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
,
404 KM_PUSHPAGE
| KM_NODEBUG
);
406 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
407 blkid
= dbuf_whichblock(dn
, offset
);
408 for (i
= 0; i
< nblks
; i
++) {
409 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+i
, tag
);
411 rw_exit(&dn
->dn_struct_rwlock
);
412 dmu_buf_rele_array(dbp
, nblks
, tag
);
414 return (SET_ERROR(EIO
));
416 /* initiate async i/o */
418 (void) dbuf_read(db
, zio
, dbuf_flags
);
422 rw_exit(&dn
->dn_struct_rwlock
);
424 /* wait for async i/o */
427 dmu_buf_rele_array(dbp
, nblks
, tag
);
431 /* wait for other io to complete */
433 for (i
= 0; i
< nblks
; i
++) {
434 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
435 mutex_enter(&db
->db_mtx
);
436 while (db
->db_state
== DB_READ
||
437 db
->db_state
== DB_FILL
)
438 cv_wait(&db
->db_changed
, &db
->db_mtx
);
439 if (db
->db_state
== DB_UNCACHED
)
440 err
= SET_ERROR(EIO
);
441 mutex_exit(&db
->db_mtx
);
443 dmu_buf_rele_array(dbp
, nblks
, tag
);
455 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
456 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
461 err
= dnode_hold(os
, object
, FTAG
, &dn
);
465 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
466 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
468 dnode_rele(dn
, FTAG
);
474 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
475 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
477 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
483 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
484 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
491 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
494 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
499 for (i
= 0; i
< numbufs
; i
++) {
501 dbuf_rele(dbp
[i
], tag
);
504 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
508 * Issue prefetch i/os for the given blocks.
510 * Note: The assumption is that we *know* these blocks will be needed
511 * almost immediately. Therefore, the prefetch i/os will be issued at
512 * ZIO_PRIORITY_SYNC_READ
514 * Note: indirect blocks and other metadata will be read synchronously,
515 * causing this function to block if they are not already cached.
518 dmu_prefetch(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t len
)
524 if (zfs_prefetch_disable
)
527 if (len
== 0) { /* they're interested in the bonus buffer */
528 dn
= DMU_META_DNODE(os
);
530 if (object
== 0 || object
>= DN_MAX_OBJECT
)
533 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
534 blkid
= dbuf_whichblock(dn
, object
* sizeof (dnode_phys_t
));
535 dbuf_prefetch(dn
, blkid
, ZIO_PRIORITY_SYNC_READ
);
536 rw_exit(&dn
->dn_struct_rwlock
);
541 * XXX - Note, if the dnode for the requested object is not
542 * already cached, we will do a *synchronous* read in the
543 * dnode_hold() call. The same is true for any indirects.
545 err
= dnode_hold(os
, object
, FTAG
, &dn
);
549 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
550 if (dn
->dn_datablkshift
) {
551 int blkshift
= dn
->dn_datablkshift
;
552 nblks
= (P2ROUNDUP(offset
+ len
, 1 << blkshift
) -
553 P2ALIGN(offset
, 1 << blkshift
)) >> blkshift
;
555 nblks
= (offset
< dn
->dn_datablksz
);
561 blkid
= dbuf_whichblock(dn
, offset
);
562 for (i
= 0; i
< nblks
; i
++)
563 dbuf_prefetch(dn
, blkid
+ i
, ZIO_PRIORITY_SYNC_READ
);
566 rw_exit(&dn
->dn_struct_rwlock
);
568 dnode_rele(dn
, FTAG
);
572 * Get the next "chunk" of file data to free. We traverse the file from
573 * the end so that the file gets shorter over time (if we crashes in the
574 * middle, this will leave us in a better state). We find allocated file
575 * data by simply searching the allocated level 1 indirects.
577 * On input, *start should be the first offset that does not need to be
578 * freed (e.g. "offset + length"). On return, *start will be the first
579 * offset that should be freed.
582 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
584 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
585 /* bytes of data covered by a level-1 indirect block */
587 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
590 ASSERT3U(minimum
, <=, *start
);
592 if (*start
- minimum
<= iblkrange
* maxblks
) {
596 ASSERT(ISP2(iblkrange
));
598 for (blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
602 * dnode_next_offset(BACKWARDS) will find an allocated L1
603 * indirect block at or before the input offset. We must
604 * decrement *start so that it is at the end of the region
608 err
= dnode_next_offset(dn
,
609 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
611 /* if there are no indirect blocks before start, we are done */
615 } else if (err
!= 0) {
619 /* set start to the beginning of this L1 indirect */
620 *start
= P2ALIGN(*start
, iblkrange
);
622 if (*start
< minimum
)
628 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
631 uint64_t object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
634 if (offset
>= object_size
)
637 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
638 length
= object_size
- offset
;
640 while (length
!= 0) {
641 uint64_t chunk_end
, chunk_begin
;
644 chunk_end
= chunk_begin
= offset
+ length
;
646 /* move chunk_begin backwards to the beginning of this chunk */
647 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
650 ASSERT3U(chunk_begin
, >=, offset
);
651 ASSERT3U(chunk_begin
, <=, chunk_end
);
653 tx
= dmu_tx_create(os
);
654 dmu_tx_hold_free(tx
, dn
->dn_object
,
655 chunk_begin
, chunk_end
- chunk_begin
);
656 err
= dmu_tx_assign(tx
, TXG_WAIT
);
661 dnode_free_range(dn
, chunk_begin
, chunk_end
- chunk_begin
, tx
);
664 length
-= chunk_end
- chunk_begin
;
670 dmu_free_long_range(objset_t
*os
, uint64_t object
,
671 uint64_t offset
, uint64_t length
)
676 err
= dnode_hold(os
, object
, FTAG
, &dn
);
679 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
682 * It is important to zero out the maxblkid when freeing the entire
683 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
684 * will take the fast path, and (b) dnode_reallocate() can verify
685 * that the entire file has been freed.
687 if (offset
== 0 && length
== DMU_OBJECT_END
)
690 dnode_rele(dn
, FTAG
);
695 dmu_free_long_object(objset_t
*os
, uint64_t object
)
700 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
704 tx
= dmu_tx_create(os
);
705 dmu_tx_hold_bonus(tx
, object
);
706 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
707 err
= dmu_tx_assign(tx
, TXG_WAIT
);
709 err
= dmu_object_free(os
, object
, tx
);
719 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
720 uint64_t size
, dmu_tx_t
*tx
)
723 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
726 ASSERT(offset
< UINT64_MAX
);
727 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
728 dnode_free_range(dn
, offset
, size
, tx
);
729 dnode_rele(dn
, FTAG
);
734 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
735 void *buf
, uint32_t flags
)
741 err
= dnode_hold(os
, object
, FTAG
, &dn
);
746 * Deal with odd block sizes, where there can't be data past the first
747 * block. If we ever do the tail block optimization, we will need to
748 * handle that here as well.
750 if (dn
->dn_maxblkid
== 0) {
751 int newsz
= offset
> dn
->dn_datablksz
? 0 :
752 MIN(size
, dn
->dn_datablksz
- offset
);
753 bzero((char *)buf
+ newsz
, size
- newsz
);
758 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
762 * NB: we could do this block-at-a-time, but it's nice
763 * to be reading in parallel.
765 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
766 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
770 for (i
= 0; i
< numbufs
; i
++) {
773 dmu_buf_t
*db
= dbp
[i
];
777 bufoff
= offset
- db
->db_offset
;
778 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
780 bcopy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
784 buf
= (char *)buf
+ tocpy
;
786 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
788 dnode_rele(dn
, FTAG
);
793 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
794 const void *buf
, dmu_tx_t
*tx
)
802 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
803 FALSE
, FTAG
, &numbufs
, &dbp
));
805 for (i
= 0; i
< numbufs
; i
++) {
808 dmu_buf_t
*db
= dbp
[i
];
812 bufoff
= offset
- db
->db_offset
;
813 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
815 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
817 if (tocpy
== db
->db_size
)
818 dmu_buf_will_fill(db
, tx
);
820 dmu_buf_will_dirty(db
, tx
);
822 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
824 if (tocpy
== db
->db_size
)
825 dmu_buf_fill_done(db
, tx
);
829 buf
= (char *)buf
+ tocpy
;
831 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
835 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
844 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
845 FALSE
, FTAG
, &numbufs
, &dbp
));
847 for (i
= 0; i
< numbufs
; i
++) {
848 dmu_buf_t
*db
= dbp
[i
];
850 dmu_buf_will_not_fill(db
, tx
);
852 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
856 * DMU support for xuio
858 kstat_t
*xuio_ksp
= NULL
;
860 typedef struct xuio_stats
{
861 /* loaned yet not returned arc_buf */
862 kstat_named_t xuiostat_onloan_rbuf
;
863 kstat_named_t xuiostat_onloan_wbuf
;
864 /* whether a copy is made when loaning out a read buffer */
865 kstat_named_t xuiostat_rbuf_copied
;
866 kstat_named_t xuiostat_rbuf_nocopy
;
867 /* whether a copy is made when assigning a write buffer */
868 kstat_named_t xuiostat_wbuf_copied
;
869 kstat_named_t xuiostat_wbuf_nocopy
;
872 static xuio_stats_t xuio_stats
= {
873 { "onloan_read_buf", KSTAT_DATA_UINT64
},
874 { "onloan_write_buf", KSTAT_DATA_UINT64
},
875 { "read_buf_copied", KSTAT_DATA_UINT64
},
876 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
877 { "write_buf_copied", KSTAT_DATA_UINT64
},
878 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
881 #define XUIOSTAT_INCR(stat, val) \
882 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
883 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
886 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
889 uio_t
*uio
= &xuio
->xu_uio
;
891 uio
->uio_iovcnt
= nblk
;
892 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_PUSHPAGE
);
894 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_PUSHPAGE
);
896 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_PUSHPAGE
);
897 priv
->iovp
= uio
->uio_iov
;
898 XUIO_XUZC_PRIV(xuio
) = priv
;
900 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
901 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
903 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
909 dmu_xuio_fini(xuio_t
*xuio
)
911 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
912 int nblk
= priv
->cnt
;
914 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
915 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
916 kmem_free(priv
, sizeof (dmu_xuio_t
));
918 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
919 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
921 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
925 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
926 * and increase priv->next by 1.
929 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
932 uio_t
*uio
= &xuio
->xu_uio
;
933 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
934 int i
= priv
->next
++;
936 ASSERT(i
< priv
->cnt
);
937 ASSERT(off
+ n
<= arc_buf_size(abuf
));
938 iov
= uio
->uio_iov
+ i
;
939 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
941 priv
->bufs
[i
] = abuf
;
946 dmu_xuio_cnt(xuio_t
*xuio
)
948 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
953 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
955 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
957 ASSERT(i
< priv
->cnt
);
958 return (priv
->bufs
[i
]);
962 dmu_xuio_clear(xuio_t
*xuio
, int i
)
964 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
966 ASSERT(i
< priv
->cnt
);
967 priv
->bufs
[i
] = NULL
;
973 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
974 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
976 if (xuio_ksp
!= NULL
) {
977 xuio_ksp
->ks_data
= &xuio_stats
;
978 kstat_install(xuio_ksp
);
985 if (xuio_ksp
!= NULL
) {
986 kstat_delete(xuio_ksp
);
992 xuio_stat_wbuf_copied()
994 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
998 xuio_stat_wbuf_nocopy()
1000 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1006 * Copy up to size bytes between arg_buf and req based on the data direction
1007 * described by the req. If an entire req's data cannot be transfered in one
1008 * pass, you should pass in @req_offset to indicate where to continue. The
1009 * return value is the number of bytes successfully copied to arg_buf.
1012 dmu_req_copy(void *arg_buf
, int size
, struct request
*req
, size_t req_offset
)
1015 struct req_iterator iter
;
1017 int tocpy
, bv_len
, bv_offset
;
1020 rq_for_each_segment(bv
, req
, iter
) {
1022 * Fully consumed the passed arg_buf. We use goto here because
1023 * rq_for_each_segment is a double loop
1025 ASSERT3S(offset
, <=, size
);
1029 /* Skip already copied bv */
1030 if (req_offset
>= bv
->bv_len
) {
1031 req_offset
-= bv
->bv_len
;
1035 bv_len
= bv
->bv_len
- req_offset
;
1036 bv_offset
= bv
->bv_offset
+ req_offset
;
1039 tocpy
= MIN(bv_len
, size
- offset
);
1040 ASSERT3S(tocpy
, >=, 0);
1042 bv_buf
= page_address(bv
->bv_page
) + bv_offset
;
1043 ASSERT3P(bv_buf
, !=, NULL
);
1045 if (rq_data_dir(req
) == WRITE
)
1046 memcpy(arg_buf
+ offset
, bv_buf
, tocpy
);
1048 memcpy(bv_buf
, arg_buf
+ offset
, tocpy
);
1057 dmu_bio_put(struct bio
*bio
)
1059 struct bio
*bio_next
;
1062 bio_next
= bio
->bi_next
;
1069 dmu_bio_clone(struct bio
*bio
, struct bio
**bio_copy
)
1071 struct bio
*bio_root
= NULL
;
1072 struct bio
*bio_last
= NULL
;
1073 struct bio
*bio_new
;
1079 bio_new
= bio_clone(bio
, GFP_NOIO
);
1080 if (bio_new
== NULL
) {
1081 dmu_bio_put(bio_root
);
1086 bio_last
->bi_next
= bio_new
;
1096 *bio_copy
= bio_root
;
1102 dmu_read_req(objset_t
*os
, uint64_t object
, struct request
*req
)
1104 uint64_t size
= blk_rq_bytes(req
);
1105 uint64_t offset
= blk_rq_pos(req
) << 9;
1106 struct bio
*bio_saved
= req
->bio
;
1108 int numbufs
, i
, err
;
1112 * NB: we could do this block-at-a-time, but it's nice
1113 * to be reading in parallel.
1115 err
= dmu_buf_hold_array(os
, object
, offset
, size
, TRUE
, FTAG
,
1121 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1122 * can be safely modified. The original bio list is relinked in to
1123 * the request when the function exits. This is required because
1124 * some file systems blindly assume that these values will remain
1125 * constant between bio_submit() and the IO completion callback.
1127 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1132 for (i
= 0; i
< numbufs
; i
++) {
1133 int tocpy
, didcpy
, bufoff
;
1134 dmu_buf_t
*db
= dbp
[i
];
1136 bufoff
= offset
- db
->db_offset
;
1137 ASSERT3S(bufoff
, >=, 0);
1139 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1143 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1154 req_offset
+= didcpy
;
1158 dmu_bio_put(req
->bio
);
1159 req
->bio
= bio_saved
;
1161 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1167 dmu_write_req(objset_t
*os
, uint64_t object
, struct request
*req
, dmu_tx_t
*tx
)
1169 uint64_t size
= blk_rq_bytes(req
);
1170 uint64_t offset
= blk_rq_pos(req
) << 9;
1171 struct bio
*bio_saved
= req
->bio
;
1173 int numbufs
, i
, err
;
1179 err
= dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1185 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1186 * can be safely modified. The original bio list is relinked in to
1187 * the request when the function exits. This is required because
1188 * some file systems blindly assume that these values will remain
1189 * constant between bio_submit() and the IO completion callback.
1191 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1196 for (i
= 0; i
< numbufs
; i
++) {
1197 int tocpy
, didcpy
, bufoff
;
1198 dmu_buf_t
*db
= dbp
[i
];
1200 bufoff
= offset
- db
->db_offset
;
1201 ASSERT3S(bufoff
, >=, 0);
1203 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1207 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1209 if (tocpy
== db
->db_size
)
1210 dmu_buf_will_fill(db
, tx
);
1212 dmu_buf_will_dirty(db
, tx
);
1214 didcpy
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, req
,
1217 if (tocpy
== db
->db_size
)
1218 dmu_buf_fill_done(db
, tx
);
1228 req_offset
+= didcpy
;
1232 dmu_bio_put(req
->bio
);
1233 req
->bio
= bio_saved
;
1235 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1241 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1244 int numbufs
, i
, err
;
1245 xuio_t
*xuio
= NULL
;
1248 * NB: we could do this block-at-a-time, but it's nice
1249 * to be reading in parallel.
1251 err
= dmu_buf_hold_array(os
, object
, uio
->uio_loffset
, size
, TRUE
, FTAG
,
1256 for (i
= 0; i
< numbufs
; i
++) {
1259 dmu_buf_t
*db
= dbp
[i
];
1263 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1264 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1267 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1268 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1269 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1270 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1272 uio
->uio_resid
-= tocpy
;
1273 uio
->uio_loffset
+= tocpy
;
1276 if (abuf
== dbuf_abuf
)
1277 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1279 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1281 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1289 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1295 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1302 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1303 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1307 for (i
= 0; i
< numbufs
; i
++) {
1310 dmu_buf_t
*db
= dbp
[i
];
1314 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1315 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1317 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1319 if (tocpy
== db
->db_size
)
1320 dmu_buf_will_fill(db
, tx
);
1322 dmu_buf_will_dirty(db
, tx
);
1325 * XXX uiomove could block forever (eg.nfs-backed
1326 * pages). There needs to be a uiolockdown() function
1327 * to lock the pages in memory, so that uiomove won't
1330 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1333 if (tocpy
== db
->db_size
)
1334 dmu_buf_fill_done(db
, tx
);
1342 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1347 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1350 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1359 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1366 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1375 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1379 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1381 dnode_rele(dn
, FTAG
);
1385 #endif /* _KERNEL */
1388 * Allocate a loaned anonymous arc buffer.
1391 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1393 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1396 DB_GET_SPA(&spa
, db
);
1397 return (arc_loan_buf(spa
, size
));
1401 * Free a loaned arc buffer.
1404 dmu_return_arcbuf(arc_buf_t
*buf
)
1406 arc_return_buf(buf
, FTAG
);
1407 VERIFY(arc_buf_remove_ref(buf
, FTAG
));
1411 * When possible directly assign passed loaned arc buffer to a dbuf.
1412 * If this is not possible copy the contents of passed arc buf via
1416 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1419 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1422 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1425 DB_DNODE_ENTER(dbuf
);
1426 dn
= DB_DNODE(dbuf
);
1427 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1428 blkid
= dbuf_whichblock(dn
, offset
);
1429 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1430 rw_exit(&dn
->dn_struct_rwlock
);
1431 DB_DNODE_EXIT(dbuf
);
1433 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1434 dbuf_assign_arcbuf(db
, buf
, tx
);
1435 dbuf_rele(db
, FTAG
);
1440 DB_DNODE_ENTER(dbuf
);
1441 dn
= DB_DNODE(dbuf
);
1443 object
= dn
->dn_object
;
1444 DB_DNODE_EXIT(dbuf
);
1446 dbuf_rele(db
, FTAG
);
1447 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1448 dmu_return_arcbuf(buf
);
1449 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1454 dbuf_dirty_record_t
*dsa_dr
;
1455 dmu_sync_cb_t
*dsa_done
;
1462 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1464 dmu_sync_arg_t
*dsa
= varg
;
1465 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1466 blkptr_t
*bp
= zio
->io_bp
;
1468 if (zio
->io_error
== 0) {
1469 if (BP_IS_HOLE(bp
)) {
1471 * A block of zeros may compress to a hole, but the
1472 * block size still needs to be known for replay.
1474 BP_SET_LSIZE(bp
, db
->db_size
);
1476 ASSERT(BP_GET_LEVEL(bp
) == 0);
1483 dmu_sync_late_arrival_ready(zio_t
*zio
)
1485 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1490 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1492 dmu_sync_arg_t
*dsa
= varg
;
1493 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1494 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1496 mutex_enter(&db
->db_mtx
);
1497 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1498 if (zio
->io_error
== 0) {
1499 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1500 if (dr
->dt
.dl
.dr_nopwrite
) {
1501 ASSERTV(blkptr_t
*bp
= zio
->io_bp
);
1502 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1503 ASSERTV(uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
));
1505 ASSERT(BP_EQUAL(bp
, bp_orig
));
1506 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1507 ASSERT(zio_checksum_table
[chksum
].ci_dedup
);
1509 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1510 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1511 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1512 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
))
1513 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1515 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1517 cv_broadcast(&db
->db_changed
);
1518 mutex_exit(&db
->db_mtx
);
1520 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1522 kmem_free(dsa
, sizeof (*dsa
));
1526 dmu_sync_late_arrival_done(zio_t
*zio
)
1528 blkptr_t
*bp
= zio
->io_bp
;
1529 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1530 ASSERTV(blkptr_t
*bp_orig
= &zio
->io_bp_orig
);
1532 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1534 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1535 * then there is nothing to do here. Otherwise, free the
1536 * newly allocated block in this txg.
1538 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1539 ASSERT(BP_EQUAL(bp
, bp_orig
));
1541 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1542 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1543 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1544 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1548 dmu_tx_commit(dsa
->dsa_tx
);
1550 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1552 kmem_free(dsa
, sizeof (*dsa
));
1556 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1557 zio_prop_t
*zp
, zbookmark_t
*zb
)
1559 dmu_sync_arg_t
*dsa
;
1562 tx
= dmu_tx_create(os
);
1563 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1564 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1566 /* Make zl_get_data do txg_waited_synced() */
1567 return (SET_ERROR(EIO
));
1570 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1572 dsa
->dsa_done
= done
;
1576 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1577 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1578 dmu_sync_late_arrival_ready
, NULL
, dmu_sync_late_arrival_done
, dsa
,
1579 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
|ZIO_FLAG_FASTWRITE
, zb
));
1585 * Intent log support: sync the block associated with db to disk.
1586 * N.B. and XXX: the caller is responsible for making sure that the
1587 * data isn't changing while dmu_sync() is writing it.
1591 * EEXIST: this txg has already been synced, so there's nothing to do.
1592 * The caller should not log the write.
1594 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1595 * The caller should not log the write.
1597 * EALREADY: this block is already in the process of being synced.
1598 * The caller should track its progress (somehow).
1600 * EIO: could not do the I/O.
1601 * The caller should do a txg_wait_synced().
1603 * 0: the I/O has been initiated.
1604 * The caller should log this blkptr in the done callback.
1605 * It is possible that the I/O will fail, in which case
1606 * the error will be reported to the done callback and
1607 * propagated to pio from zio_done().
1610 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1612 blkptr_t
*bp
= zgd
->zgd_bp
;
1613 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1614 objset_t
*os
= db
->db_objset
;
1615 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1616 dbuf_dirty_record_t
*dr
;
1617 dmu_sync_arg_t
*dsa
;
1622 ASSERT(pio
!= NULL
);
1625 SET_BOOKMARK(&zb
, ds
->ds_object
,
1626 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1630 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1634 * If we're frozen (running ziltest), we always need to generate a bp.
1636 if (txg
> spa_freeze_txg(os
->os_spa
))
1637 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1640 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1641 * and us. If we determine that this txg is not yet syncing,
1642 * but it begins to sync a moment later, that's OK because the
1643 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1645 mutex_enter(&db
->db_mtx
);
1647 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1649 * This txg has already synced. There's nothing to do.
1651 mutex_exit(&db
->db_mtx
);
1652 return (SET_ERROR(EEXIST
));
1655 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1657 * This txg is currently syncing, so we can't mess with
1658 * the dirty record anymore; just write a new log block.
1660 mutex_exit(&db
->db_mtx
);
1661 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1664 dr
= db
->db_last_dirty
;
1665 while (dr
&& dr
->dr_txg
!= txg
)
1670 * There's no dr for this dbuf, so it must have been freed.
1671 * There's no need to log writes to freed blocks, so we're done.
1673 mutex_exit(&db
->db_mtx
);
1674 return (SET_ERROR(ENOENT
));
1677 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1680 * Assume the on-disk data is X, the current syncing data is Y,
1681 * and the current in-memory data is Z (currently in dmu_sync).
1682 * X and Z are identical but Y is has been modified. Normally,
1683 * when X and Z are the same we will perform a nopwrite but if Y
1684 * is different we must disable nopwrite since the resulting write
1685 * of Y to disk can free the block containing X. If we allowed a
1686 * nopwrite to occur the block pointing to Z would reference a freed
1687 * block. Since this is a rare case we simplify this by disabling
1688 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1689 * a previous transaction.
1692 zp
.zp_nopwrite
= B_FALSE
;
1694 ASSERT(dr
->dr_txg
== txg
);
1695 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1696 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1698 * We have already issued a sync write for this buffer,
1699 * or this buffer has already been synced. It could not
1700 * have been dirtied since, or we would have cleared the state.
1702 mutex_exit(&db
->db_mtx
);
1703 return (SET_ERROR(EALREADY
));
1706 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1707 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1708 mutex_exit(&db
->db_mtx
);
1710 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1712 dsa
->dsa_done
= done
;
1716 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1717 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1718 DBUF_IS_L2COMPRESSIBLE(db
), &zp
, dmu_sync_ready
,
1719 NULL
, dmu_sync_done
, dsa
, ZIO_PRIORITY_SYNC_WRITE
,
1720 ZIO_FLAG_CANFAIL
, &zb
));
1726 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1732 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1735 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1736 dnode_rele(dn
, FTAG
);
1741 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1746 /* XXX assumes dnode_hold will not get an i/o error */
1747 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1748 ASSERT(checksum
< ZIO_CHECKSUM_FUNCTIONS
);
1749 dn
->dn_checksum
= checksum
;
1750 dnode_setdirty(dn
, tx
);
1751 dnode_rele(dn
, FTAG
);
1755 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1760 /* XXX assumes dnode_hold will not get an i/o error */
1761 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1762 ASSERT(compress
< ZIO_COMPRESS_FUNCTIONS
);
1763 dn
->dn_compress
= compress
;
1764 dnode_setdirty(dn
, tx
);
1765 dnode_rele(dn
, FTAG
);
1768 int zfs_mdcomp_disable
= 0;
1771 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1773 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1774 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1776 enum zio_checksum checksum
= os
->os_checksum
;
1777 enum zio_compress compress
= os
->os_compress
;
1778 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1779 boolean_t dedup
= B_FALSE
;
1780 boolean_t nopwrite
= B_FALSE
;
1781 boolean_t dedup_verify
= os
->os_dedup_verify
;
1782 int copies
= os
->os_copies
;
1785 * We maintain different write policies for each of the following
1788 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1789 * 3. all other level 0 blocks
1793 * XXX -- we should design a compression algorithm
1794 * that specializes in arrays of bps.
1796 compress
= zfs_mdcomp_disable
? ZIO_COMPRESS_EMPTY
:
1800 * Metadata always gets checksummed. If the data
1801 * checksum is multi-bit correctable, and it's not a
1802 * ZBT-style checksum, then it's suitable for metadata
1803 * as well. Otherwise, the metadata checksum defaults
1806 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1807 zio_checksum_table
[checksum
].ci_eck
)
1808 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1809 } else if (wp
& WP_NOFILL
) {
1813 * If we're writing preallocated blocks, we aren't actually
1814 * writing them so don't set any policy properties. These
1815 * blocks are currently only used by an external subsystem
1816 * outside of zfs (i.e. dump) and not written by the zio
1819 compress
= ZIO_COMPRESS_OFF
;
1820 checksum
= ZIO_CHECKSUM_OFF
;
1822 compress
= zio_compress_select(dn
->dn_compress
, compress
);
1824 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1825 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1829 * Determine dedup setting. If we are in dmu_sync(),
1830 * we won't actually dedup now because that's all
1831 * done in syncing context; but we do want to use the
1832 * dedup checkum. If the checksum is not strong
1833 * enough to ensure unique signatures, force
1836 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1837 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1838 if (!zio_checksum_table
[checksum
].ci_dedup
)
1839 dedup_verify
= B_TRUE
;
1843 * Enable nopwrite if we have a cryptographically secure
1844 * checksum that has no known collisions (i.e. SHA-256)
1845 * and compression is enabled. We don't enable nopwrite if
1846 * dedup is enabled as the two features are mutually exclusive.
1848 nopwrite
= (!dedup
&& zio_checksum_table
[checksum
].ci_dedup
&&
1849 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1852 zp
->zp_checksum
= checksum
;
1853 zp
->zp_compress
= compress
;
1854 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1855 zp
->zp_level
= level
;
1856 zp
->zp_copies
= MIN(copies
+ ismd
, spa_max_replication(os
->os_spa
));
1857 zp
->zp_dedup
= dedup
;
1858 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1859 zp
->zp_nopwrite
= nopwrite
;
1863 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1868 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1872 * Sync any current changes before
1873 * we go trundling through the block pointers.
1875 for (i
= 0; i
< TXG_SIZE
; i
++) {
1876 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1879 if (i
!= TXG_SIZE
) {
1880 dnode_rele(dn
, FTAG
);
1881 txg_wait_synced(dmu_objset_pool(os
), 0);
1882 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1887 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1888 dnode_rele(dn
, FTAG
);
1894 __dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1896 dnode_phys_t
*dnp
= dn
->dn_phys
;
1899 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1900 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1901 1ULL << dn
->dn_indblkshift
: 0;
1902 doi
->doi_type
= dn
->dn_type
;
1903 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1904 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1905 doi
->doi_indirection
= dn
->dn_nlevels
;
1906 doi
->doi_checksum
= dn
->dn_checksum
;
1907 doi
->doi_compress
= dn
->dn_compress
;
1908 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1909 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1910 doi
->doi_fill_count
= 0;
1911 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1912 doi
->doi_fill_count
+= dnp
->dn_blkptr
[i
].blk_fill
;
1916 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1918 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1919 mutex_enter(&dn
->dn_mtx
);
1921 __dmu_object_info_from_dnode(dn
, doi
);
1923 mutex_exit(&dn
->dn_mtx
);
1924 rw_exit(&dn
->dn_struct_rwlock
);
1928 * Get information on a DMU object.
1929 * If doi is NULL, just indicates whether the object exists.
1932 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1935 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1941 dmu_object_info_from_dnode(dn
, doi
);
1943 dnode_rele(dn
, FTAG
);
1948 * As above, but faster; can be used when you have a held dbuf in hand.
1951 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1953 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1956 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1961 * Faster still when you only care about the size.
1962 * This is specifically optimized for zfs_getattr().
1965 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1966 u_longlong_t
*nblk512
)
1968 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1974 *blksize
= dn
->dn_datablksz
;
1975 /* add 1 for dnode space */
1976 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1977 SPA_MINBLOCKSHIFT
) + 1;
1982 byteswap_uint64_array(void *vbuf
, size_t size
)
1984 uint64_t *buf
= vbuf
;
1985 size_t count
= size
>> 3;
1988 ASSERT((size
& 7) == 0);
1990 for (i
= 0; i
< count
; i
++)
1991 buf
[i
] = BSWAP_64(buf
[i
]);
1995 byteswap_uint32_array(void *vbuf
, size_t size
)
1997 uint32_t *buf
= vbuf
;
1998 size_t count
= size
>> 2;
2001 ASSERT((size
& 3) == 0);
2003 for (i
= 0; i
< count
; i
++)
2004 buf
[i
] = BSWAP_32(buf
[i
]);
2008 byteswap_uint16_array(void *vbuf
, size_t size
)
2010 uint16_t *buf
= vbuf
;
2011 size_t count
= size
>> 1;
2014 ASSERT((size
& 1) == 0);
2016 for (i
= 0; i
< count
; i
++)
2017 buf
[i
] = BSWAP_16(buf
[i
]);
2022 byteswap_uint8_array(void *vbuf
, size_t size
)
2044 arc_fini(); /* arc depends on l2arc, so arc must go first */
2056 #if defined(_KERNEL) && defined(HAVE_SPL)
2057 EXPORT_SYMBOL(dmu_bonus_hold
);
2058 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
2059 EXPORT_SYMBOL(dmu_buf_rele_array
);
2060 EXPORT_SYMBOL(dmu_prefetch
);
2061 EXPORT_SYMBOL(dmu_free_range
);
2062 EXPORT_SYMBOL(dmu_free_long_range
);
2063 EXPORT_SYMBOL(dmu_free_long_object
);
2064 EXPORT_SYMBOL(dmu_read
);
2065 EXPORT_SYMBOL(dmu_write
);
2066 EXPORT_SYMBOL(dmu_prealloc
);
2067 EXPORT_SYMBOL(dmu_object_info
);
2068 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
2069 EXPORT_SYMBOL(dmu_object_info_from_db
);
2070 EXPORT_SYMBOL(dmu_object_size_from_db
);
2071 EXPORT_SYMBOL(dmu_object_set_blocksize
);
2072 EXPORT_SYMBOL(dmu_object_set_checksum
);
2073 EXPORT_SYMBOL(dmu_object_set_compress
);
2074 EXPORT_SYMBOL(dmu_write_policy
);
2075 EXPORT_SYMBOL(dmu_sync
);
2076 EXPORT_SYMBOL(dmu_request_arcbuf
);
2077 EXPORT_SYMBOL(dmu_return_arcbuf
);
2078 EXPORT_SYMBOL(dmu_assign_arcbuf
);
2079 EXPORT_SYMBOL(dmu_buf_hold
);
2080 EXPORT_SYMBOL(dmu_ot
);
2082 module_param(zfs_mdcomp_disable
, int, 0644);
2083 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");
2085 module_param(zfs_nopwrite_enabled
, int, 0644);
2086 MODULE_PARM_DESC(zfs_nopwrite_enabled
, "Enable NOP writes");