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, 2014 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
28 #include <sys/zfs_context.h>
31 #include <sys/dmu_send.h>
32 #include <sys/dmu_impl.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dmu_tx.h>
40 #include <sys/dmu_zfetch.h>
42 #include <sys/sa_impl.h>
43 #include <sys/zfeature.h>
44 #include <sys/blkptr.h>
45 #include <sys/range_tree.h>
46 #include <sys/trace_dbuf.h>
48 struct dbuf_hold_impl_data
{
49 /* Function arguments */
55 dmu_buf_impl_t
**dh_dbp
;
57 dmu_buf_impl_t
*dh_db
;
58 dmu_buf_impl_t
*dh_parent
;
61 dbuf_dirty_record_t
*dh_dr
;
62 arc_buf_contents_t dh_type
;
66 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
67 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, int fail_sparse
,
68 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
69 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
72 * Number of times that zfs_free_range() took the slow path while doing
73 * a zfs receive. A nonzero value indicates a potential performance problem.
75 uint64_t zfs_free_range_recv_miss
;
77 static void dbuf_destroy(dmu_buf_impl_t
*db
);
78 static boolean_t
dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
79 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
82 * Global data structures and functions for the dbuf cache.
84 static kmem_cache_t
*dbuf_cache
;
88 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
90 dmu_buf_impl_t
*db
= vdb
;
91 bzero(db
, sizeof (dmu_buf_impl_t
));
93 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
94 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
95 refcount_create(&db
->db_holds
);
96 list_link_init(&db
->db_link
);
102 dbuf_dest(void *vdb
, void *unused
)
104 dmu_buf_impl_t
*db
= vdb
;
105 mutex_destroy(&db
->db_mtx
);
106 cv_destroy(&db
->db_changed
);
107 refcount_destroy(&db
->db_holds
);
111 * dbuf hash table routines
113 static dbuf_hash_table_t dbuf_hash_table
;
115 static uint64_t dbuf_hash_count
;
118 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
120 uintptr_t osv
= (uintptr_t)os
;
121 uint64_t crc
= -1ULL;
123 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
124 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
125 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
126 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
127 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
128 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
129 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
131 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
136 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
138 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
139 ((dbuf)->db.db_object == (obj) && \
140 (dbuf)->db_objset == (os) && \
141 (dbuf)->db_level == (level) && \
142 (dbuf)->db_blkid == (blkid))
145 dbuf_find(dnode_t
*dn
, uint8_t level
, uint64_t blkid
)
147 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
148 objset_t
*os
= dn
->dn_objset
;
155 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
156 idx
= hv
& h
->hash_table_mask
;
158 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
159 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
160 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
161 mutex_enter(&db
->db_mtx
);
162 if (db
->db_state
!= DB_EVICTING
) {
163 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
166 mutex_exit(&db
->db_mtx
);
169 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
174 * Insert an entry into the hash table. If there is already an element
175 * equal to elem in the hash table, then the already existing element
176 * will be returned and the new element will not be inserted.
177 * Otherwise returns NULL.
179 static dmu_buf_impl_t
*
180 dbuf_hash_insert(dmu_buf_impl_t
*db
)
182 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
183 objset_t
*os
= db
->db_objset
;
184 uint64_t obj
= db
->db
.db_object
;
185 int level
= db
->db_level
;
186 uint64_t blkid
, hv
, idx
;
189 blkid
= db
->db_blkid
;
190 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
191 idx
= hv
& h
->hash_table_mask
;
193 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
194 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
195 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
196 mutex_enter(&dbf
->db_mtx
);
197 if (dbf
->db_state
!= DB_EVICTING
) {
198 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
201 mutex_exit(&dbf
->db_mtx
);
205 mutex_enter(&db
->db_mtx
);
206 db
->db_hash_next
= h
->hash_table
[idx
];
207 h
->hash_table
[idx
] = db
;
208 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
209 atomic_add_64(&dbuf_hash_count
, 1);
215 * Remove an entry from the hash table. It must be in the EVICTING state.
218 dbuf_hash_remove(dmu_buf_impl_t
*db
)
220 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
222 dmu_buf_impl_t
*dbf
, **dbp
;
224 hv
= DBUF_HASH(db
->db_objset
, db
->db
.db_object
,
225 db
->db_level
, db
->db_blkid
);
226 idx
= hv
& h
->hash_table_mask
;
229 * We musn't hold db_mtx to maintain lock ordering:
230 * DBUF_HASH_MUTEX > db_mtx.
232 ASSERT(refcount_is_zero(&db
->db_holds
));
233 ASSERT(db
->db_state
== DB_EVICTING
);
234 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
236 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
237 dbp
= &h
->hash_table
[idx
];
238 while ((dbf
= *dbp
) != db
) {
239 dbp
= &dbf
->db_hash_next
;
242 *dbp
= db
->db_hash_next
;
243 db
->db_hash_next
= NULL
;
244 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
245 atomic_add_64(&dbuf_hash_count
, -1);
248 static arc_evict_func_t dbuf_do_evict
;
251 dbuf_evict_user(dmu_buf_impl_t
*db
)
253 ASSERT(MUTEX_HELD(&db
->db_mtx
));
255 if (db
->db_level
!= 0 || db
->db_evict_func
== NULL
)
258 if (db
->db_user_data_ptr_ptr
)
259 *db
->db_user_data_ptr_ptr
= db
->db
.db_data
;
260 db
->db_evict_func(&db
->db
, db
->db_user_ptr
);
261 db
->db_user_ptr
= NULL
;
262 db
->db_user_data_ptr_ptr
= NULL
;
263 db
->db_evict_func
= NULL
;
267 dbuf_is_metadata(dmu_buf_impl_t
*db
)
270 * Consider indirect blocks and spill blocks to be meta data.
272 if (db
->db_level
> 0 || db
->db_blkid
== DMU_SPILL_BLKID
) {
275 boolean_t is_metadata
;
278 is_metadata
= DMU_OT_IS_METADATA(DB_DNODE(db
)->dn_type
);
281 return (is_metadata
);
286 dbuf_evict(dmu_buf_impl_t
*db
)
288 ASSERT(MUTEX_HELD(&db
->db_mtx
));
289 ASSERT(db
->db_buf
== NULL
);
290 ASSERT(db
->db_data_pending
== NULL
);
299 uint64_t hsize
= 1ULL << 16;
300 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
304 * The hash table is big enough to fill all of physical memory
305 * with an average 4K block size. The table will take up
306 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
308 while (hsize
* 4096 < physmem
* PAGESIZE
)
312 h
->hash_table_mask
= hsize
- 1;
313 #if defined(_KERNEL) && defined(HAVE_SPL)
315 * Large allocations which do not require contiguous pages
316 * should be using vmem_alloc() in the linux kernel
318 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_PUSHPAGE
);
320 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
322 if (h
->hash_table
== NULL
) {
323 /* XXX - we should really return an error instead of assert */
324 ASSERT(hsize
> (1ULL << 10));
329 dbuf_cache
= kmem_cache_create("dmu_buf_impl_t",
330 sizeof (dmu_buf_impl_t
),
331 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
333 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
334 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
342 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
345 dbuf_stats_destroy();
347 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
348 mutex_destroy(&h
->hash_mutexes
[i
]);
349 #if defined(_KERNEL) && defined(HAVE_SPL)
351 * Large allocations which do not require contiguous pages
352 * should be using vmem_free() in the linux kernel
354 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
356 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
358 kmem_cache_destroy(dbuf_cache
);
367 dbuf_verify(dmu_buf_impl_t
*db
)
370 dbuf_dirty_record_t
*dr
;
372 ASSERT(MUTEX_HELD(&db
->db_mtx
));
374 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
377 ASSERT(db
->db_objset
!= NULL
);
381 ASSERT(db
->db_parent
== NULL
);
382 ASSERT(db
->db_blkptr
== NULL
);
384 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
385 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
386 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
387 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
388 db
->db_blkid
== DMU_SPILL_BLKID
||
389 !list_is_empty(&dn
->dn_dbufs
));
391 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
393 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
394 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
395 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
397 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
398 ASSERT0(db
->db
.db_offset
);
400 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
403 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
404 ASSERT(dr
->dr_dbuf
== db
);
406 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
407 ASSERT(dr
->dr_dbuf
== db
);
410 * We can't assert that db_size matches dn_datablksz because it
411 * can be momentarily different when another thread is doing
414 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
415 dr
= db
->db_data_pending
;
417 * It should only be modified in syncing context, so
418 * make sure we only have one copy of the data.
420 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
423 /* verify db->db_blkptr */
425 if (db
->db_parent
== dn
->dn_dbuf
) {
426 /* db is pointed to by the dnode */
427 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
428 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
429 ASSERT(db
->db_parent
== NULL
);
431 ASSERT(db
->db_parent
!= NULL
);
432 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
433 ASSERT3P(db
->db_blkptr
, ==,
434 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
436 /* db is pointed to by an indirect block */
437 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
439 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
440 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
443 * dnode_grow_indblksz() can make this fail if we don't
444 * have the struct_rwlock. XXX indblksz no longer
445 * grows. safe to do this now?
447 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
448 ASSERT3P(db
->db_blkptr
, ==,
449 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
450 db
->db_blkid
% epb
));
454 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
455 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
456 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
457 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
459 * If the blkptr isn't set but they have nonzero data,
460 * it had better be dirty, otherwise we'll lose that
461 * data when we evict this buffer.
463 if (db
->db_dirtycnt
== 0) {
464 ASSERTV(uint64_t *buf
= db
->db
.db_data
);
467 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
477 dbuf_update_data(dmu_buf_impl_t
*db
)
479 ASSERT(MUTEX_HELD(&db
->db_mtx
));
480 if (db
->db_level
== 0 && db
->db_user_data_ptr_ptr
) {
481 ASSERT(!refcount_is_zero(&db
->db_holds
));
482 *db
->db_user_data_ptr_ptr
= db
->db
.db_data
;
487 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
489 ASSERT(MUTEX_HELD(&db
->db_mtx
));
492 ASSERT(buf
->b_data
!= NULL
);
493 db
->db
.db_data
= buf
->b_data
;
494 if (!arc_released(buf
))
495 arc_set_callback(buf
, dbuf_do_evict
, db
);
496 dbuf_update_data(db
);
499 db
->db
.db_data
= NULL
;
500 if (db
->db_state
!= DB_NOFILL
)
501 db
->db_state
= DB_UNCACHED
;
506 * Loan out an arc_buf for read. Return the loaned arc_buf.
509 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
513 mutex_enter(&db
->db_mtx
);
514 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
515 int blksz
= db
->db
.db_size
;
516 spa_t
*spa
= db
->db_objset
->os_spa
;
518 mutex_exit(&db
->db_mtx
);
519 abuf
= arc_loan_buf(spa
, blksz
);
520 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
523 arc_loan_inuse_buf(abuf
, db
);
524 dbuf_set_data(db
, NULL
);
525 mutex_exit(&db
->db_mtx
);
531 dbuf_whichblock(dnode_t
*dn
, uint64_t offset
)
533 if (dn
->dn_datablkshift
) {
534 return (offset
>> dn
->dn_datablkshift
);
536 ASSERT3U(offset
, <, dn
->dn_datablksz
);
542 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
544 dmu_buf_impl_t
*db
= vdb
;
546 mutex_enter(&db
->db_mtx
);
547 ASSERT3U(db
->db_state
, ==, DB_READ
);
549 * All reads are synchronous, so we must have a hold on the dbuf
551 ASSERT(refcount_count(&db
->db_holds
) > 0);
552 ASSERT(db
->db_buf
== NULL
);
553 ASSERT(db
->db
.db_data
== NULL
);
554 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
555 /* we were freed in flight; disregard any error */
556 arc_release(buf
, db
);
557 bzero(buf
->b_data
, db
->db
.db_size
);
559 db
->db_freed_in_flight
= FALSE
;
560 dbuf_set_data(db
, buf
);
561 db
->db_state
= DB_CACHED
;
562 } else if (zio
== NULL
|| zio
->io_error
== 0) {
563 dbuf_set_data(db
, buf
);
564 db
->db_state
= DB_CACHED
;
566 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
567 ASSERT3P(db
->db_buf
, ==, NULL
);
568 VERIFY(arc_buf_remove_ref(buf
, db
));
569 db
->db_state
= DB_UNCACHED
;
571 cv_broadcast(&db
->db_changed
);
572 dbuf_rele_and_unlock(db
, NULL
);
576 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t *flags
)
580 uint32_t aflags
= ARC_NOWAIT
;
585 ASSERT(!refcount_is_zero(&db
->db_holds
));
586 /* We need the struct_rwlock to prevent db_blkptr from changing. */
587 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
588 ASSERT(MUTEX_HELD(&db
->db_mtx
));
589 ASSERT(db
->db_state
== DB_UNCACHED
);
590 ASSERT(db
->db_buf
== NULL
);
592 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
593 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
595 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
596 db
->db
.db_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
597 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
598 if (bonuslen
< DN_MAX_BONUSLEN
)
599 bzero(db
->db
.db_data
, DN_MAX_BONUSLEN
);
601 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
603 dbuf_update_data(db
);
604 db
->db_state
= DB_CACHED
;
605 mutex_exit(&db
->db_mtx
);
610 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
611 * processes the delete record and clears the bp while we are waiting
612 * for the dn_mtx (resulting in a "no" from block_freed).
614 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
615 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
616 BP_IS_HOLE(db
->db_blkptr
)))) {
617 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
620 dbuf_set_data(db
, arc_buf_alloc(db
->db_objset
->os_spa
,
621 db
->db
.db_size
, db
, type
));
622 bzero(db
->db
.db_data
, db
->db
.db_size
);
623 db
->db_state
= DB_CACHED
;
624 *flags
|= DB_RF_CACHED
;
625 mutex_exit(&db
->db_mtx
);
631 db
->db_state
= DB_READ
;
632 mutex_exit(&db
->db_mtx
);
634 if (DBUF_IS_L2CACHEABLE(db
))
635 aflags
|= ARC_L2CACHE
;
636 if (DBUF_IS_L2COMPRESSIBLE(db
))
637 aflags
|= ARC_L2COMPRESS
;
639 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
640 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
641 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
643 dbuf_add_ref(db
, NULL
);
645 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
646 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
647 (*flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
649 if (aflags
& ARC_CACHED
)
650 *flags
|= DB_RF_CACHED
;
652 return (SET_ERROR(err
));
656 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
659 boolean_t havepzio
= (zio
!= NULL
);
664 * We don't have to hold the mutex to check db_state because it
665 * can't be freed while we have a hold on the buffer.
667 ASSERT(!refcount_is_zero(&db
->db_holds
));
669 if (db
->db_state
== DB_NOFILL
)
670 return (SET_ERROR(EIO
));
674 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
675 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
677 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
678 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
679 DBUF_IS_CACHEABLE(db
);
681 mutex_enter(&db
->db_mtx
);
682 if (db
->db_state
== DB_CACHED
) {
683 mutex_exit(&db
->db_mtx
);
685 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
686 db
->db
.db_size
, TRUE
);
687 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
688 rw_exit(&dn
->dn_struct_rwlock
);
690 } else if (db
->db_state
== DB_UNCACHED
) {
691 spa_t
*spa
= dn
->dn_objset
->os_spa
;
694 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
696 err
= dbuf_read_impl(db
, zio
, &flags
);
698 /* dbuf_read_impl has dropped db_mtx for us */
700 if (!err
&& prefetch
)
701 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
702 db
->db
.db_size
, flags
& DB_RF_CACHED
);
704 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
705 rw_exit(&dn
->dn_struct_rwlock
);
708 if (!err
&& !havepzio
)
712 * Another reader came in while the dbuf was in flight
713 * between UNCACHED and CACHED. Either a writer will finish
714 * writing the buffer (sending the dbuf to CACHED) or the
715 * first reader's request will reach the read_done callback
716 * and send the dbuf to CACHED. Otherwise, a failure
717 * occurred and the dbuf went to UNCACHED.
719 mutex_exit(&db
->db_mtx
);
721 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
722 db
->db
.db_size
, TRUE
);
723 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
724 rw_exit(&dn
->dn_struct_rwlock
);
727 /* Skip the wait per the caller's request. */
728 mutex_enter(&db
->db_mtx
);
729 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
730 while (db
->db_state
== DB_READ
||
731 db
->db_state
== DB_FILL
) {
732 ASSERT(db
->db_state
== DB_READ
||
733 (flags
& DB_RF_HAVESTRUCT
) == 0);
734 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
736 cv_wait(&db
->db_changed
, &db
->db_mtx
);
738 if (db
->db_state
== DB_UNCACHED
)
739 err
= SET_ERROR(EIO
);
741 mutex_exit(&db
->db_mtx
);
744 ASSERT(err
|| havepzio
|| db
->db_state
== DB_CACHED
);
749 dbuf_noread(dmu_buf_impl_t
*db
)
751 ASSERT(!refcount_is_zero(&db
->db_holds
));
752 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
753 mutex_enter(&db
->db_mtx
);
754 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
755 cv_wait(&db
->db_changed
, &db
->db_mtx
);
756 if (db
->db_state
== DB_UNCACHED
) {
757 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
758 spa_t
*spa
= db
->db_objset
->os_spa
;
760 ASSERT(db
->db_buf
== NULL
);
761 ASSERT(db
->db
.db_data
== NULL
);
762 dbuf_set_data(db
, arc_buf_alloc(spa
, db
->db
.db_size
, db
, type
));
763 db
->db_state
= DB_FILL
;
764 } else if (db
->db_state
== DB_NOFILL
) {
765 dbuf_set_data(db
, NULL
);
767 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
769 mutex_exit(&db
->db_mtx
);
773 * This is our just-in-time copy function. It makes a copy of
774 * buffers, that have been modified in a previous transaction
775 * group, before we modify them in the current active group.
777 * This function is used in two places: when we are dirtying a
778 * buffer for the first time in a txg, and when we are freeing
779 * a range in a dnode that includes this buffer.
781 * Note that when we are called from dbuf_free_range() we do
782 * not put a hold on the buffer, we just traverse the active
783 * dbuf list for the dnode.
786 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
788 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
790 ASSERT(MUTEX_HELD(&db
->db_mtx
));
791 ASSERT(db
->db
.db_data
!= NULL
);
792 ASSERT(db
->db_level
== 0);
793 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
796 (dr
->dt
.dl
.dr_data
!=
797 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
801 * If the last dirty record for this dbuf has not yet synced
802 * and its referencing the dbuf data, either:
803 * reset the reference to point to a new copy,
804 * or (if there a no active holders)
805 * just null out the current db_data pointer.
807 ASSERT(dr
->dr_txg
>= txg
- 2);
808 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
809 /* Note that the data bufs here are zio_bufs */
810 dr
->dt
.dl
.dr_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
811 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
812 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, DN_MAX_BONUSLEN
);
813 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
814 int size
= db
->db
.db_size
;
815 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
816 spa_t
*spa
= db
->db_objset
->os_spa
;
818 dr
->dt
.dl
.dr_data
= arc_buf_alloc(spa
, size
, db
, type
);
819 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
821 dbuf_set_data(db
, NULL
);
826 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
828 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
829 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
830 uint64_t txg
= dr
->dr_txg
;
832 ASSERT(MUTEX_HELD(&db
->db_mtx
));
833 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
834 ASSERT(db
->db_level
== 0);
836 if (db
->db_blkid
== DMU_BONUS_BLKID
||
837 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
840 ASSERT(db
->db_data_pending
!= dr
);
842 /* free this block */
843 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
844 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
846 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
847 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
850 * Release the already-written buffer, so we leave it in
851 * a consistent dirty state. Note that all callers are
852 * modifying the buffer, so they will immediately do
853 * another (redundant) arc_release(). Therefore, leave
854 * the buf thawed to save the effort of freezing &
855 * immediately re-thawing it.
857 arc_release(dr
->dt
.dl
.dr_data
, db
);
861 * Evict (if its unreferenced) or clear (if its referenced) any level-0
862 * data blocks in the free range, so that any future readers will find
865 * This is a no-op if the dataset is in the middle of an incremental
866 * receive; see comment below for details.
869 dbuf_free_range(dnode_t
*dn
, uint64_t start
, uint64_t end
, dmu_tx_t
*tx
)
871 dmu_buf_impl_t
*db
, *db_next
;
872 uint64_t txg
= tx
->tx_txg
;
873 boolean_t freespill
=
874 (start
== DMU_SPILL_BLKID
|| end
== DMU_SPILL_BLKID
);
876 if (end
> dn
->dn_maxblkid
&& !freespill
)
877 end
= dn
->dn_maxblkid
;
878 dprintf_dnode(dn
, "start=%llu end=%llu\n", start
, end
);
880 mutex_enter(&dn
->dn_dbufs_mtx
);
881 if (start
>= dn
->dn_unlisted_l0_blkid
* dn
->dn_datablksz
&&
883 /* There can't be any dbufs in this range; no need to search. */
884 mutex_exit(&dn
->dn_dbufs_mtx
);
886 } else if (dmu_objset_is_receiving(dn
->dn_objset
)) {
888 * If we are receiving, we expect there to be no dbufs in
889 * the range to be freed, because receive modifies each
890 * block at most once, and in offset order. If this is
891 * not the case, it can lead to performance problems,
892 * so note that we unexpectedly took the slow path.
894 atomic_inc_64(&zfs_free_range_recv_miss
);
897 for (db
= list_head(&dn
->dn_dbufs
); db
!= NULL
; db
= db_next
) {
898 db_next
= list_next(&dn
->dn_dbufs
, db
);
899 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
901 if (db
->db_level
!= 0)
903 if ((db
->db_blkid
< start
|| db
->db_blkid
> end
) && !freespill
)
906 /* found a level 0 buffer in the range */
907 mutex_enter(&db
->db_mtx
);
908 if (dbuf_undirty(db
, tx
)) {
909 /* mutex has been dropped and dbuf destroyed */
913 if (db
->db_state
== DB_UNCACHED
||
914 db
->db_state
== DB_NOFILL
||
915 db
->db_state
== DB_EVICTING
) {
916 ASSERT(db
->db
.db_data
== NULL
);
917 mutex_exit(&db
->db_mtx
);
920 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
921 /* will be handled in dbuf_read_done or dbuf_rele */
922 db
->db_freed_in_flight
= TRUE
;
923 mutex_exit(&db
->db_mtx
);
926 if (refcount_count(&db
->db_holds
) == 0) {
931 /* The dbuf is referenced */
933 if (db
->db_last_dirty
!= NULL
) {
934 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
936 if (dr
->dr_txg
== txg
) {
938 * This buffer is "in-use", re-adjust the file
939 * size to reflect that this buffer may
940 * contain new data when we sync.
942 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
943 db
->db_blkid
> dn
->dn_maxblkid
)
944 dn
->dn_maxblkid
= db
->db_blkid
;
948 * This dbuf is not dirty in the open context.
949 * Either uncache it (if its not referenced in
950 * the open context) or reset its contents to
953 dbuf_fix_old_data(db
, txg
);
956 /* clear the contents if its cached */
957 if (db
->db_state
== DB_CACHED
) {
958 ASSERT(db
->db
.db_data
!= NULL
);
959 arc_release(db
->db_buf
, db
);
960 bzero(db
->db
.db_data
, db
->db
.db_size
);
961 arc_buf_freeze(db
->db_buf
);
964 mutex_exit(&db
->db_mtx
);
966 mutex_exit(&dn
->dn_dbufs_mtx
);
970 dbuf_block_freeable(dmu_buf_impl_t
*db
)
972 dsl_dataset_t
*ds
= db
->db_objset
->os_dsl_dataset
;
973 uint64_t birth_txg
= 0;
976 * We don't need any locking to protect db_blkptr:
977 * If it's syncing, then db_last_dirty will be set
978 * so we'll ignore db_blkptr.
980 * This logic ensures that only block births for
981 * filled blocks are considered.
983 ASSERT(MUTEX_HELD(&db
->db_mtx
));
984 if (db
->db_last_dirty
&& (db
->db_blkptr
== NULL
||
985 !BP_IS_HOLE(db
->db_blkptr
))) {
986 birth_txg
= db
->db_last_dirty
->dr_txg
;
987 } else if (db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
988 birth_txg
= db
->db_blkptr
->blk_birth
;
992 * If this block don't exist or is in a snapshot, it can't be freed.
993 * Don't pass the bp to dsl_dataset_block_freeable() since we
994 * are holding the db_mtx lock and might deadlock if we are
995 * prefetching a dedup-ed block.
998 return (ds
== NULL
||
999 dsl_dataset_block_freeable(ds
, NULL
, birth_txg
));
1005 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1007 arc_buf_t
*buf
, *obuf
;
1008 int osize
= db
->db
.db_size
;
1009 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1012 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1017 /* XXX does *this* func really need the lock? */
1018 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1021 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1022 * is OK, because there can be no other references to the db
1023 * when we are changing its size, so no concurrent DB_FILL can
1027 * XXX we should be doing a dbuf_read, checking the return
1028 * value and returning that up to our callers
1030 dmu_buf_will_dirty(&db
->db
, tx
);
1032 /* create the data buffer for the new block */
1033 buf
= arc_buf_alloc(dn
->dn_objset
->os_spa
, size
, db
, type
);
1035 /* copy old block data to the new block */
1037 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1038 /* zero the remainder */
1040 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1042 mutex_enter(&db
->db_mtx
);
1043 dbuf_set_data(db
, buf
);
1044 VERIFY(arc_buf_remove_ref(obuf
, db
));
1045 db
->db
.db_size
= size
;
1047 if (db
->db_level
== 0) {
1048 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1049 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1051 mutex_exit(&db
->db_mtx
);
1053 dnode_willuse_space(dn
, size
-osize
, tx
);
1058 dbuf_release_bp(dmu_buf_impl_t
*db
)
1060 ASSERTV(objset_t
*os
= db
->db_objset
);
1062 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1063 ASSERT(arc_released(os
->os_phys_buf
) ||
1064 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1065 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1067 (void) arc_release(db
->db_buf
, db
);
1070 dbuf_dirty_record_t
*
1071 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1075 dbuf_dirty_record_t
**drp
, *dr
;
1076 int drop_struct_lock
= FALSE
;
1077 boolean_t do_free_accounting
= B_FALSE
;
1078 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1080 ASSERT(tx
->tx_txg
!= 0);
1081 ASSERT(!refcount_is_zero(&db
->db_holds
));
1082 DMU_TX_DIRTY_BUF(tx
, db
);
1087 * Shouldn't dirty a regular buffer in syncing context. Private
1088 * objects may be dirtied in syncing context, but only if they
1089 * were already pre-dirtied in open context.
1091 ASSERT(!dmu_tx_is_syncing(tx
) ||
1092 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1093 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1094 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1096 * We make this assert for private objects as well, but after we
1097 * check if we're already dirty. They are allowed to re-dirty
1098 * in syncing context.
1100 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1101 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1102 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1104 mutex_enter(&db
->db_mtx
);
1106 * XXX make this true for indirects too? The problem is that
1107 * transactions created with dmu_tx_create_assigned() from
1108 * syncing context don't bother holding ahead.
1110 ASSERT(db
->db_level
!= 0 ||
1111 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1112 db
->db_state
== DB_NOFILL
);
1114 mutex_enter(&dn
->dn_mtx
);
1116 * Don't set dirtyctx to SYNC if we're just modifying this as we
1117 * initialize the objset.
1119 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
&&
1120 !BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1122 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1123 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1124 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_PUSHPAGE
);
1126 mutex_exit(&dn
->dn_mtx
);
1128 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1129 dn
->dn_have_spill
= B_TRUE
;
1132 * If this buffer is already dirty, we're done.
1134 drp
= &db
->db_last_dirty
;
1135 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1136 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1137 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1139 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1142 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1144 * If this buffer has already been written out,
1145 * we now need to reset its state.
1147 dbuf_unoverride(dr
);
1148 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1149 db
->db_state
!= DB_NOFILL
)
1150 arc_buf_thaw(db
->db_buf
);
1152 mutex_exit(&db
->db_mtx
);
1157 * Only valid if not already dirty.
1159 ASSERT(dn
->dn_object
== 0 ||
1160 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1161 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1163 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1164 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1165 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1166 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1167 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1168 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1171 * We should only be dirtying in syncing context if it's the
1172 * mos or we're initializing the os or it's a special object.
1173 * However, we are allowed to dirty in syncing context provided
1174 * we already dirtied it in open context. Hence we must make
1175 * this assertion only if we're not already dirty.
1178 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1179 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1180 ASSERT(db
->db
.db_size
!= 0);
1182 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1184 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1186 * Update the accounting.
1187 * Note: we delay "free accounting" until after we drop
1188 * the db_mtx. This keeps us from grabbing other locks
1189 * (and possibly deadlocking) in bp_get_dsize() while
1190 * also holding the db_mtx.
1192 dnode_willuse_space(dn
, db
->db
.db_size
, tx
);
1193 do_free_accounting
= dbuf_block_freeable(db
);
1197 * If this buffer is dirty in an old transaction group we need
1198 * to make a copy of it so that the changes we make in this
1199 * transaction group won't leak out when we sync the older txg.
1201 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_PUSHPAGE
);
1202 list_link_init(&dr
->dr_dirty_node
);
1203 if (db
->db_level
== 0) {
1204 void *data_old
= db
->db_buf
;
1206 if (db
->db_state
!= DB_NOFILL
) {
1207 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1208 dbuf_fix_old_data(db
, tx
->tx_txg
);
1209 data_old
= db
->db
.db_data
;
1210 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1212 * Release the data buffer from the cache so
1213 * that we can modify it without impacting
1214 * possible other users of this cached data
1215 * block. Note that indirect blocks and
1216 * private objects are not released until the
1217 * syncing state (since they are only modified
1220 arc_release(db
->db_buf
, db
);
1221 dbuf_fix_old_data(db
, tx
->tx_txg
);
1222 data_old
= db
->db_buf
;
1224 ASSERT(data_old
!= NULL
);
1226 dr
->dt
.dl
.dr_data
= data_old
;
1228 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
1229 list_create(&dr
->dt
.di
.dr_children
,
1230 sizeof (dbuf_dirty_record_t
),
1231 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1233 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1234 dr
->dr_accounted
= db
->db
.db_size
;
1236 dr
->dr_txg
= tx
->tx_txg
;
1241 * We could have been freed_in_flight between the dbuf_noread
1242 * and dbuf_dirty. We win, as though the dbuf_noread() had
1243 * happened after the free.
1245 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1246 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1247 mutex_enter(&dn
->dn_mtx
);
1248 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1249 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1252 mutex_exit(&dn
->dn_mtx
);
1253 db
->db_freed_in_flight
= FALSE
;
1257 * This buffer is now part of this txg
1259 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1260 db
->db_dirtycnt
+= 1;
1261 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1263 mutex_exit(&db
->db_mtx
);
1265 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1266 db
->db_blkid
== DMU_SPILL_BLKID
) {
1267 mutex_enter(&dn
->dn_mtx
);
1268 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1269 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1270 mutex_exit(&dn
->dn_mtx
);
1271 dnode_setdirty(dn
, tx
);
1274 } else if (do_free_accounting
) {
1275 blkptr_t
*bp
= db
->db_blkptr
;
1276 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1277 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1279 * This is only a guess -- if the dbuf is dirty
1280 * in a previous txg, we don't know how much
1281 * space it will use on disk yet. We should
1282 * really have the struct_rwlock to access
1283 * db_blkptr, but since this is just a guess,
1284 * it's OK if we get an odd answer.
1286 ddt_prefetch(os
->os_spa
, bp
);
1287 dnode_willuse_space(dn
, -willfree
, tx
);
1290 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1291 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1292 drop_struct_lock
= TRUE
;
1295 if (db
->db_level
== 0) {
1296 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1297 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1300 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1301 dmu_buf_impl_t
*parent
= db
->db_parent
;
1302 dbuf_dirty_record_t
*di
;
1303 int parent_held
= FALSE
;
1305 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1306 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1308 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1309 db
->db_blkid
>> epbs
, FTAG
);
1310 ASSERT(parent
!= NULL
);
1313 if (drop_struct_lock
)
1314 rw_exit(&dn
->dn_struct_rwlock
);
1315 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1316 di
= dbuf_dirty(parent
, tx
);
1318 dbuf_rele(parent
, FTAG
);
1320 mutex_enter(&db
->db_mtx
);
1322 * Since we've dropped the mutex, it's possible that
1323 * dbuf_undirty() might have changed this out from under us.
1325 if (db
->db_last_dirty
== dr
||
1326 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1327 mutex_enter(&di
->dt
.di
.dr_mtx
);
1328 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1329 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1330 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1331 mutex_exit(&di
->dt
.di
.dr_mtx
);
1334 mutex_exit(&db
->db_mtx
);
1336 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1337 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1338 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1339 mutex_enter(&dn
->dn_mtx
);
1340 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1341 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1342 mutex_exit(&dn
->dn_mtx
);
1343 if (drop_struct_lock
)
1344 rw_exit(&dn
->dn_struct_rwlock
);
1347 dnode_setdirty(dn
, tx
);
1353 * Undirty a buffer in the transaction group referenced by the given
1354 * transaction. Return whether this evicted the dbuf.
1357 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1360 uint64_t txg
= tx
->tx_txg
;
1361 dbuf_dirty_record_t
*dr
, **drp
;
1364 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1365 ASSERT0(db
->db_level
);
1366 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1369 * If this buffer is not dirty, we're done.
1371 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1372 if (dr
->dr_txg
<= txg
)
1374 if (dr
== NULL
|| dr
->dr_txg
< txg
)
1376 ASSERT(dr
->dr_txg
== txg
);
1377 ASSERT(dr
->dr_dbuf
== db
);
1382 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1384 ASSERT(db
->db
.db_size
!= 0);
1387 * Any space we accounted for in dp_dirty_* will be cleaned up by
1388 * dsl_pool_sync(). This is relatively rare so the discrepancy
1389 * is not a big deal.
1395 * Note that there are three places in dbuf_dirty()
1396 * where this dirty record may be put on a list.
1397 * Make sure to do a list_remove corresponding to
1398 * every one of those list_insert calls.
1400 if (dr
->dr_parent
) {
1401 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1402 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1403 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1404 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1405 db
->db_level
+1 == dn
->dn_nlevels
) {
1406 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1407 mutex_enter(&dn
->dn_mtx
);
1408 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1409 mutex_exit(&dn
->dn_mtx
);
1413 if (db
->db_state
!= DB_NOFILL
) {
1414 dbuf_unoverride(dr
);
1416 ASSERT(db
->db_buf
!= NULL
);
1417 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1418 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1419 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
, db
));
1421 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1423 ASSERT(db
->db_dirtycnt
> 0);
1424 db
->db_dirtycnt
-= 1;
1426 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1427 arc_buf_t
*buf
= db
->db_buf
;
1429 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(buf
));
1430 dbuf_set_data(db
, NULL
);
1431 VERIFY(arc_buf_remove_ref(buf
, db
));
1440 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1442 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1443 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1445 ASSERT(tx
->tx_txg
!= 0);
1446 ASSERT(!refcount_is_zero(&db
->db_holds
));
1449 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1450 rf
|= DB_RF_HAVESTRUCT
;
1452 (void) dbuf_read(db
, NULL
, rf
);
1453 (void) dbuf_dirty(db
, tx
);
1457 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1459 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1461 db
->db_state
= DB_NOFILL
;
1463 dmu_buf_will_fill(db_fake
, tx
);
1467 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1469 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1471 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1472 ASSERT(tx
->tx_txg
!= 0);
1473 ASSERT(db
->db_level
== 0);
1474 ASSERT(!refcount_is_zero(&db
->db_holds
));
1476 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1477 dmu_tx_private_ok(tx
));
1480 (void) dbuf_dirty(db
, tx
);
1483 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1486 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1488 mutex_enter(&db
->db_mtx
);
1491 if (db
->db_state
== DB_FILL
) {
1492 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1493 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1494 /* we were freed while filling */
1495 /* XXX dbuf_undirty? */
1496 bzero(db
->db
.db_data
, db
->db
.db_size
);
1497 db
->db_freed_in_flight
= FALSE
;
1499 db
->db_state
= DB_CACHED
;
1500 cv_broadcast(&db
->db_changed
);
1502 mutex_exit(&db
->db_mtx
);
1506 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
1507 bp_embedded_type_t etype
, enum zio_compress comp
,
1508 int uncompressed_size
, int compressed_size
, int byteorder
,
1511 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
1512 struct dirty_leaf
*dl
;
1513 dmu_object_type_t type
;
1516 type
= DB_DNODE(db
)->dn_type
;
1519 ASSERT0(db
->db_level
);
1520 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1522 dmu_buf_will_not_fill(dbuf
, tx
);
1524 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1525 dl
= &db
->db_last_dirty
->dt
.dl
;
1526 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
1527 data
, comp
, uncompressed_size
, compressed_size
);
1528 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
1529 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
1530 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
1531 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
1533 dl
->dr_override_state
= DR_OVERRIDDEN
;
1534 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
1538 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1539 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1542 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1544 ASSERT(!refcount_is_zero(&db
->db_holds
));
1545 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1546 ASSERT(db
->db_level
== 0);
1547 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1548 ASSERT(buf
!= NULL
);
1549 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1550 ASSERT(tx
->tx_txg
!= 0);
1552 arc_return_buf(buf
, db
);
1553 ASSERT(arc_released(buf
));
1555 mutex_enter(&db
->db_mtx
);
1557 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1558 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1560 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1562 if (db
->db_state
== DB_CACHED
&&
1563 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1564 mutex_exit(&db
->db_mtx
);
1565 (void) dbuf_dirty(db
, tx
);
1566 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1567 VERIFY(arc_buf_remove_ref(buf
, db
));
1568 xuio_stat_wbuf_copied();
1572 xuio_stat_wbuf_nocopy();
1573 if (db
->db_state
== DB_CACHED
) {
1574 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1576 ASSERT(db
->db_buf
!= NULL
);
1577 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
1578 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
1579 if (!arc_released(db
->db_buf
)) {
1580 ASSERT(dr
->dt
.dl
.dr_override_state
==
1582 arc_release(db
->db_buf
, db
);
1584 dr
->dt
.dl
.dr_data
= buf
;
1585 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1586 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
1587 arc_release(db
->db_buf
, db
);
1588 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
));
1592 ASSERT(db
->db_buf
== NULL
);
1593 dbuf_set_data(db
, buf
);
1594 db
->db_state
= DB_FILL
;
1595 mutex_exit(&db
->db_mtx
);
1596 (void) dbuf_dirty(db
, tx
);
1597 dmu_buf_fill_done(&db
->db
, tx
);
1601 * "Clear" the contents of this dbuf. This will mark the dbuf
1602 * EVICTING and clear *most* of its references. Unfortunately,
1603 * when we are not holding the dn_dbufs_mtx, we can't clear the
1604 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1605 * in this case. For callers from the DMU we will usually see:
1606 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1607 * For the arc callback, we will usually see:
1608 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1609 * Sometimes, though, we will get a mix of these two:
1610 * DMU: dbuf_clear()->arc_clear_callback()
1611 * ARC: dbuf_do_evict()->dbuf_destroy()
1613 * This routine will dissociate the dbuf from the arc, by calling
1614 * arc_clear_callback(), but will not evict the data from the ARC.
1617 dbuf_clear(dmu_buf_impl_t
*db
)
1620 dmu_buf_impl_t
*parent
= db
->db_parent
;
1621 dmu_buf_impl_t
*dndb
;
1622 boolean_t dbuf_gone
= B_FALSE
;
1624 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1625 ASSERT(refcount_is_zero(&db
->db_holds
));
1627 dbuf_evict_user(db
);
1629 if (db
->db_state
== DB_CACHED
) {
1630 ASSERT(db
->db
.db_data
!= NULL
);
1631 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1632 zio_buf_free(db
->db
.db_data
, DN_MAX_BONUSLEN
);
1633 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1635 db
->db
.db_data
= NULL
;
1636 db
->db_state
= DB_UNCACHED
;
1639 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1640 ASSERT(db
->db_data_pending
== NULL
);
1642 db
->db_state
= DB_EVICTING
;
1643 db
->db_blkptr
= NULL
;
1648 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& MUTEX_HELD(&dn
->dn_dbufs_mtx
)) {
1649 list_remove(&dn
->dn_dbufs
, db
);
1650 (void) atomic_dec_32_nv(&dn
->dn_dbufs_count
);
1654 * Decrementing the dbuf count means that the hold corresponding
1655 * to the removed dbuf is no longer discounted in dnode_move(),
1656 * so the dnode cannot be moved until after we release the hold.
1657 * The membar_producer() ensures visibility of the decremented
1658 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1662 db
->db_dnode_handle
= NULL
;
1668 dbuf_gone
= arc_clear_callback(db
->db_buf
);
1671 mutex_exit(&db
->db_mtx
);
1674 * If this dbuf is referenced from an indirect dbuf,
1675 * decrement the ref count on the indirect dbuf.
1677 if (parent
&& parent
!= dndb
)
1678 dbuf_rele(parent
, db
);
1681 __attribute__((always_inline
))
1683 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
1684 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
1691 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1693 if (blkid
== DMU_SPILL_BLKID
) {
1694 mutex_enter(&dn
->dn_mtx
);
1695 if (dn
->dn_have_spill
&&
1696 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1697 *bpp
= &dn
->dn_phys
->dn_spill
;
1700 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1701 *parentp
= dn
->dn_dbuf
;
1702 mutex_exit(&dn
->dn_mtx
);
1706 if (dn
->dn_phys
->dn_nlevels
== 0)
1709 nlevels
= dn
->dn_phys
->dn_nlevels
;
1711 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1713 ASSERT3U(level
* epbs
, <, 64);
1714 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1715 if (level
>= nlevels
||
1716 (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
1717 /* the buffer has no parent yet */
1718 return (SET_ERROR(ENOENT
));
1719 } else if (level
< nlevels
-1) {
1720 /* this block is referenced from an indirect block */
1723 err
= dbuf_hold_impl(dn
, level
+1, blkid
>> epbs
,
1724 fail_sparse
, NULL
, parentp
);
1726 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
1727 blkid
>> epbs
, fail_sparse
, NULL
,
1728 parentp
, dh
->dh_depth
+ 1);
1729 err
= __dbuf_hold_impl(dh
+ 1);
1733 err
= dbuf_read(*parentp
, NULL
,
1734 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
1736 dbuf_rele(*parentp
, NULL
);
1740 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
1741 (blkid
& ((1ULL << epbs
) - 1));
1744 /* the block is referenced from the dnode */
1745 ASSERT3U(level
, ==, nlevels
-1);
1746 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
1747 blkid
< dn
->dn_phys
->dn_nblkptr
);
1749 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1750 *parentp
= dn
->dn_dbuf
;
1752 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
1757 static dmu_buf_impl_t
*
1758 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
1759 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
1761 objset_t
*os
= dn
->dn_objset
;
1762 dmu_buf_impl_t
*db
, *odb
;
1764 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1765 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
1767 db
= kmem_cache_alloc(dbuf_cache
, KM_PUSHPAGE
);
1770 db
->db
.db_object
= dn
->dn_object
;
1771 db
->db_level
= level
;
1772 db
->db_blkid
= blkid
;
1773 db
->db_last_dirty
= NULL
;
1774 db
->db_dirtycnt
= 0;
1775 db
->db_dnode_handle
= dn
->dn_handle
;
1776 db
->db_parent
= parent
;
1777 db
->db_blkptr
= blkptr
;
1779 db
->db_user_ptr
= NULL
;
1780 db
->db_user_data_ptr_ptr
= NULL
;
1781 db
->db_evict_func
= NULL
;
1782 db
->db_immediate_evict
= 0;
1783 db
->db_freed_in_flight
= 0;
1785 if (blkid
== DMU_BONUS_BLKID
) {
1786 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
1787 db
->db
.db_size
= DN_MAX_BONUSLEN
-
1788 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
1789 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1790 db
->db
.db_offset
= DMU_BONUS_BLKID
;
1791 db
->db_state
= DB_UNCACHED
;
1792 /* the bonus dbuf is not placed in the hash table */
1793 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1795 } else if (blkid
== DMU_SPILL_BLKID
) {
1796 db
->db
.db_size
= (blkptr
!= NULL
) ?
1797 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
1798 db
->db
.db_offset
= 0;
1801 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
1802 db
->db
.db_size
= blocksize
;
1803 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
1807 * Hold the dn_dbufs_mtx while we get the new dbuf
1808 * in the hash table *and* added to the dbufs list.
1809 * This prevents a possible deadlock with someone
1810 * trying to look up this dbuf before its added to the
1813 mutex_enter(&dn
->dn_dbufs_mtx
);
1814 db
->db_state
= DB_EVICTING
;
1815 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
1816 /* someone else inserted it first */
1817 kmem_cache_free(dbuf_cache
, db
);
1818 mutex_exit(&dn
->dn_dbufs_mtx
);
1821 list_insert_head(&dn
->dn_dbufs
, db
);
1822 if (db
->db_level
== 0 && db
->db_blkid
>=
1823 dn
->dn_unlisted_l0_blkid
)
1824 dn
->dn_unlisted_l0_blkid
= db
->db_blkid
+ 1;
1825 db
->db_state
= DB_UNCACHED
;
1826 mutex_exit(&dn
->dn_dbufs_mtx
);
1827 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1829 if (parent
&& parent
!= dn
->dn_dbuf
)
1830 dbuf_add_ref(parent
, db
);
1832 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1833 refcount_count(&dn
->dn_holds
) > 0);
1834 (void) refcount_add(&dn
->dn_holds
, db
);
1835 (void) atomic_inc_32_nv(&dn
->dn_dbufs_count
);
1837 dprintf_dbuf(db
, "db=%p\n", db
);
1843 dbuf_do_evict(void *private)
1845 dmu_buf_impl_t
*db
= private;
1847 if (!MUTEX_HELD(&db
->db_mtx
))
1848 mutex_enter(&db
->db_mtx
);
1850 ASSERT(refcount_is_zero(&db
->db_holds
));
1852 if (db
->db_state
!= DB_EVICTING
) {
1853 ASSERT(db
->db_state
== DB_CACHED
);
1858 mutex_exit(&db
->db_mtx
);
1865 dbuf_destroy(dmu_buf_impl_t
*db
)
1867 ASSERT(refcount_is_zero(&db
->db_holds
));
1869 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1871 * If this dbuf is still on the dn_dbufs list,
1872 * remove it from that list.
1874 if (db
->db_dnode_handle
!= NULL
) {
1879 mutex_enter(&dn
->dn_dbufs_mtx
);
1880 list_remove(&dn
->dn_dbufs
, db
);
1881 (void) atomic_dec_32_nv(&dn
->dn_dbufs_count
);
1882 mutex_exit(&dn
->dn_dbufs_mtx
);
1885 * Decrementing the dbuf count means that the hold
1886 * corresponding to the removed dbuf is no longer
1887 * discounted in dnode_move(), so the dnode cannot be
1888 * moved until after we release the hold.
1891 db
->db_dnode_handle
= NULL
;
1893 dbuf_hash_remove(db
);
1895 db
->db_parent
= NULL
;
1898 ASSERT(!list_link_active(&db
->db_link
));
1899 ASSERT(db
->db
.db_data
== NULL
);
1900 ASSERT(db
->db_hash_next
== NULL
);
1901 ASSERT(db
->db_blkptr
== NULL
);
1902 ASSERT(db
->db_data_pending
== NULL
);
1904 kmem_cache_free(dbuf_cache
, db
);
1905 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1909 dbuf_prefetch(dnode_t
*dn
, uint64_t blkid
, zio_priority_t prio
)
1911 dmu_buf_impl_t
*db
= NULL
;
1912 blkptr_t
*bp
= NULL
;
1914 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1915 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1917 if (dnode_block_freed(dn
, blkid
))
1920 /* dbuf_find() returns with db_mtx held */
1921 if ((db
= dbuf_find(dn
, 0, blkid
))) {
1923 * This dbuf is already in the cache. We assume that
1924 * it is already CACHED, or else about to be either
1927 mutex_exit(&db
->db_mtx
);
1931 if (dbuf_findbp(dn
, 0, blkid
, TRUE
, &db
, &bp
, NULL
) == 0) {
1932 if (bp
&& !BP_IS_HOLE(bp
) && !BP_IS_EMBEDDED(bp
)) {
1933 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
1934 uint32_t aflags
= ARC_NOWAIT
| ARC_PREFETCH
;
1935 zbookmark_phys_t zb
;
1937 SET_BOOKMARK(&zb
, ds
? ds
->ds_object
: DMU_META_OBJSET
,
1938 dn
->dn_object
, 0, blkid
);
1940 (void) arc_read(NULL
, dn
->dn_objset
->os_spa
,
1941 bp
, NULL
, NULL
, prio
,
1942 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
1946 dbuf_rele(db
, NULL
);
1950 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
1953 * Returns with db_holds incremented, and db_mtx not held.
1954 * Note: dn_struct_rwlock must be held.
1957 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
1959 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
1960 dh
->dh_parent
= NULL
;
1962 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
1963 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
1964 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
1966 *(dh
->dh_dbp
) = NULL
;
1968 /* dbuf_find() returns with db_mtx held */
1969 dh
->dh_db
= dbuf_find(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
);
1971 if (dh
->dh_db
== NULL
) {
1974 ASSERT3P(dh
->dh_parent
, ==, NULL
);
1975 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
1976 dh
->dh_fail_sparse
, &dh
->dh_parent
,
1978 if (dh
->dh_fail_sparse
) {
1979 if (dh
->dh_err
== 0 &&
1980 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
1981 dh
->dh_err
= SET_ERROR(ENOENT
);
1984 dbuf_rele(dh
->dh_parent
, NULL
);
1985 return (dh
->dh_err
);
1988 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
1989 return (dh
->dh_err
);
1990 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
1991 dh
->dh_parent
, dh
->dh_bp
);
1994 if (dh
->dh_db
->db_buf
&& refcount_is_zero(&dh
->dh_db
->db_holds
)) {
1995 arc_buf_add_ref(dh
->dh_db
->db_buf
, dh
->dh_db
);
1996 if (dh
->dh_db
->db_buf
->b_data
== NULL
) {
1997 dbuf_clear(dh
->dh_db
);
1998 if (dh
->dh_parent
) {
1999 dbuf_rele(dh
->dh_parent
, NULL
);
2000 dh
->dh_parent
= NULL
;
2004 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
2007 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
2010 * If this buffer is currently syncing out, and we are are
2011 * still referencing it from db_data, we need to make a copy
2012 * of it in case we decide we want to dirty it again in this txg.
2014 if (dh
->dh_db
->db_level
== 0 &&
2015 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
2016 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2017 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
2018 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
2020 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
2021 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
2023 dbuf_set_data(dh
->dh_db
,
2024 arc_buf_alloc(dh
->dh_dn
->dn_objset
->os_spa
,
2025 dh
->dh_db
->db
.db_size
, dh
->dh_db
, dh
->dh_type
));
2026 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
2027 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
2031 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
2032 dbuf_update_data(dh
->dh_db
);
2033 DBUF_VERIFY(dh
->dh_db
);
2034 mutex_exit(&dh
->dh_db
->db_mtx
);
2036 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2038 dbuf_rele(dh
->dh_parent
, NULL
);
2040 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2041 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2042 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2043 *(dh
->dh_dbp
) = dh
->dh_db
;
2049 * The following code preserves the recursive function dbuf_hold_impl()
2050 * but moves the local variables AND function arguments to the heap to
2051 * minimize the stack frame size. Enough space is initially allocated
2052 * on the stack for 20 levels of recursion.
2055 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
, int fail_sparse
,
2056 void *tag
, dmu_buf_impl_t
**dbp
)
2058 struct dbuf_hold_impl_data
*dh
;
2061 dh
= kmem_zalloc(sizeof (struct dbuf_hold_impl_data
) *
2062 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_PUSHPAGE
);
2063 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
, tag
, dbp
, 0);
2065 error
= __dbuf_hold_impl(dh
);
2067 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
2068 DBUF_HOLD_IMPL_MAX_DEPTH
);
2074 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2075 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, int fail_sparse
,
2076 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2079 dh
->dh_level
= level
;
2080 dh
->dh_blkid
= blkid
;
2081 dh
->dh_fail_sparse
= fail_sparse
;
2084 dh
->dh_depth
= depth
;
2088 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2091 int err
= dbuf_hold_impl(dn
, 0, blkid
, FALSE
, tag
, &db
);
2092 return (err
? NULL
: db
);
2096 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2099 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, tag
, &db
);
2100 return (err
? NULL
: db
);
2104 dbuf_create_bonus(dnode_t
*dn
)
2106 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2108 ASSERT(dn
->dn_bonus
== NULL
);
2109 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2113 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2115 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2118 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2119 return (SET_ERROR(ENOTSUP
));
2121 blksz
= SPA_MINBLOCKSIZE
;
2122 if (blksz
> SPA_MAXBLOCKSIZE
)
2123 blksz
= SPA_MAXBLOCKSIZE
;
2125 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2129 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2130 dbuf_new_size(db
, blksz
, tx
);
2131 rw_exit(&dn
->dn_struct_rwlock
);
2138 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2140 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2143 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2145 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2147 VERIFY(refcount_add(&db
->db_holds
, tag
) > 1);
2151 * If you call dbuf_rele() you had better not be referencing the dnode handle
2152 * unless you have some other direct or indirect hold on the dnode. (An indirect
2153 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2154 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2155 * dnode's parent dbuf evicting its dnode handles.
2158 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2160 mutex_enter(&db
->db_mtx
);
2161 dbuf_rele_and_unlock(db
, tag
);
2165 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
2167 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
2171 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2172 * db_dirtycnt and db_holds to be updated atomically.
2175 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2179 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2183 * Remove the reference to the dbuf before removing its hold on the
2184 * dnode so we can guarantee in dnode_move() that a referenced bonus
2185 * buffer has a corresponding dnode hold.
2187 holds
= refcount_remove(&db
->db_holds
, tag
);
2191 * We can't freeze indirects if there is a possibility that they
2192 * may be modified in the current syncing context.
2194 if (db
->db_buf
&& holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0))
2195 arc_buf_freeze(db
->db_buf
);
2197 if (holds
== db
->db_dirtycnt
&&
2198 db
->db_level
== 0 && db
->db_immediate_evict
)
2199 dbuf_evict_user(db
);
2202 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2203 mutex_exit(&db
->db_mtx
);
2206 * If the dnode moves here, we cannot cross this barrier
2207 * until the move completes.
2210 (void) atomic_dec_32_nv(&DB_DNODE(db
)->dn_dbufs_count
);
2213 * The bonus buffer's dnode hold is no longer discounted
2214 * in dnode_move(). The dnode cannot move until after
2217 dnode_rele(DB_DNODE(db
), db
);
2218 } else if (db
->db_buf
== NULL
) {
2220 * This is a special case: we never associated this
2221 * dbuf with any data allocated from the ARC.
2223 ASSERT(db
->db_state
== DB_UNCACHED
||
2224 db
->db_state
== DB_NOFILL
);
2226 } else if (arc_released(db
->db_buf
)) {
2227 arc_buf_t
*buf
= db
->db_buf
;
2229 * This dbuf has anonymous data associated with it.
2231 dbuf_set_data(db
, NULL
);
2232 VERIFY(arc_buf_remove_ref(buf
, db
));
2235 VERIFY(!arc_buf_remove_ref(db
->db_buf
, db
));
2238 * A dbuf will be eligible for eviction if either the
2239 * 'primarycache' property is set or a duplicate
2240 * copy of this buffer is already cached in the arc.
2242 * In the case of the 'primarycache' a buffer
2243 * is considered for eviction if it matches the
2244 * criteria set in the property.
2246 * To decide if our buffer is considered a
2247 * duplicate, we must call into the arc to determine
2248 * if multiple buffers are referencing the same
2249 * block on-disk. If so, then we simply evict
2252 if (!DBUF_IS_CACHEABLE(db
)) {
2253 if (db
->db_blkptr
!= NULL
&&
2254 !BP_IS_HOLE(db
->db_blkptr
) &&
2255 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
2257 dmu_objset_spa(db
->db_objset
);
2258 blkptr_t bp
= *db
->db_blkptr
;
2260 arc_freed(spa
, &bp
);
2264 } else if (arc_buf_eviction_needed(db
->db_buf
)) {
2267 mutex_exit(&db
->db_mtx
);
2271 mutex_exit(&db
->db_mtx
);
2275 #pragma weak dmu_buf_refcount = dbuf_refcount
2277 dbuf_refcount(dmu_buf_impl_t
*db
)
2279 return (refcount_count(&db
->db_holds
));
2283 dmu_buf_set_user(dmu_buf_t
*db_fake
, void *user_ptr
, void *user_data_ptr_ptr
,
2284 dmu_buf_evict_func_t
*evict_func
)
2286 return (dmu_buf_update_user(db_fake
, NULL
, user_ptr
,
2287 user_data_ptr_ptr
, evict_func
));
2291 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, void *user_ptr
, void *user_data_ptr_ptr
,
2292 dmu_buf_evict_func_t
*evict_func
)
2294 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2296 db
->db_immediate_evict
= TRUE
;
2297 return (dmu_buf_update_user(db_fake
, NULL
, user_ptr
,
2298 user_data_ptr_ptr
, evict_func
));
2302 dmu_buf_update_user(dmu_buf_t
*db_fake
, void *old_user_ptr
, void *user_ptr
,
2303 void *user_data_ptr_ptr
, dmu_buf_evict_func_t
*evict_func
)
2305 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2306 ASSERT(db
->db_level
== 0);
2308 ASSERT((user_ptr
== NULL
) == (evict_func
== NULL
));
2310 mutex_enter(&db
->db_mtx
);
2312 if (db
->db_user_ptr
== old_user_ptr
) {
2313 db
->db_user_ptr
= user_ptr
;
2314 db
->db_user_data_ptr_ptr
= user_data_ptr_ptr
;
2315 db
->db_evict_func
= evict_func
;
2317 dbuf_update_data(db
);
2319 old_user_ptr
= db
->db_user_ptr
;
2322 mutex_exit(&db
->db_mtx
);
2323 return (old_user_ptr
);
2327 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2329 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2330 ASSERT(!refcount_is_zero(&db
->db_holds
));
2332 return (db
->db_user_ptr
);
2336 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2338 boolean_t res
= B_FALSE
;
2339 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2342 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2343 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2349 dmu_buf_get_blkptr(dmu_buf_t
*db
)
2351 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
2352 return (dbi
->db_blkptr
);
2356 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2358 /* ASSERT(dmu_tx_is_syncing(tx) */
2359 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2361 if (db
->db_blkptr
!= NULL
)
2364 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2365 db
->db_blkptr
= &dn
->dn_phys
->dn_spill
;
2366 BP_ZERO(db
->db_blkptr
);
2369 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2371 * This buffer was allocated at a time when there was
2372 * no available blkptrs from the dnode, or it was
2373 * inappropriate to hook it in (i.e., nlevels mis-match).
2375 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2376 ASSERT(db
->db_parent
== NULL
);
2377 db
->db_parent
= dn
->dn_dbuf
;
2378 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2381 dmu_buf_impl_t
*parent
= db
->db_parent
;
2382 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2384 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2385 if (parent
== NULL
) {
2386 mutex_exit(&db
->db_mtx
);
2387 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2388 (void) dbuf_hold_impl(dn
, db
->db_level
+1,
2389 db
->db_blkid
>> epbs
, FALSE
, db
, &parent
);
2390 rw_exit(&dn
->dn_struct_rwlock
);
2391 mutex_enter(&db
->db_mtx
);
2392 db
->db_parent
= parent
;
2394 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2395 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2401 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2402 * is critical the we not allow the compiler to inline this function in to
2403 * dbuf_sync_list() thereby drastically bloating the stack usage.
2405 noinline
static void
2406 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2408 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2412 ASSERT(dmu_tx_is_syncing(tx
));
2414 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2416 mutex_enter(&db
->db_mtx
);
2418 ASSERT(db
->db_level
> 0);
2421 /* Read the block if it hasn't been read yet. */
2422 if (db
->db_buf
== NULL
) {
2423 mutex_exit(&db
->db_mtx
);
2424 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2425 mutex_enter(&db
->db_mtx
);
2427 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2428 ASSERT(db
->db_buf
!= NULL
);
2432 /* Indirect block size must match what the dnode thinks it is. */
2433 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2434 dbuf_check_blkptr(dn
, db
);
2437 /* Provide the pending dirty record to child dbufs */
2438 db
->db_data_pending
= dr
;
2440 mutex_exit(&db
->db_mtx
);
2441 dbuf_write(dr
, db
->db_buf
, tx
);
2444 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2445 dbuf_sync_list(&dr
->dt
.di
.dr_children
, tx
);
2446 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2447 mutex_exit(&dr
->dt
.di
.dr_mtx
);
2452 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
2453 * critical the we not allow the compiler to inline this function in to
2454 * dbuf_sync_list() thereby drastically bloating the stack usage.
2456 noinline
static void
2457 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2459 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
2460 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2463 uint64_t txg
= tx
->tx_txg
;
2465 ASSERT(dmu_tx_is_syncing(tx
));
2467 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2469 mutex_enter(&db
->db_mtx
);
2471 * To be synced, we must be dirtied. But we
2472 * might have been freed after the dirty.
2474 if (db
->db_state
== DB_UNCACHED
) {
2475 /* This buffer has been freed since it was dirtied */
2476 ASSERT(db
->db
.db_data
== NULL
);
2477 } else if (db
->db_state
== DB_FILL
) {
2478 /* This buffer was freed and is now being re-filled */
2479 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
2481 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
2488 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2489 mutex_enter(&dn
->dn_mtx
);
2490 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
2491 mutex_exit(&dn
->dn_mtx
);
2495 * If this is a bonus buffer, simply copy the bonus data into the
2496 * dnode. It will be written out when the dnode is synced (and it
2497 * will be synced, since it must have been dirty for dbuf_sync to
2500 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2501 dbuf_dirty_record_t
**drp
;
2503 ASSERT(*datap
!= NULL
);
2504 ASSERT0(db
->db_level
);
2505 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
2506 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
2509 if (*datap
!= db
->db
.db_data
) {
2510 zio_buf_free(*datap
, DN_MAX_BONUSLEN
);
2511 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
2513 db
->db_data_pending
= NULL
;
2514 drp
= &db
->db_last_dirty
;
2516 drp
= &(*drp
)->dr_next
;
2517 ASSERT(dr
->dr_next
== NULL
);
2518 ASSERT(dr
->dr_dbuf
== db
);
2520 if (dr
->dr_dbuf
->db_level
!= 0) {
2521 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2522 list_destroy(&dr
->dt
.di
.dr_children
);
2524 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2525 ASSERT(db
->db_dirtycnt
> 0);
2526 db
->db_dirtycnt
-= 1;
2527 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
2534 * This function may have dropped the db_mtx lock allowing a dmu_sync
2535 * operation to sneak in. As a result, we need to ensure that we
2536 * don't check the dr_override_state until we have returned from
2537 * dbuf_check_blkptr.
2539 dbuf_check_blkptr(dn
, db
);
2542 * If this buffer is in the middle of an immediate write,
2543 * wait for the synchronous IO to complete.
2545 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
2546 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
2547 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2548 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
2551 if (db
->db_state
!= DB_NOFILL
&&
2552 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2553 refcount_count(&db
->db_holds
) > 1 &&
2554 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
2555 *datap
== db
->db_buf
) {
2557 * If this buffer is currently "in use" (i.e., there
2558 * are active holds and db_data still references it),
2559 * then make a copy before we start the write so that
2560 * any modifications from the open txg will not leak
2563 * NOTE: this copy does not need to be made for
2564 * objects only modified in the syncing context (e.g.
2565 * DNONE_DNODE blocks).
2567 int blksz
= arc_buf_size(*datap
);
2568 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2569 *datap
= arc_buf_alloc(os
->os_spa
, blksz
, db
, type
);
2570 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
2572 db
->db_data_pending
= dr
;
2574 mutex_exit(&db
->db_mtx
);
2576 dbuf_write(dr
, *datap
, tx
);
2578 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2579 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2580 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
2584 * Although zio_nowait() does not "wait for an IO", it does
2585 * initiate the IO. If this is an empty write it seems plausible
2586 * that the IO could actually be completed before the nowait
2587 * returns. We need to DB_DNODE_EXIT() first in case
2588 * zio_nowait() invalidates the dbuf.
2591 zio_nowait(dr
->dr_zio
);
2596 dbuf_sync_list(list_t
*list
, dmu_tx_t
*tx
)
2598 dbuf_dirty_record_t
*dr
;
2600 while ((dr
= list_head(list
))) {
2601 if (dr
->dr_zio
!= NULL
) {
2603 * If we find an already initialized zio then we
2604 * are processing the meta-dnode, and we have finished.
2605 * The dbufs for all dnodes are put back on the list
2606 * during processing, so that we can zio_wait()
2607 * these IOs after initiating all child IOs.
2609 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
2610 DMU_META_DNODE_OBJECT
);
2613 list_remove(list
, dr
);
2614 if (dr
->dr_dbuf
->db_level
> 0)
2615 dbuf_sync_indirect(dr
, tx
);
2617 dbuf_sync_leaf(dr
, tx
);
2623 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
2625 dmu_buf_impl_t
*db
= vdb
;
2627 blkptr_t
*bp
= zio
->io_bp
;
2628 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2629 spa_t
*spa
= zio
->io_spa
;
2634 ASSERT3P(db
->db_blkptr
, ==, bp
);
2638 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
2639 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
2640 zio
->io_prev_space_delta
= delta
;
2642 if (bp
->blk_birth
!= 0) {
2643 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
2644 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
2645 (db
->db_blkid
== DMU_SPILL_BLKID
&&
2646 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
2647 BP_IS_EMBEDDED(bp
));
2648 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
2651 mutex_enter(&db
->db_mtx
);
2654 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2655 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
2656 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
2657 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
2661 if (db
->db_level
== 0) {
2662 mutex_enter(&dn
->dn_mtx
);
2663 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
2664 db
->db_blkid
!= DMU_SPILL_BLKID
)
2665 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
2666 mutex_exit(&dn
->dn_mtx
);
2668 if (dn
->dn_type
== DMU_OT_DNODE
) {
2669 dnode_phys_t
*dnp
= db
->db
.db_data
;
2670 for (i
= db
->db
.db_size
>> DNODE_SHIFT
; i
> 0;
2672 if (dnp
->dn_type
!= DMU_OT_NONE
)
2676 if (BP_IS_HOLE(bp
)) {
2683 blkptr_t
*ibp
= db
->db
.db_data
;
2684 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2685 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
2686 if (BP_IS_HOLE(ibp
))
2688 fill
+= BP_GET_FILL(ibp
);
2693 if (!BP_IS_EMBEDDED(bp
))
2694 bp
->blk_fill
= fill
;
2696 mutex_exit(&db
->db_mtx
);
2700 * The SPA will call this callback several times for each zio - once
2701 * for every physical child i/o (zio->io_phys_children times). This
2702 * allows the DMU to monitor the progress of each logical i/o. For example,
2703 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2704 * block. There may be a long delay before all copies/fragments are completed,
2705 * so this callback allows us to retire dirty space gradually, as the physical
2710 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
2712 dmu_buf_impl_t
*db
= arg
;
2713 objset_t
*os
= db
->db_objset
;
2714 dsl_pool_t
*dp
= dmu_objset_pool(os
);
2715 dbuf_dirty_record_t
*dr
;
2718 dr
= db
->db_data_pending
;
2719 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
2722 * The callback will be called io_phys_children times. Retire one
2723 * portion of our dirty space each time we are called. Any rounding
2724 * error will be cleaned up by dsl_pool_sync()'s call to
2725 * dsl_pool_undirty_space().
2727 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
2728 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
2733 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
2735 dmu_buf_impl_t
*db
= vdb
;
2736 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2737 blkptr_t
*bp
= db
->db_blkptr
;
2738 objset_t
*os
= db
->db_objset
;
2739 dmu_tx_t
*tx
= os
->os_synctx
;
2740 dbuf_dirty_record_t
**drp
, *dr
;
2742 ASSERT0(zio
->io_error
);
2743 ASSERT(db
->db_blkptr
== bp
);
2746 * For nopwrites and rewrites we ensure that the bp matches our
2747 * original and bypass all the accounting.
2749 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
2750 ASSERT(BP_EQUAL(bp
, bp_orig
));
2752 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
2753 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
2754 dsl_dataset_block_born(ds
, bp
, tx
);
2757 mutex_enter(&db
->db_mtx
);
2761 drp
= &db
->db_last_dirty
;
2762 while ((dr
= *drp
) != db
->db_data_pending
)
2764 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2765 ASSERT(dr
->dr_dbuf
== db
);
2766 ASSERT(dr
->dr_next
== NULL
);
2770 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2775 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
2776 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
2777 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
2782 if (db
->db_level
== 0) {
2783 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2784 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
2785 if (db
->db_state
!= DB_NOFILL
) {
2786 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2787 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
2789 else if (!arc_released(db
->db_buf
))
2790 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
2797 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2798 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
2799 if (!BP_IS_HOLE(db
->db_blkptr
)) {
2800 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
2802 ASSERT3U(db
->db_blkid
, <=,
2803 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
2804 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
2806 if (!arc_released(db
->db_buf
))
2807 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
2810 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2811 list_destroy(&dr
->dt
.di
.dr_children
);
2813 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2815 cv_broadcast(&db
->db_changed
);
2816 ASSERT(db
->db_dirtycnt
> 0);
2817 db
->db_dirtycnt
-= 1;
2818 db
->db_data_pending
= NULL
;
2819 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
2823 dbuf_write_nofill_ready(zio_t
*zio
)
2825 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
2829 dbuf_write_nofill_done(zio_t
*zio
)
2831 dbuf_write_done(zio
, NULL
, zio
->io_private
);
2835 dbuf_write_override_ready(zio_t
*zio
)
2837 dbuf_dirty_record_t
*dr
= zio
->io_private
;
2838 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2840 dbuf_write_ready(zio
, NULL
, db
);
2844 dbuf_write_override_done(zio_t
*zio
)
2846 dbuf_dirty_record_t
*dr
= zio
->io_private
;
2847 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2848 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
2850 mutex_enter(&db
->db_mtx
);
2851 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
2852 if (!BP_IS_HOLE(obp
))
2853 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
2854 arc_release(dr
->dt
.dl
.dr_data
, db
);
2856 mutex_exit(&db
->db_mtx
);
2858 dbuf_write_done(zio
, NULL
, db
);
2861 /* Issue I/O to commit a dirty buffer to disk. */
2863 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
2865 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2868 dmu_buf_impl_t
*parent
= db
->db_parent
;
2869 uint64_t txg
= tx
->tx_txg
;
2870 zbookmark_phys_t zb
;
2879 if (db
->db_state
!= DB_NOFILL
) {
2880 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
2882 * Private object buffers are released here rather
2883 * than in dbuf_dirty() since they are only modified
2884 * in the syncing context and we don't want the
2885 * overhead of making multiple copies of the data.
2887 if (BP_IS_HOLE(db
->db_blkptr
)) {
2890 dbuf_release_bp(db
);
2895 if (parent
!= dn
->dn_dbuf
) {
2896 /* Our parent is an indirect block. */
2897 /* We have a dirty parent that has been scheduled for write. */
2898 ASSERT(parent
&& parent
->db_data_pending
);
2899 /* Our parent's buffer is one level closer to the dnode. */
2900 ASSERT(db
->db_level
== parent
->db_level
-1);
2902 * We're about to modify our parent's db_data by modifying
2903 * our block pointer, so the parent must be released.
2905 ASSERT(arc_released(parent
->db_buf
));
2906 zio
= parent
->db_data_pending
->dr_zio
;
2908 /* Our parent is the dnode itself. */
2909 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
2910 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
2911 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
2912 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2913 ASSERT3P(db
->db_blkptr
, ==,
2914 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
2918 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
2919 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
2922 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
2923 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
2924 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
2926 if (db
->db_blkid
== DMU_SPILL_BLKID
)
2928 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
2930 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
2933 if (db
->db_level
== 0 &&
2934 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
2936 * The BP for this block has been provided by open context
2937 * (by dmu_sync() or dmu_buf_write_embedded()).
2939 void *contents
= (data
!= NULL
) ? data
->b_data
: NULL
;
2941 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
2942 db
->db_blkptr
, contents
, db
->db
.db_size
, &zp
,
2943 dbuf_write_override_ready
, NULL
, dbuf_write_override_done
,
2944 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
2945 mutex_enter(&db
->db_mtx
);
2946 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
2947 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
2948 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
2949 mutex_exit(&db
->db_mtx
);
2950 } else if (db
->db_state
== DB_NOFILL
) {
2951 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
);
2952 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
2953 db
->db_blkptr
, NULL
, db
->db
.db_size
, &zp
,
2954 dbuf_write_nofill_ready
, NULL
, dbuf_write_nofill_done
, db
,
2955 ZIO_PRIORITY_ASYNC_WRITE
,
2956 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
2958 ASSERT(arc_released(data
));
2959 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
2960 db
->db_blkptr
, data
, DBUF_IS_L2CACHEABLE(db
),
2961 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dbuf_write_ready
,
2962 dbuf_write_physdone
, dbuf_write_done
, db
,
2963 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
2967 #if defined(_KERNEL) && defined(HAVE_SPL)
2968 EXPORT_SYMBOL(dbuf_find
);
2969 EXPORT_SYMBOL(dbuf_is_metadata
);
2970 EXPORT_SYMBOL(dbuf_evict
);
2971 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
2972 EXPORT_SYMBOL(dbuf_whichblock
);
2973 EXPORT_SYMBOL(dbuf_read
);
2974 EXPORT_SYMBOL(dbuf_unoverride
);
2975 EXPORT_SYMBOL(dbuf_free_range
);
2976 EXPORT_SYMBOL(dbuf_new_size
);
2977 EXPORT_SYMBOL(dbuf_release_bp
);
2978 EXPORT_SYMBOL(dbuf_dirty
);
2979 EXPORT_SYMBOL(dmu_buf_will_dirty
);
2980 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
2981 EXPORT_SYMBOL(dmu_buf_will_fill
);
2982 EXPORT_SYMBOL(dmu_buf_fill_done
);
2983 EXPORT_SYMBOL(dmu_buf_rele
);
2984 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
2985 EXPORT_SYMBOL(dbuf_clear
);
2986 EXPORT_SYMBOL(dbuf_prefetch
);
2987 EXPORT_SYMBOL(dbuf_hold_impl
);
2988 EXPORT_SYMBOL(dbuf_hold
);
2989 EXPORT_SYMBOL(dbuf_hold_level
);
2990 EXPORT_SYMBOL(dbuf_create_bonus
);
2991 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
2992 EXPORT_SYMBOL(dbuf_rm_spill
);
2993 EXPORT_SYMBOL(dbuf_add_ref
);
2994 EXPORT_SYMBOL(dbuf_rele
);
2995 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
2996 EXPORT_SYMBOL(dbuf_refcount
);
2997 EXPORT_SYMBOL(dbuf_sync_list
);
2998 EXPORT_SYMBOL(dmu_buf_set_user
);
2999 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
3000 EXPORT_SYMBOL(dmu_buf_update_user
);
3001 EXPORT_SYMBOL(dmu_buf_get_user
);
3002 EXPORT_SYMBOL(dmu_buf_freeable
);
3003 EXPORT_SYMBOL(dmu_buf_get_blkptr
);