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 && read
&&
489 length
<= zfetch_array_rd_sz
) {
490 dmu_zfetch(&dn
->dn_zfetch
, blkid
, nblks
);
492 rw_exit(&dn
->dn_struct_rwlock
);
494 /* wait for async i/o */
497 dmu_buf_rele_array(dbp
, nblks
, tag
);
501 /* wait for other io to complete */
503 for (i
= 0; i
< nblks
; i
++) {
504 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
505 mutex_enter(&db
->db_mtx
);
506 while (db
->db_state
== DB_READ
||
507 db
->db_state
== DB_FILL
)
508 cv_wait(&db
->db_changed
, &db
->db_mtx
);
509 if (db
->db_state
== DB_UNCACHED
)
510 err
= SET_ERROR(EIO
);
511 mutex_exit(&db
->db_mtx
);
513 dmu_buf_rele_array(dbp
, nblks
, tag
);
525 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
526 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
531 err
= dnode_hold(os
, object
, FTAG
, &dn
);
535 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
536 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
538 dnode_rele(dn
, FTAG
);
544 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
545 uint64_t length
, boolean_t read
, void *tag
, int *numbufsp
,
548 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
554 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
555 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
562 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
565 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
570 for (i
= 0; i
< numbufs
; i
++) {
572 dbuf_rele(dbp
[i
], tag
);
575 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
579 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
580 * indirect blocks prefeteched will be those that point to the blocks containing
581 * the data starting at offset, and continuing to offset + len.
583 * Note that if the indirect blocks above the blocks being prefetched are not in
584 * cache, they will be asychronously read in.
587 dmu_prefetch(objset_t
*os
, uint64_t object
, int64_t level
, uint64_t offset
,
588 uint64_t len
, zio_priority_t pri
)
594 if (len
== 0) { /* they're interested in the bonus buffer */
595 dn
= DMU_META_DNODE(os
);
597 if (object
== 0 || object
>= DN_MAX_OBJECT
)
600 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
601 blkid
= dbuf_whichblock(dn
, level
,
602 object
* sizeof (dnode_phys_t
));
603 dbuf_prefetch(dn
, level
, blkid
, pri
, 0);
604 rw_exit(&dn
->dn_struct_rwlock
);
609 * XXX - Note, if the dnode for the requested object is not
610 * already cached, we will do a *synchronous* read in the
611 * dnode_hold() call. The same is true for any indirects.
613 err
= dnode_hold(os
, object
, FTAG
, &dn
);
617 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
619 * offset + len - 1 is the last byte we want to prefetch for, and offset
620 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
621 * last block we want to prefetch, and dbuf_whichblock(dn, level,
622 * offset) is the first. Then the number we need to prefetch is the
625 if (level
> 0 || dn
->dn_datablkshift
!= 0) {
626 nblks
= dbuf_whichblock(dn
, level
, offset
+ len
- 1) -
627 dbuf_whichblock(dn
, level
, offset
) + 1;
629 nblks
= (offset
< dn
->dn_datablksz
);
635 blkid
= dbuf_whichblock(dn
, level
, offset
);
636 for (i
= 0; i
< nblks
; i
++)
637 dbuf_prefetch(dn
, level
, blkid
+ i
, pri
, 0);
640 rw_exit(&dn
->dn_struct_rwlock
);
642 dnode_rele(dn
, FTAG
);
646 * Get the next "chunk" of file data to free. We traverse the file from
647 * the end so that the file gets shorter over time (if we crashes in the
648 * middle, this will leave us in a better state). We find allocated file
649 * data by simply searching the allocated level 1 indirects.
651 * On input, *start should be the first offset that does not need to be
652 * freed (e.g. "offset + length"). On return, *start will be the first
653 * offset that should be freed.
656 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
658 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
659 /* bytes of data covered by a level-1 indirect block */
661 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
664 ASSERT3U(minimum
, <=, *start
);
666 if (*start
- minimum
<= iblkrange
* maxblks
) {
670 ASSERT(ISP2(iblkrange
));
672 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
676 * dnode_next_offset(BACKWARDS) will find an allocated L1
677 * indirect block at or before the input offset. We must
678 * decrement *start so that it is at the end of the region
682 err
= dnode_next_offset(dn
,
683 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
685 /* if there are no indirect blocks before start, we are done */
689 } else if (err
!= 0) {
693 /* set start to the beginning of this L1 indirect */
694 *start
= P2ALIGN(*start
, iblkrange
);
696 if (*start
< minimum
)
702 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
705 uint64_t object_size
;
709 return (SET_ERROR(EINVAL
));
711 object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
712 if (offset
>= object_size
)
715 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
716 length
= object_size
- offset
;
718 while (length
!= 0) {
719 uint64_t chunk_end
, chunk_begin
;
722 chunk_end
= chunk_begin
= offset
+ length
;
724 /* move chunk_begin backwards to the beginning of this chunk */
725 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
728 ASSERT3U(chunk_begin
, >=, offset
);
729 ASSERT3U(chunk_begin
, <=, chunk_end
);
731 tx
= dmu_tx_create(os
);
732 dmu_tx_hold_free(tx
, dn
->dn_object
,
733 chunk_begin
, chunk_end
- chunk_begin
);
736 * Mark this transaction as typically resulting in a net
737 * reduction in space used.
739 dmu_tx_mark_netfree(tx
);
740 err
= dmu_tx_assign(tx
, TXG_WAIT
);
745 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
748 length
-= chunk_end
- chunk_begin
;
754 dmu_free_long_range(objset_t
*os
, uint64_t object
,
755 uint64_t offset
, uint64_t length
)
760 err
= dnode_hold(os
, object
, FTAG
, &dn
);
763 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
766 * It is important to zero out the maxblkid when freeing the entire
767 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
768 * will take the fast path, and (b) dnode_reallocate() can verify
769 * that the entire file has been freed.
771 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
774 dnode_rele(dn
, FTAG
);
779 dmu_free_long_object(objset_t
*os
, uint64_t object
)
784 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
788 tx
= dmu_tx_create(os
);
789 dmu_tx_hold_bonus(tx
, object
);
790 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
791 dmu_tx_mark_netfree(tx
);
792 err
= dmu_tx_assign(tx
, TXG_WAIT
);
794 err
= dmu_object_free(os
, object
, tx
);
804 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
805 uint64_t size
, dmu_tx_t
*tx
)
808 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
811 ASSERT(offset
< UINT64_MAX
);
812 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
813 dnode_free_range(dn
, offset
, size
, tx
);
814 dnode_rele(dn
, FTAG
);
819 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
820 void *buf
, uint32_t flags
)
826 err
= dnode_hold(os
, object
, FTAG
, &dn
);
831 * Deal with odd block sizes, where there can't be data past the first
832 * block. If we ever do the tail block optimization, we will need to
833 * handle that here as well.
835 if (dn
->dn_maxblkid
== 0) {
836 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
837 MIN(size
, dn
->dn_datablksz
- offset
);
838 bzero((char *)buf
+ newsz
, size
- newsz
);
843 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
847 * NB: we could do this block-at-a-time, but it's nice
848 * to be reading in parallel.
850 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
851 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
855 for (i
= 0; i
< numbufs
; i
++) {
858 dmu_buf_t
*db
= dbp
[i
];
862 bufoff
= offset
- db
->db_offset
;
863 tocpy
= MIN(db
->db_size
- bufoff
, size
);
865 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
869 buf
= (char *)buf
+ tocpy
;
871 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
873 dnode_rele(dn
, FTAG
);
878 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
879 const void *buf
, dmu_tx_t
*tx
)
887 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
888 FALSE
, FTAG
, &numbufs
, &dbp
));
890 for (i
= 0; i
< numbufs
; i
++) {
893 dmu_buf_t
*db
= dbp
[i
];
897 bufoff
= offset
- db
->db_offset
;
898 tocpy
= MIN(db
->db_size
- bufoff
, size
);
900 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
902 if (tocpy
== db
->db_size
)
903 dmu_buf_will_fill(db
, tx
);
905 dmu_buf_will_dirty(db
, tx
);
907 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
909 if (tocpy
== db
->db_size
)
910 dmu_buf_fill_done(db
, tx
);
914 buf
= (char *)buf
+ tocpy
;
916 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
920 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
929 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
930 FALSE
, FTAG
, &numbufs
, &dbp
));
932 for (i
= 0; i
< numbufs
; i
++) {
933 dmu_buf_t
*db
= dbp
[i
];
935 dmu_buf_will_not_fill(db
, tx
);
937 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
941 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
942 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
943 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
947 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
948 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
949 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
952 dmu_buf_write_embedded(db
,
953 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
954 uncompressed_size
, compressed_size
, byteorder
, tx
);
956 dmu_buf_rele(db
, FTAG
);
960 * DMU support for xuio
962 kstat_t
*xuio_ksp
= NULL
;
964 typedef struct xuio_stats
{
965 /* loaned yet not returned arc_buf */
966 kstat_named_t xuiostat_onloan_rbuf
;
967 kstat_named_t xuiostat_onloan_wbuf
;
968 /* whether a copy is made when loaning out a read buffer */
969 kstat_named_t xuiostat_rbuf_copied
;
970 kstat_named_t xuiostat_rbuf_nocopy
;
971 /* whether a copy is made when assigning a write buffer */
972 kstat_named_t xuiostat_wbuf_copied
;
973 kstat_named_t xuiostat_wbuf_nocopy
;
976 static xuio_stats_t xuio_stats
= {
977 { "onloan_read_buf", KSTAT_DATA_UINT64
},
978 { "onloan_write_buf", KSTAT_DATA_UINT64
},
979 { "read_buf_copied", KSTAT_DATA_UINT64
},
980 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
981 { "write_buf_copied", KSTAT_DATA_UINT64
},
982 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
985 #define XUIOSTAT_INCR(stat, val) \
986 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
987 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
990 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
993 uio_t
*uio
= &xuio
->xu_uio
;
995 uio
->uio_iovcnt
= nblk
;
996 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
998 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
1000 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
1001 priv
->iovp
= (iovec_t
*)uio
->uio_iov
;
1002 XUIO_XUZC_PRIV(xuio
) = priv
;
1004 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1005 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
1007 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
1013 dmu_xuio_fini(xuio_t
*xuio
)
1015 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1016 int nblk
= priv
->cnt
;
1018 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
1019 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
1020 kmem_free(priv
, sizeof (dmu_xuio_t
));
1022 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1023 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
1025 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
1029 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1030 * and increase priv->next by 1.
1033 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
1036 uio_t
*uio
= &xuio
->xu_uio
;
1037 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1038 int i
= priv
->next
++;
1040 ASSERT(i
< priv
->cnt
);
1041 ASSERT(off
+ n
<= arc_buf_size(abuf
));
1042 iov
= (iovec_t
*)uio
->uio_iov
+ i
;
1043 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
1045 priv
->bufs
[i
] = abuf
;
1050 dmu_xuio_cnt(xuio_t
*xuio
)
1052 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1057 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
1059 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1061 ASSERT(i
< priv
->cnt
);
1062 return (priv
->bufs
[i
]);
1066 dmu_xuio_clear(xuio_t
*xuio
, int i
)
1068 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1070 ASSERT(i
< priv
->cnt
);
1071 priv
->bufs
[i
] = NULL
;
1075 xuio_stat_init(void)
1077 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1078 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1079 KSTAT_FLAG_VIRTUAL
);
1080 if (xuio_ksp
!= NULL
) {
1081 xuio_ksp
->ks_data
= &xuio_stats
;
1082 kstat_install(xuio_ksp
);
1087 xuio_stat_fini(void)
1089 if (xuio_ksp
!= NULL
) {
1090 kstat_delete(xuio_ksp
);
1096 xuio_stat_wbuf_copied()
1098 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1102 xuio_stat_wbuf_nocopy()
1104 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1109 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1112 int numbufs
, i
, err
;
1113 xuio_t
*xuio
= NULL
;
1116 * NB: we could do this block-at-a-time, but it's nice
1117 * to be reading in parallel.
1119 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1120 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1124 for (i
= 0; i
< numbufs
; i
++) {
1127 dmu_buf_t
*db
= dbp
[i
];
1131 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1132 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1135 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1136 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1137 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1138 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1140 uio
->uio_resid
-= tocpy
;
1141 uio
->uio_loffset
+= tocpy
;
1144 if (abuf
== dbuf_abuf
)
1145 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1147 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1149 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1157 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1163 * Read 'size' bytes into the uio buffer.
1164 * From object zdb->db_object.
1165 * Starting at offset uio->uio_loffset.
1167 * If the caller already has a dbuf in the target object
1168 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1169 * because we don't have to find the dnode_t for the object.
1172 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1174 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1183 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1190 * Read 'size' bytes into the uio buffer.
1191 * From the specified object
1192 * Starting at offset uio->uio_loffset.
1195 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1203 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1207 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1209 dnode_rele(dn
, FTAG
);
1215 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1222 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1223 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1227 for (i
= 0; i
< numbufs
; i
++) {
1230 dmu_buf_t
*db
= dbp
[i
];
1234 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1235 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1237 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1239 if (tocpy
== db
->db_size
)
1240 dmu_buf_will_fill(db
, tx
);
1242 dmu_buf_will_dirty(db
, tx
);
1245 * XXX uiomove could block forever (eg.nfs-backed
1246 * pages). There needs to be a uiolockdown() function
1247 * to lock the pages in memory, so that uiomove won't
1250 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1253 if (tocpy
== db
->db_size
)
1254 dmu_buf_fill_done(db
, tx
);
1262 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1267 * Write 'size' bytes from the uio buffer.
1268 * To object zdb->db_object.
1269 * Starting at offset uio->uio_loffset.
1271 * If the caller already has a dbuf in the target object
1272 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1273 * because we don't have to find the dnode_t for the object.
1276 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1279 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1288 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1295 * Write 'size' bytes from the uio buffer.
1296 * To the specified object.
1297 * Starting at offset uio->uio_loffset.
1300 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1309 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1313 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1315 dnode_rele(dn
, FTAG
);
1319 #endif /* _KERNEL */
1322 * Allocate a loaned anonymous arc buffer.
1325 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1327 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1329 return (arc_loan_buf(db
->db_objset
->os_spa
, size
));
1333 * Free a loaned arc buffer.
1336 dmu_return_arcbuf(arc_buf_t
*buf
)
1338 arc_return_buf(buf
, FTAG
);
1339 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1343 * When possible directly assign passed loaned arc buffer to a dbuf.
1344 * If this is not possible copy the contents of passed arc buf via
1348 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1351 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1354 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1357 DB_DNODE_ENTER(dbuf
);
1358 dn
= DB_DNODE(dbuf
);
1359 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1360 blkid
= dbuf_whichblock(dn
, 0, offset
);
1361 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1362 rw_exit(&dn
->dn_struct_rwlock
);
1363 DB_DNODE_EXIT(dbuf
);
1366 * We can only assign if the offset is aligned, the arc buf is the
1367 * same size as the dbuf, and the dbuf is not metadata. It
1368 * can't be metadata because the loaned arc buf comes from the
1369 * user-data kmem area.
1371 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
&&
1372 DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
) {
1373 dbuf_assign_arcbuf(db
, buf
, tx
);
1374 dbuf_rele(db
, FTAG
);
1379 DB_DNODE_ENTER(dbuf
);
1380 dn
= DB_DNODE(dbuf
);
1382 object
= dn
->dn_object
;
1383 DB_DNODE_EXIT(dbuf
);
1385 dbuf_rele(db
, FTAG
);
1386 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1387 dmu_return_arcbuf(buf
);
1388 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1393 dbuf_dirty_record_t
*dsa_dr
;
1394 dmu_sync_cb_t
*dsa_done
;
1401 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1403 dmu_sync_arg_t
*dsa
= varg
;
1404 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1405 blkptr_t
*bp
= zio
->io_bp
;
1407 if (zio
->io_error
== 0) {
1408 if (BP_IS_HOLE(bp
)) {
1410 * A block of zeros may compress to a hole, but the
1411 * block size still needs to be known for replay.
1413 BP_SET_LSIZE(bp
, db
->db_size
);
1414 } else if (!BP_IS_EMBEDDED(bp
)) {
1415 ASSERT(BP_GET_LEVEL(bp
) == 0);
1422 dmu_sync_late_arrival_ready(zio_t
*zio
)
1424 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1429 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1431 dmu_sync_arg_t
*dsa
= varg
;
1432 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1433 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1435 mutex_enter(&db
->db_mtx
);
1436 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1437 if (zio
->io_error
== 0) {
1438 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1439 if (dr
->dt
.dl
.dr_nopwrite
) {
1440 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1441 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1442 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1444 ASSERT(BP_EQUAL(bp
, bp_orig
));
1445 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1446 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1448 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1449 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1450 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1453 * Old style holes are filled with all zeros, whereas
1454 * new-style holes maintain their lsize, type, level,
1455 * and birth time (see zio_write_compress). While we
1456 * need to reset the BP_SET_LSIZE() call that happened
1457 * in dmu_sync_ready for old style holes, we do *not*
1458 * want to wipe out the information contained in new
1459 * style holes. Thus, only zero out the block pointer if
1460 * it's an old style hole.
1462 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1463 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1464 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1466 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1468 cv_broadcast(&db
->db_changed
);
1469 mutex_exit(&db
->db_mtx
);
1471 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1473 kmem_free(dsa
, sizeof (*dsa
));
1477 dmu_sync_late_arrival_done(zio_t
*zio
)
1479 blkptr_t
*bp
= zio
->io_bp
;
1480 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1481 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1483 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1485 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1486 * then there is nothing to do here. Otherwise, free the
1487 * newly allocated block in this txg.
1489 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1490 ASSERT(BP_EQUAL(bp
, bp_orig
));
1492 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1493 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1494 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1495 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1499 dmu_tx_commit(dsa
->dsa_tx
);
1501 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1503 kmem_free(dsa
, sizeof (*dsa
));
1507 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1508 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1510 dmu_sync_arg_t
*dsa
;
1513 tx
= dmu_tx_create(os
);
1514 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1515 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1517 /* Make zl_get_data do txg_waited_synced() */
1518 return (SET_ERROR(EIO
));
1521 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1523 dsa
->dsa_done
= done
;
1527 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
),
1528 zgd
->zgd_bp
, zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
,
1529 zp
, dmu_sync_late_arrival_ready
, NULL
,
1530 NULL
, dmu_sync_late_arrival_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1531 ZIO_FLAG_CANFAIL
, zb
));
1537 * Intent log support: sync the block associated with db to disk.
1538 * N.B. and XXX: the caller is responsible for making sure that the
1539 * data isn't changing while dmu_sync() is writing it.
1543 * EEXIST: this txg has already been synced, so there's nothing to do.
1544 * The caller should not log the write.
1546 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1547 * The caller should not log the write.
1549 * EALREADY: this block is already in the process of being synced.
1550 * The caller should track its progress (somehow).
1552 * EIO: could not do the I/O.
1553 * The caller should do a txg_wait_synced().
1555 * 0: the I/O has been initiated.
1556 * The caller should log this blkptr in the done callback.
1557 * It is possible that the I/O will fail, in which case
1558 * the error will be reported to the done callback and
1559 * propagated to pio from zio_done().
1562 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1564 blkptr_t
*bp
= zgd
->zgd_bp
;
1565 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1566 objset_t
*os
= db
->db_objset
;
1567 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1568 dbuf_dirty_record_t
*dr
;
1569 dmu_sync_arg_t
*dsa
;
1570 zbookmark_phys_t zb
;
1574 ASSERT(pio
!= NULL
);
1577 SET_BOOKMARK(&zb
, ds
->ds_object
,
1578 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1582 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1586 * If we're frozen (running ziltest), we always need to generate a bp.
1588 if (txg
> spa_freeze_txg(os
->os_spa
))
1589 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1592 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1593 * and us. If we determine that this txg is not yet syncing,
1594 * but it begins to sync a moment later, that's OK because the
1595 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1597 mutex_enter(&db
->db_mtx
);
1599 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1601 * This txg has already synced. There's nothing to do.
1603 mutex_exit(&db
->db_mtx
);
1604 return (SET_ERROR(EEXIST
));
1607 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1609 * This txg is currently syncing, so we can't mess with
1610 * the dirty record anymore; just write a new log block.
1612 mutex_exit(&db
->db_mtx
);
1613 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1616 dr
= db
->db_last_dirty
;
1617 while (dr
&& dr
->dr_txg
!= txg
)
1622 * There's no dr for this dbuf, so it must have been freed.
1623 * There's no need to log writes to freed blocks, so we're done.
1625 mutex_exit(&db
->db_mtx
);
1626 return (SET_ERROR(ENOENT
));
1629 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1632 * Assume the on-disk data is X, the current syncing data (in
1633 * txg - 1) is Y, and the current in-memory data is Z (currently
1636 * We usually want to perform a nopwrite if X and Z are the
1637 * same. However, if Y is different (i.e. the BP is going to
1638 * change before this write takes effect), then a nopwrite will
1639 * be incorrect - we would override with X, which could have
1640 * been freed when Y was written.
1642 * (Note that this is not a concern when we are nop-writing from
1643 * syncing context, because X and Y must be identical, because
1644 * all previous txgs have been synced.)
1646 * Therefore, we disable nopwrite if the current BP could change
1647 * before this TXG. There are two ways it could change: by
1648 * being dirty (dr_next is non-NULL), or by being freed
1649 * (dnode_block_freed()). This behavior is verified by
1650 * zio_done(), which VERIFYs that the override BP is identical
1651 * to the on-disk BP.
1655 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
1656 zp
.zp_nopwrite
= B_FALSE
;
1659 ASSERT(dr
->dr_txg
== txg
);
1660 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1661 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1663 * We have already issued a sync write for this buffer,
1664 * or this buffer has already been synced. It could not
1665 * have been dirtied since, or we would have cleared the state.
1667 mutex_exit(&db
->db_mtx
);
1668 return (SET_ERROR(EALREADY
));
1671 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1672 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1673 mutex_exit(&db
->db_mtx
);
1675 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1677 dsa
->dsa_done
= done
;
1681 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1682 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1683 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1684 NULL
, NULL
, dmu_sync_done
, dsa
,
1685 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, &zb
));
1691 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1697 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1700 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1701 dnode_rele(dn
, FTAG
);
1706 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1712 * Send streams include each object's checksum function. This
1713 * check ensures that the receiving system can understand the
1714 * checksum function transmitted.
1716 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1718 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1719 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1720 dn
->dn_checksum
= checksum
;
1721 dnode_setdirty(dn
, tx
);
1722 dnode_rele(dn
, FTAG
);
1726 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1732 * Send streams include each object's compression function. This
1733 * check ensures that the receiving system can understand the
1734 * compression function transmitted.
1736 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1738 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1739 dn
->dn_compress
= compress
;
1740 dnode_setdirty(dn
, tx
);
1741 dnode_rele(dn
, FTAG
);
1744 int zfs_mdcomp_disable
= 0;
1747 * When the "redundant_metadata" property is set to "most", only indirect
1748 * blocks of this level and higher will have an additional ditto block.
1750 int zfs_redundant_metadata_most_ditto_level
= 2;
1753 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1755 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1756 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1758 enum zio_checksum checksum
= os
->os_checksum
;
1759 enum zio_compress compress
= os
->os_compress
;
1760 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1761 boolean_t dedup
= B_FALSE
;
1762 boolean_t nopwrite
= B_FALSE
;
1763 boolean_t dedup_verify
= os
->os_dedup_verify
;
1764 int copies
= os
->os_copies
;
1767 * We maintain different write policies for each of the following
1770 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1771 * 3. all other level 0 blocks
1774 if (zfs_mdcomp_disable
) {
1775 compress
= ZIO_COMPRESS_EMPTY
;
1778 * XXX -- we should design a compression algorithm
1779 * that specializes in arrays of bps.
1781 compress
= zio_compress_select(os
->os_spa
,
1782 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
1786 * Metadata always gets checksummed. If the data
1787 * checksum is multi-bit correctable, and it's not a
1788 * ZBT-style checksum, then it's suitable for metadata
1789 * as well. Otherwise, the metadata checksum defaults
1792 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1793 zio_checksum_table
[checksum
].ci_eck
)
1794 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1796 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
1797 (os
->os_redundant_metadata
==
1798 ZFS_REDUNDANT_METADATA_MOST
&&
1799 (level
>= zfs_redundant_metadata_most_ditto_level
||
1800 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
1802 } else if (wp
& WP_NOFILL
) {
1806 * If we're writing preallocated blocks, we aren't actually
1807 * writing them so don't set any policy properties. These
1808 * blocks are currently only used by an external subsystem
1809 * outside of zfs (i.e. dump) and not written by the zio
1812 compress
= ZIO_COMPRESS_OFF
;
1813 checksum
= ZIO_CHECKSUM_OFF
;
1815 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
1818 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1819 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1823 * Determine dedup setting. If we are in dmu_sync(),
1824 * we won't actually dedup now because that's all
1825 * done in syncing context; but we do want to use the
1826 * dedup checkum. If the checksum is not strong
1827 * enough to ensure unique signatures, force
1830 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1831 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1832 if (!zio_checksum_table
[checksum
].ci_dedup
)
1833 dedup_verify
= B_TRUE
;
1837 * Enable nopwrite if we have a cryptographically secure
1838 * checksum that has no known collisions (i.e. SHA-256)
1839 * and compression is enabled. We don't enable nopwrite if
1840 * dedup is enabled as the two features are mutually exclusive.
1842 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1843 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1846 zp
->zp_checksum
= checksum
;
1847 zp
->zp_compress
= compress
;
1848 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1849 zp
->zp_level
= level
;
1850 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
1851 zp
->zp_dedup
= dedup
;
1852 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1853 zp
->zp_nopwrite
= nopwrite
;
1857 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1862 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1866 * Sync any current changes before
1867 * we go trundling through the block pointers.
1869 for (i
= 0; i
< TXG_SIZE
; i
++) {
1870 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1873 if (i
!= TXG_SIZE
) {
1874 dnode_rele(dn
, FTAG
);
1875 txg_wait_synced(dmu_objset_pool(os
), 0);
1876 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1881 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1882 dnode_rele(dn
, FTAG
);
1888 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1890 dnode_phys_t
*dnp
= dn
->dn_phys
;
1893 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1894 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1895 1ULL << dn
->dn_indblkshift
: 0;
1896 doi
->doi_type
= dn
->dn_type
;
1897 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1898 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1899 doi
->doi_dnodesize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
1900 doi
->doi_indirection
= dn
->dn_nlevels
;
1901 doi
->doi_checksum
= dn
->dn_checksum
;
1902 doi
->doi_compress
= dn
->dn_compress
;
1903 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
1904 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1905 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1906 doi
->doi_fill_count
= 0;
1907 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1908 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
1912 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1914 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1915 mutex_enter(&dn
->dn_mtx
);
1917 __dmu_object_info_from_dnode(dn
, doi
);
1919 mutex_exit(&dn
->dn_mtx
);
1920 rw_exit(&dn
->dn_struct_rwlock
);
1924 * Get information on a DMU object.
1925 * If doi is NULL, just indicates whether the object exists.
1928 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1931 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1937 dmu_object_info_from_dnode(dn
, doi
);
1939 dnode_rele(dn
, FTAG
);
1944 * As above, but faster; can be used when you have a held dbuf in hand.
1947 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1949 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1952 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1957 * Faster still when you only care about the size.
1958 * This is specifically optimized for zfs_getattr().
1961 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1962 u_longlong_t
*nblk512
)
1964 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1970 *blksize
= dn
->dn_datablksz
;
1971 /* add in number of slots used for the dnode itself */
1972 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1973 SPA_MINBLOCKSHIFT
) + dn
->dn_num_slots
;
1978 dmu_object_dnsize_from_db(dmu_buf_t
*db_fake
, int *dnsize
)
1980 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1985 *dnsize
= dn
->dn_num_slots
<< DNODE_SHIFT
;
1990 byteswap_uint64_array(void *vbuf
, size_t size
)
1992 uint64_t *buf
= vbuf
;
1993 size_t count
= size
>> 3;
1996 ASSERT((size
& 7) == 0);
1998 for (i
= 0; i
< count
; i
++)
1999 buf
[i
] = BSWAP_64(buf
[i
]);
2003 byteswap_uint32_array(void *vbuf
, size_t size
)
2005 uint32_t *buf
= vbuf
;
2006 size_t count
= size
>> 2;
2009 ASSERT((size
& 3) == 0);
2011 for (i
= 0; i
< count
; i
++)
2012 buf
[i
] = BSWAP_32(buf
[i
]);
2016 byteswap_uint16_array(void *vbuf
, size_t size
)
2018 uint16_t *buf
= vbuf
;
2019 size_t count
= size
>> 1;
2022 ASSERT((size
& 1) == 0);
2024 for (i
= 0; i
< count
; i
++)
2025 buf
[i
] = BSWAP_16(buf
[i
]);
2030 byteswap_uint8_array(void *vbuf
, size_t size
)
2052 arc_fini(); /* arc depends on l2arc, so arc must go first */
2064 #if defined(_KERNEL) && defined(HAVE_SPL)
2065 EXPORT_SYMBOL(dmu_bonus_hold
);
2066 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2067 EXPORT_SYMBOL(dmu_buf_rele_array
);
2068 EXPORT_SYMBOL(dmu_prefetch
);
2069 EXPORT_SYMBOL(dmu_free_range
);
2070 EXPORT_SYMBOL(dmu_free_long_range
);
2071 EXPORT_SYMBOL(dmu_free_long_object
);
2072 EXPORT_SYMBOL(dmu_read
);
2073 EXPORT_SYMBOL(dmu_write
);
2074 EXPORT_SYMBOL(dmu_prealloc
);
2075 EXPORT_SYMBOL(dmu_object_info
);
2076 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2077 EXPORT_SYMBOL(dmu_object_info_from_db
);
2078 EXPORT_SYMBOL(dmu_object_size_from_db
);
2079 EXPORT_SYMBOL(dmu_object_dnsize_from_db
);
2080 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2081 EXPORT_SYMBOL(dmu_object_set_checksum
);
2082 EXPORT_SYMBOL(dmu_object_set_compress
);
2083 EXPORT_SYMBOL(dmu_write_policy
);
2084 EXPORT_SYMBOL(dmu_sync
);
2085 EXPORT_SYMBOL(dmu_request_arcbuf
);
2086 EXPORT_SYMBOL(dmu_return_arcbuf
);
2087 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2088 EXPORT_SYMBOL(dmu_buf_hold
);
2089 EXPORT_SYMBOL(dmu_ot
);
2091 module_param(zfs_mdcomp_disable
, int, 0644);
2092 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2094 module_param(zfs_nopwrite_enabled
, int, 0644);
2095 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");