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.
26 #include <sys/zfs_context.h>
29 #include <sys/dmu_impl.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dmu_tx.h>
37 #include <sys/dmu_zfetch.h>
39 #include <sys/sa_impl.h>
41 struct dbuf_hold_impl_data
{
42 /* Function arguments */
48 dmu_buf_impl_t
**dh_dbp
;
50 dmu_buf_impl_t
*dh_db
;
51 dmu_buf_impl_t
*dh_parent
;
54 dbuf_dirty_record_t
*dh_dr
;
55 arc_buf_contents_t dh_type
;
59 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
60 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, int fail_sparse
,
61 void *tag
, dmu_buf_impl_t
**dbp
, int depth
);
62 static int __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
);
64 static void dbuf_destroy(dmu_buf_impl_t
*db
);
65 static int dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
);
66 static void dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
);
69 * Global data structures and functions for the dbuf cache.
71 static kmem_cache_t
*dbuf_cache
;
75 dbuf_cons(void *vdb
, void *unused
, int kmflag
)
77 dmu_buf_impl_t
*db
= vdb
;
78 bzero(db
, sizeof (dmu_buf_impl_t
));
80 mutex_init(&db
->db_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
81 cv_init(&db
->db_changed
, NULL
, CV_DEFAULT
, NULL
);
82 refcount_create(&db
->db_holds
);
83 list_link_init(&db
->db_link
);
89 dbuf_dest(void *vdb
, void *unused
)
91 dmu_buf_impl_t
*db
= vdb
;
92 mutex_destroy(&db
->db_mtx
);
93 cv_destroy(&db
->db_changed
);
94 refcount_destroy(&db
->db_holds
);
98 * dbuf hash table routines
100 static dbuf_hash_table_t dbuf_hash_table
;
102 static uint64_t dbuf_hash_count
;
105 dbuf_hash(void *os
, uint64_t obj
, uint8_t lvl
, uint64_t blkid
)
107 uintptr_t osv
= (uintptr_t)os
;
108 uint64_t crc
= -1ULL;
110 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
111 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (lvl
)) & 0xFF];
112 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (osv
>> 6)) & 0xFF];
113 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 0)) & 0xFF];
114 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (obj
>> 8)) & 0xFF];
115 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 0)) & 0xFF];
116 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ (blkid
>> 8)) & 0xFF];
118 crc
^= (osv
>>14) ^ (obj
>>16) ^ (blkid
>>16);
123 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
125 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
126 ((dbuf)->db.db_object == (obj) && \
127 (dbuf)->db_objset == (os) && \
128 (dbuf)->db_level == (level) && \
129 (dbuf)->db_blkid == (blkid))
132 dbuf_find(dnode_t
*dn
, uint8_t level
, uint64_t blkid
)
134 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
135 objset_t
*os
= dn
->dn_objset
;
142 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
143 idx
= hv
& h
->hash_table_mask
;
145 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
146 for (db
= h
->hash_table
[idx
]; db
!= NULL
; db
= db
->db_hash_next
) {
147 if (DBUF_EQUAL(db
, os
, obj
, level
, blkid
)) {
148 mutex_enter(&db
->db_mtx
);
149 if (db
->db_state
!= DB_EVICTING
) {
150 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
153 mutex_exit(&db
->db_mtx
);
156 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
161 * Insert an entry into the hash table. If there is already an element
162 * equal to elem in the hash table, then the already existing element
163 * will be returned and the new element will not be inserted.
164 * Otherwise returns NULL.
166 static dmu_buf_impl_t
*
167 dbuf_hash_insert(dmu_buf_impl_t
*db
)
169 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
170 objset_t
*os
= db
->db_objset
;
171 uint64_t obj
= db
->db
.db_object
;
172 int level
= db
->db_level
;
173 uint64_t blkid
, hv
, idx
;
176 blkid
= db
->db_blkid
;
177 hv
= DBUF_HASH(os
, obj
, level
, blkid
);
178 idx
= hv
& h
->hash_table_mask
;
180 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
181 for (dbf
= h
->hash_table
[idx
]; dbf
!= NULL
; dbf
= dbf
->db_hash_next
) {
182 if (DBUF_EQUAL(dbf
, os
, obj
, level
, blkid
)) {
183 mutex_enter(&dbf
->db_mtx
);
184 if (dbf
->db_state
!= DB_EVICTING
) {
185 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
188 mutex_exit(&dbf
->db_mtx
);
192 mutex_enter(&db
->db_mtx
);
193 db
->db_hash_next
= h
->hash_table
[idx
];
194 h
->hash_table
[idx
] = db
;
195 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
196 atomic_add_64(&dbuf_hash_count
, 1);
202 * Remove an entry from the hash table. This operation will
203 * fail if there are any existing holds on the db.
206 dbuf_hash_remove(dmu_buf_impl_t
*db
)
208 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
210 dmu_buf_impl_t
*dbf
, **dbp
;
212 hv
= DBUF_HASH(db
->db_objset
, db
->db
.db_object
,
213 db
->db_level
, db
->db_blkid
);
214 idx
= hv
& h
->hash_table_mask
;
217 * We musn't hold db_mtx to maintin lock ordering:
218 * DBUF_HASH_MUTEX > db_mtx.
220 ASSERT(refcount_is_zero(&db
->db_holds
));
221 ASSERT(db
->db_state
== DB_EVICTING
);
222 ASSERT(!MUTEX_HELD(&db
->db_mtx
));
224 mutex_enter(DBUF_HASH_MUTEX(h
, idx
));
225 dbp
= &h
->hash_table
[idx
];
226 while ((dbf
= *dbp
) != db
) {
227 dbp
= &dbf
->db_hash_next
;
230 *dbp
= db
->db_hash_next
;
231 db
->db_hash_next
= NULL
;
232 mutex_exit(DBUF_HASH_MUTEX(h
, idx
));
233 atomic_add_64(&dbuf_hash_count
, -1);
236 static arc_evict_func_t dbuf_do_evict
;
239 dbuf_evict_user(dmu_buf_impl_t
*db
)
241 ASSERT(MUTEX_HELD(&db
->db_mtx
));
243 if (db
->db_level
!= 0 || db
->db_evict_func
== NULL
)
246 if (db
->db_user_data_ptr_ptr
)
247 *db
->db_user_data_ptr_ptr
= db
->db
.db_data
;
248 db
->db_evict_func(&db
->db
, db
->db_user_ptr
);
249 db
->db_user_ptr
= NULL
;
250 db
->db_user_data_ptr_ptr
= NULL
;
251 db
->db_evict_func
= NULL
;
255 dbuf_is_metadata(dmu_buf_impl_t
*db
)
257 if (db
->db_level
> 0) {
260 boolean_t is_metadata
;
263 is_metadata
= dmu_ot
[DB_DNODE(db
)->dn_type
].ot_metadata
;
266 return (is_metadata
);
271 dbuf_evict(dmu_buf_impl_t
*db
)
273 ASSERT(MUTEX_HELD(&db
->db_mtx
));
274 ASSERT(db
->db_buf
== NULL
);
275 ASSERT(db
->db_data_pending
== NULL
);
284 uint64_t hsize
= 1ULL << 16;
285 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
289 * The hash table is big enough to fill all of physical memory
290 * with an average 4K block size. The table will take up
291 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
293 while (hsize
* 4096 < physmem
* PAGESIZE
)
297 h
->hash_table_mask
= hsize
- 1;
298 #if defined(_KERNEL) && defined(HAVE_SPL)
299 /* Large allocations which do not require contiguous pages
300 * should be using vmem_alloc() in the linux kernel */
301 h
->hash_table
= vmem_zalloc(hsize
* sizeof (void *), KM_PUSHPAGE
);
303 h
->hash_table
= kmem_zalloc(hsize
* sizeof (void *), KM_NOSLEEP
);
305 if (h
->hash_table
== NULL
) {
306 /* XXX - we should really return an error instead of assert */
307 ASSERT(hsize
> (1ULL << 10));
312 dbuf_cache
= kmem_cache_create("dmu_buf_impl_t",
313 sizeof (dmu_buf_impl_t
),
314 0, dbuf_cons
, dbuf_dest
, NULL
, NULL
, NULL
, 0);
316 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
317 mutex_init(&h
->hash_mutexes
[i
], NULL
, MUTEX_DEFAULT
, NULL
);
323 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
326 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
327 mutex_destroy(&h
->hash_mutexes
[i
]);
328 #if defined(_KERNEL) && defined(HAVE_SPL)
329 /* Large allocations which do not require contiguous pages
330 * should be using vmem_free() in the linux kernel */
331 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
333 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
335 kmem_cache_destroy(dbuf_cache
);
344 dbuf_verify(dmu_buf_impl_t
*db
)
347 dbuf_dirty_record_t
*dr
;
349 ASSERT(MUTEX_HELD(&db
->db_mtx
));
351 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
354 ASSERT(db
->db_objset
!= NULL
);
358 ASSERT(db
->db_parent
== NULL
);
359 ASSERT(db
->db_blkptr
== NULL
);
361 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
362 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
363 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
364 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
365 db
->db_blkid
== DMU_SPILL_BLKID
||
366 !list_is_empty(&dn
->dn_dbufs
));
368 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
370 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
371 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
372 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
374 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
375 ASSERT3U(db
->db
.db_offset
, ==, 0);
377 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
380 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
381 ASSERT(dr
->dr_dbuf
== db
);
383 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
384 ASSERT(dr
->dr_dbuf
== db
);
387 * We can't assert that db_size matches dn_datablksz because it
388 * can be momentarily different when another thread is doing
391 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
392 dr
= db
->db_data_pending
;
394 * It should only be modified in syncing context, so
395 * make sure we only have one copy of the data.
397 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
400 /* verify db->db_blkptr */
402 if (db
->db_parent
== dn
->dn_dbuf
) {
403 /* db is pointed to by the dnode */
404 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
405 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
406 ASSERT(db
->db_parent
== NULL
);
408 ASSERT(db
->db_parent
!= NULL
);
409 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
410 ASSERT3P(db
->db_blkptr
, ==,
411 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
413 /* db is pointed to by an indirect block */
414 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
416 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
417 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
420 * dnode_grow_indblksz() can make this fail if we don't
421 * have the struct_rwlock. XXX indblksz no longer
422 * grows. safe to do this now?
424 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
425 ASSERT3P(db
->db_blkptr
, ==,
426 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
427 db
->db_blkid
% epb
));
431 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
432 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
433 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
434 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
436 * If the blkptr isn't set but they have nonzero data,
437 * it had better be dirty, otherwise we'll lose that
438 * data when we evict this buffer.
440 if (db
->db_dirtycnt
== 0) {
441 ASSERTV(uint64_t *buf
= db
->db
.db_data
);
444 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
454 dbuf_update_data(dmu_buf_impl_t
*db
)
456 ASSERT(MUTEX_HELD(&db
->db_mtx
));
457 if (db
->db_level
== 0 && db
->db_user_data_ptr_ptr
) {
458 ASSERT(!refcount_is_zero(&db
->db_holds
));
459 *db
->db_user_data_ptr_ptr
= db
->db
.db_data
;
464 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
466 ASSERT(MUTEX_HELD(&db
->db_mtx
));
467 ASSERT(db
->db_buf
== NULL
|| !arc_has_callback(db
->db_buf
));
470 ASSERT(buf
->b_data
!= NULL
);
471 db
->db
.db_data
= buf
->b_data
;
472 if (!arc_released(buf
))
473 arc_set_callback(buf
, dbuf_do_evict
, db
);
474 dbuf_update_data(db
);
477 db
->db
.db_data
= NULL
;
478 if (db
->db_state
!= DB_NOFILL
)
479 db
->db_state
= DB_UNCACHED
;
484 * Loan out an arc_buf for read. Return the loaned arc_buf.
487 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
491 mutex_enter(&db
->db_mtx
);
492 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
493 int blksz
= db
->db
.db_size
;
496 mutex_exit(&db
->db_mtx
);
497 DB_GET_SPA(&spa
, db
);
498 abuf
= arc_loan_buf(spa
, blksz
);
499 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
502 arc_loan_inuse_buf(abuf
, db
);
503 dbuf_set_data(db
, NULL
);
504 mutex_exit(&db
->db_mtx
);
510 dbuf_whichblock(dnode_t
*dn
, uint64_t offset
)
512 if (dn
->dn_datablkshift
) {
513 return (offset
>> dn
->dn_datablkshift
);
515 ASSERT3U(offset
, <, dn
->dn_datablksz
);
521 dbuf_read_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
523 dmu_buf_impl_t
*db
= vdb
;
525 mutex_enter(&db
->db_mtx
);
526 ASSERT3U(db
->db_state
, ==, DB_READ
);
528 * All reads are synchronous, so we must have a hold on the dbuf
530 ASSERT(refcount_count(&db
->db_holds
) > 0);
531 ASSERT(db
->db_buf
== NULL
);
532 ASSERT(db
->db
.db_data
== NULL
);
533 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
534 /* we were freed in flight; disregard any error */
535 arc_release(buf
, db
);
536 bzero(buf
->b_data
, db
->db
.db_size
);
538 db
->db_freed_in_flight
= FALSE
;
539 dbuf_set_data(db
, buf
);
540 db
->db_state
= DB_CACHED
;
541 } else if (zio
== NULL
|| zio
->io_error
== 0) {
542 dbuf_set_data(db
, buf
);
543 db
->db_state
= DB_CACHED
;
545 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
546 ASSERT3P(db
->db_buf
, ==, NULL
);
547 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
548 db
->db_state
= DB_UNCACHED
;
550 cv_broadcast(&db
->db_changed
);
551 dbuf_rele_and_unlock(db
, NULL
);
555 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t *flags
)
560 uint32_t aflags
= ARC_NOWAIT
;
565 ASSERT(!refcount_is_zero(&db
->db_holds
));
566 /* We need the struct_rwlock to prevent db_blkptr from changing. */
567 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
568 ASSERT(MUTEX_HELD(&db
->db_mtx
));
569 ASSERT(db
->db_state
== DB_UNCACHED
);
570 ASSERT(db
->db_buf
== NULL
);
572 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
573 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
575 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
576 db
->db
.db_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
577 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
578 if (bonuslen
< DN_MAX_BONUSLEN
)
579 bzero(db
->db
.db_data
, DN_MAX_BONUSLEN
);
581 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
583 dbuf_update_data(db
);
584 db
->db_state
= DB_CACHED
;
585 mutex_exit(&db
->db_mtx
);
590 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
591 * processes the delete record and clears the bp while we are waiting
592 * for the dn_mtx (resulting in a "no" from block_freed).
594 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
595 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
596 BP_IS_HOLE(db
->db_blkptr
)))) {
597 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
599 dbuf_set_data(db
, arc_buf_alloc(dn
->dn_objset
->os_spa
,
600 db
->db
.db_size
, db
, type
));
602 bzero(db
->db
.db_data
, db
->db
.db_size
);
603 db
->db_state
= DB_CACHED
;
604 *flags
|= DB_RF_CACHED
;
605 mutex_exit(&db
->db_mtx
);
609 spa
= dn
->dn_objset
->os_spa
;
612 db
->db_state
= DB_READ
;
613 mutex_exit(&db
->db_mtx
);
615 if (DBUF_IS_L2CACHEABLE(db
))
616 aflags
|= ARC_L2CACHE
;
618 SET_BOOKMARK(&zb
, db
->db_objset
->os_dsl_dataset
?
619 db
->db_objset
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
620 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
622 dbuf_add_ref(db
, NULL
);
623 /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */
626 pbuf
= db
->db_parent
->db_buf
;
628 pbuf
= db
->db_objset
->os_phys_buf
;
630 (void) dsl_read(zio
, spa
, db
->db_blkptr
, pbuf
,
631 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
,
632 (*flags
& DB_RF_CANFAIL
) ? ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
,
634 if (aflags
& ARC_CACHED
)
635 *flags
|= DB_RF_CACHED
;
639 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
642 int havepzio
= (zio
!= NULL
);
647 * We don't have to hold the mutex to check db_state because it
648 * can't be freed while we have a hold on the buffer.
650 ASSERT(!refcount_is_zero(&db
->db_holds
));
652 if (db
->db_state
== DB_NOFILL
)
657 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
658 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
660 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
661 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
662 DBUF_IS_CACHEABLE(db
);
664 mutex_enter(&db
->db_mtx
);
665 if (db
->db_state
== DB_CACHED
) {
666 mutex_exit(&db
->db_mtx
);
668 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
669 db
->db
.db_size
, TRUE
);
670 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
671 rw_exit(&dn
->dn_struct_rwlock
);
673 } else if (db
->db_state
== DB_UNCACHED
) {
674 spa_t
*spa
= dn
->dn_objset
->os_spa
;
677 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
678 dbuf_read_impl(db
, zio
, &flags
);
680 /* dbuf_read_impl has dropped db_mtx for us */
683 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
684 db
->db
.db_size
, flags
& DB_RF_CACHED
);
686 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
687 rw_exit(&dn
->dn_struct_rwlock
);
693 mutex_exit(&db
->db_mtx
);
695 dmu_zfetch(&dn
->dn_zfetch
, db
->db
.db_offset
,
696 db
->db
.db_size
, TRUE
);
697 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
698 rw_exit(&dn
->dn_struct_rwlock
);
701 mutex_enter(&db
->db_mtx
);
702 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
703 while (db
->db_state
== DB_READ
||
704 db
->db_state
== DB_FILL
) {
705 ASSERT(db
->db_state
== DB_READ
||
706 (flags
& DB_RF_HAVESTRUCT
) == 0);
707 cv_wait(&db
->db_changed
, &db
->db_mtx
);
709 if (db
->db_state
== DB_UNCACHED
)
712 mutex_exit(&db
->db_mtx
);
715 ASSERT(err
|| havepzio
|| db
->db_state
== DB_CACHED
);
720 dbuf_noread(dmu_buf_impl_t
*db
)
722 ASSERT(!refcount_is_zero(&db
->db_holds
));
723 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
724 mutex_enter(&db
->db_mtx
);
725 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
726 cv_wait(&db
->db_changed
, &db
->db_mtx
);
727 if (db
->db_state
== DB_UNCACHED
) {
728 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
731 ASSERT(db
->db_buf
== NULL
);
732 ASSERT(db
->db
.db_data
== NULL
);
733 DB_GET_SPA(&spa
, db
);
734 dbuf_set_data(db
, arc_buf_alloc(spa
, db
->db
.db_size
, db
, type
));
735 db
->db_state
= DB_FILL
;
736 } else if (db
->db_state
== DB_NOFILL
) {
737 dbuf_set_data(db
, NULL
);
739 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
741 mutex_exit(&db
->db_mtx
);
745 * This is our just-in-time copy function. It makes a copy of
746 * buffers, that have been modified in a previous transaction
747 * group, before we modify them in the current active group.
749 * This function is used in two places: when we are dirtying a
750 * buffer for the first time in a txg, and when we are freeing
751 * a range in a dnode that includes this buffer.
753 * Note that when we are called from dbuf_free_range() we do
754 * not put a hold on the buffer, we just traverse the active
755 * dbuf list for the dnode.
758 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
760 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
762 ASSERT(MUTEX_HELD(&db
->db_mtx
));
763 ASSERT(db
->db
.db_data
!= NULL
);
764 ASSERT(db
->db_level
== 0);
765 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
768 (dr
->dt
.dl
.dr_data
!=
769 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
773 * If the last dirty record for this dbuf has not yet synced
774 * and its referencing the dbuf data, either:
775 * reset the reference to point to a new copy,
776 * or (if there a no active holders)
777 * just null out the current db_data pointer.
779 ASSERT(dr
->dr_txg
>= txg
- 2);
780 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
781 /* Note that the data bufs here are zio_bufs */
782 dr
->dt
.dl
.dr_data
= zio_buf_alloc(DN_MAX_BONUSLEN
);
783 arc_space_consume(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
784 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, DN_MAX_BONUSLEN
);
785 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
786 int size
= db
->db
.db_size
;
787 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
790 DB_GET_SPA(&spa
, db
);
791 dr
->dt
.dl
.dr_data
= arc_buf_alloc(spa
, size
, db
, type
);
792 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
794 dbuf_set_data(db
, NULL
);
799 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
801 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
802 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
803 uint64_t txg
= dr
->dr_txg
;
805 ASSERT(MUTEX_HELD(&db
->db_mtx
));
806 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
807 ASSERT(db
->db_level
== 0);
809 if (db
->db_blkid
== DMU_BONUS_BLKID
||
810 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
813 ASSERT(db
->db_data_pending
!= dr
);
815 /* free this block */
816 if (!BP_IS_HOLE(bp
)) {
819 DB_GET_SPA(&spa
, db
);
820 zio_free(spa
, txg
, bp
);
822 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
824 * Release the already-written buffer, so we leave it in
825 * a consistent dirty state. Note that all callers are
826 * modifying the buffer, so they will immediately do
827 * another (redundant) arc_release(). Therefore, leave
828 * the buf thawed to save the effort of freezing &
829 * immediately re-thawing it.
831 arc_release(dr
->dt
.dl
.dr_data
, db
);
835 * Evict (if its unreferenced) or clear (if its referenced) any level-0
836 * data blocks in the free range, so that any future readers will find
837 * empty blocks. Also, if we happen accross any level-1 dbufs in the
838 * range that have not already been marked dirty, mark them dirty so
839 * they stay in memory.
842 dbuf_free_range(dnode_t
*dn
, uint64_t start
, uint64_t end
, dmu_tx_t
*tx
)
844 dmu_buf_impl_t
*db
, *db_next
;
845 uint64_t txg
= tx
->tx_txg
;
846 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
847 uint64_t first_l1
= start
>> epbs
;
848 uint64_t last_l1
= end
>> epbs
;
850 if (end
> dn
->dn_maxblkid
&& (end
!= DMU_SPILL_BLKID
)) {
851 end
= dn
->dn_maxblkid
;
852 last_l1
= end
>> epbs
;
854 dprintf_dnode(dn
, "start=%llu end=%llu\n", start
, end
);
855 mutex_enter(&dn
->dn_dbufs_mtx
);
856 for (db
= list_head(&dn
->dn_dbufs
); db
; db
= db_next
) {
857 db_next
= list_next(&dn
->dn_dbufs
, db
);
858 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
860 if (db
->db_level
== 1 &&
861 db
->db_blkid
>= first_l1
&& db
->db_blkid
<= last_l1
) {
862 mutex_enter(&db
->db_mtx
);
863 if (db
->db_last_dirty
&&
864 db
->db_last_dirty
->dr_txg
< txg
) {
865 dbuf_add_ref(db
, FTAG
);
866 mutex_exit(&db
->db_mtx
);
867 dbuf_will_dirty(db
, tx
);
870 mutex_exit(&db
->db_mtx
);
874 if (db
->db_level
!= 0)
876 dprintf_dbuf(db
, "found buf %s\n", "");
877 if (db
->db_blkid
< start
|| db
->db_blkid
> end
)
880 /* found a level 0 buffer in the range */
881 if (dbuf_undirty(db
, tx
))
884 mutex_enter(&db
->db_mtx
);
885 if (db
->db_state
== DB_UNCACHED
||
886 db
->db_state
== DB_NOFILL
||
887 db
->db_state
== DB_EVICTING
) {
888 ASSERT(db
->db
.db_data
== NULL
);
889 mutex_exit(&db
->db_mtx
);
892 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
893 /* will be handled in dbuf_read_done or dbuf_rele */
894 db
->db_freed_in_flight
= TRUE
;
895 mutex_exit(&db
->db_mtx
);
898 if (refcount_count(&db
->db_holds
) == 0) {
903 /* The dbuf is referenced */
905 if (db
->db_last_dirty
!= NULL
) {
906 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
908 if (dr
->dr_txg
== txg
) {
910 * This buffer is "in-use", re-adjust the file
911 * size to reflect that this buffer may
912 * contain new data when we sync.
914 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
915 db
->db_blkid
> dn
->dn_maxblkid
)
916 dn
->dn_maxblkid
= db
->db_blkid
;
920 * This dbuf is not dirty in the open context.
921 * Either uncache it (if its not referenced in
922 * the open context) or reset its contents to
925 dbuf_fix_old_data(db
, txg
);
928 /* clear the contents if its cached */
929 if (db
->db_state
== DB_CACHED
) {
930 ASSERT(db
->db
.db_data
!= NULL
);
931 arc_release(db
->db_buf
, db
);
932 bzero(db
->db
.db_data
, db
->db
.db_size
);
933 arc_buf_freeze(db
->db_buf
);
936 mutex_exit(&db
->db_mtx
);
938 mutex_exit(&dn
->dn_dbufs_mtx
);
942 dbuf_block_freeable(dmu_buf_impl_t
*db
)
944 dsl_dataset_t
*ds
= db
->db_objset
->os_dsl_dataset
;
945 uint64_t birth_txg
= 0;
948 * We don't need any locking to protect db_blkptr:
949 * If it's syncing, then db_last_dirty will be set
950 * so we'll ignore db_blkptr.
952 ASSERT(MUTEX_HELD(&db
->db_mtx
));
953 if (db
->db_last_dirty
)
954 birth_txg
= db
->db_last_dirty
->dr_txg
;
955 else if (db
->db_blkptr
)
956 birth_txg
= db
->db_blkptr
->blk_birth
;
959 * If we don't exist or are in a snapshot, we can't be freed.
960 * Don't pass the bp to dsl_dataset_block_freeable() since we
961 * are holding the db_mtx lock and might deadlock if we are
962 * prefetching a dedup-ed block.
965 return (ds
== NULL
||
966 dsl_dataset_block_freeable(ds
, NULL
, birth_txg
));
972 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
974 arc_buf_t
*buf
, *obuf
;
975 int osize
= db
->db
.db_size
;
976 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
979 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
984 /* XXX does *this* func really need the lock? */
985 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
988 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
989 * is OK, because there can be no other references to the db
990 * when we are changing its size, so no concurrent DB_FILL can
994 * XXX we should be doing a dbuf_read, checking the return
995 * value and returning that up to our callers
997 dbuf_will_dirty(db
, tx
);
999 /* create the data buffer for the new block */
1000 buf
= arc_buf_alloc(dn
->dn_objset
->os_spa
, size
, db
, type
);
1002 /* copy old block data to the new block */
1004 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1005 /* zero the remainder */
1007 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1009 mutex_enter(&db
->db_mtx
);
1010 dbuf_set_data(db
, buf
);
1011 VERIFY(arc_buf_remove_ref(obuf
, db
) == 1);
1012 db
->db
.db_size
= size
;
1014 if (db
->db_level
== 0) {
1015 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1016 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1018 mutex_exit(&db
->db_mtx
);
1020 dnode_willuse_space(dn
, size
-osize
, tx
);
1025 dbuf_release_bp(dmu_buf_impl_t
*db
)
1030 DB_GET_OBJSET(&os
, db
);
1031 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1032 ASSERT(arc_released(os
->os_phys_buf
) ||
1033 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1034 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1036 zb
.zb_objset
= os
->os_dsl_dataset
?
1037 os
->os_dsl_dataset
->ds_object
: 0;
1038 zb
.zb_object
= db
->db
.db_object
;
1039 zb
.zb_level
= db
->db_level
;
1040 zb
.zb_blkid
= db
->db_blkid
;
1041 (void) arc_release_bp(db
->db_buf
, db
,
1042 db
->db_blkptr
, os
->os_spa
, &zb
);
1045 dbuf_dirty_record_t
*
1046 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1050 dbuf_dirty_record_t
**drp
, *dr
;
1051 int drop_struct_lock
= FALSE
;
1052 boolean_t do_free_accounting
= B_FALSE
;
1053 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1055 ASSERT(tx
->tx_txg
!= 0);
1056 ASSERT(!refcount_is_zero(&db
->db_holds
));
1057 DMU_TX_DIRTY_BUF(tx
, db
);
1062 * Shouldn't dirty a regular buffer in syncing context. Private
1063 * objects may be dirtied in syncing context, but only if they
1064 * were already pre-dirtied in open context.
1066 ASSERT(!dmu_tx_is_syncing(tx
) ||
1067 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1068 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1069 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1071 * We make this assert for private objects as well, but after we
1072 * check if we're already dirty. They are allowed to re-dirty
1073 * in syncing context.
1075 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1076 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1077 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1079 mutex_enter(&db
->db_mtx
);
1081 * XXX make this true for indirects too? The problem is that
1082 * transactions created with dmu_tx_create_assigned() from
1083 * syncing context don't bother holding ahead.
1085 ASSERT(db
->db_level
!= 0 ||
1086 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1087 db
->db_state
== DB_NOFILL
);
1089 mutex_enter(&dn
->dn_mtx
);
1091 * Don't set dirtyctx to SYNC if we're just modifying this as we
1092 * initialize the objset.
1094 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
&&
1095 !BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1097 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1098 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1099 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_PUSHPAGE
);
1101 mutex_exit(&dn
->dn_mtx
);
1103 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1104 dn
->dn_have_spill
= B_TRUE
;
1107 * If this buffer is already dirty, we're done.
1109 drp
= &db
->db_last_dirty
;
1110 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1111 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1112 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1114 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1117 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1119 * If this buffer has already been written out,
1120 * we now need to reset its state.
1122 dbuf_unoverride(dr
);
1123 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1124 db
->db_state
!= DB_NOFILL
)
1125 arc_buf_thaw(db
->db_buf
);
1127 mutex_exit(&db
->db_mtx
);
1132 * Only valid if not already dirty.
1134 ASSERT(dn
->dn_object
== 0 ||
1135 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1136 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1138 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1139 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1140 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1141 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1142 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1143 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1146 * We should only be dirtying in syncing context if it's the
1147 * mos or we're initializing the os or it's a special object.
1148 * However, we are allowed to dirty in syncing context provided
1149 * we already dirtied it in open context. Hence we must make
1150 * this assertion only if we're not already dirty.
1153 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1154 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1155 ASSERT(db
->db
.db_size
!= 0);
1157 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1159 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1161 * Update the accounting.
1162 * Note: we delay "free accounting" until after we drop
1163 * the db_mtx. This keeps us from grabbing other locks
1164 * (and possibly deadlocking) in bp_get_dsize() while
1165 * also holding the db_mtx.
1167 dnode_willuse_space(dn
, db
->db
.db_size
, tx
);
1168 do_free_accounting
= dbuf_block_freeable(db
);
1172 * If this buffer is dirty in an old transaction group we need
1173 * to make a copy of it so that the changes we make in this
1174 * transaction group won't leak out when we sync the older txg.
1176 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_PUSHPAGE
);
1177 list_link_init(&dr
->dr_dirty_node
);
1178 if (db
->db_level
== 0) {
1179 void *data_old
= db
->db_buf
;
1181 if (db
->db_state
!= DB_NOFILL
) {
1182 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1183 dbuf_fix_old_data(db
, tx
->tx_txg
);
1184 data_old
= db
->db
.db_data
;
1185 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1187 * Release the data buffer from the cache so
1188 * that we can modify it without impacting
1189 * possible other users of this cached data
1190 * block. Note that indirect blocks and
1191 * private objects are not released until the
1192 * syncing state (since they are only modified
1195 arc_release(db
->db_buf
, db
);
1196 dbuf_fix_old_data(db
, tx
->tx_txg
);
1197 data_old
= db
->db_buf
;
1199 ASSERT(data_old
!= NULL
);
1201 dr
->dt
.dl
.dr_data
= data_old
;
1203 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
1204 list_create(&dr
->dt
.di
.dr_children
,
1205 sizeof (dbuf_dirty_record_t
),
1206 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1209 dr
->dr_txg
= tx
->tx_txg
;
1214 * We could have been freed_in_flight between the dbuf_noread
1215 * and dbuf_dirty. We win, as though the dbuf_noread() had
1216 * happened after the free.
1218 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1219 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1220 mutex_enter(&dn
->dn_mtx
);
1221 dnode_clear_range(dn
, db
->db_blkid
, 1, tx
);
1222 mutex_exit(&dn
->dn_mtx
);
1223 db
->db_freed_in_flight
= FALSE
;
1227 * This buffer is now part of this txg
1229 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1230 db
->db_dirtycnt
+= 1;
1231 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1233 mutex_exit(&db
->db_mtx
);
1235 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1236 db
->db_blkid
== DMU_SPILL_BLKID
) {
1237 mutex_enter(&dn
->dn_mtx
);
1238 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1239 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1240 mutex_exit(&dn
->dn_mtx
);
1241 dnode_setdirty(dn
, tx
);
1244 } else if (do_free_accounting
) {
1245 blkptr_t
*bp
= db
->db_blkptr
;
1246 int64_t willfree
= (bp
&& !BP_IS_HOLE(bp
)) ?
1247 bp_get_dsize(os
->os_spa
, bp
) : db
->db
.db_size
;
1249 * This is only a guess -- if the dbuf is dirty
1250 * in a previous txg, we don't know how much
1251 * space it will use on disk yet. We should
1252 * really have the struct_rwlock to access
1253 * db_blkptr, but since this is just a guess,
1254 * it's OK if we get an odd answer.
1256 ddt_prefetch(os
->os_spa
, bp
);
1257 dnode_willuse_space(dn
, -willfree
, tx
);
1260 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1261 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1262 drop_struct_lock
= TRUE
;
1265 if (db
->db_level
== 0) {
1266 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1267 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1270 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1271 dmu_buf_impl_t
*parent
= db
->db_parent
;
1272 dbuf_dirty_record_t
*di
;
1273 int parent_held
= FALSE
;
1275 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1276 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1278 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
1279 db
->db_blkid
>> epbs
, FTAG
);
1280 ASSERT(parent
!= NULL
);
1283 if (drop_struct_lock
)
1284 rw_exit(&dn
->dn_struct_rwlock
);
1285 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
1286 di
= dbuf_dirty(parent
, tx
);
1288 dbuf_rele(parent
, FTAG
);
1290 mutex_enter(&db
->db_mtx
);
1291 /* possible race with dbuf_undirty() */
1292 if (db
->db_last_dirty
== dr
||
1293 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
1294 mutex_enter(&di
->dt
.di
.dr_mtx
);
1295 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
1296 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1297 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
1298 mutex_exit(&di
->dt
.di
.dr_mtx
);
1301 mutex_exit(&db
->db_mtx
);
1303 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
1304 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
1305 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1306 mutex_enter(&dn
->dn_mtx
);
1307 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1308 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1309 mutex_exit(&dn
->dn_mtx
);
1310 if (drop_struct_lock
)
1311 rw_exit(&dn
->dn_struct_rwlock
);
1314 dnode_setdirty(dn
, tx
);
1320 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1323 uint64_t txg
= tx
->tx_txg
;
1324 dbuf_dirty_record_t
*dr
, **drp
;
1327 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1329 mutex_enter(&db
->db_mtx
);
1331 * If this buffer is not dirty, we're done.
1333 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
1334 if (dr
->dr_txg
<= txg
)
1336 if (dr
== NULL
|| dr
->dr_txg
< txg
) {
1337 mutex_exit(&db
->db_mtx
);
1340 ASSERT(dr
->dr_txg
== txg
);
1341 ASSERT(dr
->dr_dbuf
== db
);
1347 * If this buffer is currently held, we cannot undirty
1348 * it, since one of the current holders may be in the
1349 * middle of an update. Note that users of dbuf_undirty()
1350 * should not place a hold on the dbuf before the call.
1351 * Also note: we can get here with a spill block, so
1352 * test for that similar to how dbuf_dirty does.
1354 if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1355 mutex_exit(&db
->db_mtx
);
1356 /* Make sure we don't toss this buffer at sync phase */
1357 if (db
->db_blkid
!= DMU_SPILL_BLKID
) {
1358 mutex_enter(&dn
->dn_mtx
);
1359 dnode_clear_range(dn
, db
->db_blkid
, 1, tx
);
1360 mutex_exit(&dn
->dn_mtx
);
1366 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1368 ASSERT(db
->db
.db_size
!= 0);
1370 /* XXX would be nice to fix up dn_towrite_space[] */
1375 * Note that there are three places in dbuf_dirty()
1376 * where this dirty record may be put on a list.
1377 * Make sure to do a list_remove corresponding to
1378 * every one of those list_insert calls.
1380 if (dr
->dr_parent
) {
1381 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1382 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
1383 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
1384 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
1385 db
->db_level
+1 == dn
->dn_nlevels
) {
1386 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
1387 mutex_enter(&dn
->dn_mtx
);
1388 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
1389 mutex_exit(&dn
->dn_mtx
);
1393 if (db
->db_level
== 0) {
1394 if (db
->db_state
!= DB_NOFILL
) {
1395 dbuf_unoverride(dr
);
1397 ASSERT(db
->db_buf
!= NULL
);
1398 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
1399 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
1400 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
1404 ASSERT(db
->db_buf
!= NULL
);
1405 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
1406 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
1407 list_destroy(&dr
->dt
.di
.dr_children
);
1409 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
1411 ASSERT(db
->db_dirtycnt
> 0);
1412 db
->db_dirtycnt
-= 1;
1414 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
1415 arc_buf_t
*buf
= db
->db_buf
;
1417 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(buf
));
1418 dbuf_set_data(db
, NULL
);
1419 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
1424 mutex_exit(&db
->db_mtx
);
1428 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1430 dbuf_will_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1432 int rf
= DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
;
1434 ASSERT(tx
->tx_txg
!= 0);
1435 ASSERT(!refcount_is_zero(&db
->db_holds
));
1438 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
1439 rf
|= DB_RF_HAVESTRUCT
;
1441 (void) dbuf_read(db
, NULL
, rf
);
1442 (void) dbuf_dirty(db
, tx
);
1446 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1448 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1450 db
->db_state
= DB_NOFILL
;
1452 dmu_buf_will_fill(db_fake
, tx
);
1456 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
1458 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1460 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1461 ASSERT(tx
->tx_txg
!= 0);
1462 ASSERT(db
->db_level
== 0);
1463 ASSERT(!refcount_is_zero(&db
->db_holds
));
1465 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
1466 dmu_tx_private_ok(tx
));
1469 (void) dbuf_dirty(db
, tx
);
1472 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1475 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1477 mutex_enter(&db
->db_mtx
);
1480 if (db
->db_state
== DB_FILL
) {
1481 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1482 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1483 /* we were freed while filling */
1484 /* XXX dbuf_undirty? */
1485 bzero(db
->db
.db_data
, db
->db
.db_size
);
1486 db
->db_freed_in_flight
= FALSE
;
1488 db
->db_state
= DB_CACHED
;
1489 cv_broadcast(&db
->db_changed
);
1491 mutex_exit(&db
->db_mtx
);
1495 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1496 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1499 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
1501 ASSERT(!refcount_is_zero(&db
->db_holds
));
1502 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1503 ASSERT(db
->db_level
== 0);
1504 ASSERT(DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
);
1505 ASSERT(buf
!= NULL
);
1506 ASSERT(arc_buf_size(buf
) == db
->db
.db_size
);
1507 ASSERT(tx
->tx_txg
!= 0);
1509 arc_return_buf(buf
, db
);
1510 ASSERT(arc_released(buf
));
1512 mutex_enter(&db
->db_mtx
);
1514 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1515 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1517 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
1519 if (db
->db_state
== DB_CACHED
&&
1520 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
1521 mutex_exit(&db
->db_mtx
);
1522 (void) dbuf_dirty(db
, tx
);
1523 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
1524 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
1525 xuio_stat_wbuf_copied();
1529 xuio_stat_wbuf_nocopy();
1530 if (db
->db_state
== DB_CACHED
) {
1531 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1533 ASSERT(db
->db_buf
!= NULL
);
1534 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
1535 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
1536 if (!arc_released(db
->db_buf
)) {
1537 ASSERT(dr
->dt
.dl
.dr_override_state
==
1539 arc_release(db
->db_buf
, db
);
1541 dr
->dt
.dl
.dr_data
= buf
;
1542 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
) == 1);
1543 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
1544 arc_release(db
->db_buf
, db
);
1545 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
) == 1);
1549 ASSERT(db
->db_buf
== NULL
);
1550 dbuf_set_data(db
, buf
);
1551 db
->db_state
= DB_FILL
;
1552 mutex_exit(&db
->db_mtx
);
1553 (void) dbuf_dirty(db
, tx
);
1554 dbuf_fill_done(db
, tx
);
1558 * "Clear" the contents of this dbuf. This will mark the dbuf
1559 * EVICTING and clear *most* of its references. Unfortunetely,
1560 * when we are not holding the dn_dbufs_mtx, we can't clear the
1561 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1562 * in this case. For callers from the DMU we will usually see:
1563 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1564 * For the arc callback, we will usually see:
1565 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1566 * Sometimes, though, we will get a mix of these two:
1567 * DMU: dbuf_clear()->arc_buf_evict()
1568 * ARC: dbuf_do_evict()->dbuf_destroy()
1571 dbuf_clear(dmu_buf_impl_t
*db
)
1574 dmu_buf_impl_t
*parent
= db
->db_parent
;
1575 dmu_buf_impl_t
*dndb
;
1576 int dbuf_gone
= FALSE
;
1578 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1579 ASSERT(refcount_is_zero(&db
->db_holds
));
1581 dbuf_evict_user(db
);
1583 if (db
->db_state
== DB_CACHED
) {
1584 ASSERT(db
->db
.db_data
!= NULL
);
1585 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1586 zio_buf_free(db
->db
.db_data
, DN_MAX_BONUSLEN
);
1587 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
1589 db
->db
.db_data
= NULL
;
1590 db
->db_state
= DB_UNCACHED
;
1593 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
1594 ASSERT(db
->db_data_pending
== NULL
);
1596 db
->db_state
= DB_EVICTING
;
1597 db
->db_blkptr
= NULL
;
1602 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& MUTEX_HELD(&dn
->dn_dbufs_mtx
)) {
1603 list_remove(&dn
->dn_dbufs
, db
);
1604 (void) atomic_dec_32_nv(&dn
->dn_dbufs_count
);
1608 * Decrementing the dbuf count means that the hold corresponding
1609 * to the removed dbuf is no longer discounted in dnode_move(),
1610 * so the dnode cannot be moved until after we release the hold.
1611 * The membar_producer() ensures visibility of the decremented
1612 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1616 db
->db_dnode_handle
= NULL
;
1622 dbuf_gone
= arc_buf_evict(db
->db_buf
);
1625 mutex_exit(&db
->db_mtx
);
1628 * If this dbuf is referenced from an indirect dbuf,
1629 * decrement the ref count on the indirect dbuf.
1631 if (parent
&& parent
!= dndb
)
1632 dbuf_rele(parent
, db
);
1635 __attribute__((always_inline
))
1637 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
1638 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
1645 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1647 if (blkid
== DMU_SPILL_BLKID
) {
1648 mutex_enter(&dn
->dn_mtx
);
1649 if (dn
->dn_have_spill
&&
1650 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
1651 *bpp
= &dn
->dn_phys
->dn_spill
;
1654 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1655 *parentp
= dn
->dn_dbuf
;
1656 mutex_exit(&dn
->dn_mtx
);
1660 if (dn
->dn_phys
->dn_nlevels
== 0)
1663 nlevels
= dn
->dn_phys
->dn_nlevels
;
1665 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
1667 ASSERT3U(level
* epbs
, <, 64);
1668 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1669 if (level
>= nlevels
||
1670 (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
1671 /* the buffer has no parent yet */
1673 } else if (level
< nlevels
-1) {
1674 /* this block is referenced from an indirect block */
1677 err
= dbuf_hold_impl(dn
, level
+1, blkid
>> epbs
,
1678 fail_sparse
, NULL
, parentp
);
1681 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
1682 blkid
>> epbs
, fail_sparse
, NULL
,
1683 parentp
, dh
->dh_depth
+ 1);
1684 err
= __dbuf_hold_impl(dh
+ 1);
1688 err
= dbuf_read(*parentp
, NULL
,
1689 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
1691 dbuf_rele(*parentp
, NULL
);
1695 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
1696 (blkid
& ((1ULL << epbs
) - 1));
1699 /* the block is referenced from the dnode */
1700 ASSERT3U(level
, ==, nlevels
-1);
1701 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
1702 blkid
< dn
->dn_phys
->dn_nblkptr
);
1704 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
1705 *parentp
= dn
->dn_dbuf
;
1707 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
1712 static dmu_buf_impl_t
*
1713 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
1714 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
1716 objset_t
*os
= dn
->dn_objset
;
1717 dmu_buf_impl_t
*db
, *odb
;
1719 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1720 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
1722 db
= kmem_cache_alloc(dbuf_cache
, KM_PUSHPAGE
);
1725 db
->db
.db_object
= dn
->dn_object
;
1726 db
->db_level
= level
;
1727 db
->db_blkid
= blkid
;
1728 db
->db_last_dirty
= NULL
;
1729 db
->db_dirtycnt
= 0;
1730 db
->db_dnode_handle
= dn
->dn_handle
;
1731 db
->db_parent
= parent
;
1732 db
->db_blkptr
= blkptr
;
1734 db
->db_user_ptr
= NULL
;
1735 db
->db_user_data_ptr_ptr
= NULL
;
1736 db
->db_evict_func
= NULL
;
1737 db
->db_immediate_evict
= 0;
1738 db
->db_freed_in_flight
= 0;
1740 if (blkid
== DMU_BONUS_BLKID
) {
1741 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
1742 db
->db
.db_size
= DN_MAX_BONUSLEN
-
1743 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
1744 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
1745 db
->db
.db_offset
= DMU_BONUS_BLKID
;
1746 db
->db_state
= DB_UNCACHED
;
1747 /* the bonus dbuf is not placed in the hash table */
1748 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1750 } else if (blkid
== DMU_SPILL_BLKID
) {
1751 db
->db
.db_size
= (blkptr
!= NULL
) ?
1752 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
1753 db
->db
.db_offset
= 0;
1756 db
->db_level
? 1<<dn
->dn_indblkshift
: dn
->dn_datablksz
;
1757 db
->db
.db_size
= blocksize
;
1758 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
1762 * Hold the dn_dbufs_mtx while we get the new dbuf
1763 * in the hash table *and* added to the dbufs list.
1764 * This prevents a possible deadlock with someone
1765 * trying to look up this dbuf before its added to the
1768 mutex_enter(&dn
->dn_dbufs_mtx
);
1769 db
->db_state
= DB_EVICTING
;
1770 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
1771 /* someone else inserted it first */
1772 kmem_cache_free(dbuf_cache
, db
);
1773 mutex_exit(&dn
->dn_dbufs_mtx
);
1776 list_insert_head(&dn
->dn_dbufs
, db
);
1777 db
->db_state
= DB_UNCACHED
;
1778 mutex_exit(&dn
->dn_dbufs_mtx
);
1779 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1781 if (parent
&& parent
!= dn
->dn_dbuf
)
1782 dbuf_add_ref(parent
, db
);
1784 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1785 refcount_count(&dn
->dn_holds
) > 0);
1786 (void) refcount_add(&dn
->dn_holds
, db
);
1787 (void) atomic_inc_32_nv(&dn
->dn_dbufs_count
);
1789 dprintf_dbuf(db
, "db=%p\n", db
);
1795 dbuf_do_evict(void *private)
1797 arc_buf_t
*buf
= private;
1798 dmu_buf_impl_t
*db
= buf
->b_private
;
1800 if (!MUTEX_HELD(&db
->db_mtx
))
1801 mutex_enter(&db
->db_mtx
);
1803 ASSERT(refcount_is_zero(&db
->db_holds
));
1805 if (db
->db_state
!= DB_EVICTING
) {
1806 ASSERT(db
->db_state
== DB_CACHED
);
1811 mutex_exit(&db
->db_mtx
);
1818 dbuf_destroy(dmu_buf_impl_t
*db
)
1820 ASSERT(refcount_is_zero(&db
->db_holds
));
1822 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1824 * If this dbuf is still on the dn_dbufs list,
1825 * remove it from that list.
1827 if (db
->db_dnode_handle
!= NULL
) {
1832 mutex_enter(&dn
->dn_dbufs_mtx
);
1833 list_remove(&dn
->dn_dbufs
, db
);
1834 (void) atomic_dec_32_nv(&dn
->dn_dbufs_count
);
1835 mutex_exit(&dn
->dn_dbufs_mtx
);
1838 * Decrementing the dbuf count means that the hold
1839 * corresponding to the removed dbuf is no longer
1840 * discounted in dnode_move(), so the dnode cannot be
1841 * moved until after we release the hold.
1844 db
->db_dnode_handle
= NULL
;
1846 dbuf_hash_remove(db
);
1848 db
->db_parent
= NULL
;
1851 ASSERT(!list_link_active(&db
->db_link
));
1852 ASSERT(db
->db
.db_data
== NULL
);
1853 ASSERT(db
->db_hash_next
== NULL
);
1854 ASSERT(db
->db_blkptr
== NULL
);
1855 ASSERT(db
->db_data_pending
== NULL
);
1857 kmem_cache_free(dbuf_cache
, db
);
1858 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_OTHER
);
1862 dbuf_prefetch(dnode_t
*dn
, uint64_t blkid
)
1864 dmu_buf_impl_t
*db
= NULL
;
1865 blkptr_t
*bp
= NULL
;
1867 ASSERT(blkid
!= DMU_BONUS_BLKID
);
1868 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1870 if (dnode_block_freed(dn
, blkid
))
1873 /* dbuf_find() returns with db_mtx held */
1874 if ((db
= dbuf_find(dn
, 0, blkid
))) {
1876 * This dbuf is already in the cache. We assume that
1877 * it is already CACHED, or else about to be either
1880 mutex_exit(&db
->db_mtx
);
1884 if (dbuf_findbp(dn
, 0, blkid
, TRUE
, &db
, &bp
, NULL
) == 0) {
1885 if (bp
&& !BP_IS_HOLE(bp
)) {
1886 int priority
= dn
->dn_type
== DMU_OT_DDT_ZAP
?
1887 ZIO_PRIORITY_DDT_PREFETCH
: ZIO_PRIORITY_ASYNC_READ
;
1889 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
1890 uint32_t aflags
= ARC_NOWAIT
| ARC_PREFETCH
;
1893 SET_BOOKMARK(&zb
, ds
? ds
->ds_object
: DMU_META_OBJSET
,
1894 dn
->dn_object
, 0, blkid
);
1899 pbuf
= dn
->dn_objset
->os_phys_buf
;
1901 (void) dsl_read(NULL
, dn
->dn_objset
->os_spa
,
1902 bp
, pbuf
, NULL
, NULL
, priority
,
1903 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
1907 dbuf_rele(db
, NULL
);
1911 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
1914 * Returns with db_holds incremented, and db_mtx not held.
1915 * Note: dn_struct_rwlock must be held.
1918 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
1920 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
1921 dh
->dh_parent
= NULL
;
1923 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
1924 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
1925 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
1927 *(dh
->dh_dbp
) = NULL
;
1929 /* dbuf_find() returns with db_mtx held */
1930 dh
->dh_db
= dbuf_find(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
);
1932 if (dh
->dh_db
== NULL
) {
1935 ASSERT3P(dh
->dh_parent
, ==, NULL
);
1936 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
1937 dh
->dh_fail_sparse
, &dh
->dh_parent
,
1939 if (dh
->dh_fail_sparse
) {
1940 if (dh
->dh_err
== 0 && dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
1941 dh
->dh_err
= ENOENT
;
1944 dbuf_rele(dh
->dh_parent
, NULL
);
1945 return (dh
->dh_err
);
1948 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
1949 return (dh
->dh_err
);
1950 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
1951 dh
->dh_parent
, dh
->dh_bp
);
1954 if (dh
->dh_db
->db_buf
&& refcount_is_zero(&dh
->dh_db
->db_holds
)) {
1955 arc_buf_add_ref(dh
->dh_db
->db_buf
, dh
->dh_db
);
1956 if (dh
->dh_db
->db_buf
->b_data
== NULL
) {
1957 dbuf_clear(dh
->dh_db
);
1958 if (dh
->dh_parent
) {
1959 dbuf_rele(dh
->dh_parent
, NULL
);
1960 dh
->dh_parent
= NULL
;
1964 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
1967 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
1970 * If this buffer is currently syncing out, and we are are
1971 * still referencing it from db_data, we need to make a copy
1972 * of it in case we decide we want to dirty it again in this txg.
1974 if (dh
->dh_db
->db_level
== 0 &&
1975 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
1976 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
1977 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
1978 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
1980 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
) {
1981 dh
->dh_type
= DBUF_GET_BUFC_TYPE(dh
->dh_db
);
1983 dbuf_set_data(dh
->dh_db
,
1984 arc_buf_alloc(dh
->dh_dn
->dn_objset
->os_spa
,
1985 dh
->dh_db
->db
.db_size
, dh
->dh_db
, dh
->dh_type
));
1986 bcopy(dh
->dh_dr
->dt
.dl
.dr_data
->b_data
,
1987 dh
->dh_db
->db
.db_data
, dh
->dh_db
->db
.db_size
);
1991 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
1992 dbuf_update_data(dh
->dh_db
);
1993 DBUF_VERIFY(dh
->dh_db
);
1994 mutex_exit(&dh
->dh_db
->db_mtx
);
1996 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1998 dbuf_rele(dh
->dh_parent
, NULL
);
2000 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
2001 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
2002 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
2003 *(dh
->dh_dbp
) = dh
->dh_db
;
2009 * The following code preserves the recursive function dbuf_hold_impl()
2010 * but moves the local variables AND function arguments to the heap to
2011 * minimize the stack frame size. Enough space is initially allocated
2012 * on the stack for 20 levels of recursion.
2015 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
, int fail_sparse
,
2016 void *tag
, dmu_buf_impl_t
**dbp
)
2018 struct dbuf_hold_impl_data
*dh
;
2021 dh
= kmem_zalloc(sizeof(struct dbuf_hold_impl_data
) *
2022 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_PUSHPAGE
);
2023 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
, tag
, dbp
, 0);
2025 error
= __dbuf_hold_impl(dh
);
2027 kmem_free(dh
, sizeof(struct dbuf_hold_impl_data
) *
2028 DBUF_HOLD_IMPL_MAX_DEPTH
);
2034 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
2035 dnode_t
*dn
, uint8_t level
, uint64_t blkid
, int fail_sparse
,
2036 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
2039 dh
->dh_level
= level
;
2040 dh
->dh_blkid
= blkid
;
2041 dh
->dh_fail_sparse
= fail_sparse
;
2044 dh
->dh_depth
= depth
;
2048 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
2051 int err
= dbuf_hold_impl(dn
, 0, blkid
, FALSE
, tag
, &db
);
2052 return (err
? NULL
: db
);
2056 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
2059 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, tag
, &db
);
2060 return (err
? NULL
: db
);
2064 dbuf_create_bonus(dnode_t
*dn
)
2066 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
2068 ASSERT(dn
->dn_bonus
== NULL
);
2069 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
2073 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
2075 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2078 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2081 blksz
= SPA_MINBLOCKSIZE
;
2082 if (blksz
> SPA_MAXBLOCKSIZE
)
2083 blksz
= SPA_MAXBLOCKSIZE
;
2085 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
2089 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2090 dbuf_new_size(db
, blksz
, tx
);
2091 rw_exit(&dn
->dn_struct_rwlock
);
2098 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
2100 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
2103 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2105 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
2107 VERIFY(refcount_add(&db
->db_holds
, tag
) > 1);
2111 * If you call dbuf_rele() you had better not be referencing the dnode handle
2112 * unless you have some other direct or indirect hold on the dnode. (An indirect
2113 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2114 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2115 * dnode's parent dbuf evicting its dnode handles.
2117 #pragma weak dmu_buf_rele = dbuf_rele
2119 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
2121 mutex_enter(&db
->db_mtx
);
2122 dbuf_rele_and_unlock(db
, tag
);
2126 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2127 * db_dirtycnt and db_holds to be updated atomically.
2130 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
2134 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2138 * Remove the reference to the dbuf before removing its hold on the
2139 * dnode so we can guarantee in dnode_move() that a referenced bonus
2140 * buffer has a corresponding dnode hold.
2142 holds
= refcount_remove(&db
->db_holds
, tag
);
2146 * We can't freeze indirects if there is a possibility that they
2147 * may be modified in the current syncing context.
2149 if (db
->db_buf
&& holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0))
2150 arc_buf_freeze(db
->db_buf
);
2152 if (holds
== db
->db_dirtycnt
&&
2153 db
->db_level
== 0 && db
->db_immediate_evict
)
2154 dbuf_evict_user(db
);
2157 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2158 mutex_exit(&db
->db_mtx
);
2161 * If the dnode moves here, we cannot cross this barrier
2162 * until the move completes.
2165 (void) atomic_dec_32_nv(&DB_DNODE(db
)->dn_dbufs_count
);
2168 * The bonus buffer's dnode hold is no longer discounted
2169 * in dnode_move(). The dnode cannot move until after
2172 dnode_rele(DB_DNODE(db
), db
);
2173 } else if (db
->db_buf
== NULL
) {
2175 * This is a special case: we never associated this
2176 * dbuf with any data allocated from the ARC.
2178 ASSERT(db
->db_state
== DB_UNCACHED
||
2179 db
->db_state
== DB_NOFILL
);
2181 } else if (arc_released(db
->db_buf
)) {
2182 arc_buf_t
*buf
= db
->db_buf
;
2184 * This dbuf has anonymous data associated with it.
2186 dbuf_set_data(db
, NULL
);
2187 VERIFY(arc_buf_remove_ref(buf
, db
) == 1);
2190 VERIFY(arc_buf_remove_ref(db
->db_buf
, db
) == 0);
2191 if (!DBUF_IS_CACHEABLE(db
))
2194 mutex_exit(&db
->db_mtx
);
2197 mutex_exit(&db
->db_mtx
);
2201 #pragma weak dmu_buf_refcount = dbuf_refcount
2203 dbuf_refcount(dmu_buf_impl_t
*db
)
2205 return (refcount_count(&db
->db_holds
));
2209 dmu_buf_set_user(dmu_buf_t
*db_fake
, void *user_ptr
, void *user_data_ptr_ptr
,
2210 dmu_buf_evict_func_t
*evict_func
)
2212 return (dmu_buf_update_user(db_fake
, NULL
, user_ptr
,
2213 user_data_ptr_ptr
, evict_func
));
2217 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, void *user_ptr
, void *user_data_ptr_ptr
,
2218 dmu_buf_evict_func_t
*evict_func
)
2220 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2222 db
->db_immediate_evict
= TRUE
;
2223 return (dmu_buf_update_user(db_fake
, NULL
, user_ptr
,
2224 user_data_ptr_ptr
, evict_func
));
2228 dmu_buf_update_user(dmu_buf_t
*db_fake
, void *old_user_ptr
, void *user_ptr
,
2229 void *user_data_ptr_ptr
, dmu_buf_evict_func_t
*evict_func
)
2231 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2232 ASSERT(db
->db_level
== 0);
2234 ASSERT((user_ptr
== NULL
) == (evict_func
== NULL
));
2236 mutex_enter(&db
->db_mtx
);
2238 if (db
->db_user_ptr
== old_user_ptr
) {
2239 db
->db_user_ptr
= user_ptr
;
2240 db
->db_user_data_ptr_ptr
= user_data_ptr_ptr
;
2241 db
->db_evict_func
= evict_func
;
2243 dbuf_update_data(db
);
2245 old_user_ptr
= db
->db_user_ptr
;
2248 mutex_exit(&db
->db_mtx
);
2249 return (old_user_ptr
);
2253 dmu_buf_get_user(dmu_buf_t
*db_fake
)
2255 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2256 ASSERT(!refcount_is_zero(&db
->db_holds
));
2258 return (db
->db_user_ptr
);
2262 dmu_buf_freeable(dmu_buf_t
*dbuf
)
2264 boolean_t res
= B_FALSE
;
2265 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2268 res
= dsl_dataset_block_freeable(db
->db_objset
->os_dsl_dataset
,
2269 db
->db_blkptr
, db
->db_blkptr
->blk_birth
);
2275 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
2277 /* ASSERT(dmu_tx_is_syncing(tx) */
2278 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2280 if (db
->db_blkptr
!= NULL
)
2283 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2284 db
->db_blkptr
= &dn
->dn_phys
->dn_spill
;
2285 BP_ZERO(db
->db_blkptr
);
2288 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
2290 * This buffer was allocated at a time when there was
2291 * no available blkptrs from the dnode, or it was
2292 * inappropriate to hook it in (i.e., nlevels mis-match).
2294 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
2295 ASSERT(db
->db_parent
== NULL
);
2296 db
->db_parent
= dn
->dn_dbuf
;
2297 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
2300 dmu_buf_impl_t
*parent
= db
->db_parent
;
2301 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2303 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
2304 if (parent
== NULL
) {
2305 mutex_exit(&db
->db_mtx
);
2306 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2307 (void) dbuf_hold_impl(dn
, db
->db_level
+1,
2308 db
->db_blkid
>> epbs
, FALSE
, db
, &parent
);
2309 rw_exit(&dn
->dn_struct_rwlock
);
2310 mutex_enter(&db
->db_mtx
);
2311 db
->db_parent
= parent
;
2313 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
2314 (db
->db_blkid
& ((1ULL << epbs
) - 1));
2319 /* dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2320 * is critical the we not allow the compiler to inline this function in to
2321 * dbuf_sync_list() thereby drastically bloating the stack usage.
2323 noinline
static void
2324 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2326 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2330 ASSERT(dmu_tx_is_syncing(tx
));
2332 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2334 mutex_enter(&db
->db_mtx
);
2336 ASSERT(db
->db_level
> 0);
2339 if (db
->db_buf
== NULL
) {
2340 mutex_exit(&db
->db_mtx
);
2341 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
2342 mutex_enter(&db
->db_mtx
);
2344 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
2345 ASSERT(db
->db_buf
!= NULL
);
2349 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2350 dbuf_check_blkptr(dn
, db
);
2353 db
->db_data_pending
= dr
;
2355 mutex_exit(&db
->db_mtx
);
2356 dbuf_write(dr
, db
->db_buf
, tx
);
2359 mutex_enter(&dr
->dt
.di
.dr_mtx
);
2360 dbuf_sync_list(&dr
->dt
.di
.dr_children
, tx
);
2361 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2362 mutex_exit(&dr
->dt
.di
.dr_mtx
);
2366 /* dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
2367 * critical the we not allow the compiler to inline this function in to
2368 * dbuf_sync_list() thereby drastically bloating the stack usage.
2370 noinline
static void
2371 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
2373 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
2374 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2377 uint64_t txg
= tx
->tx_txg
;
2379 ASSERT(dmu_tx_is_syncing(tx
));
2381 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
2383 mutex_enter(&db
->db_mtx
);
2385 * To be synced, we must be dirtied. But we
2386 * might have been freed after the dirty.
2388 if (db
->db_state
== DB_UNCACHED
) {
2389 /* This buffer has been freed since it was dirtied */
2390 ASSERT(db
->db
.db_data
== NULL
);
2391 } else if (db
->db_state
== DB_FILL
) {
2392 /* This buffer was freed and is now being re-filled */
2393 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
2395 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
2402 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2403 mutex_enter(&dn
->dn_mtx
);
2404 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
2405 mutex_exit(&dn
->dn_mtx
);
2409 * If this is a bonus buffer, simply copy the bonus data into the
2410 * dnode. It will be written out when the dnode is synced (and it
2411 * will be synced, since it must have been dirty for dbuf_sync to
2414 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2415 dbuf_dirty_record_t
**drp
;
2417 ASSERT(*datap
!= NULL
);
2418 ASSERT3U(db
->db_level
, ==, 0);
2419 ASSERT3U(dn
->dn_phys
->dn_bonuslen
, <=, DN_MAX_BONUSLEN
);
2420 bcopy(*datap
, DN_BONUS(dn
->dn_phys
), dn
->dn_phys
->dn_bonuslen
);
2423 if (*datap
!= db
->db
.db_data
) {
2424 zio_buf_free(*datap
, DN_MAX_BONUSLEN
);
2425 arc_space_return(DN_MAX_BONUSLEN
, ARC_SPACE_OTHER
);
2427 db
->db_data_pending
= NULL
;
2428 drp
= &db
->db_last_dirty
;
2430 drp
= &(*drp
)->dr_next
;
2431 ASSERT(dr
->dr_next
== NULL
);
2432 ASSERT(dr
->dr_dbuf
== db
);
2434 if (dr
->dr_dbuf
->db_level
!= 0) {
2435 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2436 list_destroy(&dr
->dt
.di
.dr_children
);
2438 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2439 ASSERT(db
->db_dirtycnt
> 0);
2440 db
->db_dirtycnt
-= 1;
2441 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
2448 * This function may have dropped the db_mtx lock allowing a dmu_sync
2449 * operation to sneak in. As a result, we need to ensure that we
2450 * don't check the dr_override_state until we have returned from
2451 * dbuf_check_blkptr.
2453 dbuf_check_blkptr(dn
, db
);
2456 * If this buffer is in the middle of an immediate write,
2457 * wait for the synchronous IO to complete.
2459 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
2460 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
2461 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2462 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
2465 if (db
->db_state
!= DB_NOFILL
&&
2466 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
2467 refcount_count(&db
->db_holds
) > 1 &&
2468 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
2469 *datap
== db
->db_buf
) {
2471 * If this buffer is currently "in use" (i.e., there
2472 * are active holds and db_data still references it),
2473 * then make a copy before we start the write so that
2474 * any modifications from the open txg will not leak
2477 * NOTE: this copy does not need to be made for
2478 * objects only modified in the syncing context (e.g.
2479 * DNONE_DNODE blocks).
2481 int blksz
= arc_buf_size(*datap
);
2482 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
2483 *datap
= arc_buf_alloc(os
->os_spa
, blksz
, db
, type
);
2484 bcopy(db
->db
.db_data
, (*datap
)->b_data
, blksz
);
2486 db
->db_data_pending
= dr
;
2488 mutex_exit(&db
->db_mtx
);
2490 dbuf_write(dr
, *datap
, tx
);
2492 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2493 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2494 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
2498 * Although zio_nowait() does not "wait for an IO", it does
2499 * initiate the IO. If this is an empty write it seems plausible
2500 * that the IO could actually be completed before the nowait
2501 * returns. We need to DB_DNODE_EXIT() first in case
2502 * zio_nowait() invalidates the dbuf.
2505 zio_nowait(dr
->dr_zio
);
2510 dbuf_sync_list(list_t
*list
, dmu_tx_t
*tx
)
2512 dbuf_dirty_record_t
*dr
;
2514 while ((dr
= list_head(list
))) {
2515 if (dr
->dr_zio
!= NULL
) {
2517 * If we find an already initialized zio then we
2518 * are processing the meta-dnode, and we have finished.
2519 * The dbufs for all dnodes are put back on the list
2520 * during processing, so that we can zio_wait()
2521 * these IOs after initiating all child IOs.
2523 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
2524 DMU_META_DNODE_OBJECT
);
2527 list_remove(list
, dr
);
2528 if (dr
->dr_dbuf
->db_level
> 0)
2529 dbuf_sync_indirect(dr
, tx
);
2531 dbuf_sync_leaf(dr
, tx
);
2537 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
2539 dmu_buf_impl_t
*db
= vdb
;
2541 blkptr_t
*bp
= zio
->io_bp
;
2542 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2543 spa_t
*spa
= zio
->io_spa
;
2548 ASSERT(db
->db_blkptr
== bp
);
2552 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
2553 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
2554 zio
->io_prev_space_delta
= delta
;
2556 if (BP_IS_HOLE(bp
)) {
2557 ASSERT(bp
->blk_fill
== 0);
2562 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
2563 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
2564 (db
->db_blkid
== DMU_SPILL_BLKID
&&
2565 BP_GET_TYPE(bp
) == dn
->dn_bonustype
));
2566 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
2568 mutex_enter(&db
->db_mtx
);
2571 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2572 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
2573 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
2574 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
2578 if (db
->db_level
== 0) {
2579 mutex_enter(&dn
->dn_mtx
);
2580 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
2581 db
->db_blkid
!= DMU_SPILL_BLKID
)
2582 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
2583 mutex_exit(&dn
->dn_mtx
);
2585 if (dn
->dn_type
== DMU_OT_DNODE
) {
2586 dnode_phys_t
*dnp
= db
->db
.db_data
;
2587 for (i
= db
->db
.db_size
>> DNODE_SHIFT
; i
> 0;
2589 if (dnp
->dn_type
!= DMU_OT_NONE
)
2596 blkptr_t
*ibp
= db
->db
.db_data
;
2597 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2598 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
2599 if (BP_IS_HOLE(ibp
))
2601 fill
+= ibp
->blk_fill
;
2606 bp
->blk_fill
= fill
;
2608 mutex_exit(&db
->db_mtx
);
2613 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
2615 dmu_buf_impl_t
*db
= vdb
;
2616 blkptr_t
*bp
= zio
->io_bp
;
2617 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2618 uint64_t txg
= zio
->io_txg
;
2619 dbuf_dirty_record_t
**drp
, *dr
;
2621 ASSERT3U(zio
->io_error
, ==, 0);
2622 ASSERT(db
->db_blkptr
== bp
);
2624 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
2625 ASSERT(BP_EQUAL(bp
, bp_orig
));
2631 DB_GET_OBJSET(&os
, db
);
2632 ds
= os
->os_dsl_dataset
;
2635 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
2636 dsl_dataset_block_born(ds
, bp
, tx
);
2639 mutex_enter(&db
->db_mtx
);
2643 drp
= &db
->db_last_dirty
;
2644 while ((dr
= *drp
) != db
->db_data_pending
)
2646 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2647 ASSERT(dr
->dr_txg
== txg
);
2648 ASSERT(dr
->dr_dbuf
== db
);
2649 ASSERT(dr
->dr_next
== NULL
);
2653 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
2658 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
2659 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
2660 db
->db_blkptr
== &dn
->dn_phys
->dn_spill
);
2665 if (db
->db_level
== 0) {
2666 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2667 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
2668 if (db
->db_state
!= DB_NOFILL
) {
2669 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2670 VERIFY(arc_buf_remove_ref(dr
->dt
.dl
.dr_data
,
2672 else if (!arc_released(db
->db_buf
))
2673 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
2680 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
2681 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
2682 if (!BP_IS_HOLE(db
->db_blkptr
)) {
2683 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
2685 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
2687 ASSERT3U(dn
->dn_phys
->dn_maxblkid
2688 >> (db
->db_level
* epbs
), >=, db
->db_blkid
);
2689 arc_set_callback(db
->db_buf
, dbuf_do_evict
, db
);
2692 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
2693 list_destroy(&dr
->dt
.di
.dr_children
);
2695 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2697 cv_broadcast(&db
->db_changed
);
2698 ASSERT(db
->db_dirtycnt
> 0);
2699 db
->db_dirtycnt
-= 1;
2700 db
->db_data_pending
= NULL
;
2701 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
2705 dbuf_write_nofill_ready(zio_t
*zio
)
2707 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
2711 dbuf_write_nofill_done(zio_t
*zio
)
2713 dbuf_write_done(zio
, NULL
, zio
->io_private
);
2717 dbuf_write_override_ready(zio_t
*zio
)
2719 dbuf_dirty_record_t
*dr
= zio
->io_private
;
2720 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2722 dbuf_write_ready(zio
, NULL
, db
);
2726 dbuf_write_override_done(zio_t
*zio
)
2728 dbuf_dirty_record_t
*dr
= zio
->io_private
;
2729 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2730 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
2732 mutex_enter(&db
->db_mtx
);
2733 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
2734 if (!BP_IS_HOLE(obp
))
2735 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
2736 arc_release(dr
->dt
.dl
.dr_data
, db
);
2738 mutex_exit(&db
->db_mtx
);
2740 dbuf_write_done(zio
, NULL
, db
);
2744 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
2746 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
2749 dmu_buf_impl_t
*parent
= db
->db_parent
;
2750 uint64_t txg
= tx
->tx_txg
;
2760 if (db
->db_state
!= DB_NOFILL
) {
2761 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
2763 * Private object buffers are released here rather
2764 * than in dbuf_dirty() since they are only modified
2765 * in the syncing context and we don't want the
2766 * overhead of making multiple copies of the data.
2768 if (BP_IS_HOLE(db
->db_blkptr
)) {
2771 dbuf_release_bp(db
);
2776 if (parent
!= dn
->dn_dbuf
) {
2777 ASSERT(parent
&& parent
->db_data_pending
);
2778 ASSERT(db
->db_level
== parent
->db_level
-1);
2779 ASSERT(arc_released(parent
->db_buf
));
2780 zio
= parent
->db_data_pending
->dr_zio
;
2782 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
2783 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
2784 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
2785 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
2786 ASSERT3P(db
->db_blkptr
, ==,
2787 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
2791 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
2792 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
2795 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
2796 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
2797 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
2799 if (db
->db_blkid
== DMU_SPILL_BLKID
)
2801 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
2803 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
2806 if (db
->db_level
== 0 && dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
2807 ASSERT(db
->db_state
!= DB_NOFILL
);
2808 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
2809 db
->db_blkptr
, data
->b_data
, arc_buf_size(data
), &zp
,
2810 dbuf_write_override_ready
, dbuf_write_override_done
, dr
,
2811 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
2812 mutex_enter(&db
->db_mtx
);
2813 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
2814 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
2815 dr
->dt
.dl
.dr_copies
);
2816 mutex_exit(&db
->db_mtx
);
2817 } else if (db
->db_state
== DB_NOFILL
) {
2818 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
);
2819 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
2820 db
->db_blkptr
, NULL
, db
->db
.db_size
, &zp
,
2821 dbuf_write_nofill_ready
, dbuf_write_nofill_done
, db
,
2822 ZIO_PRIORITY_ASYNC_WRITE
,
2823 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
2825 ASSERT(arc_released(data
));
2826 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
2827 db
->db_blkptr
, data
, DBUF_IS_L2CACHEABLE(db
), &zp
,
2828 dbuf_write_ready
, dbuf_write_done
, db
,
2829 ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
2833 #if defined(_KERNEL) && defined(HAVE_SPL)
2834 EXPORT_SYMBOL(dbuf_find
);
2835 EXPORT_SYMBOL(dbuf_is_metadata
);
2836 EXPORT_SYMBOL(dbuf_evict
);
2837 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
2838 EXPORT_SYMBOL(dbuf_whichblock
);
2839 EXPORT_SYMBOL(dbuf_read
);
2840 EXPORT_SYMBOL(dbuf_unoverride
);
2841 EXPORT_SYMBOL(dbuf_free_range
);
2842 EXPORT_SYMBOL(dbuf_new_size
);
2843 EXPORT_SYMBOL(dbuf_release_bp
);
2844 EXPORT_SYMBOL(dbuf_dirty
);
2845 EXPORT_SYMBOL(dmu_buf_will_dirty
);
2846 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
2847 EXPORT_SYMBOL(dmu_buf_will_fill
);
2848 EXPORT_SYMBOL(dmu_buf_fill_done
);
2849 EXPORT_SYMBOL(dmu_buf_rele
);
2850 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
2851 EXPORT_SYMBOL(dbuf_clear
);
2852 EXPORT_SYMBOL(dbuf_prefetch
);
2853 EXPORT_SYMBOL(dbuf_hold_impl
);
2854 EXPORT_SYMBOL(dbuf_hold
);
2855 EXPORT_SYMBOL(dbuf_hold_level
);
2856 EXPORT_SYMBOL(dbuf_create_bonus
);
2857 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
2858 EXPORT_SYMBOL(dbuf_rm_spill
);
2859 EXPORT_SYMBOL(dbuf_add_ref
);
2860 EXPORT_SYMBOL(dbuf_rele
);
2861 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
2862 EXPORT_SYMBOL(dbuf_refcount
);
2863 EXPORT_SYMBOL(dbuf_sync_list
);
2864 EXPORT_SYMBOL(dmu_buf_set_user
);
2865 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
2866 EXPORT_SYMBOL(dmu_buf_update_user
);
2867 EXPORT_SYMBOL(dmu_buf_get_user
);
2868 EXPORT_SYMBOL(dmu_buf_freeable
);