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) 2013 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(objset_t
*os
, uint64_t object
, uint64_t offset
,
128 void *tag
, dmu_buf_t
**dbp
, int flags
)
134 int db_flags
= DB_RF_CANFAIL
;
136 if (flags
& DMU_READ_NO_PREFETCH
)
137 db_flags
|= DB_RF_NOPREFETCH
;
139 err
= dnode_hold(os
, object
, FTAG
, &dn
);
142 blkid
= dbuf_whichblock(dn
, offset
);
143 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
144 db
= dbuf_hold(dn
, blkid
, tag
);
145 rw_exit(&dn
->dn_struct_rwlock
);
147 err
= SET_ERROR(EIO
);
149 err
= dbuf_read(db
, NULL
, db_flags
);
156 dnode_rele(dn
, FTAG
);
157 *dbp
= &db
->db
; /* NULL db plus first field offset is NULL */
164 return (DN_MAX_BONUSLEN
);
168 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
170 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
177 if (dn
->dn_bonus
!= db
) {
178 error
= SET_ERROR(EINVAL
);
179 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
180 error
= SET_ERROR(EINVAL
);
182 dnode_setbonuslen(dn
, newsize
, tx
);
191 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
193 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
200 if (!DMU_OT_IS_VALID(type
)) {
201 error
= SET_ERROR(EINVAL
);
202 } else if (dn
->dn_bonus
!= db
) {
203 error
= SET_ERROR(EINVAL
);
205 dnode_setbonus_type(dn
, type
, tx
);
214 dmu_get_bonustype(dmu_buf_t
*db_fake
)
216 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
218 dmu_object_type_t type
;
222 type
= dn
->dn_bonustype
;
229 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
234 error
= dnode_hold(os
, object
, FTAG
, &dn
);
235 dbuf_rm_spill(dn
, tx
);
236 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
237 dnode_rm_spill(dn
, tx
);
238 rw_exit(&dn
->dn_struct_rwlock
);
239 dnode_rele(dn
, FTAG
);
244 * returns ENOENT, EIO, or 0.
247 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
253 error
= dnode_hold(os
, object
, FTAG
, &dn
);
257 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
258 if (dn
->dn_bonus
== NULL
) {
259 rw_exit(&dn
->dn_struct_rwlock
);
260 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
261 if (dn
->dn_bonus
== NULL
)
262 dbuf_create_bonus(dn
);
266 /* as long as the bonus buf is held, the dnode will be held */
267 if (refcount_add(&db
->db_holds
, tag
) == 1) {
268 VERIFY(dnode_add_ref(dn
, db
));
269 (void) atomic_inc_32_nv(&dn
->dn_dbufs_count
);
273 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
274 * hold and incrementing the dbuf count to ensure that dnode_move() sees
275 * a dnode hold for every dbuf.
277 rw_exit(&dn
->dn_struct_rwlock
);
279 dnode_rele(dn
, FTAG
);
281 VERIFY(0 == dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
));
288 * returns ENOENT, EIO, or 0.
290 * This interface will allocate a blank spill dbuf when a spill blk
291 * doesn't already exist on the dnode.
293 * if you only want to find an already existing spill db, then
294 * dmu_spill_hold_existing() should be used.
297 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
299 dmu_buf_impl_t
*db
= NULL
;
302 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
303 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
305 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
307 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
308 rw_exit(&dn
->dn_struct_rwlock
);
311 err
= dbuf_read(db
, NULL
, flags
);
320 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
322 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
329 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
330 err
= SET_ERROR(EINVAL
);
332 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
334 if (!dn
->dn_have_spill
) {
335 err
= SET_ERROR(ENOENT
);
337 err
= dmu_spill_hold_by_dnode(dn
,
338 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
341 rw_exit(&dn
->dn_struct_rwlock
);
349 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
351 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
357 err
= dmu_spill_hold_by_dnode(dn
, DB_RF_CANFAIL
, tag
, dbp
);
364 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
365 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
366 * and can induce severe lock contention when writing to several files
367 * whose dnodes are in the same block.
370 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
371 int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
374 uint64_t blkid
, nblks
, i
;
379 ASSERT(length
<= DMU_MAX_ACCESS
);
381 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
;
382 if (flags
& DMU_READ_NO_PREFETCH
|| length
> zfetch_array_rd_sz
)
383 dbuf_flags
|= DB_RF_NOPREFETCH
;
385 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
386 if (dn
->dn_datablkshift
) {
387 int blkshift
= dn
->dn_datablkshift
;
388 nblks
= (P2ROUNDUP(offset
+length
, 1ULL<<blkshift
) -
389 P2ALIGN(offset
, 1ULL<<blkshift
)) >> blkshift
;
391 if (offset
+ length
> dn
->dn_datablksz
) {
392 zfs_panic_recover("zfs: accessing past end of object "
393 "%llx/%llx (size=%u access=%llu+%llu)",
394 (longlong_t
)dn
->dn_objset
->
395 os_dsl_dataset
->ds_object
,
396 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
397 (longlong_t
)offset
, (longlong_t
)length
);
398 rw_exit(&dn
->dn_struct_rwlock
);
399 return (SET_ERROR(EIO
));
403 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_PUSHPAGE
| KM_NODEBUG
);
405 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
406 blkid
= dbuf_whichblock(dn
, offset
);
407 for (i
= 0; i
< nblks
; i
++) {
408 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+i
, tag
);
410 rw_exit(&dn
->dn_struct_rwlock
);
411 dmu_buf_rele_array(dbp
, nblks
, tag
);
413 return (SET_ERROR(EIO
));
415 /* initiate async i/o */
417 (void) dbuf_read(db
, zio
, dbuf_flags
);
421 rw_exit(&dn
->dn_struct_rwlock
);
423 /* wait for async i/o */
426 dmu_buf_rele_array(dbp
, nblks
, tag
);
430 /* wait for other io to complete */
432 for (i
= 0; i
< nblks
; i
++) {
433 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
434 mutex_enter(&db
->db_mtx
);
435 while (db
->db_state
== DB_READ
||
436 db
->db_state
== DB_FILL
)
437 cv_wait(&db
->db_changed
, &db
->db_mtx
);
438 if (db
->db_state
== DB_UNCACHED
)
439 err
= SET_ERROR(EIO
);
440 mutex_exit(&db
->db_mtx
);
442 dmu_buf_rele_array(dbp
, nblks
, tag
);
454 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
455 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
460 err
= dnode_hold(os
, object
, FTAG
, &dn
);
464 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
465 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
467 dnode_rele(dn
, FTAG
);
473 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
474 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
476 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
482 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
483 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
490 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
493 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
498 for (i
= 0; i
< numbufs
; i
++) {
500 dbuf_rele(dbp
[i
], tag
);
503 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
507 * Issue prefetch i/os for the given blocks.
509 * Note: The assumption is that we *know* these blocks will be needed
510 * almost immediately. Therefore, the prefetch i/os will be issued at
511 * ZIO_PRIORITY_SYNC_READ
513 * Note: indirect blocks and other metadata will be read synchronously,
514 * causing this function to block if they are not already cached.
517 dmu_prefetch(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t len
)
523 if (zfs_prefetch_disable
)
526 if (len
== 0) { /* they're interested in the bonus buffer */
527 dn
= DMU_META_DNODE(os
);
529 if (object
== 0 || object
>= DN_MAX_OBJECT
)
532 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
533 blkid
= dbuf_whichblock(dn
, object
* sizeof (dnode_phys_t
));
534 dbuf_prefetch(dn
, blkid
, ZIO_PRIORITY_SYNC_READ
);
535 rw_exit(&dn
->dn_struct_rwlock
);
540 * XXX - Note, if the dnode for the requested object is not
541 * already cached, we will do a *synchronous* read in the
542 * dnode_hold() call. The same is true for any indirects.
544 err
= dnode_hold(os
, object
, FTAG
, &dn
);
548 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
549 if (dn
->dn_datablkshift
) {
550 int blkshift
= dn
->dn_datablkshift
;
551 nblks
= (P2ROUNDUP(offset
+ len
, 1 << blkshift
) -
552 P2ALIGN(offset
, 1 << blkshift
)) >> blkshift
;
554 nblks
= (offset
< dn
->dn_datablksz
);
560 blkid
= dbuf_whichblock(dn
, offset
);
561 for (i
= 0; i
< nblks
; i
++)
562 dbuf_prefetch(dn
, blkid
+ i
, ZIO_PRIORITY_SYNC_READ
);
565 rw_exit(&dn
->dn_struct_rwlock
);
567 dnode_rele(dn
, FTAG
);
571 * Get the next "chunk" of file data to free. We traverse the file from
572 * the end so that the file gets shorter over time (if we crashes in the
573 * middle, this will leave us in a better state). We find allocated file
574 * data by simply searching the allocated level 1 indirects.
576 * On input, *start should be the first offset that does not need to be
577 * freed (e.g. "offset + length"). On return, *start will be the first
578 * offset that should be freed.
581 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
583 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
584 /* bytes of data covered by a level-1 indirect block */
586 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
589 ASSERT3U(minimum
, <=, *start
);
591 if (*start
- minimum
<= iblkrange
* maxblks
) {
595 ASSERT(ISP2(iblkrange
));
597 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
601 * dnode_next_offset(BACKWARDS) will find an allocated L1
602 * indirect block at or before the input offset. We must
603 * decrement *start so that it is at the end of the region
607 err
= dnode_next_offset(dn
,
608 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
610 /* if there are no indirect blocks before start, we are done */
614 } else if (err
!= 0) {
618 /* set start to the beginning of this L1 indirect */
619 *start
= P2ALIGN(*start
, iblkrange
);
621 if (*start
< minimum
)
627 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
630 uint64_t object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
633 if (offset
>= object_size
)
636 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
637 length
= object_size
- offset
;
639 while (length
!= 0) {
640 uint64_t chunk_end
, chunk_begin
;
643 chunk_end
= chunk_begin
= offset
+ length
;
645 /* move chunk_begin backwards to the beginning of this chunk */
646 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
649 ASSERT3U(chunk_begin
, >=, offset
);
650 ASSERT3U(chunk_begin
, <=, chunk_end
);
652 tx
= dmu_tx_create(os
);
653 dmu_tx_hold_free(tx
, dn
->dn_object
,
654 chunk_begin
, chunk_end
- chunk_begin
);
655 err
= dmu_tx_assign(tx
, TXG_WAIT
);
660 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
663 length
-= chunk_end
- chunk_begin
;
669 dmu_free_long_range(objset_t
*os
, uint64_t object
,
670 uint64_t offset
, uint64_t length
)
675 err
= dnode_hold(os
, object
, FTAG
, &dn
);
678 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
681 * It is important to zero out the maxblkid when freeing the entire
682 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
683 * will take the fast path, and (b) dnode_reallocate() can verify
684 * that the entire file has been freed.
686 if (offset
== 0 && length
== DMU_OBJECT_END
)
689 dnode_rele(dn
, FTAG
);
694 dmu_free_long_object(objset_t
*os
, uint64_t object
)
699 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
703 tx
= dmu_tx_create(os
);
704 dmu_tx_hold_bonus(tx
, object
);
705 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
706 err
= dmu_tx_assign(tx
, TXG_WAIT
);
708 err
= dmu_object_free(os
, object
, tx
);
718 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
719 uint64_t size
, dmu_tx_t
*tx
)
722 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
725 ASSERT(offset
< UINT64_MAX
);
726 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
727 dnode_free_range(dn
, offset
, size
, tx
);
728 dnode_rele(dn
, FTAG
);
733 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
734 void *buf
, uint32_t flags
)
740 err
= dnode_hold(os
, object
, FTAG
, &dn
);
745 * Deal with odd block sizes, where there can't be data past the first
746 * block. If we ever do the tail block optimization, we will need to
747 * handle that here as well.
749 if (dn
->dn_maxblkid
== 0) {
750 int newsz
= offset
> dn
->dn_datablksz
? 0 :
751 MIN(size
, dn
->dn_datablksz
- offset
);
752 bzero((char *)buf
+ newsz
, size
- newsz
);
757 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
761 * NB: we could do this block-at-a-time, but it's nice
762 * to be reading in parallel.
764 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
765 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
769 for (i
= 0; i
< numbufs
; i
++) {
772 dmu_buf_t
*db
= dbp
[i
];
776 bufoff
= offset
- db
->db_offset
;
777 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
779 bcopy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
783 buf
= (char *)buf
+ tocpy
;
785 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
787 dnode_rele(dn
, FTAG
);
792 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
793 const void *buf
, dmu_tx_t
*tx
)
801 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
802 FALSE
, FTAG
, &numbufs
, &dbp
));
804 for (i
= 0; i
< numbufs
; i
++) {
807 dmu_buf_t
*db
= dbp
[i
];
811 bufoff
= offset
- db
->db_offset
;
812 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
814 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
816 if (tocpy
== db
->db_size
)
817 dmu_buf_will_fill(db
, tx
);
819 dmu_buf_will_dirty(db
, tx
);
821 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
823 if (tocpy
== db
->db_size
)
824 dmu_buf_fill_done(db
, tx
);
828 buf
= (char *)buf
+ tocpy
;
830 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
834 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
843 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
844 FALSE
, FTAG
, &numbufs
, &dbp
));
846 for (i
= 0; i
< numbufs
; i
++) {
847 dmu_buf_t
*db
= dbp
[i
];
849 dmu_buf_will_not_fill(db
, tx
);
851 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
855 * DMU support for xuio
857 kstat_t
*xuio_ksp
= NULL
;
859 typedef struct xuio_stats
{
860 /* loaned yet not returned arc_buf */
861 kstat_named_t xuiostat_onloan_rbuf
;
862 kstat_named_t xuiostat_onloan_wbuf
;
863 /* whether a copy is made when loaning out a read buffer */
864 kstat_named_t xuiostat_rbuf_copied
;
865 kstat_named_t xuiostat_rbuf_nocopy
;
866 /* whether a copy is made when assigning a write buffer */
867 kstat_named_t xuiostat_wbuf_copied
;
868 kstat_named_t xuiostat_wbuf_nocopy
;
871 static xuio_stats_t xuio_stats
= {
872 { "onloan_read_buf", KSTAT_DATA_UINT64
},
873 { "onloan_write_buf", KSTAT_DATA_UINT64
},
874 { "read_buf_copied", KSTAT_DATA_UINT64
},
875 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
876 { "write_buf_copied", KSTAT_DATA_UINT64
},
877 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
880 #define XUIOSTAT_INCR(stat, val) \
881 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
882 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
885 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
888 uio_t
*uio
= &xuio
->xu_uio
;
890 uio
->uio_iovcnt
= nblk
;
891 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_PUSHPAGE
);
893 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_PUSHPAGE
);
895 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_PUSHPAGE
);
896 priv
->iovp
= uio
->uio_iov
;
897 XUIO_XUZC_PRIV(xuio
) = priv
;
899 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
900 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
902 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
908 dmu_xuio_fini(xuio_t
*xuio
)
910 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
911 int nblk
= priv
->cnt
;
913 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
914 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
915 kmem_free(priv
, sizeof (dmu_xuio_t
));
917 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
918 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
920 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
924 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
925 * and increase priv->next by 1.
928 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
931 uio_t
*uio
= &xuio
->xu_uio
;
932 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
933 int i
= priv
->next
++;
935 ASSERT(i
< priv
->cnt
);
936 ASSERT(off
+ n
<= arc_buf_size(abuf
));
937 iov
= uio
->uio_iov
+ i
;
938 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
940 priv
->bufs
[i
] = abuf
;
945 dmu_xuio_cnt(xuio_t
*xuio
)
947 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
952 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
954 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
956 ASSERT(i
< priv
->cnt
);
957 return (priv
->bufs
[i
]);
961 dmu_xuio_clear(xuio_t
*xuio
, int i
)
963 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
965 ASSERT(i
< priv
->cnt
);
966 priv
->bufs
[i
] = NULL
;
972 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
973 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
975 if (xuio_ksp
!= NULL
) {
976 xuio_ksp
->ks_data
= &xuio_stats
;
977 kstat_install(xuio_ksp
);
984 if (xuio_ksp
!= NULL
) {
985 kstat_delete(xuio_ksp
);
991 xuio_stat_wbuf_copied()
993 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
997 xuio_stat_wbuf_nocopy()
999 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1005 * Copy up to size bytes between arg_buf and req based on the data direction
1006 * described by the req. If an entire req's data cannot be transfered the
1007 * req's is updated such that it's current index and bv offsets correctly
1008 * reference any residual data which could not be copied. The return value
1009 * is the number of bytes successfully copied to arg_buf.
1012 dmu_req_copy(void *arg_buf
, int size
, int *offset
, struct request
*req
)
1015 struct req_iterator iter
;
1020 rq_for_each_segment(bv
, req
, iter
) {
1022 /* Fully consumed the passed arg_buf */
1023 ASSERT3S(*offset
, <=, size
);
1024 if (size
== *offset
)
1027 /* Skip fully consumed bv's */
1028 if (bv
->bv_len
== 0)
1031 tocpy
= MIN(bv
->bv_len
, size
- *offset
);
1032 ASSERT3S(tocpy
, >=, 0);
1034 bv_buf
= page_address(bv
->bv_page
) + bv
->bv_offset
;
1035 ASSERT3P(bv_buf
, !=, NULL
);
1037 if (rq_data_dir(req
) == WRITE
)
1038 memcpy(arg_buf
+ *offset
, bv_buf
, tocpy
);
1040 memcpy(bv_buf
, arg_buf
+ *offset
, tocpy
);
1043 bv
->bv_offset
+= tocpy
;
1044 bv
->bv_len
-= tocpy
;
1051 dmu_bio_put(struct bio
*bio
)
1053 struct bio
*bio_next
;
1056 bio_next
= bio
->bi_next
;
1063 dmu_bio_clone(struct bio
*bio
, struct bio
**bio_copy
)
1065 struct bio
*bio_root
= NULL
;
1066 struct bio
*bio_last
= NULL
;
1067 struct bio
*bio_new
;
1073 bio_new
= bio_clone(bio
, GFP_NOIO
);
1074 if (bio_new
== NULL
) {
1075 dmu_bio_put(bio_root
);
1080 bio_last
->bi_next
= bio_new
;
1090 *bio_copy
= bio_root
;
1096 dmu_read_req(objset_t
*os
, uint64_t object
, struct request
*req
)
1098 uint64_t size
= blk_rq_bytes(req
);
1099 uint64_t offset
= blk_rq_pos(req
) << 9;
1100 struct bio
*bio_saved
= req
->bio
;
1102 int numbufs
, i
, err
;
1105 * NB: we could do this block-at-a-time, but it's nice
1106 * to be reading in parallel.
1108 err
= dmu_buf_hold_array(os
, object
, offset
, size
, TRUE
, FTAG
,
1114 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1115 * can be safely modified. The original bio list is relinked in to
1116 * the request when the function exits. This is required because
1117 * some file systems blindly assume that these values will remain
1118 * constant between bio_submit() and the IO completion callback.
1120 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1124 for (i
= 0; i
< numbufs
; i
++) {
1125 int tocpy
, didcpy
, bufoff
;
1126 dmu_buf_t
*db
= dbp
[i
];
1128 bufoff
= offset
- db
->db_offset
;
1129 ASSERT3S(bufoff
, >=, 0);
1131 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1135 err
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, &didcpy
, req
);
1148 dmu_bio_put(req
->bio
);
1149 req
->bio
= bio_saved
;
1151 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1157 dmu_write_req(objset_t
*os
, uint64_t object
, struct request
*req
, dmu_tx_t
*tx
)
1159 uint64_t size
= blk_rq_bytes(req
);
1160 uint64_t offset
= blk_rq_pos(req
) << 9;
1161 struct bio
*bio_saved
= req
->bio
;
1170 err
= dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1176 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1177 * can be safely modified. The original bio list is relinked in to
1178 * the request when the function exits. This is required because
1179 * some file systems blindly assume that these values will remain
1180 * constant between bio_submit() and the IO completion callback.
1182 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1186 for (i
= 0; i
< numbufs
; i
++) {
1187 int tocpy
, didcpy
, bufoff
;
1188 dmu_buf_t
*db
= dbp
[i
];
1190 bufoff
= offset
- db
->db_offset
;
1191 ASSERT3S(bufoff
, >=, 0);
1193 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1197 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1199 if (tocpy
== db
->db_size
)
1200 dmu_buf_will_fill(db
, tx
);
1202 dmu_buf_will_dirty(db
, tx
);
1204 err
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, &didcpy
, req
);
1206 if (tocpy
== db
->db_size
)
1207 dmu_buf_fill_done(db
, tx
);
1220 dmu_bio_put(req
->bio
);
1221 req
->bio
= bio_saved
;
1223 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1229 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1232 int numbufs
, i
, err
;
1233 xuio_t
*xuio
= NULL
;
1236 * NB: we could do this block-at-a-time, but it's nice
1237 * to be reading in parallel.
1239 err
= dmu_buf_hold_array(os
, object
, uio
->uio_loffset
, size
, TRUE
, FTAG
,
1244 for (i
= 0; i
< numbufs
; i
++) {
1247 dmu_buf_t
*db
= dbp
[i
];
1251 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1252 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1255 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1256 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1257 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1258 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1260 uio
->uio_resid
-= tocpy
;
1261 uio
->uio_loffset
+= tocpy
;
1264 if (abuf
== dbuf_abuf
)
1265 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1267 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1269 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1277 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1283 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1290 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1291 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1295 for (i
= 0; i
< numbufs
; i
++) {
1298 dmu_buf_t
*db
= dbp
[i
];
1302 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1303 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1305 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1307 if (tocpy
== db
->db_size
)
1308 dmu_buf_will_fill(db
, tx
);
1310 dmu_buf_will_dirty(db
, tx
);
1313 * XXX uiomove could block forever (eg.nfs-backed
1314 * pages). There needs to be a uiolockdown() function
1315 * to lock the pages in memory, so that uiomove won't
1318 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1321 if (tocpy
== db
->db_size
)
1322 dmu_buf_fill_done(db
, tx
);
1330 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1335 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1338 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1347 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1354 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1363 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1367 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1369 dnode_rele(dn
, FTAG
);
1373 #endif /* _KERNEL */
1376 * Allocate a loaned anonymous arc buffer.
1379 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1381 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1384 DB_GET_SPA(&spa
, db
);
1385 return (arc_loan_buf(spa
, size
));
1389 * Free a loaned arc buffer.
1392 dmu_return_arcbuf(arc_buf_t
*buf
)
1394 arc_return_buf(buf
, FTAG
);
1395 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1399 * When possible directly assign passed loaned arc buffer to a dbuf.
1400 * If this is not possible copy the contents of passed arc buf via
1404 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1407 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1410 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1413 DB_DNODE_ENTER(dbuf
);
1414 dn
= DB_DNODE(dbuf
);
1415 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1416 blkid
= dbuf_whichblock(dn
, offset
);
1417 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1418 rw_exit(&dn
->dn_struct_rwlock
);
1419 DB_DNODE_EXIT(dbuf
);
1421 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1422 dbuf_assign_arcbuf(db
, buf
, tx
);
1423 dbuf_rele(db
, FTAG
);
1428 DB_DNODE_ENTER(dbuf
);
1429 dn
= DB_DNODE(dbuf
);
1431 object
= dn
->dn_object
;
1432 DB_DNODE_EXIT(dbuf
);
1434 dbuf_rele(db
, FTAG
);
1435 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1436 dmu_return_arcbuf(buf
);
1437 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1442 dbuf_dirty_record_t
*dsa_dr
;
1443 dmu_sync_cb_t
*dsa_done
;
1450 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1452 dmu_sync_arg_t
*dsa
= varg
;
1453 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1454 blkptr_t
*bp
= zio
->io_bp
;
1456 if (zio
->io_error
== 0) {
1457 if (BP_IS_HOLE(bp
)) {
1459 * A block of zeros may compress to a hole, but the
1460 * block size still needs to be known for replay.
1462 BP_SET_LSIZE(bp
, db
->db_size
);
1464 ASSERT(BP_GET_LEVEL(bp
) == 0);
1471 dmu_sync_late_arrival_ready(zio_t
*zio
)
1473 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1478 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1480 dmu_sync_arg_t
*dsa
= varg
;
1481 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1482 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1484 mutex_enter(&db
->db_mtx
);
1485 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1486 if (zio
->io_error
== 0) {
1487 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1488 if (dr
->dt
.dl
.dr_nopwrite
) {
1489 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1490 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1491 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1493 ASSERT(BP_EQUAL(bp
, bp_orig
));
1494 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1495 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1497 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1498 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1499 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1500 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
))
1501 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1503 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1505 cv_broadcast(&db
->db_changed
);
1506 mutex_exit(&db
->db_mtx
);
1508 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1510 kmem_free(dsa
, sizeof (*dsa
));
1514 dmu_sync_late_arrival_done(zio_t
*zio
)
1516 blkptr_t
*bp
= zio
->io_bp
;
1517 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1518 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1520 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1522 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1523 * then there is nothing to do here. Otherwise, free the
1524 * newly allocated block in this txg.
1526 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1527 ASSERT(BP_EQUAL(bp
, bp_orig
));
1529 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1530 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1531 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1532 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1536 dmu_tx_commit(dsa
->dsa_tx
);
1538 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1540 kmem_free(dsa
, sizeof (*dsa
));
1544 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1545 zio_prop_t
*zp
, zbookmark_t
*zb
)
1547 dmu_sync_arg_t
*dsa
;
1550 tx
= dmu_tx_create(os
);
1551 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1552 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1554 /* Make zl_get_data do txg_waited_synced() */
1555 return (SET_ERROR(EIO
));
1558 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1560 dsa
->dsa_done
= done
;
1564 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1565 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1566 dmu_sync_late_arrival_ready
, NULL
, dmu_sync_late_arrival_done
, dsa
,
1567 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_FASTWRITE
, zb
));
1573 * Intent log support: sync the block associated with db to disk.
1574 * N.B. and XXX: the caller is responsible for making sure that the
1575 * data isn't changing while dmu_sync() is writing it.
1579 * EEXIST: this txg has already been synced, so there's nothing to do.
1580 * The caller should not log the write.
1582 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1583 * The caller should not log the write.
1585 * EALREADY: this block is already in the process of being synced.
1586 * The caller should track its progress (somehow).
1588 * EIO: could not do the I/O.
1589 * The caller should do a txg_wait_synced().
1591 * 0: the I/O has been initiated.
1592 * The caller should log this blkptr in the done callback.
1593 * It is possible that the I/O will fail, in which case
1594 * the error will be reported to the done callback and
1595 * propagated to pio from zio_done().
1598 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1600 blkptr_t
*bp
= zgd
->zgd_bp
;
1601 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1602 objset_t
*os
= db
->db_objset
;
1603 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1604 dbuf_dirty_record_t
*dr
;
1605 dmu_sync_arg_t
*dsa
;
1610 ASSERT(pio
!= NULL
);
1613 SET_BOOKMARK(&zb
, ds
->ds_object
,
1614 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1618 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1622 * If we're frozen (running ziltest), we always need to generate a bp.
1624 if (txg
> spa_freeze_txg(os
->os_spa
))
1625 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1628 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1629 * and us. If we determine that this txg is not yet syncing,
1630 * but it begins to sync a moment later, that's OK because the
1631 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1633 mutex_enter(&db
->db_mtx
);
1635 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1637 * This txg has already synced. There's nothing to do.
1639 mutex_exit(&db
->db_mtx
);
1640 return (SET_ERROR(EEXIST
));
1643 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1645 * This txg is currently syncing, so we can't mess with
1646 * the dirty record anymore; just write a new log block.
1648 mutex_exit(&db
->db_mtx
);
1649 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1652 dr
= db
->db_last_dirty
;
1653 while (dr
&& dr
->dr_txg
!= txg
)
1658 * There's no dr for this dbuf, so it must have been freed.
1659 * There's no need to log writes to freed blocks, so we're done.
1661 mutex_exit(&db
->db_mtx
);
1662 return (SET_ERROR(ENOENT
));
1665 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1668 * Assume the on-disk data is X, the current syncing data is Y,
1669 * and the current in-memory data is Z (currently in dmu_sync).
1670 * X and Z are identical but Y is has been modified. Normally,
1671 * when X and Z are the same we will perform a nopwrite but if Y
1672 * is different we must disable nopwrite since the resulting write
1673 * of Y to disk can free the block containing X. If we allowed a
1674 * nopwrite to occur the block pointing to Z would reference a freed
1675 * block. Since this is a rare case we simplify this by disabling
1676 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1677 * a previous transaction.
1680 zp
.zp_nopwrite
= B_FALSE
;
1682 ASSERT(dr
->dr_txg
== txg
);
1683 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1684 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1686 * We have already issued a sync write for this buffer,
1687 * or this buffer has already been synced. It could not
1688 * have been dirtied since, or we would have cleared the state.
1690 mutex_exit(&db
->db_mtx
);
1691 return (SET_ERROR(EALREADY
));
1694 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1695 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1696 mutex_exit(&db
->db_mtx
);
1698 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1700 dsa
->dsa_done
= done
;
1704 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1705 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1706 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1707 NULL
, dmu_sync_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1708 ZIO_FLAG_CANFAIL
, &zb
));
1714 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1720 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1723 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1724 dnode_rele(dn
, FTAG
);
1729 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1734 /* XXX assumes dnode_hold will not get an i/o error */
1735 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1736 ASSERT(checksum
< ZIO_CHECKSUM_FUNCTIONS
);
1737 dn
->dn_checksum
= checksum
;
1738 dnode_setdirty(dn
, tx
);
1739 dnode_rele(dn
, FTAG
);
1743 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1748 /* XXX assumes dnode_hold will not get an i/o error */
1749 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1750 ASSERT(compress
< ZIO_COMPRESS_FUNCTIONS
);
1751 dn
->dn_compress
= compress
;
1752 dnode_setdirty(dn
, tx
);
1753 dnode_rele(dn
, FTAG
);
1756 int zfs_mdcomp_disable
= 0;
1759 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1761 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1762 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1764 enum zio_checksum checksum
= os
->os_checksum
;
1765 enum zio_compress compress
= os
->os_compress
;
1766 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1767 boolean_t dedup
= B_FALSE
;
1768 boolean_t nopwrite
= B_FALSE
;
1769 boolean_t dedup_verify
= os
->os_dedup_verify
;
1770 int copies
= os
->os_copies
;
1773 * We maintain different write policies for each of the following
1776 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1777 * 3. all other level 0 blocks
1781 * XXX -- we should design a compression algorithm
1782 * that specializes in arrays of bps.
1784 compress
= zfs_mdcomp_disable
? ZIO_COMPRESS_EMPTY
:
1788 * Metadata always gets checksummed. If the data
1789 * checksum is multi-bit correctable, and it's not a
1790 * ZBT-style checksum, then it's suitable for metadata
1791 * as well. Otherwise, the metadata checksum defaults
1794 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1795 zio_checksum_table
[checksum
].ci_eck
)
1796 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1797 } else if (wp
& WP_NOFILL
) {
1801 * If we're writing preallocated blocks, we aren't actually
1802 * writing them so don't set any policy properties. These
1803 * blocks are currently only used by an external subsystem
1804 * outside of zfs (i.e. dump) and not written by the zio
1807 compress
= ZIO_COMPRESS_OFF
;
1808 checksum
= ZIO_CHECKSUM_OFF
;
1810 compress
= zio_compress_select(dn
->dn_compress
, compress
);
1812 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1813 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1817 * Determine dedup setting. If we are in dmu_sync(),
1818 * we won't actually dedup now because that's all
1819 * done in syncing context; but we do want to use the
1820 * dedup checkum. If the checksum is not strong
1821 * enough to ensure unique signatures, force
1824 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1825 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1826 if (!zio_checksum_table
[checksum
].ci_dedup
)
1827 dedup_verify
= B_TRUE
;
1831 * Enable nopwrite if we have a cryptographically secure
1832 * checksum that has no known collisions (i.e. SHA-256)
1833 * and compression is enabled. We don't enable nopwrite if
1834 * dedup is enabled as the two features are mutually exclusive.
1836 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1837 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1840 zp
->zp_checksum
= checksum
;
1841 zp
->zp_compress
= compress
;
1842 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1843 zp
->zp_level
= level
;
1844 zp
->zp_copies
= MIN(copies
+ ismd
, spa_max_replication(os
->os_spa
));
1845 zp
->zp_dedup
= dedup
;
1846 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1847 zp
->zp_nopwrite
= nopwrite
;
1851 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1856 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1860 * Sync any current changes before
1861 * we go trundling through the block pointers.
1863 for (i
= 0; i
< TXG_SIZE
; i
++) {
1864 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1867 if (i
!= TXG_SIZE
) {
1868 dnode_rele(dn
, FTAG
);
1869 txg_wait_synced(dmu_objset_pool(os
), 0);
1870 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1875 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1876 dnode_rele(dn
, FTAG
);
1882 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1884 dnode_phys_t
*dnp
= dn
->dn_phys
;
1887 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1888 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1889 1ULL << dn
->dn_indblkshift
: 0;
1890 doi
->doi_type
= dn
->dn_type
;
1891 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1892 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1893 doi
->doi_indirection
= dn
->dn_nlevels
;
1894 doi
->doi_checksum
= dn
->dn_checksum
;
1895 doi
->doi_compress
= dn
->dn_compress
;
1896 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1897 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1898 doi
->doi_fill_count
= 0;
1899 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1900 doi
->doi_fill_count
+= dnp
->dn_blkptr
[i
].blk_fill
;
1904 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1906 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1907 mutex_enter(&dn
->dn_mtx
);
1909 __dmu_object_info_from_dnode(dn
, doi
);
1911 mutex_exit(&dn
->dn_mtx
);
1912 rw_exit(&dn
->dn_struct_rwlock
);
1916 * Get information on a DMU object.
1917 * If doi is NULL, just indicates whether the object exists.
1920 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1923 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1929 dmu_object_info_from_dnode(dn
, doi
);
1931 dnode_rele(dn
, FTAG
);
1936 * As above, but faster; can be used when you have a held dbuf in hand.
1939 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1941 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1944 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1949 * Faster still when you only care about the size.
1950 * This is specifically optimized for zfs_getattr().
1953 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1954 u_longlong_t
*nblk512
)
1956 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1962 *blksize
= dn
->dn_datablksz
;
1963 /* add 1 for dnode space */
1964 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1965 SPA_MINBLOCKSHIFT
) + 1;
1970 byteswap_uint64_array(void *vbuf
, size_t size
)
1972 uint64_t *buf
= vbuf
;
1973 size_t count
= size
>> 3;
1976 ASSERT((size
& 7) == 0);
1978 for (i
= 0; i
< count
; i
++)
1979 buf
[i
] = BSWAP_64(buf
[i
]);
1983 byteswap_uint32_array(void *vbuf
, size_t size
)
1985 uint32_t *buf
= vbuf
;
1986 size_t count
= size
>> 2;
1989 ASSERT((size
& 3) == 0);
1991 for (i
= 0; i
< count
; i
++)
1992 buf
[i
] = BSWAP_32(buf
[i
]);
1996 byteswap_uint16_array(void *vbuf
, size_t size
)
1998 uint16_t *buf
= vbuf
;
1999 size_t count
= size
>> 1;
2002 ASSERT((size
& 1) == 0);
2004 for (i
= 0; i
< count
; i
++)
2005 buf
[i
] = BSWAP_16(buf
[i
]);
2010 byteswap_uint8_array(void *vbuf
, size_t size
)
2032 arc_fini(); /* arc depends on l2arc, so arc must go first */
2044 #if defined(_KERNEL) && defined(HAVE_SPL)
2045 EXPORT_SYMBOL(dmu_bonus_hold
);
2046 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2047 EXPORT_SYMBOL(dmu_buf_rele_array
);
2048 EXPORT_SYMBOL(dmu_prefetch
);
2049 EXPORT_SYMBOL(dmu_free_range
);
2050 EXPORT_SYMBOL(dmu_free_long_range
);
2051 EXPORT_SYMBOL(dmu_free_long_object
);
2052 EXPORT_SYMBOL(dmu_read
);
2053 EXPORT_SYMBOL(dmu_write
);
2054 EXPORT_SYMBOL(dmu_prealloc
);
2055 EXPORT_SYMBOL(dmu_object_info
);
2056 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2057 EXPORT_SYMBOL(dmu_object_info_from_db
);
2058 EXPORT_SYMBOL(dmu_object_size_from_db
);
2059 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2060 EXPORT_SYMBOL(dmu_object_set_checksum
);
2061 EXPORT_SYMBOL(dmu_object_set_compress
);
2062 EXPORT_SYMBOL(dmu_write_policy
);
2063 EXPORT_SYMBOL(dmu_sync
);
2064 EXPORT_SYMBOL(dmu_request_arcbuf
);
2065 EXPORT_SYMBOL(dmu_return_arcbuf
);
2066 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2067 EXPORT_SYMBOL(dmu_buf_hold
);
2068 EXPORT_SYMBOL(dmu_ot
);
2070 module_param(zfs_mdcomp_disable
, int, 0644);
2071 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2073 module_param(zfs_nopwrite_enabled
, int, 0644);
2074 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");