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.
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
34 #include <sys/dnode.h>
35 #include <sys/zfs_context.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dmu_traverse.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/dsl_synctask.h>
42 #include <sys/dsl_prop.h>
43 #include <sys/dmu_zfetch.h>
44 #include <sys/zfs_ioctl.h>
46 #include <sys/zio_checksum.h>
47 #include <sys/zio_compress.h>
49 #include <sys/zfeature.h>
51 #include <sys/trace_dmu.h>
52 #include <sys/zfs_rlock.h>
54 #include <sys/vmsystm.h>
55 #include <sys/zfs_znode.h>
59 * Enable/disable nopwrite feature.
61 int zfs_nopwrite_enabled
= 1;
64 * Tunable to control percentage of dirtied blocks from frees in one TXG.
65 * After this threshold is crossed, additional dirty blocks from frees
66 * wait until the next TXG.
67 * A value of zero will disable this throttle.
69 unsigned long zfs_per_txg_dirty_frees_percent
= 30;
72 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
74 int zfs_dmu_offset_next_sync
= 0;
77 * This can be used for testing, to ensure that certain actions happen
78 * while in the middle of a remap (which might otherwise complete too
81 int zfs_object_remap_one_indirect_delay_ticks
= 0;
83 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
84 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "unallocated" },
85 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "object directory" },
86 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "object array" },
87 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "packed nvlist" },
88 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "packed nvlist size" },
89 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj" },
90 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj header" },
91 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA space map header" },
92 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA space map" },
93 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, TRUE
, "ZIL intent log" },
94 {DMU_BSWAP_DNODE
, TRUE
, FALSE
, TRUE
, "DMU dnode" },
95 {DMU_BSWAP_OBJSET
, TRUE
, TRUE
, FALSE
, "DMU objset" },
96 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL directory" },
97 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL directory child map"},
98 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dataset snap map" },
99 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL props" },
100 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL dataset" },
101 {DMU_BSWAP_ZNODE
, TRUE
, FALSE
, FALSE
, "ZFS znode" },
102 {DMU_BSWAP_OLDACL
, TRUE
, FALSE
, TRUE
, "ZFS V0 ACL" },
103 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "ZFS plain file" },
104 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS directory" },
105 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "ZFS master node" },
106 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS delete queue" },
107 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "zvol object" },
108 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "zvol prop" },
109 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "other uint8[]" },
110 {DMU_BSWAP_UINT64
, FALSE
, FALSE
, TRUE
, "other uint64[]" },
111 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "other ZAP" },
112 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "persistent error log" },
113 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "SPA history" },
114 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA history offsets" },
115 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "Pool properties" },
116 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL permissions" },
117 {DMU_BSWAP_ACL
, TRUE
, FALSE
, TRUE
, "ZFS ACL" },
118 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "ZFS SYSACL" },
119 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "FUID table" },
120 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "FUID table size" },
121 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dataset next clones"},
122 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "scan work queue" },
123 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS user/group/project used" },
124 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS user/group/project quota"},
125 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "snapshot refcount tags"},
126 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "DDT ZAP algorithm" },
127 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "DDT statistics" },
128 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "System attributes" },
129 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA master node" },
130 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA attr registration" },
131 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA attr layouts" },
132 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "scan translations" },
133 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "deduplicated block" },
134 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL deadlist map" },
135 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL deadlist map hdr" },
136 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dir clones" },
137 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj subobj" }
140 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
141 { byteswap_uint8_array
, "uint8" },
142 { byteswap_uint16_array
, "uint16" },
143 { byteswap_uint32_array
, "uint32" },
144 { byteswap_uint64_array
, "uint64" },
145 { zap_byteswap
, "zap" },
146 { dnode_buf_byteswap
, "dnode" },
147 { dmu_objset_byteswap
, "objset" },
148 { zfs_znode_byteswap
, "znode" },
149 { zfs_oldacl_byteswap
, "oldacl" },
150 { zfs_acl_byteswap
, "acl" }
154 dmu_buf_hold_noread_by_dnode(dnode_t
*dn
, uint64_t offset
,
155 void *tag
, dmu_buf_t
**dbp
)
160 blkid
= dbuf_whichblock(dn
, 0, offset
);
161 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
162 db
= dbuf_hold(dn
, blkid
, tag
);
163 rw_exit(&dn
->dn_struct_rwlock
);
167 return (SET_ERROR(EIO
));
174 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
175 void *tag
, dmu_buf_t
**dbp
)
182 err
= dnode_hold(os
, object
, FTAG
, &dn
);
185 blkid
= dbuf_whichblock(dn
, 0, offset
);
186 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
187 db
= dbuf_hold(dn
, blkid
, tag
);
188 rw_exit(&dn
->dn_struct_rwlock
);
189 dnode_rele(dn
, FTAG
);
193 return (SET_ERROR(EIO
));
201 dmu_buf_hold_by_dnode(dnode_t
*dn
, uint64_t offset
,
202 void *tag
, dmu_buf_t
**dbp
, int flags
)
205 int db_flags
= DB_RF_CANFAIL
;
207 if (flags
& DMU_READ_NO_PREFETCH
)
208 db_flags
|= DB_RF_NOPREFETCH
;
209 if (flags
& DMU_READ_NO_DECRYPT
)
210 db_flags
|= DB_RF_NO_DECRYPT
;
212 err
= dmu_buf_hold_noread_by_dnode(dn
, offset
, tag
, dbp
);
214 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
215 err
= dbuf_read(db
, NULL
, db_flags
);
226 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
227 void *tag
, dmu_buf_t
**dbp
, int flags
)
230 int db_flags
= DB_RF_CANFAIL
;
232 if (flags
& DMU_READ_NO_PREFETCH
)
233 db_flags
|= DB_RF_NOPREFETCH
;
234 if (flags
& DMU_READ_NO_DECRYPT
)
235 db_flags
|= DB_RF_NO_DECRYPT
;
237 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
239 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
240 err
= dbuf_read(db
, NULL
, db_flags
);
253 return (DN_OLD_MAX_BONUSLEN
);
257 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
259 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
266 if (dn
->dn_bonus
!= db
) {
267 error
= SET_ERROR(EINVAL
);
268 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
269 error
= SET_ERROR(EINVAL
);
271 dnode_setbonuslen(dn
, newsize
, tx
);
280 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
282 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
289 if (!DMU_OT_IS_VALID(type
)) {
290 error
= SET_ERROR(EINVAL
);
291 } else if (dn
->dn_bonus
!= db
) {
292 error
= SET_ERROR(EINVAL
);
294 dnode_setbonus_type(dn
, type
, tx
);
303 dmu_get_bonustype(dmu_buf_t
*db_fake
)
305 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
307 dmu_object_type_t type
;
311 type
= dn
->dn_bonustype
;
318 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
323 error
= dnode_hold(os
, object
, FTAG
, &dn
);
324 dbuf_rm_spill(dn
, tx
);
325 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
326 dnode_rm_spill(dn
, tx
);
327 rw_exit(&dn
->dn_struct_rwlock
);
328 dnode_rele(dn
, FTAG
);
333 * returns ENOENT, EIO, or 0.
336 dmu_bonus_hold_impl(objset_t
*os
, uint64_t object
, void *tag
, uint32_t flags
,
342 uint32_t db_flags
= DB_RF_MUST_SUCCEED
;
344 if (flags
& DMU_READ_NO_PREFETCH
)
345 db_flags
|= DB_RF_NOPREFETCH
;
346 if (flags
& DMU_READ_NO_DECRYPT
)
347 db_flags
|= DB_RF_NO_DECRYPT
;
349 error
= dnode_hold(os
, object
, FTAG
, &dn
);
353 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
354 if (dn
->dn_bonus
== NULL
) {
355 rw_exit(&dn
->dn_struct_rwlock
);
356 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
357 if (dn
->dn_bonus
== NULL
)
358 dbuf_create_bonus(dn
);
362 /* as long as the bonus buf is held, the dnode will be held */
363 if (zfs_refcount_add(&db
->db_holds
, tag
) == 1) {
364 VERIFY(dnode_add_ref(dn
, db
));
365 atomic_inc_32(&dn
->dn_dbufs_count
);
369 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
370 * hold and incrementing the dbuf count to ensure that dnode_move() sees
371 * a dnode hold for every dbuf.
373 rw_exit(&dn
->dn_struct_rwlock
);
375 dnode_rele(dn
, FTAG
);
377 error
= dbuf_read(db
, NULL
, db_flags
);
379 dnode_evict_bonus(dn
);
390 dmu_bonus_hold(objset_t
*os
, uint64_t obj
, void *tag
, dmu_buf_t
**dbp
)
392 return (dmu_bonus_hold_impl(os
, obj
, tag
, DMU_READ_NO_PREFETCH
, dbp
));
396 * returns ENOENT, EIO, or 0.
398 * This interface will allocate a blank spill dbuf when a spill blk
399 * doesn't already exist on the dnode.
401 * if you only want to find an already existing spill db, then
402 * dmu_spill_hold_existing() should be used.
405 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
407 dmu_buf_impl_t
*db
= NULL
;
410 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
411 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
413 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
415 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
416 rw_exit(&dn
->dn_struct_rwlock
);
420 return (SET_ERROR(EIO
));
422 err
= dbuf_read(db
, NULL
, flags
);
433 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
435 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
442 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
443 err
= SET_ERROR(EINVAL
);
445 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
447 if (!dn
->dn_have_spill
) {
448 err
= SET_ERROR(ENOENT
);
450 err
= dmu_spill_hold_by_dnode(dn
,
451 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
454 rw_exit(&dn
->dn_struct_rwlock
);
462 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, uint32_t flags
, void *tag
,
465 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
468 uint32_t db_flags
= DB_RF_CANFAIL
;
470 if (flags
& DMU_READ_NO_DECRYPT
)
471 db_flags
|= DB_RF_NO_DECRYPT
;
475 err
= dmu_spill_hold_by_dnode(dn
, db_flags
, tag
, dbp
);
482 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
483 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
484 * and can induce severe lock contention when writing to several files
485 * whose dnodes are in the same block.
488 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
489 boolean_t read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
492 uint64_t blkid
, nblks
, i
;
497 ASSERT(length
<= DMU_MAX_ACCESS
);
500 * Note: We directly notify the prefetch code of this read, so that
501 * we can tell it about the multi-block read. dbuf_read() only knows
502 * about the one block it is accessing.
504 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
|
507 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
508 if (dn
->dn_datablkshift
) {
509 int blkshift
= dn
->dn_datablkshift
;
510 nblks
= (P2ROUNDUP(offset
+ length
, 1ULL << blkshift
) -
511 P2ALIGN(offset
, 1ULL << blkshift
)) >> blkshift
;
513 if (offset
+ length
> dn
->dn_datablksz
) {
514 zfs_panic_recover("zfs: accessing past end of object "
515 "%llx/%llx (size=%u access=%llu+%llu)",
516 (longlong_t
)dn
->dn_objset
->
517 os_dsl_dataset
->ds_object
,
518 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
519 (longlong_t
)offset
, (longlong_t
)length
);
520 rw_exit(&dn
->dn_struct_rwlock
);
521 return (SET_ERROR(EIO
));
525 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_SLEEP
);
527 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
528 blkid
= dbuf_whichblock(dn
, 0, offset
);
529 for (i
= 0; i
< nblks
; i
++) {
530 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+ i
, tag
);
532 rw_exit(&dn
->dn_struct_rwlock
);
533 dmu_buf_rele_array(dbp
, nblks
, tag
);
535 return (SET_ERROR(EIO
));
538 /* initiate async i/o */
540 (void) dbuf_read(db
, zio
, dbuf_flags
);
544 if ((flags
& DMU_READ_NO_PREFETCH
) == 0 &&
545 DNODE_META_IS_CACHEABLE(dn
) && length
<= zfetch_array_rd_sz
) {
546 dmu_zfetch(&dn
->dn_zfetch
, blkid
, nblks
,
547 read
&& DNODE_IS_CACHEABLE(dn
));
549 rw_exit(&dn
->dn_struct_rwlock
);
551 /* wait for async i/o */
554 dmu_buf_rele_array(dbp
, nblks
, tag
);
558 /* wait for other io to complete */
560 for (i
= 0; i
< nblks
; i
++) {
561 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
562 mutex_enter(&db
->db_mtx
);
563 while (db
->db_state
== DB_READ
||
564 db
->db_state
== DB_FILL
)
565 cv_wait(&db
->db_changed
, &db
->db_mtx
);
566 if (db
->db_state
== DB_UNCACHED
)
567 err
= SET_ERROR(EIO
);
568 mutex_exit(&db
->db_mtx
);
570 dmu_buf_rele_array(dbp
, nblks
, tag
);
582 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
583 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
588 err
= dnode_hold(os
, object
, FTAG
, &dn
);
592 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
593 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
595 dnode_rele(dn
, FTAG
);
601 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
602 uint64_t length
, boolean_t read
, void *tag
, int *numbufsp
,
605 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
611 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
612 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
619 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
622 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
627 for (i
= 0; i
< numbufs
; i
++) {
629 dbuf_rele(dbp
[i
], tag
);
632 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
636 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
637 * indirect blocks prefeteched will be those that point to the blocks containing
638 * the data starting at offset, and continuing to offset + len.
640 * Note that if the indirect blocks above the blocks being prefetched are not
641 * in cache, they will be asychronously read in.
644 dmu_prefetch(objset_t
*os
, uint64_t object
, int64_t level
, uint64_t offset
,
645 uint64_t len
, zio_priority_t pri
)
651 if (len
== 0) { /* they're interested in the bonus buffer */
652 dn
= DMU_META_DNODE(os
);
654 if (object
== 0 || object
>= DN_MAX_OBJECT
)
657 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
658 blkid
= dbuf_whichblock(dn
, level
,
659 object
* sizeof (dnode_phys_t
));
660 dbuf_prefetch(dn
, level
, blkid
, pri
, 0);
661 rw_exit(&dn
->dn_struct_rwlock
);
666 * XXX - Note, if the dnode for the requested object is not
667 * already cached, we will do a *synchronous* read in the
668 * dnode_hold() call. The same is true for any indirects.
670 err
= dnode_hold(os
, object
, FTAG
, &dn
);
674 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
676 * offset + len - 1 is the last byte we want to prefetch for, and offset
677 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
678 * last block we want to prefetch, and dbuf_whichblock(dn, level,
679 * offset) is the first. Then the number we need to prefetch is the
682 if (level
> 0 || dn
->dn_datablkshift
!= 0) {
683 nblks
= dbuf_whichblock(dn
, level
, offset
+ len
- 1) -
684 dbuf_whichblock(dn
, level
, offset
) + 1;
686 nblks
= (offset
< dn
->dn_datablksz
);
690 blkid
= dbuf_whichblock(dn
, level
, offset
);
691 for (int i
= 0; i
< nblks
; i
++)
692 dbuf_prefetch(dn
, level
, blkid
+ i
, pri
, 0);
695 rw_exit(&dn
->dn_struct_rwlock
);
697 dnode_rele(dn
, FTAG
);
701 * Get the next "chunk" of file data to free. We traverse the file from
702 * the end so that the file gets shorter over time (if we crashes in the
703 * middle, this will leave us in a better state). We find allocated file
704 * data by simply searching the allocated level 1 indirects.
706 * On input, *start should be the first offset that does not need to be
707 * freed (e.g. "offset + length"). On return, *start will be the first
708 * offset that should be freed.
711 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
713 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
714 /* bytes of data covered by a level-1 indirect block */
716 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
718 ASSERT3U(minimum
, <=, *start
);
720 if (*start
- minimum
<= iblkrange
* maxblks
) {
724 ASSERT(ISP2(iblkrange
));
726 for (uint64_t blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
730 * dnode_next_offset(BACKWARDS) will find an allocated L1
731 * indirect block at or before the input offset. We must
732 * decrement *start so that it is at the end of the region
736 err
= dnode_next_offset(dn
,
737 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
739 /* if there are no indirect blocks before start, we are done */
743 } else if (err
!= 0) {
747 /* set start to the beginning of this L1 indirect */
748 *start
= P2ALIGN(*start
, iblkrange
);
750 if (*start
< minimum
)
756 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
757 * otherwise return false.
758 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
762 dmu_objset_zfs_unmounting(objset_t
*os
)
765 if (dmu_objset_type(os
) == DMU_OST_ZFS
)
766 return (zfs_get_vfs_flag_unmounted(os
));
772 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
775 uint64_t object_size
;
777 uint64_t dirty_frees_threshold
;
778 dsl_pool_t
*dp
= dmu_objset_pool(os
);
781 return (SET_ERROR(EINVAL
));
783 object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
784 if (offset
>= object_size
)
787 if (zfs_per_txg_dirty_frees_percent
<= 100)
788 dirty_frees_threshold
=
789 zfs_per_txg_dirty_frees_percent
* zfs_dirty_data_max
/ 100;
791 dirty_frees_threshold
= zfs_dirty_data_max
/ 4;
793 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
794 length
= object_size
- offset
;
796 while (length
!= 0) {
797 uint64_t chunk_end
, chunk_begin
, chunk_len
;
798 uint64_t long_free_dirty_all_txgs
= 0;
801 if (dmu_objset_zfs_unmounting(dn
->dn_objset
))
802 return (SET_ERROR(EINTR
));
804 chunk_end
= chunk_begin
= offset
+ length
;
806 /* move chunk_begin backwards to the beginning of this chunk */
807 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
810 ASSERT3U(chunk_begin
, >=, offset
);
811 ASSERT3U(chunk_begin
, <=, chunk_end
);
813 chunk_len
= chunk_end
- chunk_begin
;
815 mutex_enter(&dp
->dp_lock
);
816 for (int t
= 0; t
< TXG_SIZE
; t
++) {
817 long_free_dirty_all_txgs
+=
818 dp
->dp_long_free_dirty_pertxg
[t
];
820 mutex_exit(&dp
->dp_lock
);
823 * To avoid filling up a TXG with just frees wait for
824 * the next TXG to open before freeing more chunks if
825 * we have reached the threshold of frees
827 if (dirty_frees_threshold
!= 0 &&
828 long_free_dirty_all_txgs
>= dirty_frees_threshold
) {
829 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay
);
830 txg_wait_open(dp
, 0);
834 tx
= dmu_tx_create(os
);
835 dmu_tx_hold_free(tx
, dn
->dn_object
, chunk_begin
, chunk_len
);
838 * Mark this transaction as typically resulting in a net
839 * reduction in space used.
841 dmu_tx_mark_netfree(tx
);
842 err
= dmu_tx_assign(tx
, TXG_WAIT
);
848 mutex_enter(&dp
->dp_lock
);
849 dp
->dp_long_free_dirty_pertxg
[dmu_tx_get_txg(tx
) & TXG_MASK
] +=
851 mutex_exit(&dp
->dp_lock
);
852 DTRACE_PROBE3(free__long__range
,
853 uint64_t, long_free_dirty_all_txgs
, uint64_t, chunk_len
,
854 uint64_t, dmu_tx_get_txg(tx
));
855 dnode_free_range(dn
, chunk_begin
, chunk_len
, tx
);
865 dmu_free_long_range(objset_t
*os
, uint64_t object
,
866 uint64_t offset
, uint64_t length
)
871 err
= dnode_hold(os
, object
, FTAG
, &dn
);
874 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
877 * It is important to zero out the maxblkid when freeing the entire
878 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
879 * will take the fast path, and (b) dnode_reallocate() can verify
880 * that the entire file has been freed.
882 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
885 dnode_rele(dn
, FTAG
);
890 dmu_free_long_object(objset_t
*os
, uint64_t object
)
895 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
899 tx
= dmu_tx_create(os
);
900 dmu_tx_hold_bonus(tx
, object
);
901 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
902 dmu_tx_mark_netfree(tx
);
903 err
= dmu_tx_assign(tx
, TXG_WAIT
);
906 err
= dmu_object_free(os
, object
, tx
);
917 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
918 uint64_t size
, dmu_tx_t
*tx
)
921 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
924 ASSERT(offset
< UINT64_MAX
);
925 ASSERT(size
== DMU_OBJECT_END
|| size
<= UINT64_MAX
- offset
);
926 dnode_free_range(dn
, offset
, size
, tx
);
927 dnode_rele(dn
, FTAG
);
932 dmu_read_impl(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
933 void *buf
, uint32_t flags
)
936 int numbufs
, err
= 0;
939 * Deal with odd block sizes, where there can't be data past the first
940 * block. If we ever do the tail block optimization, we will need to
941 * handle that here as well.
943 if (dn
->dn_maxblkid
== 0) {
944 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
945 MIN(size
, dn
->dn_datablksz
- offset
);
946 bzero((char *)buf
+ newsz
, size
- newsz
);
951 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
955 * NB: we could do this block-at-a-time, but it's nice
956 * to be reading in parallel.
958 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
959 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
963 for (i
= 0; i
< numbufs
; i
++) {
966 dmu_buf_t
*db
= dbp
[i
];
970 bufoff
= offset
- db
->db_offset
;
971 tocpy
= MIN(db
->db_size
- bufoff
, size
);
973 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
977 buf
= (char *)buf
+ tocpy
;
979 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
985 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
986 void *buf
, uint32_t flags
)
991 err
= dnode_hold(os
, object
, FTAG
, &dn
);
995 err
= dmu_read_impl(dn
, offset
, size
, buf
, flags
);
996 dnode_rele(dn
, FTAG
);
1001 dmu_read_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
, void *buf
,
1004 return (dmu_read_impl(dn
, offset
, size
, buf
, flags
));
1008 dmu_write_impl(dmu_buf_t
**dbp
, int numbufs
, uint64_t offset
, uint64_t size
,
1009 const void *buf
, dmu_tx_t
*tx
)
1013 for (i
= 0; i
< numbufs
; i
++) {
1016 dmu_buf_t
*db
= dbp
[i
];
1020 bufoff
= offset
- db
->db_offset
;
1021 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1023 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1025 if (tocpy
== db
->db_size
)
1026 dmu_buf_will_fill(db
, tx
);
1028 dmu_buf_will_dirty(db
, tx
);
1030 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
1032 if (tocpy
== db
->db_size
)
1033 dmu_buf_fill_done(db
, tx
);
1037 buf
= (char *)buf
+ tocpy
;
1042 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1043 const void *buf
, dmu_tx_t
*tx
)
1051 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
1052 FALSE
, FTAG
, &numbufs
, &dbp
));
1053 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
1054 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1058 dmu_write_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
1059 const void *buf
, dmu_tx_t
*tx
)
1067 VERIFY0(dmu_buf_hold_array_by_dnode(dn
, offset
, size
,
1068 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
));
1069 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
1070 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1074 dmu_object_remap_one_indirect(objset_t
*os
, dnode_t
*dn
,
1075 uint64_t last_removal_txg
, uint64_t offset
)
1077 uint64_t l1blkid
= dbuf_whichblock(dn
, 1, offset
);
1080 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1081 dmu_buf_impl_t
*dbuf
= dbuf_hold_level(dn
, 1, l1blkid
, FTAG
);
1082 ASSERT3P(dbuf
, !=, NULL
);
1085 * If the block hasn't been written yet, this default will ensure
1086 * we don't try to remap it.
1088 uint64_t birth
= UINT64_MAX
;
1089 ASSERT3U(last_removal_txg
, !=, UINT64_MAX
);
1090 if (dbuf
->db_blkptr
!= NULL
)
1091 birth
= dbuf
->db_blkptr
->blk_birth
;
1092 rw_exit(&dn
->dn_struct_rwlock
);
1095 * If this L1 was already written after the last removal, then we've
1096 * already tried to remap it.
1098 if (birth
<= last_removal_txg
&&
1099 dbuf_read(dbuf
, NULL
, DB_RF_MUST_SUCCEED
) == 0 &&
1100 dbuf_can_remap(dbuf
)) {
1101 dmu_tx_t
*tx
= dmu_tx_create(os
);
1102 dmu_tx_hold_remap_l1indirect(tx
, dn
->dn_object
);
1103 err
= dmu_tx_assign(tx
, TXG_WAIT
);
1105 (void) dbuf_dirty(dbuf
, tx
);
1112 dbuf_rele(dbuf
, FTAG
);
1114 delay(zfs_object_remap_one_indirect_delay_ticks
);
1120 * Remap all blockpointers in the object, if possible, so that they reference
1121 * only concrete vdevs.
1123 * To do this, iterate over the L0 blockpointers and remap any that reference
1124 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1125 * cannot guarantee that we can remap all blockpointer anyways (due to split
1126 * blocks), we do not want to make the code unnecessarily complicated to
1127 * catch the unlikely case that there is an L1 block on an indirect vdev that
1128 * contains no indirect blockpointers.
1131 dmu_object_remap_indirects(objset_t
*os
, uint64_t object
,
1132 uint64_t last_removal_txg
)
1134 uint64_t offset
, l1span
;
1138 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1143 if (dn
->dn_nlevels
<= 1) {
1144 if (issig(JUSTLOOKING
) && issig(FORREAL
)) {
1145 err
= SET_ERROR(EINTR
);
1149 * If the dnode has no indirect blocks, we cannot dirty them.
1150 * We still want to remap the blkptr(s) in the dnode if
1151 * appropriate, so mark it as dirty.
1153 if (err
== 0 && dnode_needs_remap(dn
)) {
1154 dmu_tx_t
*tx
= dmu_tx_create(os
);
1155 dmu_tx_hold_bonus(tx
, dn
->dn_object
);
1156 if ((err
= dmu_tx_assign(tx
, TXG_WAIT
)) == 0) {
1157 dnode_setdirty(dn
, tx
);
1164 dnode_rele(dn
, FTAG
);
1169 l1span
= 1ULL << (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
+
1170 dn
->dn_datablkshift
);
1172 * Find the next L1 indirect that is not a hole.
1174 while (dnode_next_offset(dn
, 0, &offset
, 2, 1, 0) == 0) {
1175 if (issig(JUSTLOOKING
) && issig(FORREAL
)) {
1176 err
= SET_ERROR(EINTR
);
1179 if ((err
= dmu_object_remap_one_indirect(os
, dn
,
1180 last_removal_txg
, offset
)) != 0) {
1186 dnode_rele(dn
, FTAG
);
1191 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1200 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
1201 FALSE
, FTAG
, &numbufs
, &dbp
));
1203 for (i
= 0; i
< numbufs
; i
++) {
1204 dmu_buf_t
*db
= dbp
[i
];
1206 dmu_buf_will_not_fill(db
, tx
);
1208 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1212 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
1213 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
1214 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
1218 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
1219 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
1220 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
1223 dmu_buf_write_embedded(db
,
1224 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
1225 uncompressed_size
, compressed_size
, byteorder
, tx
);
1227 dmu_buf_rele(db
, FTAG
);
1231 * DMU support for xuio
1233 kstat_t
*xuio_ksp
= NULL
;
1235 typedef struct xuio_stats
{
1236 /* loaned yet not returned arc_buf */
1237 kstat_named_t xuiostat_onloan_rbuf
;
1238 kstat_named_t xuiostat_onloan_wbuf
;
1239 /* whether a copy is made when loaning out a read buffer */
1240 kstat_named_t xuiostat_rbuf_copied
;
1241 kstat_named_t xuiostat_rbuf_nocopy
;
1242 /* whether a copy is made when assigning a write buffer */
1243 kstat_named_t xuiostat_wbuf_copied
;
1244 kstat_named_t xuiostat_wbuf_nocopy
;
1247 static xuio_stats_t xuio_stats
= {
1248 { "onloan_read_buf", KSTAT_DATA_UINT64
},
1249 { "onloan_write_buf", KSTAT_DATA_UINT64
},
1250 { "read_buf_copied", KSTAT_DATA_UINT64
},
1251 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
1252 { "write_buf_copied", KSTAT_DATA_UINT64
},
1253 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
1256 #define XUIOSTAT_INCR(stat, val) \
1257 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
1258 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
1260 #ifdef HAVE_UIO_ZEROCOPY
1262 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
1265 uio_t
*uio
= &xuio
->xu_uio
;
1267 uio
->uio_iovcnt
= nblk
;
1268 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
1270 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
1272 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
1273 priv
->iovp
= (iovec_t
*)uio
->uio_iov
;
1274 XUIO_XUZC_PRIV(xuio
) = priv
;
1276 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1277 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
1279 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
1285 dmu_xuio_fini(xuio_t
*xuio
)
1287 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1288 int nblk
= priv
->cnt
;
1290 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
1291 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
1292 kmem_free(priv
, sizeof (dmu_xuio_t
));
1294 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1295 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
1297 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
1301 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1302 * and increase priv->next by 1.
1305 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
1308 uio_t
*uio
= &xuio
->xu_uio
;
1309 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1310 int i
= priv
->next
++;
1312 ASSERT(i
< priv
->cnt
);
1313 ASSERT(off
+ n
<= arc_buf_lsize(abuf
));
1314 iov
= (iovec_t
*)uio
->uio_iov
+ i
;
1315 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
1317 priv
->bufs
[i
] = abuf
;
1322 dmu_xuio_cnt(xuio_t
*xuio
)
1324 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1329 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
1331 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1333 ASSERT(i
< priv
->cnt
);
1334 return (priv
->bufs
[i
]);
1338 dmu_xuio_clear(xuio_t
*xuio
, int i
)
1340 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1342 ASSERT(i
< priv
->cnt
);
1343 priv
->bufs
[i
] = NULL
;
1345 #endif /* HAVE_UIO_ZEROCOPY */
1348 xuio_stat_init(void)
1350 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1351 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1352 KSTAT_FLAG_VIRTUAL
);
1353 if (xuio_ksp
!= NULL
) {
1354 xuio_ksp
->ks_data
= &xuio_stats
;
1355 kstat_install(xuio_ksp
);
1360 xuio_stat_fini(void)
1362 if (xuio_ksp
!= NULL
) {
1363 kstat_delete(xuio_ksp
);
1369 xuio_stat_wbuf_copied(void)
1371 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1375 xuio_stat_wbuf_nocopy(void)
1377 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1382 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1385 int numbufs
, i
, err
;
1386 #ifdef HAVE_UIO_ZEROCOPY
1387 xuio_t
*xuio
= NULL
;
1391 * NB: we could do this block-at-a-time, but it's nice
1392 * to be reading in parallel.
1394 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1395 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1399 for (i
= 0; i
< numbufs
; i
++) {
1402 dmu_buf_t
*db
= dbp
[i
];
1406 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1407 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1409 #ifdef HAVE_UIO_ZEROCOPY
1411 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1412 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1413 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1414 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1416 uio
->uio_resid
-= tocpy
;
1417 uio
->uio_loffset
+= tocpy
;
1420 if (abuf
== dbuf_abuf
)
1421 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1423 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1426 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1433 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1439 * Read 'size' bytes into the uio buffer.
1440 * From object zdb->db_object.
1441 * Starting at offset uio->uio_loffset.
1443 * If the caller already has a dbuf in the target object
1444 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1445 * because we don't have to find the dnode_t for the object.
1448 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1450 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1459 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1466 * Read 'size' bytes into the uio buffer.
1467 * From the specified object
1468 * Starting at offset uio->uio_loffset.
1471 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1479 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1483 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1485 dnode_rele(dn
, FTAG
);
1491 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1498 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1499 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1503 for (i
= 0; i
< numbufs
; i
++) {
1506 dmu_buf_t
*db
= dbp
[i
];
1510 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1511 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1513 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1515 if (tocpy
== db
->db_size
)
1516 dmu_buf_will_fill(db
, tx
);
1518 dmu_buf_will_dirty(db
, tx
);
1521 * XXX uiomove could block forever (eg.nfs-backed
1522 * pages). There needs to be a uiolockdown() function
1523 * to lock the pages in memory, so that uiomove won't
1526 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1529 if (tocpy
== db
->db_size
)
1530 dmu_buf_fill_done(db
, tx
);
1538 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1543 * Write 'size' bytes from the uio buffer.
1544 * To object zdb->db_object.
1545 * Starting at offset uio->uio_loffset.
1547 * If the caller already has a dbuf in the target object
1548 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1549 * because we don't have to find the dnode_t for the object.
1552 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1555 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1564 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1571 * Write 'size' bytes from the uio buffer.
1572 * To the specified object.
1573 * Starting at offset uio->uio_loffset.
1576 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1585 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1589 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1591 dnode_rele(dn
, FTAG
);
1595 #endif /* _KERNEL */
1598 * Allocate a loaned anonymous arc buffer.
1601 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1603 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1605 return (arc_loan_buf(db
->db_objset
->os_spa
, B_FALSE
, size
));
1609 * Free a loaned arc buffer.
1612 dmu_return_arcbuf(arc_buf_t
*buf
)
1614 arc_return_buf(buf
, FTAG
);
1615 arc_buf_destroy(buf
, FTAG
);
1619 dmu_copy_from_buf(objset_t
*os
, uint64_t object
, uint64_t offset
,
1620 dmu_buf_t
*handle
, dmu_tx_t
*tx
)
1622 dmu_buf_t
*dst_handle
;
1623 dmu_buf_impl_t
*dstdb
;
1624 dmu_buf_impl_t
*srcdb
= (dmu_buf_impl_t
*)handle
;
1625 dmu_object_type_t type
;
1628 boolean_t byteorder
;
1629 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1630 uint8_t iv
[ZIO_DATA_IV_LEN
];
1631 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1633 ASSERT3P(srcdb
->db_buf
, !=, NULL
);
1635 /* hold the db that we want to write to */
1636 VERIFY0(dmu_buf_hold(os
, object
, offset
, FTAG
, &dst_handle
,
1637 DMU_READ_NO_DECRYPT
));
1638 dstdb
= (dmu_buf_impl_t
*)dst_handle
;
1639 datalen
= arc_buf_size(srcdb
->db_buf
);
1641 DB_DNODE_ENTER(dstdb
);
1642 type
= DB_DNODE(dstdb
)->dn_type
;
1643 DB_DNODE_EXIT(dstdb
);
1645 /* allocated an arc buffer that matches the type of srcdb->db_buf */
1646 if (arc_is_encrypted(srcdb
->db_buf
)) {
1647 arc_get_raw_params(srcdb
->db_buf
, &byteorder
, salt
, iv
, mac
);
1648 abuf
= arc_loan_raw_buf(os
->os_spa
, dmu_objset_id(os
),
1649 byteorder
, salt
, iv
, mac
, type
,
1650 datalen
, arc_buf_lsize(srcdb
->db_buf
),
1651 arc_get_compression(srcdb
->db_buf
));
1653 /* we won't get a compressed db back from dmu_buf_hold() */
1654 ASSERT3U(arc_get_compression(srcdb
->db_buf
),
1655 ==, ZIO_COMPRESS_OFF
);
1656 abuf
= arc_loan_buf(os
->os_spa
,
1657 DMU_OT_IS_METADATA(type
), datalen
);
1660 ASSERT3U(datalen
, ==, arc_buf_size(abuf
));
1662 /* copy the data to the new buffer and assign it to the dstdb */
1663 bcopy(srcdb
->db_buf
->b_data
, abuf
->b_data
, datalen
);
1664 dbuf_assign_arcbuf(dstdb
, abuf
, tx
);
1665 dmu_buf_rele(dst_handle
, FTAG
);
1669 * When possible directly assign passed loaned arc buffer to a dbuf.
1670 * If this is not possible copy the contents of passed arc buf via
1674 dmu_assign_arcbuf_by_dnode(dnode_t
*dn
, uint64_t offset
, arc_buf_t
*buf
,
1678 objset_t
*os
= dn
->dn_objset
;
1679 uint64_t object
= dn
->dn_object
;
1680 uint32_t blksz
= (uint32_t)arc_buf_lsize(buf
);
1683 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1684 blkid
= dbuf_whichblock(dn
, 0, offset
);
1685 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1686 rw_exit(&dn
->dn_struct_rwlock
);
1689 * We can only assign if the offset is aligned, the arc buf is the
1690 * same size as the dbuf, and the dbuf is not metadata.
1692 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1693 dbuf_assign_arcbuf(db
, buf
, tx
);
1694 dbuf_rele(db
, FTAG
);
1696 /* compressed bufs must always be assignable to their dbuf */
1697 ASSERT3U(arc_get_compression(buf
), ==, ZIO_COMPRESS_OFF
);
1698 ASSERT(!(buf
->b_flags
& ARC_BUF_FLAG_COMPRESSED
));
1700 dbuf_rele(db
, FTAG
);
1701 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1702 dmu_return_arcbuf(buf
);
1703 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1708 dmu_assign_arcbuf_by_dbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1711 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1713 DB_DNODE_ENTER(dbuf
);
1714 dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf
), offset
, buf
, tx
);
1715 DB_DNODE_EXIT(dbuf
);
1719 dbuf_dirty_record_t
*dsa_dr
;
1720 dmu_sync_cb_t
*dsa_done
;
1727 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1729 dmu_sync_arg_t
*dsa
= varg
;
1730 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1731 blkptr_t
*bp
= zio
->io_bp
;
1733 if (zio
->io_error
== 0) {
1734 if (BP_IS_HOLE(bp
)) {
1736 * A block of zeros may compress to a hole, but the
1737 * block size still needs to be known for replay.
1739 BP_SET_LSIZE(bp
, db
->db_size
);
1740 } else if (!BP_IS_EMBEDDED(bp
)) {
1741 ASSERT(BP_GET_LEVEL(bp
) == 0);
1748 dmu_sync_late_arrival_ready(zio_t
*zio
)
1750 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1755 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1757 dmu_sync_arg_t
*dsa
= varg
;
1758 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1759 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1760 zgd_t
*zgd
= dsa
->dsa_zgd
;
1763 * Record the vdev(s) backing this blkptr so they can be flushed after
1764 * the writes for the lwb have completed.
1766 if (zio
->io_error
== 0) {
1767 zil_lwb_add_block(zgd
->zgd_lwb
, zgd
->zgd_bp
);
1770 mutex_enter(&db
->db_mtx
);
1771 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1772 if (zio
->io_error
== 0) {
1773 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1774 if (dr
->dt
.dl
.dr_nopwrite
) {
1775 blkptr_t
*bp
= zio
->io_bp
;
1776 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
1777 uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
);
1779 ASSERT(BP_EQUAL(bp
, bp_orig
));
1780 VERIFY(BP_EQUAL(bp
, db
->db_blkptr
));
1781 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1782 VERIFY(zio_checksum_table
[chksum
].ci_flags
&
1783 ZCHECKSUM_FLAG_NOPWRITE
);
1785 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1786 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1787 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1790 * Old style holes are filled with all zeros, whereas
1791 * new-style holes maintain their lsize, type, level,
1792 * and birth time (see zio_write_compress). While we
1793 * need to reset the BP_SET_LSIZE() call that happened
1794 * in dmu_sync_ready for old style holes, we do *not*
1795 * want to wipe out the information contained in new
1796 * style holes. Thus, only zero out the block pointer if
1797 * it's an old style hole.
1799 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1800 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1801 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1803 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1805 cv_broadcast(&db
->db_changed
);
1806 mutex_exit(&db
->db_mtx
);
1808 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1810 kmem_free(dsa
, sizeof (*dsa
));
1814 dmu_sync_late_arrival_done(zio_t
*zio
)
1816 blkptr_t
*bp
= zio
->io_bp
;
1817 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1818 zgd_t
*zgd
= dsa
->dsa_zgd
;
1820 if (zio
->io_error
== 0) {
1822 * Record the vdev(s) backing this blkptr so they can be
1823 * flushed after the writes for the lwb have completed.
1825 zil_lwb_add_block(zgd
->zgd_lwb
, zgd
->zgd_bp
);
1827 if (!BP_IS_HOLE(bp
)) {
1828 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1829 ASSERT(!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
1830 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1831 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1832 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1833 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1837 dmu_tx_commit(dsa
->dsa_tx
);
1839 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1841 abd_put(zio
->io_abd
);
1842 kmem_free(dsa
, sizeof (*dsa
));
1846 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1847 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1849 dmu_sync_arg_t
*dsa
;
1852 tx
= dmu_tx_create(os
);
1853 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1854 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1856 /* Make zl_get_data do txg_waited_synced() */
1857 return (SET_ERROR(EIO
));
1861 * In order to prevent the zgd's lwb from being free'd prior to
1862 * dmu_sync_late_arrival_done() being called, we have to ensure
1863 * the lwb's "max txg" takes this tx's txg into account.
1865 zil_lwb_add_txg(zgd
->zgd_lwb
, dmu_tx_get_txg(tx
));
1867 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1869 dsa
->dsa_done
= done
;
1874 * Since we are currently syncing this txg, it's nontrivial to
1875 * determine what BP to nopwrite against, so we disable nopwrite.
1877 * When syncing, the db_blkptr is initially the BP of the previous
1878 * txg. We can not nopwrite against it because it will be changed
1879 * (this is similar to the non-late-arrival case where the dbuf is
1880 * dirty in a future txg).
1882 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1883 * We can not nopwrite against it because although the BP will not
1884 * (typically) be changed, the data has not yet been persisted to this
1887 * Finally, when dbuf_write_done() is called, it is theoretically
1888 * possible to always nopwrite, because the data that was written in
1889 * this txg is the same data that we are trying to write. However we
1890 * would need to check that this dbuf is not dirty in any future
1891 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1892 * don't nopwrite in this case.
1894 zp
->zp_nopwrite
= B_FALSE
;
1896 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1897 abd_get_from_buf(zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
),
1898 zgd
->zgd_db
->db_size
, zgd
->zgd_db
->db_size
, zp
,
1899 dmu_sync_late_arrival_ready
, NULL
, NULL
, dmu_sync_late_arrival_done
,
1900 dsa
, ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, zb
));
1906 * Intent log support: sync the block associated with db to disk.
1907 * N.B. and XXX: the caller is responsible for making sure that the
1908 * data isn't changing while dmu_sync() is writing it.
1912 * EEXIST: this txg has already been synced, so there's nothing to do.
1913 * The caller should not log the write.
1915 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1916 * The caller should not log the write.
1918 * EALREADY: this block is already in the process of being synced.
1919 * The caller should track its progress (somehow).
1921 * EIO: could not do the I/O.
1922 * The caller should do a txg_wait_synced().
1924 * 0: the I/O has been initiated.
1925 * The caller should log this blkptr in the done callback.
1926 * It is possible that the I/O will fail, in which case
1927 * the error will be reported to the done callback and
1928 * propagated to pio from zio_done().
1931 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1933 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1934 objset_t
*os
= db
->db_objset
;
1935 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1936 dbuf_dirty_record_t
*dr
;
1937 dmu_sync_arg_t
*dsa
;
1938 zbookmark_phys_t zb
;
1942 ASSERT(pio
!= NULL
);
1945 SET_BOOKMARK(&zb
, ds
->ds_object
,
1946 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1950 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1954 * If we're frozen (running ziltest), we always need to generate a bp.
1956 if (txg
> spa_freeze_txg(os
->os_spa
))
1957 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1960 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1961 * and us. If we determine that this txg is not yet syncing,
1962 * but it begins to sync a moment later, that's OK because the
1963 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1965 mutex_enter(&db
->db_mtx
);
1967 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1969 * This txg has already synced. There's nothing to do.
1971 mutex_exit(&db
->db_mtx
);
1972 return (SET_ERROR(EEXIST
));
1975 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1977 * This txg is currently syncing, so we can't mess with
1978 * the dirty record anymore; just write a new log block.
1980 mutex_exit(&db
->db_mtx
);
1981 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1984 dr
= db
->db_last_dirty
;
1985 while (dr
&& dr
->dr_txg
!= txg
)
1990 * There's no dr for this dbuf, so it must have been freed.
1991 * There's no need to log writes to freed blocks, so we're done.
1993 mutex_exit(&db
->db_mtx
);
1994 return (SET_ERROR(ENOENT
));
1997 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1999 if (db
->db_blkptr
!= NULL
) {
2001 * We need to fill in zgd_bp with the current blkptr so that
2002 * the nopwrite code can check if we're writing the same
2003 * data that's already on disk. We can only nopwrite if we
2004 * are sure that after making the copy, db_blkptr will not
2005 * change until our i/o completes. We ensure this by
2006 * holding the db_mtx, and only allowing nopwrite if the
2007 * block is not already dirty (see below). This is verified
2008 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2011 *zgd
->zgd_bp
= *db
->db_blkptr
;
2015 * Assume the on-disk data is X, the current syncing data (in
2016 * txg - 1) is Y, and the current in-memory data is Z (currently
2019 * We usually want to perform a nopwrite if X and Z are the
2020 * same. However, if Y is different (i.e. the BP is going to
2021 * change before this write takes effect), then a nopwrite will
2022 * be incorrect - we would override with X, which could have
2023 * been freed when Y was written.
2025 * (Note that this is not a concern when we are nop-writing from
2026 * syncing context, because X and Y must be identical, because
2027 * all previous txgs have been synced.)
2029 * Therefore, we disable nopwrite if the current BP could change
2030 * before this TXG. There are two ways it could change: by
2031 * being dirty (dr_next is non-NULL), or by being freed
2032 * (dnode_block_freed()). This behavior is verified by
2033 * zio_done(), which VERIFYs that the override BP is identical
2034 * to the on-disk BP.
2038 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
2039 zp
.zp_nopwrite
= B_FALSE
;
2042 ASSERT(dr
->dr_txg
== txg
);
2043 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
2044 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
2046 * We have already issued a sync write for this buffer,
2047 * or this buffer has already been synced. It could not
2048 * have been dirtied since, or we would have cleared the state.
2050 mutex_exit(&db
->db_mtx
);
2051 return (SET_ERROR(EALREADY
));
2054 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
2055 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
2056 mutex_exit(&db
->db_mtx
);
2058 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
2060 dsa
->dsa_done
= done
;
2064 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
2065 zgd
->zgd_bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
2066 &zp
, dmu_sync_ready
, NULL
, NULL
, dmu_sync_done
, dsa
,
2067 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, &zb
));
2073 dmu_object_set_nlevels(objset_t
*os
, uint64_t object
, int nlevels
, dmu_tx_t
*tx
)
2078 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2081 err
= dnode_set_nlevels(dn
, nlevels
, tx
);
2082 dnode_rele(dn
, FTAG
);
2087 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
2093 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2096 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
2097 dnode_rele(dn
, FTAG
);
2102 dmu_object_set_maxblkid(objset_t
*os
, uint64_t object
, uint64_t maxblkid
,
2108 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2111 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
2112 dnode_new_blkid(dn
, maxblkid
, tx
, B_FALSE
);
2113 rw_exit(&dn
->dn_struct_rwlock
);
2114 dnode_rele(dn
, FTAG
);
2119 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
2125 * Send streams include each object's checksum function. This
2126 * check ensures that the receiving system can understand the
2127 * checksum function transmitted.
2129 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
2131 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
2132 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
2133 dn
->dn_checksum
= checksum
;
2134 dnode_setdirty(dn
, tx
);
2135 dnode_rele(dn
, FTAG
);
2139 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
2145 * Send streams include each object's compression function. This
2146 * check ensures that the receiving system can understand the
2147 * compression function transmitted.
2149 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
2151 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
2152 dn
->dn_compress
= compress
;
2153 dnode_setdirty(dn
, tx
);
2154 dnode_rele(dn
, FTAG
);
2158 * When the "redundant_metadata" property is set to "most", only indirect
2159 * blocks of this level and higher will have an additional ditto block.
2161 int zfs_redundant_metadata_most_ditto_level
= 2;
2164 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
2166 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
2167 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
2169 enum zio_checksum checksum
= os
->os_checksum
;
2170 enum zio_compress compress
= os
->os_compress
;
2171 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
2172 boolean_t dedup
= B_FALSE
;
2173 boolean_t nopwrite
= B_FALSE
;
2174 boolean_t dedup_verify
= os
->os_dedup_verify
;
2175 boolean_t encrypt
= B_FALSE
;
2176 int copies
= os
->os_copies
;
2179 * We maintain different write policies for each of the following
2182 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2183 * 3. all other level 0 blocks
2187 * XXX -- we should design a compression algorithm
2188 * that specializes in arrays of bps.
2190 compress
= zio_compress_select(os
->os_spa
,
2191 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
2194 * Metadata always gets checksummed. If the data
2195 * checksum is multi-bit correctable, and it's not a
2196 * ZBT-style checksum, then it's suitable for metadata
2197 * as well. Otherwise, the metadata checksum defaults
2200 if (!(zio_checksum_table
[checksum
].ci_flags
&
2201 ZCHECKSUM_FLAG_METADATA
) ||
2202 (zio_checksum_table
[checksum
].ci_flags
&
2203 ZCHECKSUM_FLAG_EMBEDDED
))
2204 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
2206 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
2207 (os
->os_redundant_metadata
==
2208 ZFS_REDUNDANT_METADATA_MOST
&&
2209 (level
>= zfs_redundant_metadata_most_ditto_level
||
2210 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
2212 } else if (wp
& WP_NOFILL
) {
2216 * If we're writing preallocated blocks, we aren't actually
2217 * writing them so don't set any policy properties. These
2218 * blocks are currently only used by an external subsystem
2219 * outside of zfs (i.e. dump) and not written by the zio
2222 compress
= ZIO_COMPRESS_OFF
;
2223 checksum
= ZIO_CHECKSUM_OFF
;
2225 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
2228 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
2229 zio_checksum_select(dn
->dn_checksum
, checksum
) :
2233 * Determine dedup setting. If we are in dmu_sync(),
2234 * we won't actually dedup now because that's all
2235 * done in syncing context; but we do want to use the
2236 * dedup checkum. If the checksum is not strong
2237 * enough to ensure unique signatures, force
2240 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
2241 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
2242 if (!(zio_checksum_table
[checksum
].ci_flags
&
2243 ZCHECKSUM_FLAG_DEDUP
))
2244 dedup_verify
= B_TRUE
;
2248 * Enable nopwrite if we have secure enough checksum
2249 * algorithm (see comment in zio_nop_write) and
2250 * compression is enabled. We don't enable nopwrite if
2251 * dedup is enabled as the two features are mutually
2254 nopwrite
= (!dedup
&& (zio_checksum_table
[checksum
].ci_flags
&
2255 ZCHECKSUM_FLAG_NOPWRITE
) &&
2256 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
2260 * All objects in an encrypted objset are protected from modification
2261 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2262 * in the bp, so we cannot use all copies. Encrypted objects are also
2263 * not subject to nopwrite since writing the same data will still
2264 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2265 * to avoid ambiguity in the dedup code since the DDT does not store
2268 if (os
->os_encrypted
&& (wp
& WP_NOFILL
) == 0) {
2271 if (DMU_OT_IS_ENCRYPTED(type
)) {
2272 copies
= MIN(copies
, SPA_DVAS_PER_BP
- 1);
2279 (type
== DMU_OT_DNODE
|| type
== DMU_OT_OBJSET
)) {
2280 compress
= ZIO_COMPRESS_EMPTY
;
2284 zp
->zp_compress
= compress
;
2285 zp
->zp_checksum
= checksum
;
2286 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
2287 zp
->zp_level
= level
;
2288 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
2289 zp
->zp_dedup
= dedup
;
2290 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
2291 zp
->zp_nopwrite
= nopwrite
;
2292 zp
->zp_encrypt
= encrypt
;
2293 zp
->zp_byteorder
= ZFS_HOST_BYTEORDER
;
2294 bzero(zp
->zp_salt
, ZIO_DATA_SALT_LEN
);
2295 bzero(zp
->zp_iv
, ZIO_DATA_IV_LEN
);
2296 bzero(zp
->zp_mac
, ZIO_DATA_MAC_LEN
);
2297 zp
->zp_zpl_smallblk
= DMU_OT_IS_FILE(zp
->zp_type
) ?
2298 os
->os_zpl_special_smallblock
: 0;
2300 ASSERT3U(zp
->zp_compress
, !=, ZIO_COMPRESS_INHERIT
);
2304 * This function is only called from zfs_holey_common() for zpl_llseek()
2305 * in order to determine the location of holes. In order to accurately
2306 * report holes all dirty data must be synced to disk. This causes extremely
2307 * poor performance when seeking for holes in a dirty file. As a compromise,
2308 * only provide hole data when the dnode is clean. When a dnode is dirty
2309 * report the dnode as having no holes which is always a safe thing to do.
2312 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
2316 boolean_t clean
= B_TRUE
;
2318 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2323 * Check if dnode is dirty
2325 for (i
= 0; i
< TXG_SIZE
; i
++) {
2326 if (multilist_link_active(&dn
->dn_dirty_link
[i
])) {
2333 * If compatibility option is on, sync any current changes before
2334 * we go trundling through the block pointers.
2336 if (!clean
&& zfs_dmu_offset_next_sync
) {
2338 dnode_rele(dn
, FTAG
);
2339 txg_wait_synced(dmu_objset_pool(os
), 0);
2340 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2346 err
= dnode_next_offset(dn
,
2347 (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
2349 err
= SET_ERROR(EBUSY
);
2351 dnode_rele(dn
, FTAG
);
2357 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2359 dnode_phys_t
*dnp
= dn
->dn_phys
;
2361 doi
->doi_data_block_size
= dn
->dn_datablksz
;
2362 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
2363 1ULL << dn
->dn_indblkshift
: 0;
2364 doi
->doi_type
= dn
->dn_type
;
2365 doi
->doi_bonus_type
= dn
->dn_bonustype
;
2366 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
2367 doi
->doi_dnodesize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
2368 doi
->doi_indirection
= dn
->dn_nlevels
;
2369 doi
->doi_checksum
= dn
->dn_checksum
;
2370 doi
->doi_compress
= dn
->dn_compress
;
2371 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
2372 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
2373 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
2374 doi
->doi_fill_count
= 0;
2375 for (int i
= 0; i
< dnp
->dn_nblkptr
; i
++)
2376 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
2380 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2382 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2383 mutex_enter(&dn
->dn_mtx
);
2385 __dmu_object_info_from_dnode(dn
, doi
);
2387 mutex_exit(&dn
->dn_mtx
);
2388 rw_exit(&dn
->dn_struct_rwlock
);
2392 * Get information on a DMU object.
2393 * If doi is NULL, just indicates whether the object exists.
2396 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
2399 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
2405 dmu_object_info_from_dnode(dn
, doi
);
2407 dnode_rele(dn
, FTAG
);
2412 * As above, but faster; can be used when you have a held dbuf in hand.
2415 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
2417 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2420 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
2425 * Faster still when you only care about the size.
2426 * This is specifically optimized for zfs_getattr().
2429 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
2430 u_longlong_t
*nblk512
)
2432 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2438 *blksize
= dn
->dn_datablksz
;
2439 /* add in number of slots used for the dnode itself */
2440 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
2441 SPA_MINBLOCKSHIFT
) + dn
->dn_num_slots
;
2446 dmu_object_dnsize_from_db(dmu_buf_t
*db_fake
, int *dnsize
)
2448 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2453 *dnsize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
2458 byteswap_uint64_array(void *vbuf
, size_t size
)
2460 uint64_t *buf
= vbuf
;
2461 size_t count
= size
>> 3;
2464 ASSERT((size
& 7) == 0);
2466 for (i
= 0; i
< count
; i
++)
2467 buf
[i
] = BSWAP_64(buf
[i
]);
2471 byteswap_uint32_array(void *vbuf
, size_t size
)
2473 uint32_t *buf
= vbuf
;
2474 size_t count
= size
>> 2;
2477 ASSERT((size
& 3) == 0);
2479 for (i
= 0; i
< count
; i
++)
2480 buf
[i
] = BSWAP_32(buf
[i
]);
2484 byteswap_uint16_array(void *vbuf
, size_t size
)
2486 uint16_t *buf
= vbuf
;
2487 size_t count
= size
>> 1;
2490 ASSERT((size
& 1) == 0);
2492 for (i
= 0; i
< count
; i
++)
2493 buf
[i
] = BSWAP_16(buf
[i
]);
2498 byteswap_uint8_array(void *vbuf
, size_t size
)
2521 arc_fini(); /* arc depends on l2arc, so arc must go first */
2534 #if defined(_KERNEL)
2535 EXPORT_SYMBOL(dmu_bonus_hold
);
2536 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2537 EXPORT_SYMBOL(dmu_buf_rele_array
);
2538 EXPORT_SYMBOL(dmu_prefetch
);
2539 EXPORT_SYMBOL(dmu_free_range
);
2540 EXPORT_SYMBOL(dmu_free_long_range
);
2541 EXPORT_SYMBOL(dmu_free_long_object
);
2542 EXPORT_SYMBOL(dmu_read
);
2543 EXPORT_SYMBOL(dmu_read_by_dnode
);
2544 EXPORT_SYMBOL(dmu_write
);
2545 EXPORT_SYMBOL(dmu_write_by_dnode
);
2546 EXPORT_SYMBOL(dmu_prealloc
);
2547 EXPORT_SYMBOL(dmu_object_info
);
2548 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2549 EXPORT_SYMBOL(dmu_object_info_from_db
);
2550 EXPORT_SYMBOL(dmu_object_size_from_db
);
2551 EXPORT_SYMBOL(dmu_object_dnsize_from_db
);
2552 EXPORT_SYMBOL(dmu_object_set_nlevels
);
2553 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2554 EXPORT_SYMBOL(dmu_object_set_maxblkid
);
2555 EXPORT_SYMBOL(dmu_object_set_checksum
);
2556 EXPORT_SYMBOL(dmu_object_set_compress
);
2557 EXPORT_SYMBOL(dmu_write_policy
);
2558 EXPORT_SYMBOL(dmu_sync
);
2559 EXPORT_SYMBOL(dmu_request_arcbuf
);
2560 EXPORT_SYMBOL(dmu_return_arcbuf
);
2561 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode
);
2562 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf
);
2563 EXPORT_SYMBOL(dmu_buf_hold
);
2564 EXPORT_SYMBOL(dmu_ot
);
2567 module_param(zfs_nopwrite_enabled
, int, 0644);
2568 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");
2570 module_param(zfs_per_txg_dirty_frees_percent
, ulong
, 0644);
2571 MODULE_PARM_DESC(zfs_per_txg_dirty_frees_percent
,
2572 "percentage of dirtied blocks from frees in one TXG");
2574 module_param(zfs_dmu_offset_next_sync
, int, 0644);
2575 MODULE_PARM_DESC(zfs_dmu_offset_next_sync
,
2576 "Enable forcing txg sync to find holes");