4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_dbuf.h>
49 struct dbuf_hold_impl_data
{
50 /* Function arguments */
54 boolean_t dh_fail_sparse
;
55 boolean_t dh_fail_uncached
;
57 dmu_buf_impl_t
**dh_dbp
;
59 dmu_buf_impl_t
*dh_db
;
60 dmu_buf_impl_t
*dh_parent
;
63 dbuf_dirty_record_t
*dh_dr
;
64 arc_buf_contents_t dh_type
;
68 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
69 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, boolean_t fail_sparse
,
70 boolean_t fail_uncached
,
71 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
72 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
75 * Number of times that zfs_free_range() took the slow path while doing
76 * a zfs receive. A nonzero value indicates a potential performance problem.
78 uint64_t zfs_free_range_recv_miss
;
80 static void dbuf_destroy(dmu_buf_impl_t
*db
);
81 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
82 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
85 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
86 dmu_buf_evict_func_t
*evict_func
, dmu_buf_t
**clear_on_evict_dbufp
);
90 * Global data structures and functions for the dbuf cache.
92 static kmem_cache_t
*dbuf_cache
;
93 static taskq_t
*dbu_evict_taskq
;
97 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
99 dmu_buf_impl_t
*db
= vdb
;
100 bzero(db
, sizeof (dmu_buf_impl_t
));
102 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
103 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
104 refcount_create(&db
->db_holds
);
111 dbuf_dest(void *vdb
, void *unused
)
113 dmu_buf_impl_t
*db
= vdb
;
114 mutex_destroy(&db
->db_mtx
);
115 cv_destroy(&db
->db_changed
);
116 refcount_destroy(&db
->db_holds
);
120 * dbuf hash table routines
122 static dbuf_hash_table_t dbuf_hash_table
;
124 static uint64_t dbuf_hash_count
;
127 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
129 uintptr_t osv
= (uintptr_t)os
;
130 uint64_t crc
= -1ULL;
132 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
133 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
134 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
135 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
136 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
137 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
138 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
140 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
145 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
147 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
148 ((dbuf)->db.db_object == (obj) && \
149 (dbuf)->db_objset == (os) && \
150 (dbuf)->db_level == (level) && \
151 (dbuf)->db_blkid == (blkid))
154 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
156 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
161 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
162 idx
= hv
& h
->hash_table_mask
;
164 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
165 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
166 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
167 mutex_enter(&db
->db_mtx
);
168 if (db
->db_state
!= DB_EVICTING
) {
169 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
172 mutex_exit(&db
->db_mtx
);
175 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
179 static dmu_buf_impl_t
*
180 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
183 dmu_buf_impl_t
*db
= NULL
;
185 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
186 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
187 if (dn
->dn_bonus
!= NULL
) {
189 mutex_enter(&db
->db_mtx
);
191 rw_exit(&dn
->dn_struct_rwlock
);
192 dnode_rele(dn
, FTAG
);
198 * Insert an entry into the hash table. If there is already an element
199 * equal to elem in the hash table, then the already existing element
200 * will be returned and the new element will not be inserted.
201 * Otherwise returns NULL.
203 static dmu_buf_impl_t
*
204 dbuf_hash_insert(dmu_buf_impl_t
*db
)
206 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
207 objset_t
*os
= db
->db_objset
;
208 uint64_t obj
= db
->db
.db_object
;
209 int level
= db
->db_level
;
210 uint64_t blkid
, hv
, idx
;
213 blkid
= db
->db_blkid
;
214 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
215 idx
= hv
& h
->hash_table_mask
;
217 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
218 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
219 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
220 mutex_enter(&dbf
->db_mtx
);
221 if (dbf
->db_state
!= DB_EVICTING
) {
222 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
225 mutex_exit(&dbf
->db_mtx
);
229 mutex_enter(&db
->db_mtx
);
230 db
->db_hash_next
= h
->hash_table
[idx
];
231 h
->hash_table
[idx
] = db
;
232 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
233 atomic_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);
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);
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);
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
);
1471 } else if (do_free_accounting
) {
1472 blkptr_t
*bp
= db
->db_blkptr
;
1473 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1474 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1476 * This is only a guess -- if the dbuf is dirty
1477 * in a previous txg, we don't know how much
1478 * space it will use on disk yet. We should
1479 * really have the struct_rwlock to access
1480 * db_blkptr, but since this is just a guess,
1481 * it's OK if we get an odd answer.
1483 ddt_prefetch(os
->os_spa
, bp
);
1484 dnode_willuse_space(dn
, -willfree
, tx
);
1487 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1488 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1489 drop_struct_lock
= TRUE
;
1492 if (db
->db_level
== 0) {
1493 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1494 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1497 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1498 dmu_buf_impl_t
*parent
= db
->db_parent
;
1499 dbuf_dirty_record_t
*di
;
1500 int parent_held
= FALSE
;
1502 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1503 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1505 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1506 db
->db_blkid
>> epbs
, FTAG
);
1507 ASSERT(parent
!= NULL
);
1510 if (drop_struct_lock
)
1511 rw_exit(&dn
->dn_struct_rwlock
);
1512 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1513 di
= dbuf_dirty(parent
, tx
);
1515 dbuf_rele(parent
, FTAG
);
1517 mutex_enter(&db
->db_mtx
);
1519 * Since we've dropped the mutex, it's possible that
1520 * dbuf_undirty() might have changed this out from under us.
1522 if (db
->db_last_dirty
== dr
||
1523 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1524 mutex_enter(&di
->dt
.di
.dr_mtx
);
1525 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1526 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1527 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1528 mutex_exit(&di
->dt
.di
.dr_mtx
);
1531 mutex_exit(&db
->db_mtx
);
1533 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1534 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1535 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1536 mutex_enter(&dn
->dn_mtx
);
1537 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1538 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1539 mutex_exit(&dn
->dn_mtx
);
1540 if (drop_struct_lock
)
1541 rw_exit(&dn
->dn_struct_rwlock
);
1544 dnode_setdirty(dn
, tx
);
1550 * Undirty a buffer in the transaction group referenced by the given
1551 * transaction. Return whether this evicted the dbuf.
1554 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1557 uint64_t txg
= tx
->tx_txg
;
1558 dbuf_dirty_record_t
*dr
, **drp
;
1563 * Due to our use of dn_nlevels below, this can only be called
1564 * in open context, unless we are operating on the MOS.
1565 * From syncing context, dn_nlevels may be different from the
1566 * dn_nlevels used when dbuf was dirtied.
1568 ASSERT(db
->db_objset
==
1569 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1570 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1571 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1572 ASSERT0(db
->db_level
);
1573 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1576 * If this buffer is not dirty, we're done.
1578 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1579 if (dr
->dr_txg
<= txg
)
1581 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1583 ASSERT(dr
->dr_txg
== txg
);
1584 ASSERT(dr
->dr_dbuf
== db
);
1589 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1591 ASSERT(db
->db
.db_size
!= 0);
1593 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1594 dr
->dr_accounted
, txg
);
1599 * Note that there are three places in dbuf_dirty()
1600 * where this dirty record may be put on a list.
1601 * Make sure to do a list_remove corresponding to
1602 * every one of those list_insert calls.
1604 if (dr
->dr_parent
) {
1605 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1606 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1607 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1608 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1609 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1610 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1611 mutex_enter(&dn
->dn_mtx
);
1612 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1613 mutex_exit(&dn
->dn_mtx
);
1617 if (db
->db_state
!= DB_NOFILL
) {
1618 dbuf_unoverride(dr
);
1620 ASSERT(db
->db_buf
!= NULL
);
1621 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1622 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1623 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
, db
));
1626 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1628 ASSERT(db
->db_dirtycnt
> 0);
1629 db
->db_dirtycnt
-= 1;
1631 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1632 arc_buf_t
*buf
= db
->db_buf
;
1634 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(buf
));
1635 dbuf_clear_data(db
);
1636 VERIFY(arc_buf_remove_ref(buf
, db
));
1645 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1647 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1648 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1649 dbuf_dirty_record_t
*dr
;
1651 ASSERT(tx
->tx_txg
!= 0);
1652 ASSERT(!refcount_is_zero(&db
->db_holds
));
1655 * Quick check for dirtyness. For already dirty blocks, this
1656 * reduces runtime of this function by >90%, and overall performance
1657 * by 50% for some workloads (e.g. file deletion with indirect blocks
1660 mutex_enter(&db
->db_mtx
);
1662 for (dr
= db
->db_last_dirty
;
1663 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1665 * It's possible that it is already dirty but not cached,
1666 * because there are some calls to dbuf_dirty() that don't
1667 * go through dmu_buf_will_dirty().
1669 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1670 /* This dbuf is already dirty and cached. */
1672 mutex_exit(&db
->db_mtx
);
1676 mutex_exit(&db
->db_mtx
);
1679 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1680 rf
|= DB_RF_HAVESTRUCT
;
1682 (void) dbuf_read(db
, NULL
, rf
);
1683 (void) dbuf_dirty(db
, tx
);
1687 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1689 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1691 db
->db_state
= DB_NOFILL
;
1693 dmu_buf_will_fill(db_fake
, tx
);
1697 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1699 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1701 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1702 ASSERT(tx
->tx_txg
!= 0);
1703 ASSERT(db
->db_level
== 0);
1704 ASSERT(!refcount_is_zero(&db
->db_holds
));
1706 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1707 dmu_tx_private_ok(tx
));
1710 (void) dbuf_dirty(db
, tx
);
1713 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1716 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1718 mutex_enter(&db
->db_mtx
);
1721 if (db
->db_state
== DB_FILL
) {
1722 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1723 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1724 /* we were freed while filling */
1725 /* XXX dbuf_undirty? */
1726 bzero(db
->db
.db_data
, db
->db
.db_size
);
1727 db
->db_freed_in_flight
= FALSE
;
1729 db
->db_state
= DB_CACHED
;
1730 cv_broadcast(&db
->db_changed
);
1732 mutex_exit(&db
->db_mtx
);
1736 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
1737 bp_embedded_type_t etype
, enum zio_compress comp
,
1738 int uncompressed_size
, int compressed_size
, int byteorder
,
1741 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
1742 struct dirty_leaf
*dl
;
1743 dmu_object_type_t type
;
1745 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
1746 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
1747 SPA_FEATURE_EMBEDDED_DATA
));
1751 type
= DB_DNODE(db
)->dn_type
;
1754 ASSERT0(db
->db_level
);
1755 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1757 dmu_buf_will_not_fill(dbuf
, tx
);
1759 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1760 dl
= &db
->db_last_dirty
->dt
.dl
;
1761 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
1762 data
, comp
, uncompressed_size
, compressed_size
);
1763 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
1764 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
1765 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
1766 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
1768 dl
->dr_override_state
= DR_OVERRIDDEN
;
1769 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
1773 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1774 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1777 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1779 ASSERT(!refcount_is_zero(&db
->db_holds
));
1780 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1781 ASSERT(db
->db_level
== 0);
1782 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1783 ASSERT(buf
!= NULL
);
1784 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1785 ASSERT(tx
->tx_txg
!= 0);
1787 arc_return_buf(buf
, db
);
1788 ASSERT(arc_released(buf
));
1790 mutex_enter(&db
->db_mtx
);
1792 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1793 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1795 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1797 if (db
->db_state
== DB_CACHED
&&
1798 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1799 mutex_exit(&db
->db_mtx
);
1800 (void) dbuf_dirty(db
, tx
);
1801 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1802 VERIFY(arc_buf_remove_ref(buf
, db
));
1803 xuio_stat_wbuf_copied();
1807 xuio_stat_wbuf_nocopy();
1808 if (db
->db_state
== DB_CACHED
) {
1809 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1811 ASSERT(db
->db_buf
!= NULL
);
1812 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
1813 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
1814 if (!arc_released(db
->db_buf
)) {
1815 ASSERT(dr
->dt
.dl
.dr_override_state
==
1817 arc_release(db
->db_buf
, db
);
1819 dr
->dt
.dl
.dr_data
= buf
;
1820 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1821 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
1822 arc_release(db
->db_buf
, db
);
1823 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1827 ASSERT(db
->db_buf
== NULL
);
1828 dbuf_set_data(db
, buf
);
1829 db
->db_state
= DB_FILL
;
1830 mutex_exit(&db
->db_mtx
);
1831 (void) dbuf_dirty(db
, tx
);
1832 dmu_buf_fill_done(&db
->db
, tx
);
1836 * "Clear" the contents of this dbuf. This will mark the dbuf
1837 * EVICTING and clear *most* of its references. Unfortunately,
1838 * when we are not holding the dn_dbufs_mtx, we can't clear the
1839 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1840 * in this case. For callers from the DMU we will usually see:
1841 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1842 * For the arc callback, we will usually see:
1843 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1844 * Sometimes, though, we will get a mix of these two:
1845 * DMU: dbuf_clear()->arc_clear_callback()
1846 * ARC: dbuf_do_evict()->dbuf_destroy()
1848 * This routine will dissociate the dbuf from the arc, by calling
1849 * arc_clear_callback(), but will not evict the data from the ARC.
1852 dbuf_clear(dmu_buf_impl_t
*db
)
1855 dmu_buf_impl_t
*parent
= db
->db_parent
;
1856 dmu_buf_impl_t
*dndb
;
1857 boolean_t dbuf_gone
= B_FALSE
;
1859 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1860 ASSERT(refcount_is_zero(&db
->db_holds
));
1862 dbuf_evict_user(db
);
1864 if (db
->db_state
== DB_CACHED
) {
1865 ASSERT(db
->db
.db_data
!= NULL
);
1866 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1867 int slots
= DB_DNODE(db
)->dn_num_slots
;
1868 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
1869 zio_buf_free(db
->db
.db_data
, bonuslen
);
1870 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
1872 db
->db
.db_data
= NULL
;
1873 db
->db_state
= DB_UNCACHED
;
1876 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1877 ASSERT(db
->db_data_pending
== NULL
);
1879 db
->db_state
= DB_EVICTING
;
1880 db
->db_blkptr
= NULL
;
1885 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& MUTEX_HELD(&dn
->dn_dbufs_mtx
)) {
1886 avl_remove(&dn
->dn_dbufs
, db
);
1887 atomic_dec_32(&dn
->dn_dbufs_count
);
1891 * Decrementing the dbuf count means that the hold corresponding
1892 * to the removed dbuf is no longer discounted in dnode_move(),
1893 * so the dnode cannot be moved until after we release the hold.
1894 * The membar_producer() ensures visibility of the decremented
1895 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1899 db
->db_dnode_handle
= NULL
;
1905 dbuf_gone
= arc_clear_callback(db
->db_buf
);
1908 mutex_exit(&db
->db_mtx
);
1911 * If this dbuf is referenced from an indirect dbuf,
1912 * decrement the ref count on the indirect dbuf.
1914 if (parent
&& parent
!= dndb
)
1915 dbuf_rele(parent
, db
);
1919 * Note: While bpp will always be updated if the function returns success,
1920 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1921 * this happens when the dnode is the meta-dnode, or a userused or groupused
1924 __attribute__((always_inline
))
1926 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
1927 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
1934 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1936 if (blkid
== DMU_SPILL_BLKID
) {
1937 mutex_enter(&dn
->dn_mtx
);
1938 if (dn
->dn_have_spill
&&
1939 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1940 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
1943 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1944 *parentp
= dn
->dn_dbuf
;
1945 mutex_exit(&dn
->dn_mtx
);
1949 if (dn
->dn_phys
->dn_nlevels
== 0)
1952 nlevels
= dn
->dn_phys
->dn_nlevels
;
1954 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1956 ASSERT3U(level
* epbs
, <, 64);
1957 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1958 if (level
>= nlevels
||
1959 (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
1960 /* the buffer has no parent yet */
1961 return (SET_ERROR(ENOENT
));
1962 } else if (level
< nlevels
-1) {
1963 /* this block is referenced from an indirect block */
1966 err
= dbuf_hold_impl(dn
, level
+1,
1967 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
1969 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
1970 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
1971 parentp
, dh
->dh_depth
+ 1);
1972 err
= __dbuf_hold_impl(dh
+ 1);
1976 err
= dbuf_read(*parentp
, NULL
,
1977 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
1979 dbuf_rele(*parentp
, NULL
);
1983 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
1984 (blkid
& ((1ULL << epbs
) - 1));
1987 /* the block is referenced from the dnode */
1988 ASSERT3U(level
, ==, nlevels
-1);
1989 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
1990 blkid
< dn
->dn_phys
->dn_nblkptr
);
1992 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1993 *parentp
= dn
->dn_dbuf
;
1995 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2000 static dmu_buf_impl_t
*
2001 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2002 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2004 objset_t
*os
= dn
->dn_objset
;
2005 dmu_buf_impl_t
*db
, *odb
;
2007 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2008 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2010 db
= kmem_cache_alloc(dbuf_cache
, KM_SLEEP
);
2013 db
->db
.db_object
= dn
->dn_object
;
2014 db
->db_level
= level
;
2015 db
->db_blkid
= blkid
;
2016 db
->db_last_dirty
= NULL
;
2017 db
->db_dirtycnt
= 0;
2018 db
->db_dnode_handle
= dn
->dn_handle
;
2019 db
->db_parent
= parent
;
2020 db
->db_blkptr
= blkptr
;
2023 db
->db_user_immediate_evict
= FALSE
;
2024 db
->db_freed_in_flight
= FALSE
;
2025 db
->db_pending_evict
= FALSE
;
2027 if (blkid
== DMU_BONUS_BLKID
) {
2028 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2029 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2030 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2031 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2032 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2033 db
->db_state
= DB_UNCACHED
;
2034 /* the bonus dbuf is not placed in the hash table */
2035 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2037 } else if (blkid
== DMU_SPILL_BLKID
) {
2038 db
->db
.db_size
= (blkptr
!= NULL
) ?
2039 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2040 db
->db
.db_offset
= 0;
2043 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2044 db
->db
.db_size
= blocksize
;
2045 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2049 * Hold the dn_dbufs_mtx while we get the new dbuf
2050 * in the hash table *and* added to the dbufs list.
2051 * This prevents a possible deadlock with someone
2052 * trying to look up this dbuf before its added to the
2055 mutex_enter(&dn
->dn_dbufs_mtx
);
2056 db
->db_state
= DB_EVICTING
;
2057 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2058 /* someone else inserted it first */
2059 kmem_cache_free(dbuf_cache
, db
);
2060 mutex_exit(&dn
->dn_dbufs_mtx
);
2063 avl_add(&dn
->dn_dbufs
, db
);
2064 if (db
->db_level
== 0 && db
->db_blkid
>=
2065 dn
->dn_unlisted_l0_blkid
)
2066 dn
->dn_unlisted_l0_blkid
= db
->db_blkid
+ 1;
2067 db
->db_state
= DB_UNCACHED
;
2068 mutex_exit(&dn
->dn_dbufs_mtx
);
2069 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2071 if (parent
&& parent
!= dn
->dn_dbuf
)
2072 dbuf_add_ref(parent
, db
);
2074 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2075 refcount_count(&dn
->dn_holds
) > 0);
2076 (void) refcount_add(&dn
->dn_holds
, db
);
2077 atomic_inc_32(&dn
->dn_dbufs_count
);
2079 dprintf_dbuf(db
, "db=%p\n", db
);
2085 dbuf_do_evict(void *private)
2087 dmu_buf_impl_t
*db
= private;
2089 if (!MUTEX_HELD(&db
->db_mtx
))
2090 mutex_enter(&db
->db_mtx
);
2092 ASSERT(refcount_is_zero(&db
->db_holds
));
2094 if (db
->db_state
!= DB_EVICTING
) {
2095 ASSERT(db
->db_state
== DB_CACHED
);
2100 mutex_exit(&db
->db_mtx
);
2107 dbuf_destroy(dmu_buf_impl_t
*db
)
2109 ASSERT(refcount_is_zero(&db
->db_holds
));
2111 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2113 * If this dbuf is still on the dn_dbufs list,
2114 * remove it from that list.
2116 if (db
->db_dnode_handle
!= NULL
) {
2121 mutex_enter(&dn
->dn_dbufs_mtx
);
2122 avl_remove(&dn
->dn_dbufs
, db
);
2123 atomic_dec_32(&dn
->dn_dbufs_count
);
2124 mutex_exit(&dn
->dn_dbufs_mtx
);
2127 * Decrementing the dbuf count means that the hold
2128 * corresponding to the removed dbuf is no longer
2129 * discounted in dnode_move(), so the dnode cannot be
2130 * moved until after we release the hold.
2133 db
->db_dnode_handle
= NULL
;
2135 dbuf_hash_remove(db
);
2137 db
->db_parent
= NULL
;
2140 ASSERT(db
->db
.db_data
== NULL
);
2141 ASSERT(db
->db_hash_next
== NULL
);
2142 ASSERT(db
->db_blkptr
== NULL
);
2143 ASSERT(db
->db_data_pending
== NULL
);
2145 kmem_cache_free(dbuf_cache
, db
);
2146 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2149 typedef struct dbuf_prefetch_arg
{
2150 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2151 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2152 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2153 int dpa_curlevel
; /* The current level that we're reading */
2154 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2155 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2156 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2157 } dbuf_prefetch_arg_t
;
2160 * Actually issue the prefetch read for the block given.
2163 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2166 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2169 aflags
= dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2171 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2172 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2173 ASSERT(dpa
->dpa_zio
!= NULL
);
2174 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2175 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2176 &aflags
, &dpa
->dpa_zb
);
2180 * Called when an indirect block above our prefetch target is read in. This
2181 * will either read in the next indirect block down the tree or issue the actual
2182 * prefetch if the next block down is our target.
2185 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2187 dbuf_prefetch_arg_t
*dpa
= private;
2191 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2192 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2194 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2195 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2196 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2199 dpa
->dpa_curlevel
--;
2201 nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2202 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2203 bp
= ((blkptr_t
*)abuf
->b_data
) +
2204 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2205 if (BP_IS_HOLE(bp
) || (zio
!= NULL
&& zio
->io_error
!= 0)) {
2206 kmem_free(dpa
, sizeof (*dpa
));
2207 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2208 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2209 dbuf_issue_final_prefetch(dpa
, bp
);
2210 kmem_free(dpa
, sizeof (*dpa
));
2212 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2213 zbookmark_phys_t zb
;
2215 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2217 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2218 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2220 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2221 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2222 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2225 (void) arc_buf_remove_ref(abuf
, private);
2229 * Issue prefetch reads for the given block on the given level. If the indirect
2230 * blocks above that block are not in memory, we will read them in
2231 * asynchronously. As a result, this call never blocks waiting for a read to
2235 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2239 int epbs
, nlevels
, curlevel
;
2243 dbuf_prefetch_arg_t
*dpa
;
2246 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2247 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2249 if (blkid
> dn
->dn_maxblkid
)
2252 if (dnode_block_freed(dn
, blkid
))
2256 * This dnode hasn't been written to disk yet, so there's nothing to
2259 nlevels
= dn
->dn_phys
->dn_nlevels
;
2260 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2263 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2264 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2267 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2270 mutex_exit(&db
->db_mtx
);
2272 * This dbuf already exists. It is either CACHED, or
2273 * (we assume) about to be read or filled.
2279 * Find the closest ancestor (indirect block) of the target block
2280 * that is present in the cache. In this indirect block, we will
2281 * find the bp that is at curlevel, curblkid.
2285 while (curlevel
< nlevels
- 1) {
2286 int parent_level
= curlevel
+ 1;
2287 uint64_t parent_blkid
= curblkid
>> epbs
;
2290 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2291 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2292 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2293 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2294 dbuf_rele(db
, FTAG
);
2298 curlevel
= parent_level
;
2299 curblkid
= parent_blkid
;
2302 if (curlevel
== nlevels
- 1) {
2303 /* No cached indirect blocks found. */
2304 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2305 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2307 if (BP_IS_HOLE(&bp
))
2310 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2312 pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2315 dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2316 ds
= dn
->dn_objset
->os_dsl_dataset
;
2317 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2318 dn
->dn_object
, level
, blkid
);
2319 dpa
->dpa_curlevel
= curlevel
;
2320 dpa
->dpa_prio
= prio
;
2321 dpa
->dpa_aflags
= aflags
;
2322 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2323 dpa
->dpa_epbs
= epbs
;
2327 * If we have the indirect just above us, no need to do the asynchronous
2328 * prefetch chain; we'll just run the last step ourselves. If we're at
2329 * a higher level, though, we want to issue the prefetches for all the
2330 * indirect blocks asynchronously, so we can go on with whatever we were
2333 if (curlevel
== level
) {
2334 ASSERT3U(curblkid
, ==, blkid
);
2335 dbuf_issue_final_prefetch(dpa
, &bp
);
2336 kmem_free(dpa
, sizeof (*dpa
));
2338 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2339 zbookmark_phys_t zb
;
2341 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2342 dn
->dn_object
, curlevel
, curblkid
);
2343 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2344 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2345 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2349 * We use pio here instead of dpa_zio since it's possible that
2350 * dpa may have already been freed.
2355 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2358 * Returns with db_holds incremented, and db_mtx not held.
2359 * Note: dn_struct_rwlock must be held.
2362 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2364 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2365 dh
->dh_parent
= NULL
;
2367 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2368 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2369 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2371 *(dh
->dh_dbp
) = NULL
;
2373 /* dbuf_find() returns with db_mtx held */
2374 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2375 dh
->dh_level
, dh
->dh_blkid
);
2377 if (dh
->dh_db
== NULL
) {
2380 if (dh
->dh_fail_uncached
)
2381 return (SET_ERROR(ENOENT
));
2383 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2384 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2385 dh
->dh_fail_sparse
, &dh
->dh_parent
,
2387 if (dh
->dh_fail_sparse
) {
2388 if (dh
->dh_err
== 0 &&
2389 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2390 dh
->dh_err
= SET_ERROR(ENOENT
);
2393 dbuf_rele(dh
->dh_parent
, NULL
);
2394 return (dh
->dh_err
);
2397 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2398 return (dh
->dh_err
);
2399 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2400 dh
->dh_parent
, dh
->dh_bp
);
2403 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2404 mutex_exit(&dh
->dh_db
->db_mtx
);
2405 return (SET_ERROR(ENOENT
));
2408 if (dh
->dh_db
->db_buf
&& refcount_is_zero(&dh
->dh_db
->db_holds
)) {
2409 arc_buf_add_ref(dh
->dh_db
->db_buf
, dh
->dh_db
);
2410 if (dh
->dh_db
->db_buf
->b_data
== NULL
) {
2411 dbuf_clear(dh
->dh_db
);
2412 if (dh
->dh_parent
) {
2413 dbuf_rele(dh
->dh_parent
, NULL
);
2414 dh
->dh_parent
= NULL
;
2418 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2421 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2424 * If this buffer is currently syncing out, and we are are
2425 * still referencing it from db_data, we need to make a copy
2426 * of it in case we decide we want to dirty it again in this txg.
2428 if (dh
->dh_db
->db_level
== 0 &&
2429 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2430 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2431 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2432 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2434 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2435 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2437 dbuf_set_data(dh
->dh_db
,
2438 arc_buf_alloc(dh
->dh_dn
->dn_objset
->os_spa
,
2439 dh
->dh_db
->db
.db_size
, dh
->dh_db
, dh
->dh_type
));
2440 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2441 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2445 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2446 DBUF_VERIFY(dh
->dh_db
);
2447 mutex_exit(&dh
->dh_db
->db_mtx
);
2449 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2451 dbuf_rele(dh
->dh_parent
, NULL
);
2453 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2454 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2455 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2456 *(dh
->dh_dbp
) = dh
->dh_db
;
2462 * The following code preserves the recursive function dbuf_hold_impl()
2463 * but moves the local variables AND function arguments to the heap to
2464 * minimize the stack frame size. Enough space is initially allocated
2465 * on the stack for 20 levels of recursion.
2468 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2469 boolean_t fail_sparse
, boolean_t fail_uncached
,
2470 void *tag
, dmu_buf_impl_t
**dbp
)
2472 struct dbuf_hold_impl_data
*dh
;
2475 dh
= kmem_zalloc(sizeof (struct dbuf_hold_impl_data
) *
2476 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2477 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2478 fail_uncached
, tag
, dbp
, 0);
2480 error
= __dbuf_hold_impl(dh
);
2482 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2483 DBUF_HOLD_IMPL_MAX_DEPTH
);
2489 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2490 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2491 boolean_t fail_sparse
, boolean_t fail_uncached
,
2492 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2495 dh
->dh_level
= level
;
2496 dh
->dh_blkid
= blkid
;
2498 dh
->dh_fail_sparse
= fail_sparse
;
2499 dh
->dh_fail_uncached
= fail_uncached
;
2503 dh
->dh_depth
= depth
;
2507 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2509 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2513 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2516 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2517 return (err
? NULL
: db
);
2521 dbuf_create_bonus(dnode_t
*dn
)
2523 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2525 ASSERT(dn
->dn_bonus
== NULL
);
2526 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2530 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2532 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2535 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2536 return (SET_ERROR(ENOTSUP
));
2538 blksz
= SPA_MINBLOCKSIZE
;
2539 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2540 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2544 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2545 dbuf_new_size(db
, blksz
, tx
);
2546 rw_exit(&dn
->dn_struct_rwlock
);
2553 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2555 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2558 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2560 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2562 VERIFY(refcount_add(&db
->db_holds
, tag
) > 1);
2565 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2567 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2570 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2571 dmu_buf_impl_t
*found_db
;
2572 boolean_t result
= B_FALSE
;
2574 if (blkid
== DMU_BONUS_BLKID
)
2575 found_db
= dbuf_find_bonus(os
, obj
);
2577 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2579 if (found_db
!= NULL
) {
2580 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2581 (void) refcount_add(&db
->db_holds
, tag
);
2584 mutex_exit(&found_db
->db_mtx
);
2590 * If you call dbuf_rele() you had better not be referencing the dnode handle
2591 * unless you have some other direct or indirect hold on the dnode. (An indirect
2592 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2593 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2594 * dnode's parent dbuf evicting its dnode handles.
2597 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2599 mutex_enter(&db
->db_mtx
);
2600 dbuf_rele_and_unlock(db
, tag
);
2604 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2606 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2610 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2611 * db_dirtycnt and db_holds to be updated atomically.
2614 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2618 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2622 * Remove the reference to the dbuf before removing its hold on the
2623 * dnode so we can guarantee in dnode_move() that a referenced bonus
2624 * buffer has a corresponding dnode hold.
2626 holds
= refcount_remove(&db
->db_holds
, tag
);
2630 * We can't freeze indirects if there is a possibility that they
2631 * may be modified in the current syncing context.
2633 if (db
->db_buf
&& holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0))
2634 arc_buf_freeze(db
->db_buf
);
2636 if (holds
== db
->db_dirtycnt
&&
2637 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2638 dbuf_evict_user(db
);
2641 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2643 boolean_t evict_dbuf
= db
->db_pending_evict
;
2646 * If the dnode moves here, we cannot cross this
2647 * barrier until the move completes.
2652 atomic_dec_32(&dn
->dn_dbufs_count
);
2655 * Decrementing the dbuf count means that the bonus
2656 * buffer's dnode hold is no longer discounted in
2657 * dnode_move(). The dnode cannot move until after
2658 * the dnode_rele() below.
2663 * Do not reference db after its lock is dropped.
2664 * Another thread may evict it.
2666 mutex_exit(&db
->db_mtx
);
2669 dnode_evict_bonus(dn
);
2672 } else if (db
->db_buf
== NULL
) {
2674 * This is a special case: we never associated this
2675 * dbuf with any data allocated from the ARC.
2677 ASSERT(db
->db_state
== DB_UNCACHED
||
2678 db
->db_state
== DB_NOFILL
);
2680 } else if (arc_released(db
->db_buf
)) {
2681 arc_buf_t
*buf
= db
->db_buf
;
2683 * This dbuf has anonymous data associated with it.
2685 dbuf_clear_data(db
);
2686 VERIFY(arc_buf_remove_ref(buf
, db
));
2689 VERIFY(!arc_buf_remove_ref(db
->db_buf
, db
));
2692 * A dbuf will be eligible for eviction if either the
2693 * 'primarycache' property is set or a duplicate
2694 * copy of this buffer is already cached in the arc.
2696 * In the case of the 'primarycache' a buffer
2697 * is considered for eviction if it matches the
2698 * criteria set in the property.
2700 * To decide if our buffer is considered a
2701 * duplicate, we must call into the arc to determine
2702 * if multiple buffers are referencing the same
2703 * block on-disk. If so, then we simply evict
2706 if (!DBUF_IS_CACHEABLE(db
)) {
2707 if (db
->db_blkptr
!= NULL
&&
2708 !BP_IS_HOLE(db
->db_blkptr
) &&
2709 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2711 dmu_objset_spa(db
->db_objset
);
2712 blkptr_t bp
= *db
->db_blkptr
;
2714 arc_freed(spa
, &bp
);
2718 } else if (db
->db_pending_evict
||
2719 arc_buf_eviction_needed(db
->db_buf
)) {
2722 mutex_exit(&db
->db_mtx
);
2726 mutex_exit(&db
->db_mtx
);
2730 #pragma weak dmu_buf_refcount = dbuf_refcount
2732 dbuf_refcount(dmu_buf_impl_t
*db
)
2734 return (refcount_count(&db
->db_holds
));
2738 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
2739 dmu_buf_user_t
*new_user
)
2741 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2743 mutex_enter(&db
->db_mtx
);
2744 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2745 if (db
->db_user
== old_user
)
2746 db
->db_user
= new_user
;
2748 old_user
= db
->db_user
;
2749 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2750 mutex_exit(&db
->db_mtx
);
2756 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2758 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
2762 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2764 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2766 db
->db_user_immediate_evict
= TRUE
;
2767 return (dmu_buf_set_user(db_fake
, user
));
2771 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2773 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
2777 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2779 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2781 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2782 return (db
->db_user
);
2786 dmu_buf_user_evict_wait()
2788 taskq_wait(dbu_evict_taskq
);
2792 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2794 boolean_t res
= B_FALSE
;
2795 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2798 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2799 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2805 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2807 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2808 return (dbi
->db_blkptr
);
2812 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2814 /* ASSERT(dmu_tx_is_syncing(tx) */
2815 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2817 if (db
->db_blkptr
!= NULL
)
2820 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2821 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2822 BP_ZERO(db
->db_blkptr
);
2825 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2827 * This buffer was allocated at a time when there was
2828 * no available blkptrs from the dnode, or it was
2829 * inappropriate to hook it in (i.e., nlevels mis-match).
2831 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2832 ASSERT(db
->db_parent
== NULL
);
2833 db
->db_parent
= dn
->dn_dbuf
;
2834 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2837 dmu_buf_impl_t
*parent
= db
->db_parent
;
2838 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2840 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2841 if (parent
== NULL
) {
2842 mutex_exit(&db
->db_mtx
);
2843 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2844 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2845 db
->db_blkid
>> epbs
, db
);
2846 rw_exit(&dn
->dn_struct_rwlock
);
2847 mutex_enter(&db
->db_mtx
);
2848 db
->db_parent
= parent
;
2850 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2851 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2857 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2858 * is critical the we not allow the compiler to inline this function in to
2859 * dbuf_sync_list() thereby drastically bloating the stack usage.
2861 noinline
static void
2862 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2864 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2868 ASSERT(dmu_tx_is_syncing(tx
));
2870 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2872 mutex_enter(&db
->db_mtx
);
2874 ASSERT(db
->db_level
> 0);
2877 /* Read the block if it hasn't been read yet. */
2878 if (db
->db_buf
== NULL
) {
2879 mutex_exit(&db
->db_mtx
);
2880 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2881 mutex_enter(&db
->db_mtx
);
2883 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2884 ASSERT(db
->db_buf
!= NULL
);
2888 /* Indirect block size must match what the dnode thinks it is. */
2889 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2890 dbuf_check_blkptr(dn
, db
);
2893 /* Provide the pending dirty record to child dbufs */
2894 db
->db_data_pending
= dr
;
2896 mutex_exit(&db
->db_mtx
);
2897 dbuf_write(dr
, db
->db_buf
, tx
);
2900 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2901 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
2902 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2903 mutex_exit(&dr
->dt
.di
.dr_mtx
);
2908 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
2909 * critical the we not allow the compiler to inline this function in to
2910 * dbuf_sync_list() thereby drastically bloating the stack usage.
2912 noinline
static void
2913 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2915 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
2916 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2919 uint64_t txg
= tx
->tx_txg
;
2921 ASSERT(dmu_tx_is_syncing(tx
));
2923 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2925 mutex_enter(&db
->db_mtx
);
2927 * To be synced, we must be dirtied. But we
2928 * might have been freed after the dirty.
2930 if (db
->db_state
== DB_UNCACHED
) {
2931 /* This buffer has been freed since it was dirtied */
2932 ASSERT(db
->db
.db_data
== NULL
);
2933 } else if (db
->db_state
== DB_FILL
) {
2934 /* This buffer was freed and is now being re-filled */
2935 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
2937 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
2944 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2945 mutex_enter(&dn
->dn_mtx
);
2946 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
2948 * In the previous transaction group, the bonus buffer
2949 * was entirely used to store the attributes for the
2950 * dnode which overrode the dn_spill field. However,
2951 * when adding more attributes to the file a spill
2952 * block was required to hold the extra attributes.
2954 * Make sure to clear the garbage left in the dn_spill
2955 * field from the previous attributes in the bonus
2956 * buffer. Otherwise, after writing out the spill
2957 * block to the new allocated dva, it will free
2958 * the old block pointed to by the invalid dn_spill.
2960 db
->db_blkptr
= NULL
;
2962 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
2963 mutex_exit(&dn
->dn_mtx
);
2967 * If this is a bonus buffer, simply copy the bonus data into the
2968 * dnode. It will be written out when the dnode is synced (and it
2969 * will be synced, since it must have been dirty for dbuf_sync to
2972 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2973 dbuf_dirty_record_t
**drp
;
2975 ASSERT(*datap
!= NULL
);
2976 ASSERT0(db
->db_level
);
2977 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=,
2978 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
2979 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
2982 if (*datap
!= db
->db
.db_data
) {
2983 int slots
= DB_DNODE(db
)->dn_num_slots
;
2984 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2985 zio_buf_free(*datap
, bonuslen
);
2986 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2988 db
->db_data_pending
= NULL
;
2989 drp
= &db
->db_last_dirty
;
2991 drp
= &(*drp
)->dr_next
;
2992 ASSERT(dr
->dr_next
== NULL
);
2993 ASSERT(dr
->dr_dbuf
== db
);
2995 if (dr
->dr_dbuf
->db_level
!= 0) {
2996 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2997 list_destroy(&dr
->dt
.di
.dr_children
);
2999 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3000 ASSERT(db
->db_dirtycnt
> 0);
3001 db
->db_dirtycnt
-= 1;
3002 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3009 * This function may have dropped the db_mtx lock allowing a dmu_sync
3010 * operation to sneak in. As a result, we need to ensure that we
3011 * don't check the dr_override_state until we have returned from
3012 * dbuf_check_blkptr.
3014 dbuf_check_blkptr(dn
, db
);
3017 * If this buffer is in the middle of an immediate write,
3018 * wait for the synchronous IO to complete.
3020 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3021 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3022 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3023 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3026 if (db
->db_state
!= DB_NOFILL
&&
3027 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3028 refcount_count(&db
->db_holds
) > 1 &&
3029 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3030 *datap
== db
->db_buf
) {
3032 * If this buffer is currently "in use" (i.e., there
3033 * are active holds and db_data still references it),
3034 * then make a copy before we start the write so that
3035 * any modifications from the open txg will not leak
3038 * NOTE: this copy does not need to be made for
3039 * objects only modified in the syncing context (e.g.
3040 * DNONE_DNODE blocks).
3042 int blksz
= arc_buf_size(*datap
);
3043 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3044 *datap
= arc_buf_alloc(os
->os_spa
, blksz
, db
, type
);
3045 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
3047 db
->db_data_pending
= dr
;
3049 mutex_exit(&db
->db_mtx
);
3051 dbuf_write(dr
, *datap
, tx
);
3053 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3054 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3055 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3059 * Although zio_nowait() does not "wait for an IO", it does
3060 * initiate the IO. If this is an empty write it seems plausible
3061 * that the IO could actually be completed before the nowait
3062 * returns. We need to DB_DNODE_EXIT() first in case
3063 * zio_nowait() invalidates the dbuf.
3066 zio_nowait(dr
->dr_zio
);
3071 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3073 dbuf_dirty_record_t
*dr
;
3075 while ((dr
= list_head(list
))) {
3076 if (dr
->dr_zio
!= NULL
) {
3078 * If we find an already initialized zio then we
3079 * are processing the meta-dnode, and we have finished.
3080 * The dbufs for all dnodes are put back on the list
3081 * during processing, so that we can zio_wait()
3082 * these IOs after initiating all child IOs.
3084 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3085 DMU_META_DNODE_OBJECT
);
3088 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3089 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3090 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3092 list_remove(list
, dr
);
3093 if (dr
->dr_dbuf
->db_level
> 0)
3094 dbuf_sync_indirect(dr
, tx
);
3096 dbuf_sync_leaf(dr
, tx
);
3102 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3104 dmu_buf_impl_t
*db
= vdb
;
3106 blkptr_t
*bp
= zio
->io_bp
;
3107 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3108 spa_t
*spa
= zio
->io_spa
;
3113 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3114 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3118 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3119 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3120 zio
->io_prev_space_delta
= delta
;
3122 if (bp
->blk_birth
!= 0) {
3123 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3124 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3125 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3126 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3127 BP_IS_EMBEDDED(bp
));
3128 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3131 mutex_enter(&db
->db_mtx
);
3134 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3135 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3136 ASSERT(!(BP_IS_HOLE(bp
)) &&
3137 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3141 if (db
->db_level
== 0) {
3142 mutex_enter(&dn
->dn_mtx
);
3143 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3144 db
->db_blkid
!= DMU_SPILL_BLKID
)
3145 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3146 mutex_exit(&dn
->dn_mtx
);
3148 if (dn
->dn_type
== DMU_OT_DNODE
) {
3150 while (i
< db
->db
.db_size
) {
3151 dnode_phys_t
*dnp
= db
->db
.db_data
+ i
;
3153 i
+= DNODE_MIN_SIZE
;
3154 if (dnp
->dn_type
!= DMU_OT_NONE
) {
3156 i
+= dnp
->dn_extra_slots
*
3161 if (BP_IS_HOLE(bp
)) {
3168 blkptr_t
*ibp
= db
->db
.db_data
;
3169 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3170 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3171 if (BP_IS_HOLE(ibp
))
3173 fill
+= BP_GET_FILL(ibp
);
3178 if (!BP_IS_EMBEDDED(bp
))
3179 bp
->blk_fill
= fill
;
3181 mutex_exit(&db
->db_mtx
);
3183 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3184 *db
->db_blkptr
= *bp
;
3185 rw_exit(&dn
->dn_struct_rwlock
);
3190 * This function gets called just prior to running through the compression
3191 * stage of the zio pipeline. If we're an indirect block comprised of only
3192 * holes, then we want this indirect to be compressed away to a hole. In
3193 * order to do that we must zero out any information about the holes that
3194 * this indirect points to prior to before we try to compress it.
3197 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3199 dmu_buf_impl_t
*db
= vdb
;
3205 ASSERT3U(db
->db_level
, >, 0);
3208 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3210 /* Determine if all our children are holes */
3211 for (i
= 0, bp
= db
->db
.db_data
; i
< 1 << epbs
; i
++, bp
++) {
3212 if (!BP_IS_HOLE(bp
))
3217 * If all the children are holes, then zero them all out so that
3218 * we may get compressed away.
3220 if (i
== 1 << epbs
) {
3221 /* didn't find any non-holes */
3222 bzero(db
->db
.db_data
, db
->db
.db_size
);
3228 * The SPA will call this callback several times for each zio - once
3229 * for every physical child i/o (zio->io_phys_children times). This
3230 * allows the DMU to monitor the progress of each logical i/o. For example,
3231 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3232 * block. There may be a long delay before all copies/fragments are completed,
3233 * so this callback allows us to retire dirty space gradually, as the physical
3238 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3240 dmu_buf_impl_t
*db
= arg
;
3241 objset_t
*os
= db
->db_objset
;
3242 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3243 dbuf_dirty_record_t
*dr
;
3246 dr
= db
->db_data_pending
;
3247 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3250 * The callback will be called io_phys_children times. Retire one
3251 * portion of our dirty space each time we are called. Any rounding
3252 * error will be cleaned up by dsl_pool_sync()'s call to
3253 * dsl_pool_undirty_space().
3255 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3256 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3261 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3263 dmu_buf_impl_t
*db
= vdb
;
3264 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3265 blkptr_t
*bp
= db
->db_blkptr
;
3266 objset_t
*os
= db
->db_objset
;
3267 dmu_tx_t
*tx
= os
->os_synctx
;
3268 dbuf_dirty_record_t
**drp
, *dr
;
3270 ASSERT0(zio
->io_error
);
3271 ASSERT(db
->db_blkptr
== bp
);
3274 * For nopwrites and rewrites we ensure that the bp matches our
3275 * original and bypass all the accounting.
3277 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3278 ASSERT(BP_EQUAL(bp
, bp_orig
));
3280 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3281 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3282 dsl_dataset_block_born(ds
, bp
, tx
);
3285 mutex_enter(&db
->db_mtx
);
3289 drp
= &db
->db_last_dirty
;
3290 while ((dr
= *drp
) != db
->db_data_pending
)
3292 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3293 ASSERT(dr
->dr_dbuf
== db
);
3294 ASSERT(dr
->dr_next
== NULL
);
3298 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3303 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3304 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3305 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3310 if (db
->db_level
== 0) {
3311 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3312 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3313 if (db
->db_state
!= DB_NOFILL
) {
3314 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3315 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
3317 else if (!arc_released(db
->db_buf
))
3318 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3325 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3326 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3327 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3328 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3330 ASSERT3U(db
->db_blkid
, <=,
3331 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3332 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3334 if (!arc_released(db
->db_buf
))
3335 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3338 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3339 list_destroy(&dr
->dt
.di
.dr_children
);
3341 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3343 cv_broadcast(&db
->db_changed
);
3344 ASSERT(db
->db_dirtycnt
> 0);
3345 db
->db_dirtycnt
-= 1;
3346 db
->db_data_pending
= NULL
;
3347 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3351 dbuf_write_nofill_ready(zio_t
*zio
)
3353 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3357 dbuf_write_nofill_done(zio_t
*zio
)
3359 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3363 dbuf_write_override_ready(zio_t
*zio
)
3365 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3366 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3368 dbuf_write_ready(zio
, NULL
, db
);
3372 dbuf_write_override_done(zio_t
*zio
)
3374 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3375 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3376 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3378 mutex_enter(&db
->db_mtx
);
3379 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3380 if (!BP_IS_HOLE(obp
))
3381 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3382 arc_release(dr
->dt
.dl
.dr_data
, db
);
3384 mutex_exit(&db
->db_mtx
);
3386 dbuf_write_done(zio
, NULL
, db
);
3389 /* Issue I/O to commit a dirty buffer to disk. */
3391 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3393 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3396 dmu_buf_impl_t
*parent
= db
->db_parent
;
3397 uint64_t txg
= tx
->tx_txg
;
3398 zbookmark_phys_t zb
;
3403 ASSERT(dmu_tx_is_syncing(tx
));
3409 if (db
->db_state
!= DB_NOFILL
) {
3410 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3412 * Private object buffers are released here rather
3413 * than in dbuf_dirty() since they are only modified
3414 * in the syncing context and we don't want the
3415 * overhead of making multiple copies of the data.
3417 if (BP_IS_HOLE(db
->db_blkptr
)) {
3420 dbuf_release_bp(db
);
3425 if (parent
!= dn
->dn_dbuf
) {
3426 /* Our parent is an indirect block. */
3427 /* We have a dirty parent that has been scheduled for write. */
3428 ASSERT(parent
&& parent
->db_data_pending
);
3429 /* Our parent's buffer is one level closer to the dnode. */
3430 ASSERT(db
->db_level
== parent
->db_level
-1);
3432 * We're about to modify our parent's db_data by modifying
3433 * our block pointer, so the parent must be released.
3435 ASSERT(arc_released(parent
->db_buf
));
3436 zio
= parent
->db_data_pending
->dr_zio
;
3438 /* Our parent is the dnode itself. */
3439 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3440 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3441 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3442 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3443 ASSERT3P(db
->db_blkptr
, ==,
3444 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3448 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3449 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3452 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3453 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3454 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3456 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3458 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3460 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3464 * We copy the blkptr now (rather than when we instantiate the dirty
3465 * record), because its value can change between open context and
3466 * syncing context. We do not need to hold dn_struct_rwlock to read
3467 * db_blkptr because we are in syncing context.
3469 dr
->dr_bp_copy
= *db
->db_blkptr
;
3471 if (db
->db_level
== 0 &&
3472 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3474 * The BP for this block has been provided by open context
3475 * (by dmu_sync() or dmu_buf_write_embedded()).
3477 void *contents
= (data
!= NULL
) ? data
->b_data
: NULL
;
3479 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3480 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, &zp
,
3481 dbuf_write_override_ready
, NULL
, NULL
,
3482 dbuf_write_override_done
,
3483 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3484 mutex_enter(&db
->db_mtx
);
3485 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3486 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3487 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3488 mutex_exit(&db
->db_mtx
);
3489 } else if (db
->db_state
== DB_NOFILL
) {
3490 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
);
3491 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3492 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, &zp
,
3493 dbuf_write_nofill_ready
, NULL
, NULL
,
3494 dbuf_write_nofill_done
, db
,
3495 ZIO_PRIORITY_ASYNC_WRITE
,
3496 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3498 arc_done_func_t
*children_ready_cb
= NULL
;
3499 ASSERT(arc_released(data
));
3502 * For indirect blocks, we want to setup the children
3503 * ready callback so that we can properly handle an indirect
3504 * block that only contains holes.
3506 if (db
->db_level
!= 0)
3507 children_ready_cb
= dbuf_write_children_ready
;
3509 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3510 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
3511 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dbuf_write_ready
,
3513 dbuf_write_physdone
, dbuf_write_done
, db
,
3514 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3518 #if defined(_KERNEL) && defined(HAVE_SPL)
3519 EXPORT_SYMBOL(dbuf_find
);
3520 EXPORT_SYMBOL(dbuf_is_metadata
);
3521 EXPORT_SYMBOL(dbuf_evict
);
3522 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
3523 EXPORT_SYMBOL(dbuf_whichblock
);
3524 EXPORT_SYMBOL(dbuf_read
);
3525 EXPORT_SYMBOL(dbuf_unoverride
);
3526 EXPORT_SYMBOL(dbuf_free_range
);
3527 EXPORT_SYMBOL(dbuf_new_size
);
3528 EXPORT_SYMBOL(dbuf_release_bp
);
3529 EXPORT_SYMBOL(dbuf_dirty
);
3530 EXPORT_SYMBOL(dmu_buf_will_dirty
);
3531 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
3532 EXPORT_SYMBOL(dmu_buf_will_fill
);
3533 EXPORT_SYMBOL(dmu_buf_fill_done
);
3534 EXPORT_SYMBOL(dmu_buf_rele
);
3535 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
3536 EXPORT_SYMBOL(dbuf_clear
);
3537 EXPORT_SYMBOL(dbuf_prefetch
);
3538 EXPORT_SYMBOL(dbuf_hold_impl
);
3539 EXPORT_SYMBOL(dbuf_hold
);
3540 EXPORT_SYMBOL(dbuf_hold_level
);
3541 EXPORT_SYMBOL(dbuf_create_bonus
);
3542 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
3543 EXPORT_SYMBOL(dbuf_rm_spill
);
3544 EXPORT_SYMBOL(dbuf_add_ref
);
3545 EXPORT_SYMBOL(dbuf_rele
);
3546 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
3547 EXPORT_SYMBOL(dbuf_refcount
);
3548 EXPORT_SYMBOL(dbuf_sync_list
);
3549 EXPORT_SYMBOL(dmu_buf_set_user
);
3550 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3551 EXPORT_SYMBOL(dmu_buf_get_user
);
3552 EXPORT_SYMBOL(dmu_buf_freeable
);
3553 EXPORT_SYMBOL(dmu_buf_get_blkptr
);