty does.
1310 if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1311 mutex_exit(&db
->db_mtx
);
1312 /* Make sure we don't toss this buffer at sync phase */
1313 if (db
->db_blkid
!= DMU_SPILL_BLKID
) {
1314 mutex_enter(&dn
->dn_mtx
);
1315 dnode_clear_range(dn
, db
->db_blkid
, 1, tx
);
1316 mutex_exit(&dn
->dn_mtx
);
1322 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1324 ASSERT(db
->db
.db_size
!= 0);
1326 /* XXX would be nice to fix up dn_towrite_space[] */
1331 * Note that there are three places in dbuf_dirty()
1332 * where this dirty record may be put on a list.
1333 * Make sure to do a list_remove corresponding to
1334 * every one of those list_insert calls.
1336 if (dr
->dr_parent
) {
1337 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1338 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1339 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1340 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1341 db
->db_level
+1 == dn
->dn_nlevels
) {
1342 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1343 mutex_enter(&dn
->dn_mtx
);
1344 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1345 mutex_exit(&dn
->dn_mtx
);
1349 if (db
->db_level
== 0) {
1350 if (db
->db_state
!= DB_NOFILL
) {
1351 dbuf_unoverride(dr
);
1353 ASSERT(db
->db_buf
!= NULL
);
1354 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1355 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1356 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
1360 ASSERT(db
->db_buf
!= NULL
);
1361 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
1362 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
1363 list_destroy(&dr
->dt
.di
.dr_children
);
1365 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1367 ASSERT(db
->db_dirtycnt
> 0);
1368 db
->db_dirtycnt
-= 1;
1370 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1371 arc_buf_t
*buf
= db
->db_buf
;
1373 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(buf
));
1374 dbuf_set_data(db
, NULL
);
1375 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
1380 mutex_exit(&db
->db_mtx
);
1384 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1386 dbuf_will_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1388 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1390 ASSERT(tx
->tx_txg
!= 0);
1391 ASSERT(!refcount_is_zero(&db
->db_holds
));
1394 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1395 rf
|= DB_RF_HAVESTRUCT
;
1397 (void) dbuf_read(db
, NULL
, rf
);
1398 (void) dbuf_dirty(db
, tx
);
1402 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1404 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1406 db
->db_state
= DB_NOFILL
;
1408 dmu_buf_will_fill(db_fake
, tx
);
1412 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1414 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1416 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1417 ASSERT(tx
->tx_txg
!= 0);
1418 ASSERT(db
->db_level
== 0);
1419 ASSERT(!refcount_is_zero(&db
->db_holds
));
1421 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1422 dmu_tx_private_ok(tx
));
1425 (void) dbuf_dirty(db
, tx
);
1428 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1431 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1433 mutex_enter(&db
->db_mtx
);
1436 if (db
->db_state
== DB_FILL
) {
1437 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1438 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1439 /* we were freed while filling */
1440 /* XXX dbuf_undirty? */
1441 bzero(db
->db
.db_data
, db
->db
.db_size
);
1442 db
->db_freed_in_flight
= FALSE
;
1444 db
->db_state
= DB_CACHED
;
1445 cv_broadcast(&db
->db_changed
);
1447 mutex_exit(&db
->db_mtx
);
1451 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1452 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1455 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1457 ASSERT(!refcount_is_zero(&db
->db_holds
));
1458 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1459 ASSERT(db
->db_level
== 0);
1460 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1461 ASSERT(buf
!= NULL
);
1462 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1463 ASSERT(tx
->tx_txg
!= 0);
1465 arc_return_buf(buf
, db
);
1466 ASSERT(arc_released(buf
));
1468 mutex_enter(&db
->db_mtx
);
1470 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1471 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1473 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1475 if (db
->db_state
== DB_CACHED
&&
1476 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1477 mutex_exit(&db
->db_mtx
);
1478 (void) dbuf_dirty(db
, tx
);
1479 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1480 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
1481 xuio_stat_wbuf_copied();
1485 xuio_stat_wbuf_nocopy();
1486 if (db
->db_state
== DB_CACHED
) {
1487 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1489 ASSERT(db
->db_buf
!= NULL
);
1490 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
1491 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
1492 if (!arc_released(db
->db_buf
)) {
1493 ASSERT(dr
->dt
.dl
.dr_override_state
==
1495 arc_release(db
->db_buf
, db
);
1497 dr
->dt
.dl
.dr_data
= buf
;
1498 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
) == 1);
1499 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
1500 arc_release(db
->db_buf
, db
);
1501 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
) == 1);
1505 ASSERT(db
->db_buf
== NULL
);
1506 dbuf_set_data(db
, buf
);
1507 db
->db_state
= DB_FILL
;
1508 mutex_exit(&db
->db_mtx
);
1509 (void) dbuf_dirty(db
, tx
);
1510 dbuf_fill_done(db
, tx
);
1514 * "Clear" the contents of this dbuf. This will mark the dbuf
1515 * EVICTING and clear *most* of its references. Unfortunetely,
1516 * when we are not holding the dn_dbufs_mtx, we can't clear the
1517 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1518 * in this case. For callers from the DMU we will usually see:
1519 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1520 * For the arc callback, we will usually see:
1521 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1522 * Sometimes, though, we will get a mix of these two:
1523 * DMU: dbuf_clear()->arc_buf_evict()
1524 * ARC: dbuf_do_evict()->dbuf_destroy()
1527 dbuf_clear(dmu_buf_impl_t
*db
)
1530 dmu_buf_impl_t
*parent
= db
->db_parent
;
1531 dmu_buf_impl_t
*dndb
;
1532 int dbuf_gone
= FALSE
;
1534 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1535 ASSERT(refcount_is_zero(&db
->db_holds
));
1537 dbuf_evict_user(db
);
1539 if (db
->db_state
== DB_CACHED
) {
1540 ASSERT(db
->db
.db_data
!= NULL
);
1541 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1542 zio_buf_free(db
->db
.db_data
, DN_MAX_BONUSLEN
);
1543 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1545 db
->db
.db_data
= NULL
;
1546 db
->db_state
= DB_UNCACHED
;
1549 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1550 ASSERT(db
->db_data_pending
== NULL
);
1552 db
->db_state
= DB_EVICTING
;
1553 db
->db_blkptr
= NULL
;
1558 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& MUTEX_HELD(&dn
->dn_dbufs_mtx
)) {
1559 list_remove(&dn
->dn_dbufs
, db
);
1560 (void) atomic_dec_32_nv(&dn
->dn_dbufs_count
);
1564 * Decrementing the dbuf count means that the hold corresponding
1565 * to the removed dbuf is no longer discounted in dnode_move(),
1566 * so the dnode cannot be moved until after we release the hold.
1567 * The membar_producer() ensures visibility of the decremented
1568 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1572 db
->db_dnode_handle
= NULL
;
1578 dbuf_gone
= arc_buf_evict(db
->db_buf
);
1581 mutex_exit(&db
->db_mtx
);
1584 * If this dbuf is referenced from an indirect dbuf,
1585 * decrement the ref count on the indirect dbuf.
1587 if (parent
&& parent
!= dndb
)
1588 dbuf_rele(parent
, db
);
1592 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
1593 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
)
1600 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1602 if (blkid
== DMU_SPILL_BLKID
) {
1603 mutex_enter(&dn
->dn_mtx
);
1604 if (dn
->dn_have_spill
&&
1605 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1606 *bpp
= &dn
->dn_phys
->dn_spill
;
1609 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1610 *parentp
= dn
->dn_dbuf
;
1611 mutex_exit(&dn
->dn_mtx
);
1615 if (dn
->dn_phys
->dn_nlevels
== 0)
1618 nlevels
= dn
->dn_phys
->dn_nlevels
;
1620 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1622 ASSERT3U(level
* epbs
, <, 64);
1623 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1624 if (level
>= nlevels
||
1625 (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
1626 /* the buffer has no parent yet */
1628 } else if (level
< nlevels
-1) {
1629 /* this block is referenced from an indirect block */
1630 int err
= dbuf_hold_impl(dn
, level
+1,
1631 blkid
>> epbs
, fail_sparse
, NULL
, parentp
);
1634 err
= dbuf_read(*parentp
, NULL
,
1635 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
1637 dbuf_rele(*parentp
, NULL
);
1641 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
1642 (blkid
& ((1ULL << epbs
) - 1));
1645 /* the block is referenced from the dnode */
1646 ASSERT3U(level
, ==, nlevels
-1);
1647 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
1648 blkid
< dn
->dn_phys
->dn_nblkptr
);
1650 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1651 *parentp
= dn
->dn_dbuf
;
1653 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
1658 static dmu_buf_impl_t
*
1659 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
1660 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
1662 objset_t
*os
= dn
->dn_objset
;
1663 dmu_buf_impl_t
*db
, *odb
;
1665 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1666 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
1668 db
= kmem_cache_alloc(dbuf_cache
, KM_SLEEP
);
1671 db
->db
.db_object
= dn
->dn_object
;
1672 db
->db_level
= level
;
1673 db
->db_blkid
= blkid
;
1674 db
->db_last_dirty
= NULL
;
1675 db
->db_dirtycnt
= 0;
1676 db
->db_dnode_handle
= dn
->dn_handle
;
1677 db
->db_parent
= parent
;
1678 db
->db_blkptr
= blkptr
;
1680 db
->db_user_ptr
= NULL
;
1681 db
->db_user_data_ptr_ptr
= NULL
;
1682 db
->db_evict_func
= NULL
;
1683 db
->db_immediate_evict
= 0;
1684 db
->db_freed_in_flight
= 0;
1686 if (blkid
== DMU_BONUS_BLKID
) {
1687 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
1688 db
->db
.db_size
= DN_MAX_BONUSLEN
-
1689 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
1690 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1691 db
->db
.db_offset
= DMU_BONUS_BLKID
;
1692 db
->db_state
= DB_UNCACHED
;
1693 /* the bonus dbuf is not placed in the hash table */
1694 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1696 } else if (blkid
== DMU_SPILL_BLKID
) {
1697 db
->db
.db_size
= (blkptr
!= NULL
) ?
1698 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
1699 db
->db
.db_offset
= 0;
1702 db
->db_level
? 1<<dn
->dn_indblkshift
: dn
->dn_datablksz
;
1703 db
->db
.db_size
= blocksize
;
1704 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
1708 * Hold the dn_dbufs_mtx while we get the new dbuf
1709 * in the hash table *and* added to the dbufs list.
1710 * This prevents a possible deadlock with someone
1711 * trying to look up this dbuf before its added to the
1714 mutex_enter(&dn
->dn_dbufs_mtx
);
1715 db
->db_state
= DB_EVICTING
;
1716 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
1717 /* someone else inserted it first */
1718 kmem_cache_free(dbuf_cache
, db
);
1719 mutex_exit(&dn
->dn_dbufs_mtx
);
1722 list_insert_head(&dn
->dn_dbufs
, db
);
1723 db
->db_state
= DB_UNCACHED
;
1724 mutex_exit(&dn
->dn_dbufs_mtx
);
1725 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1727 if (parent
&& parent
!= dn
->dn_dbuf
)
1728 dbuf_add_ref(parent
, db
);
1730 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1731 refcount_count(&dn
->dn_holds
) > 0);
1732 (void) refcount_add(&dn
->dn_holds
, db
);
1733 (void) atomic_inc_32_nv(&dn
->dn_dbufs_count
);
1735 dprintf_dbuf(db
, "db=%p\n", db
);
1741 dbuf_do_evict(void *private)
1743 arc_buf_t
*buf
= private;
1744 dmu_buf_impl_t
*db
= buf
->b_private
;
1746 if (!MUTEX_HELD(&db
->db_mtx
))
1747 mutex_enter(&db
->db_mtx
);
1749 ASSERT(refcount_is_zero(&db
->db_holds
));
1751 if (db
->db_state
!= DB_EVICTING
) {
1752 ASSERT(db
->db_state
== DB_CACHED
);
1757 mutex_exit(&db
->db_mtx
);
1764 dbuf_destroy(dmu_buf_impl_t
*db
)
1766 ASSERT(refcount_is_zero(&db
->db_holds
));
1768 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1770 * If this dbuf is still on the dn_dbufs list,
1771 * remove it from that list.
1773 if (db
->db_dnode_handle
!= NULL
) {
1778 mutex_enter(&dn
->dn_dbufs_mtx
);
1779 list_remove(&dn
->dn_dbufs
, db
);
1780 (void) atomic_dec_32_nv(&dn
->dn_dbufs_count
);
1781 mutex_exit(&dn
->dn_dbufs_mtx
);
1784 * Decrementing the dbuf count means that the hold
1785 * corresponding to the removed dbuf is no longer
1786 * discounted in dnode_move(), so the dnode cannot be
1787 * moved until after we release the hold.
1790 db
->db_dnode_handle
= NULL
;
1792 dbuf_hash_remove(db
);
1794 db
->db_parent
= NULL
;
1797 ASSERT(!list_link_active(&db
->db_link
));
1798 ASSERT(db
->db
.db_data
== NULL
);
1799 ASSERT(db
->db_hash_next
== NULL
);
1800 ASSERT(db
->db_blkptr
== NULL
);
1801 ASSERT(db
->db_data_pending
== NULL
);
1803 kmem_cache_free(dbuf_cache
, db
);
1804 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1808 dbuf_prefetch(dnode_t
*dn
, uint64_t blkid
)
1810 dmu_buf_impl_t
*db
= NULL
;
1811 blkptr_t
*bp
= NULL
;
1813 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1814 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1816 if (dnode_block_freed(dn
, blkid
))
1819 /* dbuf_find() returns with db_mtx held */
1820 if (db
= dbuf_find(dn
, 0, blkid
)) {
1822 * This dbuf is already in the cache. We assume that
1823 * it is already CACHED, or else about to be either
1826 mutex_exit(&db
->db_mtx
);
1830 if (dbuf_findbp(dn
, 0, blkid
, TRUE
, &db
, &bp
) == 0) {
1831 if (bp
&& !BP_IS_HOLE(bp
)) {
1832 int priority
= dn
->dn_type
== DMU_OT_DDT_ZAP
?
1833 ZIO_PRIORITY_DDT_PREFETCH
: ZIO_PRIORITY_ASYNC_READ
;
1835 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
1836 uint32_t aflags
= ARC_NOWAIT
| ARC_PREFETCH
;
1839 SET_BOOKMARK(&zb
, ds
? ds
->ds_object
: DMU_META_OBJSET
,
1840 dn
->dn_object
, 0, blkid
);
1845 pbuf
= dn
->dn_objset
->os_phys_buf
;
1847 (void) dsl_read(NULL
, dn
->dn_objset
->os_spa
,
1848 bp
, pbuf
, NULL
, NULL
, priority
,
1849 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
1853 dbuf_rele(db
, NULL
);
1858 * Returns with db_holds incremented, and db_mtx not held.
1859 * Note: dn_struct_rwlock must be held.
1862 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
, int fail_sparse
,
1863 void *tag
, dmu_buf_impl_t
**dbp
)
1865 dmu_buf_impl_t
*db
, *parent
= NULL
;
1867 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1868 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1869 ASSERT3U(dn
->dn_nlevels
, >, level
);
1873 /* dbuf_find() returns with db_mtx held */
1874 db
= dbuf_find(dn
, level
, blkid
);
1877 blkptr_t
*bp
= NULL
;
1880 ASSERT3P(parent
, ==, NULL
);
1881 err
= dbuf_findbp(dn
, level
, blkid
, fail_sparse
, &parent
, &bp
);
1883 if (err
== 0 && bp
&& BP_IS_HOLE(bp
))
1887 dbuf_rele(parent
, NULL
);
1891 if (err
&& err
!= ENOENT
)
1893 db
= dbuf_create(dn
, level
, blkid
, parent
, bp
);
1896 if (db
->db_buf
&& refcount_is_zero(&db
->db_holds
)) {
1897 arc_buf_add_ref(db
->db_buf
, db
);
1898 if (db
->db_buf
->b_data
== NULL
) {
1901 dbuf_rele(parent
, NULL
);
1906 ASSERT3P(db
->db
.db_data
, ==, db
->db_buf
->b_data
);
1909 ASSERT(db
->db_buf
== NULL
|| arc_referenced(db
->db_buf
));
1912 * If this buffer is currently syncing out, and we are are
1913 * still referencing it from db_data, we need to make a copy
1914 * of it in case we decide we want to dirty it again in this txg.
1916 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1917 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
1918 db
->db_state
== DB_CACHED
&& db
->db_data_pending
) {
1919 dbuf_dirty_record_t
*dr
= db
->db_data_pending
;
1921 if (dr
->dt
.dl
.dr_data
== db
->db_buf
) {
1922 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1925 arc_buf_alloc(dn
->dn_objset
->os_spa
,
1926 db
->db
.db_size
, db
, type
));
1927 bcopy(dr
->dt
.dl
.dr_data
->b_data
, db
->db
.db_data
,
1932 (void) refcount_add(&db
->db_holds
, tag
);
1933 dbuf_update_data(db
);
1935 mutex_exit(&db
->db_mtx
);
1937 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1939 dbuf_rele(parent
, NULL
);
1941 ASSERT3P(DB_DNODE(db
), ==, dn
);
1942 ASSERT3U(db
->db_blkid
, ==, blkid
);
1943 ASSERT3U(db
->db_level
, ==, level
);
1950 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
1953 int err
= dbuf_hold_impl(dn
, 0, blkid
, FALSE
, tag
, &db
);
1954 return (err
? NULL
: db
);
1958 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
1961 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, tag
, &db
);
1962 return (err
? NULL
: db
);
1966 dbuf_create_bonus(dnode_t
*dn
)
1968 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1970 ASSERT(dn
->dn_bonus
== NULL
);
1971 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
1975 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
1977 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1980 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
1983 blksz
= SPA_MINBLOCKSIZE
;
1984 if (blksz
> SPA_MAXBLOCKSIZE
)
1985 blksz
= SPA_MAXBLOCKSIZE
;
1987 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
1991 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1992 dbuf_new_size(db
, blksz
, tx
);
1993 rw_exit(&dn
->dn_struct_rwlock
);
2000 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2002 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2005 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2007 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2009 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
2014 * If you call dbuf_rele() you had better not be referencing the dnode handle
2015 * unless you have some other direct or indirect hold on the dnode. (An indirect
2016 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2017 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2018 * dnode's parent dbuf evicting its dnode handles.
2020 #pragma weak dmu_buf_rele = dbuf_rele
2022 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2024 mutex_enter(&db
->db_mtx
);
2025 dbuf_rele_and_unlock(db
, tag
);
2029 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2030 * db_dirtycnt and db_holds to be updated atomically.
2033 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2037 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2041 * Remove the reference to the dbuf before removing its hold on the
2042 * dnode so we can guarantee in dnode_move() that a referenced bonus
2043 * buffer has a corresponding dnode hold.
2045 holds
= refcount_remove(&db
->db_holds
, tag
);
2049 * We can't freeze indirects if there is a possibility that they
2050 * may be modified in the current syncing context.
2052 if (db
->db_buf
&& holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0))
2053 arc_buf_freeze(db
->db_buf
);
2055 if (holds
== db
->db_dirtycnt
&&
2056 db
->db_level
== 0 && db
->db_immediate_evict
)
2057 dbuf_evict_user(db
);
2060 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2061 mutex_exit(&db
->db_mtx
);
2064 * If the dnode moves here, we cannot cross this barrier
2065 * until the move completes.
2068 (void) atomic_dec_32_nv(&DB_DNODE(db
)->dn_dbufs_count
);
2071 * The bonus buffer's dnode hold is no longer discounted
2072 * in dnode_move(). The dnode cannot move until after
2075 dnode_rele(DB_DNODE(db
), db
);
2076 } else if (db
->db_buf
== NULL
) {
2078 * This is a special case: we never associated this
2079 * dbuf with any data allocated from the ARC.
2081 ASSERT(db
->db_state
== DB_UNCACHED
||
2082 db
->db_state
== DB_NOFILL
);
2084 } else if (arc_released(db
->db_buf
)) {
2085 arc_buf_t
*buf
= db
->db_buf
;
2087 * This dbuf has anonymous data associated with it.
2089 dbuf_set_data(db
, NULL
);
2090 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
2093 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
) == 0);
2096 * A dbuf will be eligible for eviction if either the
2097 * 'primarycache' property is set or a duplicate
2098 * copy of this buffer is already cached in the arc.
2100 * In the case of the 'primarycache' a buffer
2101 * is considered for eviction if it matches the
2102 * criteria set in the property.
2104 * To decide if our buffer is considered a
2105 * duplicate, we must call into the arc to determine
2106 * if multiple buffers are referencing the same
2107 * block on-disk. If so, then we simply evict
2110 if (!DBUF_IS_CACHEABLE(db
) ||
2111 arc_buf_eviction_needed(db
->db_buf
))
2114 mutex_exit(&db
->db_mtx
);
2117 mutex_exit(&db
->db_mtx
);
2121 #pragma weak dmu_buf_refcount = dbuf_refcount
2123 dbuf_refcount(dmu_buf_impl_t
*db
)
2125 return (refcount_count(&db
->db_holds
));
2129 dmu_buf_set_user(dmu_buf_t
*db_fake
, void *user_ptr
, void *user_data_ptr_ptr
,
2130 dmu_buf_evict_func_t
*evict_func
)
2132 return (dmu_buf_update_user(db_fake
, NULL
, user_ptr
,
2133 user_data_ptr_ptr
, evict_func
));
2137 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, void *user_ptr
, void *user_data_ptr_ptr
,
2138 dmu_buf_evict_func_t
*evict_func
)
2140 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2142 db
->db_immediate_evict
= TRUE
;
2143 return (dmu_buf_update_user(db_fake
, NULL
, user_ptr
,
2144 user_data_ptr_ptr
, evict_func
));
2148 dmu_buf_update_user(dmu_buf_t
*db_fake
, void *old_user_ptr
, void *user_ptr
,
2149 void *user_data_ptr_ptr
, dmu_buf_evict_func_t
*evict_func
)
2151 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2152 ASSERT(db
->db_level
== 0);
2154 ASSERT((user_ptr
== NULL
) == (evict_func
== NULL
));
2156 mutex_enter(&db
->db_mtx
);
2158 if (db
->db_user_ptr
== old_user_ptr
) {
2159 db
->db_user_ptr
= user_ptr
;
2160 db
->db_user_data_ptr_ptr
= user_data_ptr_ptr
;
2161 db
->db_evict_func
= evict_func
;
2163 dbuf_update_data(db
);
2165 old_user_ptr
= db
->db_user_ptr
;
2168 mutex_exit(&db
->db_mtx
);
2169 return (old_user_ptr
);
2173 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2175 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2176 ASSERT(!refcount_is_zero(&db
->db_holds
));
2178 return (db
->db_user_ptr
);
2182 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2184 boolean_t res
= B_FALSE
;
2185 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2188 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2189 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2195 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2197 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2198 return (dbi
->db_blkptr
);
2202 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2204 /* ASSERT(dmu_tx_is_syncing(tx) */
2205 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2207 if (db
->db_blkptr
!= NULL
)
2210 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2211 db
->db_blkptr
= &dn
->dn_phys
->dn_spill
;
2212 BP_ZERO(db
->db_blkptr
);
2215 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2217 * This buffer was allocated at a time when there was
2218 * no available blkptrs from the dnode, or it was
2219 * inappropriate to hook it in (i.e., nlevels mis-match).
2221 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2222 ASSERT(db
->db_parent
== NULL
);
2223 db
->db_parent
= dn
->dn_dbuf
;
2224 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2227 dmu_buf_impl_t
*parent
= db
->db_parent
;
2228 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2230 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2231 if (parent
== NULL
) {
2232 mutex_exit(&db
->db_mtx
);
2233 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2234 (void) dbuf_hold_impl(dn
, db
->db_level
+1,
2235 db
->db_blkid
>> epbs
, FALSE
, db
, &parent
);
2236 rw_exit(&dn
->dn_struct_rwlock
);
2237 mutex_enter(&db
->db_mtx
);
2238 db
->db_parent
= parent
;
2240 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2241 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2247 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2249 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2253 ASSERT(dmu_tx_is_syncing(tx
));
2255 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2257 mutex_enter(&db
->db_mtx
);
2259 ASSERT(db
->db_level
> 0);
2262 if (db
->db_buf
== NULL
) {
2263 mutex_exit(&db
->db_mtx
);
2264 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2265 mutex_enter(&db
->db_mtx
);
2267 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2268 ASSERT(db
->db_buf
!= NULL
);
2272 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2273 dbuf_check_blkptr(dn
, db
);
2276 db
->db_data_pending
= dr
;
2278 mutex_exit(&db
->db_mtx
);
2279 dbuf_write(dr
, db
->db_buf
, tx
);
2282 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2283 dbuf_sync_list(&dr
->dt
.di
.dr_children
, tx
);
2284 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2285 mutex_exit(&dr
->dt
.di
.dr_mtx
);
2290 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2292 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
2293 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2296 uint64_t txg
= tx
->tx_txg
;
2298 ASSERT(dmu_tx_is_syncing(tx
));
2300 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2302 mutex_enter(&db
->db_mtx
);
2304 * To be synced, we must be dirtied. But we
2305 * might have been freed after the dirty.
2307 if (db
->db_state
== DB_UNCACHED
) {
2308 /* This buffer has been freed since it was dirtied */
2309 ASSERT(db
->db
.db_data
== NULL
);
2310 } else if (db
->db_state
== DB_FILL
) {
2311 /* This buffer was freed and is now being re-filled */
2312 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
2314 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
2321 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2322 mutex_enter(&dn
->dn_mtx
);
2323 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
2324 mutex_exit(&dn
->dn_mtx
);
2328 * If this is a bonus buffer, simply copy the bonus data into the
2329 * dnode. It will be written out when the dnode is synced (and it
2330 * will be synced, since it must have been dirty for dbuf_sync to
2333 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2334 dbuf_dirty_record_t
**drp
;
2336 ASSERT(*datap
!= NULL
);
2337 ASSERT0(db
->db_level
);
2338 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
2339 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
2342 if (*datap
!= db
->db
.db_data
) {
2343 zio_buf_free(*datap
, DN_MAX_BONUSLEN
);
2344 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
2346 db
->db_data_pending
= NULL
;
2347 drp
= &db
->db_last_dirty
;
2349 drp
= &(*drp
)->dr_next
;
2350 ASSERT(dr
->dr_next
== NULL
);
2351 ASSERT(dr
->dr_dbuf
== db
);
2353 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2354 ASSERT(db
->db_dirtycnt
> 0);
2355 db
->db_dirtycnt
-= 1;
2356 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
2363 * This function may have dropped the db_mtx lock allowing a dmu_sync
2364 * operation to sneak in. As a result, we need to ensure that we
2365 * don't check the dr_override_state until we have returned from
2366 * dbuf_check_blkptr.
2368 dbuf_check_blkptr(dn
, db
);
2371 * If this buffer is in the middle of an immediate write,
2372 * wait for the synchronous IO to complete.
2374 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
2375 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
2376 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2377 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
2380 if (db
->db_state
!= DB_NOFILL
&&
2381 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2382 refcount_count(&db
->db_holds
) > 1 &&
2383 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
2384 *datap
== db
->db_buf
) {
2386 * If this buffer is currently "in use" (i.e., there
2387 * are active holds and db_data still references it),
2388 * then make a copy before we start the write so that
2389 * any modifications from the open txg will not leak
2392 * NOTE: this copy does not need to be made for
2393 * objects only modified in the syncing context (e.g.
2394 * DNONE_DNODE blocks).
2396 int blksz
= arc_buf_size(*datap
);
2397 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2398 *datap
= arc_buf_alloc(os
->os_spa
, blksz
, db
, type
);
2399 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
2401 db
->db_data_pending
= dr
;
2403 mutex_exit(&db
->db_mtx
);
2405 dbuf_write(dr
, *datap
, tx
);
2407 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2408 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2409 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
2413 * Although zio_nowait() does not "wait for an IO", it does
2414 * initiate the IO. If this is an empty write it seems plausible
2415 * that the IO could actually be completed before the nowait
2416 * returns. We need to DB_DNODE_EXIT() first in case
2417 * zio_nowait() invalidates the dbuf.
2420 zio_nowait(dr
->dr_zio
);
2425 dbuf_sync_list(list_t
*list
, dmu_tx_t
*tx
)
2427 dbuf_dirty_record_t
*dr
;
2429 while (dr
= list_head(list
)) {
2430 if (dr
->dr_zio
!= NULL
) {
2432 * If we find an already initialized zio then we
2433 * are processing the meta-dnode, and we have finished.
2434 * The dbufs for all dnodes are put back on the list
2435 * during processing, so that we can zio_wait()
2436 * these IOs after initiating all child IOs.
2438 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
2439 DMU_META_DNODE_OBJECT
);
2442 list_remove(list
, dr
);
2443 if (dr
->dr_dbuf
->db_level
> 0)
2444 dbuf_sync_indirect(dr
, tx
);
2446 dbuf_sync_leaf(dr
, tx
);
2452 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
2454 dmu_buf_impl_t
*db
= vdb
;
2456 blkptr_t
*bp
= zio
->io_bp
;
2457 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2458 spa_t
*spa
= zio
->io_spa
;
2463 ASSERT(db
->db_blkptr
== bp
);
2467 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
2468 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
2469 zio
->io_prev_space_delta
= delta
;
2471 if (BP_IS_HOLE(bp
)) {
2472 ASSERT(bp
->blk_fill
== 0);
2477 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
2478 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
2479 (db
->db_blkid
== DMU_SPILL_BLKID
&&
2480 BP_GET_TYPE(bp
) == dn
->dn_bonustype
));
2481 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
2483 mutex_enter(&db
->db_mtx
);
2486 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2487 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
2488 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
2489 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
2493 if (db
->db_level
== 0) {
2494 mutex_enter(&dn
->dn_mtx
);
2495 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
2496 db
->db_blkid
!= DMU_SPILL_BLKID
)
2497 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
2498 mutex_exit(&dn
->dn_mtx
);
2500 if (dn
->dn_type
== DMU_OT_DNODE
) {
2501 dnode_phys_t
*dnp
= db
->db
.db_data
;
2502 for (i
= db
->db
.db_size
>> DNODE_SHIFT
; i
> 0;
2504 if (dnp
->dn_type
!= DMU_OT_NONE
)
2511 blkptr_t
*ibp
= db
->db
.db_data
;
2512 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2513 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
2514 if (BP_IS_HOLE(ibp
))
2516 fill
+= ibp
->blk_fill
;
2521 bp
->blk_fill
= fill
;
2523 mutex_exit(&db
->db_mtx
);
2528 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
2530 dmu_buf_impl_t
*db
= vdb
;
2531 blkptr_t
*bp
= zio
->io_bp
;
2532 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2533 uint64_t txg
= zio
->io_txg
;
2534 dbuf_dirty_record_t
**drp
, *dr
;
2536 ASSERT0(zio
->io_error
);
2537 ASSERT(db
->db_blkptr
== bp
);
2540 * For nopwrites and rewrites we ensure that the bp matches our
2541 * original and bypass all the accounting.
2543 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
2544 ASSERT(BP_EQUAL(bp
, bp_orig
));
2550 DB_GET_OBJSET(&os
, db
);
2551 ds
= os
->os_dsl_dataset
;
2554 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
2555 dsl_dataset_block_born(ds
, bp
, tx
);
2558 mutex_enter(&db
->db_mtx
);
2562 drp
= &db
->db_last_dirty
;
2563 while ((dr
= *drp
) != db
->db_data_pending
)
2565 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2566 ASSERT(dr
->dr_txg
== txg
);
2567 ASSERT(dr
->dr_dbuf
== db
);
2568 ASSERT(dr
->dr_next
== NULL
);
2572 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2577 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
2578 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
2579 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
2584 if (db
->db_level
== 0) {
2585 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2586 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
2587 if (db
->db_state
!= DB_NOFILL
) {
2588 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2589 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
2591 else if (!arc_released(db
->db_buf
))
2592 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
2599 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2600 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2601 if (!BP_IS_HOLE(db
->db_blkptr
)) {
2603 dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2604 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
2606 ASSERT3U(dn
->dn_phys
->dn_maxblkid
2607 >> (db
->db_level
* epbs
), >=, db
->db_blkid
);
2608 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
2611 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2612 list_destroy(&dr
->dt
.di
.dr_children
);
2614 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2616 cv_broadcast(&db
->db_changed
);
2617 ASSERT(db
->db_dirtycnt
> 0);
2618 db
->db_dirtycnt
-= 1;
2619 db
->db_data_pending
= NULL
;
2620 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
2624 dbuf_write_nofill_ready(zio_t
*zio
)
2626 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
2630 dbuf_write_nofill_done(zio_t
*zio
)
2632 dbuf_write_done(zio
, NULL
, zio
->io_private
);
2636 dbuf_write_override_ready(zio_t
*zio
)
2638 dbuf_dirty_record_t
*dr
= zio
->io_private
;
2639 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2641 dbuf_write_ready(zio
, NULL
, db
);
2645 dbuf_write_override_done(zio_t
*zio
)
2647 dbuf_dirty_record_t
*dr
= zio
->io_private
;
2648 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2649 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
2651 mutex_enter(&db
->db_mtx
);
2652 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
2653 if (!BP_IS_HOLE(obp
))
2654 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
2655 arc_release(dr
->dt
.dl
.dr_data
, db
);
2657 mutex_exit(&db
->db_mtx
);
2659 dbuf_write_done(zio
, NULL
, db
);
2663 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
2665 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2668 dmu_buf_impl_t
*parent
= db
->db_parent
;
2669 uint64_t txg
= tx
->tx_txg
;
2679 if (db
->db_state
!= DB_NOFILL
) {
2680 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
2682 * Private object buffers are released here rather
2683 * than in dbuf_dirty() since they are only modified
2684 * in the syncing context and we don't want the
2685 * overhead of making multiple copies of the data.
2687 if (BP_IS_HOLE(db
->db_blkptr
)) {
2690 dbuf_release_bp(db
);
2695 if (parent
!= dn
->dn_dbuf
) {
2696 ASSERT(parent
&& parent
->db_data_pending
);
2697 ASSERT(db
->db_level
== parent
->db_level
-1);
2698 ASSERT(arc_released(parent
->db_buf
));
2699 zio
= parent
->db_data_pending
->dr_zio
;
2701 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
2702 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
2703 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
2704 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2705 ASSERT3P(db
->db_blkptr
, ==,
2706 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
2710 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
2711 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
2714 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
2715 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
2716 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
2718 if (db
->db_blkid
== DMU_SPILL_BLKID
)
2720 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
2722 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
2725 if (db
->db_level
== 0 && dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
2726 ASSERT(db
->db_state
!= DB_NOFILL
);
2727 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
2728 db
->db_blkptr
, data
->b_data
, arc_buf_size(data
), &zp
,
2729 dbuf_write_override_ready
, dbuf_write_override_done
, dr
,
2730 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
2731 mutex_enter(&db
->db_mtx
);
2732 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
2733 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
2734 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
2735 mutex_exit(&db
->db_mtx
);
2736 } else if (db
->db_state
== DB_NOFILL
) {
2737 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
);
2738 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
2739 db
->db_blkptr
, NULL
, db
->db
.db_size
, &zp
,
2740 dbuf_write_nofill_ready
, dbuf_write_nofill_done
, db
,
2741 ZIO_PRIORITY_ASYNC_WRITE
,
2742 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
2744 ASSERT(arc_released(data
));
2745 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
2746 db
->db_blkptr
, data
, DBUF_IS_L2CACHEABLE(db
), &zp
,
2747 dbuf_write_ready
, dbuf_write_done
, db
,
2748 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);