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
, 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
, 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.
527 * Note: The assumption is that we *know* these blocks will be needed
528 * almost immediately. Therefore, the prefetch i/os will be issued at
529 * ZIO_PRIORITY_SYNC_READ
531 * Note: indirect blocks and other metadata will be read synchronously,
532 * causing this function to block if they are not already cached.
535 dmu_prefetch(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t len
)
541 if (zfs_prefetch_disable
)
544 if (len
== 0) { /* they're interested in the bonus buffer */
545 dn
= DMU_META_DNODE(os
);
547 if (object
== 0 || object
>= DN_MAX_OBJECT
)
550 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
551 blkid
= dbuf_whichblock(dn
, object
* sizeof (dnode_phys_t
));
552 dbuf_prefetch(dn
, blkid
, ZIO_PRIORITY_SYNC_READ
);
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
);
567 if (dn
->dn_datablkshift
) {
568 int blkshift
= dn
->dn_datablkshift
;
569 nblks
= (P2ROUNDUP(offset
+ len
, 1 << blkshift
) -
570 P2ALIGN(offset
, 1 << blkshift
)) >> blkshift
;
572 nblks
= (offset
< dn
->dn_datablksz
);
578 blkid
= dbuf_whichblock(dn
, offset
);
579 for (i
= 0; i
< nblks
; i
++)
580 dbuf_prefetch(dn
, blkid
+ i
, ZIO_PRIORITY_SYNC_READ
);
583 rw_exit(&dn
->dn_struct_rwlock
);
585 dnode_rele(dn
, FTAG
);
589 * Get the next "chunk" of file data to free. We traverse the file from
590 * the end so that the file gets shorter over time (if we crashes in the
591 * middle, this will leave us in a better state). We find allocated file
592 * data by simply searching the allocated level 1 indirects.
594 * On input, *start should be the first offset that does not need to be
595 * freed (e.g. "offset + length"). On return, *start will be the first
596 * offset that should be freed.
599 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
601 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
602 /* bytes of data covered by a level-1 indirect block */
604 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
607 ASSERT3U(minimum
, <=, *start
);
609 if (*start
- minimum
<= iblkrange
* maxblks
) {
613 ASSERT(ISP2(iblkrange
));
615 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
619 * dnode_next_offset(BACKWARDS) will find an allocated L1
620 * indirect block at or before the input offset. We must
621 * decrement *start so that it is at the end of the region
625 err
= dnode_next_offset(dn
,
626 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
628 /* if there are no indirect blocks before start, we are done */
632 } else if (err
!= 0) {
636 /* set start to the beginning of this L1 indirect */
637 *start
= P2ALIGN(*start
, iblkrange
);
639 if (*start
< minimum
)
645 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
648 uint64_t object_size
;
652 return (SET_ERROR(EINVAL
));
654 object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
655 if (offset
>= object_size
)
658 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
659 length
= object_size
- offset
;
661 while (length
!= 0) {
662 uint64_t chunk_end
, chunk_begin
;
665 chunk_end
= chunk_begin
= offset
+ length
;
667 /* move chunk_begin backwards to the beginning of this chunk */
668 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
671 ASSERT3U(chunk_begin
, >=, offset
);
672 ASSERT3U(chunk_begin
, <=, chunk_end
);
674 tx
= dmu_tx_create(os
);
675 dmu_tx_hold_free(tx
, dn
->dn_object
,
676 chunk_begin
, chunk_end
- chunk_begin
);
677 err
= dmu_tx_assign(tx
, TXG_WAIT
);
682 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
685 length
-= chunk_end
- chunk_begin
;
691 dmu_free_long_range(objset_t
*os
, uint64_t object
,
692 uint64_t offset
, uint64_t length
)
697 err
= dnode_hold(os
, object
, FTAG
, &dn
);
700 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
703 * It is important to zero out the maxblkid when freeing the entire
704 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
705 * will take the fast path, and (b) dnode_reallocate() can verify
706 * that the entire file has been freed.
708 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
711 dnode_rele(dn
, FTAG
);
716 dmu_free_long_object(objset_t
*os
, uint64_t object
)
721 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
725 tx
= dmu_tx_create(os
);
726 dmu_tx_hold_bonus(tx
, object
);
727 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
728 err
= dmu_tx_assign(tx
, TXG_WAIT
);
730 err
= dmu_object_free(os
, object
, tx
);
740 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
741 uint64_t size
, dmu_tx_t
*tx
)
744 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
747 ASSERT(offset
< UINT64_MAX
);
748 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
749 dnode_free_range(dn
, offset
, size
, tx
);
750 dnode_rele(dn
, FTAG
);
755 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
756 void *buf
, uint32_t flags
)
762 err
= dnode_hold(os
, object
, FTAG
, &dn
);
767 * Deal with odd block sizes, where there can't be data past the first
768 * block. If we ever do the tail block optimization, we will need to
769 * handle that here as well.
771 if (dn
->dn_maxblkid
== 0) {
772 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
773 MIN(size
, dn
->dn_datablksz
- offset
);
774 bzero((char *)buf
+ newsz
, size
- newsz
);
779 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
783 * NB: we could do this block-at-a-time, but it's nice
784 * to be reading in parallel.
786 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
787 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
791 for (i
= 0; i
< numbufs
; i
++) {
794 dmu_buf_t
*db
= dbp
[i
];
798 bufoff
= offset
- db
->db_offset
;
799 tocpy
= MIN(db
->db_size
- bufoff
, size
);
801 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
805 buf
= (char *)buf
+ tocpy
;
807 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
809 dnode_rele(dn
, FTAG
);
814 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
815 const void *buf
, dmu_tx_t
*tx
)
823 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
824 FALSE
, FTAG
, &numbufs
, &dbp
));
826 for (i
= 0; i
< numbufs
; i
++) {
829 dmu_buf_t
*db
= dbp
[i
];
833 bufoff
= offset
- db
->db_offset
;
834 tocpy
= MIN(db
->db_size
- bufoff
, size
);
836 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
838 if (tocpy
== db
->db_size
)
839 dmu_buf_will_fill(db
, tx
);
841 dmu_buf_will_dirty(db
, tx
);
843 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
845 if (tocpy
== db
->db_size
)
846 dmu_buf_fill_done(db
, tx
);
850 buf
= (char *)buf
+ tocpy
;
852 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
856 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
865 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
866 FALSE
, FTAG
, &numbufs
, &dbp
));
868 for (i
= 0; i
< numbufs
; i
++) {
869 dmu_buf_t
*db
= dbp
[i
];
871 dmu_buf_will_not_fill(db
, tx
);
873 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
877 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
878 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
879 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
883 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
884 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
885 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
888 dmu_buf_write_embedded(db
,
889 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
890 uncompressed_size
, compressed_size
, byteorder
, tx
);
892 dmu_buf_rele(db
, FTAG
);
896 * DMU support for xuio
898 kstat_t
*xuio_ksp
= NULL
;
900 typedef struct xuio_stats
{
901 /* loaned yet not returned arc_buf */
902 kstat_named_t xuiostat_onloan_rbuf
;
903 kstat_named_t xuiostat_onloan_wbuf
;
904 /* whether a copy is made when loaning out a read buffer */
905 kstat_named_t xuiostat_rbuf_copied
;
906 kstat_named_t xuiostat_rbuf_nocopy
;
907 /* whether a copy is made when assigning a write buffer */
908 kstat_named_t xuiostat_wbuf_copied
;
909 kstat_named_t xuiostat_wbuf_nocopy
;
912 static xuio_stats_t xuio_stats
= {
913 { "onloan_read_buf", KSTAT_DATA_UINT64
},
914 { "onloan_write_buf", KSTAT_DATA_UINT64
},
915 { "read_buf_copied", KSTAT_DATA_UINT64
},
916 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
917 { "write_buf_copied", KSTAT_DATA_UINT64
},
918 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
921 #define XUIOSTAT_INCR(stat, val) \
922 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
923 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
926 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
929 uio_t
*uio
= &xuio
->xu_uio
;
931 uio
->uio_iovcnt
= nblk
;
932 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
934 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
936 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
937 priv
->iovp
= (iovec_t
*)uio
->uio_iov
;
938 XUIO_XUZC_PRIV(xuio
) = priv
;
940 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
941 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
943 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
949 dmu_xuio_fini(xuio_t
*xuio
)
951 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
952 int nblk
= priv
->cnt
;
954 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
955 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
956 kmem_free(priv
, sizeof (dmu_xuio_t
));
958 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
959 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
961 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
965 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
966 * and increase priv->next by 1.
969 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
972 uio_t
*uio
= &xuio
->xu_uio
;
973 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
974 int i
= priv
->next
++;
976 ASSERT(i
< priv
->cnt
);
977 ASSERT(off
+ n
<= arc_buf_size(abuf
));
978 iov
= (iovec_t
*)uio
->uio_iov
+ i
;
979 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
981 priv
->bufs
[i
] = abuf
;
986 dmu_xuio_cnt(xuio_t
*xuio
)
988 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
993 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
995 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
997 ASSERT(i
< priv
->cnt
);
998 return (priv
->bufs
[i
]);
1002 dmu_xuio_clear(xuio_t
*xuio
, int i
)
1004 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1006 ASSERT(i
< priv
->cnt
);
1007 priv
->bufs
[i
] = NULL
;
1011 xuio_stat_init(void)
1013 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1014 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1015 KSTAT_FLAG_VIRTUAL
);
1016 if (xuio_ksp
!= NULL
) {
1017 xuio_ksp
->ks_data
= &xuio_stats
;
1018 kstat_install(xuio_ksp
);
1023 xuio_stat_fini(void)
1025 if (xuio_ksp
!= NULL
) {
1026 kstat_delete(xuio_ksp
);
1032 xuio_stat_wbuf_copied()
1034 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1038 xuio_stat_wbuf_nocopy()
1040 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1046 * Copy up to size bytes between arg_buf and req based on the data direction
1047 * described by the req. If an entire req's data cannot be transfered in one
1048 * pass, you should pass in @req_offset to indicate where to continue. The
1049 * return value is the number of bytes successfully copied to arg_buf.
1052 dmu_req_copy(void *arg_buf
, int size
, struct request
*req
, size_t req_offset
)
1054 struct bio_vec bv
, *bvp
;
1055 struct req_iterator iter
;
1057 int tocpy
, bv_len
, bv_offset
;
1060 rq_for_each_segment4(bv
, bvp
, req
, iter
) {
1062 * Fully consumed the passed arg_buf. We use goto here because
1063 * rq_for_each_segment is a double loop
1065 ASSERT3S(offset
, <=, size
);
1069 /* Skip already copied bv */
1070 if (req_offset
>= bv
.bv_len
) {
1071 req_offset
-= bv
.bv_len
;
1075 bv_len
= bv
.bv_len
- req_offset
;
1076 bv_offset
= bv
.bv_offset
+ req_offset
;
1079 tocpy
= MIN(bv_len
, size
- offset
);
1080 ASSERT3S(tocpy
, >=, 0);
1082 bv_buf
= page_address(bv
.bv_page
) + bv_offset
;
1083 ASSERT3P(bv_buf
, !=, NULL
);
1085 if (rq_data_dir(req
) == WRITE
)
1086 memcpy(arg_buf
+ offset
, bv_buf
, tocpy
);
1088 memcpy(bv_buf
, arg_buf
+ offset
, tocpy
);
1097 dmu_read_req(objset_t
*os
, uint64_t object
, struct request
*req
)
1099 uint64_t size
= blk_rq_bytes(req
);
1100 uint64_t offset
= blk_rq_pos(req
) << 9;
1102 int numbufs
, i
, err
;
1106 * NB: we could do this block-at-a-time, but it's nice
1107 * to be reading in parallel.
1109 err
= dmu_buf_hold_array(os
, object
, offset
, size
, TRUE
, FTAG
,
1115 for (i
= 0; i
< numbufs
; i
++) {
1119 dmu_buf_t
*db
= dbp
[i
];
1121 bufoff
= offset
- db
->db_offset
;
1122 ASSERT3S(bufoff
, >=, 0);
1124 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1128 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1139 req_offset
+= didcpy
;
1142 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1148 dmu_write_req(objset_t
*os
, uint64_t object
, struct request
*req
, dmu_tx_t
*tx
)
1150 uint64_t size
= blk_rq_bytes(req
);
1151 uint64_t offset
= blk_rq_pos(req
) << 9;
1153 int numbufs
, i
, err
;
1159 err
= dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1165 for (i
= 0; i
< numbufs
; i
++) {
1169 dmu_buf_t
*db
= dbp
[i
];
1171 bufoff
= offset
- db
->db_offset
;
1172 ASSERT3S(bufoff
, >=, 0);
1174 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1178 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1180 if (tocpy
== db
->db_size
)
1181 dmu_buf_will_fill(db
, tx
);
1183 dmu_buf_will_dirty(db
, tx
);
1185 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1188 if (tocpy
== db
->db_size
)
1189 dmu_buf_fill_done(db
, tx
);
1199 req_offset
+= didcpy
;
1203 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1208 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1211 int numbufs
, i
, err
;
1212 xuio_t
*xuio
= NULL
;
1215 * NB: we could do this block-at-a-time, but it's nice
1216 * to be reading in parallel.
1218 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1219 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1223 for (i
= 0; i
< numbufs
; i
++) {
1226 dmu_buf_t
*db
= dbp
[i
];
1230 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1231 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1234 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1235 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1236 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1237 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1239 uio
->uio_resid
-= tocpy
;
1240 uio
->uio_loffset
+= tocpy
;
1243 if (abuf
== dbuf_abuf
)
1244 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1246 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1248 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1256 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1262 * Read 'size' bytes into the uio buffer.
1263 * From object zdb->db_object.
1264 * Starting at offset uio->uio_loffset.
1266 * If the caller already has a dbuf in the target object
1267 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1268 * because we don't have to find the dnode_t for the object.
1271 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1273 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1282 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1289 * Read 'size' bytes into the uio buffer.
1290 * From the specified object
1291 * Starting at offset uio->uio_loffset.
1294 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1302 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1306 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1308 dnode_rele(dn
, FTAG
);
1314 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1321 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1322 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1326 for (i
= 0; i
< numbufs
; i
++) {
1329 dmu_buf_t
*db
= dbp
[i
];
1333 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1334 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1336 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1338 if (tocpy
== db
->db_size
)
1339 dmu_buf_will_fill(db
, tx
);
1341 dmu_buf_will_dirty(db
, tx
);
1344 * XXX uiomove could block forever (eg.nfs-backed
1345 * pages). There needs to be a uiolockdown() function
1346 * to lock the pages in memory, so that uiomove won't
1349 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1352 if (tocpy
== db
->db_size
)
1353 dmu_buf_fill_done(db
, tx
);
1361 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1366 * Write 'size' bytes from the uio buffer.
1367 * To object zdb->db_object.
1368 * Starting at offset uio->uio_loffset.
1370 * If the caller already has a dbuf in the target object
1371 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1372 * because we don't have to find the dnode_t for the object.
1375 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1378 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1387 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1394 * Write 'size' bytes from the uio buffer.
1395 * To the specified object.
1396 * Starting at offset uio->uio_loffset.
1399 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1408 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1412 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1414 dnode_rele(dn
, FTAG
);
1418 #endif /* _KERNEL */
1421 * Allocate a loaned anonymous arc buffer.
1424 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1426 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1428 return (arc_loan_buf(db
->db_objset
->os_spa
, size
));
1432 * Free a loaned arc buffer.
1435 dmu_return_arcbuf(arc_buf_t
*buf
)
1437 arc_return_buf(buf
, FTAG
);
1438 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1442 * When possible directly assign passed loaned arc buffer to a dbuf.
1443 * If this is not possible copy the contents of passed arc buf via
1447 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1450 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1453 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1456 DB_DNODE_ENTER(dbuf
);
1457 dn
= DB_DNODE(dbuf
);
1458 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1459 blkid
= dbuf_whichblock(dn
, offset
);
1460 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1461 rw_exit(&dn
->dn_struct_rwlock
);
1462 DB_DNODE_EXIT(dbuf
);
1465 * We can only assign if the offset is aligned, the arc buf is the
1466 * same size as the dbuf, and the dbuf is not metadata. It
1467 * can't be metadata because the loaned arc buf comes from the
1468 * user-data kmem area.
1470 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
&&
1471 DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
) {
1472 dbuf_assign_arcbuf(db
, buf
, tx
);
1473 dbuf_rele(db
, FTAG
);
1478 DB_DNODE_ENTER(dbuf
);
1479 dn
= DB_DNODE(dbuf
);
1481 object
= dn
->dn_object
;
1482 DB_DNODE_EXIT(dbuf
);
1484 dbuf_rele(db
, FTAG
);
1485 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1486 dmu_return_arcbuf(buf
);
1487 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1492 dbuf_dirty_record_t
*dsa_dr
;
1493 dmu_sync_cb_t
*dsa_done
;
1500 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1502 dmu_sync_arg_t
*dsa
= varg
;
1503 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1504 blkptr_t
*bp
= zio
->io_bp
;
1506 if (zio
->io_error
== 0) {
1507 if (BP_IS_HOLE(bp
)) {
1509 * A block of zeros may compress to a hole, but the
1510 * block size still needs to be known for replay.
1512 BP_SET_LSIZE(bp
, db
->db_size
);
1513 } else if (!BP_IS_EMBEDDED(bp
)) {
1514 ASSERT(BP_GET_LEVEL(bp
) == 0);
1521 dmu_sync_late_arrival_ready(zio_t
*zio
)
1523 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1528 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1530 dmu_sync_arg_t
*dsa
= varg
;
1531 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1532 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1534 mutex_enter(&db
->db_mtx
);
1535 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1536 if (zio
->io_error
== 0) {
1537 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1538 if (dr
->dt
.dl
.dr_nopwrite
) {
1539 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1540 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1541 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1543 ASSERT(BP_EQUAL(bp
, bp_orig
));
1544 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1545 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1547 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1548 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1549 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1552 * Old style holes are filled with all zeros, whereas
1553 * new-style holes maintain their lsize, type, level,
1554 * and birth time (see zio_write_compress). While we
1555 * need to reset the BP_SET_LSIZE() call that happened
1556 * in dmu_sync_ready for old style holes, we do *not*
1557 * want to wipe out the information contained in new
1558 * style holes. Thus, only zero out the block pointer if
1559 * it's an old style hole.
1561 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1562 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1563 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1565 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1567 cv_broadcast(&db
->db_changed
);
1568 mutex_exit(&db
->db_mtx
);
1570 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1572 kmem_free(dsa
, sizeof (*dsa
));
1576 dmu_sync_late_arrival_done(zio_t
*zio
)
1578 blkptr_t
*bp
= zio
->io_bp
;
1579 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1580 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1582 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1584 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1585 * then there is nothing to do here. Otherwise, free the
1586 * newly allocated block in this txg.
1588 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1589 ASSERT(BP_EQUAL(bp
, bp_orig
));
1591 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1592 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1593 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1594 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1598 dmu_tx_commit(dsa
->dsa_tx
);
1600 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1602 kmem_free(dsa
, sizeof (*dsa
));
1606 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1607 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1609 dmu_sync_arg_t
*dsa
;
1612 tx
= dmu_tx_create(os
);
1613 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1614 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1616 /* Make zl_get_data do txg_waited_synced() */
1617 return (SET_ERROR(EIO
));
1620 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1622 dsa
->dsa_done
= done
;
1626 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1627 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1628 dmu_sync_late_arrival_ready
, NULL
, dmu_sync_late_arrival_done
, dsa
,
1629 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
|ZIO_FLAG_FASTWRITE
, zb
));
1635 * Intent log support: sync the block associated with db to disk.
1636 * N.B. and XXX: the caller is responsible for making sure that the
1637 * data isn't changing while dmu_sync() is writing it.
1641 * EEXIST: this txg has already been synced, so there's nothing to do.
1642 * The caller should not log the write.
1644 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1645 * The caller should not log the write.
1647 * EALREADY: this block is already in the process of being synced.
1648 * The caller should track its progress (somehow).
1650 * EIO: could not do the I/O.
1651 * The caller should do a txg_wait_synced().
1653 * 0: the I/O has been initiated.
1654 * The caller should log this blkptr in the done callback.
1655 * It is possible that the I/O will fail, in which case
1656 * the error will be reported to the done callback and
1657 * propagated to pio from zio_done().
1660 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1662 blkptr_t
*bp
= zgd
->zgd_bp
;
1663 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1664 objset_t
*os
= db
->db_objset
;
1665 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1666 dbuf_dirty_record_t
*dr
;
1667 dmu_sync_arg_t
*dsa
;
1668 zbookmark_phys_t zb
;
1672 ASSERT(pio
!= NULL
);
1675 SET_BOOKMARK(&zb
, ds
->ds_object
,
1676 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1680 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1684 * If we're frozen (running ziltest), we always need to generate a bp.
1686 if (txg
> spa_freeze_txg(os
->os_spa
))
1687 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1690 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1691 * and us. If we determine that this txg is not yet syncing,
1692 * but it begins to sync a moment later, that's OK because the
1693 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1695 mutex_enter(&db
->db_mtx
);
1697 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1699 * This txg has already synced. There's nothing to do.
1701 mutex_exit(&db
->db_mtx
);
1702 return (SET_ERROR(EEXIST
));
1705 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1707 * This txg is currently syncing, so we can't mess with
1708 * the dirty record anymore; just write a new log block.
1710 mutex_exit(&db
->db_mtx
);
1711 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1714 dr
= db
->db_last_dirty
;
1715 while (dr
&& dr
->dr_txg
!= txg
)
1720 * There's no dr for this dbuf, so it must have been freed.
1721 * There's no need to log writes to freed blocks, so we're done.
1723 mutex_exit(&db
->db_mtx
);
1724 return (SET_ERROR(ENOENT
));
1727 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1730 * Assume the on-disk data is X, the current syncing data (in
1731 * txg - 1) is Y, and the current in-memory data is Z (currently
1734 * We usually want to perform a nopwrite if X and Z are the
1735 * same. However, if Y is different (i.e. the BP is going to
1736 * change before this write takes effect), then a nopwrite will
1737 * be incorrect - we would override with X, which could have
1738 * been freed when Y was written.
1740 * (Note that this is not a concern when we are nop-writing from
1741 * syncing context, because X and Y must be identical, because
1742 * all previous txgs have been synced.)
1744 * Therefore, we disable nopwrite if the current BP could change
1745 * before this TXG. There are two ways it could change: by
1746 * being dirty (dr_next is non-NULL), or by being freed
1747 * (dnode_block_freed()). This behavior is verified by
1748 * zio_done(), which VERIFYs that the override BP is identical
1749 * to the on-disk BP.
1753 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
1754 zp
.zp_nopwrite
= B_FALSE
;
1757 ASSERT(dr
->dr_txg
== txg
);
1758 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1759 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1761 * We have already issued a sync write for this buffer,
1762 * or this buffer has already been synced. It could not
1763 * have been dirtied since, or we would have cleared the state.
1765 mutex_exit(&db
->db_mtx
);
1766 return (SET_ERROR(EALREADY
));
1769 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1770 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1771 mutex_exit(&db
->db_mtx
);
1773 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1775 dsa
->dsa_done
= done
;
1779 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1780 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1781 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1782 NULL
, dmu_sync_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1783 ZIO_FLAG_CANFAIL
, &zb
));
1789 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1795 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1798 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1799 dnode_rele(dn
, FTAG
);
1804 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1810 * Send streams include each object's checksum function. This
1811 * check ensures that the receiving system can understand the
1812 * checksum function transmitted.
1814 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1816 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1817 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1818 dn
->dn_checksum
= checksum
;
1819 dnode_setdirty(dn
, tx
);
1820 dnode_rele(dn
, FTAG
);
1824 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1830 * Send streams include each object's compression function. This
1831 * check ensures that the receiving system can understand the
1832 * compression function transmitted.
1834 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1836 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1837 dn
->dn_compress
= compress
;
1838 dnode_setdirty(dn
, tx
);
1839 dnode_rele(dn
, FTAG
);
1842 int zfs_mdcomp_disable
= 0;
1845 * When the "redundant_metadata" property is set to "most", only indirect
1846 * blocks of this level and higher will have an additional ditto block.
1848 int zfs_redundant_metadata_most_ditto_level
= 2;
1851 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1853 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1854 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1856 enum zio_checksum checksum
= os
->os_checksum
;
1857 enum zio_compress compress
= os
->os_compress
;
1858 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1859 boolean_t dedup
= B_FALSE
;
1860 boolean_t nopwrite
= B_FALSE
;
1861 boolean_t dedup_verify
= os
->os_dedup_verify
;
1862 int copies
= os
->os_copies
;
1865 * We maintain different write policies for each of the following
1868 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1869 * 3. all other level 0 blocks
1872 if (zfs_mdcomp_disable
) {
1873 compress
= ZIO_COMPRESS_EMPTY
;
1876 * XXX -- we should design a compression algorithm
1877 * that specializes in arrays of bps.
1879 compress
= zio_compress_select(os
->os_spa
,
1880 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
1884 * Metadata always gets checksummed. If the data
1885 * checksum is multi-bit correctable, and it's not a
1886 * ZBT-style checksum, then it's suitable for metadata
1887 * as well. Otherwise, the metadata checksum defaults
1890 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1891 zio_checksum_table
[checksum
].ci_eck
)
1892 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1894 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
1895 (os
->os_redundant_metadata
==
1896 ZFS_REDUNDANT_METADATA_MOST
&&
1897 (level
>= zfs_redundant_metadata_most_ditto_level
||
1898 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
1900 } else if (wp
& WP_NOFILL
) {
1904 * If we're writing preallocated blocks, we aren't actually
1905 * writing them so don't set any policy properties. These
1906 * blocks are currently only used by an external subsystem
1907 * outside of zfs (i.e. dump) and not written by the zio
1910 compress
= ZIO_COMPRESS_OFF
;
1911 checksum
= ZIO_CHECKSUM_OFF
;
1913 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
1916 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1917 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1921 * Determine dedup setting. If we are in dmu_sync(),
1922 * we won't actually dedup now because that's all
1923 * done in syncing context; but we do want to use the
1924 * dedup checkum. If the checksum is not strong
1925 * enough to ensure unique signatures, force
1928 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1929 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1930 if (!zio_checksum_table
[checksum
].ci_dedup
)
1931 dedup_verify
= B_TRUE
;
1935 * Enable nopwrite if we have a cryptographically secure
1936 * checksum that has no known collisions (i.e. SHA-256)
1937 * and compression is enabled. We don't enable nopwrite if
1938 * dedup is enabled as the two features are mutually exclusive.
1940 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1941 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1944 zp
->zp_checksum
= checksum
;
1945 zp
->zp_compress
= compress
;
1946 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1947 zp
->zp_level
= level
;
1948 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
1949 zp
->zp_dedup
= dedup
;
1950 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1951 zp
->zp_nopwrite
= nopwrite
;
1955 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1960 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1964 * Sync any current changes before
1965 * we go trundling through the block pointers.
1967 for (i
= 0; i
< TXG_SIZE
; i
++) {
1968 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1971 if (i
!= TXG_SIZE
) {
1972 dnode_rele(dn
, FTAG
);
1973 txg_wait_synced(dmu_objset_pool(os
), 0);
1974 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1979 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1980 dnode_rele(dn
, FTAG
);
1986 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1988 dnode_phys_t
*dnp
= dn
->dn_phys
;
1991 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1992 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1993 1ULL << dn
->dn_indblkshift
: 0;
1994 doi
->doi_type
= dn
->dn_type
;
1995 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1996 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1997 doi
->doi_indirection
= dn
->dn_nlevels
;
1998 doi
->doi_checksum
= dn
->dn_checksum
;
1999 doi
->doi_compress
= dn
->dn_compress
;
2000 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
2001 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
2002 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
2003 doi
->doi_fill_count
= 0;
2004 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
2005 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
2009 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2011 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2012 mutex_enter(&dn
->dn_mtx
);
2014 __dmu_object_info_from_dnode(dn
, doi
);
2016 mutex_exit(&dn
->dn_mtx
);
2017 rw_exit(&dn
->dn_struct_rwlock
);
2021 * Get information on a DMU object.
2022 * If doi is NULL, just indicates whether the object exists.
2025 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
2028 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
2034 dmu_object_info_from_dnode(dn
, doi
);
2036 dnode_rele(dn
, FTAG
);
2041 * As above, but faster; can be used when you have a held dbuf in hand.
2044 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
2046 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2049 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
2054 * Faster still when you only care about the size.
2055 * This is specifically optimized for zfs_getattr().
2058 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
2059 u_longlong_t
*nblk512
)
2061 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2067 *blksize
= dn
->dn_datablksz
;
2068 /* add 1 for dnode space */
2069 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
2070 SPA_MINBLOCKSHIFT
) + 1;
2075 byteswap_uint64_array(void *vbuf
, size_t size
)
2077 uint64_t *buf
= vbuf
;
2078 size_t count
= size
>> 3;
2081 ASSERT((size
& 7) == 0);
2083 for (i
= 0; i
< count
; i
++)
2084 buf
[i
] = BSWAP_64(buf
[i
]);
2088 byteswap_uint32_array(void *vbuf
, size_t size
)
2090 uint32_t *buf
= vbuf
;
2091 size_t count
= size
>> 2;
2094 ASSERT((size
& 3) == 0);
2096 for (i
= 0; i
< count
; i
++)
2097 buf
[i
] = BSWAP_32(buf
[i
]);
2101 byteswap_uint16_array(void *vbuf
, size_t size
)
2103 uint16_t *buf
= vbuf
;
2104 size_t count
= size
>> 1;
2107 ASSERT((size
& 1) == 0);
2109 for (i
= 0; i
< count
; i
++)
2110 buf
[i
] = BSWAP_16(buf
[i
]);
2115 byteswap_uint8_array(void *vbuf
, size_t size
)
2137 arc_fini(); /* arc depends on l2arc, so arc must go first */
2149 #if defined(_KERNEL) && defined(HAVE_SPL)
2150 EXPORT_SYMBOL(dmu_bonus_hold
);
2151 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2152 EXPORT_SYMBOL(dmu_buf_rele_array
);
2153 EXPORT_SYMBOL(dmu_prefetch
);
2154 EXPORT_SYMBOL(dmu_free_range
);
2155 EXPORT_SYMBOL(dmu_free_long_range
);
2156 EXPORT_SYMBOL(dmu_free_long_object
);
2157 EXPORT_SYMBOL(dmu_read
);
2158 EXPORT_SYMBOL(dmu_write
);
2159 EXPORT_SYMBOL(dmu_prealloc
);
2160 EXPORT_SYMBOL(dmu_object_info
);
2161 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2162 EXPORT_SYMBOL(dmu_object_info_from_db
);
2163 EXPORT_SYMBOL(dmu_object_size_from_db
);
2164 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2165 EXPORT_SYMBOL(dmu_object_set_checksum
);
2166 EXPORT_SYMBOL(dmu_object_set_compress
);
2167 EXPORT_SYMBOL(dmu_write_policy
);
2168 EXPORT_SYMBOL(dmu_sync
);
2169 EXPORT_SYMBOL(dmu_request_arcbuf
);
2170 EXPORT_SYMBOL(dmu_return_arcbuf
);
2171 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2172 EXPORT_SYMBOL(dmu_buf_hold
);
2173 EXPORT_SYMBOL(dmu_ot
);
2175 module_param(zfs_mdcomp_disable
, int, 0644);
2176 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2178 module_param(zfs_nopwrite_enabled
, int, 0644);
2179 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");