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 (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
26 * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
28 * Copyright (c) 2019 Datto Inc.
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
35 #include <sys/dnode.h>
36 #include <sys/zfs_context.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/dmu_traverse.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/dsl_dir.h>
41 #include <sys/dsl_pool.h>
42 #include <sys/dsl_synctask.h>
43 #include <sys/dsl_prop.h>
44 #include <sys/dmu_zfetch.h>
45 #include <sys/zfs_ioctl.h>
47 #include <sys/zio_checksum.h>
48 #include <sys/zio_compress.h>
50 #include <sys/zfeature.h>
52 #include <sys/trace_dmu.h>
53 #include <sys/zfs_rlock.h>
55 #include <sys/vmsystm.h>
56 #include <sys/zfs_znode.h>
60 * Enable/disable nopwrite feature.
62 int zfs_nopwrite_enabled
= 1;
65 * Tunable to control percentage of dirtied L1 blocks from frees allowed into
66 * one TXG. After this threshold is crossed, additional dirty blocks from frees
67 * will wait until the next TXG.
68 * A value of zero will disable this throttle.
70 unsigned long zfs_per_txg_dirty_frees_percent
= 5;
73 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
75 int zfs_dmu_offset_next_sync
= 0;
78 * This can be used for testing, to ensure that certain actions happen
79 * while in the middle of a remap (which might otherwise complete too
80 * quickly). Used by ztest(8).
82 int zfs_object_remap_one_indirect_delay_ms
= 0;
84 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
85 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "unallocated" },
86 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "object directory" },
87 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "object array" },
88 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "packed nvlist" },
89 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "packed nvlist size" },
90 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj" },
91 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj header" },
92 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA space map header" },
93 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA space map" },
94 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, TRUE
, "ZIL intent log" },
95 {DMU_BSWAP_DNODE
, TRUE
, FALSE
, TRUE
, "DMU dnode" },
96 {DMU_BSWAP_OBJSET
, TRUE
, TRUE
, FALSE
, "DMU objset" },
97 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL directory" },
98 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL directory child map"},
99 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dataset snap map" },
100 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL props" },
101 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL dataset" },
102 {DMU_BSWAP_ZNODE
, TRUE
, FALSE
, FALSE
, "ZFS znode" },
103 {DMU_BSWAP_OLDACL
, TRUE
, FALSE
, TRUE
, "ZFS V0 ACL" },
104 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "ZFS plain file" },
105 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS directory" },
106 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "ZFS master node" },
107 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS delete queue" },
108 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "zvol object" },
109 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "zvol prop" },
110 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "other uint8[]" },
111 {DMU_BSWAP_UINT64
, FALSE
, FALSE
, TRUE
, "other uint64[]" },
112 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "other ZAP" },
113 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "persistent error log" },
114 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "SPA history" },
115 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA history offsets" },
116 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "Pool properties" },
117 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL permissions" },
118 {DMU_BSWAP_ACL
, TRUE
, FALSE
, TRUE
, "ZFS ACL" },
119 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "ZFS SYSACL" },
120 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "FUID table" },
121 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "FUID table size" },
122 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dataset next clones"},
123 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "scan work queue" },
124 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS user/group/project used" },
125 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS user/group/project quota"},
126 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "snapshot refcount tags"},
127 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "DDT ZAP algorithm" },
128 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "DDT statistics" },
129 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "System attributes" },
130 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA master node" },
131 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA attr registration" },
132 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA attr layouts" },
133 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "scan translations" },
134 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "deduplicated block" },
135 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL deadlist map" },
136 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL deadlist map hdr" },
137 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dir clones" },
138 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj subobj" }
141 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
142 { byteswap_uint8_array
, "uint8" },
143 { byteswap_uint16_array
, "uint16" },
144 { byteswap_uint32_array
, "uint32" },
145 { byteswap_uint64_array
, "uint64" },
146 { zap_byteswap
, "zap" },
147 { dnode_buf_byteswap
, "dnode" },
148 { dmu_objset_byteswap
, "objset" },
149 { zfs_znode_byteswap
, "znode" },
150 { zfs_oldacl_byteswap
, "oldacl" },
151 { zfs_acl_byteswap
, "acl" }
155 dmu_buf_hold_noread_by_dnode(dnode_t
*dn
, uint64_t offset
,
156 void *tag
, dmu_buf_t
**dbp
)
161 blkid
= dbuf_whichblock(dn
, 0, offset
);
162 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
163 db
= dbuf_hold(dn
, blkid
, tag
);
164 rw_exit(&dn
->dn_struct_rwlock
);
168 return (SET_ERROR(EIO
));
175 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
176 void *tag
, dmu_buf_t
**dbp
)
183 err
= dnode_hold(os
, object
, FTAG
, &dn
);
186 blkid
= dbuf_whichblock(dn
, 0, offset
);
187 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
188 db
= dbuf_hold(dn
, blkid
, tag
);
189 rw_exit(&dn
->dn_struct_rwlock
);
190 dnode_rele(dn
, FTAG
);
194 return (SET_ERROR(EIO
));
202 dmu_buf_hold_by_dnode(dnode_t
*dn
, uint64_t offset
,
203 void *tag
, dmu_buf_t
**dbp
, int flags
)
206 int db_flags
= DB_RF_CANFAIL
;
208 if (flags
& DMU_READ_NO_PREFETCH
)
209 db_flags
|= DB_RF_NOPREFETCH
;
210 if (flags
& DMU_READ_NO_DECRYPT
)
211 db_flags
|= DB_RF_NO_DECRYPT
;
213 err
= dmu_buf_hold_noread_by_dnode(dn
, offset
, tag
, dbp
);
215 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
216 err
= dbuf_read(db
, NULL
, db_flags
);
227 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
228 void *tag
, dmu_buf_t
**dbp
, int flags
)
231 int db_flags
= DB_RF_CANFAIL
;
233 if (flags
& DMU_READ_NO_PREFETCH
)
234 db_flags
|= DB_RF_NOPREFETCH
;
235 if (flags
& DMU_READ_NO_DECRYPT
)
236 db_flags
|= DB_RF_NO_DECRYPT
;
238 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
240 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
241 err
= dbuf_read(db
, NULL
, db_flags
);
254 return (DN_OLD_MAX_BONUSLEN
);
258 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
260 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
267 if (dn
->dn_bonus
!= db
) {
268 error
= SET_ERROR(EINVAL
);
269 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
270 error
= SET_ERROR(EINVAL
);
272 dnode_setbonuslen(dn
, newsize
, tx
);
281 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
283 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
290 if (!DMU_OT_IS_VALID(type
)) {
291 error
= SET_ERROR(EINVAL
);
292 } else if (dn
->dn_bonus
!= db
) {
293 error
= SET_ERROR(EINVAL
);
295 dnode_setbonus_type(dn
, type
, tx
);
304 dmu_get_bonustype(dmu_buf_t
*db_fake
)
306 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
308 dmu_object_type_t type
;
312 type
= dn
->dn_bonustype
;
319 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
324 error
= dnode_hold(os
, object
, FTAG
, &dn
);
325 dbuf_rm_spill(dn
, tx
);
326 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
327 dnode_rm_spill(dn
, tx
);
328 rw_exit(&dn
->dn_struct_rwlock
);
329 dnode_rele(dn
, FTAG
);
334 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
335 * has not yet been allocated a new bonus dbuf a will be allocated.
336 * Returns ENOENT, EIO, or 0.
338 int dmu_bonus_hold_by_dnode(dnode_t
*dn
, void *tag
, dmu_buf_t
**dbp
,
343 uint32_t db_flags
= DB_RF_MUST_SUCCEED
;
345 if (flags
& DMU_READ_NO_PREFETCH
)
346 db_flags
|= DB_RF_NOPREFETCH
;
347 if (flags
& DMU_READ_NO_DECRYPT
)
348 db_flags
|= DB_RF_NO_DECRYPT
;
350 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
351 if (dn
->dn_bonus
== NULL
) {
352 rw_exit(&dn
->dn_struct_rwlock
);
353 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
354 if (dn
->dn_bonus
== NULL
)
355 dbuf_create_bonus(dn
);
359 /* as long as the bonus buf is held, the dnode will be held */
360 if (zfs_refcount_add(&db
->db_holds
, tag
) == 1) {
361 VERIFY(dnode_add_ref(dn
, db
));
362 atomic_inc_32(&dn
->dn_dbufs_count
);
366 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
367 * hold and incrementing the dbuf count to ensure that dnode_move() sees
368 * a dnode hold for every dbuf.
370 rw_exit(&dn
->dn_struct_rwlock
);
372 error
= dbuf_read(db
, NULL
, db_flags
);
374 dnode_evict_bonus(dn
);
385 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
390 error
= dnode_hold(os
, object
, FTAG
, &dn
);
394 error
= dmu_bonus_hold_by_dnode(dn
, tag
, dbp
, DMU_READ_NO_PREFETCH
);
395 dnode_rele(dn
, FTAG
);
401 * returns ENOENT, EIO, or 0.
403 * This interface will allocate a blank spill dbuf when a spill blk
404 * doesn't already exist on the dnode.
406 * if you only want to find an already existing spill db, then
407 * dmu_spill_hold_existing() should be used.
410 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
412 dmu_buf_impl_t
*db
= NULL
;
415 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
416 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
418 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
420 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
421 rw_exit(&dn
->dn_struct_rwlock
);
425 return (SET_ERROR(EIO
));
427 err
= dbuf_read(db
, NULL
, flags
);
438 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
440 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
447 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
448 err
= SET_ERROR(EINVAL
);
450 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
452 if (!dn
->dn_have_spill
) {
453 err
= SET_ERROR(ENOENT
);
455 err
= dmu_spill_hold_by_dnode(dn
,
456 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
459 rw_exit(&dn
->dn_struct_rwlock
);
467 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, uint32_t flags
, void *tag
,
470 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
473 uint32_t db_flags
= DB_RF_CANFAIL
;
475 if (flags
& DMU_READ_NO_DECRYPT
)
476 db_flags
|= DB_RF_NO_DECRYPT
;
480 err
= dmu_spill_hold_by_dnode(dn
, db_flags
, tag
, dbp
);
487 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
488 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
489 * and can induce severe lock contention when writing to several files
490 * whose dnodes are in the same block.
493 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
494 boolean_t read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
497 uint64_t blkid
, nblks
, i
;
502 ASSERT(length
<= DMU_MAX_ACCESS
);
505 * Note: We directly notify the prefetch code of this read, so that
506 * we can tell it about the multi-block read. dbuf_read() only knows
507 * about the one block it is accessing.
509 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
|
512 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
513 if (dn
->dn_datablkshift
) {
514 int blkshift
= dn
->dn_datablkshift
;
515 nblks
= (P2ROUNDUP(offset
+ length
, 1ULL << blkshift
) -
516 P2ALIGN(offset
, 1ULL << blkshift
)) >> blkshift
;
518 if (offset
+ length
> dn
->dn_datablksz
) {
519 zfs_panic_recover("zfs: accessing past end of object "
520 "%llx/%llx (size=%u access=%llu+%llu)",
521 (longlong_t
)dn
->dn_objset
->
522 os_dsl_dataset
->ds_object
,
523 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
524 (longlong_t
)offset
, (longlong_t
)length
);
525 rw_exit(&dn
->dn_struct_rwlock
);
526 return (SET_ERROR(EIO
));
530 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_SLEEP
);
532 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
533 blkid
= dbuf_whichblock(dn
, 0, offset
);
534 for (i
= 0; i
< nblks
; i
++) {
535 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+ i
, tag
);
537 rw_exit(&dn
->dn_struct_rwlock
);
538 dmu_buf_rele_array(dbp
, nblks
, tag
);
540 return (SET_ERROR(EIO
));
543 /* initiate async i/o */
545 (void) dbuf_read(db
, zio
, dbuf_flags
);
549 if ((flags
& DMU_READ_NO_PREFETCH
) == 0 &&
550 DNODE_META_IS_CACHEABLE(dn
) && length
<= zfetch_array_rd_sz
) {
551 dmu_zfetch(&dn
->dn_zfetch
, blkid
, nblks
,
552 read
&& DNODE_IS_CACHEABLE(dn
));
554 rw_exit(&dn
->dn_struct_rwlock
);
556 /* wait for async i/o */
559 dmu_buf_rele_array(dbp
, nblks
, tag
);
563 /* wait for other io to complete */
565 for (i
= 0; i
< nblks
; i
++) {
566 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
567 mutex_enter(&db
->db_mtx
);
568 while (db
->db_state
== DB_READ
||
569 db
->db_state
== DB_FILL
)
570 cv_wait(&db
->db_changed
, &db
->db_mtx
);
571 if (db
->db_state
== DB_UNCACHED
)
572 err
= SET_ERROR(EIO
);
573 mutex_exit(&db
->db_mtx
);
575 dmu_buf_rele_array(dbp
, nblks
, tag
);
587 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
588 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
593 err
= dnode_hold(os
, object
, FTAG
, &dn
);
597 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
598 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
600 dnode_rele(dn
, FTAG
);
606 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
607 uint64_t length
, boolean_t read
, void *tag
, int *numbufsp
,
610 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
616 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
617 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
624 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
627 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
632 for (i
= 0; i
< numbufs
; i
++) {
634 dbuf_rele(dbp
[i
], tag
);
637 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
641 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
642 * indirect blocks prefeteched will be those that point to the blocks containing
643 * the data starting at offset, and continuing to offset + len.
645 * Note that if the indirect blocks above the blocks being prefetched are not
646 * in cache, they will be asychronously read in.
649 dmu_prefetch(objset_t
*os
, uint64_t object
, int64_t level
, uint64_t offset
,
650 uint64_t len
, zio_priority_t pri
)
656 if (len
== 0) { /* they're interested in the bonus buffer */
657 dn
= DMU_META_DNODE(os
);
659 if (object
== 0 || object
>= DN_MAX_OBJECT
)
662 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
663 blkid
= dbuf_whichblock(dn
, level
,
664 object
* sizeof (dnode_phys_t
));
665 dbuf_prefetch(dn
, level
, blkid
, pri
, 0);
666 rw_exit(&dn
->dn_struct_rwlock
);
671 * XXX - Note, if the dnode for the requested object is not
672 * already cached, we will do a *synchronous* read in the
673 * dnode_hold() call. The same is true for any indirects.
675 err
= dnode_hold(os
, object
, FTAG
, &dn
);
679 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
681 * offset + len - 1 is the last byte we want to prefetch for, and offset
682 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
683 * last block we want to prefetch, and dbuf_whichblock(dn, level,
684 * offset) is the first. Then the number we need to prefetch is the
687 if (level
> 0 || dn
->dn_datablkshift
!= 0) {
688 nblks
= dbuf_whichblock(dn
, level
, offset
+ len
- 1) -
689 dbuf_whichblock(dn
, level
, offset
) + 1;
691 nblks
= (offset
< dn
->dn_datablksz
);
695 blkid
= dbuf_whichblock(dn
, level
, offset
);
696 for (int i
= 0; i
< nblks
; i
++)
697 dbuf_prefetch(dn
, level
, blkid
+ i
, pri
, 0);
700 rw_exit(&dn
->dn_struct_rwlock
);
702 dnode_rele(dn
, FTAG
);
706 * Get the next "chunk" of file data to free. We traverse the file from
707 * the end so that the file gets shorter over time (if we crashes in the
708 * middle, this will leave us in a better state). We find allocated file
709 * data by simply searching the allocated level 1 indirects.
711 * On input, *start should be the first offset that does not need to be
712 * freed (e.g. "offset + length"). On return, *start will be the first
713 * offset that should be freed and l1blks is set to the number of level 1
714 * indirect blocks found within the chunk.
717 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
, uint64_t *l1blks
)
720 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
721 /* bytes of data covered by a level-1 indirect block */
723 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
725 ASSERT3U(minimum
, <=, *start
);
727 if (*start
- minimum
<= iblkrange
* maxblks
) {
730 * Assume full L1 blocks and 128k recordsize to approximate the
731 * expected number of L1 blocks in this chunk
733 *l1blks
= minimum
/ (1024 * 128 * 1024);
736 ASSERT(ISP2(iblkrange
));
738 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
742 * dnode_next_offset(BACKWARDS) will find an allocated L1
743 * indirect block at or before the input offset. We must
744 * decrement *start so that it is at the end of the region
748 err
= dnode_next_offset(dn
,
749 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
751 /* if there are no indirect blocks before start, we are done */
755 } else if (err
!= 0) {
760 /* set start to the beginning of this L1 indirect */
761 *start
= P2ALIGN(*start
, iblkrange
);
763 if (*start
< minimum
)
770 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
771 * otherwise return false.
772 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
776 dmu_objset_zfs_unmounting(objset_t
*os
)
779 if (dmu_objset_type(os
) == DMU_OST_ZFS
)
780 return (zfs_get_vfs_flag_unmounted(os
));
786 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
789 uint64_t object_size
;
791 uint64_t dirty_frees_threshold
;
792 dsl_pool_t
*dp
= dmu_objset_pool(os
);
795 return (SET_ERROR(EINVAL
));
797 object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
798 if (offset
>= object_size
)
801 if (zfs_per_txg_dirty_frees_percent
<= 100)
802 dirty_frees_threshold
=
803 zfs_per_txg_dirty_frees_percent
* zfs_dirty_data_max
/ 100;
805 dirty_frees_threshold
= zfs_dirty_data_max
/ 20;
807 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
808 length
= object_size
- offset
;
810 while (length
!= 0) {
811 uint64_t chunk_end
, chunk_begin
, chunk_len
;
812 uint64_t long_free_dirty_all_txgs
= 0;
816 if (dmu_objset_zfs_unmounting(dn
->dn_objset
))
817 return (SET_ERROR(EINTR
));
819 chunk_end
= chunk_begin
= offset
+ length
;
821 /* move chunk_begin backwards to the beginning of this chunk */
822 err
= get_next_chunk(dn
, &chunk_begin
, offset
, &l1blks
);
825 ASSERT3U(chunk_begin
, >=, offset
);
826 ASSERT3U(chunk_begin
, <=, chunk_end
);
828 chunk_len
= chunk_end
- chunk_begin
;
830 mutex_enter(&dp
->dp_lock
);
831 for (int t
= 0; t
< TXG_SIZE
; t
++) {
832 long_free_dirty_all_txgs
+=
833 dp
->dp_long_free_dirty_pertxg
[t
];
835 mutex_exit(&dp
->dp_lock
);
838 * To avoid filling up a TXG with just frees wait for
839 * the next TXG to open before freeing more chunks if
840 * we have reached the threshold of frees
842 if (dirty_frees_threshold
!= 0 &&
843 long_free_dirty_all_txgs
>= dirty_frees_threshold
) {
844 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay
);
845 txg_wait_open(dp
, 0);
849 tx
= dmu_tx_create(os
);
850 dmu_tx_hold_free(tx
, dn
->dn_object
, chunk_begin
, chunk_len
);
853 * Mark this transaction as typically resulting in a net
854 * reduction in space used.
856 dmu_tx_mark_netfree(tx
);
857 err
= dmu_tx_assign(tx
, TXG_WAIT
);
864 * In order to prevent unnecessary write throttling, for each
865 * TXG, we track the cumulative size of L1 blocks being dirtied
866 * in dnode_free_range() below. We compare this number to a
867 * tunable threshold, past which we prevent new L1 dirty freeing
868 * blocks from being added into the open TXG. See
869 * dmu_free_long_range_impl() for details. The threshold
870 * prevents write throttle activation due to dirty freeing L1
871 * blocks taking up a large percentage of zfs_dirty_data_max.
873 mutex_enter(&dp
->dp_lock
);
874 dp
->dp_long_free_dirty_pertxg
[dmu_tx_get_txg(tx
) & TXG_MASK
] +=
875 l1blks
<< dn
->dn_indblkshift
;
876 mutex_exit(&dp
->dp_lock
);
877 DTRACE_PROBE3(free__long__range
,
878 uint64_t, long_free_dirty_all_txgs
, uint64_t, chunk_len
,
879 uint64_t, dmu_tx_get_txg(tx
));
880 dnode_free_range(dn
, chunk_begin
, chunk_len
, tx
);
890 dmu_free_long_range(objset_t
*os
, uint64_t object
,
891 uint64_t offset
, uint64_t length
)
896 err
= dnode_hold(os
, object
, FTAG
, &dn
);
899 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
902 * It is important to zero out the maxblkid when freeing the entire
903 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
904 * will take the fast path, and (b) dnode_reallocate() can verify
905 * that the entire file has been freed.
907 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
910 dnode_rele(dn
, FTAG
);
915 dmu_free_long_object(objset_t
*os
, uint64_t object
)
920 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
924 tx
= dmu_tx_create(os
);
925 dmu_tx_hold_bonus(tx
, object
);
926 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
927 dmu_tx_mark_netfree(tx
);
928 err
= dmu_tx_assign(tx
, TXG_WAIT
);
931 err
= dmu_object_free(os
, object
, tx
);
942 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
943 uint64_t size
, dmu_tx_t
*tx
)
946 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
949 ASSERT(offset
< UINT64_MAX
);
950 ASSERT(size
== DMU_OBJECT_END
|| size
<= UINT64_MAX
- offset
);
951 dnode_free_range(dn
, offset
, size
, tx
);
952 dnode_rele(dn
, FTAG
);
957 dmu_read_impl(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
958 void *buf
, uint32_t flags
)
961 int numbufs
, err
= 0;
964 * Deal with odd block sizes, where there can't be data past the first
965 * block. If we ever do the tail block optimization, we will need to
966 * handle that here as well.
968 if (dn
->dn_maxblkid
== 0) {
969 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
970 MIN(size
, dn
->dn_datablksz
- offset
);
971 bzero((char *)buf
+ newsz
, size
- newsz
);
976 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
980 * NB: we could do this block-at-a-time, but it's nice
981 * to be reading in parallel.
983 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
984 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
988 for (i
= 0; i
< numbufs
; i
++) {
991 dmu_buf_t
*db
= dbp
[i
];
995 bufoff
= offset
- db
->db_offset
;
996 tocpy
= MIN(db
->db_size
- bufoff
, size
);
998 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
1002 buf
= (char *)buf
+ tocpy
;
1004 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1010 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1011 void *buf
, uint32_t flags
)
1016 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1020 err
= dmu_read_impl(dn
, offset
, size
, buf
, flags
);
1021 dnode_rele(dn
, FTAG
);
1026 dmu_read_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
, void *buf
,
1029 return (dmu_read_impl(dn
, offset
, size
, buf
, flags
));
1033 dmu_write_impl(dmu_buf_t
**dbp
, int numbufs
, uint64_t offset
, uint64_t size
,
1034 const void *buf
, dmu_tx_t
*tx
)
1038 for (i
= 0; i
< numbufs
; i
++) {
1041 dmu_buf_t
*db
= dbp
[i
];
1045 bufoff
= offset
- db
->db_offset
;
1046 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1048 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1050 if (tocpy
== db
->db_size
)
1051 dmu_buf_will_fill(db
, tx
);
1053 dmu_buf_will_dirty(db
, tx
);
1055 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
1057 if (tocpy
== db
->db_size
)
1058 dmu_buf_fill_done(db
, tx
);
1062 buf
= (char *)buf
+ tocpy
;
1067 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1068 const void *buf
, dmu_tx_t
*tx
)
1076 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
1077 FALSE
, FTAG
, &numbufs
, &dbp
));
1078 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
1079 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1083 dmu_write_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
1084 const void *buf
, dmu_tx_t
*tx
)
1092 VERIFY0(dmu_buf_hold_array_by_dnode(dn
, offset
, size
,
1093 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
));
1094 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
1095 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1099 dmu_object_remap_one_indirect(objset_t
*os
, dnode_t
*dn
,
1100 uint64_t last_removal_txg
, uint64_t offset
)
1102 uint64_t l1blkid
= dbuf_whichblock(dn
, 1, offset
);
1106 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1107 dmu_buf_impl_t
*dbuf
= dbuf_hold_level(dn
, 1, l1blkid
, FTAG
);
1108 ASSERT3P(dbuf
, !=, NULL
);
1111 * If the block hasn't been written yet, this default will ensure
1112 * we don't try to remap it.
1114 uint64_t birth
= UINT64_MAX
;
1115 ASSERT3U(last_removal_txg
, !=, UINT64_MAX
);
1116 if (dbuf
->db_blkptr
!= NULL
)
1117 birth
= dbuf
->db_blkptr
->blk_birth
;
1118 rw_exit(&dn
->dn_struct_rwlock
);
1121 * If this L1 was already written after the last removal, then we've
1122 * already tried to remap it. An additional hold is taken after the
1123 * dmu_tx_assign() to handle the case where the dnode is freed while
1124 * waiting for the next open txg.
1126 if (birth
<= last_removal_txg
&&
1127 dbuf_read(dbuf
, NULL
, DB_RF_MUST_SUCCEED
) == 0 &&
1128 dbuf_can_remap(dbuf
)) {
1129 dmu_tx_t
*tx
= dmu_tx_create(os
);
1130 dmu_tx_hold_remap_l1indirect(tx
, dn
->dn_object
);
1131 err
= dmu_tx_assign(tx
, TXG_WAIT
);
1133 err
= dnode_hold(os
, dn
->dn_object
, FTAG
, &dn_tx
);
1135 (void) dbuf_dirty(dbuf
, tx
);
1136 dnode_rele(dn_tx
, FTAG
);
1144 dbuf_rele(dbuf
, FTAG
);
1146 delay(MSEC_TO_TICK(zfs_object_remap_one_indirect_delay_ms
));
1152 * Remap all blockpointers in the object, if possible, so that they reference
1153 * only concrete vdevs.
1155 * To do this, iterate over the L0 blockpointers and remap any that reference
1156 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1157 * cannot guarantee that we can remap all blockpointer anyways (due to split
1158 * blocks), we do not want to make the code unnecessarily complicated to
1159 * catch the unlikely case that there is an L1 block on an indirect vdev that
1160 * contains no indirect blockpointers.
1163 dmu_object_remap_indirects(objset_t
*os
, uint64_t object
,
1164 uint64_t last_removal_txg
)
1166 uint64_t offset
, l1span
;
1168 dnode_t
*dn
, *dn_tx
;
1170 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1175 if (dn
->dn_nlevels
<= 1) {
1176 if (issig(JUSTLOOKING
) && issig(FORREAL
)) {
1177 err
= SET_ERROR(EINTR
);
1181 * If the dnode has no indirect blocks, we cannot dirty them.
1182 * We still want to remap the blkptr(s) in the dnode if
1183 * appropriate, so mark it as dirty. An additional hold is
1184 * taken after the dmu_tx_assign() to handle the case where
1185 * the dnode is freed while waiting for the next open txg.
1187 if (err
== 0 && dnode_needs_remap(dn
)) {
1188 dmu_tx_t
*tx
= dmu_tx_create(os
);
1189 dmu_tx_hold_bonus(tx
, object
);
1190 err
= dmu_tx_assign(tx
, TXG_WAIT
);
1192 err
= dnode_hold(os
, object
, FTAG
, &dn_tx
);
1194 dnode_setdirty(dn_tx
, tx
);
1195 dnode_rele(dn_tx
, FTAG
);
1203 dnode_rele(dn
, FTAG
);
1208 l1span
= 1ULL << (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
+
1209 dn
->dn_datablkshift
);
1211 * Find the next L1 indirect that is not a hole.
1213 while (dnode_next_offset(dn
, 0, &offset
, 2, 1, 0) == 0) {
1214 if (issig(JUSTLOOKING
) && issig(FORREAL
)) {
1215 err
= SET_ERROR(EINTR
);
1218 if ((err
= dmu_object_remap_one_indirect(os
, dn
,
1219 last_removal_txg
, offset
)) != 0) {
1225 dnode_rele(dn
, FTAG
);
1230 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1239 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
1240 FALSE
, FTAG
, &numbufs
, &dbp
));
1242 for (i
= 0; i
< numbufs
; i
++) {
1243 dmu_buf_t
*db
= dbp
[i
];
1245 dmu_buf_will_not_fill(db
, tx
);
1247 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1251 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
1252 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
1253 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
1257 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
1258 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
1259 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
1262 dmu_buf_write_embedded(db
,
1263 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
1264 uncompressed_size
, compressed_size
, byteorder
, tx
);
1266 dmu_buf_rele(db
, FTAG
);
1270 * DMU support for xuio
1272 kstat_t
*xuio_ksp
= NULL
;
1274 typedef struct xuio_stats
{
1275 /* loaned yet not returned arc_buf */
1276 kstat_named_t xuiostat_onloan_rbuf
;
1277 kstat_named_t xuiostat_onloan_wbuf
;
1278 /* whether a copy is made when loaning out a read buffer */
1279 kstat_named_t xuiostat_rbuf_copied
;
1280 kstat_named_t xuiostat_rbuf_nocopy
;
1281 /* whether a copy is made when assigning a write buffer */
1282 kstat_named_t xuiostat_wbuf_copied
;
1283 kstat_named_t xuiostat_wbuf_nocopy
;
1286 static xuio_stats_t xuio_stats
= {
1287 { "onloan_read_buf", KSTAT_DATA_UINT64
},
1288 { "onloan_write_buf", KSTAT_DATA_UINT64
},
1289 { "read_buf_copied", KSTAT_DATA_UINT64
},
1290 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
1291 { "write_buf_copied", KSTAT_DATA_UINT64
},
1292 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
1295 #define XUIOSTAT_INCR(stat, val) \
1296 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
1297 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
1299 #ifdef HAVE_UIO_ZEROCOPY
1301 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
1304 uio_t
*uio
= &xuio
->xu_uio
;
1306 uio
->uio_iovcnt
= nblk
;
1307 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
1309 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
1311 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
1312 priv
->iovp
= (iovec_t
*)uio
->uio_iov
;
1313 XUIO_XUZC_PRIV(xuio
) = priv
;
1315 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1316 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
1318 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
1324 dmu_xuio_fini(xuio_t
*xuio
)
1326 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1327 int nblk
= priv
->cnt
;
1329 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
1330 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
1331 kmem_free(priv
, sizeof (dmu_xuio_t
));
1333 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1334 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
1336 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
1340 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1341 * and increase priv->next by 1.
1344 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
1347 uio_t
*uio
= &xuio
->xu_uio
;
1348 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1349 int i
= priv
->next
++;
1351 ASSERT(i
< priv
->cnt
);
1352 ASSERT(off
+ n
<= arc_buf_lsize(abuf
));
1353 iov
= (iovec_t
*)uio
->uio_iov
+ i
;
1354 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
1356 priv
->bufs
[i
] = abuf
;
1361 dmu_xuio_cnt(xuio_t
*xuio
)
1363 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1368 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
1370 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1372 ASSERT(i
< priv
->cnt
);
1373 return (priv
->bufs
[i
]);
1377 dmu_xuio_clear(xuio_t
*xuio
, int i
)
1379 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1381 ASSERT(i
< priv
->cnt
);
1382 priv
->bufs
[i
] = NULL
;
1384 #endif /* HAVE_UIO_ZEROCOPY */
1387 xuio_stat_init(void)
1389 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1390 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1391 KSTAT_FLAG_VIRTUAL
);
1392 if (xuio_ksp
!= NULL
) {
1393 xuio_ksp
->ks_data
= &xuio_stats
;
1394 kstat_install(xuio_ksp
);
1399 xuio_stat_fini(void)
1401 if (xuio_ksp
!= NULL
) {
1402 kstat_delete(xuio_ksp
);
1408 xuio_stat_wbuf_copied(void)
1410 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1414 xuio_stat_wbuf_nocopy(void)
1416 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1421 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1424 int numbufs
, i
, err
;
1425 #ifdef HAVE_UIO_ZEROCOPY
1426 xuio_t
*xuio
= NULL
;
1430 * NB: we could do this block-at-a-time, but it's nice
1431 * to be reading in parallel.
1433 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1434 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1438 for (i
= 0; i
< numbufs
; i
++) {
1441 dmu_buf_t
*db
= dbp
[i
];
1445 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1446 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1448 #ifdef HAVE_UIO_ZEROCOPY
1450 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1451 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1452 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1453 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1455 uio
->uio_resid
-= tocpy
;
1456 uio
->uio_loffset
+= tocpy
;
1459 if (abuf
== dbuf_abuf
)
1460 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1462 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1465 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1472 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1478 * Read 'size' bytes into the uio buffer.
1479 * From object zdb->db_object.
1480 * Starting at offset uio->uio_loffset.
1482 * If the caller already has a dbuf in the target object
1483 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1484 * because we don't have to find the dnode_t for the object.
1487 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1489 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1498 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1505 * Read 'size' bytes into the uio buffer.
1506 * From the specified object
1507 * Starting at offset uio->uio_loffset.
1510 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1518 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1522 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1524 dnode_rele(dn
, FTAG
);
1530 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1537 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1538 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1542 for (i
= 0; i
< numbufs
; i
++) {
1545 dmu_buf_t
*db
= dbp
[i
];
1549 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1550 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1552 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1554 if (tocpy
== db
->db_size
)
1555 dmu_buf_will_fill(db
, tx
);
1557 dmu_buf_will_dirty(db
, tx
);
1560 * XXX uiomove could block forever (eg.nfs-backed
1561 * pages). There needs to be a uiolockdown() function
1562 * to lock the pages in memory, so that uiomove won't
1565 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1568 if (tocpy
== db
->db_size
)
1569 dmu_buf_fill_done(db
, tx
);
1577 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1582 * Write 'size' bytes from the uio buffer.
1583 * To object zdb->db_object.
1584 * Starting at offset uio->uio_loffset.
1586 * If the caller already has a dbuf in the target object
1587 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1588 * because we don't have to find the dnode_t for the object.
1591 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1594 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1603 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1610 * Write 'size' bytes from the uio buffer.
1611 * To the specified object.
1612 * Starting at offset uio->uio_loffset.
1615 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1624 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1628 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1630 dnode_rele(dn
, FTAG
);
1634 #endif /* _KERNEL */
1637 * Allocate a loaned anonymous arc buffer.
1640 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1642 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1644 return (arc_loan_buf(db
->db_objset
->os_spa
, B_FALSE
, size
));
1648 * Free a loaned arc buffer.
1651 dmu_return_arcbuf(arc_buf_t
*buf
)
1653 arc_return_buf(buf
, FTAG
);
1654 arc_buf_destroy(buf
, FTAG
);
1658 dmu_copy_from_buf(objset_t
*os
, uint64_t object
, uint64_t offset
,
1659 dmu_buf_t
*handle
, dmu_tx_t
*tx
)
1661 dmu_buf_t
*dst_handle
;
1662 dmu_buf_impl_t
*dstdb
;
1663 dmu_buf_impl_t
*srcdb
= (dmu_buf_impl_t
*)handle
;
1664 dmu_object_type_t type
;
1667 boolean_t byteorder
;
1668 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1669 uint8_t iv
[ZIO_DATA_IV_LEN
];
1670 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1672 ASSERT3P(srcdb
->db_buf
, !=, NULL
);
1674 /* hold the db that we want to write to */
1675 VERIFY0(dmu_buf_hold(os
, object
, offset
, FTAG
, &dst_handle
,
1676 DMU_READ_NO_DECRYPT
));
1677 dstdb
= (dmu_buf_impl_t
*)dst_handle
;
1678 datalen
= arc_buf_size(srcdb
->db_buf
);
1680 DB_DNODE_ENTER(dstdb
);
1681 type
= DB_DNODE(dstdb
)->dn_type
;
1682 DB_DNODE_EXIT(dstdb
);
1684 /* allocated an arc buffer that matches the type of srcdb->db_buf */
1685 if (arc_is_encrypted(srcdb
->db_buf
)) {
1686 arc_get_raw_params(srcdb
->db_buf
, &byteorder
, salt
, iv
, mac
);
1687 abuf
= arc_loan_raw_buf(os
->os_spa
, dmu_objset_id(os
),
1688 byteorder
, salt
, iv
, mac
, type
,
1689 datalen
, arc_buf_lsize(srcdb
->db_buf
),
1690 arc_get_compression(srcdb
->db_buf
));
1692 /* we won't get a compressed db back from dmu_buf_hold() */
1693 ASSERT3U(arc_get_compression(srcdb
->db_buf
),
1694 ==, ZIO_COMPRESS_OFF
);
1695 abuf
= arc_loan_buf(os
->os_spa
,
1696 DMU_OT_IS_METADATA(type
), datalen
);
1699 ASSERT3U(datalen
, ==, arc_buf_size(abuf
));
1701 /* copy the data to the new buffer and assign it to the dstdb */
1702 bcopy(srcdb
->db_buf
->b_data
, abuf
->b_data
, datalen
);
1703 dbuf_assign_arcbuf(dstdb
, abuf
, tx
);
1704 dmu_buf_rele(dst_handle
, FTAG
);
1708 * When possible directly assign passed loaned arc buffer to a dbuf.
1709 * If this is not possible copy the contents of passed arc buf via
1713 dmu_assign_arcbuf_by_dnode(dnode_t
*dn
, uint64_t offset
, arc_buf_t
*buf
,
1717 objset_t
*os
= dn
->dn_objset
;
1718 uint64_t object
= dn
->dn_object
;
1719 uint32_t blksz
= (uint32_t)arc_buf_lsize(buf
);
1722 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1723 blkid
= dbuf_whichblock(dn
, 0, offset
);
1724 db
= dbuf_hold(dn
, blkid
, FTAG
);
1726 return (SET_ERROR(EIO
));
1727 rw_exit(&dn
->dn_struct_rwlock
);
1730 * We can only assign if the offset is aligned, the arc buf is the
1731 * same size as the dbuf, and the dbuf is not metadata.
1733 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1734 dbuf_assign_arcbuf(db
, buf
, tx
);
1735 dbuf_rele(db
, FTAG
);
1737 /* compressed bufs must always be assignable to their dbuf */
1738 ASSERT3U(arc_get_compression(buf
), ==, ZIO_COMPRESS_OFF
);
1739 ASSERT(!(buf
->b_flags
& ARC_BUF_FLAG_COMPRESSED
));
1741 dbuf_rele(db
, FTAG
);
1742 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1743 dmu_return_arcbuf(buf
);
1744 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1751 dmu_assign_arcbuf_by_dbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1755 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1757 DB_DNODE_ENTER(dbuf
);
1758 err
= dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf
), offset
, buf
, tx
);
1759 DB_DNODE_EXIT(dbuf
);
1765 dbuf_dirty_record_t
*dsa_dr
;
1766 dmu_sync_cb_t
*dsa_done
;
1773 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1775 dmu_sync_arg_t
*dsa
= varg
;
1776 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1777 blkptr_t
*bp
= zio
->io_bp
;
1779 if (zio
->io_error
== 0) {
1780 if (BP_IS_HOLE(bp
)) {
1782 * A block of zeros may compress to a hole, but the
1783 * block size still needs to be known for replay.
1785 BP_SET_LSIZE(bp
, db
->db_size
);
1786 } else if (!BP_IS_EMBEDDED(bp
)) {
1787 ASSERT(BP_GET_LEVEL(bp
) == 0);
1794 dmu_sync_late_arrival_ready(zio_t
*zio
)
1796 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1801 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1803 dmu_sync_arg_t
*dsa
= varg
;
1804 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1805 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1806 zgd_t
*zgd
= dsa
->dsa_zgd
;
1809 * Record the vdev(s) backing this blkptr so they can be flushed after
1810 * the writes for the lwb have completed.
1812 if (zio
->io_error
== 0) {
1813 zil_lwb_add_block(zgd
->zgd_lwb
, zgd
->zgd_bp
);
1816 mutex_enter(&db
->db_mtx
);
1817 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1818 if (zio
->io_error
== 0) {
1819 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1820 if (dr
->dt
.dl
.dr_nopwrite
) {
1821 blkptr_t
*bp
= zio
->io_bp
;
1822 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
1823 uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
);
1825 ASSERT(BP_EQUAL(bp
, bp_orig
));
1826 VERIFY(BP_EQUAL(bp
, db
->db_blkptr
));
1827 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1828 VERIFY(zio_checksum_table
[chksum
].ci_flags
&
1829 ZCHECKSUM_FLAG_NOPWRITE
);
1831 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1832 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1833 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1836 * Old style holes are filled with all zeros, whereas
1837 * new-style holes maintain their lsize, type, level,
1838 * and birth time (see zio_write_compress). While we
1839 * need to reset the BP_SET_LSIZE() call that happened
1840 * in dmu_sync_ready for old style holes, we do *not*
1841 * want to wipe out the information contained in new
1842 * style holes. Thus, only zero out the block pointer if
1843 * it's an old style hole.
1845 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1846 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1847 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1849 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1851 cv_broadcast(&db
->db_changed
);
1852 mutex_exit(&db
->db_mtx
);
1854 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1856 kmem_free(dsa
, sizeof (*dsa
));
1860 dmu_sync_late_arrival_done(zio_t
*zio
)
1862 blkptr_t
*bp
= zio
->io_bp
;
1863 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1864 zgd_t
*zgd
= dsa
->dsa_zgd
;
1866 if (zio
->io_error
== 0) {
1868 * Record the vdev(s) backing this blkptr so they can be
1869 * flushed after the writes for the lwb have completed.
1871 zil_lwb_add_block(zgd
->zgd_lwb
, zgd
->zgd_bp
);
1873 if (!BP_IS_HOLE(bp
)) {
1874 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1875 ASSERT(!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
1876 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1877 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1878 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1879 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1883 dmu_tx_commit(dsa
->dsa_tx
);
1885 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1887 abd_put(zio
->io_abd
);
1888 kmem_free(dsa
, sizeof (*dsa
));
1892 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1893 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1895 dmu_sync_arg_t
*dsa
;
1898 tx
= dmu_tx_create(os
);
1899 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1900 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1902 /* Make zl_get_data do txg_waited_synced() */
1903 return (SET_ERROR(EIO
));
1907 * In order to prevent the zgd's lwb from being free'd prior to
1908 * dmu_sync_late_arrival_done() being called, we have to ensure
1909 * the lwb's "max txg" takes this tx's txg into account.
1911 zil_lwb_add_txg(zgd
->zgd_lwb
, dmu_tx_get_txg(tx
));
1913 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1915 dsa
->dsa_done
= done
;
1920 * Since we are currently syncing this txg, it's nontrivial to
1921 * determine what BP to nopwrite against, so we disable nopwrite.
1923 * When syncing, the db_blkptr is initially the BP of the previous
1924 * txg. We can not nopwrite against it because it will be changed
1925 * (this is similar to the non-late-arrival case where the dbuf is
1926 * dirty in a future txg).
1928 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1929 * We can not nopwrite against it because although the BP will not
1930 * (typically) be changed, the data has not yet been persisted to this
1933 * Finally, when dbuf_write_done() is called, it is theoretically
1934 * possible to always nopwrite, because the data that was written in
1935 * this txg is the same data that we are trying to write. However we
1936 * would need to check that this dbuf is not dirty in any future
1937 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1938 * don't nopwrite in this case.
1940 zp
->zp_nopwrite
= B_FALSE
;
1942 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1943 abd_get_from_buf(zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
),
1944 zgd
->zgd_db
->db_size
, zgd
->zgd_db
->db_size
, zp
,
1945 dmu_sync_late_arrival_ready
, NULL
, NULL
, dmu_sync_late_arrival_done
,
1946 dsa
, ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, zb
));
1952 * Intent log support: sync the block associated with db to disk.
1953 * N.B. and XXX: the caller is responsible for making sure that the
1954 * data isn't changing while dmu_sync() is writing it.
1958 * EEXIST: this txg has already been synced, so there's nothing to do.
1959 * The caller should not log the write.
1961 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1962 * The caller should not log the write.
1964 * EALREADY: this block is already in the process of being synced.
1965 * The caller should track its progress (somehow).
1967 * EIO: could not do the I/O.
1968 * The caller should do a txg_wait_synced().
1970 * 0: the I/O has been initiated.
1971 * The caller should log this blkptr in the done callback.
1972 * It is possible that the I/O will fail, in which case
1973 * the error will be reported to the done callback and
1974 * propagated to pio from zio_done().
1977 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1979 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1980 objset_t
*os
= db
->db_objset
;
1981 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1982 dbuf_dirty_record_t
*dr
;
1983 dmu_sync_arg_t
*dsa
;
1984 zbookmark_phys_t zb
;
1988 ASSERT(pio
!= NULL
);
1991 SET_BOOKMARK(&zb
, ds
->ds_object
,
1992 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1996 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
2000 * If we're frozen (running ziltest), we always need to generate a bp.
2002 if (txg
> spa_freeze_txg(os
->os_spa
))
2003 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
2006 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2007 * and us. If we determine that this txg is not yet syncing,
2008 * but it begins to sync a moment later, that's OK because the
2009 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2011 mutex_enter(&db
->db_mtx
);
2013 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
2015 * This txg has already synced. There's nothing to do.
2017 mutex_exit(&db
->db_mtx
);
2018 return (SET_ERROR(EEXIST
));
2021 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
2023 * This txg is currently syncing, so we can't mess with
2024 * the dirty record anymore; just write a new log block.
2026 mutex_exit(&db
->db_mtx
);
2027 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
2030 dr
= db
->db_last_dirty
;
2031 while (dr
&& dr
->dr_txg
!= txg
)
2036 * There's no dr for this dbuf, so it must have been freed.
2037 * There's no need to log writes to freed blocks, so we're done.
2039 mutex_exit(&db
->db_mtx
);
2040 return (SET_ERROR(ENOENT
));
2043 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
2045 if (db
->db_blkptr
!= NULL
) {
2047 * We need to fill in zgd_bp with the current blkptr so that
2048 * the nopwrite code can check if we're writing the same
2049 * data that's already on disk. We can only nopwrite if we
2050 * are sure that after making the copy, db_blkptr will not
2051 * change until our i/o completes. We ensure this by
2052 * holding the db_mtx, and only allowing nopwrite if the
2053 * block is not already dirty (see below). This is verified
2054 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2057 *zgd
->zgd_bp
= *db
->db_blkptr
;
2061 * Assume the on-disk data is X, the current syncing data (in
2062 * txg - 1) is Y, and the current in-memory data is Z (currently
2065 * We usually want to perform a nopwrite if X and Z are the
2066 * same. However, if Y is different (i.e. the BP is going to
2067 * change before this write takes effect), then a nopwrite will
2068 * be incorrect - we would override with X, which could have
2069 * been freed when Y was written.
2071 * (Note that this is not a concern when we are nop-writing from
2072 * syncing context, because X and Y must be identical, because
2073 * all previous txgs have been synced.)
2075 * Therefore, we disable nopwrite if the current BP could change
2076 * before this TXG. There are two ways it could change: by
2077 * being dirty (dr_next is non-NULL), or by being freed
2078 * (dnode_block_freed()). This behavior is verified by
2079 * zio_done(), which VERIFYs that the override BP is identical
2080 * to the on-disk BP.
2084 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
2085 zp
.zp_nopwrite
= B_FALSE
;
2088 ASSERT(dr
->dr_txg
== txg
);
2089 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
2090 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
2092 * We have already issued a sync write for this buffer,
2093 * or this buffer has already been synced. It could not
2094 * have been dirtied since, or we would have cleared the state.
2096 mutex_exit(&db
->db_mtx
);
2097 return (SET_ERROR(EALREADY
));
2100 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
2101 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
2102 mutex_exit(&db
->db_mtx
);
2104 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
2106 dsa
->dsa_done
= done
;
2110 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
2111 zgd
->zgd_bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
2112 &zp
, dmu_sync_ready
, NULL
, NULL
, dmu_sync_done
, dsa
,
2113 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, &zb
));
2119 dmu_object_set_nlevels(objset_t
*os
, uint64_t object
, int nlevels
, dmu_tx_t
*tx
)
2124 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2127 err
= dnode_set_nlevels(dn
, nlevels
, tx
);
2128 dnode_rele(dn
, FTAG
);
2133 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
2139 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2142 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
2143 dnode_rele(dn
, FTAG
);
2148 dmu_object_set_maxblkid(objset_t
*os
, uint64_t object
, uint64_t maxblkid
,
2154 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2157 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2158 dnode_new_blkid(dn
, maxblkid
, tx
, B_FALSE
);
2159 rw_exit(&dn
->dn_struct_rwlock
);
2160 dnode_rele(dn
, FTAG
);
2165 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
2171 * Send streams include each object's checksum function. This
2172 * check ensures that the receiving system can understand the
2173 * checksum function transmitted.
2175 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
2177 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
2178 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
2179 dn
->dn_checksum
= checksum
;
2180 dnode_setdirty(dn
, tx
);
2181 dnode_rele(dn
, FTAG
);
2185 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
2191 * Send streams include each object's compression function. This
2192 * check ensures that the receiving system can understand the
2193 * compression function transmitted.
2195 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
2197 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
2198 dn
->dn_compress
= compress
;
2199 dnode_setdirty(dn
, tx
);
2200 dnode_rele(dn
, FTAG
);
2204 * When the "redundant_metadata" property is set to "most", only indirect
2205 * blocks of this level and higher will have an additional ditto block.
2207 int zfs_redundant_metadata_most_ditto_level
= 2;
2210 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
2212 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
2213 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
2215 enum zio_checksum checksum
= os
->os_checksum
;
2216 enum zio_compress compress
= os
->os_compress
;
2217 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
2218 boolean_t dedup
= B_FALSE
;
2219 boolean_t nopwrite
= B_FALSE
;
2220 boolean_t dedup_verify
= os
->os_dedup_verify
;
2221 boolean_t encrypt
= B_FALSE
;
2222 int copies
= os
->os_copies
;
2225 * We maintain different write policies for each of the following
2228 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2229 * 3. all other level 0 blocks
2233 * XXX -- we should design a compression algorithm
2234 * that specializes in arrays of bps.
2236 compress
= zio_compress_select(os
->os_spa
,
2237 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
2240 * Metadata always gets checksummed. If the data
2241 * checksum is multi-bit correctable, and it's not a
2242 * ZBT-style checksum, then it's suitable for metadata
2243 * as well. Otherwise, the metadata checksum defaults
2246 if (!(zio_checksum_table
[checksum
].ci_flags
&
2247 ZCHECKSUM_FLAG_METADATA
) ||
2248 (zio_checksum_table
[checksum
].ci_flags
&
2249 ZCHECKSUM_FLAG_EMBEDDED
))
2250 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
2252 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
2253 (os
->os_redundant_metadata
==
2254 ZFS_REDUNDANT_METADATA_MOST
&&
2255 (level
>= zfs_redundant_metadata_most_ditto_level
||
2256 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
2258 } else if (wp
& WP_NOFILL
) {
2262 * If we're writing preallocated blocks, we aren't actually
2263 * writing them so don't set any policy properties. These
2264 * blocks are currently only used by an external subsystem
2265 * outside of zfs (i.e. dump) and not written by the zio
2268 compress
= ZIO_COMPRESS_OFF
;
2269 checksum
= ZIO_CHECKSUM_OFF
;
2271 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
2274 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
2275 zio_checksum_select(dn
->dn_checksum
, checksum
) :
2279 * Determine dedup setting. If we are in dmu_sync(),
2280 * we won't actually dedup now because that's all
2281 * done in syncing context; but we do want to use the
2282 * dedup checkum. If the checksum is not strong
2283 * enough to ensure unique signatures, force
2286 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
2287 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
2288 if (!(zio_checksum_table
[checksum
].ci_flags
&
2289 ZCHECKSUM_FLAG_DEDUP
))
2290 dedup_verify
= B_TRUE
;
2294 * Enable nopwrite if we have secure enough checksum
2295 * algorithm (see comment in zio_nop_write) and
2296 * compression is enabled. We don't enable nopwrite if
2297 * dedup is enabled as the two features are mutually
2300 nopwrite
= (!dedup
&& (zio_checksum_table
[checksum
].ci_flags
&
2301 ZCHECKSUM_FLAG_NOPWRITE
) &&
2302 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
2306 * All objects in an encrypted objset are protected from modification
2307 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2308 * in the bp, so we cannot use all copies. Encrypted objects are also
2309 * not subject to nopwrite since writing the same data will still
2310 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2311 * to avoid ambiguity in the dedup code since the DDT does not store
2314 if (os
->os_encrypted
&& (wp
& WP_NOFILL
) == 0) {
2317 if (DMU_OT_IS_ENCRYPTED(type
)) {
2318 copies
= MIN(copies
, SPA_DVAS_PER_BP
- 1);
2325 (type
== DMU_OT_DNODE
|| type
== DMU_OT_OBJSET
)) {
2326 compress
= ZIO_COMPRESS_EMPTY
;
2330 zp
->zp_compress
= compress
;
2331 zp
->zp_checksum
= checksum
;
2332 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
2333 zp
->zp_level
= level
;
2334 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
2335 zp
->zp_dedup
= dedup
;
2336 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
2337 zp
->zp_nopwrite
= nopwrite
;
2338 zp
->zp_encrypt
= encrypt
;
2339 zp
->zp_byteorder
= ZFS_HOST_BYTEORDER
;
2340 bzero(zp
->zp_salt
, ZIO_DATA_SALT_LEN
);
2341 bzero(zp
->zp_iv
, ZIO_DATA_IV_LEN
);
2342 bzero(zp
->zp_mac
, ZIO_DATA_MAC_LEN
);
2343 zp
->zp_zpl_smallblk
= DMU_OT_IS_FILE(zp
->zp_type
) ?
2344 os
->os_zpl_special_smallblock
: 0;
2346 ASSERT3U(zp
->zp_compress
, !=, ZIO_COMPRESS_INHERIT
);
2350 * This function is only called from zfs_holey_common() for zpl_llseek()
2351 * in order to determine the location of holes. In order to accurately
2352 * report holes all dirty data must be synced to disk. This causes extremely
2353 * poor performance when seeking for holes in a dirty file. As a compromise,
2354 * only provide hole data when the dnode is clean. When a dnode is dirty
2355 * report the dnode as having no holes which is always a safe thing to do.
2358 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
2362 boolean_t clean
= B_TRUE
;
2364 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2369 * Check if dnode is dirty
2371 for (i
= 0; i
< TXG_SIZE
; i
++) {
2372 if (multilist_link_active(&dn
->dn_dirty_link
[i
])) {
2379 * If compatibility option is on, sync any current changes before
2380 * we go trundling through the block pointers.
2382 if (!clean
&& zfs_dmu_offset_next_sync
) {
2384 dnode_rele(dn
, FTAG
);
2385 txg_wait_synced(dmu_objset_pool(os
), 0);
2386 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2392 err
= dnode_next_offset(dn
,
2393 (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
2395 err
= SET_ERROR(EBUSY
);
2397 dnode_rele(dn
, FTAG
);
2403 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2405 dnode_phys_t
*dnp
= dn
->dn_phys
;
2407 doi
->doi_data_block_size
= dn
->dn_datablksz
;
2408 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
2409 1ULL << dn
->dn_indblkshift
: 0;
2410 doi
->doi_type
= dn
->dn_type
;
2411 doi
->doi_bonus_type
= dn
->dn_bonustype
;
2412 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
2413 doi
->doi_dnodesize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
2414 doi
->doi_indirection
= dn
->dn_nlevels
;
2415 doi
->doi_checksum
= dn
->dn_checksum
;
2416 doi
->doi_compress
= dn
->dn_compress
;
2417 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
2418 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
2419 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
2420 doi
->doi_fill_count
= 0;
2421 for (int i
= 0; i
< dnp
->dn_nblkptr
; i
++)
2422 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
2426 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2428 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2429 mutex_enter(&dn
->dn_mtx
);
2431 __dmu_object_info_from_dnode(dn
, doi
);
2433 mutex_exit(&dn
->dn_mtx
);
2434 rw_exit(&dn
->dn_struct_rwlock
);
2438 * Get information on a DMU object.
2439 * If doi is NULL, just indicates whether the object exists.
2442 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
2445 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
2451 dmu_object_info_from_dnode(dn
, doi
);
2453 dnode_rele(dn
, FTAG
);
2458 * As above, but faster; can be used when you have a held dbuf in hand.
2461 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
2463 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2466 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
2471 * Faster still when you only care about the size.
2472 * This is specifically optimized for zfs_getattr().
2475 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
2476 u_longlong_t
*nblk512
)
2478 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2484 *blksize
= dn
->dn_datablksz
;
2485 /* add in number of slots used for the dnode itself */
2486 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
2487 SPA_MINBLOCKSHIFT
) + dn
->dn_num_slots
;
2492 dmu_object_dnsize_from_db(dmu_buf_t
*db_fake
, int *dnsize
)
2494 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2499 *dnsize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
2504 byteswap_uint64_array(void *vbuf
, size_t size
)
2506 uint64_t *buf
= vbuf
;
2507 size_t count
= size
>> 3;
2510 ASSERT((size
& 7) == 0);
2512 for (i
= 0; i
< count
; i
++)
2513 buf
[i
] = BSWAP_64(buf
[i
]);
2517 byteswap_uint32_array(void *vbuf
, size_t size
)
2519 uint32_t *buf
= vbuf
;
2520 size_t count
= size
>> 2;
2523 ASSERT((size
& 3) == 0);
2525 for (i
= 0; i
< count
; i
++)
2526 buf
[i
] = BSWAP_32(buf
[i
]);
2530 byteswap_uint16_array(void *vbuf
, size_t size
)
2532 uint16_t *buf
= vbuf
;
2533 size_t count
= size
>> 1;
2536 ASSERT((size
& 1) == 0);
2538 for (i
= 0; i
< count
; i
++)
2539 buf
[i
] = BSWAP_16(buf
[i
]);
2544 byteswap_uint8_array(void *vbuf
, size_t size
)
2567 arc_fini(); /* arc depends on l2arc, so arc must go first */
2580 #if defined(_KERNEL)
2581 EXPORT_SYMBOL(dmu_bonus_hold
);
2582 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode
);
2583 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2584 EXPORT_SYMBOL(dmu_buf_rele_array
);
2585 EXPORT_SYMBOL(dmu_prefetch
);
2586 EXPORT_SYMBOL(dmu_free_range
);
2587 EXPORT_SYMBOL(dmu_free_long_range
);
2588 EXPORT_SYMBOL(dmu_free_long_object
);
2589 EXPORT_SYMBOL(dmu_read
);
2590 EXPORT_SYMBOL(dmu_read_by_dnode
);
2591 EXPORT_SYMBOL(dmu_write
);
2592 EXPORT_SYMBOL(dmu_write_by_dnode
);
2593 EXPORT_SYMBOL(dmu_prealloc
);
2594 EXPORT_SYMBOL(dmu_object_info
);
2595 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2596 EXPORT_SYMBOL(dmu_object_info_from_db
);
2597 EXPORT_SYMBOL(dmu_object_size_from_db
);
2598 EXPORT_SYMBOL(dmu_object_dnsize_from_db
);
2599 EXPORT_SYMBOL(dmu_object_set_nlevels
);
2600 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2601 EXPORT_SYMBOL(dmu_object_set_maxblkid
);
2602 EXPORT_SYMBOL(dmu_object_set_checksum
);
2603 EXPORT_SYMBOL(dmu_object_set_compress
);
2604 EXPORT_SYMBOL(dmu_write_policy
);
2605 EXPORT_SYMBOL(dmu_sync
);
2606 EXPORT_SYMBOL(dmu_request_arcbuf
);
2607 EXPORT_SYMBOL(dmu_return_arcbuf
);
2608 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode
);
2609 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf
);
2610 EXPORT_SYMBOL(dmu_buf_hold
);
2611 EXPORT_SYMBOL(dmu_ot
);
2614 module_param(zfs_nopwrite_enabled
, int, 0644);
2615 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");
2617 module_param(zfs_per_txg_dirty_frees_percent
, ulong
, 0644);
2618 MODULE_PARM_DESC(zfs_per_txg_dirty_frees_percent
,
2619 "percentage of dirtied blocks from frees in one TXG");
2621 module_param(zfs_dmu_offset_next_sync
, int, 0644);
2622 MODULE_PARM_DESC(zfs_dmu_offset_next_sync
,
2623 "Enable forcing txg sync to find holes");