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, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved.
26 * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
48 #include <sys/zfeature.h>
50 #include <sys/vmsystm.h>
51 #include <sys/zfs_znode.h>
55 * Enable/disable nopwrite feature.
57 int zfs_nopwrite_enabled
= 1;
59 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
60 { DMU_BSWAP_UINT8
, TRUE
, "unallocated" },
61 { DMU_BSWAP_ZAP
, TRUE
, "object directory" },
62 { DMU_BSWAP_UINT64
, TRUE
, "object array" },
63 { DMU_BSWAP_UINT8
, TRUE
, "packed nvlist" },
64 { DMU_BSWAP_UINT64
, TRUE
, "packed nvlist size" },
65 { DMU_BSWAP_UINT64
, TRUE
, "bpobj" },
66 { DMU_BSWAP_UINT64
, TRUE
, "bpobj header" },
67 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map header" },
68 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map" },
69 { DMU_BSWAP_UINT64
, TRUE
, "ZIL intent log" },
70 { DMU_BSWAP_DNODE
, TRUE
, "DMU dnode" },
71 { DMU_BSWAP_OBJSET
, TRUE
, "DMU objset" },
72 { DMU_BSWAP_UINT64
, TRUE
, "DSL directory" },
73 { DMU_BSWAP_ZAP
, TRUE
, "DSL directory child map"},
74 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset snap map" },
75 { DMU_BSWAP_ZAP
, TRUE
, "DSL props" },
76 { DMU_BSWAP_UINT64
, TRUE
, "DSL dataset" },
77 { DMU_BSWAP_ZNODE
, TRUE
, "ZFS znode" },
78 { DMU_BSWAP_OLDACL
, TRUE
, "ZFS V0 ACL" },
79 { DMU_BSWAP_UINT8
, FALSE
, "ZFS plain file" },
80 { DMU_BSWAP_ZAP
, TRUE
, "ZFS directory" },
81 { DMU_BSWAP_ZAP
, TRUE
, "ZFS master node" },
82 { DMU_BSWAP_ZAP
, TRUE
, "ZFS delete queue" },
83 { DMU_BSWAP_UINT8
, FALSE
, "zvol object" },
84 { DMU_BSWAP_ZAP
, TRUE
, "zvol prop" },
85 { DMU_BSWAP_UINT8
, FALSE
, "other uint8[]" },
86 { DMU_BSWAP_UINT64
, FALSE
, "other uint64[]" },
87 { DMU_BSWAP_ZAP
, TRUE
, "other ZAP" },
88 { DMU_BSWAP_ZAP
, TRUE
, "persistent error log" },
89 { DMU_BSWAP_UINT8
, TRUE
, "SPA history" },
90 { DMU_BSWAP_UINT64
, TRUE
, "SPA history offsets" },
91 { DMU_BSWAP_ZAP
, TRUE
, "Pool properties" },
92 { DMU_BSWAP_ZAP
, TRUE
, "DSL permissions" },
93 { DMU_BSWAP_ACL
, TRUE
, "ZFS ACL" },
94 { DMU_BSWAP_UINT8
, TRUE
, "ZFS SYSACL" },
95 { DMU_BSWAP_UINT8
, TRUE
, "FUID table" },
96 { DMU_BSWAP_UINT64
, TRUE
, "FUID table size" },
97 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset next clones"},
98 { DMU_BSWAP_ZAP
, TRUE
, "scan work queue" },
99 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group used" },
100 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group quota" },
101 { DMU_BSWAP_ZAP
, TRUE
, "snapshot refcount tags"},
102 { DMU_BSWAP_ZAP
, TRUE
, "DDT ZAP algorithm" },
103 { DMU_BSWAP_ZAP
, TRUE
, "DDT statistics" },
104 { DMU_BSWAP_UINT8
, TRUE
, "System attributes" },
105 { DMU_BSWAP_ZAP
, TRUE
, "SA master node" },
106 { DMU_BSWAP_ZAP
, TRUE
, "SA attr registration" },
107 { DMU_BSWAP_ZAP
, TRUE
, "SA attr layouts" },
108 { DMU_BSWAP_ZAP
, TRUE
, "scan translations" },
109 { DMU_BSWAP_UINT8
, FALSE
, "deduplicated block" },
110 { DMU_BSWAP_ZAP
, TRUE
, "DSL deadlist map" },
111 { DMU_BSWAP_UINT64
, TRUE
, "DSL deadlist map hdr" },
112 { DMU_BSWAP_ZAP
, TRUE
, "DSL dir clones" },
113 { DMU_BSWAP_UINT64
, TRUE
, "bpobj subobj" }
116 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
117 { byteswap_uint8_array
, "uint8" },
118 { byteswap_uint16_array
, "uint16" },
119 { byteswap_uint32_array
, "uint32" },
120 { byteswap_uint64_array
, "uint64" },
121 { zap_byteswap
, "zap" },
122 { dnode_buf_byteswap
, "dnode" },
123 { dmu_objset_byteswap
, "objset" },
124 { zfs_znode_byteswap
, "znode" },
125 { zfs_oldacl_byteswap
, "oldacl" },
126 { zfs_acl_byteswap
, "acl" }
130 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
131 void *tag
, dmu_buf_t
**dbp
)
138 err
= dnode_hold(os
, object
, FTAG
, &dn
);
141 blkid
= dbuf_whichblock(dn
, 0, offset
);
142 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
143 db
= dbuf_hold(dn
, blkid
, tag
);
144 rw_exit(&dn
->dn_struct_rwlock
);
145 dnode_rele(dn
, FTAG
);
149 return (SET_ERROR(EIO
));
157 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
158 void *tag
, dmu_buf_t
**dbp
, int flags
)
161 int db_flags
= DB_RF_CANFAIL
;
163 if (flags
& DMU_READ_NO_PREFETCH
)
164 db_flags
|= DB_RF_NOPREFETCH
;
166 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
168 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
169 err
= dbuf_read(db
, NULL
, db_flags
);
182 return (DN_MAX_BONUSLEN
);
186 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
188 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
195 if (dn
->dn_bonus
!= db
) {
196 error
= SET_ERROR(EINVAL
);
197 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
198 error
= SET_ERROR(EINVAL
);
200 dnode_setbonuslen(dn
, newsize
, tx
);
209 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
211 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
218 if (!DMU_OT_IS_VALID(type
)) {
219 error
= SET_ERROR(EINVAL
);
220 } else if (dn
->dn_bonus
!= db
) {
221 error
= SET_ERROR(EINVAL
);
223 dnode_setbonus_type(dn
, type
, tx
);
232 dmu_get_bonustype(dmu_buf_t
*db_fake
)
234 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
236 dmu_object_type_t type
;
240 type
= dn
->dn_bonustype
;
247 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
252 error
= dnode_hold(os
, object
, FTAG
, &dn
);
253 dbuf_rm_spill(dn
, tx
);
254 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
255 dnode_rm_spill(dn
, tx
);
256 rw_exit(&dn
->dn_struct_rwlock
);
257 dnode_rele(dn
, FTAG
);
262 * returns ENOENT, EIO, or 0.
265 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
271 error
= dnode_hold(os
, object
, FTAG
, &dn
);
275 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
276 if (dn
->dn_bonus
== NULL
) {
277 rw_exit(&dn
->dn_struct_rwlock
);
278 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
279 if (dn
->dn_bonus
== NULL
)
280 dbuf_create_bonus(dn
);
284 /* as long as the bonus buf is held, the dnode will be held */
285 if (refcount_add(&db
->db_holds
, tag
) == 1) {
286 VERIFY(dnode_add_ref(dn
, db
));
287 atomic_inc_32(&dn
->dn_dbufs_count
);
291 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
292 * hold and incrementing the dbuf count to ensure that dnode_move() sees
293 * a dnode hold for every dbuf.
295 rw_exit(&dn
->dn_struct_rwlock
);
297 dnode_rele(dn
, FTAG
);
299 VERIFY(0 == dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
));
306 * returns ENOENT, EIO, or 0.
308 * This interface will allocate a blank spill dbuf when a spill blk
309 * doesn't already exist on the dnode.
311 * if you only want to find an already existing spill db, then
312 * dmu_spill_hold_existing() should be used.
315 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
317 dmu_buf_impl_t
*db
= NULL
;
320 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
321 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
323 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
325 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
326 rw_exit(&dn
->dn_struct_rwlock
);
329 err
= dbuf_read(db
, NULL
, flags
);
338 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
340 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
347 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
348 err
= SET_ERROR(EINVAL
);
350 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
352 if (!dn
->dn_have_spill
) {
353 err
= SET_ERROR(ENOENT
);
355 err
= dmu_spill_hold_by_dnode(dn
,
356 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
359 rw_exit(&dn
->dn_struct_rwlock
);
367 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
369 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
375 err
= dmu_spill_hold_by_dnode(dn
, DB_RF_CANFAIL
, tag
, dbp
);
382 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
383 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
384 * and can induce severe lock contention when writing to several files
385 * whose dnodes are in the same block.
388 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
389 int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
392 uint64_t blkid
, nblks
, i
;
397 ASSERT(length
<= DMU_MAX_ACCESS
);
399 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
;
400 if (flags
& DMU_READ_NO_PREFETCH
|| length
> zfetch_array_rd_sz
)
401 dbuf_flags
|= DB_RF_NOPREFETCH
;
403 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
404 if (dn
->dn_datablkshift
) {
405 int blkshift
= dn
->dn_datablkshift
;
406 nblks
= (P2ROUNDUP(offset
+length
, 1ULL<<blkshift
) -
407 P2ALIGN(offset
, 1ULL<<blkshift
)) >> blkshift
;
409 if (offset
+ length
> dn
->dn_datablksz
) {
410 zfs_panic_recover("zfs: accessing past end of object "
411 "%llx/%llx (size=%u access=%llu+%llu)",
412 (longlong_t
)dn
->dn_objset
->
413 os_dsl_dataset
->ds_object
,
414 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
415 (longlong_t
)offset
, (longlong_t
)length
);
416 rw_exit(&dn
->dn_struct_rwlock
);
417 return (SET_ERROR(EIO
));
421 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_SLEEP
);
423 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
424 blkid
= dbuf_whichblock(dn
, 0, offset
);
425 for (i
= 0; i
< nblks
; i
++) {
426 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+i
, tag
);
428 rw_exit(&dn
->dn_struct_rwlock
);
429 dmu_buf_rele_array(dbp
, nblks
, tag
);
431 return (SET_ERROR(EIO
));
433 /* initiate async i/o */
435 (void) dbuf_read(db
, zio
, dbuf_flags
);
439 rw_exit(&dn
->dn_struct_rwlock
);
441 /* wait for async i/o */
444 dmu_buf_rele_array(dbp
, nblks
, tag
);
448 /* wait for other io to complete */
450 for (i
= 0; i
< nblks
; i
++) {
451 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
452 mutex_enter(&db
->db_mtx
);
453 while (db
->db_state
== DB_READ
||
454 db
->db_state
== DB_FILL
)
455 cv_wait(&db
->db_changed
, &db
->db_mtx
);
456 if (db
->db_state
== DB_UNCACHED
)
457 err
= SET_ERROR(EIO
);
458 mutex_exit(&db
->db_mtx
);
460 dmu_buf_rele_array(dbp
, nblks
, tag
);
472 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
473 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
478 err
= dnode_hold(os
, object
, FTAG
, &dn
);
482 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
483 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
485 dnode_rele(dn
, FTAG
);
491 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
492 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
494 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
500 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
501 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
508 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
511 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
516 for (i
= 0; i
< numbufs
; i
++) {
518 dbuf_rele(dbp
[i
], tag
);
521 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
525 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
526 * indirect blocks prefeteched will be those that point to the blocks containing
527 * the data starting at offset, and continuing to offset + len.
529 * Note that if the indirect blocks above the blocks being prefetched are not in
530 * cache, they will be asychronously read in.
533 dmu_prefetch(objset_t
*os
, uint64_t object
, int64_t level
, uint64_t offset
,
534 uint64_t len
, zio_priority_t pri
)
540 if (zfs_prefetch_disable
)
543 if (len
== 0) { /* they're interested in the bonus buffer */
544 dn
= DMU_META_DNODE(os
);
546 if (object
== 0 || object
>= DN_MAX_OBJECT
)
549 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
550 blkid
= dbuf_whichblock(dn
, level
,
551 object
* sizeof (dnode_phys_t
));
552 dbuf_prefetch(dn
, level
, blkid
, pri
, 0);
553 rw_exit(&dn
->dn_struct_rwlock
);
558 * XXX - Note, if the dnode for the requested object is not
559 * already cached, we will do a *synchronous* read in the
560 * dnode_hold() call. The same is true for any indirects.
562 err
= dnode_hold(os
, object
, FTAG
, &dn
);
566 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
568 * offset + len - 1 is the last byte we want to prefetch for, and offset
569 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
570 * last block we want to prefetch, and dbuf_whichblock(dn, level,
571 * offset) is the first. Then the number we need to prefetch is the
574 if (level
> 0 || dn
->dn_datablkshift
!= 0) {
575 nblks
= dbuf_whichblock(dn
, level
, offset
+ len
- 1) -
576 dbuf_whichblock(dn
, level
, offset
) + 1;
578 nblks
= (offset
< dn
->dn_datablksz
);
584 blkid
= dbuf_whichblock(dn
, level
, offset
);
585 for (i
= 0; i
< nblks
; i
++)
586 dbuf_prefetch(dn
, level
, blkid
+ i
, pri
, 0);
589 rw_exit(&dn
->dn_struct_rwlock
);
591 dnode_rele(dn
, FTAG
);
595 * Get the next "chunk" of file data to free. We traverse the file from
596 * the end so that the file gets shorter over time (if we crashes in the
597 * middle, this will leave us in a better state). We find allocated file
598 * data by simply searching the allocated level 1 indirects.
600 * On input, *start should be the first offset that does not need to be
601 * freed (e.g. "offset + length"). On return, *start will be the first
602 * offset that should be freed.
605 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
607 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
608 /* bytes of data covered by a level-1 indirect block */
610 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
613 ASSERT3U(minimum
, <=, *start
);
615 if (*start
- minimum
<= iblkrange
* maxblks
) {
619 ASSERT(ISP2(iblkrange
));
621 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
625 * dnode_next_offset(BACKWARDS) will find an allocated L1
626 * indirect block at or before the input offset. We must
627 * decrement *start so that it is at the end of the region
631 err
= dnode_next_offset(dn
,
632 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
634 /* if there are no indirect blocks before start, we are done */
638 } else if (err
!= 0) {
642 /* set start to the beginning of this L1 indirect */
643 *start
= P2ALIGN(*start
, iblkrange
);
645 if (*start
< minimum
)
651 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
654 uint64_t object_size
;
658 return (SET_ERROR(EINVAL
));
660 object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
661 if (offset
>= object_size
)
664 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
665 length
= object_size
- offset
;
667 while (length
!= 0) {
668 uint64_t chunk_end
, chunk_begin
;
671 chunk_end
= chunk_begin
= offset
+ length
;
673 /* move chunk_begin backwards to the beginning of this chunk */
674 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
677 ASSERT3U(chunk_begin
, >=, offset
);
678 ASSERT3U(chunk_begin
, <=, chunk_end
);
680 tx
= dmu_tx_create(os
);
681 dmu_tx_hold_free(tx
, dn
->dn_object
,
682 chunk_begin
, chunk_end
- chunk_begin
);
683 err
= dmu_tx_assign(tx
, TXG_WAIT
);
688 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
691 length
-= chunk_end
- chunk_begin
;
697 dmu_free_long_range(objset_t
*os
, uint64_t object
,
698 uint64_t offset
, uint64_t length
)
703 err
= dnode_hold(os
, object
, FTAG
, &dn
);
706 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
709 * It is important to zero out the maxblkid when freeing the entire
710 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
711 * will take the fast path, and (b) dnode_reallocate() can verify
712 * that the entire file has been freed.
714 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
717 dnode_rele(dn
, FTAG
);
722 dmu_free_long_object(objset_t
*os
, uint64_t object
)
727 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
731 tx
= dmu_tx_create(os
);
732 dmu_tx_hold_bonus(tx
, object
);
733 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
734 err
= dmu_tx_assign(tx
, TXG_WAIT
);
736 err
= dmu_object_free(os
, object
, tx
);
746 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
747 uint64_t size
, dmu_tx_t
*tx
)
750 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
753 ASSERT(offset
< UINT64_MAX
);
754 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
755 dnode_free_range(dn
, offset
, size
, tx
);
756 dnode_rele(dn
, FTAG
);
761 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
762 void *buf
, uint32_t flags
)
768 err
= dnode_hold(os
, object
, FTAG
, &dn
);
773 * Deal with odd block sizes, where there can't be data past the first
774 * block. If we ever do the tail block optimization, we will need to
775 * handle that here as well.
777 if (dn
->dn_maxblkid
== 0) {
778 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
779 MIN(size
, dn
->dn_datablksz
- offset
);
780 bzero((char *)buf
+ newsz
, size
- newsz
);
785 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
789 * NB: we could do this block-at-a-time, but it's nice
790 * to be reading in parallel.
792 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
793 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
797 for (i
= 0; i
< numbufs
; i
++) {
800 dmu_buf_t
*db
= dbp
[i
];
804 bufoff
= offset
- db
->db_offset
;
805 tocpy
= MIN(db
->db_size
- bufoff
, size
);
807 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
811 buf
= (char *)buf
+ tocpy
;
813 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
815 dnode_rele(dn
, FTAG
);
820 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
821 const void *buf
, dmu_tx_t
*tx
)
829 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
830 FALSE
, FTAG
, &numbufs
, &dbp
));
832 for (i
= 0; i
< numbufs
; i
++) {
835 dmu_buf_t
*db
= dbp
[i
];
839 bufoff
= offset
- db
->db_offset
;
840 tocpy
= MIN(db
->db_size
- bufoff
, size
);
842 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
844 if (tocpy
== db
->db_size
)
845 dmu_buf_will_fill(db
, tx
);
847 dmu_buf_will_dirty(db
, tx
);
849 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
851 if (tocpy
== db
->db_size
)
852 dmu_buf_fill_done(db
, tx
);
856 buf
= (char *)buf
+ tocpy
;
858 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
862 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
871 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
872 FALSE
, FTAG
, &numbufs
, &dbp
));
874 for (i
= 0; i
< numbufs
; i
++) {
875 dmu_buf_t
*db
= dbp
[i
];
877 dmu_buf_will_not_fill(db
, tx
);
879 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
883 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
884 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
885 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
889 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
890 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
891 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
894 dmu_buf_write_embedded(db
,
895 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
896 uncompressed_size
, compressed_size
, byteorder
, tx
);
898 dmu_buf_rele(db
, FTAG
);
902 * DMU support for xuio
904 kstat_t
*xuio_ksp
= NULL
;
906 typedef struct xuio_stats
{
907 /* loaned yet not returned arc_buf */
908 kstat_named_t xuiostat_onloan_rbuf
;
909 kstat_named_t xuiostat_onloan_wbuf
;
910 /* whether a copy is made when loaning out a read buffer */
911 kstat_named_t xuiostat_rbuf_copied
;
912 kstat_named_t xuiostat_rbuf_nocopy
;
913 /* whether a copy is made when assigning a write buffer */
914 kstat_named_t xuiostat_wbuf_copied
;
915 kstat_named_t xuiostat_wbuf_nocopy
;
918 static xuio_stats_t xuio_stats
= {
919 { "onloan_read_buf", KSTAT_DATA_UINT64
},
920 { "onloan_write_buf", KSTAT_DATA_UINT64
},
921 { "read_buf_copied", KSTAT_DATA_UINT64
},
922 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
923 { "write_buf_copied", KSTAT_DATA_UINT64
},
924 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
927 #define XUIOSTAT_INCR(stat, val) \
928 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
929 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
932 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
935 uio_t
*uio
= &xuio
->xu_uio
;
937 uio
->uio_iovcnt
= nblk
;
938 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
940 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
942 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
943 priv
->iovp
= (iovec_t
*)uio
->uio_iov
;
944 XUIO_XUZC_PRIV(xuio
) = priv
;
946 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
947 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
949 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
955 dmu_xuio_fini(xuio_t
*xuio
)
957 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
958 int nblk
= priv
->cnt
;
960 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
961 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
962 kmem_free(priv
, sizeof (dmu_xuio_t
));
964 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
965 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
967 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
971 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
972 * and increase priv->next by 1.
975 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
978 uio_t
*uio
= &xuio
->xu_uio
;
979 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
980 int i
= priv
->next
++;
982 ASSERT(i
< priv
->cnt
);
983 ASSERT(off
+ n
<= arc_buf_size(abuf
));
984 iov
= (iovec_t
*)uio
->uio_iov
+ i
;
985 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
987 priv
->bufs
[i
] = abuf
;
992 dmu_xuio_cnt(xuio_t
*xuio
)
994 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
999 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
1001 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1003 ASSERT(i
< priv
->cnt
);
1004 return (priv
->bufs
[i
]);
1008 dmu_xuio_clear(xuio_t
*xuio
, int i
)
1010 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1012 ASSERT(i
< priv
->cnt
);
1013 priv
->bufs
[i
] = NULL
;
1017 xuio_stat_init(void)
1019 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1020 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1021 KSTAT_FLAG_VIRTUAL
);
1022 if (xuio_ksp
!= NULL
) {
1023 xuio_ksp
->ks_data
= &xuio_stats
;
1024 kstat_install(xuio_ksp
);
1029 xuio_stat_fini(void)
1031 if (xuio_ksp
!= NULL
) {
1032 kstat_delete(xuio_ksp
);
1038 xuio_stat_wbuf_copied()
1040 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1044 xuio_stat_wbuf_nocopy()
1046 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1051 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1054 int numbufs
, i
, err
;
1055 xuio_t
*xuio
= NULL
;
1058 * NB: we could do this block-at-a-time, but it's nice
1059 * to be reading in parallel.
1061 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1062 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1066 for (i
= 0; i
< numbufs
; i
++) {
1069 dmu_buf_t
*db
= dbp
[i
];
1073 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1074 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1077 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1078 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1079 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1080 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1082 uio
->uio_resid
-= tocpy
;
1083 uio
->uio_loffset
+= tocpy
;
1086 if (abuf
== dbuf_abuf
)
1087 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1089 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1091 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1099 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1105 * Read 'size' bytes into the uio buffer.
1106 * From object zdb->db_object.
1107 * Starting at offset uio->uio_loffset.
1109 * If the caller already has a dbuf in the target object
1110 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1111 * because we don't have to find the dnode_t for the object.
1114 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1116 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1125 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1132 * Read 'size' bytes into the uio buffer.
1133 * From the specified object
1134 * Starting at offset uio->uio_loffset.
1137 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1145 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1149 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1151 dnode_rele(dn
, FTAG
);
1157 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1164 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1165 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1169 for (i
= 0; i
< numbufs
; i
++) {
1172 dmu_buf_t
*db
= dbp
[i
];
1176 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1177 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1179 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1181 if (tocpy
== db
->db_size
)
1182 dmu_buf_will_fill(db
, tx
);
1184 dmu_buf_will_dirty(db
, tx
);
1187 * XXX uiomove could block forever (eg.nfs-backed
1188 * pages). There needs to be a uiolockdown() function
1189 * to lock the pages in memory, so that uiomove won't
1192 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1195 if (tocpy
== db
->db_size
)
1196 dmu_buf_fill_done(db
, tx
);
1204 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1209 * Write 'size' bytes from the uio buffer.
1210 * To object zdb->db_object.
1211 * Starting at offset uio->uio_loffset.
1213 * If the caller already has a dbuf in the target object
1214 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1215 * because we don't have to find the dnode_t for the object.
1218 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1221 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1230 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1237 * Write 'size' bytes from the uio buffer.
1238 * To the specified object.
1239 * Starting at offset uio->uio_loffset.
1242 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1251 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1255 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1257 dnode_rele(dn
, FTAG
);
1261 #endif /* _KERNEL */
1264 * Allocate a loaned anonymous arc buffer.
1267 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1269 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1271 return (arc_loan_buf(db
->db_objset
->os_spa
, size
));
1275 * Free a loaned arc buffer.
1278 dmu_return_arcbuf(arc_buf_t
*buf
)
1280 arc_return_buf(buf
, FTAG
);
1281 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1285 * When possible directly assign passed loaned arc buffer to a dbuf.
1286 * If this is not possible copy the contents of passed arc buf via
1290 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1293 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1296 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1299 DB_DNODE_ENTER(dbuf
);
1300 dn
= DB_DNODE(dbuf
);
1301 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1302 blkid
= dbuf_whichblock(dn
, 0, offset
);
1303 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1304 rw_exit(&dn
->dn_struct_rwlock
);
1305 DB_DNODE_EXIT(dbuf
);
1308 * We can only assign if the offset is aligned, the arc buf is the
1309 * same size as the dbuf, and the dbuf is not metadata. It
1310 * can't be metadata because the loaned arc buf comes from the
1311 * user-data kmem area.
1313 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
&&
1314 DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
) {
1315 dbuf_assign_arcbuf(db
, buf
, tx
);
1316 dbuf_rele(db
, FTAG
);
1321 DB_DNODE_ENTER(dbuf
);
1322 dn
= DB_DNODE(dbuf
);
1324 object
= dn
->dn_object
;
1325 DB_DNODE_EXIT(dbuf
);
1327 dbuf_rele(db
, FTAG
);
1328 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1329 dmu_return_arcbuf(buf
);
1330 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1335 dbuf_dirty_record_t
*dsa_dr
;
1336 dmu_sync_cb_t
*dsa_done
;
1343 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1345 dmu_sync_arg_t
*dsa
= varg
;
1346 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1347 blkptr_t
*bp
= zio
->io_bp
;
1349 if (zio
->io_error
== 0) {
1350 if (BP_IS_HOLE(bp
)) {
1352 * A block of zeros may compress to a hole, but the
1353 * block size still needs to be known for replay.
1355 BP_SET_LSIZE(bp
, db
->db_size
);
1356 } else if (!BP_IS_EMBEDDED(bp
)) {
1357 ASSERT(BP_GET_LEVEL(bp
) == 0);
1364 dmu_sync_late_arrival_ready(zio_t
*zio
)
1366 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1371 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1373 dmu_sync_arg_t
*dsa
= varg
;
1374 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1375 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1377 mutex_enter(&db
->db_mtx
);
1378 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1379 if (zio
->io_error
== 0) {
1380 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1381 if (dr
->dt
.dl
.dr_nopwrite
) {
1382 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1383 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1384 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1386 ASSERT(BP_EQUAL(bp
, bp_orig
));
1387 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1388 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1390 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1391 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1392 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1395 * Old style holes are filled with all zeros, whereas
1396 * new-style holes maintain their lsize, type, level,
1397 * and birth time (see zio_write_compress). While we
1398 * need to reset the BP_SET_LSIZE() call that happened
1399 * in dmu_sync_ready for old style holes, we do *not*
1400 * want to wipe out the information contained in new
1401 * style holes. Thus, only zero out the block pointer if
1402 * it's an old style hole.
1404 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1405 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1406 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1408 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1410 cv_broadcast(&db
->db_changed
);
1411 mutex_exit(&db
->db_mtx
);
1413 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1415 kmem_free(dsa
, sizeof (*dsa
));
1419 dmu_sync_late_arrival_done(zio_t
*zio
)
1421 blkptr_t
*bp
= zio
->io_bp
;
1422 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1423 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1425 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1427 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1428 * then there is nothing to do here. Otherwise, free the
1429 * newly allocated block in this txg.
1431 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1432 ASSERT(BP_EQUAL(bp
, bp_orig
));
1434 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1435 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1436 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1437 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1441 dmu_tx_commit(dsa
->dsa_tx
);
1443 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1445 kmem_free(dsa
, sizeof (*dsa
));
1449 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1450 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1452 dmu_sync_arg_t
*dsa
;
1455 tx
= dmu_tx_create(os
);
1456 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1457 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1459 /* Make zl_get_data do txg_waited_synced() */
1460 return (SET_ERROR(EIO
));
1463 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1465 dsa
->dsa_done
= done
;
1469 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1470 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1471 dmu_sync_late_arrival_ready
, NULL
, dmu_sync_late_arrival_done
, dsa
,
1472 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
|ZIO_FLAG_FASTWRITE
, zb
));
1478 * Intent log support: sync the block associated with db to disk.
1479 * N.B. and XXX: the caller is responsible for making sure that the
1480 * data isn't changing while dmu_sync() is writing it.
1484 * EEXIST: this txg has already been synced, so there's nothing to do.
1485 * The caller should not log the write.
1487 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1488 * The caller should not log the write.
1490 * EALREADY: this block is already in the process of being synced.
1491 * The caller should track its progress (somehow).
1493 * EIO: could not do the I/O.
1494 * The caller should do a txg_wait_synced().
1496 * 0: the I/O has been initiated.
1497 * The caller should log this blkptr in the done callback.
1498 * It is possible that the I/O will fail, in which case
1499 * the error will be reported to the done callback and
1500 * propagated to pio from zio_done().
1503 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1505 blkptr_t
*bp
= zgd
->zgd_bp
;
1506 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1507 objset_t
*os
= db
->db_objset
;
1508 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1509 dbuf_dirty_record_t
*dr
;
1510 dmu_sync_arg_t
*dsa
;
1511 zbookmark_phys_t zb
;
1515 ASSERT(pio
!= NULL
);
1518 SET_BOOKMARK(&zb
, ds
->ds_object
,
1519 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1523 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1527 * If we're frozen (running ziltest), we always need to generate a bp.
1529 if (txg
> spa_freeze_txg(os
->os_spa
))
1530 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1533 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1534 * and us. If we determine that this txg is not yet syncing,
1535 * but it begins to sync a moment later, that's OK because the
1536 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1538 mutex_enter(&db
->db_mtx
);
1540 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1542 * This txg has already synced. There's nothing to do.
1544 mutex_exit(&db
->db_mtx
);
1545 return (SET_ERROR(EEXIST
));
1548 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1550 * This txg is currently syncing, so we can't mess with
1551 * the dirty record anymore; just write a new log block.
1553 mutex_exit(&db
->db_mtx
);
1554 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1557 dr
= db
->db_last_dirty
;
1558 while (dr
&& dr
->dr_txg
!= txg
)
1563 * There's no dr for this dbuf, so it must have been freed.
1564 * There's no need to log writes to freed blocks, so we're done.
1566 mutex_exit(&db
->db_mtx
);
1567 return (SET_ERROR(ENOENT
));
1570 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1573 * Assume the on-disk data is X, the current syncing data (in
1574 * txg - 1) is Y, and the current in-memory data is Z (currently
1577 * We usually want to perform a nopwrite if X and Z are the
1578 * same. However, if Y is different (i.e. the BP is going to
1579 * change before this write takes effect), then a nopwrite will
1580 * be incorrect - we would override with X, which could have
1581 * been freed when Y was written.
1583 * (Note that this is not a concern when we are nop-writing from
1584 * syncing context, because X and Y must be identical, because
1585 * all previous txgs have been synced.)
1587 * Therefore, we disable nopwrite if the current BP could change
1588 * before this TXG. There are two ways it could change: by
1589 * being dirty (dr_next is non-NULL), or by being freed
1590 * (dnode_block_freed()). This behavior is verified by
1591 * zio_done(), which VERIFYs that the override BP is identical
1592 * to the on-disk BP.
1596 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
1597 zp
.zp_nopwrite
= B_FALSE
;
1600 ASSERT(dr
->dr_txg
== txg
);
1601 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1602 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1604 * We have already issued a sync write for this buffer,
1605 * or this buffer has already been synced. It could not
1606 * have been dirtied since, or we would have cleared the state.
1608 mutex_exit(&db
->db_mtx
);
1609 return (SET_ERROR(EALREADY
));
1612 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1613 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1614 mutex_exit(&db
->db_mtx
);
1616 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1618 dsa
->dsa_done
= done
;
1622 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1623 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1624 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1625 NULL
, dmu_sync_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1626 ZIO_FLAG_CANFAIL
, &zb
));
1632 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1638 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1641 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1642 dnode_rele(dn
, FTAG
);
1647 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1653 * Send streams include each object's checksum function. This
1654 * check ensures that the receiving system can understand the
1655 * checksum function transmitted.
1657 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1659 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1660 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1661 dn
->dn_checksum
= checksum
;
1662 dnode_setdirty(dn
, tx
);
1663 dnode_rele(dn
, FTAG
);
1667 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1673 * Send streams include each object's compression function. This
1674 * check ensures that the receiving system can understand the
1675 * compression function transmitted.
1677 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1679 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1680 dn
->dn_compress
= compress
;
1681 dnode_setdirty(dn
, tx
);
1682 dnode_rele(dn
, FTAG
);
1685 int zfs_mdcomp_disable
= 0;
1688 * When the "redundant_metadata" property is set to "most", only indirect
1689 * blocks of this level and higher will have an additional ditto block.
1691 int zfs_redundant_metadata_most_ditto_level
= 2;
1694 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1696 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1697 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1699 enum zio_checksum checksum
= os
->os_checksum
;
1700 enum zio_compress compress
= os
->os_compress
;
1701 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1702 boolean_t dedup
= B_FALSE
;
1703 boolean_t nopwrite
= B_FALSE
;
1704 boolean_t dedup_verify
= os
->os_dedup_verify
;
1705 int copies
= os
->os_copies
;
1708 * We maintain different write policies for each of the following
1711 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1712 * 3. all other level 0 blocks
1715 if (zfs_mdcomp_disable
) {
1716 compress
= ZIO_COMPRESS_EMPTY
;
1719 * XXX -- we should design a compression algorithm
1720 * that specializes in arrays of bps.
1722 compress
= zio_compress_select(os
->os_spa
,
1723 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
1727 * Metadata always gets checksummed. If the data
1728 * checksum is multi-bit correctable, and it's not a
1729 * ZBT-style checksum, then it's suitable for metadata
1730 * as well. Otherwise, the metadata checksum defaults
1733 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1734 zio_checksum_table
[checksum
].ci_eck
)
1735 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1737 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
1738 (os
->os_redundant_metadata
==
1739 ZFS_REDUNDANT_METADATA_MOST
&&
1740 (level
>= zfs_redundant_metadata_most_ditto_level
||
1741 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
1743 } else if (wp
& WP_NOFILL
) {
1747 * If we're writing preallocated blocks, we aren't actually
1748 * writing them so don't set any policy properties. These
1749 * blocks are currently only used by an external subsystem
1750 * outside of zfs (i.e. dump) and not written by the zio
1753 compress
= ZIO_COMPRESS_OFF
;
1754 checksum
= ZIO_CHECKSUM_OFF
;
1756 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
1759 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1760 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1764 * Determine dedup setting. If we are in dmu_sync(),
1765 * we won't actually dedup now because that's all
1766 * done in syncing context; but we do want to use the
1767 * dedup checkum. If the checksum is not strong
1768 * enough to ensure unique signatures, force
1771 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1772 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1773 if (!zio_checksum_table
[checksum
].ci_dedup
)
1774 dedup_verify
= B_TRUE
;
1778 * Enable nopwrite if we have a cryptographically secure
1779 * checksum that has no known collisions (i.e. SHA-256)
1780 * and compression is enabled. We don't enable nopwrite if
1781 * dedup is enabled as the two features are mutually exclusive.
1783 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1784 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1787 zp
->zp_checksum
= checksum
;
1788 zp
->zp_compress
= compress
;
1789 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1790 zp
->zp_level
= level
;
1791 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
1792 zp
->zp_dedup
= dedup
;
1793 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1794 zp
->zp_nopwrite
= nopwrite
;
1798 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1803 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1807 * Sync any current changes before
1808 * we go trundling through the block pointers.
1810 for (i
= 0; i
< TXG_SIZE
; i
++) {
1811 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1814 if (i
!= TXG_SIZE
) {
1815 dnode_rele(dn
, FTAG
);
1816 txg_wait_synced(dmu_objset_pool(os
), 0);
1817 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1822 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1823 dnode_rele(dn
, FTAG
);
1829 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1831 dnode_phys_t
*dnp
= dn
->dn_phys
;
1834 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1835 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1836 1ULL << dn
->dn_indblkshift
: 0;
1837 doi
->doi_type
= dn
->dn_type
;
1838 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1839 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1840 doi
->doi_indirection
= dn
->dn_nlevels
;
1841 doi
->doi_checksum
= dn
->dn_checksum
;
1842 doi
->doi_compress
= dn
->dn_compress
;
1843 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
1844 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1845 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1846 doi
->doi_fill_count
= 0;
1847 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1848 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
1852 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1854 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1855 mutex_enter(&dn
->dn_mtx
);
1857 __dmu_object_info_from_dnode(dn
, doi
);
1859 mutex_exit(&dn
->dn_mtx
);
1860 rw_exit(&dn
->dn_struct_rwlock
);
1864 * Get information on a DMU object.
1865 * If doi is NULL, just indicates whether the object exists.
1868 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1871 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1877 dmu_object_info_from_dnode(dn
, doi
);
1879 dnode_rele(dn
, FTAG
);
1884 * As above, but faster; can be used when you have a held dbuf in hand.
1887 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1889 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1892 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1897 * Faster still when you only care about the size.
1898 * This is specifically optimized for zfs_getattr().
1901 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1902 u_longlong_t
*nblk512
)
1904 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1910 *blksize
= dn
->dn_datablksz
;
1911 /* add 1 for dnode space */
1912 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1913 SPA_MINBLOCKSHIFT
) + 1;
1918 byteswap_uint64_array(void *vbuf
, size_t size
)
1920 uint64_t *buf
= vbuf
;
1921 size_t count
= size
>> 3;
1924 ASSERT((size
& 7) == 0);
1926 for (i
= 0; i
< count
; i
++)
1927 buf
[i
] = BSWAP_64(buf
[i
]);
1931 byteswap_uint32_array(void *vbuf
, size_t size
)
1933 uint32_t *buf
= vbuf
;
1934 size_t count
= size
>> 2;
1937 ASSERT((size
& 3) == 0);
1939 for (i
= 0; i
< count
; i
++)
1940 buf
[i
] = BSWAP_32(buf
[i
]);
1944 byteswap_uint16_array(void *vbuf
, size_t size
)
1946 uint16_t *buf
= vbuf
;
1947 size_t count
= size
>> 1;
1950 ASSERT((size
& 1) == 0);
1952 for (i
= 0; i
< count
; i
++)
1953 buf
[i
] = BSWAP_16(buf
[i
]);
1958 byteswap_uint8_array(void *vbuf
, size_t size
)
1980 arc_fini(); /* arc depends on l2arc, so arc must go first */
1992 #if defined(_KERNEL) && defined(HAVE_SPL)
1993 EXPORT_SYMBOL(dmu_bonus_hold
);
1994 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
1995 EXPORT_SYMBOL(dmu_buf_rele_array
);
1996 EXPORT_SYMBOL(dmu_prefetch
);
1997 EXPORT_SYMBOL(dmu_free_range
);
1998 EXPORT_SYMBOL(dmu_free_long_range
);
1999 EXPORT_SYMBOL(dmu_free_long_object
);
2000 EXPORT_SYMBOL(dmu_read
);
2001 EXPORT_SYMBOL(dmu_write
);
2002 EXPORT_SYMBOL(dmu_prealloc
);
2003 EXPORT_SYMBOL(dmu_object_info
);
2004 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2005 EXPORT_SYMBOL(dmu_object_info_from_db
);
2006 EXPORT_SYMBOL(dmu_object_size_from_db
);
2007 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2008 EXPORT_SYMBOL(dmu_object_set_checksum
);
2009 EXPORT_SYMBOL(dmu_object_set_compress
);
2010 EXPORT_SYMBOL(dmu_write_policy
);
2011 EXPORT_SYMBOL(dmu_sync
);
2012 EXPORT_SYMBOL(dmu_request_arcbuf
);
2013 EXPORT_SYMBOL(dmu_return_arcbuf
);
2014 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2015 EXPORT_SYMBOL(dmu_buf_hold
);
2016 EXPORT_SYMBOL(dmu_ot
);
2018 module_param(zfs_mdcomp_disable
, int, 0644);
2019 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2021 module_param(zfs_nopwrite_enabled
, int, 0644);
2022 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");