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, 2016 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_inc_64(&dbuf_hash_count
);
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_dec_64(&dbuf_hash_count
);
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 ASSERT0(db
->db
.db_offset
);
483 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
486 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
487 ASSERT(dr
->dr_dbuf
== db
);
489 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
490 ASSERT(dr
->dr_dbuf
== db
);
493 * We can't assert that db_size matches dn_datablksz because it
494 * can be momentarily different when another thread is doing
497 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
498 dr
= db
->db_data_pending
;
500 * It should only be modified in syncing context, so
501 * make sure we only have one copy of the data.
503 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
506 /* verify db->db_blkptr */
508 if (db
->db_parent
== dn
->dn_dbuf
) {
509 /* db is pointed to by the dnode */
510 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
511 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
512 ASSERT(db
->db_parent
== NULL
);
514 ASSERT(db
->db_parent
!= NULL
);
515 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
516 ASSERT3P(db
->db_blkptr
, ==,
517 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
519 /* db is pointed to by an indirect block */
520 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
522 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
523 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
526 * dnode_grow_indblksz() can make this fail if we don't
527 * have the struct_rwlock. XXX indblksz no longer
528 * grows. safe to do this now?
530 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
531 ASSERT3P(db
->db_blkptr
, ==,
532 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
533 db
->db_blkid
% epb
));
537 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
538 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
539 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
540 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
542 * If the blkptr isn't set but they have nonzero data,
543 * it had better be dirty, otherwise we'll lose that
544 * data when we evict this buffer.
546 * There is an exception to this rule for indirect blocks; in
547 * this case, if the indirect block is a hole, we fill in a few
548 * fields on each of the child blocks (importantly, birth time)
549 * to prevent hole birth times from being lost when you
550 * partially fill in a hole.
552 if (db
->db_dirtycnt
== 0) {
553 if (db
->db_level
== 0) {
554 uint64_t *buf
= db
->db
.db_data
;
557 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
562 blkptr_t
*bps
= db
->db
.db_data
;
563 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
566 * We want to verify that all the blkptrs in the
567 * indirect block are holes, but we may have
568 * automatically set up a few fields for them.
569 * We iterate through each blkptr and verify
570 * they only have those fields set.
573 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
575 blkptr_t
*bp
= &bps
[i
];
576 ASSERT(ZIO_CHECKSUM_IS_ZERO(
579 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
580 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
581 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
582 ASSERT0(bp
->blk_fill
);
583 ASSERT0(bp
->blk_pad
[0]);
584 ASSERT0(bp
->blk_pad
[1]);
585 ASSERT(!BP_IS_EMBEDDED(bp
));
586 ASSERT(BP_IS_HOLE(bp
));
587 ASSERT0(bp
->blk_phys_birth
);
597 dbuf_clear_data(dmu_buf_impl_t
*db
)
599 ASSERT(MUTEX_HELD(&db
->db_mtx
));
602 db
->db
.db_data
= NULL
;
603 if (db
->db_state
!= DB_NOFILL
)
604 db
->db_state
= DB_UNCACHED
;
608 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
610 ASSERT(MUTEX_HELD(&db
->db_mtx
));
614 ASSERT(buf
->b_data
!= NULL
);
615 db
->db
.db_data
= buf
->b_data
;
616 if (!arc_released(buf
))
617 arc_set_callback(buf
, dbuf_do_evict
, db
);
621 * Loan out an arc_buf for read. Return the loaned arc_buf.
624 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
628 mutex_enter(&db
->db_mtx
);
629 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
630 int blksz
= db
->db
.db_size
;
631 spa_t
*spa
= db
->db_objset
->os_spa
;
633 mutex_exit(&db
->db_mtx
);
634 abuf
= arc_loan_buf(spa
, blksz
);
635 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
638 arc_loan_inuse_buf(abuf
, db
);
640 mutex_exit(&db
->db_mtx
);
646 * Calculate which level n block references the data at the level 0 offset
650 dbuf_whichblock(dnode_t
*dn
, int64_t level
, uint64_t offset
)
652 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
654 * The level n blkid is equal to the level 0 blkid divided by
655 * the number of level 0s in a level n block.
657 * The level 0 blkid is offset >> datablkshift =
658 * offset / 2^datablkshift.
660 * The number of level 0s in a level n is the number of block
661 * pointers in an indirect block, raised to the power of level.
662 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
663 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
665 * Thus, the level n blkid is: offset /
666 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
667 * = offset / 2^(datablkshift + level *
668 * (indblkshift - SPA_BLKPTRSHIFT))
669 * = offset >> (datablkshift + level *
670 * (indblkshift - SPA_BLKPTRSHIFT))
672 return (offset
>> (dn
->dn_datablkshift
+ level
*
673 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
)));
675 ASSERT3U(offset
, <, dn
->dn_datablksz
);
681 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
683 dmu_buf_impl_t
*db
= vdb
;
685 mutex_enter(&db
->db_mtx
);
686 ASSERT3U(db
->db_state
, ==, DB_READ
);
688 * All reads are synchronous, so we must have a hold on the dbuf
690 ASSERT(refcount_count(&db
->db_holds
) > 0);
691 ASSERT(db
->db_buf
== NULL
);
692 ASSERT(db
->db
.db_data
== NULL
);
693 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
694 /* we were freed in flight; disregard any error */
695 arc_release(buf
, db
);
696 bzero(buf
->b_data
, db
->db
.db_size
);
698 db
->db_freed_in_flight
= FALSE
;
699 dbuf_set_data(db
, buf
);
700 db
->db_state
= DB_CACHED
;
701 } else if (zio
== NULL
|| zio
->io_error
== 0) {
702 dbuf_set_data(db
, buf
);
703 db
->db_state
= DB_CACHED
;
705 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
706 ASSERT3P(db
->db_buf
, ==, NULL
);
707 VERIFY(arc_buf_remove_ref(buf
, db
));
708 db
->db_state
= DB_UNCACHED
;
710 cv_broadcast(&db
->db_changed
);
711 dbuf_rele_and_unlock(db
, NULL
);
715 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
719 uint32_t aflags
= ARC_FLAG_NOWAIT
;
724 ASSERT(!refcount_is_zero(&db
->db_holds
));
725 /* We need the struct_rwlock to prevent db_blkptr from changing. */
726 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
727 ASSERT(MUTEX_HELD(&db
->db_mtx
));
728 ASSERT(db
->db_state
== DB_UNCACHED
);
729 ASSERT(db
->db_buf
== NULL
);
731 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
733 * The bonus length stored in the dnode may be less than
734 * the maximum available space in the bonus buffer.
736 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
737 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
739 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
740 db
->db
.db_data
= zio_buf_alloc(max_bonuslen
);
741 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
742 if (bonuslen
< max_bonuslen
)
743 bzero(db
->db
.db_data
, max_bonuslen
);
745 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
747 db
->db_state
= DB_CACHED
;
748 mutex_exit(&db
->db_mtx
);
753 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
754 * processes the delete record and clears the bp while we are waiting
755 * for the dn_mtx (resulting in a "no" from block_freed).
757 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
758 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
759 BP_IS_HOLE(db
->db_blkptr
)))) {
760 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
762 dbuf_set_data(db
, arc_buf_alloc(db
->db_objset
->os_spa
,
763 db
->db
.db_size
, db
, type
));
764 bzero(db
->db
.db_data
, db
->db
.db_size
);
766 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
767 BP_IS_HOLE(db
->db_blkptr
) &&
768 db
->db_blkptr
->blk_birth
!= 0) {
769 blkptr_t
*bps
= db
->db
.db_data
;
771 for (i
= 0; i
< ((1 <<
772 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
774 blkptr_t
*bp
= &bps
[i
];
775 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
776 1 << dn
->dn_indblkshift
);
778 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
780 BP_GET_LSIZE(db
->db_blkptr
));
781 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
783 BP_GET_LEVEL(db
->db_blkptr
) - 1);
784 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
788 db
->db_state
= DB_CACHED
;
789 mutex_exit(&db
->db_mtx
);
795 db
->db_state
= DB_READ
;
796 mutex_exit(&db
->db_mtx
);
798 if (DBUF_IS_L2CACHEABLE(db
))
799 aflags
|= ARC_FLAG_L2CACHE
;
800 if (DBUF_IS_L2COMPRESSIBLE(db
))
801 aflags
|= ARC_FLAG_L2COMPRESS
;
803 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
804 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
805 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
807 dbuf_add_ref(db
, NULL
);
809 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
810 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
811 (flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
814 return (SET_ERROR(err
));
818 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
821 boolean_t havepzio
= (zio
!= NULL
);
826 * We don't have to hold the mutex to check db_state because it
827 * can't be freed while we have a hold on the buffer.
829 ASSERT(!refcount_is_zero(&db
->db_holds
));
831 if (db
->db_state
== DB_NOFILL
)
832 return (SET_ERROR(EIO
));
836 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
837 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
839 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
840 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
841 DBUF_IS_CACHEABLE(db
);
843 mutex_enter(&db
->db_mtx
);
844 if (db
->db_state
== DB_CACHED
) {
845 mutex_exit(&db
->db_mtx
);
847 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
848 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
849 rw_exit(&dn
->dn_struct_rwlock
);
851 } else if (db
->db_state
== DB_UNCACHED
) {
852 spa_t
*spa
= dn
->dn_objset
->os_spa
;
855 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
857 err
= dbuf_read_impl(db
, zio
, flags
);
859 /* dbuf_read_impl has dropped db_mtx for us */
861 if (!err
&& prefetch
)
862 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
864 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
865 rw_exit(&dn
->dn_struct_rwlock
);
868 if (!err
&& !havepzio
)
872 * Another reader came in while the dbuf was in flight
873 * between UNCACHED and CACHED. Either a writer will finish
874 * writing the buffer (sending the dbuf to CACHED) or the
875 * first reader's request will reach the read_done callback
876 * and send the dbuf to CACHED. Otherwise, a failure
877 * occurred and the dbuf went to UNCACHED.
879 mutex_exit(&db
->db_mtx
);
881 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
882 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
883 rw_exit(&dn
->dn_struct_rwlock
);
886 /* Skip the wait per the caller's request. */
887 mutex_enter(&db
->db_mtx
);
888 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
889 while (db
->db_state
== DB_READ
||
890 db
->db_state
== DB_FILL
) {
891 ASSERT(db
->db_state
== DB_READ
||
892 (flags
& DB_RF_HAVESTRUCT
) == 0);
893 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
895 cv_wait(&db
->db_changed
, &db
->db_mtx
);
897 if (db
->db_state
== DB_UNCACHED
)
898 err
= SET_ERROR(EIO
);
900 mutex_exit(&db
->db_mtx
);
903 ASSERT(err
|| havepzio
|| db
->db_state
== DB_CACHED
);
908 dbuf_noread(dmu_buf_impl_t
*db
)
910 ASSERT(!refcount_is_zero(&db
->db_holds
));
911 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
912 mutex_enter(&db
->db_mtx
);
913 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
914 cv_wait(&db
->db_changed
, &db
->db_mtx
);
915 if (db
->db_state
== DB_UNCACHED
) {
916 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
917 spa_t
*spa
= db
->db_objset
->os_spa
;
919 ASSERT(db
->db_buf
== NULL
);
920 ASSERT(db
->db
.db_data
== NULL
);
921 dbuf_set_data(db
, arc_buf_alloc(spa
, db
->db
.db_size
, db
, type
));
922 db
->db_state
= DB_FILL
;
923 } else if (db
->db_state
== DB_NOFILL
) {
926 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
928 mutex_exit(&db
->db_mtx
);
932 * This is our just-in-time copy function. It makes a copy of
933 * buffers, that have been modified in a previous transaction
934 * group, before we modify them in the current active group.
936 * This function is used in two places: when we are dirtying a
937 * buffer for the first time in a txg, and when we are freeing
938 * a range in a dnode that includes this buffer.
940 * Note that when we are called from dbuf_free_range() we do
941 * not put a hold on the buffer, we just traverse the active
942 * dbuf list for the dnode.
945 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
947 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
949 ASSERT(MUTEX_HELD(&db
->db_mtx
));
950 ASSERT(db
->db
.db_data
!= NULL
);
951 ASSERT(db
->db_level
== 0);
952 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
955 (dr
->dt
.dl
.dr_data
!=
956 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
960 * If the last dirty record for this dbuf has not yet synced
961 * and its referencing the dbuf data, either:
962 * reset the reference to point to a new copy,
963 * or (if there a no active holders)
964 * just null out the current db_data pointer.
966 ASSERT(dr
->dr_txg
>= txg
- 2);
967 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
968 /* Note that the data bufs here are zio_bufs */
969 dnode_t
*dn
= DB_DNODE(db
);
970 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
971 dr
->dt
.dl
.dr_data
= zio_buf_alloc(bonuslen
);
972 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
973 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
974 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
975 int size
= db
->db
.db_size
;
976 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
977 spa_t
*spa
= db
->db_objset
->os_spa
;
979 dr
->dt
.dl
.dr_data
= arc_buf_alloc(spa
, size
, db
, type
);
980 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
987 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
989 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
990 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
991 uint64_t txg
= dr
->dr_txg
;
993 ASSERT(MUTEX_HELD(&db
->db_mtx
));
994 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
995 ASSERT(db
->db_level
== 0);
997 if (db
->db_blkid
== DMU_BONUS_BLKID
||
998 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1001 ASSERT(db
->db_data_pending
!= dr
);
1003 /* free this block */
1004 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1005 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1007 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1008 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1011 * Release the already-written buffer, so we leave it in
1012 * a consistent dirty state. Note that all callers are
1013 * modifying the buffer, so they will immediately do
1014 * another (redundant) arc_release(). Therefore, leave
1015 * the buf thawed to save the effort of freezing &
1016 * immediately re-thawing it.
1018 arc_release(dr
->dt
.dl
.dr_data
, db
);
1022 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1023 * data blocks in the free range, so that any future readers will find
1026 * This is a no-op if the dataset is in the middle of an incremental
1027 * receive; see comment below for details.
1030 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1033 dmu_buf_impl_t
*db_search
;
1034 dmu_buf_impl_t
*db
, *db_next
;
1035 uint64_t txg
= tx
->tx_txg
;
1037 boolean_t freespill
=
1038 (start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
);
1040 if (end_blkid
> dn
->dn_maxblkid
&& !freespill
)
1041 end_blkid
= dn
->dn_maxblkid
;
1042 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1044 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1045 db_search
->db_level
= 0;
1046 db_search
->db_blkid
= start_blkid
;
1047 db_search
->db_state
= DB_SEARCH
;
1049 mutex_enter(&dn
->dn_dbufs_mtx
);
1050 if (start_blkid
>= dn
->dn_unlisted_l0_blkid
&& !freespill
) {
1051 /* There can't be any dbufs in this range; no need to search. */
1053 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1054 ASSERT3P(db
, ==, NULL
);
1055 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1056 ASSERT(db
== NULL
|| db
->db_level
> 0);
1059 } else if (dmu_objset_is_receiving(dn
->dn_objset
)) {
1061 * If we are receiving, we expect there to be no dbufs in
1062 * the range to be freed, because receive modifies each
1063 * block at most once, and in offset order. If this is
1064 * not the case, it can lead to performance problems,
1065 * so note that we unexpectedly took the slow path.
1067 atomic_inc_64(&zfs_free_range_recv_miss
);
1070 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1071 ASSERT3P(db
, ==, NULL
);
1072 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1074 for (; db
!= NULL
; db
= db_next
) {
1075 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1076 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1078 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1081 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1083 /* found a level 0 buffer in the range */
1084 mutex_enter(&db
->db_mtx
);
1085 if (dbuf_undirty(db
, tx
)) {
1086 /* mutex has been dropped and dbuf destroyed */
1090 if (db
->db_state
== DB_UNCACHED
||
1091 db
->db_state
== DB_NOFILL
||
1092 db
->db_state
== DB_EVICTING
) {
1093 ASSERT(db
->db
.db_data
== NULL
);
1094 mutex_exit(&db
->db_mtx
);
1097 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1098 /* will be handled in dbuf_read_done or dbuf_rele */
1099 db
->db_freed_in_flight
= TRUE
;
1100 mutex_exit(&db
->db_mtx
);
1103 if (refcount_count(&db
->db_holds
) == 0) {
1108 /* The dbuf is referenced */
1110 if (db
->db_last_dirty
!= NULL
) {
1111 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1113 if (dr
->dr_txg
== txg
) {
1115 * This buffer is "in-use", re-adjust the file
1116 * size to reflect that this buffer may
1117 * contain new data when we sync.
1119 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1120 db
->db_blkid
> dn
->dn_maxblkid
)
1121 dn
->dn_maxblkid
= db
->db_blkid
;
1122 dbuf_unoverride(dr
);
1125 * This dbuf is not dirty in the open context.
1126 * Either uncache it (if its not referenced in
1127 * the open context) or reset its contents to
1130 dbuf_fix_old_data(db
, txg
);
1133 /* clear the contents if its cached */
1134 if (db
->db_state
== DB_CACHED
) {
1135 ASSERT(db
->db
.db_data
!= NULL
);
1136 arc_release(db
->db_buf
, db
);
1137 bzero(db
->db
.db_data
, db
->db
.db_size
);
1138 arc_buf_freeze(db
->db_buf
);
1141 mutex_exit(&db
->db_mtx
);
1145 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1146 mutex_exit(&dn
->dn_dbufs_mtx
);
1150 dbuf_block_freeable(dmu_buf_impl_t
*db
)
1152 dsl_dataset_t
*ds
= db
->db_objset
->os_dsl_dataset
;
1153 uint64_t birth_txg
= 0;
1156 * We don't need any locking to protect db_blkptr:
1157 * If it's syncing, then db_last_dirty will be set
1158 * so we'll ignore db_blkptr.
1160 * This logic ensures that only block births for
1161 * filled blocks are considered.
1163 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1164 if (db
->db_last_dirty
&& (db
->db_blkptr
== NULL
||
1165 !BP_IS_HOLE(db
->db_blkptr
))) {
1166 birth_txg
= db
->db_last_dirty
->dr_txg
;
1167 } else if (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1168 birth_txg
= db
->db_blkptr
->blk_birth
;
1172 * If this block don't exist or is in a snapshot, it can't be freed.
1173 * Don't pass the bp to dsl_dataset_block_freeable() since we
1174 * are holding the db_mtx lock and might deadlock if we are
1175 * prefetching a dedup-ed block.
1178 return (ds
== NULL
||
1179 dsl_dataset_block_freeable(ds
, NULL
, birth_txg
));
1185 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1187 arc_buf_t
*buf
, *obuf
;
1188 int osize
= db
->db
.db_size
;
1189 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1192 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1197 /* XXX does *this* func really need the lock? */
1198 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1201 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1202 * is OK, because there can be no other references to the db
1203 * when we are changing its size, so no concurrent DB_FILL can
1207 * XXX we should be doing a dbuf_read, checking the return
1208 * value and returning that up to our callers
1210 dmu_buf_will_dirty(&db
->db
, tx
);
1212 /* create the data buffer for the new block */
1213 buf
= arc_buf_alloc(dn
->dn_objset
->os_spa
, size
, db
, type
);
1215 /* copy old block data to the new block */
1217 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1218 /* zero the remainder */
1220 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1222 mutex_enter(&db
->db_mtx
);
1223 dbuf_set_data(db
, buf
);
1224 VERIFY(arc_buf_remove_ref(obuf
, db
));
1225 db
->db
.db_size
= size
;
1227 if (db
->db_level
== 0) {
1228 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1229 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1231 mutex_exit(&db
->db_mtx
);
1233 dnode_willuse_space(dn
, size
-osize
, tx
);
1238 dbuf_release_bp(dmu_buf_impl_t
*db
)
1240 ASSERTV(objset_t
*os
= db
->db_objset
);
1242 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1243 ASSERT(arc_released(os
->os_phys_buf
) ||
1244 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1245 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1247 (void) arc_release(db
->db_buf
, db
);
1251 * We already have a dirty record for this TXG, and we are being
1255 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1257 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1259 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1261 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1263 * If this buffer has already been written out,
1264 * we now need to reset its state.
1266 dbuf_unoverride(dr
);
1267 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1268 db
->db_state
!= DB_NOFILL
) {
1269 /* Already released on initial dirty, so just thaw. */
1270 ASSERT(arc_released(db
->db_buf
));
1271 arc_buf_thaw(db
->db_buf
);
1276 dbuf_dirty_record_t
*
1277 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1281 dbuf_dirty_record_t
**drp
, *dr
;
1282 int drop_struct_lock
= FALSE
;
1283 boolean_t do_free_accounting
= B_FALSE
;
1284 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1286 ASSERT(tx
->tx_txg
!= 0);
1287 ASSERT(!refcount_is_zero(&db
->db_holds
));
1288 DMU_TX_DIRTY_BUF(tx
, db
);
1293 * Shouldn't dirty a regular buffer in syncing context. Private
1294 * objects may be dirtied in syncing context, but only if they
1295 * were already pre-dirtied in open context.
1297 ASSERT(!dmu_tx_is_syncing(tx
) ||
1298 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1299 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1300 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1302 * We make this assert for private objects as well, but after we
1303 * check if we're already dirty. They are allowed to re-dirty
1304 * in syncing context.
1306 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1307 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1308 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1310 mutex_enter(&db
->db_mtx
);
1312 * XXX make this true for indirects too? The problem is that
1313 * transactions created with dmu_tx_create_assigned() from
1314 * syncing context don't bother holding ahead.
1316 ASSERT(db
->db_level
!= 0 ||
1317 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1318 db
->db_state
== DB_NOFILL
);
1320 mutex_enter(&dn
->dn_mtx
);
1322 * Don't set dirtyctx to SYNC if we're just modifying this as we
1323 * initialize the objset.
1325 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
&&
1326 !BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1328 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1329 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1330 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1332 mutex_exit(&dn
->dn_mtx
);
1334 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1335 dn
->dn_have_spill
= B_TRUE
;
1338 * If this buffer is already dirty, we're done.
1340 drp
= &db
->db_last_dirty
;
1341 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1342 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1343 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1345 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1349 mutex_exit(&db
->db_mtx
);
1354 * Only valid if not already dirty.
1356 ASSERT(dn
->dn_object
== 0 ||
1357 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1358 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1360 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1361 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1362 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1363 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1364 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1365 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1368 * We should only be dirtying in syncing context if it's the
1369 * mos or we're initializing the os or it's a special object.
1370 * However, we are allowed to dirty in syncing context provided
1371 * we already dirtied it in open context. Hence we must make
1372 * this assertion only if we're not already dirty.
1375 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1376 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1377 ASSERT(db
->db
.db_size
!= 0);
1379 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1381 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1383 * Update the accounting.
1384 * Note: we delay "free accounting" until after we drop
1385 * the db_mtx. This keeps us from grabbing other locks
1386 * (and possibly deadlocking) in bp_get_dsize() while
1387 * also holding the db_mtx.
1389 dnode_willuse_space(dn
, db
->db
.db_size
, tx
);
1390 do_free_accounting
= dbuf_block_freeable(db
);
1394 * If this buffer is dirty in an old transaction group we need
1395 * to make a copy of it so that the changes we make in this
1396 * transaction group won't leak out when we sync the older txg.
1398 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1399 list_link_init(&dr
->dr_dirty_node
);
1400 if (db
->db_level
== 0) {
1401 void *data_old
= db
->db_buf
;
1403 if (db
->db_state
!= DB_NOFILL
) {
1404 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1405 dbuf_fix_old_data(db
, tx
->tx_txg
);
1406 data_old
= db
->db
.db_data
;
1407 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1409 * Release the data buffer from the cache so
1410 * that we can modify it without impacting
1411 * possible other users of this cached data
1412 * block. Note that indirect blocks and
1413 * private objects are not released until the
1414 * syncing state (since they are only modified
1417 arc_release(db
->db_buf
, db
);
1418 dbuf_fix_old_data(db
, tx
->tx_txg
);
1419 data_old
= db
->db_buf
;
1421 ASSERT(data_old
!= NULL
);
1423 dr
->dt
.dl
.dr_data
= data_old
;
1425 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1426 list_create(&dr
->dt
.di
.dr_children
,
1427 sizeof (dbuf_dirty_record_t
),
1428 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1430 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1431 dr
->dr_accounted
= db
->db
.db_size
;
1433 dr
->dr_txg
= tx
->tx_txg
;
1438 * We could have been freed_in_flight between the dbuf_noread
1439 * and dbuf_dirty. We win, as though the dbuf_noread() had
1440 * happened after the free.
1442 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1443 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1444 mutex_enter(&dn
->dn_mtx
);
1445 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1446 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1449 mutex_exit(&dn
->dn_mtx
);
1450 db
->db_freed_in_flight
= FALSE
;
1454 * This buffer is now part of this txg
1456 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1457 db
->db_dirtycnt
+= 1;
1458 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1460 mutex_exit(&db
->db_mtx
);
1462 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1463 db
->db_blkid
== DMU_SPILL_BLKID
) {
1464 mutex_enter(&dn
->dn_mtx
);
1465 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1466 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1467 mutex_exit(&dn
->dn_mtx
);
1468 dnode_setdirty(dn
, tx
);
1474 * The dn_struct_rwlock prevents db_blkptr from changing
1475 * due to a write from syncing context completing
1476 * while we are running, so we want to acquire it before
1477 * looking at db_blkptr.
1479 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1480 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1481 drop_struct_lock
= TRUE
;
1484 if (do_free_accounting
) {
1485 blkptr_t
*bp
= db
->db_blkptr
;
1486 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1487 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1489 * This is only a guess -- if the dbuf is dirty
1490 * in a previous txg, we don't know how much
1491 * space it will use on disk yet. We should
1492 * really have the struct_rwlock to access
1493 * db_blkptr, but since this is just a guess,
1494 * it's OK if we get an odd answer.
1496 ddt_prefetch(os
->os_spa
, bp
);
1497 dnode_willuse_space(dn
, -willfree
, tx
);
1500 if (db
->db_level
== 0) {
1501 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1502 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1505 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1506 dmu_buf_impl_t
*parent
= db
->db_parent
;
1507 dbuf_dirty_record_t
*di
;
1508 int parent_held
= FALSE
;
1510 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1511 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1513 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1514 db
->db_blkid
>> epbs
, FTAG
);
1515 ASSERT(parent
!= NULL
);
1518 if (drop_struct_lock
)
1519 rw_exit(&dn
->dn_struct_rwlock
);
1520 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1521 di
= dbuf_dirty(parent
, tx
);
1523 dbuf_rele(parent
, FTAG
);
1525 mutex_enter(&db
->db_mtx
);
1527 * Since we've dropped the mutex, it's possible that
1528 * dbuf_undirty() might have changed this out from under us.
1530 if (db
->db_last_dirty
== dr
||
1531 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1532 mutex_enter(&di
->dt
.di
.dr_mtx
);
1533 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1534 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1535 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1536 mutex_exit(&di
->dt
.di
.dr_mtx
);
1539 mutex_exit(&db
->db_mtx
);
1541 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1542 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1543 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1544 mutex_enter(&dn
->dn_mtx
);
1545 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1546 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1547 mutex_exit(&dn
->dn_mtx
);
1548 if (drop_struct_lock
)
1549 rw_exit(&dn
->dn_struct_rwlock
);
1552 dnode_setdirty(dn
, tx
);
1558 * Undirty a buffer in the transaction group referenced by the given
1559 * transaction. Return whether this evicted the dbuf.
1562 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1565 uint64_t txg
= tx
->tx_txg
;
1566 dbuf_dirty_record_t
*dr
, **drp
;
1571 * Due to our use of dn_nlevels below, this can only be called
1572 * in open context, unless we are operating on the MOS.
1573 * From syncing context, dn_nlevels may be different from the
1574 * dn_nlevels used when dbuf was dirtied.
1576 ASSERT(db
->db_objset
==
1577 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1578 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1579 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1580 ASSERT0(db
->db_level
);
1581 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1584 * If this buffer is not dirty, we're done.
1586 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1587 if (dr
->dr_txg
<= txg
)
1589 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1591 ASSERT(dr
->dr_txg
== txg
);
1592 ASSERT(dr
->dr_dbuf
== db
);
1597 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1599 ASSERT(db
->db
.db_size
!= 0);
1601 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1602 dr
->dr_accounted
, txg
);
1607 * Note that there are three places in dbuf_dirty()
1608 * where this dirty record may be put on a list.
1609 * Make sure to do a list_remove corresponding to
1610 * every one of those list_insert calls.
1612 if (dr
->dr_parent
) {
1613 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1614 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1615 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1616 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1617 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1618 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1619 mutex_enter(&dn
->dn_mtx
);
1620 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1621 mutex_exit(&dn
->dn_mtx
);
1625 if (db
->db_state
!= DB_NOFILL
) {
1626 dbuf_unoverride(dr
);
1628 ASSERT(db
->db_buf
!= NULL
);
1629 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1630 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1631 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
, db
));
1634 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1636 ASSERT(db
->db_dirtycnt
> 0);
1637 db
->db_dirtycnt
-= 1;
1639 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1640 arc_buf_t
*buf
= db
->db_buf
;
1642 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(buf
));
1643 dbuf_clear_data(db
);
1644 VERIFY(arc_buf_remove_ref(buf
, db
));
1653 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1655 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1656 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1657 dbuf_dirty_record_t
*dr
;
1659 ASSERT(tx
->tx_txg
!= 0);
1660 ASSERT(!refcount_is_zero(&db
->db_holds
));
1663 * Quick check for dirtyness. For already dirty blocks, this
1664 * reduces runtime of this function by >90%, and overall performance
1665 * by 50% for some workloads (e.g. file deletion with indirect blocks
1668 mutex_enter(&db
->db_mtx
);
1670 for (dr
= db
->db_last_dirty
;
1671 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1673 * It's possible that it is already dirty but not cached,
1674 * because there are some calls to dbuf_dirty() that don't
1675 * go through dmu_buf_will_dirty().
1677 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1678 /* This dbuf is already dirty and cached. */
1680 mutex_exit(&db
->db_mtx
);
1684 mutex_exit(&db
->db_mtx
);
1687 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1688 rf
|= DB_RF_HAVESTRUCT
;
1690 (void) dbuf_read(db
, NULL
, rf
);
1691 (void) dbuf_dirty(db
, tx
);
1695 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1697 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1699 db
->db_state
= DB_NOFILL
;
1701 dmu_buf_will_fill(db_fake
, tx
);
1705 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1707 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1709 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1710 ASSERT(tx
->tx_txg
!= 0);
1711 ASSERT(db
->db_level
== 0);
1712 ASSERT(!refcount_is_zero(&db
->db_holds
));
1714 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1715 dmu_tx_private_ok(tx
));
1718 (void) dbuf_dirty(db
, tx
);
1721 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1724 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1726 mutex_enter(&db
->db_mtx
);
1729 if (db
->db_state
== DB_FILL
) {
1730 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1731 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1732 /* we were freed while filling */
1733 /* XXX dbuf_undirty? */
1734 bzero(db
->db
.db_data
, db
->db
.db_size
);
1735 db
->db_freed_in_flight
= FALSE
;
1737 db
->db_state
= DB_CACHED
;
1738 cv_broadcast(&db
->db_changed
);
1740 mutex_exit(&db
->db_mtx
);
1744 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
1745 bp_embedded_type_t etype
, enum zio_compress comp
,
1746 int uncompressed_size
, int compressed_size
, int byteorder
,
1749 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
1750 struct dirty_leaf
*dl
;
1751 dmu_object_type_t type
;
1753 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
1754 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
1755 SPA_FEATURE_EMBEDDED_DATA
));
1759 type
= DB_DNODE(db
)->dn_type
;
1762 ASSERT0(db
->db_level
);
1763 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1765 dmu_buf_will_not_fill(dbuf
, tx
);
1767 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1768 dl
= &db
->db_last_dirty
->dt
.dl
;
1769 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
1770 data
, comp
, uncompressed_size
, compressed_size
);
1771 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
1772 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
1773 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
1774 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
1776 dl
->dr_override_state
= DR_OVERRIDDEN
;
1777 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
1781 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1782 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1785 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1787 ASSERT(!refcount_is_zero(&db
->db_holds
));
1788 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1789 ASSERT(db
->db_level
== 0);
1790 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1791 ASSERT(buf
!= NULL
);
1792 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1793 ASSERT(tx
->tx_txg
!= 0);
1795 arc_return_buf(buf
, db
);
1796 ASSERT(arc_released(buf
));
1798 mutex_enter(&db
->db_mtx
);
1800 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1801 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1803 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1805 if (db
->db_state
== DB_CACHED
&&
1806 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1807 mutex_exit(&db
->db_mtx
);
1808 (void) dbuf_dirty(db
, tx
);
1809 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1810 VERIFY(arc_buf_remove_ref(buf
, db
));
1811 xuio_stat_wbuf_copied();
1815 xuio_stat_wbuf_nocopy();
1816 if (db
->db_state
== DB_CACHED
) {
1817 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1819 ASSERT(db
->db_buf
!= NULL
);
1820 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
1821 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
1822 if (!arc_released(db
->db_buf
)) {
1823 ASSERT(dr
->dt
.dl
.dr_override_state
==
1825 arc_release(db
->db_buf
, db
);
1827 dr
->dt
.dl
.dr_data
= buf
;
1828 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1829 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
1830 arc_release(db
->db_buf
, db
);
1831 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1835 ASSERT(db
->db_buf
== NULL
);
1836 dbuf_set_data(db
, buf
);
1837 db
->db_state
= DB_FILL
;
1838 mutex_exit(&db
->db_mtx
);
1839 (void) dbuf_dirty(db
, tx
);
1840 dmu_buf_fill_done(&db
->db
, tx
);
1844 * "Clear" the contents of this dbuf. This will mark the dbuf
1845 * EVICTING and clear *most* of its references. Unfortunately,
1846 * when we are not holding the dn_dbufs_mtx, we can't clear the
1847 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1848 * in this case. For callers from the DMU we will usually see:
1849 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1850 * For the arc callback, we will usually see:
1851 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1852 * Sometimes, though, we will get a mix of these two:
1853 * DMU: dbuf_clear()->arc_clear_callback()
1854 * ARC: dbuf_do_evict()->dbuf_destroy()
1856 * This routine will dissociate the dbuf from the arc, by calling
1857 * arc_clear_callback(), but will not evict the data from the ARC.
1860 dbuf_clear(dmu_buf_impl_t
*db
)
1863 dmu_buf_impl_t
*parent
= db
->db_parent
;
1864 dmu_buf_impl_t
*dndb
;
1865 boolean_t dbuf_gone
= B_FALSE
;
1867 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1868 ASSERT(refcount_is_zero(&db
->db_holds
));
1870 dbuf_evict_user(db
);
1872 if (db
->db_state
== DB_CACHED
) {
1873 ASSERT(db
->db
.db_data
!= NULL
);
1874 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1875 int slots
= DB_DNODE(db
)->dn_num_slots
;
1876 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
1877 zio_buf_free(db
->db
.db_data
, bonuslen
);
1878 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
1880 db
->db
.db_data
= NULL
;
1881 db
->db_state
= DB_UNCACHED
;
1884 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1885 ASSERT(db
->db_data_pending
== NULL
);
1887 db
->db_state
= DB_EVICTING
;
1888 db
->db_blkptr
= NULL
;
1893 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& MUTEX_HELD(&dn
->dn_dbufs_mtx
)) {
1894 avl_remove(&dn
->dn_dbufs
, db
);
1895 atomic_dec_32(&dn
->dn_dbufs_count
);
1899 * Decrementing the dbuf count means that the hold corresponding
1900 * to the removed dbuf is no longer discounted in dnode_move(),
1901 * so the dnode cannot be moved until after we release the hold.
1902 * The membar_producer() ensures visibility of the decremented
1903 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1907 db
->db_dnode_handle
= NULL
;
1913 dbuf_gone
= arc_clear_callback(db
->db_buf
);
1916 mutex_exit(&db
->db_mtx
);
1919 * If this dbuf is referenced from an indirect dbuf,
1920 * decrement the ref count on the indirect dbuf.
1922 if (parent
&& parent
!= dndb
)
1923 dbuf_rele(parent
, db
);
1927 * Note: While bpp will always be updated if the function returns success,
1928 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1929 * this happens when the dnode is the meta-dnode, or a userused or groupused
1932 __attribute__((always_inline
))
1934 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
1935 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
1942 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1944 if (blkid
== DMU_SPILL_BLKID
) {
1945 mutex_enter(&dn
->dn_mtx
);
1946 if (dn
->dn_have_spill
&&
1947 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1948 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
1951 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1952 *parentp
= dn
->dn_dbuf
;
1953 mutex_exit(&dn
->dn_mtx
);
1958 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
1959 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1961 ASSERT3U(level
* epbs
, <, 64);
1962 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1964 * This assertion shouldn't trip as long as the max indirect block size
1965 * is less than 1M. The reason for this is that up to that point,
1966 * the number of levels required to address an entire object with blocks
1967 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
1968 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
1969 * (i.e. we can address the entire object), objects will all use at most
1970 * N-1 levels and the assertion won't overflow. However, once epbs is
1971 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
1972 * enough to address an entire object, so objects will have 5 levels,
1973 * but then this assertion will overflow.
1975 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
1976 * need to redo this logic to handle overflows.
1978 ASSERT(level
>= nlevels
||
1979 ((nlevels
- level
- 1) * epbs
) +
1980 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
1981 if (level
>= nlevels
||
1982 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
1983 ((nlevels
- level
- 1) * epbs
)) ||
1985 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
1986 /* the buffer has no parent yet */
1987 return (SET_ERROR(ENOENT
));
1988 } else if (level
< nlevels
-1) {
1989 /* this block is referenced from an indirect block */
1992 err
= dbuf_hold_impl(dn
, level
+1,
1993 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
1995 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
1996 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
1997 parentp
, dh
->dh_depth
+ 1);
1998 err
= __dbuf_hold_impl(dh
+ 1);
2002 err
= dbuf_read(*parentp
, NULL
,
2003 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2005 dbuf_rele(*parentp
, NULL
);
2009 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2010 (blkid
& ((1ULL << epbs
) - 1));
2011 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2012 ASSERT(BP_IS_HOLE(*bpp
));
2015 /* the block is referenced from the dnode */
2016 ASSERT3U(level
, ==, nlevels
-1);
2017 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2018 blkid
< dn
->dn_phys
->dn_nblkptr
);
2020 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2021 *parentp
= dn
->dn_dbuf
;
2023 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2028 static dmu_buf_impl_t
*
2029 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2030 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2032 objset_t
*os
= dn
->dn_objset
;
2033 dmu_buf_impl_t
*db
, *odb
;
2035 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2036 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2038 db
= kmem_cache_alloc(dbuf_cache
, KM_SLEEP
);
2041 db
->db
.db_object
= dn
->dn_object
;
2042 db
->db_level
= level
;
2043 db
->db_blkid
= blkid
;
2044 db
->db_last_dirty
= NULL
;
2045 db
->db_dirtycnt
= 0;
2046 db
->db_dnode_handle
= dn
->dn_handle
;
2047 db
->db_parent
= parent
;
2048 db
->db_blkptr
= blkptr
;
2051 db
->db_user_immediate_evict
= FALSE
;
2052 db
->db_freed_in_flight
= FALSE
;
2053 db
->db_pending_evict
= FALSE
;
2055 if (blkid
== DMU_BONUS_BLKID
) {
2056 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2057 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2058 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2059 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2060 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2061 db
->db_state
= DB_UNCACHED
;
2062 /* the bonus dbuf is not placed in the hash table */
2063 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2065 } else if (blkid
== DMU_SPILL_BLKID
) {
2066 db
->db
.db_size
= (blkptr
!= NULL
) ?
2067 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2068 db
->db
.db_offset
= 0;
2071 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2072 db
->db
.db_size
= blocksize
;
2073 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2077 * Hold the dn_dbufs_mtx while we get the new dbuf
2078 * in the hash table *and* added to the dbufs list.
2079 * This prevents a possible deadlock with someone
2080 * trying to look up this dbuf before its added to the
2083 mutex_enter(&dn
->dn_dbufs_mtx
);
2084 db
->db_state
= DB_EVICTING
;
2085 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2086 /* someone else inserted it first */
2087 kmem_cache_free(dbuf_cache
, db
);
2088 mutex_exit(&dn
->dn_dbufs_mtx
);
2091 avl_add(&dn
->dn_dbufs
, db
);
2092 if (db
->db_level
== 0 && db
->db_blkid
>=
2093 dn
->dn_unlisted_l0_blkid
)
2094 dn
->dn_unlisted_l0_blkid
= db
->db_blkid
+ 1;
2095 db
->db_state
= DB_UNCACHED
;
2096 mutex_exit(&dn
->dn_dbufs_mtx
);
2097 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2099 if (parent
&& parent
!= dn
->dn_dbuf
)
2100 dbuf_add_ref(parent
, db
);
2102 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2103 refcount_count(&dn
->dn_holds
) > 0);
2104 (void) refcount_add(&dn
->dn_holds
, db
);
2105 atomic_inc_32(&dn
->dn_dbufs_count
);
2107 dprintf_dbuf(db
, "db=%p\n", db
);
2113 dbuf_do_evict(void *private)
2115 dmu_buf_impl_t
*db
= private;
2117 if (!MUTEX_HELD(&db
->db_mtx
))
2118 mutex_enter(&db
->db_mtx
);
2120 ASSERT(refcount_is_zero(&db
->db_holds
));
2122 if (db
->db_state
!= DB_EVICTING
) {
2123 ASSERT(db
->db_state
== DB_CACHED
);
2128 mutex_exit(&db
->db_mtx
);
2135 dbuf_destroy(dmu_buf_impl_t
*db
)
2137 ASSERT(refcount_is_zero(&db
->db_holds
));
2139 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2141 * If this dbuf is still on the dn_dbufs list,
2142 * remove it from that list.
2144 if (db
->db_dnode_handle
!= NULL
) {
2149 mutex_enter(&dn
->dn_dbufs_mtx
);
2150 avl_remove(&dn
->dn_dbufs
, db
);
2151 atomic_dec_32(&dn
->dn_dbufs_count
);
2152 mutex_exit(&dn
->dn_dbufs_mtx
);
2155 * Decrementing the dbuf count means that the hold
2156 * corresponding to the removed dbuf is no longer
2157 * discounted in dnode_move(), so the dnode cannot be
2158 * moved until after we release the hold.
2161 db
->db_dnode_handle
= NULL
;
2163 dbuf_hash_remove(db
);
2165 db
->db_parent
= NULL
;
2168 ASSERT(db
->db
.db_data
== NULL
);
2169 ASSERT(db
->db_hash_next
== NULL
);
2170 ASSERT(db
->db_blkptr
== NULL
);
2171 ASSERT(db
->db_data_pending
== NULL
);
2173 kmem_cache_free(dbuf_cache
, db
);
2174 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2177 typedef struct dbuf_prefetch_arg
{
2178 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2179 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2180 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2181 int dpa_curlevel
; /* The current level that we're reading */
2182 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2183 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2184 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2185 } dbuf_prefetch_arg_t
;
2188 * Actually issue the prefetch read for the block given.
2191 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2194 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2197 aflags
= dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2199 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2200 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2201 ASSERT(dpa
->dpa_zio
!= NULL
);
2202 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2203 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2204 &aflags
, &dpa
->dpa_zb
);
2208 * Called when an indirect block above our prefetch target is read in. This
2209 * will either read in the next indirect block down the tree or issue the actual
2210 * prefetch if the next block down is our target.
2213 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2215 dbuf_prefetch_arg_t
*dpa
= private;
2219 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2220 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2222 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2223 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2224 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2227 dpa
->dpa_curlevel
--;
2229 nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2230 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2231 bp
= ((blkptr_t
*)abuf
->b_data
) +
2232 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2233 if (BP_IS_HOLE(bp
) || (zio
!= NULL
&& zio
->io_error
!= 0)) {
2234 kmem_free(dpa
, sizeof (*dpa
));
2235 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2236 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2237 dbuf_issue_final_prefetch(dpa
, bp
);
2238 kmem_free(dpa
, sizeof (*dpa
));
2240 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2241 zbookmark_phys_t zb
;
2243 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2245 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2246 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2248 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2249 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2250 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2253 (void) arc_buf_remove_ref(abuf
, private);
2257 * Issue prefetch reads for the given block on the given level. If the indirect
2258 * blocks above that block are not in memory, we will read them in
2259 * asynchronously. As a result, this call never blocks waiting for a read to
2263 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2267 int epbs
, nlevels
, curlevel
;
2271 dbuf_prefetch_arg_t
*dpa
;
2274 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2275 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2277 if (blkid
> dn
->dn_maxblkid
)
2280 if (dnode_block_freed(dn
, blkid
))
2284 * This dnode hasn't been written to disk yet, so there's nothing to
2287 nlevels
= dn
->dn_phys
->dn_nlevels
;
2288 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2291 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2292 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2295 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2298 mutex_exit(&db
->db_mtx
);
2300 * This dbuf already exists. It is either CACHED, or
2301 * (we assume) about to be read or filled.
2307 * Find the closest ancestor (indirect block) of the target block
2308 * that is present in the cache. In this indirect block, we will
2309 * find the bp that is at curlevel, curblkid.
2313 while (curlevel
< nlevels
- 1) {
2314 int parent_level
= curlevel
+ 1;
2315 uint64_t parent_blkid
= curblkid
>> epbs
;
2318 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2319 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2320 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2321 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2322 dbuf_rele(db
, FTAG
);
2326 curlevel
= parent_level
;
2327 curblkid
= parent_blkid
;
2330 if (curlevel
== nlevels
- 1) {
2331 /* No cached indirect blocks found. */
2332 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2333 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2335 if (BP_IS_HOLE(&bp
))
2338 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2340 pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2343 dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2344 ds
= dn
->dn_objset
->os_dsl_dataset
;
2345 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2346 dn
->dn_object
, level
, blkid
);
2347 dpa
->dpa_curlevel
= curlevel
;
2348 dpa
->dpa_prio
= prio
;
2349 dpa
->dpa_aflags
= aflags
;
2350 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2351 dpa
->dpa_epbs
= epbs
;
2355 * If we have the indirect just above us, no need to do the asynchronous
2356 * prefetch chain; we'll just run the last step ourselves. If we're at
2357 * a higher level, though, we want to issue the prefetches for all the
2358 * indirect blocks asynchronously, so we can go on with whatever we were
2361 if (curlevel
== level
) {
2362 ASSERT3U(curblkid
, ==, blkid
);
2363 dbuf_issue_final_prefetch(dpa
, &bp
);
2364 kmem_free(dpa
, sizeof (*dpa
));
2366 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2367 zbookmark_phys_t zb
;
2369 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2370 dn
->dn_object
, curlevel
, curblkid
);
2371 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2372 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2373 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2377 * We use pio here instead of dpa_zio since it's possible that
2378 * dpa may have already been freed.
2383 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2386 * Returns with db_holds incremented, and db_mtx not held.
2387 * Note: dn_struct_rwlock must be held.
2390 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2392 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2393 dh
->dh_parent
= NULL
;
2395 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2396 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2397 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2399 *(dh
->dh_dbp
) = NULL
;
2401 /* dbuf_find() returns with db_mtx held */
2402 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2403 dh
->dh_level
, dh
->dh_blkid
);
2405 if (dh
->dh_db
== NULL
) {
2408 if (dh
->dh_fail_uncached
)
2409 return (SET_ERROR(ENOENT
));
2411 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2412 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2413 dh
->dh_fail_sparse
, &dh
->dh_parent
,
2415 if (dh
->dh_fail_sparse
) {
2416 if (dh
->dh_err
== 0 &&
2417 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2418 dh
->dh_err
= SET_ERROR(ENOENT
);
2421 dbuf_rele(dh
->dh_parent
, NULL
);
2422 return (dh
->dh_err
);
2425 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2426 return (dh
->dh_err
);
2427 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2428 dh
->dh_parent
, dh
->dh_bp
);
2431 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2432 mutex_exit(&dh
->dh_db
->db_mtx
);
2433 return (SET_ERROR(ENOENT
));
2436 if (dh
->dh_db
->db_buf
&& refcount_is_zero(&dh
->dh_db
->db_holds
)) {
2437 arc_buf_add_ref(dh
->dh_db
->db_buf
, dh
->dh_db
);
2438 if (dh
->dh_db
->db_buf
->b_data
== NULL
) {
2439 dbuf_clear(dh
->dh_db
);
2440 if (dh
->dh_parent
) {
2441 dbuf_rele(dh
->dh_parent
, NULL
);
2442 dh
->dh_parent
= NULL
;
2446 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2449 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2452 * If this buffer is currently syncing out, and we are are
2453 * still referencing it from db_data, we need to make a copy
2454 * of it in case we decide we want to dirty it again in this txg.
2456 if (dh
->dh_db
->db_level
== 0 &&
2457 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2458 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2459 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2460 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2462 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2463 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2465 dbuf_set_data(dh
->dh_db
,
2466 arc_buf_alloc(dh
->dh_dn
->dn_objset
->os_spa
,
2467 dh
->dh_db
->db
.db_size
, dh
->dh_db
, dh
->dh_type
));
2468 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2469 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2473 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2474 DBUF_VERIFY(dh
->dh_db
);
2475 mutex_exit(&dh
->dh_db
->db_mtx
);
2477 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2479 dbuf_rele(dh
->dh_parent
, NULL
);
2481 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2482 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2483 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2484 *(dh
->dh_dbp
) = dh
->dh_db
;
2490 * The following code preserves the recursive function dbuf_hold_impl()
2491 * but moves the local variables AND function arguments to the heap to
2492 * minimize the stack frame size. Enough space is initially allocated
2493 * on the stack for 20 levels of recursion.
2496 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2497 boolean_t fail_sparse
, boolean_t fail_uncached
,
2498 void *tag
, dmu_buf_impl_t
**dbp
)
2500 struct dbuf_hold_impl_data
*dh
;
2503 dh
= kmem_alloc(sizeof (struct dbuf_hold_impl_data
) *
2504 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2505 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2506 fail_uncached
, tag
, dbp
, 0);
2508 error
= __dbuf_hold_impl(dh
);
2510 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2511 DBUF_HOLD_IMPL_MAX_DEPTH
);
2517 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2518 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2519 boolean_t fail_sparse
, boolean_t fail_uncached
,
2520 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2523 dh
->dh_level
= level
;
2524 dh
->dh_blkid
= blkid
;
2526 dh
->dh_fail_sparse
= fail_sparse
;
2527 dh
->dh_fail_uncached
= fail_uncached
;
2533 dh
->dh_parent
= NULL
;
2539 dh
->dh_depth
= depth
;
2543 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2545 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2549 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2552 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2553 return (err
? NULL
: db
);
2557 dbuf_create_bonus(dnode_t
*dn
)
2559 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2561 ASSERT(dn
->dn_bonus
== NULL
);
2562 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2566 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2568 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2571 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2572 return (SET_ERROR(ENOTSUP
));
2574 blksz
= SPA_MINBLOCKSIZE
;
2575 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2576 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2580 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2581 dbuf_new_size(db
, blksz
, tx
);
2582 rw_exit(&dn
->dn_struct_rwlock
);
2589 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2591 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2594 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2596 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2598 VERIFY(refcount_add(&db
->db_holds
, tag
) > 1);
2601 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2603 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2606 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2607 dmu_buf_impl_t
*found_db
;
2608 boolean_t result
= B_FALSE
;
2610 if (blkid
== DMU_BONUS_BLKID
)
2611 found_db
= dbuf_find_bonus(os
, obj
);
2613 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2615 if (found_db
!= NULL
) {
2616 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2617 (void) refcount_add(&db
->db_holds
, tag
);
2620 mutex_exit(&found_db
->db_mtx
);
2626 * If you call dbuf_rele() you had better not be referencing the dnode handle
2627 * unless you have some other direct or indirect hold on the dnode. (An indirect
2628 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2629 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2630 * dnode's parent dbuf evicting its dnode handles.
2633 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2635 mutex_enter(&db
->db_mtx
);
2636 dbuf_rele_and_unlock(db
, tag
);
2640 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2642 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2646 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2647 * db_dirtycnt and db_holds to be updated atomically.
2650 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2654 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2658 * Remove the reference to the dbuf before removing its hold on the
2659 * dnode so we can guarantee in dnode_move() that a referenced bonus
2660 * buffer has a corresponding dnode hold.
2662 holds
= refcount_remove(&db
->db_holds
, tag
);
2666 * We can't freeze indirects if there is a possibility that they
2667 * may be modified in the current syncing context.
2669 if (db
->db_buf
&& holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0))
2670 arc_buf_freeze(db
->db_buf
);
2672 if (holds
== db
->db_dirtycnt
&&
2673 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2674 dbuf_evict_user(db
);
2677 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2679 boolean_t evict_dbuf
= db
->db_pending_evict
;
2682 * If the dnode moves here, we cannot cross this
2683 * barrier until the move completes.
2688 atomic_dec_32(&dn
->dn_dbufs_count
);
2691 * Decrementing the dbuf count means that the bonus
2692 * buffer's dnode hold is no longer discounted in
2693 * dnode_move(). The dnode cannot move until after
2694 * the dnode_rele() below.
2699 * Do not reference db after its lock is dropped.
2700 * Another thread may evict it.
2702 mutex_exit(&db
->db_mtx
);
2705 dnode_evict_bonus(dn
);
2708 } else if (db
->db_buf
== NULL
) {
2710 * This is a special case: we never associated this
2711 * dbuf with any data allocated from the ARC.
2713 ASSERT(db
->db_state
== DB_UNCACHED
||
2714 db
->db_state
== DB_NOFILL
);
2716 } else if (arc_released(db
->db_buf
)) {
2717 arc_buf_t
*buf
= db
->db_buf
;
2719 * This dbuf has anonymous data associated with it.
2721 dbuf_clear_data(db
);
2722 VERIFY(arc_buf_remove_ref(buf
, db
));
2725 VERIFY(!arc_buf_remove_ref(db
->db_buf
, db
));
2728 * A dbuf will be eligible for eviction if either the
2729 * 'primarycache' property is set or a duplicate
2730 * copy of this buffer is already cached in the arc.
2732 * In the case of the 'primarycache' a buffer
2733 * is considered for eviction if it matches the
2734 * criteria set in the property.
2736 * To decide if our buffer is considered a
2737 * duplicate, we must call into the arc to determine
2738 * if multiple buffers are referencing the same
2739 * block on-disk. If so, then we simply evict
2742 if (!DBUF_IS_CACHEABLE(db
)) {
2743 if (db
->db_blkptr
!= NULL
&&
2744 !BP_IS_HOLE(db
->db_blkptr
) &&
2745 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2747 dmu_objset_spa(db
->db_objset
);
2748 blkptr_t bp
= *db
->db_blkptr
;
2750 arc_freed(spa
, &bp
);
2754 } else if (db
->db_pending_evict
||
2755 arc_buf_eviction_needed(db
->db_buf
)) {
2758 mutex_exit(&db
->db_mtx
);
2762 mutex_exit(&db
->db_mtx
);
2766 #pragma weak dmu_buf_refcount = dbuf_refcount
2768 dbuf_refcount(dmu_buf_impl_t
*db
)
2770 return (refcount_count(&db
->db_holds
));
2774 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
2775 dmu_buf_user_t
*new_user
)
2777 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2779 mutex_enter(&db
->db_mtx
);
2780 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2781 if (db
->db_user
== old_user
)
2782 db
->db_user
= new_user
;
2784 old_user
= db
->db_user
;
2785 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2786 mutex_exit(&db
->db_mtx
);
2792 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2794 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
2798 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2800 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2802 db
->db_user_immediate_evict
= TRUE
;
2803 return (dmu_buf_set_user(db_fake
, user
));
2807 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2809 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
2813 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2815 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2817 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2818 return (db
->db_user
);
2822 dmu_buf_user_evict_wait()
2824 taskq_wait(dbu_evict_taskq
);
2828 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2830 boolean_t res
= B_FALSE
;
2831 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2834 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2835 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2841 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2843 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2844 return (dbi
->db_blkptr
);
2848 dmu_buf_get_objset(dmu_buf_t
*db
)
2850 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2851 return (dbi
->db_objset
);
2855 dmu_buf_dnode_enter(dmu_buf_t
*db
)
2857 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2858 DB_DNODE_ENTER(dbi
);
2859 return (DB_DNODE(dbi
));
2863 dmu_buf_dnode_exit(dmu_buf_t
*db
)
2865 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2870 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2872 /* ASSERT(dmu_tx_is_syncing(tx) */
2873 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2875 if (db
->db_blkptr
!= NULL
)
2878 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2879 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2880 BP_ZERO(db
->db_blkptr
);
2883 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2885 * This buffer was allocated at a time when there was
2886 * no available blkptrs from the dnode, or it was
2887 * inappropriate to hook it in (i.e., nlevels mis-match).
2889 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2890 ASSERT(db
->db_parent
== NULL
);
2891 db
->db_parent
= dn
->dn_dbuf
;
2892 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2895 dmu_buf_impl_t
*parent
= db
->db_parent
;
2896 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2898 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2899 if (parent
== NULL
) {
2900 mutex_exit(&db
->db_mtx
);
2901 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2902 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2903 db
->db_blkid
>> epbs
, db
);
2904 rw_exit(&dn
->dn_struct_rwlock
);
2905 mutex_enter(&db
->db_mtx
);
2906 db
->db_parent
= parent
;
2908 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2909 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2915 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2916 * is critical the we not allow the compiler to inline this function in to
2917 * dbuf_sync_list() thereby drastically bloating the stack usage.
2919 noinline
static void
2920 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2922 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2926 ASSERT(dmu_tx_is_syncing(tx
));
2928 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2930 mutex_enter(&db
->db_mtx
);
2932 ASSERT(db
->db_level
> 0);
2935 /* Read the block if it hasn't been read yet. */
2936 if (db
->db_buf
== NULL
) {
2937 mutex_exit(&db
->db_mtx
);
2938 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2939 mutex_enter(&db
->db_mtx
);
2941 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2942 ASSERT(db
->db_buf
!= NULL
);
2946 /* Indirect block size must match what the dnode thinks it is. */
2947 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2948 dbuf_check_blkptr(dn
, db
);
2951 /* Provide the pending dirty record to child dbufs */
2952 db
->db_data_pending
= dr
;
2954 mutex_exit(&db
->db_mtx
);
2955 dbuf_write(dr
, db
->db_buf
, tx
);
2958 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2959 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
2960 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2961 mutex_exit(&dr
->dt
.di
.dr_mtx
);
2966 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
2967 * critical the we not allow the compiler to inline this function in to
2968 * dbuf_sync_list() thereby drastically bloating the stack usage.
2970 noinline
static void
2971 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2973 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
2974 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2977 uint64_t txg
= tx
->tx_txg
;
2979 ASSERT(dmu_tx_is_syncing(tx
));
2981 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2983 mutex_enter(&db
->db_mtx
);
2985 * To be synced, we must be dirtied. But we
2986 * might have been freed after the dirty.
2988 if (db
->db_state
== DB_UNCACHED
) {
2989 /* This buffer has been freed since it was dirtied */
2990 ASSERT(db
->db
.db_data
== NULL
);
2991 } else if (db
->db_state
== DB_FILL
) {
2992 /* This buffer was freed and is now being re-filled */
2993 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
2995 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3002 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3003 mutex_enter(&dn
->dn_mtx
);
3004 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3006 * In the previous transaction group, the bonus buffer
3007 * was entirely used to store the attributes for the
3008 * dnode which overrode the dn_spill field. However,
3009 * when adding more attributes to the file a spill
3010 * block was required to hold the extra attributes.
3012 * Make sure to clear the garbage left in the dn_spill
3013 * field from the previous attributes in the bonus
3014 * buffer. Otherwise, after writing out the spill
3015 * block to the new allocated dva, it will free
3016 * the old block pointed to by the invalid dn_spill.
3018 db
->db_blkptr
= NULL
;
3020 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3021 mutex_exit(&dn
->dn_mtx
);
3025 * If this is a bonus buffer, simply copy the bonus data into the
3026 * dnode. It will be written out when the dnode is synced (and it
3027 * will be synced, since it must have been dirty for dbuf_sync to
3030 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3031 dbuf_dirty_record_t
**drp
;
3033 ASSERT(*datap
!= NULL
);
3034 ASSERT0(db
->db_level
);
3035 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=,
3036 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3037 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
3040 if (*datap
!= db
->db
.db_data
) {
3041 int slots
= DB_DNODE(db
)->dn_num_slots
;
3042 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3043 zio_buf_free(*datap
, bonuslen
);
3044 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3046 db
->db_data_pending
= NULL
;
3047 drp
= &db
->db_last_dirty
;
3049 drp
= &(*drp
)->dr_next
;
3050 ASSERT(dr
->dr_next
== NULL
);
3051 ASSERT(dr
->dr_dbuf
== db
);
3053 if (dr
->dr_dbuf
->db_level
!= 0) {
3054 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3055 list_destroy(&dr
->dt
.di
.dr_children
);
3057 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3058 ASSERT(db
->db_dirtycnt
> 0);
3059 db
->db_dirtycnt
-= 1;
3060 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3067 * This function may have dropped the db_mtx lock allowing a dmu_sync
3068 * operation to sneak in. As a result, we need to ensure that we
3069 * don't check the dr_override_state until we have returned from
3070 * dbuf_check_blkptr.
3072 dbuf_check_blkptr(dn
, db
);
3075 * If this buffer is in the middle of an immediate write,
3076 * wait for the synchronous IO to complete.
3078 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3079 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3080 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3081 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3084 if (db
->db_state
!= DB_NOFILL
&&
3085 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3086 refcount_count(&db
->db_holds
) > 1 &&
3087 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3088 *datap
== db
->db_buf
) {
3090 * If this buffer is currently "in use" (i.e., there
3091 * are active holds and db_data still references it),
3092 * then make a copy before we start the write so that
3093 * any modifications from the open txg will not leak
3096 * NOTE: this copy does not need to be made for
3097 * objects only modified in the syncing context (e.g.
3098 * DNONE_DNODE blocks).
3100 int blksz
= arc_buf_size(*datap
);
3101 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3102 *datap
= arc_buf_alloc(os
->os_spa
, blksz
, db
, type
);
3103 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
3105 db
->db_data_pending
= dr
;
3107 mutex_exit(&db
->db_mtx
);
3109 dbuf_write(dr
, *datap
, tx
);
3111 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3112 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3113 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3117 * Although zio_nowait() does not "wait for an IO", it does
3118 * initiate the IO. If this is an empty write it seems plausible
3119 * that the IO could actually be completed before the nowait
3120 * returns. We need to DB_DNODE_EXIT() first in case
3121 * zio_nowait() invalidates the dbuf.
3124 zio_nowait(dr
->dr_zio
);
3129 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3131 dbuf_dirty_record_t
*dr
;
3133 while ((dr
= list_head(list
))) {
3134 if (dr
->dr_zio
!= NULL
) {
3136 * If we find an already initialized zio then we
3137 * are processing the meta-dnode, and we have finished.
3138 * The dbufs for all dnodes are put back on the list
3139 * during processing, so that we can zio_wait()
3140 * these IOs after initiating all child IOs.
3142 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3143 DMU_META_DNODE_OBJECT
);
3146 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3147 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3148 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3150 list_remove(list
, dr
);
3151 if (dr
->dr_dbuf
->db_level
> 0)
3152 dbuf_sync_indirect(dr
, tx
);
3154 dbuf_sync_leaf(dr
, tx
);
3160 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3162 dmu_buf_impl_t
*db
= vdb
;
3164 blkptr_t
*bp
= zio
->io_bp
;
3165 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3166 spa_t
*spa
= zio
->io_spa
;
3171 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3172 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3176 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3177 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3178 zio
->io_prev_space_delta
= delta
;
3180 if (bp
->blk_birth
!= 0) {
3181 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3182 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3183 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3184 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3185 BP_IS_EMBEDDED(bp
));
3186 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3189 mutex_enter(&db
->db_mtx
);
3192 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3193 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3194 ASSERT(!(BP_IS_HOLE(bp
)) &&
3195 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3199 if (db
->db_level
== 0) {
3200 mutex_enter(&dn
->dn_mtx
);
3201 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3202 db
->db_blkid
!= DMU_SPILL_BLKID
)
3203 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3204 mutex_exit(&dn
->dn_mtx
);
3206 if (dn
->dn_type
== DMU_OT_DNODE
) {
3208 while (i
< db
->db
.db_size
) {
3209 dnode_phys_t
*dnp
= db
->db
.db_data
+ i
;
3211 i
+= DNODE_MIN_SIZE
;
3212 if (dnp
->dn_type
!= DMU_OT_NONE
) {
3214 i
+= dnp
->dn_extra_slots
*
3219 if (BP_IS_HOLE(bp
)) {
3226 blkptr_t
*ibp
= db
->db
.db_data
;
3227 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3228 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3229 if (BP_IS_HOLE(ibp
))
3231 fill
+= BP_GET_FILL(ibp
);
3236 if (!BP_IS_EMBEDDED(bp
))
3237 bp
->blk_fill
= fill
;
3239 mutex_exit(&db
->db_mtx
);
3241 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3242 *db
->db_blkptr
= *bp
;
3243 rw_exit(&dn
->dn_struct_rwlock
);
3248 * This function gets called just prior to running through the compression
3249 * stage of the zio pipeline. If we're an indirect block comprised of only
3250 * holes, then we want this indirect to be compressed away to a hole. In
3251 * order to do that we must zero out any information about the holes that
3252 * this indirect points to prior to before we try to compress it.
3255 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3257 dmu_buf_impl_t
*db
= vdb
;
3263 ASSERT3U(db
->db_level
, >, 0);
3266 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3268 /* Determine if all our children are holes */
3269 for (i
= 0, bp
= db
->db
.db_data
; i
< 1 << epbs
; i
++, bp
++) {
3270 if (!BP_IS_HOLE(bp
))
3275 * If all the children are holes, then zero them all out so that
3276 * we may get compressed away.
3278 if (i
== 1 << epbs
) {
3279 /* didn't find any non-holes */
3280 bzero(db
->db
.db_data
, db
->db
.db_size
);
3286 * The SPA will call this callback several times for each zio - once
3287 * for every physical child i/o (zio->io_phys_children times). This
3288 * allows the DMU to monitor the progress of each logical i/o. For example,
3289 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3290 * block. There may be a long delay before all copies/fragments are completed,
3291 * so this callback allows us to retire dirty space gradually, as the physical
3296 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3298 dmu_buf_impl_t
*db
= arg
;
3299 objset_t
*os
= db
->db_objset
;
3300 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3301 dbuf_dirty_record_t
*dr
;
3304 dr
= db
->db_data_pending
;
3305 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3308 * The callback will be called io_phys_children times. Retire one
3309 * portion of our dirty space each time we are called. Any rounding
3310 * error will be cleaned up by dsl_pool_sync()'s call to
3311 * dsl_pool_undirty_space().
3313 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3314 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3319 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3321 dmu_buf_impl_t
*db
= vdb
;
3322 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3323 blkptr_t
*bp
= db
->db_blkptr
;
3324 objset_t
*os
= db
->db_objset
;
3325 dmu_tx_t
*tx
= os
->os_synctx
;
3326 dbuf_dirty_record_t
**drp
, *dr
;
3328 ASSERT0(zio
->io_error
);
3329 ASSERT(db
->db_blkptr
== bp
);
3332 * For nopwrites and rewrites we ensure that the bp matches our
3333 * original and bypass all the accounting.
3335 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3336 ASSERT(BP_EQUAL(bp
, bp_orig
));
3338 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3339 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3340 dsl_dataset_block_born(ds
, bp
, tx
);
3343 mutex_enter(&db
->db_mtx
);
3347 drp
= &db
->db_last_dirty
;
3348 while ((dr
= *drp
) != db
->db_data_pending
)
3350 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3351 ASSERT(dr
->dr_dbuf
== db
);
3352 ASSERT(dr
->dr_next
== NULL
);
3356 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3361 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3362 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3363 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3368 if (db
->db_level
== 0) {
3369 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3370 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3371 if (db
->db_state
!= DB_NOFILL
) {
3372 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3373 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
3375 else if (!arc_released(db
->db_buf
))
3376 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3383 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3384 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3385 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3386 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3388 ASSERT3U(db
->db_blkid
, <=,
3389 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3390 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3392 if (!arc_released(db
->db_buf
))
3393 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3396 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3397 list_destroy(&dr
->dt
.di
.dr_children
);
3399 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3401 cv_broadcast(&db
->db_changed
);
3402 ASSERT(db
->db_dirtycnt
> 0);
3403 db
->db_dirtycnt
-= 1;
3404 db
->db_data_pending
= NULL
;
3405 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3409 dbuf_write_nofill_ready(zio_t
*zio
)
3411 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3415 dbuf_write_nofill_done(zio_t
*zio
)
3417 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3421 dbuf_write_override_ready(zio_t
*zio
)
3423 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3424 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3426 dbuf_write_ready(zio
, NULL
, db
);
3430 dbuf_write_override_done(zio_t
*zio
)
3432 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3433 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3434 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3436 mutex_enter(&db
->db_mtx
);
3437 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3438 if (!BP_IS_HOLE(obp
))
3439 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3440 arc_release(dr
->dt
.dl
.dr_data
, db
);
3442 mutex_exit(&db
->db_mtx
);
3444 dbuf_write_done(zio
, NULL
, db
);
3447 /* Issue I/O to commit a dirty buffer to disk. */
3449 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3451 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3454 dmu_buf_impl_t
*parent
= db
->db_parent
;
3455 uint64_t txg
= tx
->tx_txg
;
3456 zbookmark_phys_t zb
;
3461 ASSERT(dmu_tx_is_syncing(tx
));
3467 if (db
->db_state
!= DB_NOFILL
) {
3468 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3470 * Private object buffers are released here rather
3471 * than in dbuf_dirty() since they are only modified
3472 * in the syncing context and we don't want the
3473 * overhead of making multiple copies of the data.
3475 if (BP_IS_HOLE(db
->db_blkptr
)) {
3478 dbuf_release_bp(db
);
3483 if (parent
!= dn
->dn_dbuf
) {
3484 /* Our parent is an indirect block. */
3485 /* We have a dirty parent that has been scheduled for write. */
3486 ASSERT(parent
&& parent
->db_data_pending
);
3487 /* Our parent's buffer is one level closer to the dnode. */
3488 ASSERT(db
->db_level
== parent
->db_level
-1);
3490 * We're about to modify our parent's db_data by modifying
3491 * our block pointer, so the parent must be released.
3493 ASSERT(arc_released(parent
->db_buf
));
3494 zio
= parent
->db_data_pending
->dr_zio
;
3496 /* Our parent is the dnode itself. */
3497 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3498 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3499 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3500 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3501 ASSERT3P(db
->db_blkptr
, ==,
3502 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3506 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3507 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3510 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3511 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3512 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3514 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3516 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3518 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3522 * We copy the blkptr now (rather than when we instantiate the dirty
3523 * record), because its value can change between open context and
3524 * syncing context. We do not need to hold dn_struct_rwlock to read
3525 * db_blkptr because we are in syncing context.
3527 dr
->dr_bp_copy
= *db
->db_blkptr
;
3529 if (db
->db_level
== 0 &&
3530 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3532 * The BP for this block has been provided by open context
3533 * (by dmu_sync() or dmu_buf_write_embedded()).
3535 void *contents
= (data
!= NULL
) ? data
->b_data
: NULL
;
3537 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3538 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, &zp
,
3539 dbuf_write_override_ready
, NULL
, NULL
,
3540 dbuf_write_override_done
,
3541 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3542 mutex_enter(&db
->db_mtx
);
3543 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3544 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3545 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3546 mutex_exit(&db
->db_mtx
);
3547 } else if (db
->db_state
== DB_NOFILL
) {
3548 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
);
3549 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3550 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, &zp
,
3551 dbuf_write_nofill_ready
, NULL
, NULL
,
3552 dbuf_write_nofill_done
, db
,
3553 ZIO_PRIORITY_ASYNC_WRITE
,
3554 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3556 arc_done_func_t
*children_ready_cb
= NULL
;
3557 ASSERT(arc_released(data
));
3560 * For indirect blocks, we want to setup the children
3561 * ready callback so that we can properly handle an indirect
3562 * block that only contains holes.
3564 if (db
->db_level
!= 0)
3565 children_ready_cb
= dbuf_write_children_ready
;
3567 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3568 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3569 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dbuf_write_ready
,
3571 dbuf_write_physdone
, dbuf_write_done
, db
,
3572 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3576 #if defined(_KERNEL) && defined(HAVE_SPL)
3577 EXPORT_SYMBOL(dbuf_find
);
3578 EXPORT_SYMBOL(dbuf_is_metadata
);
3579 EXPORT_SYMBOL(dbuf_evict
);
3580 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
3581 EXPORT_SYMBOL(dbuf_whichblock
);
3582 EXPORT_SYMBOL(dbuf_read
);
3583 EXPORT_SYMBOL(dbuf_unoverride
);
3584 EXPORT_SYMBOL(dbuf_free_range
);
3585 EXPORT_SYMBOL(dbuf_new_size
);
3586 EXPORT_SYMBOL(dbuf_release_bp
);
3587 EXPORT_SYMBOL(dbuf_dirty
);
3588 EXPORT_SYMBOL(dmu_buf_will_dirty
);
3589 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
3590 EXPORT_SYMBOL(dmu_buf_will_fill
);
3591 EXPORT_SYMBOL(dmu_buf_fill_done
);
3592 EXPORT_SYMBOL(dmu_buf_rele
);
3593 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
3594 EXPORT_SYMBOL(dbuf_clear
);
3595 EXPORT_SYMBOL(dbuf_prefetch
);
3596 EXPORT_SYMBOL(dbuf_hold_impl
);
3597 EXPORT_SYMBOL(dbuf_hold
);
3598 EXPORT_SYMBOL(dbuf_hold_level
);
3599 EXPORT_SYMBOL(dbuf_create_bonus
);
3600 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
3601 EXPORT_SYMBOL(dbuf_rm_spill
);
3602 EXPORT_SYMBOL(dbuf_add_ref
);
3603 EXPORT_SYMBOL(dbuf_rele
);
3604 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
3605 EXPORT_SYMBOL(dbuf_refcount
);
3606 EXPORT_SYMBOL(dbuf_sync_list
);
3607 EXPORT_SYMBOL(dmu_buf_set_user
);
3608 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3609 EXPORT_SYMBOL(dmu_buf_get_user
);
3610 EXPORT_SYMBOL(dmu_buf_freeable
);
3611 EXPORT_SYMBOL(dmu_buf_get_blkptr
);