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) 2012, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
28 #include <sys/dmu_impl.h>
29 #include <sys/dmu_tx.h>
31 #include <sys/dnode.h>
32 #include <sys/zfs_context.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dmu_traverse.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dsl_pool.h>
38 #include <sys/dsl_synctask.h>
39 #include <sys/dsl_prop.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/zfs_ioctl.h>
43 #include <sys/zio_checksum.h>
44 #include <sys/zio_compress.h>
47 #include <sys/vmsystm.h>
48 #include <sys/zfs_znode.h>
52 * Enable/disable nopwrite feature.
54 int zfs_nopwrite_enabled
= 1;
56 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
57 { DMU_BSWAP_UINT8
, TRUE
, "unallocated" },
58 { DMU_BSWAP_ZAP
, TRUE
, "object directory" },
59 { DMU_BSWAP_UINT64
, TRUE
, "object array" },
60 { DMU_BSWAP_UINT8
, TRUE
, "packed nvlist" },
61 { DMU_BSWAP_UINT64
, TRUE
, "packed nvlist size" },
62 { DMU_BSWAP_UINT64
, TRUE
, "bpobj" },
63 { DMU_BSWAP_UINT64
, TRUE
, "bpobj header" },
64 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map header" },
65 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map" },
66 { DMU_BSWAP_UINT64
, TRUE
, "ZIL intent log" },
67 { DMU_BSWAP_DNODE
, TRUE
, "DMU dnode" },
68 { DMU_BSWAP_OBJSET
, TRUE
, "DMU objset" },
69 { DMU_BSWAP_UINT64
, TRUE
, "DSL directory" },
70 { DMU_BSWAP_ZAP
, TRUE
, "DSL directory child map"},
71 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset snap map" },
72 { DMU_BSWAP_ZAP
, TRUE
, "DSL props" },
73 { DMU_BSWAP_UINT64
, TRUE
, "DSL dataset" },
74 { DMU_BSWAP_ZNODE
, TRUE
, "ZFS znode" },
75 { DMU_BSWAP_OLDACL
, TRUE
, "ZFS V0 ACL" },
76 { DMU_BSWAP_UINT8
, FALSE
, "ZFS plain file" },
77 { DMU_BSWAP_ZAP
, TRUE
, "ZFS directory" },
78 { DMU_BSWAP_ZAP
, TRUE
, "ZFS master node" },
79 { DMU_BSWAP_ZAP
, TRUE
, "ZFS delete queue" },
80 { DMU_BSWAP_UINT8
, FALSE
, "zvol object" },
81 { DMU_BSWAP_ZAP
, TRUE
, "zvol prop" },
82 { DMU_BSWAP_UINT8
, FALSE
, "other uint8[]" },
83 { DMU_BSWAP_UINT64
, FALSE
, "other uint64[]" },
84 { DMU_BSWAP_ZAP
, TRUE
, "other ZAP" },
85 { DMU_BSWAP_ZAP
, TRUE
, "persistent error log" },
86 { DMU_BSWAP_UINT8
, TRUE
, "SPA history" },
87 { DMU_BSWAP_UINT64
, TRUE
, "SPA history offsets" },
88 { DMU_BSWAP_ZAP
, TRUE
, "Pool properties" },
89 { DMU_BSWAP_ZAP
, TRUE
, "DSL permissions" },
90 { DMU_BSWAP_ACL
, TRUE
, "ZFS ACL" },
91 { DMU_BSWAP_UINT8
, TRUE
, "ZFS SYSACL" },
92 { DMU_BSWAP_UINT8
, TRUE
, "FUID table" },
93 { DMU_BSWAP_UINT64
, TRUE
, "FUID table size" },
94 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset next clones"},
95 { DMU_BSWAP_ZAP
, TRUE
, "scan work queue" },
96 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group used" },
97 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group quota" },
98 { DMU_BSWAP_ZAP
, TRUE
, "snapshot refcount tags"},
99 { DMU_BSWAP_ZAP
, TRUE
, "DDT ZAP algorithm" },
100 { DMU_BSWAP_ZAP
, TRUE
, "DDT statistics" },
101 { DMU_BSWAP_UINT8
, TRUE
, "System attributes" },
102 { DMU_BSWAP_ZAP
, TRUE
, "SA master node" },
103 { DMU_BSWAP_ZAP
, TRUE
, "SA attr registration" },
104 { DMU_BSWAP_ZAP
, TRUE
, "SA attr layouts" },
105 { DMU_BSWAP_ZAP
, TRUE
, "scan translations" },
106 { DMU_BSWAP_UINT8
, FALSE
, "deduplicated block" },
107 { DMU_BSWAP_ZAP
, TRUE
, "DSL deadlist map" },
108 { DMU_BSWAP_UINT64
, TRUE
, "DSL deadlist map hdr" },
109 { DMU_BSWAP_ZAP
, TRUE
, "DSL dir clones" },
110 { DMU_BSWAP_UINT64
, TRUE
, "bpobj subobj" }
113 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
114 { byteswap_uint8_array
, "uint8" },
115 { byteswap_uint16_array
, "uint16" },
116 { byteswap_uint32_array
, "uint32" },
117 { byteswap_uint64_array
, "uint64" },
118 { zap_byteswap
, "zap" },
119 { dnode_buf_byteswap
, "dnode" },
120 { dmu_objset_byteswap
, "objset" },
121 { zfs_znode_byteswap
, "znode" },
122 { zfs_oldacl_byteswap
, "oldacl" },
123 { zfs_acl_byteswap
, "acl" }
127 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
128 void *tag
, dmu_buf_t
**dbp
)
135 err
= dnode_hold(os
, object
, FTAG
, &dn
);
138 blkid
= dbuf_whichblock(dn
, offset
);
139 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
140 db
= dbuf_hold(dn
, blkid
, tag
);
141 rw_exit(&dn
->dn_struct_rwlock
);
142 dnode_rele(dn
, FTAG
);
146 return (SET_ERROR(EIO
));
154 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
155 void *tag
, dmu_buf_t
**dbp
, int flags
)
158 int db_flags
= DB_RF_CANFAIL
;
160 if (flags
& DMU_READ_NO_PREFETCH
)
161 db_flags
|= DB_RF_NOPREFETCH
;
163 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
165 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
166 err
= dbuf_read(db
, NULL
, db_flags
);
179 return (DN_MAX_BONUSLEN
);
183 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
185 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
192 if (dn
->dn_bonus
!= db
) {
193 error
= SET_ERROR(EINVAL
);
194 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
195 error
= SET_ERROR(EINVAL
);
197 dnode_setbonuslen(dn
, newsize
, tx
);
206 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
208 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
215 if (!DMU_OT_IS_VALID(type
)) {
216 error
= SET_ERROR(EINVAL
);
217 } else if (dn
->dn_bonus
!= db
) {
218 error
= SET_ERROR(EINVAL
);
220 dnode_setbonus_type(dn
, type
, tx
);
229 dmu_get_bonustype(dmu_buf_t
*db_fake
)
231 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
233 dmu_object_type_t type
;
237 type
= dn
->dn_bonustype
;
244 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
249 error
= dnode_hold(os
, object
, FTAG
, &dn
);
250 dbuf_rm_spill(dn
, tx
);
251 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
252 dnode_rm_spill(dn
, tx
);
253 rw_exit(&dn
->dn_struct_rwlock
);
254 dnode_rele(dn
, FTAG
);
259 * returns ENOENT, EIO, or 0.
262 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
268 error
= dnode_hold(os
, object
, FTAG
, &dn
);
272 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
273 if (dn
->dn_bonus
== NULL
) {
274 rw_exit(&dn
->dn_struct_rwlock
);
275 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
276 if (dn
->dn_bonus
== NULL
)
277 dbuf_create_bonus(dn
);
281 /* as long as the bonus buf is held, the dnode will be held */
282 if (refcount_add(&db
->db_holds
, tag
) == 1) {
283 VERIFY(dnode_add_ref(dn
, db
));
284 (void) atomic_inc_32_nv(&dn
->dn_dbufs_count
);
288 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
289 * hold and incrementing the dbuf count to ensure that dnode_move() sees
290 * a dnode hold for every dbuf.
292 rw_exit(&dn
->dn_struct_rwlock
);
294 dnode_rele(dn
, FTAG
);
296 VERIFY(0 == dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
));
303 * returns ENOENT, EIO, or 0.
305 * This interface will allocate a blank spill dbuf when a spill blk
306 * doesn't already exist on the dnode.
308 * if you only want to find an already existing spill db, then
309 * dmu_spill_hold_existing() should be used.
312 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
314 dmu_buf_impl_t
*db
= NULL
;
317 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
318 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
320 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
322 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
323 rw_exit(&dn
->dn_struct_rwlock
);
326 err
= dbuf_read(db
, NULL
, flags
);
335 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
337 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
344 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
345 err
= SET_ERROR(EINVAL
);
347 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
349 if (!dn
->dn_have_spill
) {
350 err
= SET_ERROR(ENOENT
);
352 err
= dmu_spill_hold_by_dnode(dn
,
353 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
356 rw_exit(&dn
->dn_struct_rwlock
);
364 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
366 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
372 err
= dmu_spill_hold_by_dnode(dn
, DB_RF_CANFAIL
, tag
, dbp
);
379 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
380 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
381 * and can induce severe lock contention when writing to several files
382 * whose dnodes are in the same block.
385 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
386 int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
389 uint64_t blkid
, nblks
, i
;
394 ASSERT(length
<= DMU_MAX_ACCESS
);
396 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
;
397 if (flags
& DMU_READ_NO_PREFETCH
|| length
> zfetch_array_rd_sz
)
398 dbuf_flags
|= DB_RF_NOPREFETCH
;
400 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
401 if (dn
->dn_datablkshift
) {
402 int blkshift
= dn
->dn_datablkshift
;
403 nblks
= (P2ROUNDUP(offset
+length
, 1ULL<<blkshift
) -
404 P2ALIGN(offset
, 1ULL<<blkshift
)) >> blkshift
;
406 if (offset
+ length
> dn
->dn_datablksz
) {
407 zfs_panic_recover("zfs: accessing past end of object "
408 "%llx/%llx (size=%u access=%llu+%llu)",
409 (longlong_t
)dn
->dn_objset
->
410 os_dsl_dataset
->ds_object
,
411 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
412 (longlong_t
)offset
, (longlong_t
)length
);
413 rw_exit(&dn
->dn_struct_rwlock
);
414 return (SET_ERROR(EIO
));
418 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
,
419 KM_PUSHPAGE
| KM_NODEBUG
);
421 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
422 blkid
= dbuf_whichblock(dn
, offset
);
423 for (i
= 0; i
< nblks
; i
++) {
424 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+i
, tag
);
426 rw_exit(&dn
->dn_struct_rwlock
);
427 dmu_buf_rele_array(dbp
, nblks
, tag
);
429 return (SET_ERROR(EIO
));
431 /* initiate async i/o */
433 (void) dbuf_read(db
, zio
, dbuf_flags
);
437 rw_exit(&dn
->dn_struct_rwlock
);
439 /* wait for async i/o */
442 dmu_buf_rele_array(dbp
, nblks
, tag
);
446 /* wait for other io to complete */
448 for (i
= 0; i
< nblks
; i
++) {
449 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
450 mutex_enter(&db
->db_mtx
);
451 while (db
->db_state
== DB_READ
||
452 db
->db_state
== DB_FILL
)
453 cv_wait(&db
->db_changed
, &db
->db_mtx
);
454 if (db
->db_state
== DB_UNCACHED
)
455 err
= SET_ERROR(EIO
);
456 mutex_exit(&db
->db_mtx
);
458 dmu_buf_rele_array(dbp
, nblks
, tag
);
470 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
471 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
476 err
= dnode_hold(os
, object
, FTAG
, &dn
);
480 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
481 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
483 dnode_rele(dn
, FTAG
);
489 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
490 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
492 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
498 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
499 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
506 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
509 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
514 for (i
= 0; i
< numbufs
; i
++) {
516 dbuf_rele(dbp
[i
], tag
);
519 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
523 * Issue prefetch i/os for the given blocks.
525 * Note: The assumption is that we *know* these blocks will be needed
526 * almost immediately. Therefore, the prefetch i/os will be issued at
527 * ZIO_PRIORITY_SYNC_READ
529 * Note: indirect blocks and other metadata will be read synchronously,
530 * causing this function to block if they are not already cached.
533 dmu_prefetch(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t len
)
539 if (zfs_prefetch_disable
)
542 if (len
== 0) { /* they're interested in the bonus buffer */
543 dn
= DMU_META_DNODE(os
);
545 if (object
== 0 || object
>= DN_MAX_OBJECT
)
548 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
549 blkid
= dbuf_whichblock(dn
, object
* sizeof (dnode_phys_t
));
550 dbuf_prefetch(dn
, blkid
, ZIO_PRIORITY_SYNC_READ
);
551 rw_exit(&dn
->dn_struct_rwlock
);
556 * XXX - Note, if the dnode for the requested object is not
557 * already cached, we will do a *synchronous* read in the
558 * dnode_hold() call. The same is true for any indirects.
560 err
= dnode_hold(os
, object
, FTAG
, &dn
);
564 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
565 if (dn
->dn_datablkshift
) {
566 int blkshift
= dn
->dn_datablkshift
;
567 nblks
= (P2ROUNDUP(offset
+ len
, 1 << blkshift
) -
568 P2ALIGN(offset
, 1 << blkshift
)) >> blkshift
;
570 nblks
= (offset
< dn
->dn_datablksz
);
576 blkid
= dbuf_whichblock(dn
, offset
);
577 for (i
= 0; i
< nblks
; i
++)
578 dbuf_prefetch(dn
, blkid
+ i
, ZIO_PRIORITY_SYNC_READ
);
581 rw_exit(&dn
->dn_struct_rwlock
);
583 dnode_rele(dn
, FTAG
);
587 * Get the next "chunk" of file data to free. We traverse the file from
588 * the end so that the file gets shorter over time (if we crashes in the
589 * middle, this will leave us in a better state). We find allocated file
590 * data by simply searching the allocated level 1 indirects.
592 * On input, *start should be the first offset that does not need to be
593 * freed (e.g. "offset + length"). On return, *start will be the first
594 * offset that should be freed.
597 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
599 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
600 /* bytes of data covered by a level-1 indirect block */
602 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
605 ASSERT3U(minimum
, <=, *start
);
607 if (*start
- minimum
<= iblkrange
* maxblks
) {
611 ASSERT(ISP2(iblkrange
));
613 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
617 * dnode_next_offset(BACKWARDS) will find an allocated L1
618 * indirect block at or before the input offset. We must
619 * decrement *start so that it is at the end of the region
623 err
= dnode_next_offset(dn
,
624 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
626 /* if there are no indirect blocks before start, we are done */
630 } else if (err
!= 0) {
634 /* set start to the beginning of this L1 indirect */
635 *start
= P2ALIGN(*start
, iblkrange
);
637 if (*start
< minimum
)
643 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
646 uint64_t object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
649 if (offset
>= object_size
)
652 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
653 length
= object_size
- offset
;
655 while (length
!= 0) {
656 uint64_t chunk_end
, chunk_begin
;
659 chunk_end
= chunk_begin
= offset
+ length
;
661 /* move chunk_begin backwards to the beginning of this chunk */
662 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
665 ASSERT3U(chunk_begin
, >=, offset
);
666 ASSERT3U(chunk_begin
, <=, chunk_end
);
668 tx
= dmu_tx_create(os
);
669 dmu_tx_hold_free(tx
, dn
->dn_object
,
670 chunk_begin
, chunk_end
- chunk_begin
);
671 err
= dmu_tx_assign(tx
, TXG_WAIT
);
676 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
679 length
-= chunk_end
- chunk_begin
;
685 dmu_free_long_range(objset_t
*os
, uint64_t object
,
686 uint64_t offset
, uint64_t length
)
691 err
= dnode_hold(os
, object
, FTAG
, &dn
);
694 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
697 * It is important to zero out the maxblkid when freeing the entire
698 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
699 * will take the fast path, and (b) dnode_reallocate() can verify
700 * that the entire file has been freed.
702 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
705 dnode_rele(dn
, FTAG
);
710 dmu_free_long_object(objset_t
*os
, uint64_t object
)
715 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
719 tx
= dmu_tx_create(os
);
720 dmu_tx_hold_bonus(tx
, object
);
721 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
722 err
= dmu_tx_assign(tx
, TXG_WAIT
);
724 err
= dmu_object_free(os
, object
, tx
);
734 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
735 uint64_t size
, dmu_tx_t
*tx
)
738 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
741 ASSERT(offset
< UINT64_MAX
);
742 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
743 dnode_free_range(dn
, offset
, size
, tx
);
744 dnode_rele(dn
, FTAG
);
749 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
750 void *buf
, uint32_t flags
)
756 err
= dnode_hold(os
, object
, FTAG
, &dn
);
761 * Deal with odd block sizes, where there can't be data past the first
762 * block. If we ever do the tail block optimization, we will need to
763 * handle that here as well.
765 if (dn
->dn_maxblkid
== 0) {
766 int newsz
= offset
> dn
->dn_datablksz
? 0 :
767 MIN(size
, dn
->dn_datablksz
- offset
);
768 bzero((char *)buf
+ newsz
, size
- newsz
);
773 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
777 * NB: we could do this block-at-a-time, but it's nice
778 * to be reading in parallel.
780 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
781 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
785 for (i
= 0; i
< numbufs
; i
++) {
788 dmu_buf_t
*db
= dbp
[i
];
792 bufoff
= offset
- db
->db_offset
;
793 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
795 bcopy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
799 buf
= (char *)buf
+ tocpy
;
801 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
803 dnode_rele(dn
, FTAG
);
808 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
809 const void *buf
, dmu_tx_t
*tx
)
817 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
818 FALSE
, FTAG
, &numbufs
, &dbp
));
820 for (i
= 0; i
< numbufs
; i
++) {
823 dmu_buf_t
*db
= dbp
[i
];
827 bufoff
= offset
- db
->db_offset
;
828 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
830 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
832 if (tocpy
== db
->db_size
)
833 dmu_buf_will_fill(db
, tx
);
835 dmu_buf_will_dirty(db
, tx
);
837 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
839 if (tocpy
== db
->db_size
)
840 dmu_buf_fill_done(db
, tx
);
844 buf
= (char *)buf
+ tocpy
;
846 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
850 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
859 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
860 FALSE
, FTAG
, &numbufs
, &dbp
));
862 for (i
= 0; i
< numbufs
; i
++) {
863 dmu_buf_t
*db
= dbp
[i
];
865 dmu_buf_will_not_fill(db
, tx
);
867 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
871 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
872 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
873 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
877 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
878 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
879 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
882 dmu_buf_write_embedded(db
,
883 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
884 uncompressed_size
, compressed_size
, byteorder
, tx
);
886 dmu_buf_rele(db
, FTAG
);
890 * DMU support for xuio
892 kstat_t
*xuio_ksp
= NULL
;
894 typedef struct xuio_stats
{
895 /* loaned yet not returned arc_buf */
896 kstat_named_t xuiostat_onloan_rbuf
;
897 kstat_named_t xuiostat_onloan_wbuf
;
898 /* whether a copy is made when loaning out a read buffer */
899 kstat_named_t xuiostat_rbuf_copied
;
900 kstat_named_t xuiostat_rbuf_nocopy
;
901 /* whether a copy is made when assigning a write buffer */
902 kstat_named_t xuiostat_wbuf_copied
;
903 kstat_named_t xuiostat_wbuf_nocopy
;
906 static xuio_stats_t xuio_stats
= {
907 { "onloan_read_buf", KSTAT_DATA_UINT64
},
908 { "onloan_write_buf", KSTAT_DATA_UINT64
},
909 { "read_buf_copied", KSTAT_DATA_UINT64
},
910 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
911 { "write_buf_copied", KSTAT_DATA_UINT64
},
912 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
915 #define XUIOSTAT_INCR(stat, val) \
916 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
917 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
920 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
923 uio_t
*uio
= &xuio
->xu_uio
;
925 uio
->uio_iovcnt
= nblk
;
926 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_PUSHPAGE
);
928 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_PUSHPAGE
);
930 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_PUSHPAGE
);
931 priv
->iovp
= uio
->uio_iov
;
932 XUIO_XUZC_PRIV(xuio
) = priv
;
934 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
935 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
937 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
943 dmu_xuio_fini(xuio_t
*xuio
)
945 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
946 int nblk
= priv
->cnt
;
948 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
949 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
950 kmem_free(priv
, sizeof (dmu_xuio_t
));
952 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
953 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
955 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
959 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
960 * and increase priv->next by 1.
963 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
966 uio_t
*uio
= &xuio
->xu_uio
;
967 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
968 int i
= priv
->next
++;
970 ASSERT(i
< priv
->cnt
);
971 ASSERT(off
+ n
<= arc_buf_size(abuf
));
972 iov
= uio
->uio_iov
+ i
;
973 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
975 priv
->bufs
[i
] = abuf
;
980 dmu_xuio_cnt(xuio_t
*xuio
)
982 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
987 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
989 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
991 ASSERT(i
< priv
->cnt
);
992 return (priv
->bufs
[i
]);
996 dmu_xuio_clear(xuio_t
*xuio
, int i
)
998 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1000 ASSERT(i
< priv
->cnt
);
1001 priv
->bufs
[i
] = NULL
;
1005 xuio_stat_init(void)
1007 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1008 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1009 KSTAT_FLAG_VIRTUAL
);
1010 if (xuio_ksp
!= NULL
) {
1011 xuio_ksp
->ks_data
= &xuio_stats
;
1012 kstat_install(xuio_ksp
);
1017 xuio_stat_fini(void)
1019 if (xuio_ksp
!= NULL
) {
1020 kstat_delete(xuio_ksp
);
1026 xuio_stat_wbuf_copied()
1028 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1032 xuio_stat_wbuf_nocopy()
1034 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1040 * Copy up to size bytes between arg_buf and req based on the data direction
1041 * described by the req. If an entire req's data cannot be transfered in one
1042 * pass, you should pass in @req_offset to indicate where to continue. The
1043 * return value is the number of bytes successfully copied to arg_buf.
1046 dmu_req_copy(void *arg_buf
, int size
, struct request
*req
, size_t req_offset
)
1048 struct bio_vec bv
, *bvp
;
1049 struct req_iterator iter
;
1051 int tocpy
, bv_len
, bv_offset
;
1054 rq_for_each_segment4(bv
, bvp
, req
, iter
) {
1056 * Fully consumed the passed arg_buf. We use goto here because
1057 * rq_for_each_segment is a double loop
1059 ASSERT3S(offset
, <=, size
);
1063 /* Skip already copied bv */
1064 if (req_offset
>= bv
.bv_len
) {
1065 req_offset
-= bv
.bv_len
;
1069 bv_len
= bv
.bv_len
- req_offset
;
1070 bv_offset
= bv
.bv_offset
+ req_offset
;
1073 tocpy
= MIN(bv_len
, size
- offset
);
1074 ASSERT3S(tocpy
, >=, 0);
1076 bv_buf
= page_address(bv
.bv_page
) + bv_offset
;
1077 ASSERT3P(bv_buf
, !=, NULL
);
1079 if (rq_data_dir(req
) == WRITE
)
1080 memcpy(arg_buf
+ offset
, bv_buf
, tocpy
);
1082 memcpy(bv_buf
, arg_buf
+ offset
, tocpy
);
1091 dmu_read_req(objset_t
*os
, uint64_t object
, struct request
*req
)
1093 uint64_t size
= blk_rq_bytes(req
);
1094 uint64_t offset
= blk_rq_pos(req
) << 9;
1096 int numbufs
, i
, err
;
1100 * NB: we could do this block-at-a-time, but it's nice
1101 * to be reading in parallel.
1103 err
= dmu_buf_hold_array(os
, object
, offset
, size
, TRUE
, FTAG
,
1109 for (i
= 0; i
< numbufs
; i
++) {
1110 int tocpy
, didcpy
, bufoff
;
1111 dmu_buf_t
*db
= dbp
[i
];
1113 bufoff
= offset
- db
->db_offset
;
1114 ASSERT3S(bufoff
, >=, 0);
1116 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1120 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1131 req_offset
+= didcpy
;
1134 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1140 dmu_write_req(objset_t
*os
, uint64_t object
, struct request
*req
, dmu_tx_t
*tx
)
1142 uint64_t size
= blk_rq_bytes(req
);
1143 uint64_t offset
= blk_rq_pos(req
) << 9;
1145 int numbufs
, i
, err
;
1151 err
= dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1157 for (i
= 0; i
< numbufs
; i
++) {
1158 int tocpy
, didcpy
, bufoff
;
1159 dmu_buf_t
*db
= dbp
[i
];
1161 bufoff
= offset
- db
->db_offset
;
1162 ASSERT3S(bufoff
, >=, 0);
1164 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1168 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1170 if (tocpy
== db
->db_size
)
1171 dmu_buf_will_fill(db
, tx
);
1173 dmu_buf_will_dirty(db
, tx
);
1175 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1178 if (tocpy
== db
->db_size
)
1179 dmu_buf_fill_done(db
, tx
);
1189 req_offset
+= didcpy
;
1193 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1198 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1201 int numbufs
, i
, err
;
1202 xuio_t
*xuio
= NULL
;
1205 * NB: we could do this block-at-a-time, but it's nice
1206 * to be reading in parallel.
1208 err
= dmu_buf_hold_array(os
, object
, uio
->uio_loffset
, size
, TRUE
, FTAG
,
1213 for (i
= 0; i
< numbufs
; i
++) {
1216 dmu_buf_t
*db
= dbp
[i
];
1220 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1221 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1224 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1225 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1226 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1227 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1229 uio
->uio_resid
-= tocpy
;
1230 uio
->uio_loffset
+= tocpy
;
1233 if (abuf
== dbuf_abuf
)
1234 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1236 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1238 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1246 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1252 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1259 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1260 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1264 for (i
= 0; i
< numbufs
; i
++) {
1267 dmu_buf_t
*db
= dbp
[i
];
1271 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1272 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1274 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1276 if (tocpy
== db
->db_size
)
1277 dmu_buf_will_fill(db
, tx
);
1279 dmu_buf_will_dirty(db
, tx
);
1282 * XXX uiomove could block forever (eg.nfs-backed
1283 * pages). There needs to be a uiolockdown() function
1284 * to lock the pages in memory, so that uiomove won't
1287 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1290 if (tocpy
== db
->db_size
)
1291 dmu_buf_fill_done(db
, tx
);
1299 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1304 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1307 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1316 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1323 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1332 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1336 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1338 dnode_rele(dn
, FTAG
);
1342 #endif /* _KERNEL */
1345 * Allocate a loaned anonymous arc buffer.
1348 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1350 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1352 return (arc_loan_buf(db
->db_objset
->os_spa
, size
));
1356 * Free a loaned arc buffer.
1359 dmu_return_arcbuf(arc_buf_t
*buf
)
1361 arc_return_buf(buf
, FTAG
);
1362 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1366 * When possible directly assign passed loaned arc buffer to a dbuf.
1367 * If this is not possible copy the contents of passed arc buf via
1371 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1374 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1377 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1380 DB_DNODE_ENTER(dbuf
);
1381 dn
= DB_DNODE(dbuf
);
1382 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1383 blkid
= dbuf_whichblock(dn
, offset
);
1384 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1385 rw_exit(&dn
->dn_struct_rwlock
);
1386 DB_DNODE_EXIT(dbuf
);
1388 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1389 dbuf_assign_arcbuf(db
, buf
, tx
);
1390 dbuf_rele(db
, FTAG
);
1395 DB_DNODE_ENTER(dbuf
);
1396 dn
= DB_DNODE(dbuf
);
1398 object
= dn
->dn_object
;
1399 DB_DNODE_EXIT(dbuf
);
1401 dbuf_rele(db
, FTAG
);
1402 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1403 dmu_return_arcbuf(buf
);
1404 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1409 dbuf_dirty_record_t
*dsa_dr
;
1410 dmu_sync_cb_t
*dsa_done
;
1417 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1419 dmu_sync_arg_t
*dsa
= varg
;
1420 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1421 blkptr_t
*bp
= zio
->io_bp
;
1423 if (zio
->io_error
== 0) {
1424 if (BP_IS_HOLE(bp
)) {
1426 * A block of zeros may compress to a hole, but the
1427 * block size still needs to be known for replay.
1429 BP_SET_LSIZE(bp
, db
->db_size
);
1430 } else if (!BP_IS_EMBEDDED(bp
)) {
1431 ASSERT(BP_GET_LEVEL(bp
) == 0);
1438 dmu_sync_late_arrival_ready(zio_t
*zio
)
1440 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1445 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1447 dmu_sync_arg_t
*dsa
= varg
;
1448 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1449 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1451 mutex_enter(&db
->db_mtx
);
1452 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1453 if (zio
->io_error
== 0) {
1454 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1455 if (dr
->dt
.dl
.dr_nopwrite
) {
1456 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1457 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1458 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1460 ASSERT(BP_EQUAL(bp
, bp_orig
));
1461 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1462 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1464 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1465 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1466 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1467 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
))
1468 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1470 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1472 cv_broadcast(&db
->db_changed
);
1473 mutex_exit(&db
->db_mtx
);
1475 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1477 kmem_free(dsa
, sizeof (*dsa
));
1481 dmu_sync_late_arrival_done(zio_t
*zio
)
1483 blkptr_t
*bp
= zio
->io_bp
;
1484 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1485 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1487 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1489 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1490 * then there is nothing to do here. Otherwise, free the
1491 * newly allocated block in this txg.
1493 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1494 ASSERT(BP_EQUAL(bp
, bp_orig
));
1496 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1497 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1498 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1499 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1503 dmu_tx_commit(dsa
->dsa_tx
);
1505 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1507 kmem_free(dsa
, sizeof (*dsa
));
1511 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1512 zio_prop_t
*zp
, zbookmark_t
*zb
)
1514 dmu_sync_arg_t
*dsa
;
1517 tx
= dmu_tx_create(os
);
1518 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1519 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1521 /* Make zl_get_data do txg_waited_synced() */
1522 return (SET_ERROR(EIO
));
1525 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1527 dsa
->dsa_done
= done
;
1531 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1532 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1533 dmu_sync_late_arrival_ready
, NULL
, dmu_sync_late_arrival_done
, dsa
,
1534 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
|ZIO_FLAG_FASTWRITE
, zb
));
1540 * Intent log support: sync the block associated with db to disk.
1541 * N.B. and XXX: the caller is responsible for making sure that the
1542 * data isn't changing while dmu_sync() is writing it.
1546 * EEXIST: this txg has already been synced, so there's nothing to do.
1547 * The caller should not log the write.
1549 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1550 * The caller should not log the write.
1552 * EALREADY: this block is already in the process of being synced.
1553 * The caller should track its progress (somehow).
1555 * EIO: could not do the I/O.
1556 * The caller should do a txg_wait_synced().
1558 * 0: the I/O has been initiated.
1559 * The caller should log this blkptr in the done callback.
1560 * It is possible that the I/O will fail, in which case
1561 * the error will be reported to the done callback and
1562 * propagated to pio from zio_done().
1565 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1567 blkptr_t
*bp
= zgd
->zgd_bp
;
1568 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1569 objset_t
*os
= db
->db_objset
;
1570 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1571 dbuf_dirty_record_t
*dr
;
1572 dmu_sync_arg_t
*dsa
;
1577 ASSERT(pio
!= NULL
);
1580 SET_BOOKMARK(&zb
, ds
->ds_object
,
1581 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1585 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1589 * If we're frozen (running ziltest), we always need to generate a bp.
1591 if (txg
> spa_freeze_txg(os
->os_spa
))
1592 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1595 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1596 * and us. If we determine that this txg is not yet syncing,
1597 * but it begins to sync a moment later, that's OK because the
1598 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1600 mutex_enter(&db
->db_mtx
);
1602 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1604 * This txg has already synced. There's nothing to do.
1606 mutex_exit(&db
->db_mtx
);
1607 return (SET_ERROR(EEXIST
));
1610 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1612 * This txg is currently syncing, so we can't mess with
1613 * the dirty record anymore; just write a new log block.
1615 mutex_exit(&db
->db_mtx
);
1616 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1619 dr
= db
->db_last_dirty
;
1620 while (dr
&& dr
->dr_txg
!= txg
)
1625 * There's no dr for this dbuf, so it must have been freed.
1626 * There's no need to log writes to freed blocks, so we're done.
1628 mutex_exit(&db
->db_mtx
);
1629 return (SET_ERROR(ENOENT
));
1632 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1635 * Assume the on-disk data is X, the current syncing data is Y,
1636 * and the current in-memory data is Z (currently in dmu_sync).
1637 * X and Z are identical but Y is has been modified. Normally,
1638 * when X and Z are the same we will perform a nopwrite but if Y
1639 * is different we must disable nopwrite since the resulting write
1640 * of Y to disk can free the block containing X. If we allowed a
1641 * nopwrite to occur the block pointing to Z would reference a freed
1642 * block. Since this is a rare case we simplify this by disabling
1643 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1644 * a previous transaction.
1647 zp
.zp_nopwrite
= B_FALSE
;
1649 ASSERT(dr
->dr_txg
== txg
);
1650 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1651 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1653 * We have already issued a sync write for this buffer,
1654 * or this buffer has already been synced. It could not
1655 * have been dirtied since, or we would have cleared the state.
1657 mutex_exit(&db
->db_mtx
);
1658 return (SET_ERROR(EALREADY
));
1661 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1662 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1663 mutex_exit(&db
->db_mtx
);
1665 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1667 dsa
->dsa_done
= done
;
1671 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1672 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1673 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1674 NULL
, dmu_sync_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1675 ZIO_FLAG_CANFAIL
, &zb
));
1681 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1687 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1690 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1691 dnode_rele(dn
, FTAG
);
1696 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1702 * Send streams include each object's checksum function. This
1703 * check ensures that the receiving system can understand the
1704 * checksum function transmitted.
1706 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1708 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1709 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1710 dn
->dn_checksum
= checksum
;
1711 dnode_setdirty(dn
, tx
);
1712 dnode_rele(dn
, FTAG
);
1716 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1722 * Send streams include each object's compression function. This
1723 * check ensures that the receiving system can understand the
1724 * compression function transmitted.
1726 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1728 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1729 dn
->dn_compress
= compress
;
1730 dnode_setdirty(dn
, tx
);
1731 dnode_rele(dn
, FTAG
);
1734 int zfs_mdcomp_disable
= 0;
1737 * When the "redundant_metadata" property is set to "most", only indirect
1738 * blocks of this level and higher will have an additional ditto block.
1740 int zfs_redundant_metadata_most_ditto_level
= 2;
1743 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1745 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1746 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1748 enum zio_checksum checksum
= os
->os_checksum
;
1749 enum zio_compress compress
= os
->os_compress
;
1750 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1751 boolean_t dedup
= B_FALSE
;
1752 boolean_t nopwrite
= B_FALSE
;
1753 boolean_t dedup_verify
= os
->os_dedup_verify
;
1754 int copies
= os
->os_copies
;
1757 * We maintain different write policies for each of the following
1760 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1761 * 3. all other level 0 blocks
1765 * XXX -- we should design a compression algorithm
1766 * that specializes in arrays of bps.
1768 compress
= zfs_mdcomp_disable
? ZIO_COMPRESS_EMPTY
:
1772 * Metadata always gets checksummed. If the data
1773 * checksum is multi-bit correctable, and it's not a
1774 * ZBT-style checksum, then it's suitable for metadata
1775 * as well. Otherwise, the metadata checksum defaults
1778 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1779 zio_checksum_table
[checksum
].ci_eck
)
1780 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1782 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
1783 (os
->os_redundant_metadata
==
1784 ZFS_REDUNDANT_METADATA_MOST
&&
1785 (level
>= zfs_redundant_metadata_most_ditto_level
||
1786 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
1788 } else if (wp
& WP_NOFILL
) {
1792 * If we're writing preallocated blocks, we aren't actually
1793 * writing them so don't set any policy properties. These
1794 * blocks are currently only used by an external subsystem
1795 * outside of zfs (i.e. dump) and not written by the zio
1798 compress
= ZIO_COMPRESS_OFF
;
1799 checksum
= ZIO_CHECKSUM_OFF
;
1801 compress
= zio_compress_select(dn
->dn_compress
, compress
);
1803 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1804 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1808 * Determine dedup setting. If we are in dmu_sync(),
1809 * we won't actually dedup now because that's all
1810 * done in syncing context; but we do want to use the
1811 * dedup checkum. If the checksum is not strong
1812 * enough to ensure unique signatures, force
1815 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1816 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1817 if (!zio_checksum_table
[checksum
].ci_dedup
)
1818 dedup_verify
= B_TRUE
;
1822 * Enable nopwrite if we have a cryptographically secure
1823 * checksum that has no known collisions (i.e. SHA-256)
1824 * and compression is enabled. We don't enable nopwrite if
1825 * dedup is enabled as the two features are mutually exclusive.
1827 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1828 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1831 zp
->zp_checksum
= checksum
;
1832 zp
->zp_compress
= compress
;
1833 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1834 zp
->zp_level
= level
;
1835 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
1836 zp
->zp_dedup
= dedup
;
1837 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1838 zp
->zp_nopwrite
= nopwrite
;
1842 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1847 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1851 * Sync any current changes before
1852 * we go trundling through the block pointers.
1854 for (i
= 0; i
< TXG_SIZE
; i
++) {
1855 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1858 if (i
!= TXG_SIZE
) {
1859 dnode_rele(dn
, FTAG
);
1860 txg_wait_synced(dmu_objset_pool(os
), 0);
1861 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1866 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1867 dnode_rele(dn
, FTAG
);
1873 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1875 dnode_phys_t
*dnp
= dn
->dn_phys
;
1878 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1879 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1880 1ULL << dn
->dn_indblkshift
: 0;
1881 doi
->doi_type
= dn
->dn_type
;
1882 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1883 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1884 doi
->doi_indirection
= dn
->dn_nlevels
;
1885 doi
->doi_checksum
= dn
->dn_checksum
;
1886 doi
->doi_compress
= dn
->dn_compress
;
1887 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1888 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1889 doi
->doi_fill_count
= 0;
1890 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1891 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
1895 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1897 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1898 mutex_enter(&dn
->dn_mtx
);
1900 __dmu_object_info_from_dnode(dn
, doi
);
1902 mutex_exit(&dn
->dn_mtx
);
1903 rw_exit(&dn
->dn_struct_rwlock
);
1907 * Get information on a DMU object.
1908 * If doi is NULL, just indicates whether the object exists.
1911 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1914 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1920 dmu_object_info_from_dnode(dn
, doi
);
1922 dnode_rele(dn
, FTAG
);
1927 * As above, but faster; can be used when you have a held dbuf in hand.
1930 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1932 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1935 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1940 * Faster still when you only care about the size.
1941 * This is specifically optimized for zfs_getattr().
1944 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1945 u_longlong_t
*nblk512
)
1947 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1953 *blksize
= dn
->dn_datablksz
;
1954 /* add 1 for dnode space */
1955 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1956 SPA_MINBLOCKSHIFT
) + 1;
1961 byteswap_uint64_array(void *vbuf
, size_t size
)
1963 uint64_t *buf
= vbuf
;
1964 size_t count
= size
>> 3;
1967 ASSERT((size
& 7) == 0);
1969 for (i
= 0; i
< count
; i
++)
1970 buf
[i
] = BSWAP_64(buf
[i
]);
1974 byteswap_uint32_array(void *vbuf
, size_t size
)
1976 uint32_t *buf
= vbuf
;
1977 size_t count
= size
>> 2;
1980 ASSERT((size
& 3) == 0);
1982 for (i
= 0; i
< count
; i
++)
1983 buf
[i
] = BSWAP_32(buf
[i
]);
1987 byteswap_uint16_array(void *vbuf
, size_t size
)
1989 uint16_t *buf
= vbuf
;
1990 size_t count
= size
>> 1;
1993 ASSERT((size
& 1) == 0);
1995 for (i
= 0; i
< count
; i
++)
1996 buf
[i
] = BSWAP_16(buf
[i
]);
2001 byteswap_uint8_array(void *vbuf
, size_t size
)
2023 arc_fini(); /* arc depends on l2arc, so arc must go first */
2035 #if defined(_KERNEL) && defined(HAVE_SPL)
2036 EXPORT_SYMBOL(dmu_bonus_hold
);
2037 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2038 EXPORT_SYMBOL(dmu_buf_rele_array
);
2039 EXPORT_SYMBOL(dmu_prefetch
);
2040 EXPORT_SYMBOL(dmu_free_range
);
2041 EXPORT_SYMBOL(dmu_free_long_range
);
2042 EXPORT_SYMBOL(dmu_free_long_object
);
2043 EXPORT_SYMBOL(dmu_read
);
2044 EXPORT_SYMBOL(dmu_write
);
2045 EXPORT_SYMBOL(dmu_prealloc
);
2046 EXPORT_SYMBOL(dmu_object_info
);
2047 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2048 EXPORT_SYMBOL(dmu_object_info_from_db
);
2049 EXPORT_SYMBOL(dmu_object_size_from_db
);
2050 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2051 EXPORT_SYMBOL(dmu_object_set_checksum
);
2052 EXPORT_SYMBOL(dmu_object_set_compress
);
2053 EXPORT_SYMBOL(dmu_write_policy
);
2054 EXPORT_SYMBOL(dmu_sync
);
2055 EXPORT_SYMBOL(dmu_request_arcbuf
);
2056 EXPORT_SYMBOL(dmu_return_arcbuf
);
2057 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2058 EXPORT_SYMBOL(dmu_buf_hold
);
2059 EXPORT_SYMBOL(dmu_ot
);
2061 module_param(zfs_mdcomp_disable
, int, 0644);
2062 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2064 module_param(zfs_nopwrite_enabled
, int, 0644);
2065 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");