4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
27 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
30 #include <sys/dnode.h>
31 #include <sys/zfs_context.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dmu_traverse.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/dsl_synctask.h>
38 #include <sys/dsl_prop.h>
39 #include <sys/dmu_zfetch.h>
40 #include <sys/zfs_ioctl.h>
42 #include <sys/zio_checksum.h>
45 #include <sys/vmsystm.h>
46 #include <sys/zfs_znode.h>
49 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
50 { DMU_BSWAP_UINT8
, TRUE
, "unallocated" },
51 { DMU_BSWAP_ZAP
, TRUE
, "object directory" },
52 { DMU_BSWAP_UINT64
, TRUE
, "object array" },
53 { DMU_BSWAP_UINT8
, TRUE
, "packed nvlist" },
54 { DMU_BSWAP_UINT64
, TRUE
, "packed nvlist size" },
55 { DMU_BSWAP_UINT64
, TRUE
, "bpobj" },
56 { DMU_BSWAP_UINT64
, TRUE
, "bpobj header" },
57 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map header" },
58 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map" },
59 { DMU_BSWAP_UINT64
, TRUE
, "ZIL intent log" },
60 { DMU_BSWAP_DNODE
, TRUE
, "DMU dnode" },
61 { DMU_BSWAP_OBJSET
, TRUE
, "DMU objset" },
62 { DMU_BSWAP_UINT64
, TRUE
, "DSL directory" },
63 { DMU_BSWAP_ZAP
, TRUE
, "DSL directory child map"},
64 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset snap map" },
65 { DMU_BSWAP_ZAP
, TRUE
, "DSL props" },
66 { DMU_BSWAP_UINT64
, TRUE
, "DSL dataset" },
67 { DMU_BSWAP_ZNODE
, TRUE
, "ZFS znode" },
68 { DMU_BSWAP_OLDACL
, TRUE
, "ZFS V0 ACL" },
69 { DMU_BSWAP_UINT8
, FALSE
, "ZFS plain file" },
70 { DMU_BSWAP_ZAP
, TRUE
, "ZFS directory" },
71 { DMU_BSWAP_ZAP
, TRUE
, "ZFS master node" },
72 { DMU_BSWAP_ZAP
, TRUE
, "ZFS delete queue" },
73 { DMU_BSWAP_UINT8
, FALSE
, "zvol object" },
74 { DMU_BSWAP_ZAP
, TRUE
, "zvol prop" },
75 { DMU_BSWAP_UINT8
, FALSE
, "other uint8[]" },
76 { DMU_BSWAP_UINT64
, FALSE
, "other uint64[]" },
77 { DMU_BSWAP_ZAP
, TRUE
, "other ZAP" },
78 { DMU_BSWAP_ZAP
, TRUE
, "persistent error log" },
79 { DMU_BSWAP_UINT8
, TRUE
, "SPA history" },
80 { DMU_BSWAP_UINT64
, TRUE
, "SPA history offsets" },
81 { DMU_BSWAP_ZAP
, TRUE
, "Pool properties" },
82 { DMU_BSWAP_ZAP
, TRUE
, "DSL permissions" },
83 { DMU_BSWAP_ACL
, TRUE
, "ZFS ACL" },
84 { DMU_BSWAP_UINT8
, TRUE
, "ZFS SYSACL" },
85 { DMU_BSWAP_UINT8
, TRUE
, "FUID table" },
86 { DMU_BSWAP_UINT64
, TRUE
, "FUID table size" },
87 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset next clones"},
88 { DMU_BSWAP_ZAP
, TRUE
, "scan work queue" },
89 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group used" },
90 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group quota" },
91 { DMU_BSWAP_ZAP
, TRUE
, "snapshot refcount tags"},
92 { DMU_BSWAP_ZAP
, TRUE
, "DDT ZAP algorithm" },
93 { DMU_BSWAP_ZAP
, TRUE
, "DDT statistics" },
94 { DMU_BSWAP_UINT8
, TRUE
, "System attributes" },
95 { DMU_BSWAP_ZAP
, TRUE
, "SA master node" },
96 { DMU_BSWAP_ZAP
, TRUE
, "SA attr registration" },
97 { DMU_BSWAP_ZAP
, TRUE
, "SA attr layouts" },
98 { DMU_BSWAP_ZAP
, TRUE
, "scan translations" },
99 { DMU_BSWAP_UINT8
, FALSE
, "deduplicated block" },
100 { DMU_BSWAP_ZAP
, TRUE
, "DSL deadlist map" },
101 { DMU_BSWAP_UINT64
, TRUE
, "DSL deadlist map hdr" },
102 { DMU_BSWAP_ZAP
, TRUE
, "DSL dir clones" },
103 { DMU_BSWAP_UINT64
, TRUE
, "bpobj subobj" }
106 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
107 { byteswap_uint8_array
, "uint8" },
108 { byteswap_uint16_array
, "uint16" },
109 { byteswap_uint32_array
, "uint32" },
110 { byteswap_uint64_array
, "uint64" },
111 { zap_byteswap
, "zap" },
112 { dnode_buf_byteswap
, "dnode" },
113 { dmu_objset_byteswap
, "objset" },
114 { zfs_znode_byteswap
, "znode" },
115 { zfs_oldacl_byteswap
, "oldacl" },
116 { zfs_acl_byteswap
, "acl" }
120 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
121 void *tag
, dmu_buf_t
**dbp
, int flags
)
127 int db_flags
= DB_RF_CANFAIL
;
129 if (flags
& DMU_READ_NO_PREFETCH
)
130 db_flags
|= DB_RF_NOPREFETCH
;
132 err
= dnode_hold(os
, object
, FTAG
, &dn
);
135 blkid
= dbuf_whichblock(dn
, offset
);
136 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
137 db
= dbuf_hold(dn
, blkid
, tag
);
138 rw_exit(&dn
->dn_struct_rwlock
);
142 err
= dbuf_read(db
, NULL
, db_flags
);
149 dnode_rele(dn
, FTAG
);
150 *dbp
= &db
->db
; /* NULL db plus first field offset is NULL */
157 return (DN_MAX_BONUSLEN
);
161 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
163 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
170 if (dn
->dn_bonus
!= db
) {
172 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
175 dnode_setbonuslen(dn
, newsize
, tx
);
184 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
186 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
193 if (!DMU_OT_IS_VALID(type
)) {
195 } else if (dn
->dn_bonus
!= db
) {
198 dnode_setbonus_type(dn
, type
, tx
);
207 dmu_get_bonustype(dmu_buf_t
*db_fake
)
209 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
211 dmu_object_type_t type
;
215 type
= dn
->dn_bonustype
;
222 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
227 error
= dnode_hold(os
, object
, FTAG
, &dn
);
228 dbuf_rm_spill(dn
, tx
);
229 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
230 dnode_rm_spill(dn
, tx
);
231 rw_exit(&dn
->dn_struct_rwlock
);
232 dnode_rele(dn
, FTAG
);
237 * returns ENOENT, EIO, or 0.
240 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
246 error
= dnode_hold(os
, object
, FTAG
, &dn
);
250 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
251 if (dn
->dn_bonus
== NULL
) {
252 rw_exit(&dn
->dn_struct_rwlock
);
253 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
254 if (dn
->dn_bonus
== NULL
)
255 dbuf_create_bonus(dn
);
259 /* as long as the bonus buf is held, the dnode will be held */
260 if (refcount_add(&db
->db_holds
, tag
) == 1) {
261 VERIFY(dnode_add_ref(dn
, db
));
262 (void) atomic_inc_32_nv(&dn
->dn_dbufs_count
);
266 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
267 * hold and incrementing the dbuf count to ensure that dnode_move() sees
268 * a dnode hold for every dbuf.
270 rw_exit(&dn
->dn_struct_rwlock
);
272 dnode_rele(dn
, FTAG
);
274 VERIFY(0 == dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
));
281 * returns ENOENT, EIO, or 0.
283 * This interface will allocate a blank spill dbuf when a spill blk
284 * doesn't already exist on the dnode.
286 * if you only want to find an already existing spill db, then
287 * dmu_spill_hold_existing() should be used.
290 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
292 dmu_buf_impl_t
*db
= NULL
;
295 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
296 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
298 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
300 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
301 rw_exit(&dn
->dn_struct_rwlock
);
304 err
= dbuf_read(db
, NULL
, flags
);
313 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
315 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
322 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
325 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
327 if (!dn
->dn_have_spill
) {
330 err
= dmu_spill_hold_by_dnode(dn
,
331 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
334 rw_exit(&dn
->dn_struct_rwlock
);
342 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
344 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
350 err
= dmu_spill_hold_by_dnode(dn
, DB_RF_CANFAIL
, tag
, dbp
);
357 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
358 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
359 * and can induce severe lock contention when writing to several files
360 * whose dnodes are in the same block.
363 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
364 int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
366 dsl_pool_t
*dp
= NULL
;
368 uint64_t blkid
, nblks
, i
;
374 ASSERT(length
<= DMU_MAX_ACCESS
);
376 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
;
377 if (flags
& DMU_READ_NO_PREFETCH
|| length
> zfetch_array_rd_sz
)
378 dbuf_flags
|= DB_RF_NOPREFETCH
;
380 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
381 if (dn
->dn_datablkshift
) {
382 int blkshift
= dn
->dn_datablkshift
;
383 nblks
= (P2ROUNDUP(offset
+length
, 1ULL<<blkshift
) -
384 P2ALIGN(offset
, 1ULL<<blkshift
)) >> blkshift
;
386 if (offset
+ length
> dn
->dn_datablksz
) {
387 zfs_panic_recover("zfs: accessing past end of object "
388 "%llx/%llx (size=%u access=%llu+%llu)",
389 (longlong_t
)dn
->dn_objset
->
390 os_dsl_dataset
->ds_object
,
391 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
392 (longlong_t
)offset
, (longlong_t
)length
);
393 rw_exit(&dn
->dn_struct_rwlock
);
398 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_PUSHPAGE
| KM_NODEBUG
);
400 if (dn
->dn_objset
->os_dsl_dataset
)
401 dp
= dn
->dn_objset
->os_dsl_dataset
->ds_dir
->dd_pool
;
402 if (dp
&& dsl_pool_sync_context(dp
))
404 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
405 blkid
= dbuf_whichblock(dn
, offset
);
406 for (i
= 0; i
< nblks
; i
++) {
407 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+i
, tag
);
409 rw_exit(&dn
->dn_struct_rwlock
);
410 dmu_buf_rele_array(dbp
, nblks
, tag
);
414 /* initiate async i/o */
416 (void) dbuf_read(db
, zio
, dbuf_flags
);
420 rw_exit(&dn
->dn_struct_rwlock
);
422 /* wait for async i/o */
424 /* track read overhead when we are in sync context */
425 if (dp
&& dsl_pool_sync_context(dp
))
426 dp
->dp_read_overhead
+= gethrtime() - start
;
428 dmu_buf_rele_array(dbp
, nblks
, tag
);
432 /* wait for other io to complete */
434 for (i
= 0; i
< nblks
; i
++) {
435 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
436 mutex_enter(&db
->db_mtx
);
437 while (db
->db_state
== DB_READ
||
438 db
->db_state
== DB_FILL
)
439 cv_wait(&db
->db_changed
, &db
->db_mtx
);
440 if (db
->db_state
== DB_UNCACHED
)
442 mutex_exit(&db
->db_mtx
);
444 dmu_buf_rele_array(dbp
, nblks
, tag
);
456 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
457 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
462 err
= dnode_hold(os
, object
, FTAG
, &dn
);
466 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
467 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
469 dnode_rele(dn
, FTAG
);
475 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
476 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
478 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
484 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
485 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
492 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
495 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
500 for (i
= 0; i
< numbufs
; i
++) {
502 dbuf_rele(dbp
[i
], tag
);
505 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
509 dmu_prefetch(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t len
)
515 if (zfs_prefetch_disable
)
518 if (len
== 0) { /* they're interested in the bonus buffer */
519 dn
= DMU_META_DNODE(os
);
521 if (object
== 0 || object
>= DN_MAX_OBJECT
)
524 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
525 blkid
= dbuf_whichblock(dn
, object
* sizeof (dnode_phys_t
));
526 dbuf_prefetch(dn
, blkid
);
527 rw_exit(&dn
->dn_struct_rwlock
);
532 * XXX - Note, if the dnode for the requested object is not
533 * already cached, we will do a *synchronous* read in the
534 * dnode_hold() call. The same is true for any indirects.
536 err
= dnode_hold(os
, object
, FTAG
, &dn
);
540 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
541 if (dn
->dn_datablkshift
) {
542 int blkshift
= dn
->dn_datablkshift
;
543 nblks
= (P2ROUNDUP(offset
+len
, 1<<blkshift
) -
544 P2ALIGN(offset
, 1<<blkshift
)) >> blkshift
;
546 nblks
= (offset
< dn
->dn_datablksz
);
550 blkid
= dbuf_whichblock(dn
, offset
);
551 for (i
= 0; i
< nblks
; i
++)
552 dbuf_prefetch(dn
, blkid
+i
);
555 rw_exit(&dn
->dn_struct_rwlock
);
557 dnode_rele(dn
, FTAG
);
561 * Get the next "chunk" of file data to free. We traverse the file from
562 * the end so that the file gets shorter over time (if we crashes in the
563 * middle, this will leave us in a better state). We find allocated file
564 * data by simply searching the allocated level 1 indirects.
567 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t limit
)
569 uint64_t len
= *start
- limit
;
571 uint64_t maxblks
= DMU_MAX_ACCESS
/ (1ULL << (dn
->dn_indblkshift
+ 1));
573 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
575 ASSERT(limit
<= *start
);
577 if (len
<= iblkrange
* maxblks
) {
581 ASSERT(ISP2(iblkrange
));
583 while (*start
> limit
&& blkcnt
< maxblks
) {
586 /* find next allocated L1 indirect */
587 err
= dnode_next_offset(dn
,
588 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
590 /* if there are no more, then we are done */
599 /* reset offset to end of "next" block back */
600 *start
= P2ALIGN(*start
, iblkrange
);
610 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
611 uint64_t length
, boolean_t free_dnode
)
614 uint64_t object_size
, start
, end
, len
;
615 boolean_t trunc
= (length
== DMU_OBJECT_END
);
618 align
= 1 << dn
->dn_datablkshift
;
620 object_size
= align
== 1 ? dn
->dn_datablksz
:
621 (dn
->dn_maxblkid
+ 1) << dn
->dn_datablkshift
;
623 end
= offset
+ length
;
624 if (trunc
|| end
> object_size
)
628 length
= end
- offset
;
632 /* assert(offset <= start) */
633 err
= get_next_chunk(dn
, &start
, offset
);
636 len
= trunc
? DMU_OBJECT_END
: end
- start
;
638 tx
= dmu_tx_create(os
);
639 dmu_tx_hold_free(tx
, dn
->dn_object
, start
, len
);
640 err
= dmu_tx_assign(tx
, TXG_WAIT
);
646 dnode_free_range(dn
, start
, trunc
? -1 : len
, tx
);
648 if (start
== 0 && free_dnode
) {
653 length
-= end
- start
;
662 dmu_free_long_range(objset_t
*os
, uint64_t object
,
663 uint64_t offset
, uint64_t length
)
668 err
= dnode_hold(os
, object
, FTAG
, &dn
);
671 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
, FALSE
);
672 dnode_rele(dn
, FTAG
);
677 dmu_free_object(objset_t
*os
, uint64_t object
)
683 err
= dnode_hold_impl(os
, object
, DNODE_MUST_BE_ALLOCATED
,
687 if (dn
->dn_nlevels
== 1) {
688 tx
= dmu_tx_create(os
);
689 dmu_tx_hold_bonus(tx
, object
);
690 dmu_tx_hold_free(tx
, dn
->dn_object
, 0, DMU_OBJECT_END
);
691 err
= dmu_tx_assign(tx
, TXG_WAIT
);
693 dnode_free_range(dn
, 0, DMU_OBJECT_END
, tx
);
700 err
= dmu_free_long_range_impl(os
, dn
, 0, DMU_OBJECT_END
, TRUE
);
702 dnode_rele(dn
, FTAG
);
707 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
708 uint64_t size
, dmu_tx_t
*tx
)
711 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
714 ASSERT(offset
< UINT64_MAX
);
715 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
716 dnode_free_range(dn
, offset
, size
, tx
);
717 dnode_rele(dn
, FTAG
);
722 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
723 void *buf
, uint32_t flags
)
729 err
= dnode_hold(os
, object
, FTAG
, &dn
);
734 * Deal with odd block sizes, where there can't be data past the first
735 * block. If we ever do the tail block optimization, we will need to
736 * handle that here as well.
738 if (dn
->dn_maxblkid
== 0) {
739 int newsz
= offset
> dn
->dn_datablksz
? 0 :
740 MIN(size
, dn
->dn_datablksz
- offset
);
741 bzero((char *)buf
+ newsz
, size
- newsz
);
746 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
750 * NB: we could do this block-at-a-time, but it's nice
751 * to be reading in parallel.
753 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
754 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
758 for (i
= 0; i
< numbufs
; i
++) {
761 dmu_buf_t
*db
= dbp
[i
];
765 bufoff
= offset
- db
->db_offset
;
766 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
768 bcopy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
772 buf
= (char *)buf
+ tocpy
;
774 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
776 dnode_rele(dn
, FTAG
);
781 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
782 const void *buf
, dmu_tx_t
*tx
)
790 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
791 FALSE
, FTAG
, &numbufs
, &dbp
));
793 for (i
= 0; i
< numbufs
; i
++) {
796 dmu_buf_t
*db
= dbp
[i
];
800 bufoff
= offset
- db
->db_offset
;
801 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
803 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
805 if (tocpy
== db
->db_size
)
806 dmu_buf_will_fill(db
, tx
);
808 dmu_buf_will_dirty(db
, tx
);
810 (void) memcpy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
812 if (tocpy
== db
->db_size
)
813 dmu_buf_fill_done(db
, tx
);
817 buf
= (char *)buf
+ tocpy
;
819 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
823 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
832 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
833 FALSE
, FTAG
, &numbufs
, &dbp
));
835 for (i
= 0; i
< numbufs
; i
++) {
836 dmu_buf_t
*db
= dbp
[i
];
838 dmu_buf_will_not_fill(db
, tx
);
840 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
844 * DMU support for xuio
846 kstat_t
*xuio_ksp
= NULL
;
848 typedef struct xuio_stats
{
849 /* loaned yet not returned arc_buf */
850 kstat_named_t xuiostat_onloan_rbuf
;
851 kstat_named_t xuiostat_onloan_wbuf
;
852 /* whether a copy is made when loaning out a read buffer */
853 kstat_named_t xuiostat_rbuf_copied
;
854 kstat_named_t xuiostat_rbuf_nocopy
;
855 /* whether a copy is made when assigning a write buffer */
856 kstat_named_t xuiostat_wbuf_copied
;
857 kstat_named_t xuiostat_wbuf_nocopy
;
860 static xuio_stats_t xuio_stats
= {
861 { "onloan_read_buf", KSTAT_DATA_UINT64
},
862 { "onloan_write_buf", KSTAT_DATA_UINT64
},
863 { "read_buf_copied", KSTAT_DATA_UINT64
},
864 { "read_buf_nocopy", KSTAT_DATA_UINT64
},
865 { "write_buf_copied", KSTAT_DATA_UINT64
},
866 { "write_buf_nocopy", KSTAT_DATA_UINT64
}
869 #define XUIOSTAT_INCR(stat, val) \
870 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
871 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
874 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
877 uio_t
*uio
= &xuio
->xu_uio
;
879 uio
->uio_iovcnt
= nblk
;
880 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_PUSHPAGE
);
882 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_PUSHPAGE
);
884 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_PUSHPAGE
);
885 priv
->iovp
= uio
->uio_iov
;
886 XUIO_XUZC_PRIV(xuio
) = priv
;
888 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
889 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
891 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
897 dmu_xuio_fini(xuio_t
*xuio
)
899 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
900 int nblk
= priv
->cnt
;
902 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
903 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
904 kmem_free(priv
, sizeof (dmu_xuio_t
));
906 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
907 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
909 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
913 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
914 * and increase priv->next by 1.
917 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
920 uio_t
*uio
= &xuio
->xu_uio
;
921 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
922 int i
= priv
->next
++;
924 ASSERT(i
< priv
->cnt
);
925 ASSERT(off
+ n
<= arc_buf_size(abuf
));
926 iov
= uio
->uio_iov
+ i
;
927 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
929 priv
->bufs
[i
] = abuf
;
934 dmu_xuio_cnt(xuio_t
*xuio
)
936 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
941 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
943 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
945 ASSERT(i
< priv
->cnt
);
946 return (priv
->bufs
[i
]);
950 dmu_xuio_clear(xuio_t
*xuio
, int i
)
952 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
954 ASSERT(i
< priv
->cnt
);
955 priv
->bufs
[i
] = NULL
;
961 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
962 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
964 if (xuio_ksp
!= NULL
) {
965 xuio_ksp
->ks_data
= &xuio_stats
;
966 kstat_install(xuio_ksp
);
973 if (xuio_ksp
!= NULL
) {
974 kstat_delete(xuio_ksp
);
980 xuio_stat_wbuf_copied()
982 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
986 xuio_stat_wbuf_nocopy()
988 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
994 * Copy up to size bytes between arg_buf and req based on the data direction
995 * described by the req. If an entire req's data cannot be transfered the
996 * req's is updated such that it's current index and bv offsets correctly
997 * reference any residual data which could not be copied. The return value
998 * is the number of bytes successfully copied to arg_buf.
1001 dmu_req_copy(void *arg_buf
, int size
, int *offset
, struct request
*req
)
1004 struct req_iterator iter
;
1009 rq_for_each_segment(bv
, req
, iter
) {
1011 /* Fully consumed the passed arg_buf */
1012 ASSERT3S(*offset
, <=, size
);
1013 if (size
== *offset
)
1016 /* Skip fully consumed bv's */
1017 if (bv
->bv_len
== 0)
1020 tocpy
= MIN(bv
->bv_len
, size
- *offset
);
1021 ASSERT3S(tocpy
, >=, 0);
1023 bv_buf
= page_address(bv
->bv_page
) + bv
->bv_offset
;
1024 ASSERT3P(bv_buf
, !=, NULL
);
1026 if (rq_data_dir(req
) == WRITE
)
1027 memcpy(arg_buf
+ *offset
, bv_buf
, tocpy
);
1029 memcpy(bv_buf
, arg_buf
+ *offset
, tocpy
);
1032 bv
->bv_offset
+= tocpy
;
1033 bv
->bv_len
-= tocpy
;
1040 dmu_bio_put(struct bio
*bio
)
1042 struct bio
*bio_next
;
1045 bio_next
= bio
->bi_next
;
1052 dmu_bio_clone(struct bio
*bio
, struct bio
**bio_copy
)
1054 struct bio
*bio_root
= NULL
;
1055 struct bio
*bio_last
= NULL
;
1056 struct bio
*bio_new
;
1062 bio_new
= bio_clone(bio
, GFP_NOIO
);
1063 if (bio_new
== NULL
) {
1064 dmu_bio_put(bio_root
);
1069 bio_last
->bi_next
= bio_new
;
1079 *bio_copy
= bio_root
;
1085 dmu_read_req(objset_t
*os
, uint64_t object
, struct request
*req
)
1087 uint64_t size
= blk_rq_bytes(req
);
1088 uint64_t offset
= blk_rq_pos(req
) << 9;
1089 struct bio
*bio_saved
= req
->bio
;
1091 int numbufs
, i
, err
;
1094 * NB: we could do this block-at-a-time, but it's nice
1095 * to be reading in parallel.
1097 err
= dmu_buf_hold_array(os
, object
, offset
, size
, TRUE
, FTAG
,
1103 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1104 * can be safely modified. The original bio list is relinked in to
1105 * the request when the function exits. This is required because
1106 * some file systems blindly assume that these values will remain
1107 * constant between bio_submit() and the IO completion callback.
1109 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1113 for (i
= 0; i
< numbufs
; i
++) {
1114 int tocpy
, didcpy
, bufoff
;
1115 dmu_buf_t
*db
= dbp
[i
];
1117 bufoff
= offset
- db
->db_offset
;
1118 ASSERT3S(bufoff
, >=, 0);
1120 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1124 err
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, &didcpy
, req
);
1137 dmu_bio_put(req
->bio
);
1138 req
->bio
= bio_saved
;
1140 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1146 dmu_write_req(objset_t
*os
, uint64_t object
, struct request
*req
, dmu_tx_t
*tx
)
1148 uint64_t size
= blk_rq_bytes(req
);
1149 uint64_t offset
= blk_rq_pos(req
) << 9;
1150 struct bio
*bio_saved
= req
->bio
;
1159 err
= dmu_buf_hold_array(os
, object
, offset
, size
, FALSE
, FTAG
,
1165 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1166 * can be safely modified. The original bio list is relinked in to
1167 * the request when the function exits. This is required because
1168 * some file systems blindly assume that these values will remain
1169 * constant between bio_submit() and the IO completion callback.
1171 err
= dmu_bio_clone(bio_saved
, &req
->bio
);
1175 for (i
= 0; i
< numbufs
; i
++) {
1176 int tocpy
, didcpy
, bufoff
;
1177 dmu_buf_t
*db
= dbp
[i
];
1179 bufoff
= offset
- db
->db_offset
;
1180 ASSERT3S(bufoff
, >=, 0);
1182 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1186 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1188 if (tocpy
== db
->db_size
)
1189 dmu_buf_will_fill(db
, tx
);
1191 dmu_buf_will_dirty(db
, tx
);
1193 err
= dmu_req_copy(db
->db_data
+ bufoff
, tocpy
, &didcpy
, req
);
1195 if (tocpy
== db
->db_size
)
1196 dmu_buf_fill_done(db
, tx
);
1209 dmu_bio_put(req
->bio
);
1210 req
->bio
= bio_saved
;
1212 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1218 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1221 int numbufs
, i
, err
;
1222 xuio_t
*xuio
= NULL
;
1225 * NB: we could do this block-at-a-time, but it's nice
1226 * to be reading in parallel.
1228 err
= dmu_buf_hold_array(os
, object
, uio
->uio_loffset
, size
, TRUE
, FTAG
,
1233 for (i
= 0; i
< numbufs
; i
++) {
1236 dmu_buf_t
*db
= dbp
[i
];
1240 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1241 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1244 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1245 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1246 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1247 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1249 uio
->uio_resid
-= tocpy
;
1250 uio
->uio_loffset
+= tocpy
;
1253 if (abuf
== dbuf_abuf
)
1254 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1256 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1258 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1266 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1272 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1279 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1280 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1284 for (i
= 0; i
< numbufs
; i
++) {
1287 dmu_buf_t
*db
= dbp
[i
];
1291 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1292 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1294 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1296 if (tocpy
== db
->db_size
)
1297 dmu_buf_will_fill(db
, tx
);
1299 dmu_buf_will_dirty(db
, tx
);
1302 * XXX uiomove could block forever (eg.nfs-backed
1303 * pages). There needs to be a uiolockdown() function
1304 * to lock the pages in memory, so that uiomove won't
1307 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1310 if (tocpy
== db
->db_size
)
1311 dmu_buf_fill_done(db
, tx
);
1319 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1324 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1327 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1336 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1343 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1352 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1356 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1358 dnode_rele(dn
, FTAG
);
1362 #endif /* _KERNEL */
1365 * Allocate a loaned anonymous arc buffer.
1368 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1370 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1373 DB_GET_SPA(&spa
, db
);
1374 return (arc_loan_buf(spa
, size
));
1378 * Free a loaned arc buffer.
1381 dmu_return_arcbuf(arc_buf_t
*buf
)
1383 arc_return_buf(buf
, FTAG
);
1384 VERIFY(arc_buf_remove_ref(buf
, FTAG
) == 1);
1388 * When possible directly assign passed loaned arc buffer to a dbuf.
1389 * If this is not possible copy the contents of passed arc buf via
1393 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1396 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1399 uint32_t blksz
= (uint32_t)arc_buf_size(buf
);
1402 DB_DNODE_ENTER(dbuf
);
1403 dn
= DB_DNODE(dbuf
);
1404 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1405 blkid
= dbuf_whichblock(dn
, offset
);
1406 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1407 rw_exit(&dn
->dn_struct_rwlock
);
1408 DB_DNODE_EXIT(dbuf
);
1410 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1411 dbuf_assign_arcbuf(db
, buf
, tx
);
1412 dbuf_rele(db
, FTAG
);
1417 DB_DNODE_ENTER(dbuf
);
1418 dn
= DB_DNODE(dbuf
);
1420 object
= dn
->dn_object
;
1421 DB_DNODE_EXIT(dbuf
);
1423 dbuf_rele(db
, FTAG
);
1424 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1425 dmu_return_arcbuf(buf
);
1426 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1431 dbuf_dirty_record_t
*dsa_dr
;
1432 dmu_sync_cb_t
*dsa_done
;
1439 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1441 dmu_sync_arg_t
*dsa
= varg
;
1442 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1443 blkptr_t
*bp
= zio
->io_bp
;
1445 if (zio
->io_error
== 0) {
1446 if (BP_IS_HOLE(bp
)) {
1448 * A block of zeros may compress to a hole, but the
1449 * block size still needs to be known for replay.
1451 BP_SET_LSIZE(bp
, db
->db_size
);
1453 ASSERT(BP_GET_LEVEL(bp
) == 0);
1460 dmu_sync_late_arrival_ready(zio_t
*zio
)
1462 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1467 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1469 dmu_sync_arg_t
*dsa
= varg
;
1470 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1471 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1473 mutex_enter(&db
->db_mtx
);
1474 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1475 if (zio
->io_error
== 0) {
1476 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1477 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1478 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1479 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
))
1480 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1482 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1484 cv_broadcast(&db
->db_changed
);
1485 mutex_exit(&db
->db_mtx
);
1487 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1489 kmem_free(dsa
, sizeof (*dsa
));
1493 dmu_sync_late_arrival_done(zio_t
*zio
)
1495 blkptr_t
*bp
= zio
->io_bp
;
1496 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1498 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1499 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1500 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1501 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1504 dmu_tx_commit(dsa
->dsa_tx
);
1506 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1508 kmem_free(dsa
, sizeof (*dsa
));
1512 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1513 zio_prop_t
*zp
, zbookmark_t
*zb
)
1515 dmu_sync_arg_t
*dsa
;
1518 tx
= dmu_tx_create(os
);
1519 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1520 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1522 return (EIO
); /* Make zl_get_data do txg_waited_synced() */
1525 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1527 dsa
->dsa_done
= done
;
1531 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1532 zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
, zp
,
1533 dmu_sync_late_arrival_ready
, dmu_sync_late_arrival_done
, dsa
,
1534 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_FASTWRITE
, zb
));
1540 * Intent log support: sync the block associated with db to disk.
1541 * N.B. and XXX: the caller is responsible for making sure that the
1542 * data isn't changing while dmu_sync() is writing it.
1546 * EEXIST: this txg has already been synced, so there's nothing to to.
1547 * The caller should not log the write.
1549 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1550 * The caller should not log the write.
1552 * EALREADY: this block is already in the process of being synced.
1553 * The caller should track its progress (somehow).
1555 * EIO: could not do the I/O.
1556 * The caller should do a txg_wait_synced().
1558 * 0: the I/O has been initiated.
1559 * The caller should log this blkptr in the done callback.
1560 * It is possible that the I/O will fail, in which case
1561 * the error will be reported to the done callback and
1562 * propagated to pio from zio_done().
1565 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1567 blkptr_t
*bp
= zgd
->zgd_bp
;
1568 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1569 objset_t
*os
= db
->db_objset
;
1570 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1571 dbuf_dirty_record_t
*dr
;
1572 dmu_sync_arg_t
*dsa
;
1577 ASSERT(pio
!= NULL
);
1578 ASSERT(BP_IS_HOLE(bp
));
1581 SET_BOOKMARK(&zb
, ds
->ds_object
,
1582 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1586 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1590 * If we're frozen (running ziltest), we always need to generate a bp.
1592 if (txg
> spa_freeze_txg(os
->os_spa
))
1593 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1596 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1597 * and us. If we determine that this txg is not yet syncing,
1598 * but it begins to sync a moment later, that's OK because the
1599 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1601 mutex_enter(&db
->db_mtx
);
1603 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1605 * This txg has already synced. There's nothing to do.
1607 mutex_exit(&db
->db_mtx
);
1611 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1613 * This txg is currently syncing, so we can't mess with
1614 * the dirty record anymore; just write a new log block.
1616 mutex_exit(&db
->db_mtx
);
1617 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1620 dr
= db
->db_last_dirty
;
1621 while (dr
&& dr
->dr_txg
!= txg
)
1626 * There's no dr for this dbuf, so it must have been freed.
1627 * There's no need to log writes to freed blocks, so we're done.
1629 mutex_exit(&db
->db_mtx
);
1633 ASSERT(dr
->dr_txg
== txg
);
1634 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1635 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1637 * We have already issued a sync write for this buffer,
1638 * or this buffer has already been synced. It could not
1639 * have been dirtied since, or we would have cleared the state.
1641 mutex_exit(&db
->db_mtx
);
1645 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1646 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1647 mutex_exit(&db
->db_mtx
);
1649 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_PUSHPAGE
);
1651 dsa
->dsa_done
= done
;
1655 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1656 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
), &zp
,
1657 dmu_sync_ready
, dmu_sync_done
, dsa
,
1658 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_FASTWRITE
, &zb
));
1664 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1670 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1673 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1674 dnode_rele(dn
, FTAG
);
1679 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1684 /* XXX assumes dnode_hold will not get an i/o error */
1685 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1686 ASSERT(checksum
< ZIO_CHECKSUM_FUNCTIONS
);
1687 dn
->dn_checksum
= checksum
;
1688 dnode_setdirty(dn
, tx
);
1689 dnode_rele(dn
, FTAG
);
1693 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1698 /* XXX assumes dnode_hold will not get an i/o error */
1699 (void) dnode_hold(os
, object
, FTAG
, &dn
);
1700 ASSERT(compress
< ZIO_COMPRESS_FUNCTIONS
);
1701 dn
->dn_compress
= compress
;
1702 dnode_setdirty(dn
, tx
);
1703 dnode_rele(dn
, FTAG
);
1706 int zfs_mdcomp_disable
= 0;
1709 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1711 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1712 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1714 enum zio_checksum checksum
= os
->os_checksum
;
1715 enum zio_compress compress
= os
->os_compress
;
1716 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1718 boolean_t dedup_verify
= os
->os_dedup_verify
;
1719 int copies
= os
->os_copies
;
1722 * Determine checksum setting.
1726 * Metadata always gets checksummed. If the data
1727 * checksum is multi-bit correctable, and it's not a
1728 * ZBT-style checksum, then it's suitable for metadata
1729 * as well. Otherwise, the metadata checksum defaults
1732 if (zio_checksum_table
[checksum
].ci_correctable
< 1 ||
1733 zio_checksum_table
[checksum
].ci_eck
)
1734 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1736 checksum
= zio_checksum_select(dn
->dn_checksum
, checksum
);
1740 * Determine compression setting.
1744 * XXX -- we should design a compression algorithm
1745 * that specializes in arrays of bps.
1747 compress
= zfs_mdcomp_disable
? ZIO_COMPRESS_EMPTY
:
1750 compress
= zio_compress_select(dn
->dn_compress
, compress
);
1754 * Determine dedup setting. If we are in dmu_sync(), we won't
1755 * actually dedup now because that's all done in syncing context;
1756 * but we do want to use the dedup checkum. If the checksum is not
1757 * strong enough to ensure unique signatures, force dedup_verify.
1759 dedup
= (!ismd
&& dedup_checksum
!= ZIO_CHECKSUM_OFF
);
1761 checksum
= dedup_checksum
;
1762 if (!zio_checksum_table
[checksum
].ci_dedup
)
1766 if (wp
& WP_DMU_SYNC
)
1769 if (wp
& WP_NOFILL
) {
1770 ASSERT(!ismd
&& level
== 0);
1771 checksum
= ZIO_CHECKSUM_OFF
;
1772 compress
= ZIO_COMPRESS_OFF
;
1776 zp
->zp_checksum
= checksum
;
1777 zp
->zp_compress
= compress
;
1778 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1779 zp
->zp_level
= level
;
1780 zp
->zp_copies
= MIN(copies
+ ismd
, spa_max_replication(os
->os_spa
));
1781 zp
->zp_dedup
= dedup
;
1782 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1786 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
1791 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1795 * Sync any current changes before
1796 * we go trundling through the block pointers.
1798 for (i
= 0; i
< TXG_SIZE
; i
++) {
1799 if (list_link_active(&dn
->dn_dirty_link
[i
]))
1802 if (i
!= TXG_SIZE
) {
1803 dnode_rele(dn
, FTAG
);
1804 txg_wait_synced(dmu_objset_pool(os
), 0);
1805 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1810 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
1811 dnode_rele(dn
, FTAG
);
1817 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
1822 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1823 mutex_enter(&dn
->dn_mtx
);
1827 doi
->doi_data_block_size
= dn
->dn_datablksz
;
1828 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
1829 1ULL << dn
->dn_indblkshift
: 0;
1830 doi
->doi_type
= dn
->dn_type
;
1831 doi
->doi_bonus_type
= dn
->dn_bonustype
;
1832 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
1833 doi
->doi_indirection
= dn
->dn_nlevels
;
1834 doi
->doi_checksum
= dn
->dn_checksum
;
1835 doi
->doi_compress
= dn
->dn_compress
;
1836 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
1837 doi
->doi_max_offset
= (dnp
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
1838 doi
->doi_fill_count
= 0;
1839 for (i
= 0; i
< dnp
->dn_nblkptr
; i
++)
1840 doi
->doi_fill_count
+= dnp
->dn_blkptr
[i
].blk_fill
;
1842 mutex_exit(&dn
->dn_mtx
);
1843 rw_exit(&dn
->dn_struct_rwlock
);
1847 * Get information on a DMU object.
1848 * If doi is NULL, just indicates whether the object exists.
1851 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
1854 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
1860 dmu_object_info_from_dnode(dn
, doi
);
1862 dnode_rele(dn
, FTAG
);
1867 * As above, but faster; can be used when you have a held dbuf in hand.
1870 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
1872 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1875 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
1880 * Faster still when you only care about the size.
1881 * This is specifically optimized for zfs_getattr().
1884 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
1885 u_longlong_t
*nblk512
)
1887 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
1893 *blksize
= dn
->dn_datablksz
;
1894 /* add 1 for dnode space */
1895 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
1896 SPA_MINBLOCKSHIFT
) + 1;
1901 byteswap_uint64_array(void *vbuf
, size_t size
)
1903 uint64_t *buf
= vbuf
;
1904 size_t count
= size
>> 3;
1907 ASSERT((size
& 7) == 0);
1909 for (i
= 0; i
< count
; i
++)
1910 buf
[i
] = BSWAP_64(buf
[i
]);
1914 byteswap_uint32_array(void *vbuf
, size_t size
)
1916 uint32_t *buf
= vbuf
;
1917 size_t count
= size
>> 2;
1920 ASSERT((size
& 3) == 0);
1922 for (i
= 0; i
< count
; i
++)
1923 buf
[i
] = BSWAP_32(buf
[i
]);
1927 byteswap_uint16_array(void *vbuf
, size_t size
)
1929 uint16_t *buf
= vbuf
;
1930 size_t count
= size
>> 1;
1933 ASSERT((size
& 1) == 0);
1935 for (i
= 0; i
< count
; i
++)
1936 buf
[i
] = BSWAP_16(buf
[i
]);
1941 byteswap_uint8_array(void *vbuf
, size_t size
)
1975 #if defined(_KERNEL) && defined(HAVE_SPL)
1976 EXPORT_SYMBOL(dmu_bonus_hold
);
1977 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus
);
1978 EXPORT_SYMBOL(dmu_buf_rele_array
);
1979 EXPORT_SYMBOL(dmu_free_range
);
1980 EXPORT_SYMBOL(dmu_read
);
1981 EXPORT_SYMBOL(dmu_write
);
1982 EXPORT_SYMBOL(dmu_object_info
);
1983 EXPORT_SYMBOL(dmu_object_info_from_dnode
);
1984 EXPORT_SYMBOL(dmu_object_info_from_db
);
1985 EXPORT_SYMBOL(dmu_object_size_from_db
);
1986 EXPORT_SYMBOL(dmu_object_set_blocksize
);
1987 EXPORT_SYMBOL(dmu_object_set_checksum
);
1988 EXPORT_SYMBOL(dmu_object_set_compress
);
1989 EXPORT_SYMBOL(dmu_request_arcbuf
);
1990 EXPORT_SYMBOL(dmu_return_arcbuf
);
1991 EXPORT_SYMBOL(dmu_assign_arcbuf
);
1992 EXPORT_SYMBOL(dmu_buf_hold
);
1993 EXPORT_SYMBOL(dmu_ot
);
1995 module_param(zfs_mdcomp_disable
, int, 0644);
1996 MODULE_PARM_DESC(zfs_mdcomp_disable
, "Disable meta data compression");