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.
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
32 #include <sys/dnode.h>
33 #include <sys/zfs_context.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dmu_traverse.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dsl_pool.h>
39 #include <sys/dsl_synctask.h>
40 #include <sys/dsl_prop.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/zfs_ioctl.h>
44 #include <sys/zio_checksum.h>
45 #include <sys/zio_compress.h>
47 #include <sys/zfeature.h>
49 #include <sys/vmsystm.h>
50 #include <sys/zfs_znode.h>
54 * Enable/disable nopwrite feature.
56 int zfs_nopwrite_enabled
= 1;
58 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
59 { DMU_BSWAP_UINT8
, TRUE
, "unallocated" },
60 { DMU_BSWAP_ZAP
, TRUE
, "object directory" },
61 { DMU_BSWAP_UINT64
, TRUE
, "object array" },
62 { DMU_BSWAP_UINT8
, TRUE
, "packed nvlist" },
63 { DMU_BSWAP_UINT64
, TRUE
, "packed nvlist size" },
64 { DMU_BSWAP_UINT64
, TRUE
, "bpobj" },
65 { DMU_BSWAP_UINT64
, TRUE
, "bpobj header" },
66 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map header" },
67 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map" },
68 { DMU_BSWAP_UINT64
, TRUE
, "ZIL intent log" },
69 { DMU_BSWAP_DNODE
, TRUE
, "DMU dnode" },
70 { DMU_BSWAP_OBJSET
, TRUE
, "DMU objset" },
71 { DMU_BSWAP_UINT64
, TRUE
, "DSL directory" },
72 { DMU_BSWAP_ZAP
, TRUE
, "DSL directory child map"},
73 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset snap map" },
74 { DMU_BSWAP_ZAP
, TRUE
, "DSL props" },
75 { DMU_BSWAP_UINT64
, TRUE
, "DSL dataset" },
76 { DMU_BSWAP_ZNODE
, TRUE
, "ZFS znode" },
77 { DMU_BSWAP_OLDACL
, TRUE
, "ZFS V0 ACL" },
78 { DMU_BSWAP_UINT8
, FALSE
, "ZFS plain file" },
79 { DMU_BSWAP_ZAP
, TRUE
, "ZFS directory" },
80 { DMU_BSWAP_ZAP
, TRUE
, "ZFS master node" },
81 { DMU_BSWAP_ZAP
, TRUE
, "ZFS delete queue" },
82 { DMU_BSWAP_UINT8
, FALSE
, "zvol object" },
83 { DMU_BSWAP_ZAP
, TRUE
, "zvol prop" },
84 { DMU_BSWAP_UINT8
, FALSE
, "other uint8[]" },
85 { DMU_BSWAP_UINT64
, FALSE
, "other uint64[]" },
86 { DMU_BSWAP_ZAP
, TRUE
, "other ZAP" },
87 { DMU_BSWAP_ZAP
, TRUE
, "persistent error log" },
88 { DMU_BSWAP_UINT8
, TRUE
, "SPA history" },
89 { DMU_BSWAP_UINT64
, TRUE
, "SPA history offsets" },
90 { DMU_BSWAP_ZAP
, TRUE
, "Pool properties" },
91 { DMU_BSWAP_ZAP
, TRUE
, "DSL permissions" },
92 { DMU_BSWAP_ACL
, TRUE
, "ZFS ACL" },
93 { DMU_BSWAP_UINT8
, TRUE
, "ZFS SYSACL" },
94 { DMU_BSWAP_UINT8
, TRUE
, "FUID table" },
95 { DMU_BSWAP_UINT64
, TRUE
, "FUID table size" },
96 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset next clones"},
97 { DMU_BSWAP_ZAP
, TRUE
, "scan work queue" },
98 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group used" },
99 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group quota" },
100 { DMU_BSWAP_ZAP
, TRUE
, "snapshot refcount tags"},
101 { DMU_BSWAP_ZAP
, TRUE
, "DDT ZAP algorithm" },
102 { DMU_BSWAP_ZAP
, TRUE
, "DDT statistics" },
103 { DMU_BSWAP_UINT8
, TRUE
, "System attributes" },
104 { DMU_BSWAP_ZAP
, TRUE
, "SA master node" },
105 { DMU_BSWAP_ZAP
, TRUE
, "SA attr registration" },
106 { DMU_BSWAP_ZAP
, TRUE
, "SA attr layouts" },
107 { DMU_BSWAP_ZAP
, TRUE
, "scan translations" },
108 { DMU_BSWAP_UINT8
, FALSE
, "deduplicated block" },
109 { DMU_BSWAP_ZAP
, TRUE
, "DSL deadlist map" },
110 { DMU_BSWAP_UINT64
, TRUE
, "DSL deadlist map hdr" },
111 { DMU_BSWAP_ZAP
, TRUE
, "DSL dir clones" },
112 { DMU_BSWAP_UINT64
, TRUE
, "bpobj subobj" }
115 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
116 { byteswap_uint8_array
, "uint8" },
117 { byteswap_uint16_array
, "uint16" },
118 { byteswap_uint32_array
, "uint32" },
119 { byteswap_uint64_array
, "uint64" },
120 { zap_byteswap
, "zap" },
121 { dnode_buf_byteswap
, "dnode" },
122 { dmu_objset_byteswap
, "objset" },
123 { zfs_znode_byteswap
, "znode" },
124 { zfs_oldacl_byteswap
, "oldacl" },
125 { zfs_acl_byteswap
, "acl" }
129 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
130 void *tag
, dmu_buf_t
**dbp
)
137 err
= dnode_hold(os
, object
, FTAG
, &dn
);
140 blkid
= dbuf_whichblock(dn
, offset
);
141 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
142 db
= dbuf_hold(dn
, blkid
, tag
);
143 rw_exit(&dn
->dn_struct_rwlock
);
144 dnode_rele(dn
, FTAG
);
148 return (SET_ERROR(EIO
));
156 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
157 void *tag
, dmu_buf_t
**dbp
, int flags
)
160 int db_flags
= DB_RF_CANFAIL
;
162 if (flags
& DMU_READ_NO_PREFETCH
)
163 db_flags
|= DB_RF_NOPREFETCH
;
165 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
167 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
168 err
= dbuf_read(db
, NULL
, db_flags
);
181 return (DN_MAX_BONUSLEN
);
185 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
187 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
194 if (dn
->dn_bonus
!= db
) {
195 error
= SET_ERROR(EINVAL
);
196 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
197 error
= SET_ERROR(EINVAL
);
199 dnode_setbonuslen(dn
, newsize
, tx
);
208 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
210 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
217 if (!DMU_OT_IS_VALID(type
)) {
218 error
= SET_ERROR(EINVAL
);
219 } else if (dn
->dn_bonus
!= db
) {
220 error
= SET_ERROR(EINVAL
);
222 dnode_setbonus_type(dn
, type
, tx
);
231 dmu_get_bonustype(dmu_buf_t
*db_fake
)
233 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
235 dmu_object_type_t type
;
239 type
= dn
->dn_bonustype
;
246 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
251 error
= dnode_hold(os
, object
, FTAG
, &dn
);
252 dbuf_rm_spill(dn
, tx
);
253 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
254 dnode_rm_spill(dn
, tx
);
255 rw_exit(&dn
->dn_struct_rwlock
);
256 dnode_rele(dn
, FTAG
);
261 * returns ENOENT, EIO, or 0.
264 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
270 error
= dnode_hold(os
, object
, FTAG
, &dn
);
274 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
275 if (dn
->dn_bonus
== NULL
) {
276 rw_exit(&dn
->dn_struct_rwlock
);
277 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
278 if (dn
->dn_bonus
== NULL
)
279 dbuf_create_bonus(dn
);
283 /* as long as the bonus buf is held, the dnode will be held */
284 if (refcount_add(&db
->db_holds
, tag
) == 1) {
285 VERIFY(dnode_add_ref(dn
, db
));
286 atomic_inc_32(&dn
->dn_dbufs_count
);
290 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
291 * hold and incrementing the dbuf count to ensure that dnode_move() sees
292 * a dnode hold for every dbuf.
294 rw_exit(&dn
->dn_struct_rwlock
);
296 dnode_rele(dn
, FTAG
);
298 VERIFY(0 == dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
));
305 * returns ENOENT, EIO, or 0.
307 * This interface will allocate a blank spill dbuf when a spill blk
308 * doesn't already exist on the dnode.
310 * if you only want to find an already existing spill db, then
311 * dmu_spill_hold_existing() should be used.
314 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
316 dmu_buf_impl_t
*db
= NULL
;
319 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
320 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
322 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
324 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
325 rw_exit(&dn
->dn_struct_rwlock
);
328 err
= dbuf_read(db
, NULL
, flags
);
337 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
339 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
346 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
347 err
= SET_ERROR(EINVAL
);
349 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
351 if (!dn
->dn_have_spill
) {
352 err
= SET_ERROR(ENOENT
);
354 err
= dmu_spill_hold_by_dnode(dn
,
355 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
358 rw_exit(&dn
->dn_struct_rwlock
);
366 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
368 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
374 err
= dmu_spill_hold_by_dnode(dn
, DB_RF_CANFAIL
, tag
, dbp
);
381 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
382 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
383 * and can induce severe lock contention when writing to several files
384 * whose dnodes are in the same block.
387 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
388 int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
391 uint64_t blkid
, nblks
, i
;
396 ASSERT(length
<= DMU_MAX_ACCESS
);
398 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
;
399 if (flags
& DMU_READ_NO_PREFETCH
|| length
> zfetch_array_rd_sz
)
400 dbuf_flags
|= DB_RF_NOPREFETCH
;
402 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
403 if (dn
->dn_datablkshift
) {
404 int blkshift
= dn
->dn_datablkshift
;
405 nblks
= (P2ROUNDUP(offset
+length
, 1ULL<<blkshift
) -
406 P2ALIGN(offset
, 1ULL<<blkshift
)) >> blkshift
;
408 if (offset
+ length
> dn
->dn_datablksz
) {
409 zfs_panic_recover("zfs: accessing past end of object "
410 "%llx/%llx (size=%u access=%llu+%llu)",
411 (longlong_t
)dn
->dn_objset
->
412 os_dsl_dataset
->ds_object
,
413 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
414 (longlong_t
)offset
, (longlong_t
)length
);
415 rw_exit(&dn
->dn_struct_rwlock
);
416 return (SET_ERROR(EIO
));
420 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_SLEEP
);
422 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
423 blkid
= dbuf_whichblock(dn
, offset
);
424 for (i
= 0; i
< nblks
; i
++) {
425 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+i
, tag
);
427 rw_exit(&dn
->dn_struct_rwlock
);
428 dmu_buf_rele_array(dbp
, nblks
, tag
);
430 return (SET_ERROR(EIO
));
432 /* initiate async i/o */
434 (void) dbuf_read(db
, zio
, dbuf_flags
);
438 rw_exit(&dn
->dn_struct_rwlock
);
440 /* wait for async i/o */
443 dmu_buf_rele_array(dbp
, nblks
, tag
);
447 /* wait for other io to complete */
449 for (i
= 0; i
< nblks
; i
++) {
450 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
451 mutex_enter(&db
->db_mtx
);
452 while (db
->db_state
== DB_READ
||
453 db
->db_state
== DB_FILL
)
454 cv_wait(&db
->db_changed
, &db
->db_mtx
);
455 if (db
->db_state
== DB_UNCACHED
)
456 err
= SET_ERROR(EIO
);
457 mutex_exit(&db
->db_mtx
);
459 dmu_buf_rele_array(dbp
, nblks
, tag
);
471 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
472 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
477 err
= dnode_hold(os
, object
, FTAG
, &dn
);
481 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
482 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
484 dnode_rele(dn
, FTAG
);
490 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
491 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
493 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
499 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
500 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
507 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
510 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
515 for (i
= 0; i
< numbufs
; i
++) {
517 dbuf_rele(dbp
[i
], tag
);
520 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
524 * Issue prefetch i/os for the given blocks.
526 * Note: The assumption is that we *know* these blocks will be needed
527 * almost immediately. Therefore, the prefetch i/os will be issued at
528 * ZIO_PRIORITY_SYNC_READ
530 * Note: indirect blocks and other metadata will be read synchronously,
531 * causing this function to block if they are not already cached.
534 dmu_prefetch(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t len
)
540 if (zfs_prefetch_disable
)
543 if (len
== 0) { /* they're interested in the bonus buffer */
544 dn
= DMU_META_DNODE(os
);
546 if (object
== 0 || object
>= DN_MAX_OBJECT
)
549 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
550 blkid
= dbuf_whichblock(dn
, object
* sizeof (dnode_phys_t
));
551 dbuf_prefetch(dn
, blkid
, ZIO_PRIORITY_SYNC_READ
);
552 rw_exit(&dn
->dn_struct_rwlock
);
557 * XXX - Note, if the dnode for the requested object is not
558 * already cached, we will do a *synchronous* read in the
559 * dnode_hold() call. The same is true for any indirects.
561 err
= dnode_hold(os
, object
, FTAG
, &dn
);
565 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
566 if (dn
->dn_datablkshift
) {
567 int blkshift
= dn
->dn_datablkshift
;
568 nblks
= (P2ROUNDUP(offset
+ len
, 1 << blkshift
) -
569 P2ALIGN(offset
, 1 << blkshift
)) >> blkshift
;
571 nblks
= (offset
< dn
->dn_datablksz
);
577 blkid
= dbuf_whichblock(dn
, offset
);
578 for (i
= 0; i
< nblks
; i
++)
579 dbuf_prefetch(dn
, blkid
+ i
, ZIO_PRIORITY_SYNC_READ
);
582 rw_exit(&dn
->dn_struct_rwlock
);
584 dnode_rele(dn
, FTAG
);
588 * Get the next "chunk" of file data to free. We traverse the file from
589 * the end so that the file gets shorter over time (if we crashes in the
590 * middle, this will leave us in a better state). We find allocated file
591 * data by simply searching the allocated level 1 indirects.
593 * On input, *start should be the first offset that does not need to be
594 * freed (e.g. "offset + length"). On return, *start will be the first
595 * offset that should be freed.
598 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
600 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
601 /* bytes of data covered by a level-1 indirect block */
603 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
606 ASSERT3U(minimum
, <=, *start
);
608 if (*start
- minimum
<= iblkrange
* maxblks
) {
612 ASSERT(ISP2(iblkrange
));
614 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
618 * dnode_next_offset(BACKWARDS) will find an allocated L1
619 * indirect block at or before the input offset. We must
620 * decrement *start so that it is at the end of the region
624 err
= dnode_next_offset(dn
,
625 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
627 /* if there are no indirect blocks before start, we are done */
631 } else if (err
!= 0) {
635 /* set start to the beginning of this L1 indirect */
636 *start
= P2ALIGN(*start
, iblkrange
);
638 if (*start
< minimum
)
644 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
647 uint64_t object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
650 if (offset
>= object_size
)
653 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
654 length
= object_size
- offset
;
656 while (length
!= 0) {
657 uint64_t chunk_end
, chunk_begin
;
660 chunk_end
= chunk_begin
= offset
+ length
;
662 /* move chunk_begin backwards to the beginning of this chunk */
663 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
666 ASSERT3U(chunk_begin
, >=, offset
);
667 ASSERT3U(chunk_begin
, <=, chunk_end
);
669 tx
= dmu_tx_create(os
);
670 dmu_tx_hold_free(tx
, dn
->dn_object
,
671 chunk_begin
, chunk_end
- chunk_begin
);
672 err
= dmu_tx_assign(tx
, TXG_WAIT
);
677 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
680 length
-= chunk_end
- chunk_begin
;
686 dmu_free_long_range(objset_t
*os
, uint64_t object
,
687 uint64_t offset
, uint64_t length
)
692 err
= dnode_hold(os
, object
, FTAG
, &dn
);
695 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
698 * It is important to zero out the maxblkid when freeing the entire
699 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
700 * will take the fast path, and (b) dnode_reallocate() can verify
701 * that the entire file has been freed.
703 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
706 dnode_rele(dn
, FTAG
);
711 dmu_free_long_object(objset_t
*os
, uint64_t object
)
716 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
720 tx
= dmu_tx_create(os
);
721 dmu_tx_hold_bonus(tx
, object
);
722 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
723 err
= dmu_tx_assign(tx
, TXG_WAIT
);
725 err
= dmu_object_free(os
, object
, tx
);
735 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
736 uint64_t size
, dmu_tx_t
*tx
)
739 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
742 ASSERT(offset
< UINT64_MAX
);
743 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
744 dnode_free_range(dn
, offset
, size
, tx
);
745 dnode_rele(dn
, FTAG
);
750 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
751 void *buf
, uint32_t flags
)
757 err
= dnode_hold(os
, object
, FTAG
, &dn
);
762 * Deal with odd block sizes, where there can't be data past the first
763 * block. If we ever do the tail block optimization, we will need to
764 * handle that here as well.
766 if (dn
->dn_maxblkid
== 0) {
767 uint64_t newsz
= offset
> dn
->dn_datablksz
? 0 :
768 MIN(size
, dn
->dn_datablksz
- offset
);
769 bzero((char *)buf
+ newsz
, size
- newsz
);
774 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
778 * NB: we could do this block-at-a-time, but it's nice
779 * to be reading in parallel.
781 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
782 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
786 for (i
= 0; i
< numbufs
; i
++) {
789 dmu_buf_t
*db
= dbp
[i
];
793 bufoff
= offset
- db
->db_offset
;
794 tocpy
= MIN(db
->db_size
- bufoff
, size
);
796 (void) memcpy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
800 buf
= (char *)buf
+ tocpy
;
802 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
804 dnode_rele(dn
, FTAG
);
809 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
810 const void *buf
, dmu_tx_t
*tx
)
818 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
819 FALSE
, FTAG
, &numbufs
, &dbp
));
821 for (i
= 0; i
< numbufs
; i
++) {
824 dmu_buf_t
*db
= dbp
[i
];
828 bufoff
= offset
- db
->db_offset
;
829 tocpy
= MIN(db
->db_size
- bufoff
, size
);
831 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
833 if (tocpy
== db
->db_size
)
834 dmu_buf_will_fill(db
, tx
);
836 dmu_buf_will_dirty(db
, tx
);
838 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
840 if (tocpy
== db
->db_size
)
841 dmu_buf_fill_done(db
, tx
);
845 buf
= (char *)buf
+ tocpy
;
847 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
851 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
860 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
861 FALSE
, FTAG
, &numbufs
, &dbp
));
863 for (i
= 0; i
< numbufs
; i
++) {
864 dmu_buf_t
*db
= dbp
[i
];
866 dmu_buf_will_not_fill(db
, tx
);
868 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
872 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
873 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
874 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
878 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
879 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
880 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
883 dmu_buf_write_embedded(db
,
884 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
885 uncompressed_size
, compressed_size
, byteorder
, tx
);
887 dmu_buf_rele(db
, FTAG
);
891 * DMU support for xuio
893 kstat_t
*xuio_ksp
= NULL
;
895 typedef struct xuio_stats
{
896 /* loaned yet not returned arc_buf */
897 kstat_named_t xuiostat_onloan_rbuf
;
898 kstat_named_t xuiostat_onloan_wbuf
;
899 /* whether a copy is made when loaning out a read buffer */
900 kstat_named_t xuiostat_rbuf_copied
;
901 kstat_named_t xuiostat_rbuf_nocopy
;
902 /* whether a copy is made when assigning a write buffer */
903 kstat_named_t xuiostat_wbuf_copied
;
904 kstat_named_t xuiostat_wbuf_nocopy
;
907 static xuio_stats_t xuio_stats
= {
908 { "onloan_read_buf", KSTAT_DATA_UINT64
},
909 { "onloan_write_buf", KSTAT_DATA_UINT64
},
910 { "read_buf_copied", KSTAT_DATA_UINT64
},
911 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
912 { "write_buf_copied", KSTAT_DATA_UINT64
},
913 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
916 #define XUIOSTAT_INCR(stat, val) \
917 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
918 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
921 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
924 uio_t
*uio
= &xuio
->xu_uio
;
926 uio
->uio_iovcnt
= nblk
;
927 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
929 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
931 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
932 priv
->iovp
= uio
->uio_iov
;
933 XUIO_XUZC_PRIV(xuio
) = priv
;
935 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
936 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
938 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
944 dmu_xuio_fini(xuio_t
*xuio
)
946 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
947 int nblk
= priv
->cnt
;
949 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
950 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
951 kmem_free(priv
, sizeof (dmu_xuio_t
));
953 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
954 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
956 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
960 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
961 * and increase priv->next by 1.
964 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
967 uio_t
*uio
= &xuio
->xu_uio
;
968 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
969 int i
= priv
->next
++;
971 ASSERT(i
< priv
->cnt
);
972 ASSERT(off
+ n
<= arc_buf_size(abuf
));
973 iov
= uio
->uio_iov
+ i
;
974 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
976 priv
->bufs
[i
] = abuf
;
981 dmu_xuio_cnt(xuio_t
*xuio
)
983 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
988 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
990 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
992 ASSERT(i
< priv
->cnt
);
993 return (priv
->bufs
[i
]);
997 dmu_xuio_clear(xuio_t
*xuio
, int i
)
999 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1001 ASSERT(i
< priv
->cnt
);
1002 priv
->bufs
[i
] = NULL
;
1006 xuio_stat_init(void)
1008 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1009 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1010 KSTAT_FLAG_VIRTUAL
);
1011 if (xuio_ksp
!= NULL
) {
1012 xuio_ksp
->ks_data
= &xuio_stats
;
1013 kstat_install(xuio_ksp
);
1018 xuio_stat_fini(void)
1020 if (xuio_ksp
!= NULL
) {
1021 kstat_delete(xuio_ksp
);
1027 xuio_stat_wbuf_copied()
1029 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1033 xuio_stat_wbuf_nocopy()
1035 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1041 * Copy up to size bytes between arg_buf and req based on the data direction
1042 * described by the req. If an entire req's data cannot be transfered in one
1043 * pass, you should pass in @req_offset to indicate where to continue. The
1044 * return value is the number of bytes successfully copied to arg_buf.
1047 dmu_req_copy(void *arg_buf
, int size
, struct request
*req
, size_t req_offset
)
1049 struct bio_vec bv
, *bvp
;
1050 struct req_iterator iter
;
1052 int tocpy
, bv_len
, bv_offset
;
1055 rq_for_each_segment4(bv
, bvp
, req
, iter
) {
1057 * Fully consumed the passed arg_buf. We use goto here because
1058 * rq_for_each_segment is a double loop
1060 ASSERT3S(offset
, <=, size
);
1064 /* Skip already copied bv */
1065 if (req_offset
>= bv
.bv_len
) {
1066 req_offset
-= bv
.bv_len
;
1070 bv_len
= bv
.bv_len
- req_offset
;
1071 bv_offset
= bv
.bv_offset
+ req_offset
;
1074 tocpy
= MIN(bv_len
, size
- offset
);
1075 ASSERT3S(tocpy
, >=, 0);
1077 bv_buf
= page_address(bv
.bv_page
) + bv_offset
;
1078 ASSERT3P(bv_buf
, !=, NULL
);
1080 if (rq_data_dir(req
) == WRITE
)
1081 memcpy(arg_buf
+ offset
, bv_buf
, tocpy
);
1083 memcpy(bv_buf
, arg_buf
+ offset
, tocpy
);
1092 dmu_read_req(objset_t
*os
, uint64_t object
, struct request
*req
)
1094 uint64_t size
= blk_rq_bytes(req
);
1095 uint64_t offset
= blk_rq_pos(req
) << 9;
1097 int numbufs
, i
, err
;
1101 * NB: we could do this block-at-a-time, but it's nice
1102 * to be reading in parallel.
1104 err
= dmu_buf_hold_array(os
, object
, offset
, size
, TRUE
, FTAG
,
1110 for (i
= 0; i
< numbufs
; i
++) {
1114 dmu_buf_t
*db
= dbp
[i
];
1116 bufoff
= offset
- db
->db_offset
;
1117 ASSERT3S(bufoff
, >=, 0);
1119 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1123 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1134 req_offset
+= didcpy
;
1137 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1143 dmu_write_req(objset_t
*os
, uint64_t object
, struct request
*req
, dmu_tx_t
*tx
)
1145 uint64_t size
= blk_rq_bytes(req
);
1146 uint64_t offset
= blk_rq_pos(req
) << 9;
1148 int numbufs
, i
, err
;
1154 err
= dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1160 for (i
= 0; i
< numbufs
; i
++) {
1164 dmu_buf_t
*db
= dbp
[i
];
1166 bufoff
= offset
- db
->db_offset
;
1167 ASSERT3S(bufoff
, >=, 0);
1169 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1173 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1175 if (tocpy
== db
->db_size
)
1176 dmu_buf_will_fill(db
, tx
);
1178 dmu_buf_will_dirty(db
, tx
);
1180 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1183 if (tocpy
== db
->db_size
)
1184 dmu_buf_fill_done(db
, tx
);
1194 req_offset
+= didcpy
;
1198 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1203 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1206 int numbufs
, i
, err
;
1207 xuio_t
*xuio
= NULL
;
1210 * NB: we could do this block-at-a-time, but it's nice
1211 * to be reading in parallel.
1213 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1214 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1218 for (i
= 0; i
< numbufs
; i
++) {
1221 dmu_buf_t
*db
= dbp
[i
];
1225 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1226 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1229 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1230 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1231 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1232 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1234 uio
->uio_resid
-= tocpy
;
1235 uio
->uio_loffset
+= tocpy
;
1238 if (abuf
== dbuf_abuf
)
1239 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1241 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1243 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1251 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1257 * Read 'size' bytes into the uio buffer.
1258 * From object zdb->db_object.
1259 * Starting at offset uio->uio_loffset.
1261 * If the caller already has a dbuf in the target object
1262 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1263 * because we don't have to find the dnode_t for the object.
1266 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1268 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1277 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1284 * Read 'size' bytes into the uio buffer.
1285 * From the specified object
1286 * Starting at offset uio->uio_loffset.
1289 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1297 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1301 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1303 dnode_rele(dn
, FTAG
);
1309 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1316 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1317 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1321 for (i
= 0; i
< numbufs
; i
++) {
1324 dmu_buf_t
*db
= dbp
[i
];
1328 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1329 tocpy
= MIN(db
->db_size
- bufoff
, size
);
1331 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1333 if (tocpy
== db
->db_size
)
1334 dmu_buf_will_fill(db
, tx
);
1336 dmu_buf_will_dirty(db
, tx
);
1339 * XXX uiomove could block forever (eg.nfs-backed
1340 * pages). There needs to be a uiolockdown() function
1341 * to lock the pages in memory, so that uiomove won't
1344 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1347 if (tocpy
== db
->db_size
)
1348 dmu_buf_fill_done(db
, tx
);
1356 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1361 * Write 'size' bytes from the uio buffer.
1362 * To object zdb->db_object.
1363 * Starting at offset uio->uio_loffset.
1365 * If the caller already has a dbuf in the target object
1366 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1367 * because we don't have to find the dnode_t for the object.
1370 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1373 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1382 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1389 * Write 'size' bytes from the uio buffer.
1390 * To the specified object.
1391 * Starting at offset uio->uio_loffset.
1394 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1403 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1407 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1409 dnode_rele(dn
, FTAG
);
1413 #endif /* _KERNEL */
1416 * Allocate a loaned anonymous arc buffer.
1419 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1421 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1423 return (arc_loan_buf(db
->db_objset
->os_spa
, size
));
1427 * Free a loaned arc buffer.
1430 dmu_return_arcbuf(arc_buf_t
*buf
)
1432 arc_return_buf(buf
, FTAG
);
1433 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1437 * When possible directly assign passed loaned arc buffer to a dbuf.
1438 * If this is not possible copy the contents of passed arc buf via
1442 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1445 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1448 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1451 DB_DNODE_ENTER(dbuf
);
1452 dn
= DB_DNODE(dbuf
);
1453 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1454 blkid
= dbuf_whichblock(dn
, offset
);
1455 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1456 rw_exit(&dn
->dn_struct_rwlock
);
1457 DB_DNODE_EXIT(dbuf
);
1460 * We can only assign if the offset is aligned, the arc buf is the
1461 * same size as the dbuf, and the dbuf is not metadata. It
1462 * can't be metadata because the loaned arc buf comes from the
1463 * user-data kmem area.
1465 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
&&
1466 DBUF_GET_BUFC_TYPE(db
) == ARC_BUFC_DATA
) {
1467 dbuf_assign_arcbuf(db
, buf
, tx
);
1468 dbuf_rele(db
, FTAG
);
1473 DB_DNODE_ENTER(dbuf
);
1474 dn
= DB_DNODE(dbuf
);
1476 object
= dn
->dn_object
;
1477 DB_DNODE_EXIT(dbuf
);
1479 dbuf_rele(db
, FTAG
);
1480 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1481 dmu_return_arcbuf(buf
);
1482 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1487 dbuf_dirty_record_t
*dsa_dr
;
1488 dmu_sync_cb_t
*dsa_done
;
1495 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1497 dmu_sync_arg_t
*dsa
= varg
;
1498 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1499 blkptr_t
*bp
= zio
->io_bp
;
1501 if (zio
->io_error
== 0) {
1502 if (BP_IS_HOLE(bp
)) {
1504 * A block of zeros may compress to a hole, but the
1505 * block size still needs to be known for replay.
1507 BP_SET_LSIZE(bp
, db
->db_size
);
1508 } else if (!BP_IS_EMBEDDED(bp
)) {
1509 ASSERT(BP_GET_LEVEL(bp
) == 0);
1516 dmu_sync_late_arrival_ready(zio_t
*zio
)
1518 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1523 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1525 dmu_sync_arg_t
*dsa
= varg
;
1526 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1527 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1529 mutex_enter(&db
->db_mtx
);
1530 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1531 if (zio
->io_error
== 0) {
1532 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1533 if (dr
->dt
.dl
.dr_nopwrite
) {
1534 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1535 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1536 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1538 ASSERT(BP_EQUAL(bp
, bp_orig
));
1539 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1540 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1542 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1543 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1544 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1547 * Old style holes are filled with all zeros, whereas
1548 * new-style holes maintain their lsize, type, level,
1549 * and birth time (see zio_write_compress). While we
1550 * need to reset the BP_SET_LSIZE() call that happened
1551 * in dmu_sync_ready for old style holes, we do *not*
1552 * want to wipe out the information contained in new
1553 * style holes. Thus, only zero out the block pointer if
1554 * it's an old style hole.
1556 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1557 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1558 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1560 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1562 cv_broadcast(&db
->db_changed
);
1563 mutex_exit(&db
->db_mtx
);
1565 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1567 kmem_free(dsa
, sizeof (*dsa
));
1571 dmu_sync_late_arrival_done(zio_t
*zio
)
1573 blkptr_t
*bp
= zio
->io_bp
;
1574 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1575 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1577 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1579 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1580 * then there is nothing to do here. Otherwise, free the
1581 * newly allocated block in this txg.
1583 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1584 ASSERT(BP_EQUAL(bp
, bp_orig
));
1586 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1587 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1588 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1589 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1593 dmu_tx_commit(dsa
->dsa_tx
);
1595 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1597 kmem_free(dsa
, sizeof (*dsa
));
1601 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1602 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1604 dmu_sync_arg_t
*dsa
;
1607 tx
= dmu_tx_create(os
);
1608 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1609 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1611 /* Make zl_get_data do txg_waited_synced() */
1612 return (SET_ERROR(EIO
));
1615 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1617 dsa
->dsa_done
= done
;
1621 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1622 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1623 dmu_sync_late_arrival_ready
, NULL
, dmu_sync_late_arrival_done
, dsa
,
1624 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
|ZIO_FLAG_FASTWRITE
, zb
));
1630 * Intent log support: sync the block associated with db to disk.
1631 * N.B. and XXX: the caller is responsible for making sure that the
1632 * data isn't changing while dmu_sync() is writing it.
1636 * EEXIST: this txg has already been synced, so there's nothing to do.
1637 * The caller should not log the write.
1639 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1640 * The caller should not log the write.
1642 * EALREADY: this block is already in the process of being synced.
1643 * The caller should track its progress (somehow).
1645 * EIO: could not do the I/O.
1646 * The caller should do a txg_wait_synced().
1648 * 0: the I/O has been initiated.
1649 * The caller should log this blkptr in the done callback.
1650 * It is possible that the I/O will fail, in which case
1651 * the error will be reported to the done callback and
1652 * propagated to pio from zio_done().
1655 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1657 blkptr_t
*bp
= zgd
->zgd_bp
;
1658 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1659 objset_t
*os
= db
->db_objset
;
1660 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1661 dbuf_dirty_record_t
*dr
;
1662 dmu_sync_arg_t
*dsa
;
1663 zbookmark_phys_t zb
;
1667 ASSERT(pio
!= NULL
);
1670 SET_BOOKMARK(&zb
, ds
->ds_object
,
1671 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1675 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1679 * If we're frozen (running ziltest), we always need to generate a bp.
1681 if (txg
> spa_freeze_txg(os
->os_spa
))
1682 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1685 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1686 * and us. If we determine that this txg is not yet syncing,
1687 * but it begins to sync a moment later, that's OK because the
1688 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1690 mutex_enter(&db
->db_mtx
);
1692 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1694 * This txg has already synced. There's nothing to do.
1696 mutex_exit(&db
->db_mtx
);
1697 return (SET_ERROR(EEXIST
));
1700 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1702 * This txg is currently syncing, so we can't mess with
1703 * the dirty record anymore; just write a new log block.
1705 mutex_exit(&db
->db_mtx
);
1706 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1709 dr
= db
->db_last_dirty
;
1710 while (dr
&& dr
->dr_txg
!= txg
)
1715 * There's no dr for this dbuf, so it must have been freed.
1716 * There's no need to log writes to freed blocks, so we're done.
1718 mutex_exit(&db
->db_mtx
);
1719 return (SET_ERROR(ENOENT
));
1722 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1725 * Assume the on-disk data is X, the current syncing data (in
1726 * txg - 1) is Y, and the current in-memory data is Z (currently
1729 * We usually want to perform a nopwrite if X and Z are the
1730 * same. However, if Y is different (i.e. the BP is going to
1731 * change before this write takes effect), then a nopwrite will
1732 * be incorrect - we would override with X, which could have
1733 * been freed when Y was written.
1735 * (Note that this is not a concern when we are nop-writing from
1736 * syncing context, because X and Y must be identical, because
1737 * all previous txgs have been synced.)
1739 * Therefore, we disable nopwrite if the current BP could change
1740 * before this TXG. There are two ways it could change: by
1741 * being dirty (dr_next is non-NULL), or by being freed
1742 * (dnode_block_freed()). This behavior is verified by
1743 * zio_done(), which VERIFYs that the override BP is identical
1744 * to the on-disk BP.
1748 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
1749 zp
.zp_nopwrite
= B_FALSE
;
1752 ASSERT(dr
->dr_txg
== txg
);
1753 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1754 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1756 * We have already issued a sync write for this buffer,
1757 * or this buffer has already been synced. It could not
1758 * have been dirtied since, or we would have cleared the state.
1760 mutex_exit(&db
->db_mtx
);
1761 return (SET_ERROR(EALREADY
));
1764 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1765 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1766 mutex_exit(&db
->db_mtx
);
1768 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1770 dsa
->dsa_done
= done
;
1774 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1775 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1776 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1777 NULL
, dmu_sync_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1778 ZIO_FLAG_CANFAIL
, &zb
));
1784 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1790 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1793 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1794 dnode_rele(dn
, FTAG
);
1799 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1805 * Send streams include each object's checksum function. This
1806 * check ensures that the receiving system can understand the
1807 * checksum function transmitted.
1809 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1811 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1812 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1813 dn
->dn_checksum
= checksum
;
1814 dnode_setdirty(dn
, tx
);
1815 dnode_rele(dn
, FTAG
);
1819 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1825 * Send streams include each object's compression function. This
1826 * check ensures that the receiving system can understand the
1827 * compression function transmitted.
1829 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1831 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1832 dn
->dn_compress
= compress
;
1833 dnode_setdirty(dn
, tx
);
1834 dnode_rele(dn
, FTAG
);
1837 int zfs_mdcomp_disable
= 0;
1840 * When the "redundant_metadata" property is set to "most", only indirect
1841 * blocks of this level and higher will have an additional ditto block.
1843 int zfs_redundant_metadata_most_ditto_level
= 2;
1846 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1848 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1849 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1851 enum zio_checksum checksum
= os
->os_checksum
;
1852 enum zio_compress compress
= os
->os_compress
;
1853 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1854 boolean_t dedup
= B_FALSE
;
1855 boolean_t nopwrite
= B_FALSE
;
1856 boolean_t dedup_verify
= os
->os_dedup_verify
;
1857 int copies
= os
->os_copies
;
1860 * We maintain different write policies for each of the following
1863 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1864 * 3. all other level 0 blocks
1868 * XXX -- we should design a compression algorithm
1869 * that specializes in arrays of bps.
1871 boolean_t lz4_ac
= spa_feature_is_active(os
->os_spa
,
1872 SPA_FEATURE_LZ4_COMPRESS
);
1874 if (zfs_mdcomp_disable
) {
1875 compress
= ZIO_COMPRESS_EMPTY
;
1876 } else if (lz4_ac
) {
1877 compress
= ZIO_COMPRESS_LZ4
;
1879 compress
= ZIO_COMPRESS_LZJB
;
1883 * Metadata always gets checksummed. If the data
1884 * checksum is multi-bit correctable, and it's not a
1885 * ZBT-style checksum, then it's suitable for metadata
1886 * as well. Otherwise, the metadata checksum defaults
1889 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1890 zio_checksum_table
[checksum
].ci_eck
)
1891 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1893 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
1894 (os
->os_redundant_metadata
==
1895 ZFS_REDUNDANT_METADATA_MOST
&&
1896 (level
>= zfs_redundant_metadata_most_ditto_level
||
1897 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
1899 } else if (wp
& WP_NOFILL
) {
1903 * If we're writing preallocated blocks, we aren't actually
1904 * writing them so don't set any policy properties. These
1905 * blocks are currently only used by an external subsystem
1906 * outside of zfs (i.e. dump) and not written by the zio
1909 compress
= ZIO_COMPRESS_OFF
;
1910 checksum
= ZIO_CHECKSUM_OFF
;
1912 compress
= zio_compress_select(dn
->dn_compress
, compress
);
1914 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1915 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1919 * Determine dedup setting. If we are in dmu_sync(),
1920 * we won't actually dedup now because that's all
1921 * done in syncing context; but we do want to use the
1922 * dedup checkum. If the checksum is not strong
1923 * enough to ensure unique signatures, force
1926 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1927 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1928 if (!zio_checksum_table
[checksum
].ci_dedup
)
1929 dedup_verify
= B_TRUE
;
1933 * Enable nopwrite if we have a cryptographically secure
1934 * checksum that has no known collisions (i.e. SHA-256)
1935 * and compression is enabled. We don't enable nopwrite if
1936 * dedup is enabled as the two features are mutually exclusive.
1938 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1939 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1942 zp
->zp_checksum
= checksum
;
1943 zp
->zp_compress
= compress
;
1944 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1945 zp
->zp_level
= level
;
1946 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
1947 zp
->zp_dedup
= dedup
;
1948 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1949 zp
->zp_nopwrite
= nopwrite
;
1953 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1958 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1962 * Sync any current changes before
1963 * we go trundling through the block pointers.
1965 for (i
= 0; i
< TXG_SIZE
; i
++) {
1966 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1969 if (i
!= TXG_SIZE
) {
1970 dnode_rele(dn
, FTAG
);
1971 txg_wait_synced(dmu_objset_pool(os
), 0);
1972 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1977 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1978 dnode_rele(dn
, FTAG
);
1984 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1986 dnode_phys_t
*dnp
= dn
->dn_phys
;
1989 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1990 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1991 1ULL << dn
->dn_indblkshift
: 0;
1992 doi
->doi_type
= dn
->dn_type
;
1993 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1994 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1995 doi
->doi_indirection
= dn
->dn_nlevels
;
1996 doi
->doi_checksum
= dn
->dn_checksum
;
1997 doi
->doi_compress
= dn
->dn_compress
;
1998 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
1999 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
2000 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
2001 doi
->doi_fill_count
= 0;
2002 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
2003 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
2007 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2009 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2010 mutex_enter(&dn
->dn_mtx
);
2012 __dmu_object_info_from_dnode(dn
, doi
);
2014 mutex_exit(&dn
->dn_mtx
);
2015 rw_exit(&dn
->dn_struct_rwlock
);
2019 * Get information on a DMU object.
2020 * If doi is NULL, just indicates whether the object exists.
2023 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
2026 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
2032 dmu_object_info_from_dnode(dn
, doi
);
2034 dnode_rele(dn
, FTAG
);
2039 * As above, but faster; can be used when you have a held dbuf in hand.
2042 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
2044 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2047 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
2052 * Faster still when you only care about the size.
2053 * This is specifically optimized for zfs_getattr().
2056 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
2057 u_longlong_t
*nblk512
)
2059 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2065 *blksize
= dn
->dn_datablksz
;
2066 /* add 1 for dnode space */
2067 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
2068 SPA_MINBLOCKSHIFT
) + 1;
2073 byteswap_uint64_array(void *vbuf
, size_t size
)
2075 uint64_t *buf
= vbuf
;
2076 size_t count
= size
>> 3;
2079 ASSERT((size
& 7) == 0);
2081 for (i
= 0; i
< count
; i
++)
2082 buf
[i
] = BSWAP_64(buf
[i
]);
2086 byteswap_uint32_array(void *vbuf
, size_t size
)
2088 uint32_t *buf
= vbuf
;
2089 size_t count
= size
>> 2;
2092 ASSERT((size
& 3) == 0);
2094 for (i
= 0; i
< count
; i
++)
2095 buf
[i
] = BSWAP_32(buf
[i
]);
2099 byteswap_uint16_array(void *vbuf
, size_t size
)
2101 uint16_t *buf
= vbuf
;
2102 size_t count
= size
>> 1;
2105 ASSERT((size
& 1) == 0);
2107 for (i
= 0; i
< count
; i
++)
2108 buf
[i
] = BSWAP_16(buf
[i
]);
2113 byteswap_uint8_array(void *vbuf
, size_t size
)
2135 arc_fini(); /* arc depends on l2arc, so arc must go first */
2147 #if defined(_KERNEL) && defined(HAVE_SPL)
2148 EXPORT_SYMBOL(dmu_bonus_hold
);
2149 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2150 EXPORT_SYMBOL(dmu_buf_rele_array
);
2151 EXPORT_SYMBOL(dmu_prefetch
);
2152 EXPORT_SYMBOL(dmu_free_range
);
2153 EXPORT_SYMBOL(dmu_free_long_range
);
2154 EXPORT_SYMBOL(dmu_free_long_object
);
2155 EXPORT_SYMBOL(dmu_read
);
2156 EXPORT_SYMBOL(dmu_write
);
2157 EXPORT_SYMBOL(dmu_prealloc
);
2158 EXPORT_SYMBOL(dmu_object_info
);
2159 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2160 EXPORT_SYMBOL(dmu_object_info_from_db
);
2161 EXPORT_SYMBOL(dmu_object_size_from_db
);
2162 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2163 EXPORT_SYMBOL(dmu_object_set_checksum
);
2164 EXPORT_SYMBOL(dmu_object_set_compress
);
2165 EXPORT_SYMBOL(dmu_write_policy
);
2166 EXPORT_SYMBOL(dmu_sync
);
2167 EXPORT_SYMBOL(dmu_request_arcbuf
);
2168 EXPORT_SYMBOL(dmu_return_arcbuf
);
2169 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2170 EXPORT_SYMBOL(dmu_buf_hold
);
2171 EXPORT_SYMBOL(dmu_ot
);
2173 module_param(zfs_mdcomp_disable
, int, 0644);
2174 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2176 module_param(zfs_nopwrite_enabled
, int, 0644);
2177 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");