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, 2015 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>
49 struct dbuf_hold_impl_data
{
50 /* Function arguments */
54 boolean_t dh_fail_sparse
;
55 boolean_t dh_fail_uncached
;
57 dmu_buf_impl_t
**dh_dbp
;
59 dmu_buf_impl_t
*dh_db
;
60 dmu_buf_impl_t
*dh_parent
;
63 dbuf_dirty_record_t
*dh_dr
;
64 arc_buf_contents_t dh_type
;
68 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
69 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, boolean_t fail_sparse
,
70 boolean_t fail_uncached
,
71 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
72 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
75 * Number of times that zfs_free_range() took the slow path while doing
76 * a zfs receive. A nonzero value indicates a potential performance problem.
78 uint64_t zfs_free_range_recv_miss
;
80 static void dbuf_destroy(dmu_buf_impl_t
*db
);
81 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
82 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
85 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
86 dmu_buf_evict_func_t
*evict_func
, dmu_buf_t
**clear_on_evict_dbufp
);
90 * Global data structures and functions for the dbuf cache.
92 static kmem_cache_t
*dbuf_cache
;
93 static taskq_t
*dbu_evict_taskq
;
97 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
99 dmu_buf_impl_t
*db
= vdb
;
100 bzero(db
, sizeof (dmu_buf_impl_t
));
102 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
103 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
104 refcount_create(&db
->db_holds
);
111 dbuf_dest(void *vdb
, void *unused
)
113 dmu_buf_impl_t
*db
= vdb
;
114 mutex_destroy(&db
->db_mtx
);
115 cv_destroy(&db
->db_changed
);
116 refcount_destroy(&db
->db_holds
);
120 * dbuf hash table routines
122 static dbuf_hash_table_t dbuf_hash_table
;
124 static uint64_t dbuf_hash_count
;
127 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
129 uintptr_t osv
= (uintptr_t)os
;
130 uint64_t crc
= -1ULL;
132 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
133 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
134 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
135 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
136 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
137 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
138 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
140 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
145 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
147 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
148 ((dbuf)->db.db_object == (obj) && \
149 (dbuf)->db_objset == (os) && \
150 (dbuf)->db_level == (level) && \
151 (dbuf)->db_blkid == (blkid))
154 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
156 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
161 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
162 idx
= hv
& h
->hash_table_mask
;
164 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
165 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
166 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
167 mutex_enter(&db
->db_mtx
);
168 if (db
->db_state
!= DB_EVICTING
) {
169 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
172 mutex_exit(&db
->db_mtx
);
175 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
179 static dmu_buf_impl_t
*
180 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
183 dmu_buf_impl_t
*db
= NULL
;
185 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
186 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
187 if (dn
->dn_bonus
!= NULL
) {
189 mutex_enter(&db
->db_mtx
);
191 rw_exit(&dn
->dn_struct_rwlock
);
192 dnode_rele(dn
, FTAG
);
198 * Insert an entry into the hash table. If there is already an element
199 * equal to elem in the hash table, then the already existing element
200 * will be returned and the new element will not be inserted.
201 * Otherwise returns NULL.
203 static dmu_buf_impl_t
*
204 dbuf_hash_insert(dmu_buf_impl_t
*db
)
206 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
207 objset_t
*os
= db
->db_objset
;
208 uint64_t obj
= db
->db
.db_object
;
209 int level
= db
->db_level
;
210 uint64_t blkid
, hv
, idx
;
213 blkid
= db
->db_blkid
;
214 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
215 idx
= hv
& h
->hash_table_mask
;
217 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
218 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
219 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
220 mutex_enter(&dbf
->db_mtx
);
221 if (dbf
->db_state
!= DB_EVICTING
) {
222 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
225 mutex_exit(&dbf
->db_mtx
);
229 mutex_enter(&db
->db_mtx
);
230 db
->db_hash_next
= h
->hash_table
[idx
];
231 h
->hash_table
[idx
] = db
;
232 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
233 atomic_add_64(&dbuf_hash_count
, 1);
239 * Remove an entry from the hash table. It must be in the EVICTING state.
242 dbuf_hash_remove(dmu_buf_impl_t
*db
)
244 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
246 dmu_buf_impl_t
*dbf
, **dbp
;
248 hv
= DBUF_HASH(db
->db_objset
, db
->db
.db_object
,
249 db
->db_level
, db
->db_blkid
);
250 idx
= hv
& h
->hash_table_mask
;
253 * We musn't hold db_mtx to maintain lock ordering:
254 * DBUF_HASH_MUTEX > db_mtx.
256 ASSERT(refcount_is_zero(&db
->db_holds
));
257 ASSERT(db
->db_state
== DB_EVICTING
);
258 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
260 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
261 dbp
= &h
->hash_table
[idx
];
262 while ((dbf
= *dbp
) != db
) {
263 dbp
= &dbf
->db_hash_next
;
266 *dbp
= db
->db_hash_next
;
267 db
->db_hash_next
= NULL
;
268 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
269 atomic_add_64(&dbuf_hash_count
, -1);
272 static arc_evict_func_t dbuf_do_evict
;
277 } dbvu_verify_type_t
;
280 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
285 if (db
->db_user
== NULL
)
288 /* Only data blocks support the attachment of user data. */
289 ASSERT(db
->db_level
== 0);
291 /* Clients must resolve a dbuf before attaching user data. */
292 ASSERT(db
->db
.db_data
!= NULL
);
293 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
295 holds
= refcount_count(&db
->db_holds
);
296 if (verify_type
== DBVU_EVICTING
) {
298 * Immediate eviction occurs when holds == dirtycnt.
299 * For normal eviction buffers, holds is zero on
300 * eviction, except when dbuf_fix_old_data() calls
301 * dbuf_clear_data(). However, the hold count can grow
302 * during eviction even though db_mtx is held (see
303 * dmu_bonus_hold() for an example), so we can only
304 * test the generic invariant that holds >= dirtycnt.
306 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
308 if (db
->db_user_immediate_evict
== TRUE
)
309 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
311 ASSERT3U(holds
, >, 0);
317 dbuf_evict_user(dmu_buf_impl_t
*db
)
319 dmu_buf_user_t
*dbu
= db
->db_user
;
321 ASSERT(MUTEX_HELD(&db
->db_mtx
));
326 dbuf_verify_user(db
, DBVU_EVICTING
);
330 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
331 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
335 * Invoke the callback from a taskq to avoid lock order reversals
336 * and limit stack depth.
338 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func
, dbu
, 0,
343 dbuf_is_metadata(dmu_buf_impl_t
*db
)
346 * Consider indirect blocks and spill blocks to be meta data.
348 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
351 boolean_t is_metadata
;
354 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
357 return (is_metadata
);
362 dbuf_evict(dmu_buf_impl_t
*db
)
364 ASSERT(MUTEX_HELD(&db
->db_mtx
));
365 ASSERT(db
->db_buf
== NULL
);
366 ASSERT(db
->db_data_pending
== NULL
);
375 uint64_t hsize
= 1ULL << 16;
376 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
380 * The hash table is big enough to fill all of physical memory
381 * with an average block size of zfs_arc_average_blocksize (default 8K).
382 * By default, the table will take up
383 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
385 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
389 h
->hash_table_mask
= hsize
- 1;
390 #if defined(_KERNEL) && defined(HAVE_SPL)
392 * Large allocations which do not require contiguous pages
393 * should be using vmem_alloc() in the linux kernel
395 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
397 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
399 if (h
->hash_table
== NULL
) {
400 /* XXX - we should really return an error instead of assert */
401 ASSERT(hsize
> (1ULL << 10));
406 dbuf_cache
= kmem_cache_create("dmu_buf_impl_t",
407 sizeof (dmu_buf_impl_t
),
408 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
410 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
411 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
416 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
417 * configuration is not required.
419 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
425 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
428 dbuf_stats_destroy();
430 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
431 mutex_destroy(&h
->hash_mutexes
[i
]);
432 #if defined(_KERNEL) && defined(HAVE_SPL)
434 * Large allocations which do not require contiguous pages
435 * should be using vmem_free() in the linux kernel
437 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
439 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
441 kmem_cache_destroy(dbuf_cache
);
442 taskq_destroy(dbu_evict_taskq
);
451 dbuf_verify(dmu_buf_impl_t
*db
)
454 dbuf_dirty_record_t
*dr
;
456 ASSERT(MUTEX_HELD(&db
->db_mtx
));
458 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
461 ASSERT(db
->db_objset
!= NULL
);
465 ASSERT(db
->db_parent
== NULL
);
466 ASSERT(db
->db_blkptr
== NULL
);
468 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
469 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
470 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
471 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
472 db
->db_blkid
== DMU_SPILL_BLKID
||
473 !avl_is_empty(&dn
->dn_dbufs
));
475 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
477 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
478 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
479 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
481 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
482 ASSERT0(db
->db
.db_offset
);
484 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
487 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
488 ASSERT(dr
->dr_dbuf
== db
);
490 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
491 ASSERT(dr
->dr_dbuf
== db
);
494 * We can't assert that db_size matches dn_datablksz because it
495 * can be momentarily different when another thread is doing
498 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
499 dr
= db
->db_data_pending
;
501 * It should only be modified in syncing context, so
502 * make sure we only have one copy of the data.
504 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
507 /* verify db->db_blkptr */
509 if (db
->db_parent
== dn
->dn_dbuf
) {
510 /* db is pointed to by the dnode */
511 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
512 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
513 ASSERT(db
->db_parent
== NULL
);
515 ASSERT(db
->db_parent
!= NULL
);
516 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
517 ASSERT3P(db
->db_blkptr
, ==,
518 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
520 /* db is pointed to by an indirect block */
521 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
523 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
524 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
527 * dnode_grow_indblksz() can make this fail if we don't
528 * have the struct_rwlock. XXX indblksz no longer
529 * grows. safe to do this now?
531 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
532 ASSERT3P(db
->db_blkptr
, ==,
533 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
534 db
->db_blkid
% epb
));
538 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
539 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
540 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
541 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
543 * If the blkptr isn't set but they have nonzero data,
544 * it had better be dirty, otherwise we'll lose that
545 * data when we evict this buffer.
547 if (db
->db_dirtycnt
== 0) {
548 ASSERTV(uint64_t *buf
= db
->db
.db_data
);
551 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
561 dbuf_clear_data(dmu_buf_impl_t
*db
)
563 ASSERT(MUTEX_HELD(&db
->db_mtx
));
566 db
->db
.db_data
= NULL
;
567 if (db
->db_state
!= DB_NOFILL
)
568 db
->db_state
= DB_UNCACHED
;
572 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
574 ASSERT(MUTEX_HELD(&db
->db_mtx
));
578 ASSERT(buf
->b_data
!= NULL
);
579 db
->db
.db_data
= buf
->b_data
;
580 if (!arc_released(buf
))
581 arc_set_callback(buf
, dbuf_do_evict
, db
);
585 * Loan out an arc_buf for read. Return the loaned arc_buf.
588 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
592 mutex_enter(&db
->db_mtx
);
593 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
594 int blksz
= db
->db
.db_size
;
595 spa_t
*spa
= db
->db_objset
->os_spa
;
597 mutex_exit(&db
->db_mtx
);
598 abuf
= arc_loan_buf(spa
, blksz
);
599 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
602 arc_loan_inuse_buf(abuf
, db
);
604 mutex_exit(&db
->db_mtx
);
610 * Calculate which level n block references the data at the level 0 offset
614 dbuf_whichblock(dnode_t
*dn
, int64_t level
, uint64_t offset
)
616 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
618 * The level n blkid is equal to the level 0 blkid divided by
619 * the number of level 0s in a level n block.
621 * The level 0 blkid is offset >> datablkshift =
622 * offset / 2^datablkshift.
624 * The number of level 0s in a level n is the number of block
625 * pointers in an indirect block, raised to the power of level.
626 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
627 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
629 * Thus, the level n blkid is: offset /
630 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
631 * = offset / 2^(datablkshift + level *
632 * (indblkshift - SPA_BLKPTRSHIFT))
633 * = offset >> (datablkshift + level *
634 * (indblkshift - SPA_BLKPTRSHIFT))
636 return (offset
>> (dn
->dn_datablkshift
+ level
*
637 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
)));
639 ASSERT3U(offset
, <, dn
->dn_datablksz
);
645 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
647 dmu_buf_impl_t
*db
= vdb
;
649 mutex_enter(&db
->db_mtx
);
650 ASSERT3U(db
->db_state
, ==, DB_READ
);
652 * All reads are synchronous, so we must have a hold on the dbuf
654 ASSERT(refcount_count(&db
->db_holds
) > 0);
655 ASSERT(db
->db_buf
== NULL
);
656 ASSERT(db
->db
.db_data
== NULL
);
657 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
658 /* we were freed in flight; disregard any error */
659 arc_release(buf
, db
);
660 bzero(buf
->b_data
, db
->db
.db_size
);
662 db
->db_freed_in_flight
= FALSE
;
663 dbuf_set_data(db
, buf
);
664 db
->db_state
= DB_CACHED
;
665 } else if (zio
== NULL
|| zio
->io_error
== 0) {
666 dbuf_set_data(db
, buf
);
667 db
->db_state
= DB_CACHED
;
669 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
670 ASSERT3P(db
->db_buf
, ==, NULL
);
671 VERIFY(arc_buf_remove_ref(buf
, db
));
672 db
->db_state
= DB_UNCACHED
;
674 cv_broadcast(&db
->db_changed
);
675 dbuf_rele_and_unlock(db
, NULL
);
679 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
683 uint32_t aflags
= ARC_FLAG_NOWAIT
;
688 ASSERT(!refcount_is_zero(&db
->db_holds
));
689 /* We need the struct_rwlock to prevent db_blkptr from changing. */
690 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
691 ASSERT(MUTEX_HELD(&db
->db_mtx
));
692 ASSERT(db
->db_state
== DB_UNCACHED
);
693 ASSERT(db
->db_buf
== NULL
);
695 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
696 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
698 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
699 db
->db
.db_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
700 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
701 if (bonuslen
< DN_MAX_BONUSLEN
)
702 bzero(db
->db
.db_data
, DN_MAX_BONUSLEN
);
704 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
706 db
->db_state
= DB_CACHED
;
707 mutex_exit(&db
->db_mtx
);
712 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
713 * processes the delete record and clears the bp while we are waiting
714 * for the dn_mtx (resulting in a "no" from block_freed).
716 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
717 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
718 BP_IS_HOLE(db
->db_blkptr
)))) {
719 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
722 dbuf_set_data(db
, arc_buf_alloc(db
->db_objset
->os_spa
,
723 db
->db
.db_size
, db
, type
));
724 bzero(db
->db
.db_data
, db
->db
.db_size
);
725 db
->db_state
= DB_CACHED
;
726 mutex_exit(&db
->db_mtx
);
732 db
->db_state
= DB_READ
;
733 mutex_exit(&db
->db_mtx
);
735 if (DBUF_IS_L2CACHEABLE(db
))
736 aflags
|= ARC_FLAG_L2CACHE
;
737 if (DBUF_IS_L2COMPRESSIBLE(db
))
738 aflags
|= ARC_FLAG_L2COMPRESS
;
740 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
741 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
742 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
744 dbuf_add_ref(db
, NULL
);
746 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
747 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
748 (flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
751 return (SET_ERROR(err
));
755 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
758 boolean_t havepzio
= (zio
!= NULL
);
763 * We don't have to hold the mutex to check db_state because it
764 * can't be freed while we have a hold on the buffer.
766 ASSERT(!refcount_is_zero(&db
->db_holds
));
768 if (db
->db_state
== DB_NOFILL
)
769 return (SET_ERROR(EIO
));
773 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
774 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
776 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
777 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
778 DBUF_IS_CACHEABLE(db
);
780 mutex_enter(&db
->db_mtx
);
781 if (db
->db_state
== DB_CACHED
) {
782 mutex_exit(&db
->db_mtx
);
784 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1);
785 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
786 rw_exit(&dn
->dn_struct_rwlock
);
788 } else if (db
->db_state
== DB_UNCACHED
) {
789 spa_t
*spa
= dn
->dn_objset
->os_spa
;
792 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
794 err
= dbuf_read_impl(db
, zio
, flags
);
796 /* dbuf_read_impl has dropped db_mtx for us */
798 if (!err
&& prefetch
)
799 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1);
801 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
802 rw_exit(&dn
->dn_struct_rwlock
);
805 if (!err
&& !havepzio
)
809 * Another reader came in while the dbuf was in flight
810 * between UNCACHED and CACHED. Either a writer will finish
811 * writing the buffer (sending the dbuf to CACHED) or the
812 * first reader's request will reach the read_done callback
813 * and send the dbuf to CACHED. Otherwise, a failure
814 * occurred and the dbuf went to UNCACHED.
816 mutex_exit(&db
->db_mtx
);
818 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1);
819 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
820 rw_exit(&dn
->dn_struct_rwlock
);
823 /* Skip the wait per the caller's request. */
824 mutex_enter(&db
->db_mtx
);
825 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
826 while (db
->db_state
== DB_READ
||
827 db
->db_state
== DB_FILL
) {
828 ASSERT(db
->db_state
== DB_READ
||
829 (flags
& DB_RF_HAVESTRUCT
) == 0);
830 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
832 cv_wait(&db
->db_changed
, &db
->db_mtx
);
834 if (db
->db_state
== DB_UNCACHED
)
835 err
= SET_ERROR(EIO
);
837 mutex_exit(&db
->db_mtx
);
840 ASSERT(err
|| havepzio
|| db
->db_state
== DB_CACHED
);
845 dbuf_noread(dmu_buf_impl_t
*db
)
847 ASSERT(!refcount_is_zero(&db
->db_holds
));
848 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
849 mutex_enter(&db
->db_mtx
);
850 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
851 cv_wait(&db
->db_changed
, &db
->db_mtx
);
852 if (db
->db_state
== DB_UNCACHED
) {
853 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
854 spa_t
*spa
= db
->db_objset
->os_spa
;
856 ASSERT(db
->db_buf
== NULL
);
857 ASSERT(db
->db
.db_data
== NULL
);
858 dbuf_set_data(db
, arc_buf_alloc(spa
, db
->db
.db_size
, db
, type
));
859 db
->db_state
= DB_FILL
;
860 } else if (db
->db_state
== DB_NOFILL
) {
863 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
865 mutex_exit(&db
->db_mtx
);
869 * This is our just-in-time copy function. It makes a copy of
870 * buffers, that have been modified in a previous transaction
871 * group, before we modify them in the current active group.
873 * This function is used in two places: when we are dirtying a
874 * buffer for the first time in a txg, and when we are freeing
875 * a range in a dnode that includes this buffer.
877 * Note that when we are called from dbuf_free_range() we do
878 * not put a hold on the buffer, we just traverse the active
879 * dbuf list for the dnode.
882 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
884 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
886 ASSERT(MUTEX_HELD(&db
->db_mtx
));
887 ASSERT(db
->db
.db_data
!= NULL
);
888 ASSERT(db
->db_level
== 0);
889 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
892 (dr
->dt
.dl
.dr_data
!=
893 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
897 * If the last dirty record for this dbuf has not yet synced
898 * and its referencing the dbuf data, either:
899 * reset the reference to point to a new copy,
900 * or (if there a no active holders)
901 * just null out the current db_data pointer.
903 ASSERT(dr
->dr_txg
>= txg
- 2);
904 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
905 /* Note that the data bufs here are zio_bufs */
906 dr
->dt
.dl
.dr_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
907 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
908 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, DN_MAX_BONUSLEN
);
909 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
910 int size
= db
->db
.db_size
;
911 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
912 spa_t
*spa
= db
->db_objset
->os_spa
;
914 dr
->dt
.dl
.dr_data
= arc_buf_alloc(spa
, size
, db
, type
);
915 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
922 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
924 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
925 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
926 uint64_t txg
= dr
->dr_txg
;
928 ASSERT(MUTEX_HELD(&db
->db_mtx
));
929 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
930 ASSERT(db
->db_level
== 0);
932 if (db
->db_blkid
== DMU_BONUS_BLKID
||
933 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
936 ASSERT(db
->db_data_pending
!= dr
);
938 /* free this block */
939 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
940 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
942 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
943 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
946 * Release the already-written buffer, so we leave it in
947 * a consistent dirty state. Note that all callers are
948 * modifying the buffer, so they will immediately do
949 * another (redundant) arc_release(). Therefore, leave
950 * the buf thawed to save the effort of freezing &
951 * immediately re-thawing it.
953 arc_release(dr
->dt
.dl
.dr_data
, db
);
957 * Evict (if its unreferenced) or clear (if its referenced) any level-0
958 * data blocks in the free range, so that any future readers will find
961 * This is a no-op if the dataset is in the middle of an incremental
962 * receive; see comment below for details.
965 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
968 dmu_buf_impl_t
*db_search
;
969 dmu_buf_impl_t
*db
, *db_next
;
970 uint64_t txg
= tx
->tx_txg
;
972 boolean_t freespill
=
973 (start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
);
975 if (end_blkid
> dn
->dn_maxblkid
&& !freespill
)
976 end_blkid
= dn
->dn_maxblkid
;
977 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
979 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
980 db_search
->db_level
= 0;
981 db_search
->db_blkid
= start_blkid
;
982 db_search
->db_state
= DB_SEARCH
;
984 mutex_enter(&dn
->dn_dbufs_mtx
);
985 if (start_blkid
>= dn
->dn_unlisted_l0_blkid
&& !freespill
) {
986 /* There can't be any dbufs in this range; no need to search. */
988 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
989 ASSERT3P(db
, ==, NULL
);
990 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
991 ASSERT(db
== NULL
|| db
->db_level
> 0);
994 } else if (dmu_objset_is_receiving(dn
->dn_objset
)) {
996 * If we are receiving, we expect there to be no dbufs in
997 * the range to be freed, because receive modifies each
998 * block at most once, and in offset order. If this is
999 * not the case, it can lead to performance problems,
1000 * so note that we unexpectedly took the slow path.
1002 atomic_inc_64(&zfs_free_range_recv_miss
);
1005 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1006 ASSERT3P(db
, ==, NULL
);
1007 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1009 for (; db
!= NULL
; db
= db_next
) {
1010 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1011 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1013 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1016 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1018 /* found a level 0 buffer in the range */
1019 mutex_enter(&db
->db_mtx
);
1020 if (dbuf_undirty(db
, tx
)) {
1021 /* mutex has been dropped and dbuf destroyed */
1025 if (db
->db_state
== DB_UNCACHED
||
1026 db
->db_state
== DB_NOFILL
||
1027 db
->db_state
== DB_EVICTING
) {
1028 ASSERT(db
->db
.db_data
== NULL
);
1029 mutex_exit(&db
->db_mtx
);
1032 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1033 /* will be handled in dbuf_read_done or dbuf_rele */
1034 db
->db_freed_in_flight
= TRUE
;
1035 mutex_exit(&db
->db_mtx
);
1038 if (refcount_count(&db
->db_holds
) == 0) {
1043 /* The dbuf is referenced */
1045 if (db
->db_last_dirty
!= NULL
) {
1046 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1048 if (dr
->dr_txg
== txg
) {
1050 * This buffer is "in-use", re-adjust the file
1051 * size to reflect that this buffer may
1052 * contain new data when we sync.
1054 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1055 db
->db_blkid
> dn
->dn_maxblkid
)
1056 dn
->dn_maxblkid
= db
->db_blkid
;
1057 dbuf_unoverride(dr
);
1060 * This dbuf is not dirty in the open context.
1061 * Either uncache it (if its not referenced in
1062 * the open context) or reset its contents to
1065 dbuf_fix_old_data(db
, txg
);
1068 /* clear the contents if its cached */
1069 if (db
->db_state
== DB_CACHED
) {
1070 ASSERT(db
->db
.db_data
!= NULL
);
1071 arc_release(db
->db_buf
, db
);
1072 bzero(db
->db
.db_data
, db
->db
.db_size
);
1073 arc_buf_freeze(db
->db_buf
);
1076 mutex_exit(&db
->db_mtx
);
1080 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1081 mutex_exit(&dn
->dn_dbufs_mtx
);
1085 dbuf_block_freeable(dmu_buf_impl_t
*db
)
1087 dsl_dataset_t
*ds
= db
->db_objset
->os_dsl_dataset
;
1088 uint64_t birth_txg
= 0;
1091 * We don't need any locking to protect db_blkptr:
1092 * If it's syncing, then db_last_dirty will be set
1093 * so we'll ignore db_blkptr.
1095 * This logic ensures that only block births for
1096 * filled blocks are considered.
1098 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1099 if (db
->db_last_dirty
&& (db
->db_blkptr
== NULL
||
1100 !BP_IS_HOLE(db
->db_blkptr
))) {
1101 birth_txg
= db
->db_last_dirty
->dr_txg
;
1102 } else if (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1103 birth_txg
= db
->db_blkptr
->blk_birth
;
1107 * If this block don't exist or is in a snapshot, it can't be freed.
1108 * Don't pass the bp to dsl_dataset_block_freeable() since we
1109 * are holding the db_mtx lock and might deadlock if we are
1110 * prefetching a dedup-ed block.
1113 return (ds
== NULL
||
1114 dsl_dataset_block_freeable(ds
, NULL
, birth_txg
));
1120 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1122 arc_buf_t
*buf
, *obuf
;
1123 int osize
= db
->db
.db_size
;
1124 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1127 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1132 /* XXX does *this* func really need the lock? */
1133 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1136 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1137 * is OK, because there can be no other references to the db
1138 * when we are changing its size, so no concurrent DB_FILL can
1142 * XXX we should be doing a dbuf_read, checking the return
1143 * value and returning that up to our callers
1145 dmu_buf_will_dirty(&db
->db
, tx
);
1147 /* create the data buffer for the new block */
1148 buf
= arc_buf_alloc(dn
->dn_objset
->os_spa
, size
, db
, type
);
1150 /* copy old block data to the new block */
1152 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1153 /* zero the remainder */
1155 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1157 mutex_enter(&db
->db_mtx
);
1158 dbuf_set_data(db
, buf
);
1159 VERIFY(arc_buf_remove_ref(obuf
, db
));
1160 db
->db
.db_size
= size
;
1162 if (db
->db_level
== 0) {
1163 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1164 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1166 mutex_exit(&db
->db_mtx
);
1168 dnode_willuse_space(dn
, size
-osize
, tx
);
1173 dbuf_release_bp(dmu_buf_impl_t
*db
)
1175 ASSERTV(objset_t
*os
= db
->db_objset
);
1177 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1178 ASSERT(arc_released(os
->os_phys_buf
) ||
1179 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1180 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1182 (void) arc_release(db
->db_buf
, db
);
1185 dbuf_dirty_record_t
*
1186 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1190 dbuf_dirty_record_t
**drp
, *dr
;
1191 int drop_struct_lock
= FALSE
;
1192 boolean_t do_free_accounting
= B_FALSE
;
1193 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1195 ASSERT(tx
->tx_txg
!= 0);
1196 ASSERT(!refcount_is_zero(&db
->db_holds
));
1197 DMU_TX_DIRTY_BUF(tx
, db
);
1202 * Shouldn't dirty a regular buffer in syncing context. Private
1203 * objects may be dirtied in syncing context, but only if they
1204 * were already pre-dirtied in open context.
1206 ASSERT(!dmu_tx_is_syncing(tx
) ||
1207 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1208 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1209 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1211 * We make this assert for private objects as well, but after we
1212 * check if we're already dirty. They are allowed to re-dirty
1213 * in syncing context.
1215 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1216 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1217 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1219 mutex_enter(&db
->db_mtx
);
1221 * XXX make this true for indirects too? The problem is that
1222 * transactions created with dmu_tx_create_assigned() from
1223 * syncing context don't bother holding ahead.
1225 ASSERT(db
->db_level
!= 0 ||
1226 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1227 db
->db_state
== DB_NOFILL
);
1229 mutex_enter(&dn
->dn_mtx
);
1231 * Don't set dirtyctx to SYNC if we're just modifying this as we
1232 * initialize the objset.
1234 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
&&
1235 !BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1237 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1238 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1239 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1241 mutex_exit(&dn
->dn_mtx
);
1243 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1244 dn
->dn_have_spill
= B_TRUE
;
1247 * If this buffer is already dirty, we're done.
1249 drp
= &db
->db_last_dirty
;
1250 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1251 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1252 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1254 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1257 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1259 * If this buffer has already been written out,
1260 * we now need to reset its state.
1262 dbuf_unoverride(dr
);
1263 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1264 db
->db_state
!= DB_NOFILL
)
1265 arc_buf_thaw(db
->db_buf
);
1267 mutex_exit(&db
->db_mtx
);
1272 * Only valid if not already dirty.
1274 ASSERT(dn
->dn_object
== 0 ||
1275 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1276 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1278 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1279 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1280 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1281 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1282 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1283 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1286 * We should only be dirtying in syncing context if it's the
1287 * mos or we're initializing the os or it's a special object.
1288 * However, we are allowed to dirty in syncing context provided
1289 * we already dirtied it in open context. Hence we must make
1290 * this assertion only if we're not already dirty.
1293 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1294 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1295 ASSERT(db
->db
.db_size
!= 0);
1297 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1299 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1301 * Update the accounting.
1302 * Note: we delay "free accounting" until after we drop
1303 * the db_mtx. This keeps us from grabbing other locks
1304 * (and possibly deadlocking) in bp_get_dsize() while
1305 * also holding the db_mtx.
1307 dnode_willuse_space(dn
, db
->db
.db_size
, tx
);
1308 do_free_accounting
= dbuf_block_freeable(db
);
1312 * If this buffer is dirty in an old transaction group we need
1313 * to make a copy of it so that the changes we make in this
1314 * transaction group won't leak out when we sync the older txg.
1316 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1317 list_link_init(&dr
->dr_dirty_node
);
1318 if (db
->db_level
== 0) {
1319 void *data_old
= db
->db_buf
;
1321 if (db
->db_state
!= DB_NOFILL
) {
1322 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1323 dbuf_fix_old_data(db
, tx
->tx_txg
);
1324 data_old
= db
->db
.db_data
;
1325 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1327 * Release the data buffer from the cache so
1328 * that we can modify it without impacting
1329 * possible other users of this cached data
1330 * block. Note that indirect blocks and
1331 * private objects are not released until the
1332 * syncing state (since they are only modified
1335 arc_release(db
->db_buf
, db
);
1336 dbuf_fix_old_data(db
, tx
->tx_txg
);
1337 data_old
= db
->db_buf
;
1339 ASSERT(data_old
!= NULL
);
1341 dr
->dt
.dl
.dr_data
= data_old
;
1343 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1344 list_create(&dr
->dt
.di
.dr_children
,
1345 sizeof (dbuf_dirty_record_t
),
1346 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1348 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1349 dr
->dr_accounted
= db
->db
.db_size
;
1351 dr
->dr_txg
= tx
->tx_txg
;
1356 * We could have been freed_in_flight between the dbuf_noread
1357 * and dbuf_dirty. We win, as though the dbuf_noread() had
1358 * happened after the free.
1360 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1361 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1362 mutex_enter(&dn
->dn_mtx
);
1363 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1364 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1367 mutex_exit(&dn
->dn_mtx
);
1368 db
->db_freed_in_flight
= FALSE
;
1372 * This buffer is now part of this txg
1374 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1375 db
->db_dirtycnt
+= 1;
1376 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1378 mutex_exit(&db
->db_mtx
);
1380 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1381 db
->db_blkid
== DMU_SPILL_BLKID
) {
1382 mutex_enter(&dn
->dn_mtx
);
1383 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1384 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1385 mutex_exit(&dn
->dn_mtx
);
1386 dnode_setdirty(dn
, tx
);
1389 } else if (do_free_accounting
) {
1390 blkptr_t
*bp
= db
->db_blkptr
;
1391 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1392 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1394 * This is only a guess -- if the dbuf is dirty
1395 * in a previous txg, we don't know how much
1396 * space it will use on disk yet. We should
1397 * really have the struct_rwlock to access
1398 * db_blkptr, but since this is just a guess,
1399 * it's OK if we get an odd answer.
1401 ddt_prefetch(os
->os_spa
, bp
);
1402 dnode_willuse_space(dn
, -willfree
, tx
);
1405 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1406 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1407 drop_struct_lock
= TRUE
;
1410 if (db
->db_level
== 0) {
1411 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1412 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1415 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1416 dmu_buf_impl_t
*parent
= db
->db_parent
;
1417 dbuf_dirty_record_t
*di
;
1418 int parent_held
= FALSE
;
1420 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1421 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1423 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1424 db
->db_blkid
>> epbs
, FTAG
);
1425 ASSERT(parent
!= NULL
);
1428 if (drop_struct_lock
)
1429 rw_exit(&dn
->dn_struct_rwlock
);
1430 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1431 di
= dbuf_dirty(parent
, tx
);
1433 dbuf_rele(parent
, FTAG
);
1435 mutex_enter(&db
->db_mtx
);
1437 * Since we've dropped the mutex, it's possible that
1438 * dbuf_undirty() might have changed this out from under us.
1440 if (db
->db_last_dirty
== dr
||
1441 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1442 mutex_enter(&di
->dt
.di
.dr_mtx
);
1443 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1444 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1445 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1446 mutex_exit(&di
->dt
.di
.dr_mtx
);
1449 mutex_exit(&db
->db_mtx
);
1451 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1452 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1453 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1454 mutex_enter(&dn
->dn_mtx
);
1455 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1456 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1457 mutex_exit(&dn
->dn_mtx
);
1458 if (drop_struct_lock
)
1459 rw_exit(&dn
->dn_struct_rwlock
);
1462 dnode_setdirty(dn
, tx
);
1468 * Undirty a buffer in the transaction group referenced by the given
1469 * transaction. Return whether this evicted the dbuf.
1472 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1475 uint64_t txg
= tx
->tx_txg
;
1476 dbuf_dirty_record_t
*dr
, **drp
;
1481 * Due to our use of dn_nlevels below, this can only be called
1482 * in open context, unless we are operating on the MOS.
1483 * From syncing context, dn_nlevels may be different from the
1484 * dn_nlevels used when dbuf was dirtied.
1486 ASSERT(db
->db_objset
==
1487 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1488 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1489 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1490 ASSERT0(db
->db_level
);
1491 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1494 * If this buffer is not dirty, we're done.
1496 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1497 if (dr
->dr_txg
<= txg
)
1499 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1501 ASSERT(dr
->dr_txg
== txg
);
1502 ASSERT(dr
->dr_dbuf
== db
);
1507 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1509 ASSERT(db
->db
.db_size
!= 0);
1511 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1512 dr
->dr_accounted
, txg
);
1517 * Note that there are three places in dbuf_dirty()
1518 * where this dirty record may be put on a list.
1519 * Make sure to do a list_remove corresponding to
1520 * every one of those list_insert calls.
1522 if (dr
->dr_parent
) {
1523 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1524 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1525 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1526 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1527 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1528 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1529 mutex_enter(&dn
->dn_mtx
);
1530 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1531 mutex_exit(&dn
->dn_mtx
);
1535 if (db
->db_state
!= DB_NOFILL
) {
1536 dbuf_unoverride(dr
);
1538 ASSERT(db
->db_buf
!= NULL
);
1539 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1540 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1541 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
, db
));
1544 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1546 ASSERT(db
->db_dirtycnt
> 0);
1547 db
->db_dirtycnt
-= 1;
1549 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1550 arc_buf_t
*buf
= db
->db_buf
;
1552 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(buf
));
1553 dbuf_clear_data(db
);
1554 VERIFY(arc_buf_remove_ref(buf
, db
));
1563 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1565 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1566 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1568 ASSERT(tx
->tx_txg
!= 0);
1569 ASSERT(!refcount_is_zero(&db
->db_holds
));
1572 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1573 rf
|= DB_RF_HAVESTRUCT
;
1575 (void) dbuf_read(db
, NULL
, rf
);
1576 (void) dbuf_dirty(db
, tx
);
1580 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1582 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1584 db
->db_state
= DB_NOFILL
;
1586 dmu_buf_will_fill(db_fake
, tx
);
1590 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1592 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1594 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1595 ASSERT(tx
->tx_txg
!= 0);
1596 ASSERT(db
->db_level
== 0);
1597 ASSERT(!refcount_is_zero(&db
->db_holds
));
1599 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1600 dmu_tx_private_ok(tx
));
1603 (void) dbuf_dirty(db
, tx
);
1606 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1609 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1611 mutex_enter(&db
->db_mtx
);
1614 if (db
->db_state
== DB_FILL
) {
1615 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1616 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1617 /* we were freed while filling */
1618 /* XXX dbuf_undirty? */
1619 bzero(db
->db
.db_data
, db
->db
.db_size
);
1620 db
->db_freed_in_flight
= FALSE
;
1622 db
->db_state
= DB_CACHED
;
1623 cv_broadcast(&db
->db_changed
);
1625 mutex_exit(&db
->db_mtx
);
1629 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
1630 bp_embedded_type_t etype
, enum zio_compress comp
,
1631 int uncompressed_size
, int compressed_size
, int byteorder
,
1634 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
1635 struct dirty_leaf
*dl
;
1636 dmu_object_type_t type
;
1638 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
1639 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
1640 SPA_FEATURE_EMBEDDED_DATA
));
1644 type
= DB_DNODE(db
)->dn_type
;
1647 ASSERT0(db
->db_level
);
1648 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1650 dmu_buf_will_not_fill(dbuf
, tx
);
1652 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1653 dl
= &db
->db_last_dirty
->dt
.dl
;
1654 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
1655 data
, comp
, uncompressed_size
, compressed_size
);
1656 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
1657 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
1658 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
1659 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
1661 dl
->dr_override_state
= DR_OVERRIDDEN
;
1662 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
1666 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1667 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1670 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1672 ASSERT(!refcount_is_zero(&db
->db_holds
));
1673 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1674 ASSERT(db
->db_level
== 0);
1675 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1676 ASSERT(buf
!= NULL
);
1677 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1678 ASSERT(tx
->tx_txg
!= 0);
1680 arc_return_buf(buf
, db
);
1681 ASSERT(arc_released(buf
));
1683 mutex_enter(&db
->db_mtx
);
1685 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1686 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1688 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1690 if (db
->db_state
== DB_CACHED
&&
1691 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1692 mutex_exit(&db
->db_mtx
);
1693 (void) dbuf_dirty(db
, tx
);
1694 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1695 VERIFY(arc_buf_remove_ref(buf
, db
));
1696 xuio_stat_wbuf_copied();
1700 xuio_stat_wbuf_nocopy();
1701 if (db
->db_state
== DB_CACHED
) {
1702 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1704 ASSERT(db
->db_buf
!= NULL
);
1705 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
1706 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
1707 if (!arc_released(db
->db_buf
)) {
1708 ASSERT(dr
->dt
.dl
.dr_override_state
==
1710 arc_release(db
->db_buf
, db
);
1712 dr
->dt
.dl
.dr_data
= buf
;
1713 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1714 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
1715 arc_release(db
->db_buf
, db
);
1716 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1720 ASSERT(db
->db_buf
== NULL
);
1721 dbuf_set_data(db
, buf
);
1722 db
->db_state
= DB_FILL
;
1723 mutex_exit(&db
->db_mtx
);
1724 (void) dbuf_dirty(db
, tx
);
1725 dmu_buf_fill_done(&db
->db
, tx
);
1729 * "Clear" the contents of this dbuf. This will mark the dbuf
1730 * EVICTING and clear *most* of its references. Unfortunately,
1731 * when we are not holding the dn_dbufs_mtx, we can't clear the
1732 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1733 * in this case. For callers from the DMU we will usually see:
1734 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1735 * For the arc callback, we will usually see:
1736 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1737 * Sometimes, though, we will get a mix of these two:
1738 * DMU: dbuf_clear()->arc_clear_callback()
1739 * ARC: dbuf_do_evict()->dbuf_destroy()
1741 * This routine will dissociate the dbuf from the arc, by calling
1742 * arc_clear_callback(), but will not evict the data from the ARC.
1745 dbuf_clear(dmu_buf_impl_t
*db
)
1748 dmu_buf_impl_t
*parent
= db
->db_parent
;
1749 dmu_buf_impl_t
*dndb
;
1750 boolean_t dbuf_gone
= B_FALSE
;
1752 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1753 ASSERT(refcount_is_zero(&db
->db_holds
));
1755 dbuf_evict_user(db
);
1757 if (db
->db_state
== DB_CACHED
) {
1758 ASSERT(db
->db
.db_data
!= NULL
);
1759 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1760 zio_buf_free(db
->db
.db_data
, DN_MAX_BONUSLEN
);
1761 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1763 db
->db
.db_data
= NULL
;
1764 db
->db_state
= DB_UNCACHED
;
1767 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1768 ASSERT(db
->db_data_pending
== NULL
);
1770 db
->db_state
= DB_EVICTING
;
1771 db
->db_blkptr
= NULL
;
1776 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& MUTEX_HELD(&dn
->dn_dbufs_mtx
)) {
1777 avl_remove(&dn
->dn_dbufs
, db
);
1778 atomic_dec_32(&dn
->dn_dbufs_count
);
1782 * Decrementing the dbuf count means that the hold corresponding
1783 * to the removed dbuf is no longer discounted in dnode_move(),
1784 * so the dnode cannot be moved until after we release the hold.
1785 * The membar_producer() ensures visibility of the decremented
1786 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1790 db
->db_dnode_handle
= NULL
;
1796 dbuf_gone
= arc_clear_callback(db
->db_buf
);
1799 mutex_exit(&db
->db_mtx
);
1802 * If this dbuf is referenced from an indirect dbuf,
1803 * decrement the ref count on the indirect dbuf.
1805 if (parent
&& parent
!= dndb
)
1806 dbuf_rele(parent
, db
);
1810 * Note: While bpp will always be updated if the function returns success,
1811 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1812 * this happens when the dnode is the meta-dnode, or a userused or groupused
1815 __attribute__((always_inline
))
1817 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
1818 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
1825 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1827 if (blkid
== DMU_SPILL_BLKID
) {
1828 mutex_enter(&dn
->dn_mtx
);
1829 if (dn
->dn_have_spill
&&
1830 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1831 *bpp
= &dn
->dn_phys
->dn_spill
;
1834 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1835 *parentp
= dn
->dn_dbuf
;
1836 mutex_exit(&dn
->dn_mtx
);
1840 if (dn
->dn_phys
->dn_nlevels
== 0)
1843 nlevels
= dn
->dn_phys
->dn_nlevels
;
1845 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1847 ASSERT3U(level
* epbs
, <, 64);
1848 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1849 if (level
>= nlevels
||
1850 (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
1851 /* the buffer has no parent yet */
1852 return (SET_ERROR(ENOENT
));
1853 } else if (level
< nlevels
-1) {
1854 /* this block is referenced from an indirect block */
1857 err
= dbuf_hold_impl(dn
, level
+1,
1858 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
1860 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
1861 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
1862 parentp
, dh
->dh_depth
+ 1);
1863 err
= __dbuf_hold_impl(dh
+ 1);
1867 err
= dbuf_read(*parentp
, NULL
,
1868 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
1870 dbuf_rele(*parentp
, NULL
);
1874 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
1875 (blkid
& ((1ULL << epbs
) - 1));
1878 /* the block is referenced from the dnode */
1879 ASSERT3U(level
, ==, nlevels
-1);
1880 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
1881 blkid
< dn
->dn_phys
->dn_nblkptr
);
1883 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1884 *parentp
= dn
->dn_dbuf
;
1886 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
1891 static dmu_buf_impl_t
*
1892 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
1893 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
1895 objset_t
*os
= dn
->dn_objset
;
1896 dmu_buf_impl_t
*db
, *odb
;
1898 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1899 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
1901 db
= kmem_cache_alloc(dbuf_cache
, KM_SLEEP
);
1904 db
->db
.db_object
= dn
->dn_object
;
1905 db
->db_level
= level
;
1906 db
->db_blkid
= blkid
;
1907 db
->db_last_dirty
= NULL
;
1908 db
->db_dirtycnt
= 0;
1909 db
->db_dnode_handle
= dn
->dn_handle
;
1910 db
->db_parent
= parent
;
1911 db
->db_blkptr
= blkptr
;
1914 db
->db_user_immediate_evict
= FALSE
;
1915 db
->db_freed_in_flight
= FALSE
;
1916 db
->db_pending_evict
= FALSE
;
1918 if (blkid
== DMU_BONUS_BLKID
) {
1919 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
1920 db
->db
.db_size
= DN_MAX_BONUSLEN
-
1921 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
1922 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1923 db
->db
.db_offset
= DMU_BONUS_BLKID
;
1924 db
->db_state
= DB_UNCACHED
;
1925 /* the bonus dbuf is not placed in the hash table */
1926 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1928 } else if (blkid
== DMU_SPILL_BLKID
) {
1929 db
->db
.db_size
= (blkptr
!= NULL
) ?
1930 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
1931 db
->db
.db_offset
= 0;
1934 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
1935 db
->db
.db_size
= blocksize
;
1936 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
1940 * Hold the dn_dbufs_mtx while we get the new dbuf
1941 * in the hash table *and* added to the dbufs list.
1942 * This prevents a possible deadlock with someone
1943 * trying to look up this dbuf before its added to the
1946 mutex_enter(&dn
->dn_dbufs_mtx
);
1947 db
->db_state
= DB_EVICTING
;
1948 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
1949 /* someone else inserted it first */
1950 kmem_cache_free(dbuf_cache
, db
);
1951 mutex_exit(&dn
->dn_dbufs_mtx
);
1954 avl_add(&dn
->dn_dbufs
, db
);
1955 if (db
->db_level
== 0 && db
->db_blkid
>=
1956 dn
->dn_unlisted_l0_blkid
)
1957 dn
->dn_unlisted_l0_blkid
= db
->db_blkid
+ 1;
1958 db
->db_state
= DB_UNCACHED
;
1959 mutex_exit(&dn
->dn_dbufs_mtx
);
1960 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1962 if (parent
&& parent
!= dn
->dn_dbuf
)
1963 dbuf_add_ref(parent
, db
);
1965 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1966 refcount_count(&dn
->dn_holds
) > 0);
1967 (void) refcount_add(&dn
->dn_holds
, db
);
1968 atomic_inc_32(&dn
->dn_dbufs_count
);
1970 dprintf_dbuf(db
, "db=%p\n", db
);
1976 dbuf_do_evict(void *private)
1978 dmu_buf_impl_t
*db
= private;
1980 if (!MUTEX_HELD(&db
->db_mtx
))
1981 mutex_enter(&db
->db_mtx
);
1983 ASSERT(refcount_is_zero(&db
->db_holds
));
1985 if (db
->db_state
!= DB_EVICTING
) {
1986 ASSERT(db
->db_state
== DB_CACHED
);
1991 mutex_exit(&db
->db_mtx
);
1998 dbuf_destroy(dmu_buf_impl_t
*db
)
2000 ASSERT(refcount_is_zero(&db
->db_holds
));
2002 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2004 * If this dbuf is still on the dn_dbufs list,
2005 * remove it from that list.
2007 if (db
->db_dnode_handle
!= NULL
) {
2012 mutex_enter(&dn
->dn_dbufs_mtx
);
2013 avl_remove(&dn
->dn_dbufs
, db
);
2014 atomic_dec_32(&dn
->dn_dbufs_count
);
2015 mutex_exit(&dn
->dn_dbufs_mtx
);
2018 * Decrementing the dbuf count means that the hold
2019 * corresponding to the removed dbuf is no longer
2020 * discounted in dnode_move(), so the dnode cannot be
2021 * moved until after we release the hold.
2024 db
->db_dnode_handle
= NULL
;
2026 dbuf_hash_remove(db
);
2028 db
->db_parent
= NULL
;
2031 ASSERT(db
->db
.db_data
== NULL
);
2032 ASSERT(db
->db_hash_next
== NULL
);
2033 ASSERT(db
->db_blkptr
== NULL
);
2034 ASSERT(db
->db_data_pending
== NULL
);
2036 kmem_cache_free(dbuf_cache
, db
);
2037 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2040 typedef struct dbuf_prefetch_arg
{
2041 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2042 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2043 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2044 int dpa_curlevel
; /* The current level that we're reading */
2045 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2046 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2047 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2048 } dbuf_prefetch_arg_t
;
2051 * Actually issue the prefetch read for the block given.
2054 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2057 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2060 aflags
= dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2062 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2063 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2064 ASSERT(dpa
->dpa_zio
!= NULL
);
2065 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2066 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2067 &aflags
, &dpa
->dpa_zb
);
2071 * Called when an indirect block above our prefetch target is read in. This
2072 * will either read in the next indirect block down the tree or issue the actual
2073 * prefetch if the next block down is our target.
2076 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2078 dbuf_prefetch_arg_t
*dpa
= private;
2082 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2083 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2085 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2086 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2087 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2090 dpa
->dpa_curlevel
--;
2092 nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2093 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2094 bp
= ((blkptr_t
*)abuf
->b_data
) +
2095 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2096 if (BP_IS_HOLE(bp
) || (zio
!= NULL
&& zio
->io_error
!= 0)) {
2097 kmem_free(dpa
, sizeof (*dpa
));
2098 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2099 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2100 dbuf_issue_final_prefetch(dpa
, bp
);
2101 kmem_free(dpa
, sizeof (*dpa
));
2103 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2104 zbookmark_phys_t zb
;
2106 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2108 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2109 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2111 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2112 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2113 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2116 (void) arc_buf_remove_ref(abuf
, private);
2120 * Issue prefetch reads for the given block on the given level. If the indirect
2121 * blocks above that block are not in memory, we will read them in
2122 * asynchronously. As a result, this call never blocks waiting for a read to
2126 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2130 int epbs
, nlevels
, curlevel
;
2134 dbuf_prefetch_arg_t
*dpa
;
2137 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2138 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2140 if (blkid
> dn
->dn_maxblkid
)
2143 if (dnode_block_freed(dn
, blkid
))
2147 * This dnode hasn't been written to disk yet, so there's nothing to
2150 nlevels
= dn
->dn_phys
->dn_nlevels
;
2151 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2154 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2155 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2158 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2161 mutex_exit(&db
->db_mtx
);
2163 * This dbuf already exists. It is either CACHED, or
2164 * (we assume) about to be read or filled.
2170 * Find the closest ancestor (indirect block) of the target block
2171 * that is present in the cache. In this indirect block, we will
2172 * find the bp that is at curlevel, curblkid.
2176 while (curlevel
< nlevels
- 1) {
2177 int parent_level
= curlevel
+ 1;
2178 uint64_t parent_blkid
= curblkid
>> epbs
;
2181 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2182 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2183 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2184 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2185 dbuf_rele(db
, FTAG
);
2189 curlevel
= parent_level
;
2190 curblkid
= parent_blkid
;
2193 if (curlevel
== nlevels
- 1) {
2194 /* No cached indirect blocks found. */
2195 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2196 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2198 if (BP_IS_HOLE(&bp
))
2201 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2203 pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2206 dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2207 ds
= dn
->dn_objset
->os_dsl_dataset
;
2208 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2209 dn
->dn_object
, level
, blkid
);
2210 dpa
->dpa_curlevel
= curlevel
;
2211 dpa
->dpa_prio
= prio
;
2212 dpa
->dpa_aflags
= aflags
;
2213 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2214 dpa
->dpa_epbs
= epbs
;
2218 * If we have the indirect just above us, no need to do the asynchronous
2219 * prefetch chain; we'll just run the last step ourselves. If we're at
2220 * a higher level, though, we want to issue the prefetches for all the
2221 * indirect blocks asynchronously, so we can go on with whatever we were
2224 if (curlevel
== level
) {
2225 ASSERT3U(curblkid
, ==, blkid
);
2226 dbuf_issue_final_prefetch(dpa
, &bp
);
2227 kmem_free(dpa
, sizeof (*dpa
));
2229 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2230 zbookmark_phys_t zb
;
2232 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2233 dn
->dn_object
, curlevel
, curblkid
);
2234 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2235 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2236 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2240 * We use pio here instead of dpa_zio since it's possible that
2241 * dpa may have already been freed.
2246 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2249 * Returns with db_holds incremented, and db_mtx not held.
2250 * Note: dn_struct_rwlock must be held.
2253 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2255 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2256 dh
->dh_parent
= NULL
;
2258 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2259 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2260 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2262 *(dh
->dh_dbp
) = NULL
;
2264 /* dbuf_find() returns with db_mtx held */
2265 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2266 dh
->dh_level
, dh
->dh_blkid
);
2268 if (dh
->dh_db
== NULL
) {
2271 if (dh
->dh_fail_uncached
)
2272 return (SET_ERROR(ENOENT
));
2274 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2275 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2276 dh
->dh_fail_sparse
, &dh
->dh_parent
,
2278 if (dh
->dh_fail_sparse
) {
2279 if (dh
->dh_err
== 0 &&
2280 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2281 dh
->dh_err
= SET_ERROR(ENOENT
);
2284 dbuf_rele(dh
->dh_parent
, NULL
);
2285 return (dh
->dh_err
);
2288 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2289 return (dh
->dh_err
);
2290 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2291 dh
->dh_parent
, dh
->dh_bp
);
2294 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2295 mutex_exit(&dh
->dh_db
->db_mtx
);
2296 return (SET_ERROR(ENOENT
));
2299 if (dh
->dh_db
->db_buf
&& refcount_is_zero(&dh
->dh_db
->db_holds
)) {
2300 arc_buf_add_ref(dh
->dh_db
->db_buf
, dh
->dh_db
);
2301 if (dh
->dh_db
->db_buf
->b_data
== NULL
) {
2302 dbuf_clear(dh
->dh_db
);
2303 if (dh
->dh_parent
) {
2304 dbuf_rele(dh
->dh_parent
, NULL
);
2305 dh
->dh_parent
= NULL
;
2309 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2312 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2315 * If this buffer is currently syncing out, and we are are
2316 * still referencing it from db_data, we need to make a copy
2317 * of it in case we decide we want to dirty it again in this txg.
2319 if (dh
->dh_db
->db_level
== 0 &&
2320 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2321 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2322 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2323 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2325 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2326 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2328 dbuf_set_data(dh
->dh_db
,
2329 arc_buf_alloc(dh
->dh_dn
->dn_objset
->os_spa
,
2330 dh
->dh_db
->db
.db_size
, dh
->dh_db
, dh
->dh_type
));
2331 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2332 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2336 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2337 DBUF_VERIFY(dh
->dh_db
);
2338 mutex_exit(&dh
->dh_db
->db_mtx
);
2340 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2342 dbuf_rele(dh
->dh_parent
, NULL
);
2344 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2345 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2346 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2347 *(dh
->dh_dbp
) = dh
->dh_db
;
2353 * The following code preserves the recursive function dbuf_hold_impl()
2354 * but moves the local variables AND function arguments to the heap to
2355 * minimize the stack frame size. Enough space is initially allocated
2356 * on the stack for 20 levels of recursion.
2359 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2360 boolean_t fail_sparse
, boolean_t fail_uncached
,
2361 void *tag
, dmu_buf_impl_t
**dbp
)
2363 struct dbuf_hold_impl_data
*dh
;
2366 dh
= kmem_zalloc(sizeof (struct dbuf_hold_impl_data
) *
2367 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2368 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2369 fail_uncached
, tag
, dbp
, 0);
2371 error
= __dbuf_hold_impl(dh
);
2373 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2374 DBUF_HOLD_IMPL_MAX_DEPTH
);
2380 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2381 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2382 boolean_t fail_sparse
, boolean_t fail_uncached
,
2383 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2386 dh
->dh_level
= level
;
2387 dh
->dh_blkid
= blkid
;
2389 dh
->dh_fail_sparse
= fail_sparse
;
2390 dh
->dh_fail_uncached
= fail_uncached
;
2394 dh
->dh_depth
= depth
;
2398 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2400 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2404 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2407 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2408 return (err
? NULL
: db
);
2412 dbuf_create_bonus(dnode_t
*dn
)
2414 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2416 ASSERT(dn
->dn_bonus
== NULL
);
2417 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2421 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2423 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2426 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2427 return (SET_ERROR(ENOTSUP
));
2429 blksz
= SPA_MINBLOCKSIZE
;
2430 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2431 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2435 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2436 dbuf_new_size(db
, blksz
, tx
);
2437 rw_exit(&dn
->dn_struct_rwlock
);
2444 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2446 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2449 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2451 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2453 VERIFY(refcount_add(&db
->db_holds
, tag
) > 1);
2456 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2458 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2461 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2462 dmu_buf_impl_t
*found_db
;
2463 boolean_t result
= B_FALSE
;
2465 if (blkid
== DMU_BONUS_BLKID
)
2466 found_db
= dbuf_find_bonus(os
, obj
);
2468 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2470 if (found_db
!= NULL
) {
2471 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2472 (void) refcount_add(&db
->db_holds
, tag
);
2475 mutex_exit(&found_db
->db_mtx
);
2481 * If you call dbuf_rele() you had better not be referencing the dnode handle
2482 * unless you have some other direct or indirect hold on the dnode. (An indirect
2483 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2484 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2485 * dnode's parent dbuf evicting its dnode handles.
2488 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2490 mutex_enter(&db
->db_mtx
);
2491 dbuf_rele_and_unlock(db
, tag
);
2495 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2497 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2501 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2502 * db_dirtycnt and db_holds to be updated atomically.
2505 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2509 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2513 * Remove the reference to the dbuf before removing its hold on the
2514 * dnode so we can guarantee in dnode_move() that a referenced bonus
2515 * buffer has a corresponding dnode hold.
2517 holds
= refcount_remove(&db
->db_holds
, tag
);
2521 * We can't freeze indirects if there is a possibility that they
2522 * may be modified in the current syncing context.
2524 if (db
->db_buf
&& holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0))
2525 arc_buf_freeze(db
->db_buf
);
2527 if (holds
== db
->db_dirtycnt
&&
2528 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2529 dbuf_evict_user(db
);
2532 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2534 boolean_t evict_dbuf
= db
->db_pending_evict
;
2537 * If the dnode moves here, we cannot cross this
2538 * barrier until the move completes.
2543 atomic_dec_32(&dn
->dn_dbufs_count
);
2546 * Decrementing the dbuf count means that the bonus
2547 * buffer's dnode hold is no longer discounted in
2548 * dnode_move(). The dnode cannot move until after
2549 * the dnode_rele() below.
2554 * Do not reference db after its lock is dropped.
2555 * Another thread may evict it.
2557 mutex_exit(&db
->db_mtx
);
2560 dnode_evict_bonus(dn
);
2563 } else if (db
->db_buf
== NULL
) {
2565 * This is a special case: we never associated this
2566 * dbuf with any data allocated from the ARC.
2568 ASSERT(db
->db_state
== DB_UNCACHED
||
2569 db
->db_state
== DB_NOFILL
);
2571 } else if (arc_released(db
->db_buf
)) {
2572 arc_buf_t
*buf
= db
->db_buf
;
2574 * This dbuf has anonymous data associated with it.
2576 dbuf_clear_data(db
);
2577 VERIFY(arc_buf_remove_ref(buf
, db
));
2580 VERIFY(!arc_buf_remove_ref(db
->db_buf
, db
));
2583 * A dbuf will be eligible for eviction if either the
2584 * 'primarycache' property is set or a duplicate
2585 * copy of this buffer is already cached in the arc.
2587 * In the case of the 'primarycache' a buffer
2588 * is considered for eviction if it matches the
2589 * criteria set in the property.
2591 * To decide if our buffer is considered a
2592 * duplicate, we must call into the arc to determine
2593 * if multiple buffers are referencing the same
2594 * block on-disk. If so, then we simply evict
2597 if (!DBUF_IS_CACHEABLE(db
)) {
2598 if (db
->db_blkptr
!= NULL
&&
2599 !BP_IS_HOLE(db
->db_blkptr
) &&
2600 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2602 dmu_objset_spa(db
->db_objset
);
2603 blkptr_t bp
= *db
->db_blkptr
;
2605 arc_freed(spa
, &bp
);
2609 } else if (db
->db_pending_evict
||
2610 arc_buf_eviction_needed(db
->db_buf
)) {
2613 mutex_exit(&db
->db_mtx
);
2617 mutex_exit(&db
->db_mtx
);
2621 #pragma weak dmu_buf_refcount = dbuf_refcount
2623 dbuf_refcount(dmu_buf_impl_t
*db
)
2625 return (refcount_count(&db
->db_holds
));
2629 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
2630 dmu_buf_user_t
*new_user
)
2632 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2634 mutex_enter(&db
->db_mtx
);
2635 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2636 if (db
->db_user
== old_user
)
2637 db
->db_user
= new_user
;
2639 old_user
= db
->db_user
;
2640 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2641 mutex_exit(&db
->db_mtx
);
2647 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2649 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
2653 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2655 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2657 db
->db_user_immediate_evict
= TRUE
;
2658 return (dmu_buf_set_user(db_fake
, user
));
2662 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2664 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
2668 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2670 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2672 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2673 return (db
->db_user
);
2677 dmu_buf_user_evict_wait()
2679 taskq_wait(dbu_evict_taskq
);
2683 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2685 boolean_t res
= B_FALSE
;
2686 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2689 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2690 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2696 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2698 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2699 return (dbi
->db_blkptr
);
2703 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2705 /* ASSERT(dmu_tx_is_syncing(tx) */
2706 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2708 if (db
->db_blkptr
!= NULL
)
2711 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2712 db
->db_blkptr
= &dn
->dn_phys
->dn_spill
;
2713 BP_ZERO(db
->db_blkptr
);
2716 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2718 * This buffer was allocated at a time when there was
2719 * no available blkptrs from the dnode, or it was
2720 * inappropriate to hook it in (i.e., nlevels mis-match).
2722 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2723 ASSERT(db
->db_parent
== NULL
);
2724 db
->db_parent
= dn
->dn_dbuf
;
2725 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2728 dmu_buf_impl_t
*parent
= db
->db_parent
;
2729 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2731 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2732 if (parent
== NULL
) {
2733 mutex_exit(&db
->db_mtx
);
2734 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2735 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2736 db
->db_blkid
>> epbs
, db
);
2737 rw_exit(&dn
->dn_struct_rwlock
);
2738 mutex_enter(&db
->db_mtx
);
2739 db
->db_parent
= parent
;
2741 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2742 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2748 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2749 * is critical the we not allow the compiler to inline this function in to
2750 * dbuf_sync_list() thereby drastically bloating the stack usage.
2752 noinline
static void
2753 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2755 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2759 ASSERT(dmu_tx_is_syncing(tx
));
2761 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2763 mutex_enter(&db
->db_mtx
);
2765 ASSERT(db
->db_level
> 0);
2768 /* Read the block if it hasn't been read yet. */
2769 if (db
->db_buf
== NULL
) {
2770 mutex_exit(&db
->db_mtx
);
2771 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2772 mutex_enter(&db
->db_mtx
);
2774 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2775 ASSERT(db
->db_buf
!= NULL
);
2779 /* Indirect block size must match what the dnode thinks it is. */
2780 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2781 dbuf_check_blkptr(dn
, db
);
2784 /* Provide the pending dirty record to child dbufs */
2785 db
->db_data_pending
= dr
;
2787 mutex_exit(&db
->db_mtx
);
2788 dbuf_write(dr
, db
->db_buf
, tx
);
2791 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2792 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
2793 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2794 mutex_exit(&dr
->dt
.di
.dr_mtx
);
2799 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
2800 * critical the we not allow the compiler to inline this function in to
2801 * dbuf_sync_list() thereby drastically bloating the stack usage.
2803 noinline
static void
2804 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2806 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
2807 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2810 uint64_t txg
= tx
->tx_txg
;
2812 ASSERT(dmu_tx_is_syncing(tx
));
2814 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2816 mutex_enter(&db
->db_mtx
);
2818 * To be synced, we must be dirtied. But we
2819 * might have been freed after the dirty.
2821 if (db
->db_state
== DB_UNCACHED
) {
2822 /* This buffer has been freed since it was dirtied */
2823 ASSERT(db
->db
.db_data
== NULL
);
2824 } else if (db
->db_state
== DB_FILL
) {
2825 /* This buffer was freed and is now being re-filled */
2826 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
2828 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
2835 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2836 mutex_enter(&dn
->dn_mtx
);
2837 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
2838 mutex_exit(&dn
->dn_mtx
);
2842 * If this is a bonus buffer, simply copy the bonus data into the
2843 * dnode. It will be written out when the dnode is synced (and it
2844 * will be synced, since it must have been dirty for dbuf_sync to
2847 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2848 dbuf_dirty_record_t
**drp
;
2850 ASSERT(*datap
!= NULL
);
2851 ASSERT0(db
->db_level
);
2852 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
2853 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
2856 if (*datap
!= db
->db
.db_data
) {
2857 zio_buf_free(*datap
, DN_MAX_BONUSLEN
);
2858 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
2860 db
->db_data_pending
= NULL
;
2861 drp
= &db
->db_last_dirty
;
2863 drp
= &(*drp
)->dr_next
;
2864 ASSERT(dr
->dr_next
== NULL
);
2865 ASSERT(dr
->dr_dbuf
== db
);
2867 if (dr
->dr_dbuf
->db_level
!= 0) {
2868 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2869 list_destroy(&dr
->dt
.di
.dr_children
);
2871 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2872 ASSERT(db
->db_dirtycnt
> 0);
2873 db
->db_dirtycnt
-= 1;
2874 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
2881 * This function may have dropped the db_mtx lock allowing a dmu_sync
2882 * operation to sneak in. As a result, we need to ensure that we
2883 * don't check the dr_override_state until we have returned from
2884 * dbuf_check_blkptr.
2886 dbuf_check_blkptr(dn
, db
);
2889 * If this buffer is in the middle of an immediate write,
2890 * wait for the synchronous IO to complete.
2892 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
2893 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
2894 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2895 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
2898 if (db
->db_state
!= DB_NOFILL
&&
2899 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2900 refcount_count(&db
->db_holds
) > 1 &&
2901 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
2902 *datap
== db
->db_buf
) {
2904 * If this buffer is currently "in use" (i.e., there
2905 * are active holds and db_data still references it),
2906 * then make a copy before we start the write so that
2907 * any modifications from the open txg will not leak
2910 * NOTE: this copy does not need to be made for
2911 * objects only modified in the syncing context (e.g.
2912 * DNONE_DNODE blocks).
2914 int blksz
= arc_buf_size(*datap
);
2915 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2916 *datap
= arc_buf_alloc(os
->os_spa
, blksz
, db
, type
);
2917 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
2919 db
->db_data_pending
= dr
;
2921 mutex_exit(&db
->db_mtx
);
2923 dbuf_write(dr
, *datap
, tx
);
2925 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2926 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2927 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
2931 * Although zio_nowait() does not "wait for an IO", it does
2932 * initiate the IO. If this is an empty write it seems plausible
2933 * that the IO could actually be completed before the nowait
2934 * returns. We need to DB_DNODE_EXIT() first in case
2935 * zio_nowait() invalidates the dbuf.
2938 zio_nowait(dr
->dr_zio
);
2943 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
2945 dbuf_dirty_record_t
*dr
;
2947 while ((dr
= list_head(list
))) {
2948 if (dr
->dr_zio
!= NULL
) {
2950 * If we find an already initialized zio then we
2951 * are processing the meta-dnode, and we have finished.
2952 * The dbufs for all dnodes are put back on the list
2953 * during processing, so that we can zio_wait()
2954 * these IOs after initiating all child IOs.
2956 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
2957 DMU_META_DNODE_OBJECT
);
2960 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
2961 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
2962 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
2964 list_remove(list
, dr
);
2965 if (dr
->dr_dbuf
->db_level
> 0)
2966 dbuf_sync_indirect(dr
, tx
);
2968 dbuf_sync_leaf(dr
, tx
);
2974 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
2976 dmu_buf_impl_t
*db
= vdb
;
2978 blkptr_t
*bp
= zio
->io_bp
;
2979 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2980 spa_t
*spa
= zio
->io_spa
;
2985 ASSERT3P(db
->db_blkptr
, ==, bp
);
2989 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
2990 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
2991 zio
->io_prev_space_delta
= delta
;
2993 if (bp
->blk_birth
!= 0) {
2994 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
2995 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
2996 (db
->db_blkid
== DMU_SPILL_BLKID
&&
2997 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
2998 BP_IS_EMBEDDED(bp
));
2999 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3002 mutex_enter(&db
->db_mtx
);
3005 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3006 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3007 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3008 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3012 if (db
->db_level
== 0) {
3013 mutex_enter(&dn
->dn_mtx
);
3014 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3015 db
->db_blkid
!= DMU_SPILL_BLKID
)
3016 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3017 mutex_exit(&dn
->dn_mtx
);
3019 if (dn
->dn_type
== DMU_OT_DNODE
) {
3020 dnode_phys_t
*dnp
= db
->db
.db_data
;
3021 for (i
= db
->db
.db_size
>> DNODE_SHIFT
; i
> 0;
3023 if (dnp
->dn_type
!= DMU_OT_NONE
)
3027 if (BP_IS_HOLE(bp
)) {
3034 blkptr_t
*ibp
= db
->db
.db_data
;
3035 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3036 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3037 if (BP_IS_HOLE(ibp
))
3039 fill
+= BP_GET_FILL(ibp
);
3044 if (!BP_IS_EMBEDDED(bp
))
3045 bp
->blk_fill
= fill
;
3047 mutex_exit(&db
->db_mtx
);
3051 * The SPA will call this callback several times for each zio - once
3052 * for every physical child i/o (zio->io_phys_children times). This
3053 * allows the DMU to monitor the progress of each logical i/o. For example,
3054 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3055 * block. There may be a long delay before all copies/fragments are completed,
3056 * so this callback allows us to retire dirty space gradually, as the physical
3061 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3063 dmu_buf_impl_t
*db
= arg
;
3064 objset_t
*os
= db
->db_objset
;
3065 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3066 dbuf_dirty_record_t
*dr
;
3069 dr
= db
->db_data_pending
;
3070 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3073 * The callback will be called io_phys_children times. Retire one
3074 * portion of our dirty space each time we are called. Any rounding
3075 * error will be cleaned up by dsl_pool_sync()'s call to
3076 * dsl_pool_undirty_space().
3078 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3079 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3084 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3086 dmu_buf_impl_t
*db
= vdb
;
3087 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3088 blkptr_t
*bp
= db
->db_blkptr
;
3089 objset_t
*os
= db
->db_objset
;
3090 dmu_tx_t
*tx
= os
->os_synctx
;
3091 dbuf_dirty_record_t
**drp
, *dr
;
3093 ASSERT0(zio
->io_error
);
3094 ASSERT(db
->db_blkptr
== bp
);
3097 * For nopwrites and rewrites we ensure that the bp matches our
3098 * original and bypass all the accounting.
3100 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3101 ASSERT(BP_EQUAL(bp
, bp_orig
));
3103 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3104 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3105 dsl_dataset_block_born(ds
, bp
, tx
);
3108 mutex_enter(&db
->db_mtx
);
3112 drp
= &db
->db_last_dirty
;
3113 while ((dr
= *drp
) != db
->db_data_pending
)
3115 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3116 ASSERT(dr
->dr_dbuf
== db
);
3117 ASSERT(dr
->dr_next
== NULL
);
3121 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3126 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3127 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3128 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3133 if (db
->db_level
== 0) {
3134 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3135 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3136 if (db
->db_state
!= DB_NOFILL
) {
3137 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3138 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
3140 else if (!arc_released(db
->db_buf
))
3141 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3148 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3149 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3150 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3151 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3153 ASSERT3U(db
->db_blkid
, <=,
3154 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3155 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3157 if (!arc_released(db
->db_buf
))
3158 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3161 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3162 list_destroy(&dr
->dt
.di
.dr_children
);
3164 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3166 cv_broadcast(&db
->db_changed
);
3167 ASSERT(db
->db_dirtycnt
> 0);
3168 db
->db_dirtycnt
-= 1;
3169 db
->db_data_pending
= NULL
;
3170 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3174 dbuf_write_nofill_ready(zio_t
*zio
)
3176 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3180 dbuf_write_nofill_done(zio_t
*zio
)
3182 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3186 dbuf_write_override_ready(zio_t
*zio
)
3188 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3189 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3191 dbuf_write_ready(zio
, NULL
, db
);
3195 dbuf_write_override_done(zio_t
*zio
)
3197 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3198 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3199 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3201 mutex_enter(&db
->db_mtx
);
3202 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3203 if (!BP_IS_HOLE(obp
))
3204 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3205 arc_release(dr
->dt
.dl
.dr_data
, db
);
3207 mutex_exit(&db
->db_mtx
);
3209 dbuf_write_done(zio
, NULL
, db
);
3212 /* Issue I/O to commit a dirty buffer to disk. */
3214 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3216 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3219 dmu_buf_impl_t
*parent
= db
->db_parent
;
3220 uint64_t txg
= tx
->tx_txg
;
3221 zbookmark_phys_t zb
;
3230 if (db
->db_state
!= DB_NOFILL
) {
3231 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3233 * Private object buffers are released here rather
3234 * than in dbuf_dirty() since they are only modified
3235 * in the syncing context and we don't want the
3236 * overhead of making multiple copies of the data.
3238 if (BP_IS_HOLE(db
->db_blkptr
)) {
3241 dbuf_release_bp(db
);
3246 if (parent
!= dn
->dn_dbuf
) {
3247 /* Our parent is an indirect block. */
3248 /* We have a dirty parent that has been scheduled for write. */
3249 ASSERT(parent
&& parent
->db_data_pending
);
3250 /* Our parent's buffer is one level closer to the dnode. */
3251 ASSERT(db
->db_level
== parent
->db_level
-1);
3253 * We're about to modify our parent's db_data by modifying
3254 * our block pointer, so the parent must be released.
3256 ASSERT(arc_released(parent
->db_buf
));
3257 zio
= parent
->db_data_pending
->dr_zio
;
3259 /* Our parent is the dnode itself. */
3260 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3261 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3262 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3263 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3264 ASSERT3P(db
->db_blkptr
, ==,
3265 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3269 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3270 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3273 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3274 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3275 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3277 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3279 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3281 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3284 if (db
->db_level
== 0 &&
3285 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3287 * The BP for this block has been provided by open context
3288 * (by dmu_sync() or dmu_buf_write_embedded()).
3290 void *contents
= (data
!= NULL
) ? data
->b_data
: NULL
;
3292 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3293 db
->db_blkptr
, contents
, db
->db
.db_size
, &zp
,
3294 dbuf_write_override_ready
, NULL
, dbuf_write_override_done
,
3295 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3296 mutex_enter(&db
->db_mtx
);
3297 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3298 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3299 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3300 mutex_exit(&db
->db_mtx
);
3301 } else if (db
->db_state
== DB_NOFILL
) {
3302 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
);
3303 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3304 db
->db_blkptr
, NULL
, db
->db
.db_size
, &zp
,
3305 dbuf_write_nofill_ready
, NULL
, dbuf_write_nofill_done
, db
,
3306 ZIO_PRIORITY_ASYNC_WRITE
,
3307 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3309 ASSERT(arc_released(data
));
3310 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3311 db
->db_blkptr
, data
, DBUF_IS_L2CACHEABLE(db
),
3312 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dbuf_write_ready
,
3313 dbuf_write_physdone
, dbuf_write_done
, db
,
3314 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3318 #if defined(_KERNEL) && defined(HAVE_SPL)
3319 EXPORT_SYMBOL(dbuf_find
);
3320 EXPORT_SYMBOL(dbuf_is_metadata
);
3321 EXPORT_SYMBOL(dbuf_evict
);
3322 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
3323 EXPORT_SYMBOL(dbuf_whichblock
);
3324 EXPORT_SYMBOL(dbuf_read
);
3325 EXPORT_SYMBOL(dbuf_unoverride
);
3326 EXPORT_SYMBOL(dbuf_free_range
);
3327 EXPORT_SYMBOL(dbuf_new_size
);
3328 EXPORT_SYMBOL(dbuf_release_bp
);
3329 EXPORT_SYMBOL(dbuf_dirty
);
3330 EXPORT_SYMBOL(dmu_buf_will_dirty
);
3331 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
3332 EXPORT_SYMBOL(dmu_buf_will_fill
);
3333 EXPORT_SYMBOL(dmu_buf_fill_done
);
3334 EXPORT_SYMBOL(dmu_buf_rele
);
3335 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
3336 EXPORT_SYMBOL(dbuf_clear
);
3337 EXPORT_SYMBOL(dbuf_prefetch
);
3338 EXPORT_SYMBOL(dbuf_hold_impl
);
3339 EXPORT_SYMBOL(dbuf_hold
);
3340 EXPORT_SYMBOL(dbuf_hold_level
);
3341 EXPORT_SYMBOL(dbuf_create_bonus
);
3342 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
3343 EXPORT_SYMBOL(dbuf_rm_spill
);
3344 EXPORT_SYMBOL(dbuf_add_ref
);
3345 EXPORT_SYMBOL(dbuf_rele
);
3346 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
3347 EXPORT_SYMBOL(dbuf_refcount
);
3348 EXPORT_SYMBOL(dbuf_sync_list
);
3349 EXPORT_SYMBOL(dmu_buf_set_user
);
3350 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3351 EXPORT_SYMBOL(dmu_buf_get_user
);
3352 EXPORT_SYMBOL(dmu_buf_freeable
);
3353 EXPORT_SYMBOL(dmu_buf_get_blkptr
);