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, 2016 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) 2014, 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/vmsystm.h>
52 #include <sys/zfs_znode.h>
56 * Enable/disable nopwrite feature.
58 int zfs_nopwrite_enabled
= 1;
60 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
61 { DMU_BSWAP_UINT8
, TRUE
, "unallocated" },
62 { DMU_BSWAP_ZAP
, TRUE
, "object directory" },
63 { DMU_BSWAP_UINT64
, TRUE
, "object array" },
64 { DMU_BSWAP_UINT8
, TRUE
, "packed nvlist" },
65 { DMU_BSWAP_UINT64
, TRUE
, "packed nvlist size" },
66 { DMU_BSWAP_UINT64
, TRUE
, "bpobj" },
67 { DMU_BSWAP_UINT64
, TRUE
, "bpobj header" },
68 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map header" },
69 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map" },
70 { DMU_BSWAP_UINT64
, TRUE
, "ZIL intent log" },
71 { DMU_BSWAP_DNODE
, TRUE
, "DMU dnode" },
72 { DMU_BSWAP_OBJSET
, TRUE
, "DMU objset" },
73 { DMU_BSWAP_UINT64
, TRUE
, "DSL directory" },
74 { DMU_BSWAP_ZAP
, TRUE
, "DSL directory child map"},
75 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset snap map" },
76 { DMU_BSWAP_ZAP
, TRUE
, "DSL props" },
77 { DMU_BSWAP_UINT64
, TRUE
, "DSL dataset" },
78 { DMU_BSWAP_ZNODE
, TRUE
, "ZFS znode" },
79 { DMU_BSWAP_OLDACL
, TRUE
, "ZFS V0 ACL" },
80 { DMU_BSWAP_UINT8
, FALSE
, "ZFS plain file" },
81 { DMU_BSWAP_ZAP
, TRUE
, "ZFS directory" },
82 { DMU_BSWAP_ZAP
, TRUE
, "ZFS master node" },
83 { DMU_BSWAP_ZAP
, TRUE
, "ZFS delete queue" },
84 { DMU_BSWAP_UINT8
, FALSE
, "zvol object" },
85 { DMU_BSWAP_ZAP
, TRUE
, "zvol prop" },
86 { DMU_BSWAP_UINT8
, FALSE
, "other uint8[]" },
87 { DMU_BSWAP_UINT64
, FALSE
, "other uint64[]" },
88 { DMU_BSWAP_ZAP
, TRUE
, "other ZAP" },
89 { DMU_BSWAP_ZAP
, TRUE
, "persistent error log" },
90 { DMU_BSWAP_UINT8
, TRUE
, "SPA history" },
91 { DMU_BSWAP_UINT64
, TRUE
, "SPA history offsets" },
92 { DMU_BSWAP_ZAP
, TRUE
, "Pool properties" },
93 { DMU_BSWAP_ZAP
, TRUE
, "DSL permissions" },
94 { DMU_BSWAP_ACL
, TRUE
, "ZFS ACL" },
95 { DMU_BSWAP_UINT8
, TRUE
, "ZFS SYSACL" },
96 { DMU_BSWAP_UINT8
, TRUE
, "FUID table" },
97 { DMU_BSWAP_UINT64
, TRUE
, "FUID table size" },
98 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset next clones"},
99 { DMU_BSWAP_ZAP
, TRUE
, "scan work queue" },
100 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group used" },
101 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group quota" },
102 { DMU_BSWAP_ZAP
, TRUE
, "snapshot refcount tags"},
103 { DMU_BSWAP_ZAP
, TRUE
, "DDT ZAP algorithm" },
104 { DMU_BSWAP_ZAP
, TRUE
, "DDT statistics" },
105 { DMU_BSWAP_UINT8
, TRUE
, "System attributes" },
106 { DMU_BSWAP_ZAP
, TRUE
, "SA master node" },
107 { DMU_BSWAP_ZAP
, TRUE
, "SA attr registration" },
108 { DMU_BSWAP_ZAP
, TRUE
, "SA attr layouts" },
109 { DMU_BSWAP_ZAP
, TRUE
, "scan translations" },
110 { DMU_BSWAP_UINT8
, FALSE
, "deduplicated block" },
111 { DMU_BSWAP_ZAP
, TRUE
, "DSL deadlist map" },
112 { DMU_BSWAP_UINT64
, TRUE
, "DSL deadlist map hdr" },
113 { DMU_BSWAP_ZAP
, TRUE
, "DSL dir clones" },
114 { DMU_BSWAP_UINT64
, TRUE
, "bpobj subobj" }
117 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
118 { byteswap_uint8_array
, "uint8" },
119 { byteswap_uint16_array
, "uint16" },
120 { byteswap_uint32_array
, "uint32" },
121 { byteswap_uint64_array
, "uint64" },
122 { zap_byteswap
, "zap" },
123 { dnode_buf_byteswap
, "dnode" },
124 { dmu_objset_byteswap
, "objset" },
125 { zfs_znode_byteswap
, "znode" },
126 { zfs_oldacl_byteswap
, "oldacl" },
127 { zfs_acl_byteswap
, "acl" }
131 dmu_buf_hold_noread_by_dnode(dnode_t
*dn
, uint64_t offset
,
132 void *tag
, dmu_buf_t
**dbp
)
137 blkid
= dbuf_whichblock(dn
, 0, offset
);
138 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
139 db
= dbuf_hold(dn
, blkid
, tag
);
140 rw_exit(&dn
->dn_struct_rwlock
);
144 return (SET_ERROR(EIO
));
151 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
152 void *tag
, dmu_buf_t
**dbp
)
159 err
= dnode_hold(os
, object
, FTAG
, &dn
);
162 blkid
= dbuf_whichblock(dn
, 0, offset
);
163 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
164 db
= dbuf_hold(dn
, blkid
, tag
);
165 rw_exit(&dn
->dn_struct_rwlock
);
166 dnode_rele(dn
, FTAG
);
170 return (SET_ERROR(EIO
));
178 dmu_buf_hold_by_dnode(dnode_t
*dn
, uint64_t offset
,
179 void *tag
, dmu_buf_t
**dbp
, int flags
)
182 int db_flags
= DB_RF_CANFAIL
;
184 if (flags
& DMU_READ_NO_PREFETCH
)
185 db_flags
|= DB_RF_NOPREFETCH
;
187 err
= dmu_buf_hold_noread_by_dnode(dn
, offset
, tag
, dbp
);
189 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
190 err
= dbuf_read(db
, NULL
, db_flags
);
201 dmu_buf_hold(objset_t
*os
, uint64_t object
, 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
;
210 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
212 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
213 err
= dbuf_read(db
, NULL
, db_flags
);
226 return (DN_OLD_MAX_BONUSLEN
);
230 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
232 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
239 if (dn
->dn_bonus
!= db
) {
240 error
= SET_ERROR(EINVAL
);
241 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
242 error
= SET_ERROR(EINVAL
);
244 dnode_setbonuslen(dn
, newsize
, tx
);
253 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
255 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
262 if (!DMU_OT_IS_VALID(type
)) {
263 error
= SET_ERROR(EINVAL
);
264 } else if (dn
->dn_bonus
!= db
) {
265 error
= SET_ERROR(EINVAL
);
267 dnode_setbonus_type(dn
, type
, tx
);
276 dmu_get_bonustype(dmu_buf_t
*db_fake
)
278 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
280 dmu_object_type_t type
;
284 type
= dn
->dn_bonustype
;
291 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
296 error
= dnode_hold(os
, object
, FTAG
, &dn
);
297 dbuf_rm_spill(dn
, tx
);
298 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
299 dnode_rm_spill(dn
, tx
);
300 rw_exit(&dn
->dn_struct_rwlock
);
301 dnode_rele(dn
, FTAG
);
306 * returns ENOENT, EIO, or 0.
309 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
315 error
= dnode_hold(os
, object
, FTAG
, &dn
);
319 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
320 if (dn
->dn_bonus
== NULL
) {
321 rw_exit(&dn
->dn_struct_rwlock
);
322 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
323 if (dn
->dn_bonus
== NULL
)
324 dbuf_create_bonus(dn
);
328 /* as long as the bonus buf is held, the dnode will be held */
329 if (refcount_add(&db
->db_holds
, tag
) == 1) {
330 VERIFY(dnode_add_ref(dn
, db
));
331 atomic_inc_32(&dn
->dn_dbufs_count
);
335 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
336 * hold and incrementing the dbuf count to ensure that dnode_move() sees
337 * a dnode hold for every dbuf.
339 rw_exit(&dn
->dn_struct_rwlock
);
341 dnode_rele(dn
, FTAG
);
343 VERIFY(0 == dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
));
350 * returns ENOENT, EIO, or 0.
352 * This interface will allocate a blank spill dbuf when a spill blk
353 * doesn't already exist on the dnode.
355 * if you only want to find an already existing spill db, then
356 * dmu_spill_hold_existing() should be used.
359 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
361 dmu_buf_impl_t
*db
= NULL
;
364 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
365 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
367 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
369 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
370 rw_exit(&dn
->dn_struct_rwlock
);
373 err
= dbuf_read(db
, NULL
, flags
);
382 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
384 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
391 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
392 err
= SET_ERROR(EINVAL
);
394 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
396 if (!dn
->dn_have_spill
) {
397 err
= SET_ERROR(ENOENT
);
399 err
= dmu_spill_hold_by_dnode(dn
,
400 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
403 rw_exit(&dn
->dn_struct_rwlock
);
411 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
413 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
419 err
= dmu_spill_hold_by_dnode(dn
, DB_RF_CANFAIL
, tag
, dbp
);
426 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
427 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
428 * and can induce severe lock contention when writing to several files
429 * whose dnodes are in the same block.
432 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
433 boolean_t read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
436 uint64_t blkid
, nblks
, i
;
441 ASSERT(length
<= DMU_MAX_ACCESS
);
444 * Note: We directly notify the prefetch code of this read, so that
445 * we can tell it about the multi-block read. dbuf_read() only knows
446 * about the one block it is accessing.
448 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
|
451 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
452 if (dn
->dn_datablkshift
) {
453 int blkshift
= dn
->dn_datablkshift
;
454 nblks
= (P2ROUNDUP(offset
+ length
, 1ULL << blkshift
) -
455 P2ALIGN(offset
, 1ULL << blkshift
)) >> blkshift
;
457 if (offset
+ length
> dn
->dn_datablksz
) {
458 zfs_panic_recover("zfs: accessing past end of object "
459 "%llx/%llx (size=%u access=%llu+%llu)",
460 (longlong_t
)dn
->dn_objset
->
461 os_dsl_dataset
->ds_object
,
462 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
463 (longlong_t
)offset
, (longlong_t
)length
);
464 rw_exit(&dn
->dn_struct_rwlock
);
465 return (SET_ERROR(EIO
));
469 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_SLEEP
);
471 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
472 blkid
= dbuf_whichblock(dn
, 0, offset
);
473 for (i
= 0; i
< nblks
; i
++) {
474 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+ i
, tag
);
476 rw_exit(&dn
->dn_struct_rwlock
);
477 dmu_buf_rele_array(dbp
, nblks
, tag
);
479 return (SET_ERROR(EIO
));
482 /* initiate async i/o */
484 (void) dbuf_read(db
, zio
, dbuf_flags
);
488 if ((flags
& DMU_READ_NO_PREFETCH
) == 0 &&
489 DNODE_META_IS_CACHEABLE(dn
) && length
<= zfetch_array_rd_sz
) {
490 dmu_zfetch(&dn
->dn_zfetch
, blkid
, nblks
,
491 read
&& DNODE_IS_CACHEABLE(dn
));
493 rw_exit(&dn
->dn_struct_rwlock
);
495 /* wait for async i/o */
498 dmu_buf_rele_array(dbp
, nblks
, tag
);
502 /* wait for other io to complete */
504 for (i
= 0; i
< nblks
; i
++) {
505 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
506 mutex_enter(&db
->db_mtx
);
507 while (db
->db_state
== DB_READ
||
508 db
->db_state
== DB_FILL
)
509 cv_wait(&db
->db_changed
, &db
->db_mtx
);
510 if (db
->db_state
== DB_UNCACHED
)
511 err
= SET_ERROR(EIO
);
512 mutex_exit(&db
->db_mtx
);
514 dmu_buf_rele_array(dbp
, nblks
, tag
);
526 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
527 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
532 err
= dnode_hold(os
, object
, FTAG
, &dn
);
536 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
537 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
539 dnode_rele(dn
, FTAG
);
545 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
546 uint64_t length
, boolean_t read
, void *tag
, int *numbufsp
,
549 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
555 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
556 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
563 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
566 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
571 for (i
= 0; i
< numbufs
; i
++) {
573 dbuf_rele(dbp
[i
], tag
);
576 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
580 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
581 * indirect blocks prefeteched will be those that point to the blocks containing
582 * the data starting at offset, and continuing to offset + len.
584 * Note that if the indirect blocks above the blocks being prefetched are not in
585 * cache, they will be asychronously read in.
588 dmu_prefetch(objset_t
*os
, uint64_t object
, int64_t level
, uint64_t offset
,
589 uint64_t len
, zio_priority_t pri
)
595 if (len
== 0) { /* they're interested in the bonus buffer */
596 dn
= DMU_META_DNODE(os
);
598 if (object
== 0 || object
>= DN_MAX_OBJECT
)
601 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
602 blkid
= dbuf_whichblock(dn
, level
,
603 object
* sizeof (dnode_phys_t
));
604 dbuf_prefetch(dn
, level
, blkid
, pri
, 0);
605 rw_exit(&dn
->dn_struct_rwlock
);
610 * XXX - Note, if the dnode for the requested object is not
611 * already cached, we will do a *synchronous* read in the
612 * dnode_hold() call. The same is true for any indirects.
614 err
= dnode_hold(os
, object
, FTAG
, &dn
);
618 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
620 * offset + len - 1 is the last byte we want to prefetch for, and offset
621 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
622 * last block we want to prefetch, and dbuf_whichblock(dn, level,
623 * offset) is the first. Then the number we need to prefetch is the
626 if (level
> 0 || dn
->dn_datablkshift
!= 0) {
627 nblks
= dbuf_whichblock(dn
, level
, offset
+ len
- 1) -
628 dbuf_whichblock(dn
, level
, offset
) + 1;
630 nblks
= (offset
< dn
->dn_datablksz
);
636 blkid
= dbuf_whichblock(dn
, level
, offset
);
637 for (i
= 0; i
< nblks
; i
++)
638 dbuf_prefetch(dn
, level
, blkid
+ i
, pri
, 0);
641 rw_exit(&dn
->dn_struct_rwlock
);
643 dnode_rele(dn
, FTAG
);
647 * Get the next "chunk" of file data to free. We traverse the file from
648 * the end so that the file gets shorter over time (if we crashes in the
649 * middle, this will leave us in a better state). We find allocated file
650 * data by simply searching the allocated level 1 indirects.
652 * On input, *start should be the first offset that does not need to be
653 * freed (e.g. "offset + length"). On return, *start will be the first
654 * offset that should be freed.
657 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
659 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
660 /* bytes of data covered by a level-1 indirect block */
662 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
665 ASSERT3U(minimum
, <=, *start
);
667 if (*start
- minimum
<= iblkrange
* maxblks
) {
671 ASSERT(ISP2(iblkrange
));
673 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
677 * dnode_next_offset(BACKWARDS) will find an allocated L1
678 * indirect block at or before the input offset. We must
679 * decrement *start so that it is at the end of the region
683 err
= dnode_next_offset(dn
,
684 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
686 /* if there are no indirect blocks before start, we are done */
690 } else if (err
!= 0) {
694 /* set start to the beginning of this L1 indirect */
695 *start
= P2ALIGN(*start
, iblkrange
);
697 if (*start
< minimum
)
703 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
706 uint64_t object_size
;
710 return (SET_ERROR(EINVAL
));
712 object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
713 if (offset
>= object_size
)
716 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
717 length
= object_size
- offset
;
719 while (length
!= 0) {
720 uint64_t chunk_end
, chunk_begin
;
723 chunk_end
= chunk_begin
= offset
+ length
;
725 /* move chunk_begin backwards to the beginning of this chunk */
726 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
729 ASSERT3U(chunk_begin
, >=, offset
);
730 ASSERT3U(chunk_begin
, <=, chunk_end
);
732 tx
= dmu_tx_create(os
);
733 dmu_tx_hold_free(tx
, dn
->dn_object
,
734 chunk_begin
, chunk_end
- chunk_begin
);
737 * Mark this transaction as typically resulting in a net
738 * reduction in space used.
740 dmu_tx_mark_netfree(tx
);
741 err
= dmu_tx_assign(tx
, TXG_WAIT
);
746 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
749 length
-= chunk_end
- chunk_begin
;
755 dmu_free_long_range(objset_t
*os
, uint64_t object
,
756 uint64_t offset
, uint64_t length
)
761 err
= dnode_hold(os
, object
, FTAG
, &dn
);
764 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
767 * It is important to zero out the maxblkid when freeing the entire
768 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
769 * will take the fast path, and (b) dnode_reallocate() can verify
770 * that the entire file has been freed.
772 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
775 dnode_rele(dn
, FTAG
);
780 dmu_free_long_object(objset_t
*os
, uint64_t object
)
785 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
789 tx
= dmu_tx_create(os
);
790 dmu_tx_hold_bonus(tx
, object
);
791 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
792 dmu_tx_mark_netfree(tx
);
793 err
= dmu_tx_assign(tx
, TXG_WAIT
);
795 err
= dmu_object_free(os
, object
, tx
);
805 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
806 uint64_t size
, dmu_tx_t
*tx
)
809 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
812 ASSERT(offset
< UINT64_MAX
);
813 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
814 dnode_free_range(dn
, offset
, size
, tx
);
815 dnode_rele(dn
, FTAG
);
820 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
821 void *buf
, uint32_t flags
)
827 err
= dnode_hold(os
, object
, FTAG
, &dn
);
832 * Deal with odd block sizes, where there can't be data past the first
833 * block. If we ever do the tail block optimization, we will need to
834 * handle that here as well.
836 if (dn
->dn_maxblkid
== 0) {
837 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
838 MIN(size
, dn
->dn_datablksz
- offset
);
839 bzero((char *)buf
+ newsz
, size
- newsz
);
844 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
848 * NB: we could do this block-at-a-time, but it's nice
849 * to be reading in parallel.
851 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
852 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
856 for (i
= 0; i
< numbufs
; i
++) {
859 dmu_buf_t
*db
= dbp
[i
];
863 bufoff
= offset
- db
->db_offset
;
864 tocpy
= MIN(db
->db_size
- bufoff
, size
);
866 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
870 buf
= (char *)buf
+ tocpy
;
872 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
874 dnode_rele(dn
, FTAG
);
879 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
880 const void *buf
, dmu_tx_t
*tx
)
888 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
889 FALSE
, FTAG
, &numbufs
, &dbp
));
891 for (i
= 0; i
< numbufs
; i
++) {
894 dmu_buf_t
*db
= dbp
[i
];
898 bufoff
= offset
- db
->db_offset
;
899 tocpy
= MIN(db
->db_size
- bufoff
, size
);
901 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
903 if (tocpy
== db
->db_size
)
904 dmu_buf_will_fill(db
, tx
);
906 dmu_buf_will_dirty(db
, tx
);
908 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
910 if (tocpy
== db
->db_size
)
911 dmu_buf_fill_done(db
, tx
);
915 buf
= (char *)buf
+ tocpy
;
917 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
921 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
930 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
931 FALSE
, FTAG
, &numbufs
, &dbp
));
933 for (i
= 0; i
< numbufs
; i
++) {
934 dmu_buf_t
*db
= dbp
[i
];
936 dmu_buf_will_not_fill(db
, tx
);
938 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
942 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
943 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
944 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
948 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
949 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
950 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
953 dmu_buf_write_embedded(db
,
954 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
955 uncompressed_size
, compressed_size
, byteorder
, tx
);
957 dmu_buf_rele(db
, FTAG
);
961 * DMU support for xuio
963 kstat_t
*xuio_ksp
= NULL
;
965 typedef struct xuio_stats
{
966 /* loaned yet not returned arc_buf */
967 kstat_named_t xuiostat_onloan_rbuf
;
968 kstat_named_t xuiostat_onloan_wbuf
;
969 /* whether a copy is made when loaning out a read buffer */
970 kstat_named_t xuiostat_rbuf_copied
;
971 kstat_named_t xuiostat_rbuf_nocopy
;
972 /* whether a copy is made when assigning a write buffer */
973 kstat_named_t xuiostat_wbuf_copied
;
974 kstat_named_t xuiostat_wbuf_nocopy
;
977 static xuio_stats_t xuio_stats
= {
978 { "onloan_read_buf", KSTAT_DATA_UINT64
},
979 { "onloan_write_buf", KSTAT_DATA_UINT64
},
980 { "read_buf_copied", KSTAT_DATA_UINT64
},
981 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
982 { "write_buf_copied", KSTAT_DATA_UINT64
},
983 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
986 #define XUIOSTAT_INCR(stat, val) \
987 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
988 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
991 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
994 uio_t
*uio
= &xuio
->xu_uio
;
996 uio
->uio_iovcnt
= nblk
;
997 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
999 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
1001 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
1002 priv
->iovp
= (iovec_t
*)uio
->uio_iov
;
1003 XUIO_XUZC_PRIV(xuio
) = priv
;
1005 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1006 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
1008 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
1014 dmu_xuio_fini(xuio_t
*xuio
)
1016 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1017 int nblk
= priv
->cnt
;
1019 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
1020 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
1021 kmem_free(priv
, sizeof (dmu_xuio_t
));
1023 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1024 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
1026 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
1030 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1031 * and increase priv->next by 1.
1034 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
1037 uio_t
*uio
= &xuio
->xu_uio
;
1038 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1039 int i
= priv
->next
++;
1041 ASSERT(i
< priv
->cnt
);
1042 ASSERT(off
+ n
<= arc_buf_size(abuf
));
1043 iov
= (iovec_t
*)uio
->uio_iov
+ i
;
1044 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
1046 priv
->bufs
[i
] = abuf
;
1051 dmu_xuio_cnt(xuio_t
*xuio
)
1053 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1058 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
1060 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1062 ASSERT(i
< priv
->cnt
);
1063 return (priv
->bufs
[i
]);
1067 dmu_xuio_clear(xuio_t
*xuio
, int i
)
1069 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1071 ASSERT(i
< priv
->cnt
);
1072 priv
->bufs
[i
] = NULL
;
1076 xuio_stat_init(void)
1078 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1079 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1080 KSTAT_FLAG_VIRTUAL
);
1081 if (xuio_ksp
!= NULL
) {
1082 xuio_ksp
->ks_data
= &xuio_stats
;
1083 kstat_install(xuio_ksp
);
1088 xuio_stat_fini(void)
1090 if (xuio_ksp
!= NULL
) {
1091 kstat_delete(xuio_ksp
);
1097 xuio_stat_wbuf_copied()
1099 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1103 xuio_stat_wbuf_nocopy()
1105 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1110 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1113 int numbufs
, i
, err
;
1114 xuio_t
*xuio
= NULL
;
1117 * NB: we could do this block-at-a-time, but it's nice
1118 * to be reading in parallel.
1120 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1121 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1125 for (i
= 0; i
< numbufs
; i
++) {
1128 dmu_buf_t
*db
= dbp
[i
];
1132 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1133 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1136 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1137 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1138 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1139 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1141 uio
->uio_resid
-= tocpy
;
1142 uio
->uio_loffset
+= tocpy
;
1145 if (abuf
== dbuf_abuf
)
1146 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1148 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1150 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1158 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1164 * Read 'size' bytes into the uio buffer.
1165 * From object zdb->db_object.
1166 * Starting at offset uio->uio_loffset.
1168 * If the caller already has a dbuf in the target object
1169 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1170 * because we don't have to find the dnode_t for the object.
1173 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1175 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1184 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1191 * Read 'size' bytes into the uio buffer.
1192 * From the specified object
1193 * Starting at offset uio->uio_loffset.
1196 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1204 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1208 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1210 dnode_rele(dn
, FTAG
);
1216 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1223 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1224 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1228 for (i
= 0; i
< numbufs
; i
++) {
1231 dmu_buf_t
*db
= dbp
[i
];
1235 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1236 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1238 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1240 if (tocpy
== db
->db_size
)
1241 dmu_buf_will_fill(db
, tx
);
1243 dmu_buf_will_dirty(db
, tx
);
1246 * XXX uiomove could block forever (eg.nfs-backed
1247 * pages). There needs to be a uiolockdown() function
1248 * to lock the pages in memory, so that uiomove won't
1251 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1254 if (tocpy
== db
->db_size
)
1255 dmu_buf_fill_done(db
, tx
);
1263 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1268 * Write 'size' bytes from the uio buffer.
1269 * To object zdb->db_object.
1270 * Starting at offset uio->uio_loffset.
1272 * If the caller already has a dbuf in the target object
1273 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1274 * because we don't have to find the dnode_t for the object.
1277 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1280 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1289 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1296 * Write 'size' bytes from the uio buffer.
1297 * To the specified object.
1298 * Starting at offset uio->uio_loffset.
1301 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1310 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1314 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1316 dnode_rele(dn
, FTAG
);
1320 #endif /* _KERNEL */
1323 * Allocate a loaned anonymous arc buffer.
1326 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1328 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1330 return (arc_loan_buf(db
->db_objset
->os_spa
, size
));
1334 * Free a loaned arc buffer.
1337 dmu_return_arcbuf(arc_buf_t
*buf
)
1339 arc_return_buf(buf
, FTAG
);
1340 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1344 * When possible directly assign passed loaned arc buffer to a dbuf.
1345 * If this is not possible copy the contents of passed arc buf via
1349 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1352 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1355 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1358 DB_DNODE_ENTER(dbuf
);
1359 dn
= DB_DNODE(dbuf
);
1360 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1361 blkid
= dbuf_whichblock(dn
, 0, offset
);
1362 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1363 rw_exit(&dn
->dn_struct_rwlock
);
1364 DB_DNODE_EXIT(dbuf
);
1367 * We can only assign if the offset is aligned, the arc buf is the
1368 * same size as the dbuf, and the dbuf is not metadata. It
1369 * can't be metadata because the loaned arc buf comes from the
1370 * user-data kmem area.
1372 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
&&
1373 DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
) {
1374 dbuf_assign_arcbuf(db
, buf
, tx
);
1375 dbuf_rele(db
, FTAG
);
1380 DB_DNODE_ENTER(dbuf
);
1381 dn
= DB_DNODE(dbuf
);
1383 object
= dn
->dn_object
;
1384 DB_DNODE_EXIT(dbuf
);
1386 dbuf_rele(db
, FTAG
);
1387 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1388 dmu_return_arcbuf(buf
);
1389 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1394 dbuf_dirty_record_t
*dsa_dr
;
1395 dmu_sync_cb_t
*dsa_done
;
1402 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1404 dmu_sync_arg_t
*dsa
= varg
;
1405 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1406 blkptr_t
*bp
= zio
->io_bp
;
1408 if (zio
->io_error
== 0) {
1409 if (BP_IS_HOLE(bp
)) {
1411 * A block of zeros may compress to a hole, but the
1412 * block size still needs to be known for replay.
1414 BP_SET_LSIZE(bp
, db
->db_size
);
1415 } else if (!BP_IS_EMBEDDED(bp
)) {
1416 ASSERT(BP_GET_LEVEL(bp
) == 0);
1423 dmu_sync_late_arrival_ready(zio_t
*zio
)
1425 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1430 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1432 dmu_sync_arg_t
*dsa
= varg
;
1433 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1434 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1436 mutex_enter(&db
->db_mtx
);
1437 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1438 if (zio
->io_error
== 0) {
1439 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1440 if (dr
->dt
.dl
.dr_nopwrite
) {
1441 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1442 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1443 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1445 ASSERT(BP_EQUAL(bp
, bp_orig
));
1446 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1447 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1449 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1450 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1451 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1454 * Old style holes are filled with all zeros, whereas
1455 * new-style holes maintain their lsize, type, level,
1456 * and birth time (see zio_write_compress). While we
1457 * need to reset the BP_SET_LSIZE() call that happened
1458 * in dmu_sync_ready for old style holes, we do *not*
1459 * want to wipe out the information contained in new
1460 * style holes. Thus, only zero out the block pointer if
1461 * it's an old style hole.
1463 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1464 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1465 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1467 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1469 cv_broadcast(&db
->db_changed
);
1470 mutex_exit(&db
->db_mtx
);
1472 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1474 kmem_free(dsa
, sizeof (*dsa
));
1478 dmu_sync_late_arrival_done(zio_t
*zio
)
1480 blkptr_t
*bp
= zio
->io_bp
;
1481 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1482 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1484 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1486 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1487 * then there is nothing to do here. Otherwise, free the
1488 * newly allocated block in this txg.
1490 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1491 ASSERT(BP_EQUAL(bp
, bp_orig
));
1493 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1494 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1495 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1496 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1500 dmu_tx_commit(dsa
->dsa_tx
);
1502 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1504 kmem_free(dsa
, sizeof (*dsa
));
1508 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1509 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1511 dmu_sync_arg_t
*dsa
;
1514 tx
= dmu_tx_create(os
);
1515 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1516 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1518 /* Make zl_get_data do txg_waited_synced() */
1519 return (SET_ERROR(EIO
));
1522 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1524 dsa
->dsa_done
= done
;
1528 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
),
1529 zgd
->zgd_bp
, zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
,
1530 zp
, dmu_sync_late_arrival_ready
, NULL
,
1531 NULL
, dmu_sync_late_arrival_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1532 ZIO_FLAG_CANFAIL
, zb
));
1538 * Intent log support: sync the block associated with db to disk.
1539 * N.B. and XXX: the caller is responsible for making sure that the
1540 * data isn't changing while dmu_sync() is writing it.
1544 * EEXIST: this txg has already been synced, so there's nothing to do.
1545 * The caller should not log the write.
1547 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1548 * The caller should not log the write.
1550 * EALREADY: this block is already in the process of being synced.
1551 * The caller should track its progress (somehow).
1553 * EIO: could not do the I/O.
1554 * The caller should do a txg_wait_synced().
1556 * 0: the I/O has been initiated.
1557 * The caller should log this blkptr in the done callback.
1558 * It is possible that the I/O will fail, in which case
1559 * the error will be reported to the done callback and
1560 * propagated to pio from zio_done().
1563 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1565 blkptr_t
*bp
= zgd
->zgd_bp
;
1566 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1567 objset_t
*os
= db
->db_objset
;
1568 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1569 dbuf_dirty_record_t
*dr
;
1570 dmu_sync_arg_t
*dsa
;
1571 zbookmark_phys_t zb
;
1575 ASSERT(pio
!= NULL
);
1578 SET_BOOKMARK(&zb
, ds
->ds_object
,
1579 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1583 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1587 * If we're frozen (running ziltest), we always need to generate a bp.
1589 if (txg
> spa_freeze_txg(os
->os_spa
))
1590 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1593 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1594 * and us. If we determine that this txg is not yet syncing,
1595 * but it begins to sync a moment later, that's OK because the
1596 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1598 mutex_enter(&db
->db_mtx
);
1600 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1602 * This txg has already synced. There's nothing to do.
1604 mutex_exit(&db
->db_mtx
);
1605 return (SET_ERROR(EEXIST
));
1608 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1610 * This txg is currently syncing, so we can't mess with
1611 * the dirty record anymore; just write a new log block.
1613 mutex_exit(&db
->db_mtx
);
1614 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1617 dr
= db
->db_last_dirty
;
1618 while (dr
&& dr
->dr_txg
!= txg
)
1623 * There's no dr for this dbuf, so it must have been freed.
1624 * There's no need to log writes to freed blocks, so we're done.
1626 mutex_exit(&db
->db_mtx
);
1627 return (SET_ERROR(ENOENT
));
1630 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1633 * Assume the on-disk data is X, the current syncing data (in
1634 * txg - 1) is Y, and the current in-memory data is Z (currently
1637 * We usually want to perform a nopwrite if X and Z are the
1638 * same. However, if Y is different (i.e. the BP is going to
1639 * change before this write takes effect), then a nopwrite will
1640 * be incorrect - we would override with X, which could have
1641 * been freed when Y was written.
1643 * (Note that this is not a concern when we are nop-writing from
1644 * syncing context, because X and Y must be identical, because
1645 * all previous txgs have been synced.)
1647 * Therefore, we disable nopwrite if the current BP could change
1648 * before this TXG. There are two ways it could change: by
1649 * being dirty (dr_next is non-NULL), or by being freed
1650 * (dnode_block_freed()). This behavior is verified by
1651 * zio_done(), which VERIFYs that the override BP is identical
1652 * to the on-disk BP.
1656 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
1657 zp
.zp_nopwrite
= B_FALSE
;
1660 ASSERT(dr
->dr_txg
== txg
);
1661 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1662 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1664 * We have already issued a sync write for this buffer,
1665 * or this buffer has already been synced. It could not
1666 * have been dirtied since, or we would have cleared the state.
1668 mutex_exit(&db
->db_mtx
);
1669 return (SET_ERROR(EALREADY
));
1672 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1673 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1674 mutex_exit(&db
->db_mtx
);
1676 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1678 dsa
->dsa_done
= done
;
1682 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1683 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1684 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1685 NULL
, NULL
, dmu_sync_done
, dsa
,
1686 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, &zb
));
1692 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1698 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1701 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1702 dnode_rele(dn
, FTAG
);
1707 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1713 * Send streams include each object's checksum function. This
1714 * check ensures that the receiving system can understand the
1715 * checksum function transmitted.
1717 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1719 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1720 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1721 dn
->dn_checksum
= checksum
;
1722 dnode_setdirty(dn
, tx
);
1723 dnode_rele(dn
, FTAG
);
1727 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1733 * Send streams include each object's compression function. This
1734 * check ensures that the receiving system can understand the
1735 * compression function transmitted.
1737 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1739 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1740 dn
->dn_compress
= compress
;
1741 dnode_setdirty(dn
, tx
);
1742 dnode_rele(dn
, FTAG
);
1745 int zfs_mdcomp_disable
= 0;
1748 * When the "redundant_metadata" property is set to "most", only indirect
1749 * blocks of this level and higher will have an additional ditto block.
1751 int zfs_redundant_metadata_most_ditto_level
= 2;
1754 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1756 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1757 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1759 enum zio_checksum checksum
= os
->os_checksum
;
1760 enum zio_compress compress
= os
->os_compress
;
1761 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1762 boolean_t dedup
= B_FALSE
;
1763 boolean_t nopwrite
= B_FALSE
;
1764 boolean_t dedup_verify
= os
->os_dedup_verify
;
1765 int copies
= os
->os_copies
;
1768 * We maintain different write policies for each of the following
1771 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1772 * 3. all other level 0 blocks
1775 if (zfs_mdcomp_disable
) {
1776 compress
= ZIO_COMPRESS_EMPTY
;
1779 * XXX -- we should design a compression algorithm
1780 * that specializes in arrays of bps.
1782 compress
= zio_compress_select(os
->os_spa
,
1783 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
1787 * Metadata always gets checksummed. If the data
1788 * checksum is multi-bit correctable, and it's not a
1789 * ZBT-style checksum, then it's suitable for metadata
1790 * as well. Otherwise, the metadata checksum defaults
1793 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1794 zio_checksum_table
[checksum
].ci_eck
)
1795 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1797 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
1798 (os
->os_redundant_metadata
==
1799 ZFS_REDUNDANT_METADATA_MOST
&&
1800 (level
>= zfs_redundant_metadata_most_ditto_level
||
1801 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
1803 } else if (wp
& WP_NOFILL
) {
1807 * If we're writing preallocated blocks, we aren't actually
1808 * writing them so don't set any policy properties. These
1809 * blocks are currently only used by an external subsystem
1810 * outside of zfs (i.e. dump) and not written by the zio
1813 compress
= ZIO_COMPRESS_OFF
;
1814 checksum
= ZIO_CHECKSUM_OFF
;
1816 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
1819 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1820 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1824 * Determine dedup setting. If we are in dmu_sync(),
1825 * we won't actually dedup now because that's all
1826 * done in syncing context; but we do want to use the
1827 * dedup checkum. If the checksum is not strong
1828 * enough to ensure unique signatures, force
1831 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1832 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1833 if (!zio_checksum_table
[checksum
].ci_dedup
)
1834 dedup_verify
= B_TRUE
;
1838 * Enable nopwrite if we have a cryptographically secure
1839 * checksum that has no known collisions (i.e. SHA-256)
1840 * and compression is enabled. We don't enable nopwrite if
1841 * dedup is enabled as the two features are mutually exclusive.
1843 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1844 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1847 zp
->zp_checksum
= checksum
;
1848 zp
->zp_compress
= compress
;
1849 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1850 zp
->zp_level
= level
;
1851 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
1852 zp
->zp_dedup
= dedup
;
1853 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1854 zp
->zp_nopwrite
= nopwrite
;
1858 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1863 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1867 * Sync any current changes before
1868 * we go trundling through the block pointers.
1870 for (i
= 0; i
< TXG_SIZE
; i
++) {
1871 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1874 if (i
!= TXG_SIZE
) {
1875 dnode_rele(dn
, FTAG
);
1876 txg_wait_synced(dmu_objset_pool(os
), 0);
1877 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1882 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1883 dnode_rele(dn
, FTAG
);
1889 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1891 dnode_phys_t
*dnp
= dn
->dn_phys
;
1894 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1895 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1896 1ULL << dn
->dn_indblkshift
: 0;
1897 doi
->doi_type
= dn
->dn_type
;
1898 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1899 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1900 doi
->doi_dnodesize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
1901 doi
->doi_indirection
= dn
->dn_nlevels
;
1902 doi
->doi_checksum
= dn
->dn_checksum
;
1903 doi
->doi_compress
= dn
->dn_compress
;
1904 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
1905 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1906 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1907 doi
->doi_fill_count
= 0;
1908 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1909 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
1913 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1915 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1916 mutex_enter(&dn
->dn_mtx
);
1918 __dmu_object_info_from_dnode(dn
, doi
);
1920 mutex_exit(&dn
->dn_mtx
);
1921 rw_exit(&dn
->dn_struct_rwlock
);
1925 * Get information on a DMU object.
1926 * If doi is NULL, just indicates whether the object exists.
1929 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1932 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1938 dmu_object_info_from_dnode(dn
, doi
);
1940 dnode_rele(dn
, FTAG
);
1945 * As above, but faster; can be used when you have a held dbuf in hand.
1948 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1950 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1953 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1958 * Faster still when you only care about the size.
1959 * This is specifically optimized for zfs_getattr().
1962 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1963 u_longlong_t
*nblk512
)
1965 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1971 *blksize
= dn
->dn_datablksz
;
1972 /* add in number of slots used for the dnode itself */
1973 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1974 SPA_MINBLOCKSHIFT
) + dn
->dn_num_slots
;
1979 dmu_object_dnsize_from_db(dmu_buf_t
*db_fake
, int *dnsize
)
1981 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1986 *dnsize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
1991 byteswap_uint64_array(void *vbuf
, size_t size
)
1993 uint64_t *buf
= vbuf
;
1994 size_t count
= size
>> 3;
1997 ASSERT((size
& 7) == 0);
1999 for (i
= 0; i
< count
; i
++)
2000 buf
[i
] = BSWAP_64(buf
[i
]);
2004 byteswap_uint32_array(void *vbuf
, size_t size
)
2006 uint32_t *buf
= vbuf
;
2007 size_t count
= size
>> 2;
2010 ASSERT((size
& 3) == 0);
2012 for (i
= 0; i
< count
; i
++)
2013 buf
[i
] = BSWAP_32(buf
[i
]);
2017 byteswap_uint16_array(void *vbuf
, size_t size
)
2019 uint16_t *buf
= vbuf
;
2020 size_t count
= size
>> 1;
2023 ASSERT((size
& 1) == 0);
2025 for (i
= 0; i
< count
; i
++)
2026 buf
[i
] = BSWAP_16(buf
[i
]);
2031 byteswap_uint8_array(void *vbuf
, size_t size
)
2053 arc_fini(); /* arc depends on l2arc, so arc must go first */
2065 #if defined(_KERNEL) && defined(HAVE_SPL)
2066 EXPORT_SYMBOL(dmu_bonus_hold
);
2067 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2068 EXPORT_SYMBOL(dmu_buf_rele_array
);
2069 EXPORT_SYMBOL(dmu_prefetch
);
2070 EXPORT_SYMBOL(dmu_free_range
);
2071 EXPORT_SYMBOL(dmu_free_long_range
);
2072 EXPORT_SYMBOL(dmu_free_long_object
);
2073 EXPORT_SYMBOL(dmu_read
);
2074 EXPORT_SYMBOL(dmu_write
);
2075 EXPORT_SYMBOL(dmu_prealloc
);
2076 EXPORT_SYMBOL(dmu_object_info
);
2077 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2078 EXPORT_SYMBOL(dmu_object_info_from_db
);
2079 EXPORT_SYMBOL(dmu_object_size_from_db
);
2080 EXPORT_SYMBOL(dmu_object_dnsize_from_db
);
2081 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2082 EXPORT_SYMBOL(dmu_object_set_checksum
);
2083 EXPORT_SYMBOL(dmu_object_set_compress
);
2084 EXPORT_SYMBOL(dmu_write_policy
);
2085 EXPORT_SYMBOL(dmu_sync
);
2086 EXPORT_SYMBOL(dmu_request_arcbuf
);
2087 EXPORT_SYMBOL(dmu_return_arcbuf
);
2088 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2089 EXPORT_SYMBOL(dmu_buf_hold
);
2090 EXPORT_SYMBOL(dmu_ot
);
2092 module_param(zfs_mdcomp_disable
, int, 0644);
2093 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2095 module_param(zfs_nopwrite_enabled
, int, 0644);
2096 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");