4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_dbuf.h>
49 struct dbuf_hold_impl_data
{
50 /* Function arguments */
54 boolean_t dh_fail_sparse
;
55 boolean_t dh_fail_uncached
;
57 dmu_buf_impl_t
**dh_dbp
;
59 dmu_buf_impl_t
*dh_db
;
60 dmu_buf_impl_t
*dh_parent
;
63 dbuf_dirty_record_t
*dh_dr
;
64 arc_buf_contents_t dh_type
;
68 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
69 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, boolean_t fail_sparse
,
70 boolean_t fail_uncached
,
71 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
72 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
75 * Number of times that zfs_free_range() took the slow path while doing
76 * a zfs receive. A nonzero value indicates a potential performance problem.
78 uint64_t zfs_free_range_recv_miss
;
80 static void dbuf_destroy(dmu_buf_impl_t
*db
);
81 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
82 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
85 extern inline void dmu_buf_init_user(dmu_buf_user_t
*dbu
,
86 dmu_buf_evict_func_t
*evict_func
, dmu_buf_t
**clear_on_evict_dbufp
);
90 * Global data structures and functions for the dbuf cache.
92 static kmem_cache_t
*dbuf_cache
;
93 static taskq_t
*dbu_evict_taskq
;
97 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
99 dmu_buf_impl_t
*db
= vdb
;
100 bzero(db
, sizeof (dmu_buf_impl_t
));
102 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
103 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
104 refcount_create(&db
->db_holds
);
111 dbuf_dest(void *vdb
, void *unused
)
113 dmu_buf_impl_t
*db
= vdb
;
114 mutex_destroy(&db
->db_mtx
);
115 cv_destroy(&db
->db_changed
);
116 refcount_destroy(&db
->db_holds
);
120 * dbuf hash table routines
122 static dbuf_hash_table_t dbuf_hash_table
;
124 static uint64_t dbuf_hash_count
;
127 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
129 uintptr_t osv
= (uintptr_t)os
;
130 uint64_t crc
= -1ULL;
132 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
133 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
134 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
135 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
136 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
137 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
138 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
140 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
145 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
147 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
148 ((dbuf)->db.db_object == (obj) && \
149 (dbuf)->db_objset == (os) && \
150 (dbuf)->db_level == (level) && \
151 (dbuf)->db_blkid == (blkid))
154 dbuf_find(objset_t
*os
, uint64_t obj
, uint8_t level
, uint64_t blkid
)
156 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
161 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
162 idx
= hv
& h
->hash_table_mask
;
164 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
165 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
166 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
167 mutex_enter(&db
->db_mtx
);
168 if (db
->db_state
!= DB_EVICTING
) {
169 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
172 mutex_exit(&db
->db_mtx
);
175 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
179 static dmu_buf_impl_t
*
180 dbuf_find_bonus(objset_t
*os
, uint64_t object
)
183 dmu_buf_impl_t
*db
= NULL
;
185 if (dnode_hold(os
, object
, FTAG
, &dn
) == 0) {
186 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
187 if (dn
->dn_bonus
!= NULL
) {
189 mutex_enter(&db
->db_mtx
);
191 rw_exit(&dn
->dn_struct_rwlock
);
192 dnode_rele(dn
, FTAG
);
198 * Insert an entry into the hash table. If there is already an element
199 * equal to elem in the hash table, then the already existing element
200 * will be returned and the new element will not be inserted.
201 * Otherwise returns NULL.
203 static dmu_buf_impl_t
*
204 dbuf_hash_insert(dmu_buf_impl_t
*db
)
206 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
207 objset_t
*os
= db
->db_objset
;
208 uint64_t obj
= db
->db
.db_object
;
209 int level
= db
->db_level
;
210 uint64_t blkid
, hv
, idx
;
213 blkid
= db
->db_blkid
;
214 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
215 idx
= hv
& h
->hash_table_mask
;
217 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
218 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
219 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
220 mutex_enter(&dbf
->db_mtx
);
221 if (dbf
->db_state
!= DB_EVICTING
) {
222 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
225 mutex_exit(&dbf
->db_mtx
);
229 mutex_enter(&db
->db_mtx
);
230 db
->db_hash_next
= h
->hash_table
[idx
];
231 h
->hash_table
[idx
] = db
;
232 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
233 atomic_add_64(&dbuf_hash_count
, 1);
239 * Remove an entry from the hash table. It must be in the EVICTING state.
242 dbuf_hash_remove(dmu_buf_impl_t
*db
)
244 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
246 dmu_buf_impl_t
*dbf
, **dbp
;
248 hv
= DBUF_HASH(db
->db_objset
, db
->db
.db_object
,
249 db
->db_level
, db
->db_blkid
);
250 idx
= hv
& h
->hash_table_mask
;
253 * We musn't hold db_mtx to maintain lock ordering:
254 * DBUF_HASH_MUTEX > db_mtx.
256 ASSERT(refcount_is_zero(&db
->db_holds
));
257 ASSERT(db
->db_state
== DB_EVICTING
);
258 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
260 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
261 dbp
= &h
->hash_table
[idx
];
262 while ((dbf
= *dbp
) != db
) {
263 dbp
= &dbf
->db_hash_next
;
266 *dbp
= db
->db_hash_next
;
267 db
->db_hash_next
= NULL
;
268 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
269 atomic_add_64(&dbuf_hash_count
, -1);
272 static arc_evict_func_t dbuf_do_evict
;
277 } dbvu_verify_type_t
;
280 dbuf_verify_user(dmu_buf_impl_t
*db
, dbvu_verify_type_t verify_type
)
285 if (db
->db_user
== NULL
)
288 /* Only data blocks support the attachment of user data. */
289 ASSERT(db
->db_level
== 0);
291 /* Clients must resolve a dbuf before attaching user data. */
292 ASSERT(db
->db
.db_data
!= NULL
);
293 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
295 holds
= refcount_count(&db
->db_holds
);
296 if (verify_type
== DBVU_EVICTING
) {
298 * Immediate eviction occurs when holds == dirtycnt.
299 * For normal eviction buffers, holds is zero on
300 * eviction, except when dbuf_fix_old_data() calls
301 * dbuf_clear_data(). However, the hold count can grow
302 * during eviction even though db_mtx is held (see
303 * dmu_bonus_hold() for an example), so we can only
304 * test the generic invariant that holds >= dirtycnt.
306 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
308 if (db
->db_user_immediate_evict
== TRUE
)
309 ASSERT3U(holds
, >=, db
->db_dirtycnt
);
311 ASSERT3U(holds
, >, 0);
317 dbuf_evict_user(dmu_buf_impl_t
*db
)
319 dmu_buf_user_t
*dbu
= db
->db_user
;
321 ASSERT(MUTEX_HELD(&db
->db_mtx
));
326 dbuf_verify_user(db
, DBVU_EVICTING
);
330 if (dbu
->dbu_clear_on_evict_dbufp
!= NULL
)
331 *dbu
->dbu_clear_on_evict_dbufp
= NULL
;
335 * Invoke the callback from a taskq to avoid lock order reversals
336 * and limit stack depth.
338 taskq_dispatch_ent(dbu_evict_taskq
, dbu
->dbu_evict_func
, dbu
, 0,
343 dbuf_is_metadata(dmu_buf_impl_t
*db
)
346 * Consider indirect blocks and spill blocks to be meta data.
348 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
351 boolean_t is_metadata
;
354 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
357 return (is_metadata
);
362 dbuf_evict(dmu_buf_impl_t
*db
)
364 ASSERT(MUTEX_HELD(&db
->db_mtx
));
365 ASSERT(db
->db_buf
== NULL
);
366 ASSERT(db
->db_data_pending
== NULL
);
375 uint64_t hsize
= 1ULL << 16;
376 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
380 * The hash table is big enough to fill all of physical memory
381 * with an average block size of zfs_arc_average_blocksize (default 8K).
382 * By default, the table will take up
383 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
385 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
389 h
->hash_table_mask
= hsize
- 1;
390 #if defined(_KERNEL) && defined(HAVE_SPL)
392 * Large allocations which do not require contiguous pages
393 * should be using vmem_alloc() in the linux kernel
395 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_SLEEP
);
397 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
399 if (h
->hash_table
== NULL
) {
400 /* XXX - we should really return an error instead of assert */
401 ASSERT(hsize
> (1ULL << 10));
406 dbuf_cache
= kmem_cache_create("dmu_buf_impl_t",
407 sizeof (dmu_buf_impl_t
),
408 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
410 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
411 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
416 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
417 * configuration is not required.
419 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
425 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
428 dbuf_stats_destroy();
430 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
431 mutex_destroy(&h
->hash_mutexes
[i
]);
432 #if defined(_KERNEL) && defined(HAVE_SPL)
434 * Large allocations which do not require contiguous pages
435 * should be using vmem_free() in the linux kernel
437 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
439 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
441 kmem_cache_destroy(dbuf_cache
);
442 taskq_destroy(dbu_evict_taskq
);
451 dbuf_verify(dmu_buf_impl_t
*db
)
454 dbuf_dirty_record_t
*dr
;
456 ASSERT(MUTEX_HELD(&db
->db_mtx
));
458 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
461 ASSERT(db
->db_objset
!= NULL
);
465 ASSERT(db
->db_parent
== NULL
);
466 ASSERT(db
->db_blkptr
== NULL
);
468 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
469 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
470 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
471 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
472 db
->db_blkid
== DMU_SPILL_BLKID
||
473 !avl_is_empty(&dn
->dn_dbufs
));
475 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
477 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
478 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
479 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
481 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
482 ASSERT0(db
->db
.db_offset
);
484 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
487 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
488 ASSERT(dr
->dr_dbuf
== db
);
490 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
491 ASSERT(dr
->dr_dbuf
== db
);
494 * We can't assert that db_size matches dn_datablksz because it
495 * can be momentarily different when another thread is doing
498 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
499 dr
= db
->db_data_pending
;
501 * It should only be modified in syncing context, so
502 * make sure we only have one copy of the data.
504 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
507 /* verify db->db_blkptr */
509 if (db
->db_parent
== dn
->dn_dbuf
) {
510 /* db is pointed to by the dnode */
511 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
512 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
513 ASSERT(db
->db_parent
== NULL
);
515 ASSERT(db
->db_parent
!= NULL
);
516 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
517 ASSERT3P(db
->db_blkptr
, ==,
518 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
520 /* db is pointed to by an indirect block */
521 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
523 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
524 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
527 * dnode_grow_indblksz() can make this fail if we don't
528 * have the struct_rwlock. XXX indblksz no longer
529 * grows. safe to do this now?
531 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
532 ASSERT3P(db
->db_blkptr
, ==,
533 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
534 db
->db_blkid
% epb
));
538 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
539 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
540 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
541 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
543 * If the blkptr isn't set but they have nonzero data,
544 * it had better be dirty, otherwise we'll lose that
545 * data when we evict this buffer.
547 if (db
->db_dirtycnt
== 0) {
548 ASSERTV(uint64_t *buf
= db
->db
.db_data
);
551 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
561 dbuf_clear_data(dmu_buf_impl_t
*db
)
563 ASSERT(MUTEX_HELD(&db
->db_mtx
));
566 db
->db
.db_data
= NULL
;
567 if (db
->db_state
!= DB_NOFILL
)
568 db
->db_state
= DB_UNCACHED
;
572 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
574 ASSERT(MUTEX_HELD(&db
->db_mtx
));
578 ASSERT(buf
->b_data
!= NULL
);
579 db
->db
.db_data
= buf
->b_data
;
580 if (!arc_released(buf
))
581 arc_set_callback(buf
, dbuf_do_evict
, db
);
585 * Loan out an arc_buf for read. Return the loaned arc_buf.
588 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
592 mutex_enter(&db
->db_mtx
);
593 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
594 int blksz
= db
->db
.db_size
;
595 spa_t
*spa
= db
->db_objset
->os_spa
;
597 mutex_exit(&db
->db_mtx
);
598 abuf
= arc_loan_buf(spa
, blksz
);
599 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
602 arc_loan_inuse_buf(abuf
, db
);
604 mutex_exit(&db
->db_mtx
);
610 * Calculate which level n block references the data at the level 0 offset
614 dbuf_whichblock(dnode_t
*dn
, int64_t level
, uint64_t offset
)
616 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
618 * The level n blkid is equal to the level 0 blkid divided by
619 * the number of level 0s in a level n block.
621 * The level 0 blkid is offset >> datablkshift =
622 * offset / 2^datablkshift.
624 * The number of level 0s in a level n is the number of block
625 * pointers in an indirect block, raised to the power of level.
626 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
627 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
629 * Thus, the level n blkid is: offset /
630 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
631 * = offset / 2^(datablkshift + level *
632 * (indblkshift - SPA_BLKPTRSHIFT))
633 * = offset >> (datablkshift + level *
634 * (indblkshift - SPA_BLKPTRSHIFT))
636 return (offset
>> (dn
->dn_datablkshift
+ level
*
637 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
)));
639 ASSERT3U(offset
, <, dn
->dn_datablksz
);
645 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
647 dmu_buf_impl_t
*db
= vdb
;
649 mutex_enter(&db
->db_mtx
);
650 ASSERT3U(db
->db_state
, ==, DB_READ
);
652 * All reads are synchronous, so we must have a hold on the dbuf
654 ASSERT(refcount_count(&db
->db_holds
) > 0);
655 ASSERT(db
->db_buf
== NULL
);
656 ASSERT(db
->db
.db_data
== NULL
);
657 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
658 /* we were freed in flight; disregard any error */
659 arc_release(buf
, db
);
660 bzero(buf
->b_data
, db
->db
.db_size
);
662 db
->db_freed_in_flight
= FALSE
;
663 dbuf_set_data(db
, buf
);
664 db
->db_state
= DB_CACHED
;
665 } else if (zio
== NULL
|| zio
->io_error
== 0) {
666 dbuf_set_data(db
, buf
);
667 db
->db_state
= DB_CACHED
;
669 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
670 ASSERT3P(db
->db_buf
, ==, NULL
);
671 VERIFY(arc_buf_remove_ref(buf
, db
));
672 db
->db_state
= DB_UNCACHED
;
674 cv_broadcast(&db
->db_changed
);
675 dbuf_rele_and_unlock(db
, NULL
);
679 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
683 uint32_t aflags
= ARC_FLAG_NOWAIT
;
688 ASSERT(!refcount_is_zero(&db
->db_holds
));
689 /* We need the struct_rwlock to prevent db_blkptr from changing. */
690 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
691 ASSERT(MUTEX_HELD(&db
->db_mtx
));
692 ASSERT(db
->db_state
== DB_UNCACHED
);
693 ASSERT(db
->db_buf
== NULL
);
695 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
696 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
698 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
699 db
->db
.db_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
700 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
701 if (bonuslen
< DN_MAX_BONUSLEN
)
702 bzero(db
->db
.db_data
, DN_MAX_BONUSLEN
);
704 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
706 db
->db_state
= DB_CACHED
;
707 mutex_exit(&db
->db_mtx
);
712 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
713 * processes the delete record and clears the bp while we are waiting
714 * for the dn_mtx (resulting in a "no" from block_freed).
716 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
717 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
718 BP_IS_HOLE(db
->db_blkptr
)))) {
719 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
722 dbuf_set_data(db
, arc_buf_alloc(db
->db_objset
->os_spa
,
723 db
->db
.db_size
, db
, type
));
724 bzero(db
->db
.db_data
, db
->db
.db_size
);
725 db
->db_state
= DB_CACHED
;
726 mutex_exit(&db
->db_mtx
);
732 db
->db_state
= DB_READ
;
733 mutex_exit(&db
->db_mtx
);
735 if (DBUF_IS_L2CACHEABLE(db
))
736 aflags
|= ARC_FLAG_L2CACHE
;
737 if (DBUF_IS_L2COMPRESSIBLE(db
))
738 aflags
|= ARC_FLAG_L2COMPRESS
;
740 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
741 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
742 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
744 dbuf_add_ref(db
, NULL
);
746 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
747 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
748 (flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
751 return (SET_ERROR(err
));
755 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
758 boolean_t havepzio
= (zio
!= NULL
);
763 * We don't have to hold the mutex to check db_state because it
764 * can't be freed while we have a hold on the buffer.
766 ASSERT(!refcount_is_zero(&db
->db_holds
));
768 if (db
->db_state
== DB_NOFILL
)
769 return (SET_ERROR(EIO
));
773 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
774 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
776 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
777 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
778 DBUF_IS_CACHEABLE(db
);
780 mutex_enter(&db
->db_mtx
);
781 if (db
->db_state
== DB_CACHED
) {
782 mutex_exit(&db
->db_mtx
);
784 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1);
785 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
786 rw_exit(&dn
->dn_struct_rwlock
);
788 } else if (db
->db_state
== DB_UNCACHED
) {
789 spa_t
*spa
= dn
->dn_objset
->os_spa
;
792 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
794 err
= dbuf_read_impl(db
, zio
, flags
);
796 /* dbuf_read_impl has dropped db_mtx for us */
798 if (!err
&& prefetch
)
799 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1);
801 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
802 rw_exit(&dn
->dn_struct_rwlock
);
805 if (!err
&& !havepzio
)
809 * Another reader came in while the dbuf was in flight
810 * between UNCACHED and CACHED. Either a writer will finish
811 * writing the buffer (sending the dbuf to CACHED) or the
812 * first reader's request will reach the read_done callback
813 * and send the dbuf to CACHED. Otherwise, a failure
814 * occurred and the dbuf went to UNCACHED.
816 mutex_exit(&db
->db_mtx
);
818 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1);
819 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
820 rw_exit(&dn
->dn_struct_rwlock
);
823 /* Skip the wait per the caller's request. */
824 mutex_enter(&db
->db_mtx
);
825 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
826 while (db
->db_state
== DB_READ
||
827 db
->db_state
== DB_FILL
) {
828 ASSERT(db
->db_state
== DB_READ
||
829 (flags
& DB_RF_HAVESTRUCT
) == 0);
830 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
832 cv_wait(&db
->db_changed
, &db
->db_mtx
);
834 if (db
->db_state
== DB_UNCACHED
)
835 err
= SET_ERROR(EIO
);
837 mutex_exit(&db
->db_mtx
);
840 ASSERT(err
|| havepzio
|| db
->db_state
== DB_CACHED
);
845 dbuf_noread(dmu_buf_impl_t
*db
)
847 ASSERT(!refcount_is_zero(&db
->db_holds
));
848 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
849 mutex_enter(&db
->db_mtx
);
850 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
851 cv_wait(&db
->db_changed
, &db
->db_mtx
);
852 if (db
->db_state
== DB_UNCACHED
) {
853 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
854 spa_t
*spa
= db
->db_objset
->os_spa
;
856 ASSERT(db
->db_buf
== NULL
);
857 ASSERT(db
->db
.db_data
== NULL
);
858 dbuf_set_data(db
, arc_buf_alloc(spa
, db
->db
.db_size
, db
, type
));
859 db
->db_state
= DB_FILL
;
860 } else if (db
->db_state
== DB_NOFILL
) {
863 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
865 mutex_exit(&db
->db_mtx
);
869 * This is our just-in-time copy function. It makes a copy of
870 * buffers, that have been modified in a previous transaction
871 * group, before we modify them in the current active group.
873 * This function is used in two places: when we are dirtying a
874 * buffer for the first time in a txg, and when we are freeing
875 * a range in a dnode that includes this buffer.
877 * Note that when we are called from dbuf_free_range() we do
878 * not put a hold on the buffer, we just traverse the active
879 * dbuf list for the dnode.
882 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
884 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
886 ASSERT(MUTEX_HELD(&db
->db_mtx
));
887 ASSERT(db
->db
.db_data
!= NULL
);
888 ASSERT(db
->db_level
== 0);
889 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
892 (dr
->dt
.dl
.dr_data
!=
893 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
897 * If the last dirty record for this dbuf has not yet synced
898 * and its referencing the dbuf data, either:
899 * reset the reference to point to a new copy,
900 * or (if there a no active holders)
901 * just null out the current db_data pointer.
903 ASSERT(dr
->dr_txg
>= txg
- 2);
904 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
905 /* Note that the data bufs here are zio_bufs */
906 dr
->dt
.dl
.dr_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
907 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
908 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, DN_MAX_BONUSLEN
);
909 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
910 int size
= db
->db
.db_size
;
911 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
912 spa_t
*spa
= db
->db_objset
->os_spa
;
914 dr
->dt
.dl
.dr_data
= arc_buf_alloc(spa
, size
, db
, type
);
915 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
922 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
924 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
925 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
926 uint64_t txg
= dr
->dr_txg
;
928 ASSERT(MUTEX_HELD(&db
->db_mtx
));
929 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
930 ASSERT(db
->db_level
== 0);
932 if (db
->db_blkid
== DMU_BONUS_BLKID
||
933 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
936 ASSERT(db
->db_data_pending
!= dr
);
938 /* free this block */
939 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
940 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
942 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
943 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
946 * Release the already-written buffer, so we leave it in
947 * a consistent dirty state. Note that all callers are
948 * modifying the buffer, so they will immediately do
949 * another (redundant) arc_release(). Therefore, leave
950 * the buf thawed to save the effort of freezing &
951 * immediately re-thawing it.
953 arc_release(dr
->dt
.dl
.dr_data
, db
);
957 * Evict (if its unreferenced) or clear (if its referenced) any level-0
958 * data blocks in the free range, so that any future readers will find
961 * This is a no-op if the dataset is in the middle of an incremental
962 * receive; see comment below for details.
965 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
968 dmu_buf_impl_t
*db_search
;
969 dmu_buf_impl_t
*db
, *db_next
;
970 uint64_t txg
= tx
->tx_txg
;
972 boolean_t freespill
=
973 (start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
);
975 if (end_blkid
> dn
->dn_maxblkid
&& !freespill
)
976 end_blkid
= dn
->dn_maxblkid
;
977 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
979 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
980 db_search
->db_level
= 0;
981 db_search
->db_blkid
= start_blkid
;
982 db_search
->db_state
= DB_SEARCH
;
984 mutex_enter(&dn
->dn_dbufs_mtx
);
985 if (start_blkid
>= dn
->dn_unlisted_l0_blkid
&& !freespill
) {
986 /* There can't be any dbufs in this range; no need to search. */
988 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
989 ASSERT3P(db
, ==, NULL
);
990 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
991 ASSERT(db
== NULL
|| db
->db_level
> 0);
994 } else if (dmu_objset_is_receiving(dn
->dn_objset
)) {
996 * If we are receiving, we expect there to be no dbufs in
997 * the range to be freed, because receive modifies each
998 * block at most once, and in offset order. If this is
999 * not the case, it can lead to performance problems,
1000 * so note that we unexpectedly took the slow path.
1002 atomic_inc_64(&zfs_free_range_recv_miss
);
1005 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1006 ASSERT3P(db
, ==, NULL
);
1007 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1009 for (; db
!= NULL
; db
= db_next
) {
1010 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1011 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1013 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1016 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1018 /* found a level 0 buffer in the range */
1019 mutex_enter(&db
->db_mtx
);
1020 if (dbuf_undirty(db
, tx
)) {
1021 /* mutex has been dropped and dbuf destroyed */
1025 if (db
->db_state
== DB_UNCACHED
||
1026 db
->db_state
== DB_NOFILL
||
1027 db
->db_state
== DB_EVICTING
) {
1028 ASSERT(db
->db
.db_data
== NULL
);
1029 mutex_exit(&db
->db_mtx
);
1032 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1033 /* will be handled in dbuf_read_done or dbuf_rele */
1034 db
->db_freed_in_flight
= TRUE
;
1035 mutex_exit(&db
->db_mtx
);
1038 if (refcount_count(&db
->db_holds
) == 0) {
1043 /* The dbuf is referenced */
1045 if (db
->db_last_dirty
!= NULL
) {
1046 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1048 if (dr
->dr_txg
== txg
) {
1050 * This buffer is "in-use", re-adjust the file
1051 * size to reflect that this buffer may
1052 * contain new data when we sync.
1054 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1055 db
->db_blkid
> dn
->dn_maxblkid
)
1056 dn
->dn_maxblkid
= db
->db_blkid
;
1057 dbuf_unoverride(dr
);
1060 * This dbuf is not dirty in the open context.
1061 * Either uncache it (if its not referenced in
1062 * the open context) or reset its contents to
1065 dbuf_fix_old_data(db
, txg
);
1068 /* clear the contents if its cached */
1069 if (db
->db_state
== DB_CACHED
) {
1070 ASSERT(db
->db
.db_data
!= NULL
);
1071 arc_release(db
->db_buf
, db
);
1072 bzero(db
->db
.db_data
, db
->db
.db_size
);
1073 arc_buf_freeze(db
->db_buf
);
1076 mutex_exit(&db
->db_mtx
);
1080 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1081 mutex_exit(&dn
->dn_dbufs_mtx
);
1085 dbuf_block_freeable(dmu_buf_impl_t
*db
)
1087 dsl_dataset_t
*ds
= db
->db_objset
->os_dsl_dataset
;
1088 uint64_t birth_txg
= 0;
1091 * We don't need any locking to protect db_blkptr:
1092 * If it's syncing, then db_last_dirty will be set
1093 * so we'll ignore db_blkptr.
1095 * This logic ensures that only block births for
1096 * filled blocks are considered.
1098 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1099 if (db
->db_last_dirty
&& (db
->db_blkptr
== NULL
||
1100 !BP_IS_HOLE(db
->db_blkptr
))) {
1101 birth_txg
= db
->db_last_dirty
->dr_txg
;
1102 } else if (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1103 birth_txg
= db
->db_blkptr
->blk_birth
;
1107 * If this block don't exist or is in a snapshot, it can't be freed.
1108 * Don't pass the bp to dsl_dataset_block_freeable() since we
1109 * are holding the db_mtx lock and might deadlock if we are
1110 * prefetching a dedup-ed block.
1113 return (ds
== NULL
||
1114 dsl_dataset_block_freeable(ds
, NULL
, birth_txg
));
1120 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1122 arc_buf_t
*buf
, *obuf
;
1123 int osize
= db
->db
.db_size
;
1124 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1127 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1132 /* XXX does *this* func really need the lock? */
1133 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1136 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1137 * is OK, because there can be no other references to the db
1138 * when we are changing its size, so no concurrent DB_FILL can
1142 * XXX we should be doing a dbuf_read, checking the return
1143 * value and returning that up to our callers
1145 dmu_buf_will_dirty(&db
->db
, tx
);
1147 /* create the data buffer for the new block */
1148 buf
= arc_buf_alloc(dn
->dn_objset
->os_spa
, size
, db
, type
);
1150 /* copy old block data to the new block */
1152 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1153 /* zero the remainder */
1155 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1157 mutex_enter(&db
->db_mtx
);
1158 dbuf_set_data(db
, buf
);
1159 VERIFY(arc_buf_remove_ref(obuf
, db
));
1160 db
->db
.db_size
= size
;
1162 if (db
->db_level
== 0) {
1163 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1164 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1166 mutex_exit(&db
->db_mtx
);
1168 dnode_willuse_space(dn
, size
-osize
, tx
);
1173 dbuf_release_bp(dmu_buf_impl_t
*db
)
1175 ASSERTV(objset_t
*os
= db
->db_objset
);
1177 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1178 ASSERT(arc_released(os
->os_phys_buf
) ||
1179 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1180 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1182 (void) arc_release(db
->db_buf
, db
);
1186 * We already have a dirty record for this TXG, and we are being
1190 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1192 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1194 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1196 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1198 * If this buffer has already been written out,
1199 * we now need to reset its state.
1201 dbuf_unoverride(dr
);
1202 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1203 db
->db_state
!= DB_NOFILL
) {
1204 /* Already released on initial dirty, so just thaw. */
1205 ASSERT(arc_released(db
->db_buf
));
1206 arc_buf_thaw(db
->db_buf
);
1211 dbuf_dirty_record_t
*
1212 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1216 dbuf_dirty_record_t
**drp
, *dr
;
1217 int drop_struct_lock
= FALSE
;
1218 boolean_t do_free_accounting
= B_FALSE
;
1219 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1221 ASSERT(tx
->tx_txg
!= 0);
1222 ASSERT(!refcount_is_zero(&db
->db_holds
));
1223 DMU_TX_DIRTY_BUF(tx
, db
);
1228 * Shouldn't dirty a regular buffer in syncing context. Private
1229 * objects may be dirtied in syncing context, but only if they
1230 * were already pre-dirtied in open context.
1232 ASSERT(!dmu_tx_is_syncing(tx
) ||
1233 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1234 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1235 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1237 * We make this assert for private objects as well, but after we
1238 * check if we're already dirty. They are allowed to re-dirty
1239 * in syncing context.
1241 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1242 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1243 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1245 mutex_enter(&db
->db_mtx
);
1247 * XXX make this true for indirects too? The problem is that
1248 * transactions created with dmu_tx_create_assigned() from
1249 * syncing context don't bother holding ahead.
1251 ASSERT(db
->db_level
!= 0 ||
1252 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1253 db
->db_state
== DB_NOFILL
);
1255 mutex_enter(&dn
->dn_mtx
);
1257 * Don't set dirtyctx to SYNC if we're just modifying this as we
1258 * initialize the objset.
1260 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
&&
1261 !BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1263 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1264 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1265 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1267 mutex_exit(&dn
->dn_mtx
);
1269 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1270 dn
->dn_have_spill
= B_TRUE
;
1273 * If this buffer is already dirty, we're done.
1275 drp
= &db
->db_last_dirty
;
1276 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1277 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1278 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1280 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1284 mutex_exit(&db
->db_mtx
);
1289 * Only valid if not already dirty.
1291 ASSERT(dn
->dn_object
== 0 ||
1292 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1293 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1295 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1296 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1297 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1298 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1299 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1300 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1303 * We should only be dirtying in syncing context if it's the
1304 * mos or we're initializing the os or it's a special object.
1305 * However, we are allowed to dirty in syncing context provided
1306 * we already dirtied it in open context. Hence we must make
1307 * this assertion only if we're not already dirty.
1310 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1311 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1312 ASSERT(db
->db
.db_size
!= 0);
1314 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1316 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1318 * Update the accounting.
1319 * Note: we delay "free accounting" until after we drop
1320 * the db_mtx. This keeps us from grabbing other locks
1321 * (and possibly deadlocking) in bp_get_dsize() while
1322 * also holding the db_mtx.
1324 dnode_willuse_space(dn
, db
->db
.db_size
, tx
);
1325 do_free_accounting
= dbuf_block_freeable(db
);
1329 * If this buffer is dirty in an old transaction group we need
1330 * to make a copy of it so that the changes we make in this
1331 * transaction group won't leak out when we sync the older txg.
1333 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1334 list_link_init(&dr
->dr_dirty_node
);
1335 if (db
->db_level
== 0) {
1336 void *data_old
= db
->db_buf
;
1338 if (db
->db_state
!= DB_NOFILL
) {
1339 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1340 dbuf_fix_old_data(db
, tx
->tx_txg
);
1341 data_old
= db
->db
.db_data
;
1342 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1344 * Release the data buffer from the cache so
1345 * that we can modify it without impacting
1346 * possible other users of this cached data
1347 * block. Note that indirect blocks and
1348 * private objects are not released until the
1349 * syncing state (since they are only modified
1352 arc_release(db
->db_buf
, db
);
1353 dbuf_fix_old_data(db
, tx
->tx_txg
);
1354 data_old
= db
->db_buf
;
1356 ASSERT(data_old
!= NULL
);
1358 dr
->dt
.dl
.dr_data
= data_old
;
1360 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1361 list_create(&dr
->dt
.di
.dr_children
,
1362 sizeof (dbuf_dirty_record_t
),
1363 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1365 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1366 dr
->dr_accounted
= db
->db
.db_size
;
1368 dr
->dr_txg
= tx
->tx_txg
;
1373 * We could have been freed_in_flight between the dbuf_noread
1374 * and dbuf_dirty. We win, as though the dbuf_noread() had
1375 * happened after the free.
1377 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1378 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1379 mutex_enter(&dn
->dn_mtx
);
1380 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1381 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1384 mutex_exit(&dn
->dn_mtx
);
1385 db
->db_freed_in_flight
= FALSE
;
1389 * This buffer is now part of this txg
1391 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1392 db
->db_dirtycnt
+= 1;
1393 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1395 mutex_exit(&db
->db_mtx
);
1397 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1398 db
->db_blkid
== DMU_SPILL_BLKID
) {
1399 mutex_enter(&dn
->dn_mtx
);
1400 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1401 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1402 mutex_exit(&dn
->dn_mtx
);
1403 dnode_setdirty(dn
, tx
);
1406 } else if (do_free_accounting
) {
1407 blkptr_t
*bp
= db
->db_blkptr
;
1408 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1409 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1411 * This is only a guess -- if the dbuf is dirty
1412 * in a previous txg, we don't know how much
1413 * space it will use on disk yet. We should
1414 * really have the struct_rwlock to access
1415 * db_blkptr, but since this is just a guess,
1416 * it's OK if we get an odd answer.
1418 ddt_prefetch(os
->os_spa
, bp
);
1419 dnode_willuse_space(dn
, -willfree
, tx
);
1422 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1423 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1424 drop_struct_lock
= TRUE
;
1427 if (db
->db_level
== 0) {
1428 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1429 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1432 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1433 dmu_buf_impl_t
*parent
= db
->db_parent
;
1434 dbuf_dirty_record_t
*di
;
1435 int parent_held
= FALSE
;
1437 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1438 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1440 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1441 db
->db_blkid
>> epbs
, FTAG
);
1442 ASSERT(parent
!= NULL
);
1445 if (drop_struct_lock
)
1446 rw_exit(&dn
->dn_struct_rwlock
);
1447 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1448 di
= dbuf_dirty(parent
, tx
);
1450 dbuf_rele(parent
, FTAG
);
1452 mutex_enter(&db
->db_mtx
);
1454 * Since we've dropped the mutex, it's possible that
1455 * dbuf_undirty() might have changed this out from under us.
1457 if (db
->db_last_dirty
== dr
||
1458 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1459 mutex_enter(&di
->dt
.di
.dr_mtx
);
1460 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1461 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1462 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1463 mutex_exit(&di
->dt
.di
.dr_mtx
);
1466 mutex_exit(&db
->db_mtx
);
1468 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1469 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1470 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1471 mutex_enter(&dn
->dn_mtx
);
1472 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1473 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1474 mutex_exit(&dn
->dn_mtx
);
1475 if (drop_struct_lock
)
1476 rw_exit(&dn
->dn_struct_rwlock
);
1479 dnode_setdirty(dn
, tx
);
1485 * Undirty a buffer in the transaction group referenced by the given
1486 * transaction. Return whether this evicted the dbuf.
1489 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1492 uint64_t txg
= tx
->tx_txg
;
1493 dbuf_dirty_record_t
*dr
, **drp
;
1498 * Due to our use of dn_nlevels below, this can only be called
1499 * in open context, unless we are operating on the MOS.
1500 * From syncing context, dn_nlevels may be different from the
1501 * dn_nlevels used when dbuf was dirtied.
1503 ASSERT(db
->db_objset
==
1504 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
1505 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
1506 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1507 ASSERT0(db
->db_level
);
1508 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1511 * If this buffer is not dirty, we're done.
1513 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1514 if (dr
->dr_txg
<= txg
)
1516 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1518 ASSERT(dr
->dr_txg
== txg
);
1519 ASSERT(dr
->dr_dbuf
== db
);
1524 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1526 ASSERT(db
->db
.db_size
!= 0);
1528 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
1529 dr
->dr_accounted
, txg
);
1534 * Note that there are three places in dbuf_dirty()
1535 * where this dirty record may be put on a list.
1536 * Make sure to do a list_remove corresponding to
1537 * every one of those list_insert calls.
1539 if (dr
->dr_parent
) {
1540 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1541 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1542 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1543 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1544 db
->db_level
+ 1 == dn
->dn_nlevels
) {
1545 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1546 mutex_enter(&dn
->dn_mtx
);
1547 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1548 mutex_exit(&dn
->dn_mtx
);
1552 if (db
->db_state
!= DB_NOFILL
) {
1553 dbuf_unoverride(dr
);
1555 ASSERT(db
->db_buf
!= NULL
);
1556 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1557 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1558 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
, db
));
1561 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1563 ASSERT(db
->db_dirtycnt
> 0);
1564 db
->db_dirtycnt
-= 1;
1566 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1567 arc_buf_t
*buf
= db
->db_buf
;
1569 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(buf
));
1570 dbuf_clear_data(db
);
1571 VERIFY(arc_buf_remove_ref(buf
, db
));
1580 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1582 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1583 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1584 dbuf_dirty_record_t
*dr
;
1586 ASSERT(tx
->tx_txg
!= 0);
1587 ASSERT(!refcount_is_zero(&db
->db_holds
));
1590 * Quick check for dirtyness. For already dirty blocks, this
1591 * reduces runtime of this function by >90%, and overall performance
1592 * by 50% for some workloads (e.g. file deletion with indirect blocks
1595 mutex_enter(&db
->db_mtx
);
1597 for (dr
= db
->db_last_dirty
;
1598 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
1600 * It's possible that it is already dirty but not cached,
1601 * because there are some calls to dbuf_dirty() that don't
1602 * go through dmu_buf_will_dirty().
1604 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
1605 /* This dbuf is already dirty and cached. */
1607 mutex_exit(&db
->db_mtx
);
1611 mutex_exit(&db
->db_mtx
);
1614 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1615 rf
|= DB_RF_HAVESTRUCT
;
1617 (void) dbuf_read(db
, NULL
, rf
);
1618 (void) dbuf_dirty(db
, tx
);
1622 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1624 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1626 db
->db_state
= DB_NOFILL
;
1628 dmu_buf_will_fill(db_fake
, tx
);
1632 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1634 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1636 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1637 ASSERT(tx
->tx_txg
!= 0);
1638 ASSERT(db
->db_level
== 0);
1639 ASSERT(!refcount_is_zero(&db
->db_holds
));
1641 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1642 dmu_tx_private_ok(tx
));
1645 (void) dbuf_dirty(db
, tx
);
1648 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1651 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1653 mutex_enter(&db
->db_mtx
);
1656 if (db
->db_state
== DB_FILL
) {
1657 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1658 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1659 /* we were freed while filling */
1660 /* XXX dbuf_undirty? */
1661 bzero(db
->db
.db_data
, db
->db
.db_size
);
1662 db
->db_freed_in_flight
= FALSE
;
1664 db
->db_state
= DB_CACHED
;
1665 cv_broadcast(&db
->db_changed
);
1667 mutex_exit(&db
->db_mtx
);
1671 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
1672 bp_embedded_type_t etype
, enum zio_compress comp
,
1673 int uncompressed_size
, int compressed_size
, int byteorder
,
1676 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
1677 struct dirty_leaf
*dl
;
1678 dmu_object_type_t type
;
1680 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
1681 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
1682 SPA_FEATURE_EMBEDDED_DATA
));
1686 type
= DB_DNODE(db
)->dn_type
;
1689 ASSERT0(db
->db_level
);
1690 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1692 dmu_buf_will_not_fill(dbuf
, tx
);
1694 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1695 dl
= &db
->db_last_dirty
->dt
.dl
;
1696 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
1697 data
, comp
, uncompressed_size
, compressed_size
);
1698 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
1699 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
1700 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
1701 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
1703 dl
->dr_override_state
= DR_OVERRIDDEN
;
1704 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
1708 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1709 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1712 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1714 ASSERT(!refcount_is_zero(&db
->db_holds
));
1715 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1716 ASSERT(db
->db_level
== 0);
1717 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1718 ASSERT(buf
!= NULL
);
1719 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1720 ASSERT(tx
->tx_txg
!= 0);
1722 arc_return_buf(buf
, db
);
1723 ASSERT(arc_released(buf
));
1725 mutex_enter(&db
->db_mtx
);
1727 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1728 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1730 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1732 if (db
->db_state
== DB_CACHED
&&
1733 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1734 mutex_exit(&db
->db_mtx
);
1735 (void) dbuf_dirty(db
, tx
);
1736 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1737 VERIFY(arc_buf_remove_ref(buf
, db
));
1738 xuio_stat_wbuf_copied();
1742 xuio_stat_wbuf_nocopy();
1743 if (db
->db_state
== DB_CACHED
) {
1744 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1746 ASSERT(db
->db_buf
!= NULL
);
1747 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
1748 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
1749 if (!arc_released(db
->db_buf
)) {
1750 ASSERT(dr
->dt
.dl
.dr_override_state
==
1752 arc_release(db
->db_buf
, db
);
1754 dr
->dt
.dl
.dr_data
= buf
;
1755 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1756 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
1757 arc_release(db
->db_buf
, db
);
1758 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1762 ASSERT(db
->db_buf
== NULL
);
1763 dbuf_set_data(db
, buf
);
1764 db
->db_state
= DB_FILL
;
1765 mutex_exit(&db
->db_mtx
);
1766 (void) dbuf_dirty(db
, tx
);
1767 dmu_buf_fill_done(&db
->db
, tx
);
1771 * "Clear" the contents of this dbuf. This will mark the dbuf
1772 * EVICTING and clear *most* of its references. Unfortunately,
1773 * when we are not holding the dn_dbufs_mtx, we can't clear the
1774 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1775 * in this case. For callers from the DMU we will usually see:
1776 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1777 * For the arc callback, we will usually see:
1778 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1779 * Sometimes, though, we will get a mix of these two:
1780 * DMU: dbuf_clear()->arc_clear_callback()
1781 * ARC: dbuf_do_evict()->dbuf_destroy()
1783 * This routine will dissociate the dbuf from the arc, by calling
1784 * arc_clear_callback(), but will not evict the data from the ARC.
1787 dbuf_clear(dmu_buf_impl_t
*db
)
1790 dmu_buf_impl_t
*parent
= db
->db_parent
;
1791 dmu_buf_impl_t
*dndb
;
1792 boolean_t dbuf_gone
= B_FALSE
;
1794 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1795 ASSERT(refcount_is_zero(&db
->db_holds
));
1797 dbuf_evict_user(db
);
1799 if (db
->db_state
== DB_CACHED
) {
1800 ASSERT(db
->db
.db_data
!= NULL
);
1801 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1802 zio_buf_free(db
->db
.db_data
, DN_MAX_BONUSLEN
);
1803 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1805 db
->db
.db_data
= NULL
;
1806 db
->db_state
= DB_UNCACHED
;
1809 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1810 ASSERT(db
->db_data_pending
== NULL
);
1812 db
->db_state
= DB_EVICTING
;
1813 db
->db_blkptr
= NULL
;
1818 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& MUTEX_HELD(&dn
->dn_dbufs_mtx
)) {
1819 avl_remove(&dn
->dn_dbufs
, db
);
1820 atomic_dec_32(&dn
->dn_dbufs_count
);
1824 * Decrementing the dbuf count means that the hold corresponding
1825 * to the removed dbuf is no longer discounted in dnode_move(),
1826 * so the dnode cannot be moved until after we release the hold.
1827 * The membar_producer() ensures visibility of the decremented
1828 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1832 db
->db_dnode_handle
= NULL
;
1838 dbuf_gone
= arc_clear_callback(db
->db_buf
);
1841 mutex_exit(&db
->db_mtx
);
1844 * If this dbuf is referenced from an indirect dbuf,
1845 * decrement the ref count on the indirect dbuf.
1847 if (parent
&& parent
!= dndb
)
1848 dbuf_rele(parent
, db
);
1852 * Note: While bpp will always be updated if the function returns success,
1853 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1854 * this happens when the dnode is the meta-dnode, or a userused or groupused
1857 __attribute__((always_inline
))
1859 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
1860 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
1867 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1869 if (blkid
== DMU_SPILL_BLKID
) {
1870 mutex_enter(&dn
->dn_mtx
);
1871 if (dn
->dn_have_spill
&&
1872 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1873 *bpp
= &dn
->dn_phys
->dn_spill
;
1876 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1877 *parentp
= dn
->dn_dbuf
;
1878 mutex_exit(&dn
->dn_mtx
);
1882 if (dn
->dn_phys
->dn_nlevels
== 0)
1885 nlevels
= dn
->dn_phys
->dn_nlevels
;
1887 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1889 ASSERT3U(level
* epbs
, <, 64);
1890 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1891 if (level
>= nlevels
||
1892 (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
1893 /* the buffer has no parent yet */
1894 return (SET_ERROR(ENOENT
));
1895 } else if (level
< nlevels
-1) {
1896 /* this block is referenced from an indirect block */
1899 err
= dbuf_hold_impl(dn
, level
+1,
1900 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
1902 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
1903 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
1904 parentp
, dh
->dh_depth
+ 1);
1905 err
= __dbuf_hold_impl(dh
+ 1);
1909 err
= dbuf_read(*parentp
, NULL
,
1910 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
1912 dbuf_rele(*parentp
, NULL
);
1916 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
1917 (blkid
& ((1ULL << epbs
) - 1));
1920 /* the block is referenced from the dnode */
1921 ASSERT3U(level
, ==, nlevels
-1);
1922 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
1923 blkid
< dn
->dn_phys
->dn_nblkptr
);
1925 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1926 *parentp
= dn
->dn_dbuf
;
1928 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
1933 static dmu_buf_impl_t
*
1934 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
1935 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
1937 objset_t
*os
= dn
->dn_objset
;
1938 dmu_buf_impl_t
*db
, *odb
;
1940 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1941 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
1943 db
= kmem_cache_alloc(dbuf_cache
, KM_SLEEP
);
1946 db
->db
.db_object
= dn
->dn_object
;
1947 db
->db_level
= level
;
1948 db
->db_blkid
= blkid
;
1949 db
->db_last_dirty
= NULL
;
1950 db
->db_dirtycnt
= 0;
1951 db
->db_dnode_handle
= dn
->dn_handle
;
1952 db
->db_parent
= parent
;
1953 db
->db_blkptr
= blkptr
;
1956 db
->db_user_immediate_evict
= FALSE
;
1957 db
->db_freed_in_flight
= FALSE
;
1958 db
->db_pending_evict
= FALSE
;
1960 if (blkid
== DMU_BONUS_BLKID
) {
1961 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
1962 db
->db
.db_size
= DN_MAX_BONUSLEN
-
1963 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
1964 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1965 db
->db
.db_offset
= DMU_BONUS_BLKID
;
1966 db
->db_state
= DB_UNCACHED
;
1967 /* the bonus dbuf is not placed in the hash table */
1968 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1970 } else if (blkid
== DMU_SPILL_BLKID
) {
1971 db
->db
.db_size
= (blkptr
!= NULL
) ?
1972 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
1973 db
->db
.db_offset
= 0;
1976 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
1977 db
->db
.db_size
= blocksize
;
1978 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
1982 * Hold the dn_dbufs_mtx while we get the new dbuf
1983 * in the hash table *and* added to the dbufs list.
1984 * This prevents a possible deadlock with someone
1985 * trying to look up this dbuf before its added to the
1988 mutex_enter(&dn
->dn_dbufs_mtx
);
1989 db
->db_state
= DB_EVICTING
;
1990 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
1991 /* someone else inserted it first */
1992 kmem_cache_free(dbuf_cache
, db
);
1993 mutex_exit(&dn
->dn_dbufs_mtx
);
1996 avl_add(&dn
->dn_dbufs
, db
);
1997 if (db
->db_level
== 0 && db
->db_blkid
>=
1998 dn
->dn_unlisted_l0_blkid
)
1999 dn
->dn_unlisted_l0_blkid
= db
->db_blkid
+ 1;
2000 db
->db_state
= DB_UNCACHED
;
2001 mutex_exit(&dn
->dn_dbufs_mtx
);
2002 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2004 if (parent
&& parent
!= dn
->dn_dbuf
)
2005 dbuf_add_ref(parent
, db
);
2007 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2008 refcount_count(&dn
->dn_holds
) > 0);
2009 (void) refcount_add(&dn
->dn_holds
, db
);
2010 atomic_inc_32(&dn
->dn_dbufs_count
);
2012 dprintf_dbuf(db
, "db=%p\n", db
);
2018 dbuf_do_evict(void *private)
2020 dmu_buf_impl_t
*db
= private;
2022 if (!MUTEX_HELD(&db
->db_mtx
))
2023 mutex_enter(&db
->db_mtx
);
2025 ASSERT(refcount_is_zero(&db
->db_holds
));
2027 if (db
->db_state
!= DB_EVICTING
) {
2028 ASSERT(db
->db_state
== DB_CACHED
);
2033 mutex_exit(&db
->db_mtx
);
2040 dbuf_destroy(dmu_buf_impl_t
*db
)
2042 ASSERT(refcount_is_zero(&db
->db_holds
));
2044 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2046 * If this dbuf is still on the dn_dbufs list,
2047 * remove it from that list.
2049 if (db
->db_dnode_handle
!= NULL
) {
2054 mutex_enter(&dn
->dn_dbufs_mtx
);
2055 avl_remove(&dn
->dn_dbufs
, db
);
2056 atomic_dec_32(&dn
->dn_dbufs_count
);
2057 mutex_exit(&dn
->dn_dbufs_mtx
);
2060 * Decrementing the dbuf count means that the hold
2061 * corresponding to the removed dbuf is no longer
2062 * discounted in dnode_move(), so the dnode cannot be
2063 * moved until after we release the hold.
2066 db
->db_dnode_handle
= NULL
;
2068 dbuf_hash_remove(db
);
2070 db
->db_parent
= NULL
;
2073 ASSERT(db
->db
.db_data
== NULL
);
2074 ASSERT(db
->db_hash_next
== NULL
);
2075 ASSERT(db
->db_blkptr
== NULL
);
2076 ASSERT(db
->db_data_pending
== NULL
);
2078 kmem_cache_free(dbuf_cache
, db
);
2079 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
2082 typedef struct dbuf_prefetch_arg
{
2083 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2084 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2085 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2086 int dpa_curlevel
; /* The current level that we're reading */
2087 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2088 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2089 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2090 } dbuf_prefetch_arg_t
;
2093 * Actually issue the prefetch read for the block given.
2096 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2099 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2102 aflags
= dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2104 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2105 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2106 ASSERT(dpa
->dpa_zio
!= NULL
);
2107 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2108 dpa
->dpa_prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2109 &aflags
, &dpa
->dpa_zb
);
2113 * Called when an indirect block above our prefetch target is read in. This
2114 * will either read in the next indirect block down the tree or issue the actual
2115 * prefetch if the next block down is our target.
2118 dbuf_prefetch_indirect_done(zio_t
*zio
, arc_buf_t
*abuf
, void *private)
2120 dbuf_prefetch_arg_t
*dpa
= private;
2124 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2125 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2127 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2128 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2129 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2132 dpa
->dpa_curlevel
--;
2134 nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2135 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2136 bp
= ((blkptr_t
*)abuf
->b_data
) +
2137 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2138 if (BP_IS_HOLE(bp
) || (zio
!= NULL
&& zio
->io_error
!= 0)) {
2139 kmem_free(dpa
, sizeof (*dpa
));
2140 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2141 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2142 dbuf_issue_final_prefetch(dpa
, bp
);
2143 kmem_free(dpa
, sizeof (*dpa
));
2145 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2146 zbookmark_phys_t zb
;
2148 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2150 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2151 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2153 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2154 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2155 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2158 (void) arc_buf_remove_ref(abuf
, private);
2162 * Issue prefetch reads for the given block on the given level. If the indirect
2163 * blocks above that block are not in memory, we will read them in
2164 * asynchronously. As a result, this call never blocks waiting for a read to
2168 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2172 int epbs
, nlevels
, curlevel
;
2176 dbuf_prefetch_arg_t
*dpa
;
2179 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2180 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2182 if (blkid
> dn
->dn_maxblkid
)
2185 if (dnode_block_freed(dn
, blkid
))
2189 * This dnode hasn't been written to disk yet, so there's nothing to
2192 nlevels
= dn
->dn_phys
->dn_nlevels
;
2193 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2196 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2197 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2200 db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2203 mutex_exit(&db
->db_mtx
);
2205 * This dbuf already exists. It is either CACHED, or
2206 * (we assume) about to be read or filled.
2212 * Find the closest ancestor (indirect block) of the target block
2213 * that is present in the cache. In this indirect block, we will
2214 * find the bp that is at curlevel, curblkid.
2218 while (curlevel
< nlevels
- 1) {
2219 int parent_level
= curlevel
+ 1;
2220 uint64_t parent_blkid
= curblkid
>> epbs
;
2223 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2224 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2225 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2226 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2227 dbuf_rele(db
, FTAG
);
2231 curlevel
= parent_level
;
2232 curblkid
= parent_blkid
;
2235 if (curlevel
== nlevels
- 1) {
2236 /* No cached indirect blocks found. */
2237 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2238 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2240 if (BP_IS_HOLE(&bp
))
2243 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2245 pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2248 dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2249 ds
= dn
->dn_objset
->os_dsl_dataset
;
2250 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2251 dn
->dn_object
, level
, blkid
);
2252 dpa
->dpa_curlevel
= curlevel
;
2253 dpa
->dpa_prio
= prio
;
2254 dpa
->dpa_aflags
= aflags
;
2255 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2256 dpa
->dpa_epbs
= epbs
;
2260 * If we have the indirect just above us, no need to do the asynchronous
2261 * prefetch chain; we'll just run the last step ourselves. If we're at
2262 * a higher level, though, we want to issue the prefetches for all the
2263 * indirect blocks asynchronously, so we can go on with whatever we were
2266 if (curlevel
== level
) {
2267 ASSERT3U(curblkid
, ==, blkid
);
2268 dbuf_issue_final_prefetch(dpa
, &bp
);
2269 kmem_free(dpa
, sizeof (*dpa
));
2271 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2272 zbookmark_phys_t zb
;
2274 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2275 dn
->dn_object
, curlevel
, curblkid
);
2276 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2277 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2278 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2282 * We use pio here instead of dpa_zio since it's possible that
2283 * dpa may have already been freed.
2288 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2291 * Returns with db_holds incremented, and db_mtx not held.
2292 * Note: dn_struct_rwlock must be held.
2295 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2297 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2298 dh
->dh_parent
= NULL
;
2300 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2301 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2302 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2304 *(dh
->dh_dbp
) = NULL
;
2306 /* dbuf_find() returns with db_mtx held */
2307 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2308 dh
->dh_level
, dh
->dh_blkid
);
2310 if (dh
->dh_db
== NULL
) {
2313 if (dh
->dh_fail_uncached
)
2314 return (SET_ERROR(ENOENT
));
2316 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2317 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2318 dh
->dh_fail_sparse
, &dh
->dh_parent
,
2320 if (dh
->dh_fail_sparse
) {
2321 if (dh
->dh_err
== 0 &&
2322 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2323 dh
->dh_err
= SET_ERROR(ENOENT
);
2326 dbuf_rele(dh
->dh_parent
, NULL
);
2327 return (dh
->dh_err
);
2330 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2331 return (dh
->dh_err
);
2332 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2333 dh
->dh_parent
, dh
->dh_bp
);
2336 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2337 mutex_exit(&dh
->dh_db
->db_mtx
);
2338 return (SET_ERROR(ENOENT
));
2341 if (dh
->dh_db
->db_buf
&& refcount_is_zero(&dh
->dh_db
->db_holds
)) {
2342 arc_buf_add_ref(dh
->dh_db
->db_buf
, dh
->dh_db
);
2343 if (dh
->dh_db
->db_buf
->b_data
== NULL
) {
2344 dbuf_clear(dh
->dh_db
);
2345 if (dh
->dh_parent
) {
2346 dbuf_rele(dh
->dh_parent
, NULL
);
2347 dh
->dh_parent
= NULL
;
2351 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2354 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2357 * If this buffer is currently syncing out, and we are are
2358 * still referencing it from db_data, we need to make a copy
2359 * of it in case we decide we want to dirty it again in this txg.
2361 if (dh
->dh_db
->db_level
== 0 &&
2362 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2363 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2364 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2365 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2367 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2368 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2370 dbuf_set_data(dh
->dh_db
,
2371 arc_buf_alloc(dh
->dh_dn
->dn_objset
->os_spa
,
2372 dh
->dh_db
->db
.db_size
, dh
->dh_db
, dh
->dh_type
));
2373 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2374 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2378 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2379 DBUF_VERIFY(dh
->dh_db
);
2380 mutex_exit(&dh
->dh_db
->db_mtx
);
2382 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2384 dbuf_rele(dh
->dh_parent
, NULL
);
2386 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2387 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2388 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2389 *(dh
->dh_dbp
) = dh
->dh_db
;
2395 * The following code preserves the recursive function dbuf_hold_impl()
2396 * but moves the local variables AND function arguments to the heap to
2397 * minimize the stack frame size. Enough space is initially allocated
2398 * on the stack for 20 levels of recursion.
2401 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2402 boolean_t fail_sparse
, boolean_t fail_uncached
,
2403 void *tag
, dmu_buf_impl_t
**dbp
)
2405 struct dbuf_hold_impl_data
*dh
;
2408 dh
= kmem_zalloc(sizeof (struct dbuf_hold_impl_data
) *
2409 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
2410 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
2411 fail_uncached
, tag
, dbp
, 0);
2413 error
= __dbuf_hold_impl(dh
);
2415 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2416 DBUF_HOLD_IMPL_MAX_DEPTH
);
2422 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2423 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2424 boolean_t fail_sparse
, boolean_t fail_uncached
,
2425 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2428 dh
->dh_level
= level
;
2429 dh
->dh_blkid
= blkid
;
2431 dh
->dh_fail_sparse
= fail_sparse
;
2432 dh
->dh_fail_uncached
= fail_uncached
;
2436 dh
->dh_depth
= depth
;
2440 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2442 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
2446 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2449 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
2450 return (err
? NULL
: db
);
2454 dbuf_create_bonus(dnode_t
*dn
)
2456 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2458 ASSERT(dn
->dn_bonus
== NULL
);
2459 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2463 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2465 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2468 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2469 return (SET_ERROR(ENOTSUP
));
2471 blksz
= SPA_MINBLOCKSIZE
;
2472 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
2473 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2477 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2478 dbuf_new_size(db
, blksz
, tx
);
2479 rw_exit(&dn
->dn_struct_rwlock
);
2486 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2488 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2491 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2493 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2495 VERIFY(refcount_add(&db
->db_holds
, tag
) > 1);
2498 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2500 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
2503 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2504 dmu_buf_impl_t
*found_db
;
2505 boolean_t result
= B_FALSE
;
2507 if (blkid
== DMU_BONUS_BLKID
)
2508 found_db
= dbuf_find_bonus(os
, obj
);
2510 found_db
= dbuf_find(os
, obj
, 0, blkid
);
2512 if (found_db
!= NULL
) {
2513 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
2514 (void) refcount_add(&db
->db_holds
, tag
);
2517 mutex_exit(&found_db
->db_mtx
);
2523 * If you call dbuf_rele() you had better not be referencing the dnode handle
2524 * unless you have some other direct or indirect hold on the dnode. (An indirect
2525 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2526 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2527 * dnode's parent dbuf evicting its dnode handles.
2530 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2532 mutex_enter(&db
->db_mtx
);
2533 dbuf_rele_and_unlock(db
, tag
);
2537 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2539 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2543 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2544 * db_dirtycnt and db_holds to be updated atomically.
2547 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2551 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2555 * Remove the reference to the dbuf before removing its hold on the
2556 * dnode so we can guarantee in dnode_move() that a referenced bonus
2557 * buffer has a corresponding dnode hold.
2559 holds
= refcount_remove(&db
->db_holds
, tag
);
2563 * We can't freeze indirects if there is a possibility that they
2564 * may be modified in the current syncing context.
2566 if (db
->db_buf
&& holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0))
2567 arc_buf_freeze(db
->db_buf
);
2569 if (holds
== db
->db_dirtycnt
&&
2570 db
->db_level
== 0 && db
->db_user_immediate_evict
)
2571 dbuf_evict_user(db
);
2574 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2576 boolean_t evict_dbuf
= db
->db_pending_evict
;
2579 * If the dnode moves here, we cannot cross this
2580 * barrier until the move completes.
2585 atomic_dec_32(&dn
->dn_dbufs_count
);
2588 * Decrementing the dbuf count means that the bonus
2589 * buffer's dnode hold is no longer discounted in
2590 * dnode_move(). The dnode cannot move until after
2591 * the dnode_rele() below.
2596 * Do not reference db after its lock is dropped.
2597 * Another thread may evict it.
2599 mutex_exit(&db
->db_mtx
);
2602 dnode_evict_bonus(dn
);
2605 } else if (db
->db_buf
== NULL
) {
2607 * This is a special case: we never associated this
2608 * dbuf with any data allocated from the ARC.
2610 ASSERT(db
->db_state
== DB_UNCACHED
||
2611 db
->db_state
== DB_NOFILL
);
2613 } else if (arc_released(db
->db_buf
)) {
2614 arc_buf_t
*buf
= db
->db_buf
;
2616 * This dbuf has anonymous data associated with it.
2618 dbuf_clear_data(db
);
2619 VERIFY(arc_buf_remove_ref(buf
, db
));
2622 VERIFY(!arc_buf_remove_ref(db
->db_buf
, db
));
2625 * A dbuf will be eligible for eviction if either the
2626 * 'primarycache' property is set or a duplicate
2627 * copy of this buffer is already cached in the arc.
2629 * In the case of the 'primarycache' a buffer
2630 * is considered for eviction if it matches the
2631 * criteria set in the property.
2633 * To decide if our buffer is considered a
2634 * duplicate, we must call into the arc to determine
2635 * if multiple buffers are referencing the same
2636 * block on-disk. If so, then we simply evict
2639 if (!DBUF_IS_CACHEABLE(db
)) {
2640 if (db
->db_blkptr
!= NULL
&&
2641 !BP_IS_HOLE(db
->db_blkptr
) &&
2642 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2644 dmu_objset_spa(db
->db_objset
);
2645 blkptr_t bp
= *db
->db_blkptr
;
2647 arc_freed(spa
, &bp
);
2651 } else if (db
->db_pending_evict
||
2652 arc_buf_eviction_needed(db
->db_buf
)) {
2655 mutex_exit(&db
->db_mtx
);
2659 mutex_exit(&db
->db_mtx
);
2663 #pragma weak dmu_buf_refcount = dbuf_refcount
2665 dbuf_refcount(dmu_buf_impl_t
*db
)
2667 return (refcount_count(&db
->db_holds
));
2671 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
2672 dmu_buf_user_t
*new_user
)
2674 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2676 mutex_enter(&db
->db_mtx
);
2677 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2678 if (db
->db_user
== old_user
)
2679 db
->db_user
= new_user
;
2681 old_user
= db
->db_user
;
2682 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2683 mutex_exit(&db
->db_mtx
);
2689 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2691 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
2695 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2697 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2699 db
->db_user_immediate_evict
= TRUE
;
2700 return (dmu_buf_set_user(db_fake
, user
));
2704 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
2706 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
2710 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2712 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2714 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
2715 return (db
->db_user
);
2719 dmu_buf_user_evict_wait()
2721 taskq_wait(dbu_evict_taskq
);
2725 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2727 boolean_t res
= B_FALSE
;
2728 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2731 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2732 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2738 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2740 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2741 return (dbi
->db_blkptr
);
2745 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2747 /* ASSERT(dmu_tx_is_syncing(tx) */
2748 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2750 if (db
->db_blkptr
!= NULL
)
2753 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2754 db
->db_blkptr
= &dn
->dn_phys
->dn_spill
;
2755 BP_ZERO(db
->db_blkptr
);
2758 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2760 * This buffer was allocated at a time when there was
2761 * no available blkptrs from the dnode, or it was
2762 * inappropriate to hook it in (i.e., nlevels mis-match).
2764 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2765 ASSERT(db
->db_parent
== NULL
);
2766 db
->db_parent
= dn
->dn_dbuf
;
2767 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2770 dmu_buf_impl_t
*parent
= db
->db_parent
;
2771 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2773 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2774 if (parent
== NULL
) {
2775 mutex_exit(&db
->db_mtx
);
2776 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2777 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
2778 db
->db_blkid
>> epbs
, db
);
2779 rw_exit(&dn
->dn_struct_rwlock
);
2780 mutex_enter(&db
->db_mtx
);
2781 db
->db_parent
= parent
;
2783 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2784 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2790 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2791 * is critical the we not allow the compiler to inline this function in to
2792 * dbuf_sync_list() thereby drastically bloating the stack usage.
2794 noinline
static void
2795 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2797 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2801 ASSERT(dmu_tx_is_syncing(tx
));
2803 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2805 mutex_enter(&db
->db_mtx
);
2807 ASSERT(db
->db_level
> 0);
2810 /* Read the block if it hasn't been read yet. */
2811 if (db
->db_buf
== NULL
) {
2812 mutex_exit(&db
->db_mtx
);
2813 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2814 mutex_enter(&db
->db_mtx
);
2816 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2817 ASSERT(db
->db_buf
!= NULL
);
2821 /* Indirect block size must match what the dnode thinks it is. */
2822 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2823 dbuf_check_blkptr(dn
, db
);
2826 /* Provide the pending dirty record to child dbufs */
2827 db
->db_data_pending
= dr
;
2829 mutex_exit(&db
->db_mtx
);
2830 dbuf_write(dr
, db
->db_buf
, tx
);
2833 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2834 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
2835 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2836 mutex_exit(&dr
->dt
.di
.dr_mtx
);
2841 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
2842 * critical the we not allow the compiler to inline this function in to
2843 * dbuf_sync_list() thereby drastically bloating the stack usage.
2845 noinline
static void
2846 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2848 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
2849 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2852 uint64_t txg
= tx
->tx_txg
;
2854 ASSERT(dmu_tx_is_syncing(tx
));
2856 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2858 mutex_enter(&db
->db_mtx
);
2860 * To be synced, we must be dirtied. But we
2861 * might have been freed after the dirty.
2863 if (db
->db_state
== DB_UNCACHED
) {
2864 /* This buffer has been freed since it was dirtied */
2865 ASSERT(db
->db
.db_data
== NULL
);
2866 } else if (db
->db_state
== DB_FILL
) {
2867 /* This buffer was freed and is now being re-filled */
2868 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
2870 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
2877 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2878 mutex_enter(&dn
->dn_mtx
);
2879 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
2880 mutex_exit(&dn
->dn_mtx
);
2884 * If this is a bonus buffer, simply copy the bonus data into the
2885 * dnode. It will be written out when the dnode is synced (and it
2886 * will be synced, since it must have been dirty for dbuf_sync to
2889 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2890 dbuf_dirty_record_t
**drp
;
2892 ASSERT(*datap
!= NULL
);
2893 ASSERT0(db
->db_level
);
2894 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
2895 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
2898 if (*datap
!= db
->db
.db_data
) {
2899 zio_buf_free(*datap
, DN_MAX_BONUSLEN
);
2900 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
2902 db
->db_data_pending
= NULL
;
2903 drp
= &db
->db_last_dirty
;
2905 drp
= &(*drp
)->dr_next
;
2906 ASSERT(dr
->dr_next
== NULL
);
2907 ASSERT(dr
->dr_dbuf
== db
);
2909 if (dr
->dr_dbuf
->db_level
!= 0) {
2910 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2911 list_destroy(&dr
->dt
.di
.dr_children
);
2913 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2914 ASSERT(db
->db_dirtycnt
> 0);
2915 db
->db_dirtycnt
-= 1;
2916 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
2923 * This function may have dropped the db_mtx lock allowing a dmu_sync
2924 * operation to sneak in. As a result, we need to ensure that we
2925 * don't check the dr_override_state until we have returned from
2926 * dbuf_check_blkptr.
2928 dbuf_check_blkptr(dn
, db
);
2931 * If this buffer is in the middle of an immediate write,
2932 * wait for the synchronous IO to complete.
2934 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
2935 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
2936 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2937 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
2940 if (db
->db_state
!= DB_NOFILL
&&
2941 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2942 refcount_count(&db
->db_holds
) > 1 &&
2943 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
2944 *datap
== db
->db_buf
) {
2946 * If this buffer is currently "in use" (i.e., there
2947 * are active holds and db_data still references it),
2948 * then make a copy before we start the write so that
2949 * any modifications from the open txg will not leak
2952 * NOTE: this copy does not need to be made for
2953 * objects only modified in the syncing context (e.g.
2954 * DNONE_DNODE blocks).
2956 int blksz
= arc_buf_size(*datap
);
2957 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2958 *datap
= arc_buf_alloc(os
->os_spa
, blksz
, db
, type
);
2959 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
2961 db
->db_data_pending
= dr
;
2963 mutex_exit(&db
->db_mtx
);
2965 dbuf_write(dr
, *datap
, tx
);
2967 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2968 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2969 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
2973 * Although zio_nowait() does not "wait for an IO", it does
2974 * initiate the IO. If this is an empty write it seems plausible
2975 * that the IO could actually be completed before the nowait
2976 * returns. We need to DB_DNODE_EXIT() first in case
2977 * zio_nowait() invalidates the dbuf.
2980 zio_nowait(dr
->dr_zio
);
2985 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
2987 dbuf_dirty_record_t
*dr
;
2989 while ((dr
= list_head(list
))) {
2990 if (dr
->dr_zio
!= NULL
) {
2992 * If we find an already initialized zio then we
2993 * are processing the meta-dnode, and we have finished.
2994 * The dbufs for all dnodes are put back on the list
2995 * during processing, so that we can zio_wait()
2996 * these IOs after initiating all child IOs.
2998 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
2999 DMU_META_DNODE_OBJECT
);
3002 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3003 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3004 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3006 list_remove(list
, dr
);
3007 if (dr
->dr_dbuf
->db_level
> 0)
3008 dbuf_sync_indirect(dr
, tx
);
3010 dbuf_sync_leaf(dr
, tx
);
3016 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3018 dmu_buf_impl_t
*db
= vdb
;
3020 blkptr_t
*bp
= zio
->io_bp
;
3021 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3022 spa_t
*spa
= zio
->io_spa
;
3027 ASSERT3P(db
->db_blkptr
, ==, bp
);
3031 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3032 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3033 zio
->io_prev_space_delta
= delta
;
3035 if (bp
->blk_birth
!= 0) {
3036 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3037 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3038 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3039 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3040 BP_IS_EMBEDDED(bp
));
3041 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3044 mutex_enter(&db
->db_mtx
);
3047 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3048 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3049 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3050 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3054 if (db
->db_level
== 0) {
3055 mutex_enter(&dn
->dn_mtx
);
3056 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3057 db
->db_blkid
!= DMU_SPILL_BLKID
)
3058 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3059 mutex_exit(&dn
->dn_mtx
);
3061 if (dn
->dn_type
== DMU_OT_DNODE
) {
3062 dnode_phys_t
*dnp
= db
->db
.db_data
;
3063 for (i
= db
->db
.db_size
>> DNODE_SHIFT
; i
> 0;
3065 if (dnp
->dn_type
!= DMU_OT_NONE
)
3069 if (BP_IS_HOLE(bp
)) {
3076 blkptr_t
*ibp
= db
->db
.db_data
;
3077 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3078 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3079 if (BP_IS_HOLE(ibp
))
3081 fill
+= BP_GET_FILL(ibp
);
3086 if (!BP_IS_EMBEDDED(bp
))
3087 bp
->blk_fill
= fill
;
3089 mutex_exit(&db
->db_mtx
);
3093 * The SPA will call this callback several times for each zio - once
3094 * for every physical child i/o (zio->io_phys_children times). This
3095 * allows the DMU to monitor the progress of each logical i/o. For example,
3096 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3097 * block. There may be a long delay before all copies/fragments are completed,
3098 * so this callback allows us to retire dirty space gradually, as the physical
3103 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3105 dmu_buf_impl_t
*db
= arg
;
3106 objset_t
*os
= db
->db_objset
;
3107 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3108 dbuf_dirty_record_t
*dr
;
3111 dr
= db
->db_data_pending
;
3112 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3115 * The callback will be called io_phys_children times. Retire one
3116 * portion of our dirty space each time we are called. Any rounding
3117 * error will be cleaned up by dsl_pool_sync()'s call to
3118 * dsl_pool_undirty_space().
3120 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3121 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3126 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3128 dmu_buf_impl_t
*db
= vdb
;
3129 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3130 blkptr_t
*bp
= db
->db_blkptr
;
3131 objset_t
*os
= db
->db_objset
;
3132 dmu_tx_t
*tx
= os
->os_synctx
;
3133 dbuf_dirty_record_t
**drp
, *dr
;
3135 ASSERT0(zio
->io_error
);
3136 ASSERT(db
->db_blkptr
== bp
);
3139 * For nopwrites and rewrites we ensure that the bp matches our
3140 * original and bypass all the accounting.
3142 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3143 ASSERT(BP_EQUAL(bp
, bp_orig
));
3145 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3146 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3147 dsl_dataset_block_born(ds
, bp
, tx
);
3150 mutex_enter(&db
->db_mtx
);
3154 drp
= &db
->db_last_dirty
;
3155 while ((dr
= *drp
) != db
->db_data_pending
)
3157 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3158 ASSERT(dr
->dr_dbuf
== db
);
3159 ASSERT(dr
->dr_next
== NULL
);
3163 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3168 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3169 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3170 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
3175 if (db
->db_level
== 0) {
3176 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3177 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3178 if (db
->db_state
!= DB_NOFILL
) {
3179 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3180 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
3182 else if (!arc_released(db
->db_buf
))
3183 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3190 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3191 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3192 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3193 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3195 ASSERT3U(db
->db_blkid
, <=,
3196 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3197 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
3199 if (!arc_released(db
->db_buf
))
3200 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
3203 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3204 list_destroy(&dr
->dt
.di
.dr_children
);
3206 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3208 cv_broadcast(&db
->db_changed
);
3209 ASSERT(db
->db_dirtycnt
> 0);
3210 db
->db_dirtycnt
-= 1;
3211 db
->db_data_pending
= NULL
;
3212 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
3216 dbuf_write_nofill_ready(zio_t
*zio
)
3218 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
3222 dbuf_write_nofill_done(zio_t
*zio
)
3224 dbuf_write_done(zio
, NULL
, zio
->io_private
);
3228 dbuf_write_override_ready(zio_t
*zio
)
3230 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3231 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3233 dbuf_write_ready(zio
, NULL
, db
);
3237 dbuf_write_override_done(zio_t
*zio
)
3239 dbuf_dirty_record_t
*dr
= zio
->io_private
;
3240 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3241 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
3243 mutex_enter(&db
->db_mtx
);
3244 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
3245 if (!BP_IS_HOLE(obp
))
3246 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
3247 arc_release(dr
->dt
.dl
.dr_data
, db
);
3249 mutex_exit(&db
->db_mtx
);
3251 dbuf_write_done(zio
, NULL
, db
);
3254 /* Issue I/O to commit a dirty buffer to disk. */
3256 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
3258 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3261 dmu_buf_impl_t
*parent
= db
->db_parent
;
3262 uint64_t txg
= tx
->tx_txg
;
3263 zbookmark_phys_t zb
;
3272 if (db
->db_state
!= DB_NOFILL
) {
3273 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
3275 * Private object buffers are released here rather
3276 * than in dbuf_dirty() since they are only modified
3277 * in the syncing context and we don't want the
3278 * overhead of making multiple copies of the data.
3280 if (BP_IS_HOLE(db
->db_blkptr
)) {
3283 dbuf_release_bp(db
);
3288 if (parent
!= dn
->dn_dbuf
) {
3289 /* Our parent is an indirect block. */
3290 /* We have a dirty parent that has been scheduled for write. */
3291 ASSERT(parent
&& parent
->db_data_pending
);
3292 /* Our parent's buffer is one level closer to the dnode. */
3293 ASSERT(db
->db_level
== parent
->db_level
-1);
3295 * We're about to modify our parent's db_data by modifying
3296 * our block pointer, so the parent must be released.
3298 ASSERT(arc_released(parent
->db_buf
));
3299 zio
= parent
->db_data_pending
->dr_zio
;
3301 /* Our parent is the dnode itself. */
3302 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
3303 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
3304 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
3305 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3306 ASSERT3P(db
->db_blkptr
, ==,
3307 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
3311 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
3312 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
3315 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
3316 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
3317 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3319 if (db
->db_blkid
== DMU_SPILL_BLKID
)
3321 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
3323 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
3326 if (db
->db_level
== 0 &&
3327 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
3329 * The BP for this block has been provided by open context
3330 * (by dmu_sync() or dmu_buf_write_embedded()).
3332 void *contents
= (data
!= NULL
) ? data
->b_data
: NULL
;
3334 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3335 db
->db_blkptr
, contents
, db
->db
.db_size
, &zp
,
3336 dbuf_write_override_ready
, NULL
, dbuf_write_override_done
,
3337 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3338 mutex_enter(&db
->db_mtx
);
3339 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
3340 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
3341 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
3342 mutex_exit(&db
->db_mtx
);
3343 } else if (db
->db_state
== DB_NOFILL
) {
3344 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
);
3345 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
3346 db
->db_blkptr
, NULL
, db
->db
.db_size
, &zp
,
3347 dbuf_write_nofill_ready
, NULL
, dbuf_write_nofill_done
, db
,
3348 ZIO_PRIORITY_ASYNC_WRITE
,
3349 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
3351 ASSERT(arc_released(data
));
3352 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
3353 db
->db_blkptr
, data
, DBUF_IS_L2CACHEABLE(db
),
3354 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dbuf_write_ready
,
3355 dbuf_write_physdone
, dbuf_write_done
, db
,
3356 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
3360 #if defined(_KERNEL) && defined(HAVE_SPL)
3361 EXPORT_SYMBOL(dbuf_find
);
3362 EXPORT_SYMBOL(dbuf_is_metadata
);
3363 EXPORT_SYMBOL(dbuf_evict
);
3364 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
3365 EXPORT_SYMBOL(dbuf_whichblock
);
3366 EXPORT_SYMBOL(dbuf_read
);
3367 EXPORT_SYMBOL(dbuf_unoverride
);
3368 EXPORT_SYMBOL(dbuf_free_range
);
3369 EXPORT_SYMBOL(dbuf_new_size
);
3370 EXPORT_SYMBOL(dbuf_release_bp
);
3371 EXPORT_SYMBOL(dbuf_dirty
);
3372 EXPORT_SYMBOL(dmu_buf_will_dirty
);
3373 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
3374 EXPORT_SYMBOL(dmu_buf_will_fill
);
3375 EXPORT_SYMBOL(dmu_buf_fill_done
);
3376 EXPORT_SYMBOL(dmu_buf_rele
);
3377 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
3378 EXPORT_SYMBOL(dbuf_clear
);
3379 EXPORT_SYMBOL(dbuf_prefetch
);
3380 EXPORT_SYMBOL(dbuf_hold_impl
);
3381 EXPORT_SYMBOL(dbuf_hold
);
3382 EXPORT_SYMBOL(dbuf_hold_level
);
3383 EXPORT_SYMBOL(dbuf_create_bonus
);
3384 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
3385 EXPORT_SYMBOL(dbuf_rm_spill
);
3386 EXPORT_SYMBOL(dbuf_add_ref
);
3387 EXPORT_SYMBOL(dbuf_rele
);
3388 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
3389 EXPORT_SYMBOL(dbuf_refcount
);
3390 EXPORT_SYMBOL(dbuf_sync_list
);
3391 EXPORT_SYMBOL(dmu_buf_set_user
);
3392 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3393 EXPORT_SYMBOL(dmu_buf_get_user
);
3394 EXPORT_SYMBOL(dmu_buf_freeable
);
3395 EXPORT_SYMBOL(dmu_buf_get_blkptr
);