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
)
373 dsl_pool_t
*dp
= NULL
;
375 uint64_t blkid
, nblks
, i
;
381 ASSERT(length
<= DMU_MAX_ACCESS
);
383 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
;
384 if (flags
& DMU_READ_NO_PREFETCH
|| length
> zfetch_array_rd_sz
)
385 dbuf_flags
|= DB_RF_NOPREFETCH
;
387 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
388 if (dn
->dn_datablkshift
) {
389 int blkshift
= dn
->dn_datablkshift
;
390 nblks
= (P2ROUNDUP(offset
+length
, 1ULL<<blkshift
) -
391 P2ALIGN(offset
, 1ULL<<blkshift
)) >> blkshift
;
393 if (offset
+ length
> dn
->dn_datablksz
) {
394 zfs_panic_recover("zfs: accessing past end of object "
395 "%llx/%llx (size=%u access=%llu+%llu)",
396 (longlong_t
)dn
->dn_objset
->
397 os_dsl_dataset
->ds_object
,
398 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
399 (longlong_t
)offset
, (longlong_t
)length
);
400 rw_exit(&dn
->dn_struct_rwlock
);
401 return (SET_ERROR(EIO
));
405 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_PUSHPAGE
| KM_NODEBUG
);
407 if (dn
->dn_objset
->os_dsl_dataset
)
408 dp
= dn
->dn_objset
->os_dsl_dataset
->ds_dir
->dd_pool
;
410 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
411 blkid
= dbuf_whichblock(dn
, offset
);
412 for (i
= 0; i
< nblks
; i
++) {
413 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+i
, tag
);
415 rw_exit(&dn
->dn_struct_rwlock
);
416 dmu_buf_rele_array(dbp
, nblks
, tag
);
418 return (SET_ERROR(EIO
));
420 /* initiate async i/o */
422 (void) dbuf_read(db
, zio
, dbuf_flags
);
426 rw_exit(&dn
->dn_struct_rwlock
);
428 /* wait for async i/o */
430 /* track read overhead when we are in sync context */
431 if (dp
&& dsl_pool_sync_context(dp
))
432 dp
->dp_read_overhead
+= gethrtime() - start
;
434 dmu_buf_rele_array(dbp
, nblks
, tag
);
438 /* wait for other io to complete */
440 for (i
= 0; i
< nblks
; i
++) {
441 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
442 mutex_enter(&db
->db_mtx
);
443 while (db
->db_state
== DB_READ
||
444 db
->db_state
== DB_FILL
)
445 cv_wait(&db
->db_changed
, &db
->db_mtx
);
446 if (db
->db_state
== DB_UNCACHED
)
447 err
= SET_ERROR(EIO
);
448 mutex_exit(&db
->db_mtx
);
450 dmu_buf_rele_array(dbp
, nblks
, tag
);
462 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
463 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
468 err
= dnode_hold(os
, object
, FTAG
, &dn
);
472 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
473 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
475 dnode_rele(dn
, FTAG
);
481 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
482 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
484 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
490 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
491 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
498 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
501 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
506 for (i
= 0; i
< numbufs
; i
++) {
508 dbuf_rele(dbp
[i
], tag
);
511 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
515 dmu_prefetch(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t len
)
521 if (zfs_prefetch_disable
)
524 if (len
== 0) { /* they're interested in the bonus buffer */
525 dn
= DMU_META_DNODE(os
);
527 if (object
== 0 || object
>= DN_MAX_OBJECT
)
530 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
531 blkid
= dbuf_whichblock(dn
, object
* sizeof (dnode_phys_t
));
532 dbuf_prefetch(dn
, blkid
);
533 rw_exit(&dn
->dn_struct_rwlock
);
538 * XXX - Note, if the dnode for the requested object is not
539 * already cached, we will do a *synchronous* read in the
540 * dnode_hold() call. The same is true for any indirects.
542 err
= dnode_hold(os
, object
, FTAG
, &dn
);
546 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
547 if (dn
->dn_datablkshift
) {
548 int blkshift
= dn
->dn_datablkshift
;
549 nblks
= (P2ROUNDUP(offset
+len
, 1<<blkshift
) -
550 P2ALIGN(offset
, 1<<blkshift
)) >> blkshift
;
552 nblks
= (offset
< dn
->dn_datablksz
);
556 blkid
= dbuf_whichblock(dn
, offset
);
557 for (i
= 0; i
< nblks
; i
++)
558 dbuf_prefetch(dn
, blkid
+i
);
561 rw_exit(&dn
->dn_struct_rwlock
);
563 dnode_rele(dn
, FTAG
);
567 * Get the next "chunk" of file data to free. We traverse the file from
568 * the end so that the file gets shorter over time (if we crashes in the
569 * middle, this will leave us in a better state). We find allocated file
570 * data by simply searching the allocated level 1 indirects.
573 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t limit
)
575 uint64_t len
= *start
- limit
;
577 uint64_t maxblks
= DMU_MAX_ACCESS
/ (1ULL << (dn
->dn_indblkshift
+ 1));
579 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
581 ASSERT(limit
<= *start
);
583 if (len
<= iblkrange
* maxblks
) {
587 ASSERT(ISP2(iblkrange
));
589 while (*start
> limit
&& blkcnt
< maxblks
) {
592 /* find next allocated L1 indirect */
593 err
= dnode_next_offset(dn
,
594 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
596 /* if there are no more, then we are done */
605 /* reset offset to end of "next" block back */
606 *start
= P2ALIGN(*start
, iblkrange
);
616 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
617 uint64_t length
, boolean_t free_dnode
)
620 uint64_t object_size
, start
, end
, len
;
621 boolean_t trunc
= (length
== DMU_OBJECT_END
);
624 align
= 1 << dn
->dn_datablkshift
;
626 object_size
= align
== 1 ? dn
->dn_datablksz
:
627 (dn
->dn_maxblkid
+ 1) << dn
->dn_datablkshift
;
629 end
= offset
+ length
;
630 if (trunc
|| end
> object_size
)
634 length
= end
- offset
;
638 /* assert(offset <= start) */
639 err
= get_next_chunk(dn
, &start
, offset
);
642 len
= trunc
? DMU_OBJECT_END
: end
- start
;
644 tx
= dmu_tx_create(os
);
645 dmu_tx_hold_free(tx
, dn
->dn_object
, start
, len
);
646 err
= dmu_tx_assign(tx
, TXG_WAIT
);
652 dnode_free_range(dn
, start
, trunc
? -1 : len
, tx
);
654 if (start
== 0 && free_dnode
) {
659 length
-= end
- start
;
668 dmu_free_long_range(objset_t
*os
, uint64_t object
,
669 uint64_t offset
, uint64_t length
)
674 err
= dnode_hold(os
, object
, FTAG
, &dn
);
677 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
, FALSE
);
678 dnode_rele(dn
, FTAG
);
683 dmu_free_object(objset_t
*os
, uint64_t object
)
689 err
= dnode_hold_impl(os
, object
, DNODE_MUST_BE_ALLOCATED
,
693 if (dn
->dn_nlevels
== 1) {
694 tx
= dmu_tx_create(os
);
695 dmu_tx_hold_bonus(tx
, object
);
696 dmu_tx_hold_free(tx
, dn
->dn_object
, 0, DMU_OBJECT_END
);
697 err
= dmu_tx_assign(tx
, TXG_WAIT
);
699 dnode_free_range(dn
, 0, DMU_OBJECT_END
, tx
);
706 err
= dmu_free_long_range_impl(os
, dn
, 0, DMU_OBJECT_END
, TRUE
);
708 dnode_rele(dn
, FTAG
);
713 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
714 uint64_t size
, dmu_tx_t
*tx
)
717 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
720 ASSERT(offset
< UINT64_MAX
);
721 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
722 dnode_free_range(dn
, offset
, size
, tx
);
723 dnode_rele(dn
, FTAG
);
728 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
729 void *buf
, uint32_t flags
)
735 err
= dnode_hold(os
, object
, FTAG
, &dn
);
740 * Deal with odd block sizes, where there can't be data past the first
741 * block. If we ever do the tail block optimization, we will need to
742 * handle that here as well.
744 if (dn
->dn_maxblkid
== 0) {
745 int newsz
= offset
> dn
->dn_datablksz
? 0 :
746 MIN(size
, dn
->dn_datablksz
- offset
);
747 bzero((char *)buf
+ newsz
, size
- newsz
);
752 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
756 * NB: we could do this block-at-a-time, but it's nice
757 * to be reading in parallel.
759 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
760 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
764 for (i
= 0; i
< numbufs
; i
++) {
767 dmu_buf_t
*db
= dbp
[i
];
771 bufoff
= offset
- db
->db_offset
;
772 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
774 bcopy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
778 buf
= (char *)buf
+ tocpy
;
780 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
782 dnode_rele(dn
, FTAG
);
787 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
788 const void *buf
, dmu_tx_t
*tx
)
796 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
797 FALSE
, FTAG
, &numbufs
, &dbp
));
799 for (i
= 0; i
< numbufs
; i
++) {
802 dmu_buf_t
*db
= dbp
[i
];
806 bufoff
= offset
- db
->db_offset
;
807 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
809 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
811 if (tocpy
== db
->db_size
)
812 dmu_buf_will_fill(db
, tx
);
814 dmu_buf_will_dirty(db
, tx
);
816 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
818 if (tocpy
== db
->db_size
)
819 dmu_buf_fill_done(db
, tx
);
823 buf
= (char *)buf
+ tocpy
;
825 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
829 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
838 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
839 FALSE
, FTAG
, &numbufs
, &dbp
));
841 for (i
= 0; i
< numbufs
; i
++) {
842 dmu_buf_t
*db
= dbp
[i
];
844 dmu_buf_will_not_fill(db
, tx
);
846 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
850 * DMU support for xuio
852 kstat_t
*xuio_ksp
= NULL
;
854 typedef struct xuio_stats
{
855 /* loaned yet not returned arc_buf */
856 kstat_named_t xuiostat_onloan_rbuf
;
857 kstat_named_t xuiostat_onloan_wbuf
;
858 /* whether a copy is made when loaning out a read buffer */
859 kstat_named_t xuiostat_rbuf_copied
;
860 kstat_named_t xuiostat_rbuf_nocopy
;
861 /* whether a copy is made when assigning a write buffer */
862 kstat_named_t xuiostat_wbuf_copied
;
863 kstat_named_t xuiostat_wbuf_nocopy
;
866 static xuio_stats_t xuio_stats
= {
867 { "onloan_read_buf", KSTAT_DATA_UINT64
},
868 { "onloan_write_buf", KSTAT_DATA_UINT64
},
869 { "read_buf_copied", KSTAT_DATA_UINT64
},
870 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
871 { "write_buf_copied", KSTAT_DATA_UINT64
},
872 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
875 #define XUIOSTAT_INCR(stat, val) \
876 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
877 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
880 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
883 uio_t
*uio
= &xuio
->xu_uio
;
885 uio
->uio_iovcnt
= nblk
;
886 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_PUSHPAGE
);
888 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_PUSHPAGE
);
890 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_PUSHPAGE
);
891 priv
->iovp
= uio
->uio_iov
;
892 XUIO_XUZC_PRIV(xuio
) = priv
;
894 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
895 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
897 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
903 dmu_xuio_fini(xuio_t
*xuio
)
905 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
906 int nblk
= priv
->cnt
;
908 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
909 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
910 kmem_free(priv
, sizeof (dmu_xuio_t
));
912 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
913 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
915 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
919 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
920 * and increase priv->next by 1.
923 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
926 uio_t
*uio
= &xuio
->xu_uio
;
927 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
928 int i
= priv
->next
++;
930 ASSERT(i
< priv
->cnt
);
931 ASSERT(off
+ n
<= arc_buf_size(abuf
));
932 iov
= uio
->uio_iov
+ i
;
933 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
935 priv
->bufs
[i
] = abuf
;
940 dmu_xuio_cnt(xuio_t
*xuio
)
942 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
947 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
949 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
951 ASSERT(i
< priv
->cnt
);
952 return (priv
->bufs
[i
]);
956 dmu_xuio_clear(xuio_t
*xuio
, int i
)
958 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
960 ASSERT(i
< priv
->cnt
);
961 priv
->bufs
[i
] = NULL
;
967 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
968 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
970 if (xuio_ksp
!= NULL
) {
971 xuio_ksp
->ks_data
= &xuio_stats
;
972 kstat_install(xuio_ksp
);
979 if (xuio_ksp
!= NULL
) {
980 kstat_delete(xuio_ksp
);
986 xuio_stat_wbuf_copied()
988 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
992 xuio_stat_wbuf_nocopy()
994 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1000 * Copy up to size bytes between arg_buf and req based on the data direction
1001 * described by the req. If an entire req's data cannot be transfered the
1002 * req's is updated such that it's current index and bv offsets correctly
1003 * reference any residual data which could not be copied. The return value
1004 * is the number of bytes successfully copied to arg_buf.
1007 dmu_req_copy(void *arg_buf
, int size
, int *offset
, struct request
*req
)
1010 struct req_iterator iter
;
1015 rq_for_each_segment(bv
, req
, iter
) {
1017 /* Fully consumed the passed arg_buf */
1018 ASSERT3S(*offset
, <=, size
);
1019 if (size
== *offset
)
1022 /* Skip fully consumed bv's */
1023 if (bv
->bv_len
== 0)
1026 tocpy
= MIN(bv
->bv_len
, size
- *offset
);
1027 ASSERT3S(tocpy
, >=, 0);
1029 bv_buf
= page_address(bv
->bv_page
) + bv
->bv_offset
;
1030 ASSERT3P(bv_buf
, !=, NULL
);
1032 if (rq_data_dir(req
) == WRITE
)
1033 memcpy(arg_buf
+ *offset
, bv_buf
, tocpy
);
1035 memcpy(bv_buf
, arg_buf
+ *offset
, tocpy
);
1038 bv
->bv_offset
+= tocpy
;
1039 bv
->bv_len
-= tocpy
;
1046 dmu_bio_put(struct bio
*bio
)
1048 struct bio
*bio_next
;
1051 bio_next
= bio
->bi_next
;
1058 dmu_bio_clone(struct bio
*bio
, struct bio
**bio_copy
)
1060 struct bio
*bio_root
= NULL
;
1061 struct bio
*bio_last
= NULL
;
1062 struct bio
*bio_new
;
1068 bio_new
= bio_clone(bio
, GFP_NOIO
);
1069 if (bio_new
== NULL
) {
1070 dmu_bio_put(bio_root
);
1075 bio_last
->bi_next
= bio_new
;
1085 *bio_copy
= bio_root
;
1091 dmu_read_req(objset_t
*os
, uint64_t object
, struct request
*req
)
1093 uint64_t size
= blk_rq_bytes(req
);
1094 uint64_t offset
= blk_rq_pos(req
) << 9;
1095 struct bio
*bio_saved
= req
->bio
;
1097 int numbufs
, i
, err
;
1100 * NB: we could do this block-at-a-time, but it's nice
1101 * to be reading in parallel.
1103 err
= dmu_buf_hold_array(os
, object
, offset
, size
, TRUE
, FTAG
,
1109 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1110 * can be safely modified. The original bio list is relinked in to
1111 * the request when the function exits. This is required because
1112 * some file systems blindly assume that these values will remain
1113 * constant between bio_submit() and the IO completion callback.
1115 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1119 for (i
= 0; i
< numbufs
; i
++) {
1120 int tocpy
, didcpy
, bufoff
;
1121 dmu_buf_t
*db
= dbp
[i
];
1123 bufoff
= offset
- db
->db_offset
;
1124 ASSERT3S(bufoff
, >=, 0);
1126 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1130 err
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, &didcpy
, req
);
1143 dmu_bio_put(req
->bio
);
1144 req
->bio
= bio_saved
;
1146 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1152 dmu_write_req(objset_t
*os
, uint64_t object
, struct request
*req
, dmu_tx_t
*tx
)
1154 uint64_t size
= blk_rq_bytes(req
);
1155 uint64_t offset
= blk_rq_pos(req
) << 9;
1156 struct bio
*bio_saved
= req
->bio
;
1165 err
= dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1171 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1172 * can be safely modified. The original bio list is relinked in to
1173 * the request when the function exits. This is required because
1174 * some file systems blindly assume that these values will remain
1175 * constant between bio_submit() and the IO completion callback.
1177 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1181 for (i
= 0; i
< numbufs
; i
++) {
1182 int tocpy
, didcpy
, bufoff
;
1183 dmu_buf_t
*db
= dbp
[i
];
1185 bufoff
= offset
- db
->db_offset
;
1186 ASSERT3S(bufoff
, >=, 0);
1188 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1192 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1194 if (tocpy
== db
->db_size
)
1195 dmu_buf_will_fill(db
, tx
);
1197 dmu_buf_will_dirty(db
, tx
);
1199 err
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, &didcpy
, req
);
1201 if (tocpy
== db
->db_size
)
1202 dmu_buf_fill_done(db
, tx
);
1215 dmu_bio_put(req
->bio
);
1216 req
->bio
= bio_saved
;
1218 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1224 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1227 int numbufs
, i
, err
;
1228 xuio_t
*xuio
= NULL
;
1231 * NB: we could do this block-at-a-time, but it's nice
1232 * to be reading in parallel.
1234 err
= dmu_buf_hold_array(os
, object
, uio
->uio_loffset
, size
, TRUE
, FTAG
,
1239 for (i
= 0; i
< numbufs
; i
++) {
1242 dmu_buf_t
*db
= dbp
[i
];
1246 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1247 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1250 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1251 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1252 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1253 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1255 uio
->uio_resid
-= tocpy
;
1256 uio
->uio_loffset
+= tocpy
;
1259 if (abuf
== dbuf_abuf
)
1260 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1262 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1264 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1272 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1278 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1285 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1286 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1290 for (i
= 0; i
< numbufs
; i
++) {
1293 dmu_buf_t
*db
= dbp
[i
];
1297 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1298 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1300 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1302 if (tocpy
== db
->db_size
)
1303 dmu_buf_will_fill(db
, tx
);
1305 dmu_buf_will_dirty(db
, tx
);
1308 * XXX uiomove could block forever (eg.nfs-backed
1309 * pages). There needs to be a uiolockdown() function
1310 * to lock the pages in memory, so that uiomove won't
1313 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1316 if (tocpy
== db
->db_size
)
1317 dmu_buf_fill_done(db
, tx
);
1325 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1330 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1333 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1342 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1349 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1358 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1362 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1364 dnode_rele(dn
, FTAG
);
1368 #endif /* _KERNEL */
1371 * Allocate a loaned anonymous arc buffer.
1374 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1376 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1379 DB_GET_SPA(&spa
, db
);
1380 return (arc_loan_buf(spa
, size
));
1384 * Free a loaned arc buffer.
1387 dmu_return_arcbuf(arc_buf_t
*buf
)
1389 arc_return_buf(buf
, FTAG
);
1390 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1394 * When possible directly assign passed loaned arc buffer to a dbuf.
1395 * If this is not possible copy the contents of passed arc buf via
1399 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1402 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1405 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1408 DB_DNODE_ENTER(dbuf
);
1409 dn
= DB_DNODE(dbuf
);
1410 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1411 blkid
= dbuf_whichblock(dn
, offset
);
1412 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1413 rw_exit(&dn
->dn_struct_rwlock
);
1414 DB_DNODE_EXIT(dbuf
);
1416 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1417 dbuf_assign_arcbuf(db
, buf
, tx
);
1418 dbuf_rele(db
, FTAG
);
1423 DB_DNODE_ENTER(dbuf
);
1424 dn
= DB_DNODE(dbuf
);
1426 object
= dn
->dn_object
;
1427 DB_DNODE_EXIT(dbuf
);
1429 dbuf_rele(db
, FTAG
);
1430 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1431 dmu_return_arcbuf(buf
);
1432 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1437 dbuf_dirty_record_t
*dsa_dr
;
1438 dmu_sync_cb_t
*dsa_done
;
1445 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1447 dmu_sync_arg_t
*dsa
= varg
;
1448 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1449 blkptr_t
*bp
= zio
->io_bp
;
1451 if (zio
->io_error
== 0) {
1452 if (BP_IS_HOLE(bp
)) {
1454 * A block of zeros may compress to a hole, but the
1455 * block size still needs to be known for replay.
1457 BP_SET_LSIZE(bp
, db
->db_size
);
1459 ASSERT(BP_GET_LEVEL(bp
) == 0);
1466 dmu_sync_late_arrival_ready(zio_t
*zio
)
1468 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1473 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1475 dmu_sync_arg_t
*dsa
= varg
;
1476 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1477 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1479 mutex_enter(&db
->db_mtx
);
1480 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1481 if (zio
->io_error
== 0) {
1482 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1483 if (dr
->dt
.dl
.dr_nopwrite
) {
1484 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1485 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1486 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1488 ASSERT(BP_EQUAL(bp
, bp_orig
));
1489 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1490 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1492 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1493 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1494 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1495 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
))
1496 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1498 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1500 cv_broadcast(&db
->db_changed
);
1501 mutex_exit(&db
->db_mtx
);
1503 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1505 kmem_free(dsa
, sizeof (*dsa
));
1509 dmu_sync_late_arrival_done(zio_t
*zio
)
1511 blkptr_t
*bp
= zio
->io_bp
;
1512 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1513 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1515 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1517 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1518 * then there is nothing to do here. Otherwise, free the
1519 * newly allocated block in this txg.
1521 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1522 ASSERT(BP_EQUAL(bp
, bp_orig
));
1524 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1525 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1526 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1527 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1531 dmu_tx_commit(dsa
->dsa_tx
);
1533 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1535 kmem_free(dsa
, sizeof (*dsa
));
1539 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1540 zio_prop_t
*zp
, zbookmark_t
*zb
)
1542 dmu_sync_arg_t
*dsa
;
1545 tx
= dmu_tx_create(os
);
1546 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1547 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1549 /* Make zl_get_data do txg_waited_synced() */
1550 return (SET_ERROR(EIO
));
1553 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1555 dsa
->dsa_done
= done
;
1559 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1560 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1561 dmu_sync_late_arrival_ready
, dmu_sync_late_arrival_done
, dsa
,
1562 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_FASTWRITE
, zb
));
1568 * Intent log support: sync the block associated with db to disk.
1569 * N.B. and XXX: the caller is responsible for making sure that the
1570 * data isn't changing while dmu_sync() is writing it.
1574 * EEXIST: this txg has already been synced, so there's nothing to do.
1575 * The caller should not log the write.
1577 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1578 * The caller should not log the write.
1580 * EALREADY: this block is already in the process of being synced.
1581 * The caller should track its progress (somehow).
1583 * EIO: could not do the I/O.
1584 * The caller should do a txg_wait_synced().
1586 * 0: the I/O has been initiated.
1587 * The caller should log this blkptr in the done callback.
1588 * It is possible that the I/O will fail, in which case
1589 * the error will be reported to the done callback and
1590 * propagated to pio from zio_done().
1593 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1595 blkptr_t
*bp
= zgd
->zgd_bp
;
1596 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1597 objset_t
*os
= db
->db_objset
;
1598 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1599 dbuf_dirty_record_t
*dr
;
1600 dmu_sync_arg_t
*dsa
;
1605 ASSERT(pio
!= NULL
);
1608 SET_BOOKMARK(&zb
, ds
->ds_object
,
1609 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1613 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1617 * If we're frozen (running ziltest), we always need to generate a bp.
1619 if (txg
> spa_freeze_txg(os
->os_spa
))
1620 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1623 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1624 * and us. If we determine that this txg is not yet syncing,
1625 * but it begins to sync a moment later, that's OK because the
1626 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1628 mutex_enter(&db
->db_mtx
);
1630 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1632 * This txg has already synced. There's nothing to do.
1634 mutex_exit(&db
->db_mtx
);
1635 return (SET_ERROR(EEXIST
));
1638 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1640 * This txg is currently syncing, so we can't mess with
1641 * the dirty record anymore; just write a new log block.
1643 mutex_exit(&db
->db_mtx
);
1644 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1647 dr
= db
->db_last_dirty
;
1648 while (dr
&& dr
->dr_txg
!= txg
)
1653 * There's no dr for this dbuf, so it must have been freed.
1654 * There's no need to log writes to freed blocks, so we're done.
1656 mutex_exit(&db
->db_mtx
);
1657 return (SET_ERROR(ENOENT
));
1660 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1663 * Assume the on-disk data is X, the current syncing data is Y,
1664 * and the current in-memory data is Z (currently in dmu_sync).
1665 * X and Z are identical but Y is has been modified. Normally,
1666 * when X and Z are the same we will perform a nopwrite but if Y
1667 * is different we must disable nopwrite since the resulting write
1668 * of Y to disk can free the block containing X. If we allowed a
1669 * nopwrite to occur the block pointing to Z would reference a freed
1670 * block. Since this is a rare case we simplify this by disabling
1671 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1672 * a previous transaction.
1675 zp
.zp_nopwrite
= B_FALSE
;
1677 ASSERT(dr
->dr_txg
== txg
);
1678 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1679 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1681 * We have already issued a sync write for this buffer,
1682 * or this buffer has already been synced. It could not
1683 * have been dirtied since, or we would have cleared the state.
1685 mutex_exit(&db
->db_mtx
);
1686 return (SET_ERROR(EALREADY
));
1689 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1690 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1691 mutex_exit(&db
->db_mtx
);
1693 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1695 dsa
->dsa_done
= done
;
1699 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1700 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1701 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
, dmu_sync_done
,
1702 dsa
, ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_FASTWRITE
, &zb
));
1708 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1714 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1717 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1718 dnode_rele(dn
, FTAG
);
1723 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1728 /* XXX assumes dnode_hold will not get an i/o error */
1729 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1730 ASSERT(checksum
< ZIO_CHECKSUM_FUNCTIONS
);
1731 dn
->dn_checksum
= checksum
;
1732 dnode_setdirty(dn
, tx
);
1733 dnode_rele(dn
, FTAG
);
1737 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1742 /* XXX assumes dnode_hold will not get an i/o error */
1743 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1744 ASSERT(compress
< ZIO_COMPRESS_FUNCTIONS
);
1745 dn
->dn_compress
= compress
;
1746 dnode_setdirty(dn
, tx
);
1747 dnode_rele(dn
, FTAG
);
1750 int zfs_mdcomp_disable
= 0;
1753 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1755 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1756 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1758 enum zio_checksum checksum
= os
->os_checksum
;
1759 enum zio_compress compress
= os
->os_compress
;
1760 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1761 boolean_t dedup
= B_FALSE
;
1762 boolean_t nopwrite
= B_FALSE
;
1763 boolean_t dedup_verify
= os
->os_dedup_verify
;
1764 int copies
= os
->os_copies
;
1767 * We maintain different write policies for each of the following
1770 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1771 * 3. all other level 0 blocks
1775 * XXX -- we should design a compression algorithm
1776 * that specializes in arrays of bps.
1778 compress
= zfs_mdcomp_disable
? ZIO_COMPRESS_EMPTY
:
1782 * Metadata always gets checksummed. If the data
1783 * checksum is multi-bit correctable, and it's not a
1784 * ZBT-style checksum, then it's suitable for metadata
1785 * as well. Otherwise, the metadata checksum defaults
1788 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1789 zio_checksum_table
[checksum
].ci_eck
)
1790 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1791 } else if (wp
& WP_NOFILL
) {
1795 * If we're writing preallocated blocks, we aren't actually
1796 * writing them so don't set any policy properties. These
1797 * blocks are currently only used by an external subsystem
1798 * outside of zfs (i.e. dump) and not written by the zio
1801 compress
= ZIO_COMPRESS_OFF
;
1802 checksum
= ZIO_CHECKSUM_OFF
;
1804 compress
= zio_compress_select(dn
->dn_compress
, compress
);
1806 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1807 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1811 * Determine dedup setting. If we are in dmu_sync(),
1812 * we won't actually dedup now because that's all
1813 * done in syncing context; but we do want to use the
1814 * dedup checkum. If the checksum is not strong
1815 * enough to ensure unique signatures, force
1818 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1819 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1820 if (!zio_checksum_table
[checksum
].ci_dedup
)
1821 dedup_verify
= B_TRUE
;
1825 * Enable nopwrite if we have a cryptographically secure
1826 * checksum that has no known collisions (i.e. SHA-256)
1827 * and compression is enabled. We don't enable nopwrite if
1828 * dedup is enabled as the two features are mutually exclusive.
1830 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1831 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1834 zp
->zp_checksum
= checksum
;
1835 zp
->zp_compress
= compress
;
1836 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1837 zp
->zp_level
= level
;
1838 zp
->zp_copies
= MIN(copies
+ ismd
, spa_max_replication(os
->os_spa
));
1839 zp
->zp_dedup
= dedup
;
1840 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1841 zp
->zp_nopwrite
= nopwrite
;
1845 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1850 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1854 * Sync any current changes before
1855 * we go trundling through the block pointers.
1857 for (i
= 0; i
< TXG_SIZE
; i
++) {
1858 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1861 if (i
!= TXG_SIZE
) {
1862 dnode_rele(dn
, FTAG
);
1863 txg_wait_synced(dmu_objset_pool(os
), 0);
1864 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1869 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1870 dnode_rele(dn
, FTAG
);
1876 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1878 dnode_phys_t
*dnp
= dn
->dn_phys
;
1881 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1882 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1883 1ULL << dn
->dn_indblkshift
: 0;
1884 doi
->doi_type
= dn
->dn_type
;
1885 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1886 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1887 doi
->doi_indirection
= dn
->dn_nlevels
;
1888 doi
->doi_checksum
= dn
->dn_checksum
;
1889 doi
->doi_compress
= dn
->dn_compress
;
1890 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1891 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1892 doi
->doi_fill_count
= 0;
1893 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1894 doi
->doi_fill_count
+= dnp
->dn_blkptr
[i
].blk_fill
;
1898 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1900 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1901 mutex_enter(&dn
->dn_mtx
);
1903 __dmu_object_info_from_dnode(dn
, doi
);
1905 mutex_exit(&dn
->dn_mtx
);
1906 rw_exit(&dn
->dn_struct_rwlock
);
1910 * Get information on a DMU object.
1911 * If doi is NULL, just indicates whether the object exists.
1914 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1917 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1923 dmu_object_info_from_dnode(dn
, doi
);
1925 dnode_rele(dn
, FTAG
);
1930 * As above, but faster; can be used when you have a held dbuf in hand.
1933 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1935 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1938 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1943 * Faster still when you only care about the size.
1944 * This is specifically optimized for zfs_getattr().
1947 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1948 u_longlong_t
*nblk512
)
1950 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1956 *blksize
= dn
->dn_datablksz
;
1957 /* add 1 for dnode space */
1958 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1959 SPA_MINBLOCKSHIFT
) + 1;
1964 byteswap_uint64_array(void *vbuf
, size_t size
)
1966 uint64_t *buf
= vbuf
;
1967 size_t count
= size
>> 3;
1970 ASSERT((size
& 7) == 0);
1972 for (i
= 0; i
< count
; i
++)
1973 buf
[i
] = BSWAP_64(buf
[i
]);
1977 byteswap_uint32_array(void *vbuf
, size_t size
)
1979 uint32_t *buf
= vbuf
;
1980 size_t count
= size
>> 2;
1983 ASSERT((size
& 3) == 0);
1985 for (i
= 0; i
< count
; i
++)
1986 buf
[i
] = BSWAP_32(buf
[i
]);
1990 byteswap_uint16_array(void *vbuf
, size_t size
)
1992 uint16_t *buf
= vbuf
;
1993 size_t count
= size
>> 1;
1996 ASSERT((size
& 1) == 0);
1998 for (i
= 0; i
< count
; i
++)
1999 buf
[i
] = BSWAP_16(buf
[i
]);
2004 byteswap_uint8_array(void *vbuf
, size_t size
)
2026 arc_fini(); /* arc depends on l2arc, so arc must go first */
2038 #if defined(_KERNEL) && defined(HAVE_SPL)
2039 EXPORT_SYMBOL(dmu_bonus_hold
);
2040 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2041 EXPORT_SYMBOL(dmu_buf_rele_array
);
2042 EXPORT_SYMBOL(dmu_prefetch
);
2043 EXPORT_SYMBOL(dmu_free_range
);
2044 EXPORT_SYMBOL(dmu_free_long_range
);
2045 EXPORT_SYMBOL(dmu_free_object
);
2046 EXPORT_SYMBOL(dmu_read
);
2047 EXPORT_SYMBOL(dmu_write
);
2048 EXPORT_SYMBOL(dmu_prealloc
);
2049 EXPORT_SYMBOL(dmu_object_info
);
2050 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2051 EXPORT_SYMBOL(dmu_object_info_from_db
);
2052 EXPORT_SYMBOL(dmu_object_size_from_db
);
2053 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2054 EXPORT_SYMBOL(dmu_object_set_checksum
);
2055 EXPORT_SYMBOL(dmu_object_set_compress
);
2056 EXPORT_SYMBOL(dmu_write_policy
);
2057 EXPORT_SYMBOL(dmu_sync
);
2058 EXPORT_SYMBOL(dmu_request_arcbuf
);
2059 EXPORT_SYMBOL(dmu_return_arcbuf
);
2060 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2061 EXPORT_SYMBOL(dmu_buf_hold
);
2062 EXPORT_SYMBOL(dmu_ot
);
2064 module_param(zfs_mdcomp_disable
, int, 0644);
2065 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2067 module_param(zfs_nopwrite_enabled
, int, 0644);
2068 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");