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 https://opensource.org/licenses/CDDL-1.0.
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, 2020 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.
29 * Copyright (c) 2019, Klara Inc.
30 * Copyright (c) 2019, Allan Jude
31 * Copyright (c) 2022 Hewlett Packard Enterprise Development LP.
32 * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
36 #include <sys/dmu_impl.h>
37 #include <sys/dmu_tx.h>
39 #include <sys/dnode.h>
40 #include <sys/zfs_context.h>
41 #include <sys/dmu_objset.h>
42 #include <sys/dmu_traverse.h>
43 #include <sys/dsl_dataset.h>
44 #include <sys/dsl_dir.h>
45 #include <sys/dsl_pool.h>
46 #include <sys/dsl_synctask.h>
47 #include <sys/dsl_prop.h>
48 #include <sys/dmu_zfetch.h>
49 #include <sys/zfs_ioctl.h>
51 #include <sys/zio_checksum.h>
52 #include <sys/zio_compress.h>
54 #include <sys/zfeature.h>
57 #include <sys/trace_zfs.h>
58 #include <sys/zfs_racct.h>
59 #include <sys/zfs_rlock.h>
61 #include <sys/vmsystm.h>
62 #include <sys/zfs_znode.h>
66 * Enable/disable nopwrite feature.
68 static int zfs_nopwrite_enabled
= 1;
71 * Tunable to control percentage of dirtied L1 blocks from frees allowed into
72 * one TXG. After this threshold is crossed, additional dirty blocks from frees
73 * will wait until the next TXG.
74 * A value of zero will disable this throttle.
76 static uint_t zfs_per_txg_dirty_frees_percent
= 30;
79 * Enable/disable forcing txg sync when dirty checking for holes with lseek().
80 * By default this is enabled to ensure accurate hole reporting, it can result
81 * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads.
82 * Disabling this option will result in holes never being reported in dirty
83 * files which is always safe.
85 static int zfs_dmu_offset_next_sync
= 1;
88 * Limit the amount we can prefetch with one call to this amount. This
89 * helps to limit the amount of memory that can be used by prefetching.
90 * Larger objects should be prefetched a bit at a time.
93 uint_t dmu_prefetch_max
= 8 * 1024 * 1024;
95 uint_t dmu_prefetch_max
= 8 * SPA_MAXBLOCKSIZE
;
98 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
99 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "unallocated" },
100 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "object directory" },
101 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "object array" },
102 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "packed nvlist" },
103 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "packed nvlist size" },
104 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj" },
105 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj header" },
106 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA space map header" },
107 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA space map" },
108 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, TRUE
, "ZIL intent log" },
109 {DMU_BSWAP_DNODE
, TRUE
, FALSE
, TRUE
, "DMU dnode" },
110 {DMU_BSWAP_OBJSET
, TRUE
, TRUE
, FALSE
, "DMU objset" },
111 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL directory" },
112 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL directory child map"},
113 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dataset snap map" },
114 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL props" },
115 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL dataset" },
116 {DMU_BSWAP_ZNODE
, TRUE
, FALSE
, FALSE
, "ZFS znode" },
117 {DMU_BSWAP_OLDACL
, TRUE
, FALSE
, TRUE
, "ZFS V0 ACL" },
118 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "ZFS plain file" },
119 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS directory" },
120 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "ZFS master node" },
121 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS delete queue" },
122 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "zvol object" },
123 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "zvol prop" },
124 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "other uint8[]" },
125 {DMU_BSWAP_UINT64
, FALSE
, FALSE
, TRUE
, "other uint64[]" },
126 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "other ZAP" },
127 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "persistent error log" },
128 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, FALSE
, "SPA history" },
129 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "SPA history offsets" },
130 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "Pool properties" },
131 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL permissions" },
132 {DMU_BSWAP_ACL
, TRUE
, FALSE
, TRUE
, "ZFS ACL" },
133 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "ZFS SYSACL" },
134 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "FUID table" },
135 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "FUID table size" },
136 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dataset next clones"},
137 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "scan work queue" },
138 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS user/group/project used" },
139 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "ZFS user/group/project quota"},
140 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "snapshot refcount tags"},
141 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "DDT ZAP algorithm" },
142 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "DDT statistics" },
143 {DMU_BSWAP_UINT8
, TRUE
, FALSE
, TRUE
, "System attributes" },
144 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA master node" },
145 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA attr registration" },
146 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, TRUE
, "SA attr layouts" },
147 {DMU_BSWAP_ZAP
, TRUE
, FALSE
, FALSE
, "scan translations" },
148 {DMU_BSWAP_UINT8
, FALSE
, FALSE
, TRUE
, "deduplicated block" },
149 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL deadlist map" },
150 {DMU_BSWAP_UINT64
, TRUE
, TRUE
, FALSE
, "DSL deadlist map hdr" },
151 {DMU_BSWAP_ZAP
, TRUE
, TRUE
, FALSE
, "DSL dir clones" },
152 {DMU_BSWAP_UINT64
, TRUE
, FALSE
, FALSE
, "bpobj subobj" }
155 dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
156 { byteswap_uint8_array
, "uint8" },
157 { byteswap_uint16_array
, "uint16" },
158 { byteswap_uint32_array
, "uint32" },
159 { byteswap_uint64_array
, "uint64" },
160 { zap_byteswap
, "zap" },
161 { dnode_buf_byteswap
, "dnode" },
162 { dmu_objset_byteswap
, "objset" },
163 { zfs_znode_byteswap
, "znode" },
164 { zfs_oldacl_byteswap
, "oldacl" },
165 { zfs_acl_byteswap
, "acl" }
169 dmu_buf_hold_noread_by_dnode(dnode_t
*dn
, uint64_t offset
,
170 const void *tag
, dmu_buf_t
**dbp
)
175 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
176 blkid
= dbuf_whichblock(dn
, 0, offset
);
177 db
= dbuf_hold(dn
, blkid
, tag
);
178 rw_exit(&dn
->dn_struct_rwlock
);
182 return (SET_ERROR(EIO
));
190 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
191 const void *tag
, dmu_buf_t
**dbp
)
198 err
= dnode_hold(os
, object
, FTAG
, &dn
);
201 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
202 blkid
= dbuf_whichblock(dn
, 0, offset
);
203 db
= dbuf_hold(dn
, blkid
, tag
);
204 rw_exit(&dn
->dn_struct_rwlock
);
205 dnode_rele(dn
, FTAG
);
209 return (SET_ERROR(EIO
));
217 dmu_buf_hold_by_dnode(dnode_t
*dn
, uint64_t offset
,
218 const void *tag
, dmu_buf_t
**dbp
, int flags
)
221 int db_flags
= DB_RF_CANFAIL
;
223 if (flags
& DMU_READ_NO_PREFETCH
)
224 db_flags
|= DB_RF_NOPREFETCH
;
225 if (flags
& DMU_READ_NO_DECRYPT
)
226 db_flags
|= DB_RF_NO_DECRYPT
;
228 err
= dmu_buf_hold_noread_by_dnode(dn
, offset
, tag
, dbp
);
230 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
231 err
= dbuf_read(db
, NULL
, db_flags
);
242 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
243 const void *tag
, dmu_buf_t
**dbp
, int flags
)
246 int db_flags
= DB_RF_CANFAIL
;
248 if (flags
& DMU_READ_NO_PREFETCH
)
249 db_flags
|= DB_RF_NOPREFETCH
;
250 if (flags
& DMU_READ_NO_DECRYPT
)
251 db_flags
|= DB_RF_NO_DECRYPT
;
253 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
255 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
256 err
= dbuf_read(db
, NULL
, db_flags
);
269 return (DN_OLD_MAX_BONUSLEN
);
273 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
275 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
282 if (dn
->dn_bonus
!= db
) {
283 error
= SET_ERROR(EINVAL
);
284 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
285 error
= SET_ERROR(EINVAL
);
287 dnode_setbonuslen(dn
, newsize
, tx
);
296 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
298 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
305 if (!DMU_OT_IS_VALID(type
)) {
306 error
= SET_ERROR(EINVAL
);
307 } else if (dn
->dn_bonus
!= db
) {
308 error
= SET_ERROR(EINVAL
);
310 dnode_setbonus_type(dn
, type
, tx
);
319 dmu_get_bonustype(dmu_buf_t
*db_fake
)
321 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
323 dmu_object_type_t type
;
327 type
= dn
->dn_bonustype
;
334 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
339 error
= dnode_hold(os
, object
, FTAG
, &dn
);
340 dbuf_rm_spill(dn
, tx
);
341 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
342 dnode_rm_spill(dn
, tx
);
343 rw_exit(&dn
->dn_struct_rwlock
);
344 dnode_rele(dn
, FTAG
);
349 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
350 * has not yet been allocated a new bonus dbuf a will be allocated.
351 * Returns ENOENT, EIO, or 0.
353 int dmu_bonus_hold_by_dnode(dnode_t
*dn
, const void *tag
, dmu_buf_t
**dbp
,
358 uint32_t db_flags
= DB_RF_MUST_SUCCEED
;
360 if (flags
& DMU_READ_NO_PREFETCH
)
361 db_flags
|= DB_RF_NOPREFETCH
;
362 if (flags
& DMU_READ_NO_DECRYPT
)
363 db_flags
|= DB_RF_NO_DECRYPT
;
365 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
366 if (dn
->dn_bonus
== NULL
) {
367 if (!rw_tryupgrade(&dn
->dn_struct_rwlock
)) {
368 rw_exit(&dn
->dn_struct_rwlock
);
369 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
371 if (dn
->dn_bonus
== NULL
)
372 dbuf_create_bonus(dn
);
376 /* as long as the bonus buf is held, the dnode will be held */
377 if (zfs_refcount_add(&db
->db_holds
, tag
) == 1) {
378 VERIFY(dnode_add_ref(dn
, db
));
379 atomic_inc_32(&dn
->dn_dbufs_count
);
383 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
384 * hold and incrementing the dbuf count to ensure that dnode_move() sees
385 * a dnode hold for every dbuf.
387 rw_exit(&dn
->dn_struct_rwlock
);
389 error
= dbuf_read(db
, NULL
, db_flags
);
391 dnode_evict_bonus(dn
);
402 dmu_bonus_hold(objset_t
*os
, uint64_t object
, const void *tag
, dmu_buf_t
**dbp
)
407 error
= dnode_hold(os
, object
, FTAG
, &dn
);
411 error
= dmu_bonus_hold_by_dnode(dn
, tag
, dbp
, DMU_READ_NO_PREFETCH
);
412 dnode_rele(dn
, FTAG
);
418 * returns ENOENT, EIO, or 0.
420 * This interface will allocate a blank spill dbuf when a spill blk
421 * doesn't already exist on the dnode.
423 * if you only want to find an already existing spill db, then
424 * dmu_spill_hold_existing() should be used.
427 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, const void *tag
,
430 dmu_buf_impl_t
*db
= NULL
;
433 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
434 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
436 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
438 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
439 rw_exit(&dn
->dn_struct_rwlock
);
443 return (SET_ERROR(EIO
));
445 err
= dbuf_read(db
, NULL
, flags
);
456 dmu_spill_hold_existing(dmu_buf_t
*bonus
, const void *tag
, dmu_buf_t
**dbp
)
458 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
465 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
466 err
= SET_ERROR(EINVAL
);
468 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
470 if (!dn
->dn_have_spill
) {
471 err
= SET_ERROR(ENOENT
);
473 err
= dmu_spill_hold_by_dnode(dn
,
474 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
477 rw_exit(&dn
->dn_struct_rwlock
);
485 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, uint32_t flags
, const void *tag
,
488 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
491 uint32_t db_flags
= DB_RF_CANFAIL
;
493 if (flags
& DMU_READ_NO_DECRYPT
)
494 db_flags
|= DB_RF_NO_DECRYPT
;
498 err
= dmu_spill_hold_by_dnode(dn
, db_flags
, tag
, dbp
);
505 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
506 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
507 * and can induce severe lock contention when writing to several files
508 * whose dnodes are in the same block.
511 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
512 boolean_t read
, const void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
,
516 zstream_t
*zs
= NULL
;
517 uint64_t blkid
, nblks
, i
;
521 boolean_t missed
= B_FALSE
;
523 ASSERT(!read
|| length
<= DMU_MAX_ACCESS
);
526 * Note: We directly notify the prefetch code of this read, so that
527 * we can tell it about the multi-block read. dbuf_read() only knows
528 * about the one block it is accessing.
530 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
|
533 if ((flags
& DMU_READ_NO_DECRYPT
) != 0)
534 dbuf_flags
|= DB_RF_NO_DECRYPT
;
536 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
537 if (dn
->dn_datablkshift
) {
538 int blkshift
= dn
->dn_datablkshift
;
539 nblks
= (P2ROUNDUP(offset
+ length
, 1ULL << blkshift
) -
540 P2ALIGN(offset
, 1ULL << blkshift
)) >> blkshift
;
542 if (offset
+ length
> dn
->dn_datablksz
) {
543 zfs_panic_recover("zfs: accessing past end of object "
544 "%llx/%llx (size=%u access=%llu+%llu)",
545 (longlong_t
)dn
->dn_objset
->
546 os_dsl_dataset
->ds_object
,
547 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
548 (longlong_t
)offset
, (longlong_t
)length
);
549 rw_exit(&dn
->dn_struct_rwlock
);
550 return (SET_ERROR(EIO
));
554 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_SLEEP
);
557 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
,
559 blkid
= dbuf_whichblock(dn
, 0, offset
);
560 if ((flags
& DMU_READ_NO_PREFETCH
) == 0) {
562 * Prepare the zfetch before initiating the demand reads, so
563 * that if multiple threads block on same indirect block, we
564 * base predictions on the original less racy request order.
566 zs
= dmu_zfetch_prepare(&dn
->dn_zfetch
, blkid
, nblks
, read
,
569 for (i
= 0; i
< nblks
; i
++) {
570 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+ i
, tag
);
573 dmu_zfetch_run(zs
, missed
, B_TRUE
);
574 rw_exit(&dn
->dn_struct_rwlock
);
575 dmu_buf_rele_array(dbp
, nblks
, tag
);
578 return (SET_ERROR(EIO
));
582 * Initiate async demand data read.
583 * We check the db_state after calling dbuf_read() because
584 * (1) dbuf_read() may change the state to CACHED due to a
585 * hit in the ARC, and (2) on a cache miss, a child will
586 * have been added to "zio" but not yet completed, so the
587 * state will not yet be CACHED.
590 if (i
== nblks
- 1 && blkid
+ i
< dn
->dn_maxblkid
&&
591 offset
+ length
< db
->db
.db_offset
+
593 if (offset
<= db
->db
.db_offset
)
594 dbuf_flags
|= DB_RF_PARTIAL_FIRST
;
596 dbuf_flags
|= DB_RF_PARTIAL_MORE
;
598 (void) dbuf_read(db
, zio
, dbuf_flags
);
599 if (db
->db_state
!= DB_CACHED
)
606 zfs_racct_write(length
, nblks
);
609 dmu_zfetch_run(zs
, missed
, B_TRUE
);
610 rw_exit(&dn
->dn_struct_rwlock
);
613 /* wait for async read i/o */
616 dmu_buf_rele_array(dbp
, nblks
, tag
);
620 /* wait for other io to complete */
621 for (i
= 0; i
< nblks
; i
++) {
622 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
623 mutex_enter(&db
->db_mtx
);
624 while (db
->db_state
== DB_READ
||
625 db
->db_state
== DB_FILL
)
626 cv_wait(&db
->db_changed
, &db
->db_mtx
);
627 if (db
->db_state
== DB_UNCACHED
)
628 err
= SET_ERROR(EIO
);
629 mutex_exit(&db
->db_mtx
);
631 dmu_buf_rele_array(dbp
, nblks
, tag
);
643 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
644 uint64_t length
, int read
, const void *tag
, int *numbufsp
,
650 err
= dnode_hold(os
, object
, FTAG
, &dn
);
654 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
655 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
657 dnode_rele(dn
, FTAG
);
663 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
664 uint64_t length
, boolean_t read
, const void *tag
, int *numbufsp
,
667 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
673 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
674 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
681 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, const void *tag
)
684 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
689 for (i
= 0; i
< numbufs
; i
++) {
691 dbuf_rele(dbp
[i
], tag
);
694 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
698 * Issue prefetch I/Os for the given blocks. If level is greater than 0, the
699 * indirect blocks prefetched will be those that point to the blocks containing
700 * the data starting at offset, and continuing to offset + len. If the range
701 * it too long, prefetch the first dmu_prefetch_max bytes as requested, while
702 * for the rest only a higher level, also fitting within dmu_prefetch_max. It
703 * should primarily help random reads, since for long sequential reads there is
704 * a speculative prefetcher.
706 * Note that if the indirect blocks above the blocks being prefetched are not
707 * in cache, they will be asynchronously read in. Dnode read by dnode_hold()
708 * is currently synchronous.
711 dmu_prefetch(objset_t
*os
, uint64_t object
, int64_t level
, uint64_t offset
,
712 uint64_t len
, zio_priority_t pri
)
716 if (dmu_prefetch_max
== 0 || len
== 0) {
717 dmu_prefetch_dnode(os
, object
, pri
);
721 if (dnode_hold(os
, object
, FTAG
, &dn
) != 0)
724 dmu_prefetch_by_dnode(dn
, level
, offset
, len
, pri
);
726 dnode_rele(dn
, FTAG
);
730 dmu_prefetch_by_dnode(dnode_t
*dn
, int64_t level
, uint64_t offset
,
731 uint64_t len
, zio_priority_t pri
)
733 int64_t level2
= level
;
734 uint64_t start
, end
, start2
, end2
;
737 * Depending on len we may do two prefetches: blocks [start, end) at
738 * level, and following blocks [start2, end2) at higher level2.
740 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
741 if (dn
->dn_datablkshift
!= 0) {
743 * The object has multiple blocks. Calculate the full range
744 * of blocks [start, end2) and then split it into two parts,
745 * so that the first [start, end) fits into dmu_prefetch_max.
747 start
= dbuf_whichblock(dn
, level
, offset
);
748 end2
= dbuf_whichblock(dn
, level
, offset
+ len
- 1) + 1;
749 uint8_t ibs
= dn
->dn_indblkshift
;
750 uint8_t bs
= (level
== 0) ? dn
->dn_datablkshift
: ibs
;
751 uint_t limit
= P2ROUNDUP(dmu_prefetch_max
, 1 << bs
) >> bs
;
752 start2
= end
= MIN(end2
, start
+ limit
);
755 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
757 uint8_t ibps
= ibs
- SPA_BLKPTRSHIFT
;
758 limit
= P2ROUNDUP(dmu_prefetch_max
, 1 << ibs
) >> ibs
;
761 start2
= P2ROUNDUP(start2
, 1 << ibps
) >> ibps
;
762 end2
= P2ROUNDUP(end2
, 1 << ibps
) >> ibps
;
763 } while (end2
- start2
> limit
);
765 /* There is only one block. Prefetch it or nothing. */
766 start
= start2
= end2
= 0;
767 end
= start
+ (level
== 0 && offset
< dn
->dn_datablksz
);
770 for (uint64_t i
= start
; i
< end
; i
++)
771 dbuf_prefetch(dn
, level
, i
, pri
, 0);
772 for (uint64_t i
= start2
; i
< end2
; i
++)
773 dbuf_prefetch(dn
, level2
, i
, pri
, 0);
774 rw_exit(&dn
->dn_struct_rwlock
);
778 * Issue prefetch I/Os for the given object's dnode.
781 dmu_prefetch_dnode(objset_t
*os
, uint64_t object
, zio_priority_t pri
)
783 if (object
== 0 || object
>= DN_MAX_OBJECT
)
786 dnode_t
*dn
= DMU_META_DNODE(os
);
787 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
788 uint64_t blkid
= dbuf_whichblock(dn
, 0, object
* sizeof (dnode_phys_t
));
789 dbuf_prefetch(dn
, 0, blkid
, pri
, 0);
790 rw_exit(&dn
->dn_struct_rwlock
);
794 * Get the next "chunk" of file data to free. We traverse the file from
795 * the end so that the file gets shorter over time (if we crashes in the
796 * middle, this will leave us in a better state). We find allocated file
797 * data by simply searching the allocated level 1 indirects.
799 * On input, *start should be the first offset that does not need to be
800 * freed (e.g. "offset + length"). On return, *start will be the first
801 * offset that should be freed and l1blks is set to the number of level 1
802 * indirect blocks found within the chunk.
805 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
, uint64_t *l1blks
)
808 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
809 /* bytes of data covered by a level-1 indirect block */
810 uint64_t iblkrange
= (uint64_t)dn
->dn_datablksz
*
811 EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
813 ASSERT3U(minimum
, <=, *start
);
816 * Check if we can free the entire range assuming that all of the
817 * L1 blocks in this range have data. If we can, we use this
818 * worst case value as an estimate so we can avoid having to look
819 * at the object's actual data.
821 uint64_t total_l1blks
=
822 (roundup(*start
, iblkrange
) - (minimum
/ iblkrange
* iblkrange
)) /
824 if (total_l1blks
<= maxblks
) {
825 *l1blks
= total_l1blks
;
829 ASSERT(ISP2(iblkrange
));
831 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
835 * dnode_next_offset(BACKWARDS) will find an allocated L1
836 * indirect block at or before the input offset. We must
837 * decrement *start so that it is at the end of the region
842 err
= dnode_next_offset(dn
,
843 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
845 /* if there are no indirect blocks before start, we are done */
849 } else if (err
!= 0) {
854 /* set start to the beginning of this L1 indirect */
855 *start
= P2ALIGN(*start
, iblkrange
);
857 if (*start
< minimum
)
865 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
866 * otherwise return false.
867 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
870 dmu_objset_zfs_unmounting(objset_t
*os
)
873 if (dmu_objset_type(os
) == DMU_OST_ZFS
)
874 return (zfs_get_vfs_flag_unmounted(os
));
882 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
885 uint64_t object_size
;
887 uint64_t dirty_frees_threshold
;
888 dsl_pool_t
*dp
= dmu_objset_pool(os
);
891 return (SET_ERROR(EINVAL
));
893 object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
894 if (offset
>= object_size
)
897 if (zfs_per_txg_dirty_frees_percent
<= 100)
898 dirty_frees_threshold
=
899 zfs_per_txg_dirty_frees_percent
* zfs_dirty_data_max
/ 100;
901 dirty_frees_threshold
= zfs_dirty_data_max
/ 20;
903 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
904 length
= object_size
- offset
;
906 while (length
!= 0) {
907 uint64_t chunk_end
, chunk_begin
, chunk_len
;
911 if (dmu_objset_zfs_unmounting(dn
->dn_objset
))
912 return (SET_ERROR(EINTR
));
914 chunk_end
= chunk_begin
= offset
+ length
;
916 /* move chunk_begin backwards to the beginning of this chunk */
917 err
= get_next_chunk(dn
, &chunk_begin
, offset
, &l1blks
);
920 ASSERT3U(chunk_begin
, >=, offset
);
921 ASSERT3U(chunk_begin
, <=, chunk_end
);
923 chunk_len
= chunk_end
- chunk_begin
;
925 tx
= dmu_tx_create(os
);
926 dmu_tx_hold_free(tx
, dn
->dn_object
, chunk_begin
, chunk_len
);
929 * Mark this transaction as typically resulting in a net
930 * reduction in space used.
932 dmu_tx_mark_netfree(tx
);
933 err
= dmu_tx_assign(tx
, TXG_WAIT
);
939 uint64_t txg
= dmu_tx_get_txg(tx
);
941 mutex_enter(&dp
->dp_lock
);
942 uint64_t long_free_dirty
=
943 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
];
944 mutex_exit(&dp
->dp_lock
);
947 * To avoid filling up a TXG with just frees, wait for
948 * the next TXG to open before freeing more chunks if
949 * we have reached the threshold of frees.
951 if (dirty_frees_threshold
!= 0 &&
952 long_free_dirty
>= dirty_frees_threshold
) {
953 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay
);
955 txg_wait_open(dp
, 0, B_TRUE
);
960 * In order to prevent unnecessary write throttling, for each
961 * TXG, we track the cumulative size of L1 blocks being dirtied
962 * in dnode_free_range() below. We compare this number to a
963 * tunable threshold, past which we prevent new L1 dirty freeing
964 * blocks from being added into the open TXG. See
965 * dmu_free_long_range_impl() for details. The threshold
966 * prevents write throttle activation due to dirty freeing L1
967 * blocks taking up a large percentage of zfs_dirty_data_max.
969 mutex_enter(&dp
->dp_lock
);
970 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
] +=
971 l1blks
<< dn
->dn_indblkshift
;
972 mutex_exit(&dp
->dp_lock
);
973 DTRACE_PROBE3(free__long__range
,
974 uint64_t, long_free_dirty
, uint64_t, chunk_len
,
976 dnode_free_range(dn
, chunk_begin
, chunk_len
, tx
);
986 dmu_free_long_range(objset_t
*os
, uint64_t object
,
987 uint64_t offset
, uint64_t length
)
992 err
= dnode_hold(os
, object
, FTAG
, &dn
);
995 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
998 * It is important to zero out the maxblkid when freeing the entire
999 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
1000 * will take the fast path, and (b) dnode_reallocate() can verify
1001 * that the entire file has been freed.
1003 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
1004 dn
->dn_maxblkid
= 0;
1006 dnode_rele(dn
, FTAG
);
1011 dmu_free_long_object(objset_t
*os
, uint64_t object
)
1016 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
1020 tx
= dmu_tx_create(os
);
1021 dmu_tx_hold_bonus(tx
, object
);
1022 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
1023 dmu_tx_mark_netfree(tx
);
1024 err
= dmu_tx_assign(tx
, TXG_WAIT
);
1026 err
= dmu_object_free(os
, object
, tx
);
1036 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
1037 uint64_t size
, dmu_tx_t
*tx
)
1040 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1043 ASSERT(offset
< UINT64_MAX
);
1044 ASSERT(size
== DMU_OBJECT_END
|| size
<= UINT64_MAX
- offset
);
1045 dnode_free_range(dn
, offset
, size
, tx
);
1046 dnode_rele(dn
, FTAG
);
1051 dmu_read_impl(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
1052 void *buf
, uint32_t flags
)
1055 int numbufs
, err
= 0;
1058 * Deal with odd block sizes, where there can't be data past the first
1059 * block. If we ever do the tail block optimization, we will need to
1060 * handle that here as well.
1062 if (dn
->dn_maxblkid
== 0) {
1063 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
1064 MIN(size
, dn
->dn_datablksz
- offset
);
1065 memset((char *)buf
+ newsz
, 0, size
- newsz
);
1070 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
1074 * NB: we could do this block-at-a-time, but it's nice
1075 * to be reading in parallel.
1077 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
1078 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
1082 for (i
= 0; i
< numbufs
; i
++) {
1085 dmu_buf_t
*db
= dbp
[i
];
1089 bufoff
= offset
- db
->db_offset
;
1090 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1092 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
1096 buf
= (char *)buf
+ tocpy
;
1098 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1104 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1105 void *buf
, uint32_t flags
)
1110 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1114 err
= dmu_read_impl(dn
, offset
, size
, buf
, flags
);
1115 dnode_rele(dn
, FTAG
);
1120 dmu_read_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
, void *buf
,
1123 return (dmu_read_impl(dn
, offset
, size
, buf
, flags
));
1127 dmu_write_impl(dmu_buf_t
**dbp
, int numbufs
, uint64_t offset
, uint64_t size
,
1128 const void *buf
, dmu_tx_t
*tx
)
1132 for (i
= 0; i
< numbufs
; i
++) {
1135 dmu_buf_t
*db
= dbp
[i
];
1139 bufoff
= offset
- db
->db_offset
;
1140 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1142 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1144 if (tocpy
== db
->db_size
)
1145 dmu_buf_will_fill(db
, tx
, B_FALSE
);
1147 dmu_buf_will_dirty(db
, tx
);
1149 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
1151 if (tocpy
== db
->db_size
)
1152 dmu_buf_fill_done(db
, tx
, B_FALSE
);
1156 buf
= (char *)buf
+ tocpy
;
1161 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1162 const void *buf
, dmu_tx_t
*tx
)
1170 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
1171 FALSE
, FTAG
, &numbufs
, &dbp
));
1172 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
1173 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1177 * Note: Lustre is an external consumer of this interface.
1180 dmu_write_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
1181 const void *buf
, dmu_tx_t
*tx
)
1189 VERIFY0(dmu_buf_hold_array_by_dnode(dn
, offset
, size
,
1190 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
));
1191 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
1192 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1196 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1205 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
1206 FALSE
, FTAG
, &numbufs
, &dbp
));
1208 for (i
= 0; i
< numbufs
; i
++) {
1209 dmu_buf_t
*db
= dbp
[i
];
1211 dmu_buf_will_not_fill(db
, tx
);
1213 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1217 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
1218 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
1219 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
1223 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
1224 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
1225 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
1228 dmu_buf_write_embedded(db
,
1229 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
1230 uncompressed_size
, compressed_size
, byteorder
, tx
);
1232 dmu_buf_rele(db
, FTAG
);
1236 dmu_redact(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1242 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1244 for (i
= 0; i
< numbufs
; i
++)
1245 dmu_buf_redact(dbp
[i
], tx
);
1246 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1251 dmu_read_uio_dnode(dnode_t
*dn
, zfs_uio_t
*uio
, uint64_t size
)
1254 int numbufs
, i
, err
;
1257 * NB: we could do this block-at-a-time, but it's nice
1258 * to be reading in parallel.
1260 err
= dmu_buf_hold_array_by_dnode(dn
, zfs_uio_offset(uio
), size
,
1261 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1265 for (i
= 0; i
< numbufs
; i
++) {
1268 dmu_buf_t
*db
= dbp
[i
];
1272 bufoff
= zfs_uio_offset(uio
) - db
->db_offset
;
1273 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1275 err
= zfs_uio_fault_move((char *)db
->db_data
+ bufoff
, tocpy
,
1283 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1289 * Read 'size' bytes into the uio buffer.
1290 * From object zdb->db_object.
1291 * Starting at zfs_uio_offset(uio).
1293 * If the caller already has a dbuf in the target object
1294 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1295 * because we don't have to find the dnode_t for the object.
1298 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, zfs_uio_t
*uio
, uint64_t size
)
1300 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1309 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1316 * Read 'size' bytes into the uio buffer.
1317 * From the specified object
1318 * Starting at offset zfs_uio_offset(uio).
1321 dmu_read_uio(objset_t
*os
, uint64_t object
, zfs_uio_t
*uio
, uint64_t size
)
1329 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1333 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1335 dnode_rele(dn
, FTAG
);
1341 dmu_write_uio_dnode(dnode_t
*dn
, zfs_uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1348 err
= dmu_buf_hold_array_by_dnode(dn
, zfs_uio_offset(uio
), size
,
1349 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1353 for (i
= 0; i
< numbufs
; i
++) {
1356 dmu_buf_t
*db
= dbp
[i
];
1360 offset_t off
= zfs_uio_offset(uio
);
1361 bufoff
= off
- db
->db_offset
;
1362 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1364 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1366 if (tocpy
== db
->db_size
)
1367 dmu_buf_will_fill(db
, tx
, B_TRUE
);
1369 dmu_buf_will_dirty(db
, tx
);
1371 err
= zfs_uio_fault_move((char *)db
->db_data
+ bufoff
,
1372 tocpy
, UIO_WRITE
, uio
);
1374 if (tocpy
== db
->db_size
&& dmu_buf_fill_done(db
, tx
, err
)) {
1375 /* The fill was reverted. Undo any uio progress. */
1376 zfs_uio_advance(uio
, off
- zfs_uio_offset(uio
));
1385 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1390 * Write 'size' bytes from the uio buffer.
1391 * To object zdb->db_object.
1392 * Starting at offset zfs_uio_offset(uio).
1394 * If the caller already has a dbuf in the target object
1395 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1396 * because we don't have to find the dnode_t for the object.
1399 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, zfs_uio_t
*uio
, uint64_t size
,
1402 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1411 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1418 * Write 'size' bytes from the uio buffer.
1419 * To the specified object.
1420 * Starting at offset zfs_uio_offset(uio).
1423 dmu_write_uio(objset_t
*os
, uint64_t object
, zfs_uio_t
*uio
, uint64_t size
,
1432 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1436 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1438 dnode_rele(dn
, FTAG
);
1442 #endif /* _KERNEL */
1445 * Allocate a loaned anonymous arc buffer.
1448 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1450 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1452 return (arc_loan_buf(db
->db_objset
->os_spa
, B_FALSE
, size
));
1456 * Free a loaned arc buffer.
1459 dmu_return_arcbuf(arc_buf_t
*buf
)
1461 arc_return_buf(buf
, FTAG
);
1462 arc_buf_destroy(buf
, FTAG
);
1466 * A "lightweight" write is faster than a regular write (e.g.
1467 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1468 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the
1469 * data can not be read or overwritten until the transaction's txg has been
1470 * synced. This makes it appropriate for workloads that are known to be
1471 * (temporarily) write-only, like "zfs receive".
1473 * A single block is written, starting at the specified offset in bytes. If
1474 * the call is successful, it returns 0 and the provided abd has been
1475 * consumed (the caller should not free it).
1478 dmu_lightweight_write_by_dnode(dnode_t
*dn
, uint64_t offset
, abd_t
*abd
,
1479 const zio_prop_t
*zp
, zio_flag_t flags
, dmu_tx_t
*tx
)
1481 dbuf_dirty_record_t
*dr
=
1482 dbuf_dirty_lightweight(dn
, dbuf_whichblock(dn
, 0, offset
), tx
);
1484 return (SET_ERROR(EIO
));
1485 dr
->dt
.dll
.dr_abd
= abd
;
1486 dr
->dt
.dll
.dr_props
= *zp
;
1487 dr
->dt
.dll
.dr_flags
= flags
;
1492 * When possible directly assign passed loaned arc buffer to a dbuf.
1493 * If this is not possible copy the contents of passed arc buf via
1497 dmu_assign_arcbuf_by_dnode(dnode_t
*dn
, uint64_t offset
, arc_buf_t
*buf
,
1501 objset_t
*os
= dn
->dn_objset
;
1502 uint64_t object
= dn
->dn_object
;
1503 uint32_t blksz
= (uint32_t)arc_buf_lsize(buf
);
1506 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1507 blkid
= dbuf_whichblock(dn
, 0, offset
);
1508 db
= dbuf_hold(dn
, blkid
, FTAG
);
1509 rw_exit(&dn
->dn_struct_rwlock
);
1511 return (SET_ERROR(EIO
));
1514 * We can only assign if the offset is aligned and the arc buf is the
1515 * same size as the dbuf.
1517 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1518 zfs_racct_write(blksz
, 1);
1519 dbuf_assign_arcbuf(db
, buf
, tx
);
1520 dbuf_rele(db
, FTAG
);
1522 /* compressed bufs must always be assignable to their dbuf */
1523 ASSERT3U(arc_get_compression(buf
), ==, ZIO_COMPRESS_OFF
);
1524 ASSERT(!(buf
->b_flags
& ARC_BUF_FLAG_COMPRESSED
));
1526 dbuf_rele(db
, FTAG
);
1527 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1528 dmu_return_arcbuf(buf
);
1535 dmu_assign_arcbuf_by_dbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1539 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1541 DB_DNODE_ENTER(dbuf
);
1542 err
= dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf
), offset
, buf
, tx
);
1543 DB_DNODE_EXIT(dbuf
);
1549 dbuf_dirty_record_t
*dsa_dr
;
1550 dmu_sync_cb_t
*dsa_done
;
1556 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1559 dmu_sync_arg_t
*dsa
= varg
;
1560 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1561 blkptr_t
*bp
= zio
->io_bp
;
1563 if (zio
->io_error
== 0) {
1564 if (BP_IS_HOLE(bp
)) {
1566 * A block of zeros may compress to a hole, but the
1567 * block size still needs to be known for replay.
1569 BP_SET_LSIZE(bp
, db
->db_size
);
1570 } else if (!BP_IS_EMBEDDED(bp
)) {
1571 ASSERT(BP_GET_LEVEL(bp
) == 0);
1578 dmu_sync_late_arrival_ready(zio_t
*zio
)
1580 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1584 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1587 dmu_sync_arg_t
*dsa
= varg
;
1588 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1589 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1590 zgd_t
*zgd
= dsa
->dsa_zgd
;
1593 * Record the vdev(s) backing this blkptr so they can be flushed after
1594 * the writes for the lwb have completed.
1596 if (zio
->io_error
== 0) {
1597 zil_lwb_add_block(zgd
->zgd_lwb
, zgd
->zgd_bp
);
1600 mutex_enter(&db
->db_mtx
);
1601 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1602 if (zio
->io_error
== 0) {
1603 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1604 if (dr
->dt
.dl
.dr_nopwrite
) {
1605 blkptr_t
*bp
= zio
->io_bp
;
1606 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
1607 uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
);
1609 ASSERT(BP_EQUAL(bp
, bp_orig
));
1610 VERIFY(BP_EQUAL(bp
, db
->db_blkptr
));
1611 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1612 VERIFY(zio_checksum_table
[chksum
].ci_flags
&
1613 ZCHECKSUM_FLAG_NOPWRITE
);
1615 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1616 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1617 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1620 * Old style holes are filled with all zeros, whereas
1621 * new-style holes maintain their lsize, type, level,
1622 * and birth time (see zio_write_compress). While we
1623 * need to reset the BP_SET_LSIZE() call that happened
1624 * in dmu_sync_ready for old style holes, we do *not*
1625 * want to wipe out the information contained in new
1626 * style holes. Thus, only zero out the block pointer if
1627 * it's an old style hole.
1629 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1630 BP_GET_LOGICAL_BIRTH(&dr
->dt
.dl
.dr_overridden_by
) == 0)
1631 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1633 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1635 cv_broadcast(&db
->db_changed
);
1636 mutex_exit(&db
->db_mtx
);
1638 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1640 kmem_free(dsa
, sizeof (*dsa
));
1644 dmu_sync_late_arrival_done(zio_t
*zio
)
1646 blkptr_t
*bp
= zio
->io_bp
;
1647 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1648 zgd_t
*zgd
= dsa
->dsa_zgd
;
1650 if (zio
->io_error
== 0) {
1652 * Record the vdev(s) backing this blkptr so they can be
1653 * flushed after the writes for the lwb have completed.
1655 zil_lwb_add_block(zgd
->zgd_lwb
, zgd
->zgd_bp
);
1657 if (!BP_IS_HOLE(bp
)) {
1658 blkptr_t
*bp_orig __maybe_unused
= &zio
->io_bp_orig
;
1659 ASSERT(!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
1660 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1661 ASSERT(BP_GET_LOGICAL_BIRTH(zio
->io_bp
) == zio
->io_txg
);
1662 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1663 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1667 dmu_tx_commit(dsa
->dsa_tx
);
1669 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1671 abd_free(zio
->io_abd
);
1672 kmem_free(dsa
, sizeof (*dsa
));
1676 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1677 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1679 dmu_sync_arg_t
*dsa
;
1683 error
= dbuf_read((dmu_buf_impl_t
*)zgd
->zgd_db
, NULL
,
1684 DB_RF_CANFAIL
| DB_RF_NOPREFETCH
);
1688 tx
= dmu_tx_create(os
);
1689 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1691 * This transaction does not produce any dirty data or log blocks, so
1692 * it should not be throttled. All other cases wait for TXG sync, by
1693 * which time the log block we are writing will be obsolete, so we can
1694 * skip waiting and just return error here instead.
1696 if (dmu_tx_assign(tx
, TXG_NOWAIT
| TXG_NOTHROTTLE
) != 0) {
1698 /* Make zl_get_data do txg_waited_synced() */
1699 return (SET_ERROR(EIO
));
1703 * In order to prevent the zgd's lwb from being free'd prior to
1704 * dmu_sync_late_arrival_done() being called, we have to ensure
1705 * the lwb's "max txg" takes this tx's txg into account.
1707 zil_lwb_add_txg(zgd
->zgd_lwb
, dmu_tx_get_txg(tx
));
1709 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1711 dsa
->dsa_done
= done
;
1716 * Since we are currently syncing this txg, it's nontrivial to
1717 * determine what BP to nopwrite against, so we disable nopwrite.
1719 * When syncing, the db_blkptr is initially the BP of the previous
1720 * txg. We can not nopwrite against it because it will be changed
1721 * (this is similar to the non-late-arrival case where the dbuf is
1722 * dirty in a future txg).
1724 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1725 * We can not nopwrite against it because although the BP will not
1726 * (typically) be changed, the data has not yet been persisted to this
1729 * Finally, when dbuf_write_done() is called, it is theoretically
1730 * possible to always nopwrite, because the data that was written in
1731 * this txg is the same data that we are trying to write. However we
1732 * would need to check that this dbuf is not dirty in any future
1733 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1734 * don't nopwrite in this case.
1736 zp
->zp_nopwrite
= B_FALSE
;
1738 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1739 abd_get_from_buf(zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
),
1740 zgd
->zgd_db
->db_size
, zgd
->zgd_db
->db_size
, zp
,
1741 dmu_sync_late_arrival_ready
, NULL
, dmu_sync_late_arrival_done
,
1742 dsa
, ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, zb
));
1748 * Intent log support: sync the block associated with db to disk.
1749 * N.B. and XXX: the caller is responsible for making sure that the
1750 * data isn't changing while dmu_sync() is writing it.
1754 * EEXIST: this txg has already been synced, so there's nothing to do.
1755 * The caller should not log the write.
1757 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1758 * The caller should not log the write.
1760 * EALREADY: this block is already in the process of being synced.
1761 * The caller should track its progress (somehow).
1763 * EIO: could not do the I/O.
1764 * The caller should do a txg_wait_synced().
1766 * 0: the I/O has been initiated.
1767 * The caller should log this blkptr in the done callback.
1768 * It is possible that the I/O will fail, in which case
1769 * the error will be reported to the done callback and
1770 * propagated to pio from zio_done().
1773 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1775 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1776 objset_t
*os
= db
->db_objset
;
1777 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1778 dbuf_dirty_record_t
*dr
, *dr_next
;
1779 dmu_sync_arg_t
*dsa
;
1780 zbookmark_phys_t zb
;
1784 ASSERT(pio
!= NULL
);
1787 SET_BOOKMARK(&zb
, ds
->ds_object
,
1788 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1792 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1796 * If we're frozen (running ziltest), we always need to generate a bp.
1798 if (txg
> spa_freeze_txg(os
->os_spa
))
1799 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1802 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1803 * and us. If we determine that this txg is not yet syncing,
1804 * but it begins to sync a moment later, that's OK because the
1805 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1807 mutex_enter(&db
->db_mtx
);
1809 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1811 * This txg has already synced. There's nothing to do.
1813 mutex_exit(&db
->db_mtx
);
1814 return (SET_ERROR(EEXIST
));
1817 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1819 * This txg is currently syncing, so we can't mess with
1820 * the dirty record anymore; just write a new log block.
1822 mutex_exit(&db
->db_mtx
);
1823 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1826 dr
= dbuf_find_dirty_eq(db
, txg
);
1830 * There's no dr for this dbuf, so it must have been freed.
1831 * There's no need to log writes to freed blocks, so we're done.
1833 mutex_exit(&db
->db_mtx
);
1834 return (SET_ERROR(ENOENT
));
1837 dr_next
= list_next(&db
->db_dirty_records
, dr
);
1838 ASSERT(dr_next
== NULL
|| dr_next
->dr_txg
< txg
);
1840 if (db
->db_blkptr
!= NULL
) {
1842 * We need to fill in zgd_bp with the current blkptr so that
1843 * the nopwrite code can check if we're writing the same
1844 * data that's already on disk. We can only nopwrite if we
1845 * are sure that after making the copy, db_blkptr will not
1846 * change until our i/o completes. We ensure this by
1847 * holding the db_mtx, and only allowing nopwrite if the
1848 * block is not already dirty (see below). This is verified
1849 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1852 *zgd
->zgd_bp
= *db
->db_blkptr
;
1856 * Assume the on-disk data is X, the current syncing data (in
1857 * txg - 1) is Y, and the current in-memory data is Z (currently
1860 * We usually want to perform a nopwrite if X and Z are the
1861 * same. However, if Y is different (i.e. the BP is going to
1862 * change before this write takes effect), then a nopwrite will
1863 * be incorrect - we would override with X, which could have
1864 * been freed when Y was written.
1866 * (Note that this is not a concern when we are nop-writing from
1867 * syncing context, because X and Y must be identical, because
1868 * all previous txgs have been synced.)
1870 * Therefore, we disable nopwrite if the current BP could change
1871 * before this TXG. There are two ways it could change: by
1872 * being dirty (dr_next is non-NULL), or by being freed
1873 * (dnode_block_freed()). This behavior is verified by
1874 * zio_done(), which VERIFYs that the override BP is identical
1875 * to the on-disk BP.
1879 if (dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
1880 zp
.zp_nopwrite
= B_FALSE
;
1883 ASSERT(dr
->dr_txg
== txg
);
1884 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1885 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1887 * We have already issued a sync write for this buffer,
1888 * or this buffer has already been synced. It could not
1889 * have been dirtied since, or we would have cleared the state.
1891 mutex_exit(&db
->db_mtx
);
1892 return (SET_ERROR(EALREADY
));
1895 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1896 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1897 mutex_exit(&db
->db_mtx
);
1899 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1901 dsa
->dsa_done
= done
;
1905 zio_nowait(arc_write(pio
, os
->os_spa
, txg
, zgd
->zgd_bp
,
1906 dr
->dt
.dl
.dr_data
, !DBUF_IS_CACHEABLE(db
), dbuf_is_l2cacheable(db
),
1907 &zp
, dmu_sync_ready
, NULL
, dmu_sync_done
, dsa
,
1908 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, &zb
));
1914 dmu_object_set_nlevels(objset_t
*os
, uint64_t object
, int nlevels
, dmu_tx_t
*tx
)
1919 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1922 err
= dnode_set_nlevels(dn
, nlevels
, tx
);
1923 dnode_rele(dn
, FTAG
);
1928 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1934 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1937 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1938 dnode_rele(dn
, FTAG
);
1943 dmu_object_set_maxblkid(objset_t
*os
, uint64_t object
, uint64_t maxblkid
,
1949 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1952 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
1953 dnode_new_blkid(dn
, maxblkid
, tx
, B_FALSE
, B_TRUE
);
1954 rw_exit(&dn
->dn_struct_rwlock
);
1955 dnode_rele(dn
, FTAG
);
1960 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1966 * Send streams include each object's checksum function. This
1967 * check ensures that the receiving system can understand the
1968 * checksum function transmitted.
1970 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1972 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1973 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1974 dn
->dn_checksum
= checksum
;
1975 dnode_setdirty(dn
, tx
);
1976 dnode_rele(dn
, FTAG
);
1980 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1986 * Send streams include each object's compression function. This
1987 * check ensures that the receiving system can understand the
1988 * compression function transmitted.
1990 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1992 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1993 dn
->dn_compress
= compress
;
1994 dnode_setdirty(dn
, tx
);
1995 dnode_rele(dn
, FTAG
);
1999 * When the "redundant_metadata" property is set to "most", only indirect
2000 * blocks of this level and higher will have an additional ditto block.
2002 static const int zfs_redundant_metadata_most_ditto_level
= 2;
2005 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
2007 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
2008 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
2010 enum zio_checksum checksum
= os
->os_checksum
;
2011 enum zio_compress compress
= os
->os_compress
;
2012 uint8_t complevel
= os
->os_complevel
;
2013 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
2014 boolean_t dedup
= B_FALSE
;
2015 boolean_t nopwrite
= B_FALSE
;
2016 boolean_t dedup_verify
= os
->os_dedup_verify
;
2017 boolean_t encrypt
= B_FALSE
;
2018 int copies
= os
->os_copies
;
2021 * We maintain different write policies for each of the following
2024 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2025 * 3. all other level 0 blocks
2029 * XXX -- we should design a compression algorithm
2030 * that specializes in arrays of bps.
2032 compress
= zio_compress_select(os
->os_spa
,
2033 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
2036 * Metadata always gets checksummed. If the data
2037 * checksum is multi-bit correctable, and it's not a
2038 * ZBT-style checksum, then it's suitable for metadata
2039 * as well. Otherwise, the metadata checksum defaults
2042 if (!(zio_checksum_table
[checksum
].ci_flags
&
2043 ZCHECKSUM_FLAG_METADATA
) ||
2044 (zio_checksum_table
[checksum
].ci_flags
&
2045 ZCHECKSUM_FLAG_EMBEDDED
))
2046 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
2048 switch (os
->os_redundant_metadata
) {
2049 case ZFS_REDUNDANT_METADATA_ALL
:
2052 case ZFS_REDUNDANT_METADATA_MOST
:
2053 if (level
>= zfs_redundant_metadata_most_ditto_level
||
2054 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))
2057 case ZFS_REDUNDANT_METADATA_SOME
:
2058 if (DMU_OT_IS_CRITICAL(type
))
2061 case ZFS_REDUNDANT_METADATA_NONE
:
2064 } else if (wp
& WP_NOFILL
) {
2068 * If we're writing preallocated blocks, we aren't actually
2069 * writing them so don't set any policy properties. These
2070 * blocks are currently only used by an external subsystem
2071 * outside of zfs (i.e. dump) and not written by the zio
2074 compress
= ZIO_COMPRESS_OFF
;
2075 checksum
= ZIO_CHECKSUM_OFF
;
2077 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
2079 complevel
= zio_complevel_select(os
->os_spa
, compress
,
2080 complevel
, complevel
);
2082 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
2083 zio_checksum_select(dn
->dn_checksum
, checksum
) :
2087 * Determine dedup setting. If we are in dmu_sync(),
2088 * we won't actually dedup now because that's all
2089 * done in syncing context; but we do want to use the
2090 * dedup checksum. If the checksum is not strong
2091 * enough to ensure unique signatures, force
2094 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
2095 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
2096 if (!(zio_checksum_table
[checksum
].ci_flags
&
2097 ZCHECKSUM_FLAG_DEDUP
))
2098 dedup_verify
= B_TRUE
;
2102 * Enable nopwrite if we have secure enough checksum
2103 * algorithm (see comment in zio_nop_write) and
2104 * compression is enabled. We don't enable nopwrite if
2105 * dedup is enabled as the two features are mutually
2108 nopwrite
= (!dedup
&& (zio_checksum_table
[checksum
].ci_flags
&
2109 ZCHECKSUM_FLAG_NOPWRITE
) &&
2110 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
2114 * All objects in an encrypted objset are protected from modification
2115 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2116 * in the bp, so we cannot use all copies. Encrypted objects are also
2117 * not subject to nopwrite since writing the same data will still
2118 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2119 * to avoid ambiguity in the dedup code since the DDT does not store
2122 if (os
->os_encrypted
&& (wp
& WP_NOFILL
) == 0) {
2125 if (DMU_OT_IS_ENCRYPTED(type
)) {
2126 copies
= MIN(copies
, SPA_DVAS_PER_BP
- 1);
2133 (type
== DMU_OT_DNODE
|| type
== DMU_OT_OBJSET
)) {
2134 compress
= ZIO_COMPRESS_EMPTY
;
2138 zp
->zp_compress
= compress
;
2139 zp
->zp_complevel
= complevel
;
2140 zp
->zp_checksum
= checksum
;
2141 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
2142 zp
->zp_level
= level
;
2143 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
2144 zp
->zp_dedup
= dedup
;
2145 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
2146 zp
->zp_nopwrite
= nopwrite
;
2147 zp
->zp_encrypt
= encrypt
;
2148 zp
->zp_byteorder
= ZFS_HOST_BYTEORDER
;
2149 memset(zp
->zp_salt
, 0, ZIO_DATA_SALT_LEN
);
2150 memset(zp
->zp_iv
, 0, ZIO_DATA_IV_LEN
);
2151 memset(zp
->zp_mac
, 0, ZIO_DATA_MAC_LEN
);
2152 zp
->zp_zpl_smallblk
= DMU_OT_IS_FILE(zp
->zp_type
) ?
2153 os
->os_zpl_special_smallblock
: 0;
2155 ASSERT3U(zp
->zp_compress
, !=, ZIO_COMPRESS_INHERIT
);
2159 * Reports the location of data and holes in an object. In order to
2160 * accurately report holes all dirty data must be synced to disk. This
2161 * causes extremely poor performance when seeking for holes in a dirty file.
2162 * As a compromise, only provide hole data when the dnode is clean. When
2163 * a dnode is dirty report the dnode as having no holes by returning EBUSY
2164 * which is always safe to do.
2167 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
2170 int restarted
= 0, err
;
2173 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2177 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2179 if (dnode_is_dirty(dn
)) {
2181 * If the zfs_dmu_offset_next_sync module option is enabled
2182 * then hole reporting has been requested. Dirty dnodes
2183 * must be synced to disk to accurately report holes.
2185 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is
2186 * held by the caller only a single restart will be required.
2187 * We tolerate callers which do not hold the rangelock by
2188 * returning EBUSY and not reporting holes after one restart.
2190 if (zfs_dmu_offset_next_sync
) {
2191 rw_exit(&dn
->dn_struct_rwlock
);
2192 dnode_rele(dn
, FTAG
);
2195 return (SET_ERROR(EBUSY
));
2197 txg_wait_synced(dmu_objset_pool(os
), 0);
2202 err
= SET_ERROR(EBUSY
);
2204 err
= dnode_next_offset(dn
, DNODE_FIND_HAVELOCK
|
2205 (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
2208 rw_exit(&dn
->dn_struct_rwlock
);
2209 dnode_rele(dn
, FTAG
);
2215 dmu_read_l0_bps(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t length
,
2216 blkptr_t
*bps
, size_t *nbpsp
)
2218 dmu_buf_t
**dbp
, *dbuf
;
2223 error
= dmu_buf_hold_array(os
, object
, offset
, length
, FALSE
, FTAG
,
2226 if (error
== ESRCH
) {
2227 error
= SET_ERROR(ENXIO
);
2232 ASSERT3U(numbufs
, <=, *nbpsp
);
2234 for (int i
= 0; i
< numbufs
; i
++) {
2236 db
= (dmu_buf_impl_t
*)dbuf
;
2238 mutex_enter(&db
->db_mtx
);
2240 if (!list_is_empty(&db
->db_dirty_records
)) {
2241 dbuf_dirty_record_t
*dr
;
2243 dr
= list_head(&db
->db_dirty_records
);
2244 if (dr
->dt
.dl
.dr_brtwrite
) {
2246 * This is very special case where we clone a
2247 * block and in the same transaction group we
2248 * read its BP (most likely to clone the clone).
2250 bp
= &dr
->dt
.dl
.dr_overridden_by
;
2253 * The block was modified in the same
2254 * transaction group.
2256 mutex_exit(&db
->db_mtx
);
2257 error
= SET_ERROR(EAGAIN
);
2264 mutex_exit(&db
->db_mtx
);
2268 * The block was created in this transaction group,
2269 * so it has no BP yet.
2271 error
= SET_ERROR(EAGAIN
);
2275 * Make sure we clone only data blocks.
2277 if (BP_IS_METADATA(bp
) && !BP_IS_HOLE(bp
)) {
2278 error
= SET_ERROR(EINVAL
);
2283 * If the block was allocated in transaction group that is not
2284 * yet synced, we could clone it, but we couldn't write this
2285 * operation into ZIL, or it may be impossible to replay, since
2286 * the block may appear not yet allocated at that point.
2288 if (BP_GET_BIRTH(bp
) > spa_freeze_txg(os
->os_spa
)) {
2289 error
= SET_ERROR(EINVAL
);
2292 if (BP_GET_BIRTH(bp
) > spa_last_synced_txg(os
->os_spa
)) {
2293 error
= SET_ERROR(EAGAIN
);
2302 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
2308 dmu_brt_clone(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t length
,
2309 dmu_tx_t
*tx
, const blkptr_t
*bps
, size_t nbps
)
2312 dmu_buf_t
**dbp
, *dbuf
;
2314 struct dirty_leaf
*dl
;
2315 dbuf_dirty_record_t
*dr
;
2317 int error
= 0, i
, numbufs
;
2321 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, length
, FALSE
, FTAG
,
2323 ASSERT3U(nbps
, ==, numbufs
);
2326 * Before we start cloning make sure that the dbufs sizes match new BPs
2327 * sizes. If they don't, that's a no-go, as we are not able to shrink
2330 for (i
= 0; i
< numbufs
; i
++) {
2332 db
= (dmu_buf_impl_t
*)dbuf
;
2335 ASSERT0(db
->db_level
);
2336 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2337 ASSERT(db
->db_blkid
!= DMU_SPILL_BLKID
);
2339 if (!BP_IS_HOLE(bp
) && BP_GET_LSIZE(bp
) != dbuf
->db_size
) {
2340 error
= SET_ERROR(EXDEV
);
2345 for (i
= 0; i
< numbufs
; i
++) {
2347 db
= (dmu_buf_impl_t
*)dbuf
;
2350 ASSERT0(db
->db_level
);
2351 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2352 ASSERT(db
->db_blkid
!= DMU_SPILL_BLKID
);
2353 ASSERT(BP_IS_HOLE(bp
) || dbuf
->db_size
== BP_GET_LSIZE(bp
));
2355 dmu_buf_will_clone(dbuf
, tx
);
2357 mutex_enter(&db
->db_mtx
);
2359 dr
= list_head(&db
->db_dirty_records
);
2361 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2363 dl
->dr_overridden_by
= *bp
;
2364 dl
->dr_brtwrite
= B_TRUE
;
2365 dl
->dr_override_state
= DR_OVERRIDDEN
;
2366 if (BP_IS_HOLE(bp
)) {
2367 BP_SET_LOGICAL_BIRTH(&dl
->dr_overridden_by
, 0);
2368 BP_SET_PHYSICAL_BIRTH(&dl
->dr_overridden_by
, 0);
2370 BP_SET_LOGICAL_BIRTH(&dl
->dr_overridden_by
,
2372 if (!BP_IS_EMBEDDED(bp
)) {
2373 BP_SET_PHYSICAL_BIRTH(&dl
->dr_overridden_by
,
2378 mutex_exit(&db
->db_mtx
);
2381 * When data in embedded into BP there is no need to create
2382 * BRT entry as there is no data block. Just copy the BP as
2383 * it contains the data.
2385 if (!BP_IS_HOLE(bp
) && !BP_IS_EMBEDDED(bp
)) {
2386 brt_pending_add(spa
, bp
, tx
);
2390 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
2396 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2398 dnode_phys_t
*dnp
= dn
->dn_phys
;
2400 doi
->doi_data_block_size
= dn
->dn_datablksz
;
2401 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
2402 1ULL << dn
->dn_indblkshift
: 0;
2403 doi
->doi_type
= dn
->dn_type
;
2404 doi
->doi_bonus_type
= dn
->dn_bonustype
;
2405 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
2406 doi
->doi_dnodesize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
2407 doi
->doi_indirection
= dn
->dn_nlevels
;
2408 doi
->doi_checksum
= dn
->dn_checksum
;
2409 doi
->doi_compress
= dn
->dn_compress
;
2410 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
2411 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
2412 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
2413 doi
->doi_fill_count
= 0;
2414 for (int i
= 0; i
< dnp
->dn_nblkptr
; i
++)
2415 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
2419 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2421 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2422 mutex_enter(&dn
->dn_mtx
);
2424 __dmu_object_info_from_dnode(dn
, doi
);
2426 mutex_exit(&dn
->dn_mtx
);
2427 rw_exit(&dn
->dn_struct_rwlock
);
2431 * Get information on a DMU object.
2432 * If doi is NULL, just indicates whether the object exists.
2435 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
2438 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
2444 dmu_object_info_from_dnode(dn
, doi
);
2446 dnode_rele(dn
, FTAG
);
2451 * As above, but faster; can be used when you have a held dbuf in hand.
2454 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
2456 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2459 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
2464 * Faster still when you only care about the size.
2467 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
2468 u_longlong_t
*nblk512
)
2470 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2476 *blksize
= dn
->dn_datablksz
;
2477 /* add in number of slots used for the dnode itself */
2478 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
2479 SPA_MINBLOCKSHIFT
) + dn
->dn_num_slots
;
2484 dmu_object_dnsize_from_db(dmu_buf_t
*db_fake
, int *dnsize
)
2486 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2491 *dnsize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
2496 byteswap_uint64_array(void *vbuf
, size_t size
)
2498 uint64_t *buf
= vbuf
;
2499 size_t count
= size
>> 3;
2502 ASSERT((size
& 7) == 0);
2504 for (i
= 0; i
< count
; i
++)
2505 buf
[i
] = BSWAP_64(buf
[i
]);
2509 byteswap_uint32_array(void *vbuf
, size_t size
)
2511 uint32_t *buf
= vbuf
;
2512 size_t count
= size
>> 2;
2515 ASSERT((size
& 3) == 0);
2517 for (i
= 0; i
< count
; i
++)
2518 buf
[i
] = BSWAP_32(buf
[i
]);
2522 byteswap_uint16_array(void *vbuf
, size_t size
)
2524 uint16_t *buf
= vbuf
;
2525 size_t count
= size
>> 1;
2528 ASSERT((size
& 1) == 0);
2530 for (i
= 0; i
< count
; i
++)
2531 buf
[i
] = BSWAP_16(buf
[i
]);
2535 byteswap_uint8_array(void *vbuf
, size_t size
)
2537 (void) vbuf
, (void) size
;
2558 arc_fini(); /* arc depends on l2arc, so arc must go first */
2570 EXPORT_SYMBOL(dmu_bonus_hold
);
2571 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode
);
2572 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2573 EXPORT_SYMBOL(dmu_buf_rele_array
);
2574 EXPORT_SYMBOL(dmu_prefetch
);
2575 EXPORT_SYMBOL(dmu_prefetch_by_dnode
);
2576 EXPORT_SYMBOL(dmu_prefetch_dnode
);
2577 EXPORT_SYMBOL(dmu_free_range
);
2578 EXPORT_SYMBOL(dmu_free_long_range
);
2579 EXPORT_SYMBOL(dmu_free_long_object
);
2580 EXPORT_SYMBOL(dmu_read
);
2581 EXPORT_SYMBOL(dmu_read_by_dnode
);
2582 EXPORT_SYMBOL(dmu_write
);
2583 EXPORT_SYMBOL(dmu_write_by_dnode
);
2584 EXPORT_SYMBOL(dmu_prealloc
);
2585 EXPORT_SYMBOL(dmu_object_info
);
2586 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2587 EXPORT_SYMBOL(dmu_object_info_from_db
);
2588 EXPORT_SYMBOL(dmu_object_size_from_db
);
2589 EXPORT_SYMBOL(dmu_object_dnsize_from_db
);
2590 EXPORT_SYMBOL(dmu_object_set_nlevels
);
2591 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2592 EXPORT_SYMBOL(dmu_object_set_maxblkid
);
2593 EXPORT_SYMBOL(dmu_object_set_checksum
);
2594 EXPORT_SYMBOL(dmu_object_set_compress
);
2595 EXPORT_SYMBOL(dmu_offset_next
);
2596 EXPORT_SYMBOL(dmu_write_policy
);
2597 EXPORT_SYMBOL(dmu_sync
);
2598 EXPORT_SYMBOL(dmu_request_arcbuf
);
2599 EXPORT_SYMBOL(dmu_return_arcbuf
);
2600 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode
);
2601 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf
);
2602 EXPORT_SYMBOL(dmu_buf_hold
);
2603 EXPORT_SYMBOL(dmu_ot
);
2605 ZFS_MODULE_PARAM(zfs
, zfs_
, nopwrite_enabled
, INT
, ZMOD_RW
,
2606 "Enable NOP writes");
2608 ZFS_MODULE_PARAM(zfs
, zfs_
, per_txg_dirty_frees_percent
, UINT
, ZMOD_RW
,
2609 "Percentage of dirtied blocks from frees in one TXG");
2611 ZFS_MODULE_PARAM(zfs
, zfs_
, dmu_offset_next_sync
, INT
, ZMOD_RW
,
2612 "Enable forcing txg sync to find holes");
2615 ZFS_MODULE_PARAM(zfs
, , dmu_prefetch_max
, UINT
, ZMOD_RW
,
2616 "Limit one prefetch call to this size");