4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 /* Copyright 2016 Nexenta Systems, Inc. All rights reserved. */
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
48 #include <sys/zfeature.h>
51 #include <sys/vmsystm.h>
52 #include <sys/zfs_znode.h>
56 * Enable/disable nopwrite feature.
58 int zfs_nopwrite_enabled
= 1;
61 * Tunable to control percentage of dirtied blocks from frees in one TXG.
62 * After this threshold is crossed, additional dirty blocks from frees
63 * wait until the next TXG.
64 * A value of zero will disable this throttle.
66 uint32_t zfs_per_txg_dirty_frees_percent
= 30;
68 const dmu_object_type_info_t dmu_ot
[DMU_OT_NUMTYPES
] = {
69 { DMU_BSWAP_UINT8
, TRUE
, "unallocated" },
70 { DMU_BSWAP_ZAP
, TRUE
, "object directory" },
71 { DMU_BSWAP_UINT64
, TRUE
, "object array" },
72 { DMU_BSWAP_UINT8
, TRUE
, "packed nvlist" },
73 { DMU_BSWAP_UINT64
, TRUE
, "packed nvlist size" },
74 { DMU_BSWAP_UINT64
, TRUE
, "bpobj" },
75 { DMU_BSWAP_UINT64
, TRUE
, "bpobj header" },
76 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map header" },
77 { DMU_BSWAP_UINT64
, TRUE
, "SPA space map" },
78 { DMU_BSWAP_UINT64
, TRUE
, "ZIL intent log" },
79 { DMU_BSWAP_DNODE
, TRUE
, "DMU dnode" },
80 { DMU_BSWAP_OBJSET
, TRUE
, "DMU objset" },
81 { DMU_BSWAP_UINT64
, TRUE
, "DSL directory" },
82 { DMU_BSWAP_ZAP
, TRUE
, "DSL directory child map"},
83 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset snap map" },
84 { DMU_BSWAP_ZAP
, TRUE
, "DSL props" },
85 { DMU_BSWAP_UINT64
, TRUE
, "DSL dataset" },
86 { DMU_BSWAP_ZNODE
, TRUE
, "ZFS znode" },
87 { DMU_BSWAP_OLDACL
, TRUE
, "ZFS V0 ACL" },
88 { DMU_BSWAP_UINT8
, FALSE
, "ZFS plain file" },
89 { DMU_BSWAP_ZAP
, TRUE
, "ZFS directory" },
90 { DMU_BSWAP_ZAP
, TRUE
, "ZFS master node" },
91 { DMU_BSWAP_ZAP
, TRUE
, "ZFS delete queue" },
92 { DMU_BSWAP_UINT8
, FALSE
, "zvol object" },
93 { DMU_BSWAP_ZAP
, TRUE
, "zvol prop" },
94 { DMU_BSWAP_UINT8
, FALSE
, "other uint8[]" },
95 { DMU_BSWAP_UINT64
, FALSE
, "other uint64[]" },
96 { DMU_BSWAP_ZAP
, TRUE
, "other ZAP" },
97 { DMU_BSWAP_ZAP
, TRUE
, "persistent error log" },
98 { DMU_BSWAP_UINT8
, TRUE
, "SPA history" },
99 { DMU_BSWAP_UINT64
, TRUE
, "SPA history offsets" },
100 { DMU_BSWAP_ZAP
, TRUE
, "Pool properties" },
101 { DMU_BSWAP_ZAP
, TRUE
, "DSL permissions" },
102 { DMU_BSWAP_ACL
, TRUE
, "ZFS ACL" },
103 { DMU_BSWAP_UINT8
, TRUE
, "ZFS SYSACL" },
104 { DMU_BSWAP_UINT8
, TRUE
, "FUID table" },
105 { DMU_BSWAP_UINT64
, TRUE
, "FUID table size" },
106 { DMU_BSWAP_ZAP
, TRUE
, "DSL dataset next clones"},
107 { DMU_BSWAP_ZAP
, TRUE
, "scan work queue" },
108 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group used" },
109 { DMU_BSWAP_ZAP
, TRUE
, "ZFS user/group quota" },
110 { DMU_BSWAP_ZAP
, TRUE
, "snapshot refcount tags"},
111 { DMU_BSWAP_ZAP
, TRUE
, "DDT ZAP algorithm" },
112 { DMU_BSWAP_ZAP
, TRUE
, "DDT statistics" },
113 { DMU_BSWAP_UINT8
, TRUE
, "System attributes" },
114 { DMU_BSWAP_ZAP
, TRUE
, "SA master node" },
115 { DMU_BSWAP_ZAP
, TRUE
, "SA attr registration" },
116 { DMU_BSWAP_ZAP
, TRUE
, "SA attr layouts" },
117 { DMU_BSWAP_ZAP
, TRUE
, "scan translations" },
118 { DMU_BSWAP_UINT8
, FALSE
, "deduplicated block" },
119 { DMU_BSWAP_ZAP
, TRUE
, "DSL deadlist map" },
120 { DMU_BSWAP_UINT64
, TRUE
, "DSL deadlist map hdr" },
121 { DMU_BSWAP_ZAP
, TRUE
, "DSL dir clones" },
122 { DMU_BSWAP_UINT64
, TRUE
, "bpobj subobj" }
125 const dmu_object_byteswap_info_t dmu_ot_byteswap
[DMU_BSWAP_NUMFUNCS
] = {
126 { byteswap_uint8_array
, "uint8" },
127 { byteswap_uint16_array
, "uint16" },
128 { byteswap_uint32_array
, "uint32" },
129 { byteswap_uint64_array
, "uint64" },
130 { zap_byteswap
, "zap" },
131 { dnode_buf_byteswap
, "dnode" },
132 { dmu_objset_byteswap
, "objset" },
133 { zfs_znode_byteswap
, "znode" },
134 { zfs_oldacl_byteswap
, "oldacl" },
135 { zfs_acl_byteswap
, "acl" }
139 dmu_buf_hold_noread_by_dnode(dnode_t
*dn
, uint64_t offset
,
140 void *tag
, dmu_buf_t
**dbp
)
145 blkid
= dbuf_whichblock(dn
, 0, offset
);
146 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
147 db
= dbuf_hold(dn
, blkid
, tag
);
148 rw_exit(&dn
->dn_struct_rwlock
);
152 return (SET_ERROR(EIO
));
159 dmu_buf_hold_noread(objset_t
*os
, uint64_t object
, uint64_t offset
,
160 void *tag
, dmu_buf_t
**dbp
)
167 err
= dnode_hold(os
, object
, FTAG
, &dn
);
170 blkid
= dbuf_whichblock(dn
, 0, offset
);
171 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
172 db
= dbuf_hold(dn
, blkid
, tag
);
173 rw_exit(&dn
->dn_struct_rwlock
);
174 dnode_rele(dn
, FTAG
);
178 return (SET_ERROR(EIO
));
186 dmu_buf_hold_by_dnode(dnode_t
*dn
, uint64_t offset
,
187 void *tag
, dmu_buf_t
**dbp
, int flags
)
190 int db_flags
= DB_RF_CANFAIL
;
192 if (flags
& DMU_READ_NO_PREFETCH
)
193 db_flags
|= DB_RF_NOPREFETCH
;
195 err
= dmu_buf_hold_noread_by_dnode(dn
, offset
, tag
, dbp
);
197 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
198 err
= dbuf_read(db
, NULL
, db_flags
);
209 dmu_buf_hold(objset_t
*os
, uint64_t object
, uint64_t offset
,
210 void *tag
, dmu_buf_t
**dbp
, int flags
)
213 int db_flags
= DB_RF_CANFAIL
;
215 if (flags
& DMU_READ_NO_PREFETCH
)
216 db_flags
|= DB_RF_NOPREFETCH
;
218 err
= dmu_buf_hold_noread(os
, object
, offset
, tag
, dbp
);
220 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)(*dbp
);
221 err
= dbuf_read(db
, NULL
, db_flags
);
234 return (DN_MAX_BONUSLEN
);
238 dmu_set_bonus(dmu_buf_t
*db_fake
, int newsize
, dmu_tx_t
*tx
)
240 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
247 if (dn
->dn_bonus
!= db
) {
248 error
= SET_ERROR(EINVAL
);
249 } else if (newsize
< 0 || newsize
> db_fake
->db_size
) {
250 error
= SET_ERROR(EINVAL
);
252 dnode_setbonuslen(dn
, newsize
, tx
);
261 dmu_set_bonustype(dmu_buf_t
*db_fake
, dmu_object_type_t type
, dmu_tx_t
*tx
)
263 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
270 if (!DMU_OT_IS_VALID(type
)) {
271 error
= SET_ERROR(EINVAL
);
272 } else if (dn
->dn_bonus
!= db
) {
273 error
= SET_ERROR(EINVAL
);
275 dnode_setbonus_type(dn
, type
, tx
);
284 dmu_get_bonustype(dmu_buf_t
*db_fake
)
286 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
288 dmu_object_type_t type
;
292 type
= dn
->dn_bonustype
;
299 dmu_rm_spill(objset_t
*os
, uint64_t object
, dmu_tx_t
*tx
)
304 error
= dnode_hold(os
, object
, FTAG
, &dn
);
305 dbuf_rm_spill(dn
, tx
);
306 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
307 dnode_rm_spill(dn
, tx
);
308 rw_exit(&dn
->dn_struct_rwlock
);
309 dnode_rele(dn
, FTAG
);
314 * returns ENOENT, EIO, or 0.
317 dmu_bonus_hold(objset_t
*os
, uint64_t object
, void *tag
, dmu_buf_t
**dbp
)
323 error
= dnode_hold(os
, object
, FTAG
, &dn
);
327 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
328 if (dn
->dn_bonus
== NULL
) {
329 rw_exit(&dn
->dn_struct_rwlock
);
330 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
331 if (dn
->dn_bonus
== NULL
)
332 dbuf_create_bonus(dn
);
336 /* as long as the bonus buf is held, the dnode will be held */
337 if (refcount_add(&db
->db_holds
, tag
) == 1) {
338 VERIFY(dnode_add_ref(dn
, db
));
339 atomic_inc_32(&dn
->dn_dbufs_count
);
343 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
344 * hold and incrementing the dbuf count to ensure that dnode_move() sees
345 * a dnode hold for every dbuf.
347 rw_exit(&dn
->dn_struct_rwlock
);
349 dnode_rele(dn
, FTAG
);
351 VERIFY(0 == dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
));
358 * returns ENOENT, EIO, or 0.
360 * This interface will allocate a blank spill dbuf when a spill blk
361 * doesn't already exist on the dnode.
363 * if you only want to find an already existing spill db, then
364 * dmu_spill_hold_existing() should be used.
367 dmu_spill_hold_by_dnode(dnode_t
*dn
, uint32_t flags
, void *tag
, dmu_buf_t
**dbp
)
369 dmu_buf_impl_t
*db
= NULL
;
372 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
373 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
375 db
= dbuf_hold(dn
, DMU_SPILL_BLKID
, tag
);
377 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
378 rw_exit(&dn
->dn_struct_rwlock
);
381 err
= dbuf_read(db
, NULL
, flags
);
390 dmu_spill_hold_existing(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
392 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
399 if (spa_version(dn
->dn_objset
->os_spa
) < SPA_VERSION_SA
) {
400 err
= SET_ERROR(EINVAL
);
402 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
404 if (!dn
->dn_have_spill
) {
405 err
= SET_ERROR(ENOENT
);
407 err
= dmu_spill_hold_by_dnode(dn
,
408 DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
, tag
, dbp
);
411 rw_exit(&dn
->dn_struct_rwlock
);
419 dmu_spill_hold_by_bonus(dmu_buf_t
*bonus
, void *tag
, dmu_buf_t
**dbp
)
421 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)bonus
;
427 err
= dmu_spill_hold_by_dnode(dn
, DB_RF_CANFAIL
, tag
, dbp
);
434 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
435 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
436 * and can induce severe lock contention when writing to several files
437 * whose dnodes are in the same block.
440 dmu_buf_hold_array_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t length
,
441 boolean_t read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
, uint32_t flags
)
444 uint64_t blkid
, nblks
, i
;
449 ASSERT(length
<= DMU_MAX_ACCESS
);
452 * Note: We directly notify the prefetch code of this read, so that
453 * we can tell it about the multi-block read. dbuf_read() only knows
454 * about the one block it is accessing.
456 dbuf_flags
= DB_RF_CANFAIL
| DB_RF_NEVERWAIT
| DB_RF_HAVESTRUCT
|
459 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
460 if (dn
->dn_datablkshift
) {
461 int blkshift
= dn
->dn_datablkshift
;
462 nblks
= (P2ROUNDUP(offset
+ length
, 1ULL << blkshift
) -
463 P2ALIGN(offset
, 1ULL << blkshift
)) >> blkshift
;
465 if (offset
+ length
> dn
->dn_datablksz
) {
466 zfs_panic_recover("zfs: accessing past end of object "
467 "%llx/%llx (size=%u access=%llu+%llu)",
468 (longlong_t
)dn
->dn_objset
->
469 os_dsl_dataset
->ds_object
,
470 (longlong_t
)dn
->dn_object
, dn
->dn_datablksz
,
471 (longlong_t
)offset
, (longlong_t
)length
);
472 rw_exit(&dn
->dn_struct_rwlock
);
473 return (SET_ERROR(EIO
));
477 dbp
= kmem_zalloc(sizeof (dmu_buf_t
*) * nblks
, KM_SLEEP
);
479 zio
= zio_root(dn
->dn_objset
->os_spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
480 blkid
= dbuf_whichblock(dn
, 0, offset
);
481 for (i
= 0; i
< nblks
; i
++) {
482 dmu_buf_impl_t
*db
= dbuf_hold(dn
, blkid
+ i
, tag
);
484 rw_exit(&dn
->dn_struct_rwlock
);
485 dmu_buf_rele_array(dbp
, nblks
, tag
);
487 return (SET_ERROR(EIO
));
490 /* initiate async i/o */
492 (void) dbuf_read(db
, zio
, dbuf_flags
);
496 if ((flags
& DMU_READ_NO_PREFETCH
) == 0 &&
497 DNODE_META_IS_CACHEABLE(dn
) && length
<= zfetch_array_rd_sz
) {
498 dmu_zfetch(&dn
->dn_zfetch
, blkid
, nblks
,
499 read
&& DNODE_IS_CACHEABLE(dn
));
501 rw_exit(&dn
->dn_struct_rwlock
);
503 /* wait for async i/o */
506 dmu_buf_rele_array(dbp
, nblks
, tag
);
510 /* wait for other io to complete */
512 for (i
= 0; i
< nblks
; i
++) {
513 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbp
[i
];
514 mutex_enter(&db
->db_mtx
);
515 while (db
->db_state
== DB_READ
||
516 db
->db_state
== DB_FILL
)
517 cv_wait(&db
->db_changed
, &db
->db_mtx
);
518 if (db
->db_state
== DB_UNCACHED
)
519 err
= SET_ERROR(EIO
);
520 mutex_exit(&db
->db_mtx
);
522 dmu_buf_rele_array(dbp
, nblks
, tag
);
534 dmu_buf_hold_array(objset_t
*os
, uint64_t object
, uint64_t offset
,
535 uint64_t length
, int read
, void *tag
, int *numbufsp
, dmu_buf_t
***dbpp
)
540 err
= dnode_hold(os
, object
, FTAG
, &dn
);
544 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
545 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
547 dnode_rele(dn
, FTAG
);
553 dmu_buf_hold_array_by_bonus(dmu_buf_t
*db_fake
, uint64_t offset
,
554 uint64_t length
, boolean_t read
, void *tag
, int *numbufsp
,
557 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
563 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, length
, read
, tag
,
564 numbufsp
, dbpp
, DMU_READ_PREFETCH
);
571 dmu_buf_rele_array(dmu_buf_t
**dbp_fake
, int numbufs
, void *tag
)
574 dmu_buf_impl_t
**dbp
= (dmu_buf_impl_t
**)dbp_fake
;
579 for (i
= 0; i
< numbufs
; i
++) {
581 dbuf_rele(dbp
[i
], tag
);
584 kmem_free(dbp
, sizeof (dmu_buf_t
*) * numbufs
);
588 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
589 * indirect blocks prefeteched will be those that point to the blocks containing
590 * the data starting at offset, and continuing to offset + len.
592 * Note that if the indirect blocks above the blocks being prefetched are not in
593 * cache, they will be asychronously read in.
596 dmu_prefetch(objset_t
*os
, uint64_t object
, int64_t level
, uint64_t offset
,
597 uint64_t len
, zio_priority_t pri
)
603 if (len
== 0) { /* they're interested in the bonus buffer */
604 dn
= DMU_META_DNODE(os
);
606 if (object
== 0 || object
>= DN_MAX_OBJECT
)
609 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
610 blkid
= dbuf_whichblock(dn
, level
,
611 object
* sizeof (dnode_phys_t
));
612 dbuf_prefetch(dn
, level
, blkid
, pri
, 0);
613 rw_exit(&dn
->dn_struct_rwlock
);
618 * XXX - Note, if the dnode for the requested object is not
619 * already cached, we will do a *synchronous* read in the
620 * dnode_hold() call. The same is true for any indirects.
622 err
= dnode_hold(os
, object
, FTAG
, &dn
);
626 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
628 * offset + len - 1 is the last byte we want to prefetch for, and offset
629 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
630 * last block we want to prefetch, and dbuf_whichblock(dn, level,
631 * offset) is the first. Then the number we need to prefetch is the
634 if (level
> 0 || dn
->dn_datablkshift
!= 0) {
635 nblks
= dbuf_whichblock(dn
, level
, offset
+ len
- 1) -
636 dbuf_whichblock(dn
, level
, offset
) + 1;
638 nblks
= (offset
< dn
->dn_datablksz
);
642 blkid
= dbuf_whichblock(dn
, level
, offset
);
643 for (int i
= 0; i
< nblks
; i
++)
644 dbuf_prefetch(dn
, level
, blkid
+ i
, pri
, 0);
647 rw_exit(&dn
->dn_struct_rwlock
);
649 dnode_rele(dn
, FTAG
);
653 * Get the next "chunk" of file data to free. We traverse the file from
654 * the end so that the file gets shorter over time (if we crashes in the
655 * middle, this will leave us in a better state). We find allocated file
656 * data by simply searching the allocated level 1 indirects.
658 * On input, *start should be the first offset that does not need to be
659 * freed (e.g. "offset + length"). On return, *start will be the first
660 * offset that should be freed.
663 get_next_chunk(dnode_t
*dn
, uint64_t *start
, uint64_t minimum
)
665 uint64_t maxblks
= DMU_MAX_ACCESS
>> (dn
->dn_indblkshift
+ 1);
666 /* bytes of data covered by a level-1 indirect block */
668 dn
->dn_datablksz
* EPB(dn
->dn_indblkshift
, SPA_BLKPTRSHIFT
);
670 ASSERT3U(minimum
, <=, *start
);
672 if (*start
- minimum
<= iblkrange
* maxblks
) {
676 ASSERT(ISP2(iblkrange
));
678 for (uint64_t blks
= 0; *start
> minimum
&& blks
< maxblks
; blks
++) {
682 * dnode_next_offset(BACKWARDS) will find an allocated L1
683 * indirect block at or before the input offset. We must
684 * decrement *start so that it is at the end of the region
688 err
= dnode_next_offset(dn
,
689 DNODE_FIND_BACKWARDS
, start
, 2, 1, 0);
691 /* if there are no indirect blocks before start, we are done */
695 } else if (err
!= 0) {
699 /* set start to the beginning of this L1 indirect */
700 *start
= P2ALIGN(*start
, iblkrange
);
702 if (*start
< minimum
)
708 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
709 * otherwise return false.
710 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
714 dmu_objset_zfs_unmounting(objset_t
*os
)
717 if (dmu_objset_type(os
) == DMU_OST_ZFS
)
718 return (zfs_get_vfs_flag_unmounted(os
));
724 dmu_free_long_range_impl(objset_t
*os
, dnode_t
*dn
, uint64_t offset
,
727 uint64_t object_size
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
729 uint64_t dirty_frees_threshold
;
730 dsl_pool_t
*dp
= dmu_objset_pool(os
);
732 if (offset
>= object_size
)
735 if (zfs_per_txg_dirty_frees_percent
<= 100)
736 dirty_frees_threshold
=
737 zfs_per_txg_dirty_frees_percent
* zfs_dirty_data_max
/ 100;
739 dirty_frees_threshold
= zfs_dirty_data_max
/ 4;
741 if (length
== DMU_OBJECT_END
|| offset
+ length
> object_size
)
742 length
= object_size
- offset
;
744 while (length
!= 0) {
745 uint64_t chunk_end
, chunk_begin
, chunk_len
;
746 uint64_t long_free_dirty_all_txgs
= 0;
749 if (dmu_objset_zfs_unmounting(dn
->dn_objset
))
750 return (SET_ERROR(EINTR
));
752 chunk_end
= chunk_begin
= offset
+ length
;
754 /* move chunk_begin backwards to the beginning of this chunk */
755 err
= get_next_chunk(dn
, &chunk_begin
, offset
);
758 ASSERT3U(chunk_begin
, >=, offset
);
759 ASSERT3U(chunk_begin
, <=, chunk_end
);
761 chunk_len
= chunk_end
- chunk_begin
;
763 mutex_enter(&dp
->dp_lock
);
764 for (int t
= 0; t
< TXG_SIZE
; t
++) {
765 long_free_dirty_all_txgs
+=
766 dp
->dp_long_free_dirty_pertxg
[t
];
768 mutex_exit(&dp
->dp_lock
);
771 * To avoid filling up a TXG with just frees wait for
772 * the next TXG to open before freeing more chunks if
773 * we have reached the threshold of frees
775 if (dirty_frees_threshold
!= 0 &&
776 long_free_dirty_all_txgs
>= dirty_frees_threshold
) {
777 txg_wait_open(dp
, 0);
781 tx
= dmu_tx_create(os
);
782 dmu_tx_hold_free(tx
, dn
->dn_object
, chunk_begin
, chunk_len
);
785 * Mark this transaction as typically resulting in a net
786 * reduction in space used.
788 dmu_tx_mark_netfree(tx
);
789 err
= dmu_tx_assign(tx
, TXG_WAIT
);
795 mutex_enter(&dp
->dp_lock
);
796 dp
->dp_long_free_dirty_pertxg
[dmu_tx_get_txg(tx
) & TXG_MASK
] +=
798 mutex_exit(&dp
->dp_lock
);
799 DTRACE_PROBE3(free__long__range
,
800 uint64_t, long_free_dirty_all_txgs
, uint64_t, chunk_len
,
801 uint64_t, dmu_tx_get_txg(tx
));
802 dnode_free_range(dn
, chunk_begin
, chunk_len
, tx
);
811 dmu_free_long_range(objset_t
*os
, uint64_t object
,
812 uint64_t offset
, uint64_t length
)
817 err
= dnode_hold(os
, object
, FTAG
, &dn
);
820 err
= dmu_free_long_range_impl(os
, dn
, offset
, length
);
823 * It is important to zero out the maxblkid when freeing the entire
824 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
825 * will take the fast path, and (b) dnode_reallocate() can verify
826 * that the entire file has been freed.
828 if (err
== 0 && offset
== 0 && length
== DMU_OBJECT_END
)
831 dnode_rele(dn
, FTAG
);
836 dmu_free_long_object(objset_t
*os
, uint64_t object
)
841 err
= dmu_free_long_range(os
, object
, 0, DMU_OBJECT_END
);
845 tx
= dmu_tx_create(os
);
846 dmu_tx_hold_bonus(tx
, object
);
847 dmu_tx_hold_free(tx
, object
, 0, DMU_OBJECT_END
);
848 dmu_tx_mark_netfree(tx
);
849 err
= dmu_tx_assign(tx
, TXG_WAIT
);
851 err
= dmu_object_free(os
, object
, tx
);
861 dmu_free_range(objset_t
*os
, uint64_t object
, uint64_t offset
,
862 uint64_t size
, dmu_tx_t
*tx
)
865 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
868 ASSERT(offset
< UINT64_MAX
);
869 ASSERT(size
== -1ULL || size
<= UINT64_MAX
- offset
);
870 dnode_free_range(dn
, offset
, size
, tx
);
871 dnode_rele(dn
, FTAG
);
876 dmu_read_impl(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
877 void *buf
, uint32_t flags
)
880 int numbufs
, err
= 0;
883 * Deal with odd block sizes, where there can't be data past the first
884 * block. If we ever do the tail block optimization, we will need to
885 * handle that here as well.
887 if (dn
->dn_maxblkid
== 0) {
888 int newsz
= offset
> dn
->dn_datablksz
? 0 :
889 MIN(size
, dn
->dn_datablksz
- offset
);
890 bzero((char *)buf
+ newsz
, size
- newsz
);
895 uint64_t mylen
= MIN(size
, DMU_MAX_ACCESS
/ 2);
899 * NB: we could do this block-at-a-time, but it's nice
900 * to be reading in parallel.
902 err
= dmu_buf_hold_array_by_dnode(dn
, offset
, mylen
,
903 TRUE
, FTAG
, &numbufs
, &dbp
, flags
);
907 for (i
= 0; i
< numbufs
; i
++) {
910 dmu_buf_t
*db
= dbp
[i
];
914 bufoff
= offset
- db
->db_offset
;
915 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
917 bcopy((char *)db
->db_data
+ bufoff
, buf
, tocpy
);
921 buf
= (char *)buf
+ tocpy
;
923 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
929 dmu_read(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
930 void *buf
, uint32_t flags
)
935 err
= dnode_hold(os
, object
, FTAG
, &dn
);
939 err
= dmu_read_impl(dn
, offset
, size
, buf
, flags
);
940 dnode_rele(dn
, FTAG
);
945 dmu_read_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
, void *buf
,
948 return (dmu_read_impl(dn
, offset
, size
, buf
, flags
));
952 dmu_write_impl(dmu_buf_t
**dbp
, int numbufs
, uint64_t offset
, uint64_t size
,
953 const void *buf
, dmu_tx_t
*tx
)
957 for (i
= 0; i
< numbufs
; i
++) {
960 dmu_buf_t
*db
= dbp
[i
];
964 bufoff
= offset
- db
->db_offset
;
965 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
967 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
969 if (tocpy
== db
->db_size
)
970 dmu_buf_will_fill(db
, tx
);
972 dmu_buf_will_dirty(db
, tx
);
974 bcopy(buf
, (char *)db
->db_data
+ bufoff
, tocpy
);
976 if (tocpy
== db
->db_size
)
977 dmu_buf_fill_done(db
, tx
);
981 buf
= (char *)buf
+ tocpy
;
986 dmu_write(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
987 const void *buf
, dmu_tx_t
*tx
)
995 VERIFY0(dmu_buf_hold_array(os
, object
, offset
, size
,
996 FALSE
, FTAG
, &numbufs
, &dbp
));
997 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
998 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1002 dmu_write_by_dnode(dnode_t
*dn
, uint64_t offset
, uint64_t size
,
1003 const void *buf
, dmu_tx_t
*tx
)
1011 VERIFY0(dmu_buf_hold_array_by_dnode(dn
, offset
, size
,
1012 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
));
1013 dmu_write_impl(dbp
, numbufs
, offset
, size
, buf
, tx
);
1014 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1018 dmu_prealloc(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1027 VERIFY(0 == dmu_buf_hold_array(os
, object
, offset
, size
,
1028 FALSE
, FTAG
, &numbufs
, &dbp
));
1030 for (i
= 0; i
< numbufs
; i
++) {
1031 dmu_buf_t
*db
= dbp
[i
];
1033 dmu_buf_will_not_fill(db
, tx
);
1035 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1039 dmu_write_embedded(objset_t
*os
, uint64_t object
, uint64_t offset
,
1040 void *data
, uint8_t etype
, uint8_t comp
, int uncompressed_size
,
1041 int compressed_size
, int byteorder
, dmu_tx_t
*tx
)
1045 ASSERT3U(etype
, <, NUM_BP_EMBEDDED_TYPES
);
1046 ASSERT3U(comp
, <, ZIO_COMPRESS_FUNCTIONS
);
1047 VERIFY0(dmu_buf_hold_noread(os
, object
, offset
,
1050 dmu_buf_write_embedded(db
,
1051 data
, (bp_embedded_type_t
)etype
, (enum zio_compress
)comp
,
1052 uncompressed_size
, compressed_size
, byteorder
, tx
);
1054 dmu_buf_rele(db
, FTAG
);
1058 * DMU support for xuio
1060 kstat_t
*xuio_ksp
= NULL
;
1063 dmu_xuio_init(xuio_t
*xuio
, int nblk
)
1066 uio_t
*uio
= &xuio
->xu_uio
;
1068 uio
->uio_iovcnt
= nblk
;
1069 uio
->uio_iov
= kmem_zalloc(nblk
* sizeof (iovec_t
), KM_SLEEP
);
1071 priv
= kmem_zalloc(sizeof (dmu_xuio_t
), KM_SLEEP
);
1073 priv
->bufs
= kmem_zalloc(nblk
* sizeof (arc_buf_t
*), KM_SLEEP
);
1074 priv
->iovp
= uio
->uio_iov
;
1075 XUIO_XUZC_PRIV(xuio
) = priv
;
1077 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1078 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, nblk
);
1080 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, nblk
);
1086 dmu_xuio_fini(xuio_t
*xuio
)
1088 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1089 int nblk
= priv
->cnt
;
1091 kmem_free(priv
->iovp
, nblk
* sizeof (iovec_t
));
1092 kmem_free(priv
->bufs
, nblk
* sizeof (arc_buf_t
*));
1093 kmem_free(priv
, sizeof (dmu_xuio_t
));
1095 if (XUIO_XUZC_RW(xuio
) == UIO_READ
)
1096 XUIOSTAT_INCR(xuiostat_onloan_rbuf
, -nblk
);
1098 XUIOSTAT_INCR(xuiostat_onloan_wbuf
, -nblk
);
1102 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1103 * and increase priv->next by 1.
1106 dmu_xuio_add(xuio_t
*xuio
, arc_buf_t
*abuf
, offset_t off
, size_t n
)
1109 uio_t
*uio
= &xuio
->xu_uio
;
1110 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1111 int i
= priv
->next
++;
1113 ASSERT(i
< priv
->cnt
);
1114 ASSERT(off
+ n
<= arc_buf_lsize(abuf
));
1115 iov
= uio
->uio_iov
+ i
;
1116 iov
->iov_base
= (char *)abuf
->b_data
+ off
;
1118 priv
->bufs
[i
] = abuf
;
1123 dmu_xuio_cnt(xuio_t
*xuio
)
1125 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1130 dmu_xuio_arcbuf(xuio_t
*xuio
, int i
)
1132 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1134 ASSERT(i
< priv
->cnt
);
1135 return (priv
->bufs
[i
]);
1139 dmu_xuio_clear(xuio_t
*xuio
, int i
)
1141 dmu_xuio_t
*priv
= XUIO_XUZC_PRIV(xuio
);
1143 ASSERT(i
< priv
->cnt
);
1144 priv
->bufs
[i
] = NULL
;
1148 xuio_stat_init(void)
1150 xuio_ksp
= kstat_create("zfs", 0, "xuio_stats", "misc",
1151 KSTAT_TYPE_NAMED
, sizeof (xuio_stats
) / sizeof (kstat_named_t
),
1152 KSTAT_FLAG_VIRTUAL
);
1153 if (xuio_ksp
!= NULL
) {
1154 xuio_ksp
->ks_data
= &xuio_stats
;
1155 kstat_install(xuio_ksp
);
1160 xuio_stat_fini(void)
1162 if (xuio_ksp
!= NULL
) {
1163 kstat_delete(xuio_ksp
);
1169 xuio_stat_wbuf_copied(void)
1171 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1175 xuio_stat_wbuf_nocopy(void)
1177 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy
);
1182 dmu_read_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
)
1185 int numbufs
, i
, err
;
1186 xuio_t
*xuio
= NULL
;
1189 * NB: we could do this block-at-a-time, but it's nice
1190 * to be reading in parallel.
1192 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1193 TRUE
, FTAG
, &numbufs
, &dbp
, 0);
1197 if (uio
->uio_extflg
== UIO_XUIO
)
1198 xuio
= (xuio_t
*)uio
;
1200 for (i
= 0; i
< numbufs
; i
++) {
1203 dmu_buf_t
*db
= dbp
[i
];
1207 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1208 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1211 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
1212 arc_buf_t
*dbuf_abuf
= dbi
->db_buf
;
1213 arc_buf_t
*abuf
= dbuf_loan_arcbuf(dbi
);
1214 err
= dmu_xuio_add(xuio
, abuf
, bufoff
, tocpy
);
1216 uio
->uio_resid
-= tocpy
;
1217 uio
->uio_loffset
+= tocpy
;
1220 if (abuf
== dbuf_abuf
)
1221 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy
);
1223 XUIOSTAT_BUMP(xuiostat_rbuf_copied
);
1225 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1233 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1239 * Read 'size' bytes into the uio buffer.
1240 * From object zdb->db_object.
1241 * Starting at offset uio->uio_loffset.
1243 * If the caller already has a dbuf in the target object
1244 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1245 * because we don't have to find the dnode_t for the object.
1248 dmu_read_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
)
1250 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1259 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1266 * Read 'size' bytes into the uio buffer.
1267 * From the specified object
1268 * Starting at offset uio->uio_loffset.
1271 dmu_read_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
)
1279 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1283 err
= dmu_read_uio_dnode(dn
, uio
, size
);
1285 dnode_rele(dn
, FTAG
);
1291 dmu_write_uio_dnode(dnode_t
*dn
, uio_t
*uio
, uint64_t size
, dmu_tx_t
*tx
)
1298 err
= dmu_buf_hold_array_by_dnode(dn
, uio
->uio_loffset
, size
,
1299 FALSE
, FTAG
, &numbufs
, &dbp
, DMU_READ_PREFETCH
);
1303 for (i
= 0; i
< numbufs
; i
++) {
1306 dmu_buf_t
*db
= dbp
[i
];
1310 bufoff
= uio
->uio_loffset
- db
->db_offset
;
1311 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1313 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1315 if (tocpy
== db
->db_size
)
1316 dmu_buf_will_fill(db
, tx
);
1318 dmu_buf_will_dirty(db
, tx
);
1321 * XXX uiomove could block forever (eg. nfs-backed
1322 * pages). There needs to be a uiolockdown() function
1323 * to lock the pages in memory, so that uiomove won't
1326 err
= uiomove((char *)db
->db_data
+ bufoff
, tocpy
,
1329 if (tocpy
== db
->db_size
)
1330 dmu_buf_fill_done(db
, tx
);
1338 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1343 * Write 'size' bytes from the uio buffer.
1344 * To object zdb->db_object.
1345 * Starting at offset uio->uio_loffset.
1347 * If the caller already has a dbuf in the target object
1348 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1349 * because we don't have to find the dnode_t for the object.
1352 dmu_write_uio_dbuf(dmu_buf_t
*zdb
, uio_t
*uio
, uint64_t size
,
1355 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zdb
;
1364 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1371 * Write 'size' bytes from the uio buffer.
1372 * To the specified object.
1373 * Starting at offset uio->uio_loffset.
1376 dmu_write_uio(objset_t
*os
, uint64_t object
, uio_t
*uio
, uint64_t size
,
1385 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1389 err
= dmu_write_uio_dnode(dn
, uio
, size
, tx
);
1391 dnode_rele(dn
, FTAG
);
1397 dmu_write_pages(objset_t
*os
, uint64_t object
, uint64_t offset
, uint64_t size
,
1398 page_t
*pp
, dmu_tx_t
*tx
)
1407 err
= dmu_buf_hold_array(os
, object
, offset
, size
,
1408 FALSE
, FTAG
, &numbufs
, &dbp
);
1412 for (i
= 0; i
< numbufs
; i
++) {
1413 int tocpy
, copied
, thiscpy
;
1415 dmu_buf_t
*db
= dbp
[i
];
1419 ASSERT3U(db
->db_size
, >=, PAGESIZE
);
1421 bufoff
= offset
- db
->db_offset
;
1422 tocpy
= (int)MIN(db
->db_size
- bufoff
, size
);
1424 ASSERT(i
== 0 || i
== numbufs
-1 || tocpy
== db
->db_size
);
1426 if (tocpy
== db
->db_size
)
1427 dmu_buf_will_fill(db
, tx
);
1429 dmu_buf_will_dirty(db
, tx
);
1431 for (copied
= 0; copied
< tocpy
; copied
+= PAGESIZE
) {
1432 ASSERT3U(pp
->p_offset
, ==, db
->db_offset
+ bufoff
);
1433 thiscpy
= MIN(PAGESIZE
, tocpy
- copied
);
1434 va
= zfs_map_page(pp
, S_READ
);
1435 bcopy(va
, (char *)db
->db_data
+ bufoff
, thiscpy
);
1436 zfs_unmap_page(pp
, va
);
1441 if (tocpy
== db
->db_size
)
1442 dmu_buf_fill_done(db
, tx
);
1447 dmu_buf_rele_array(dbp
, numbufs
, FTAG
);
1453 * Allocate a loaned anonymous arc buffer.
1456 dmu_request_arcbuf(dmu_buf_t
*handle
, int size
)
1458 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)handle
;
1460 return (arc_loan_buf(db
->db_objset
->os_spa
, B_FALSE
, size
));
1464 * Free a loaned arc buffer.
1467 dmu_return_arcbuf(arc_buf_t
*buf
)
1469 arc_return_buf(buf
, FTAG
);
1470 arc_buf_destroy(buf
, FTAG
);
1474 * When possible directly assign passed loaned arc buffer to a dbuf.
1475 * If this is not possible copy the contents of passed arc buf via
1479 dmu_assign_arcbuf(dmu_buf_t
*handle
, uint64_t offset
, arc_buf_t
*buf
,
1482 dmu_buf_impl_t
*dbuf
= (dmu_buf_impl_t
*)handle
;
1485 uint32_t blksz
= (uint32_t)arc_buf_lsize(buf
);
1488 DB_DNODE_ENTER(dbuf
);
1489 dn
= DB_DNODE(dbuf
);
1490 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1491 blkid
= dbuf_whichblock(dn
, 0, offset
);
1492 VERIFY((db
= dbuf_hold(dn
, blkid
, FTAG
)) != NULL
);
1493 rw_exit(&dn
->dn_struct_rwlock
);
1494 DB_DNODE_EXIT(dbuf
);
1497 * We can only assign if the offset is aligned, the arc buf is the
1498 * same size as the dbuf, and the dbuf is not metadata.
1500 if (offset
== db
->db
.db_offset
&& blksz
== db
->db
.db_size
) {
1501 dbuf_assign_arcbuf(db
, buf
, tx
);
1502 dbuf_rele(db
, FTAG
);
1507 /* compressed bufs must always be assignable to their dbuf */
1508 ASSERT3U(arc_get_compression(buf
), ==, ZIO_COMPRESS_OFF
);
1509 ASSERT(!(buf
->b_flags
& ARC_BUF_FLAG_COMPRESSED
));
1511 DB_DNODE_ENTER(dbuf
);
1512 dn
= DB_DNODE(dbuf
);
1514 object
= dn
->dn_object
;
1515 DB_DNODE_EXIT(dbuf
);
1517 dbuf_rele(db
, FTAG
);
1518 dmu_write(os
, object
, offset
, blksz
, buf
->b_data
, tx
);
1519 dmu_return_arcbuf(buf
);
1520 XUIOSTAT_BUMP(xuiostat_wbuf_copied
);
1525 dbuf_dirty_record_t
*dsa_dr
;
1526 dmu_sync_cb_t
*dsa_done
;
1533 dmu_sync_ready(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1535 dmu_sync_arg_t
*dsa
= varg
;
1536 dmu_buf_t
*db
= dsa
->dsa_zgd
->zgd_db
;
1537 blkptr_t
*bp
= zio
->io_bp
;
1539 if (zio
->io_error
== 0) {
1540 if (BP_IS_HOLE(bp
)) {
1542 * A block of zeros may compress to a hole, but the
1543 * block size still needs to be known for replay.
1545 BP_SET_LSIZE(bp
, db
->db_size
);
1546 } else if (!BP_IS_EMBEDDED(bp
)) {
1547 ASSERT(BP_GET_LEVEL(bp
) == 0);
1554 dmu_sync_late_arrival_ready(zio_t
*zio
)
1556 dmu_sync_ready(zio
, NULL
, zio
->io_private
);
1561 dmu_sync_done(zio_t
*zio
, arc_buf_t
*buf
, void *varg
)
1563 dmu_sync_arg_t
*dsa
= varg
;
1564 dbuf_dirty_record_t
*dr
= dsa
->dsa_dr
;
1565 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1567 mutex_enter(&db
->db_mtx
);
1568 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
);
1569 if (zio
->io_error
== 0) {
1570 dr
->dt
.dl
.dr_nopwrite
= !!(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
1571 if (dr
->dt
.dl
.dr_nopwrite
) {
1572 blkptr_t
*bp
= zio
->io_bp
;
1573 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
1574 uint8_t chksum
= BP_GET_CHECKSUM(bp_orig
);
1576 ASSERT(BP_EQUAL(bp
, bp_orig
));
1577 ASSERT(zio
->io_prop
.zp_compress
!= ZIO_COMPRESS_OFF
);
1578 ASSERT(zio_checksum_table
[chksum
].ci_flags
&
1579 ZCHECKSUM_FLAG_NOPWRITE
);
1581 dr
->dt
.dl
.dr_overridden_by
= *zio
->io_bp
;
1582 dr
->dt
.dl
.dr_override_state
= DR_OVERRIDDEN
;
1583 dr
->dt
.dl
.dr_copies
= zio
->io_prop
.zp_copies
;
1586 * Old style holes are filled with all zeros, whereas
1587 * new-style holes maintain their lsize, type, level,
1588 * and birth time (see zio_write_compress). While we
1589 * need to reset the BP_SET_LSIZE() call that happened
1590 * in dmu_sync_ready for old style holes, we do *not*
1591 * want to wipe out the information contained in new
1592 * style holes. Thus, only zero out the block pointer if
1593 * it's an old style hole.
1595 if (BP_IS_HOLE(&dr
->dt
.dl
.dr_overridden_by
) &&
1596 dr
->dt
.dl
.dr_overridden_by
.blk_birth
== 0)
1597 BP_ZERO(&dr
->dt
.dl
.dr_overridden_by
);
1599 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1601 cv_broadcast(&db
->db_changed
);
1602 mutex_exit(&db
->db_mtx
);
1604 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1606 kmem_free(dsa
, sizeof (*dsa
));
1610 dmu_sync_late_arrival_done(zio_t
*zio
)
1612 blkptr_t
*bp
= zio
->io_bp
;
1613 dmu_sync_arg_t
*dsa
= zio
->io_private
;
1614 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
1616 if (zio
->io_error
== 0 && !BP_IS_HOLE(bp
)) {
1618 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1619 * then there is nothing to do here. Otherwise, free the
1620 * newly allocated block in this txg.
1622 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
1623 ASSERT(BP_EQUAL(bp
, bp_orig
));
1625 ASSERT(BP_IS_HOLE(bp_orig
) || !BP_EQUAL(bp
, bp_orig
));
1626 ASSERT(zio
->io_bp
->blk_birth
== zio
->io_txg
);
1627 ASSERT(zio
->io_txg
> spa_syncing_txg(zio
->io_spa
));
1628 zio_free(zio
->io_spa
, zio
->io_txg
, zio
->io_bp
);
1632 dmu_tx_commit(dsa
->dsa_tx
);
1634 dsa
->dsa_done(dsa
->dsa_zgd
, zio
->io_error
);
1636 abd_put(zio
->io_abd
);
1637 kmem_free(dsa
, sizeof (*dsa
));
1641 dmu_sync_late_arrival(zio_t
*pio
, objset_t
*os
, dmu_sync_cb_t
*done
, zgd_t
*zgd
,
1642 zio_prop_t
*zp
, zbookmark_phys_t
*zb
)
1644 dmu_sync_arg_t
*dsa
;
1647 tx
= dmu_tx_create(os
);
1648 dmu_tx_hold_space(tx
, zgd
->zgd_db
->db_size
);
1649 if (dmu_tx_assign(tx
, TXG_WAIT
) != 0) {
1651 /* Make zl_get_data do txg_waited_synced() */
1652 return (SET_ERROR(EIO
));
1655 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1657 dsa
->dsa_done
= done
;
1661 zio_nowait(zio_write(pio
, os
->os_spa
, dmu_tx_get_txg(tx
), zgd
->zgd_bp
,
1662 abd_get_from_buf(zgd
->zgd_db
->db_data
, zgd
->zgd_db
->db_size
),
1663 zgd
->zgd_db
->db_size
, zgd
->zgd_db
->db_size
, zp
,
1664 dmu_sync_late_arrival_ready
, NULL
, NULL
, dmu_sync_late_arrival_done
,
1665 dsa
, ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, zb
));
1671 * Intent log support: sync the block associated with db to disk.
1672 * N.B. and XXX: the caller is responsible for making sure that the
1673 * data isn't changing while dmu_sync() is writing it.
1677 * EEXIST: this txg has already been synced, so there's nothing to do.
1678 * The caller should not log the write.
1680 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1681 * The caller should not log the write.
1683 * EALREADY: this block is already in the process of being synced.
1684 * The caller should track its progress (somehow).
1686 * EIO: could not do the I/O.
1687 * The caller should do a txg_wait_synced().
1689 * 0: the I/O has been initiated.
1690 * The caller should log this blkptr in the done callback.
1691 * It is possible that the I/O will fail, in which case
1692 * the error will be reported to the done callback and
1693 * propagated to pio from zio_done().
1696 dmu_sync(zio_t
*pio
, uint64_t txg
, dmu_sync_cb_t
*done
, zgd_t
*zgd
)
1698 blkptr_t
*bp
= zgd
->zgd_bp
;
1699 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)zgd
->zgd_db
;
1700 objset_t
*os
= db
->db_objset
;
1701 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
1702 dbuf_dirty_record_t
*dr
;
1703 dmu_sync_arg_t
*dsa
;
1704 zbookmark_phys_t zb
;
1708 ASSERT(pio
!= NULL
);
1711 SET_BOOKMARK(&zb
, ds
->ds_object
,
1712 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1716 dmu_write_policy(os
, dn
, db
->db_level
, WP_DMU_SYNC
, &zp
);
1720 * If we're frozen (running ziltest), we always need to generate a bp.
1722 if (txg
> spa_freeze_txg(os
->os_spa
))
1723 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1726 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1727 * and us. If we determine that this txg is not yet syncing,
1728 * but it begins to sync a moment later, that's OK because the
1729 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1731 mutex_enter(&db
->db_mtx
);
1733 if (txg
<= spa_last_synced_txg(os
->os_spa
)) {
1735 * This txg has already synced. There's nothing to do.
1737 mutex_exit(&db
->db_mtx
);
1738 return (SET_ERROR(EEXIST
));
1741 if (txg
<= spa_syncing_txg(os
->os_spa
)) {
1743 * This txg is currently syncing, so we can't mess with
1744 * the dirty record anymore; just write a new log block.
1746 mutex_exit(&db
->db_mtx
);
1747 return (dmu_sync_late_arrival(pio
, os
, done
, zgd
, &zp
, &zb
));
1750 dr
= db
->db_last_dirty
;
1751 while (dr
&& dr
->dr_txg
!= txg
)
1756 * There's no dr for this dbuf, so it must have been freed.
1757 * There's no need to log writes to freed blocks, so we're done.
1759 mutex_exit(&db
->db_mtx
);
1760 return (SET_ERROR(ENOENT
));
1763 ASSERT(dr
->dr_next
== NULL
|| dr
->dr_next
->dr_txg
< txg
);
1766 * Assume the on-disk data is X, the current syncing data (in
1767 * txg - 1) is Y, and the current in-memory data is Z (currently
1770 * We usually want to perform a nopwrite if X and Z are the
1771 * same. However, if Y is different (i.e. the BP is going to
1772 * change before this write takes effect), then a nopwrite will
1773 * be incorrect - we would override with X, which could have
1774 * been freed when Y was written.
1776 * (Note that this is not a concern when we are nop-writing from
1777 * syncing context, because X and Y must be identical, because
1778 * all previous txgs have been synced.)
1780 * Therefore, we disable nopwrite if the current BP could change
1781 * before this TXG. There are two ways it could change: by
1782 * being dirty (dr_next is non-NULL), or by being freed
1783 * (dnode_block_freed()). This behavior is verified by
1784 * zio_done(), which VERIFYs that the override BP is identical
1785 * to the on-disk BP.
1789 if (dr
->dr_next
!= NULL
|| dnode_block_freed(dn
, db
->db_blkid
))
1790 zp
.zp_nopwrite
= B_FALSE
;
1793 ASSERT(dr
->dr_txg
== txg
);
1794 if (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
||
1795 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
1797 * We have already issued a sync write for this buffer,
1798 * or this buffer has already been synced. It could not
1799 * have been dirtied since, or we would have cleared the state.
1801 mutex_exit(&db
->db_mtx
);
1802 return (SET_ERROR(EALREADY
));
1805 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
1806 dr
->dt
.dl
.dr_override_state
= DR_IN_DMU_SYNC
;
1807 mutex_exit(&db
->db_mtx
);
1809 dsa
= kmem_alloc(sizeof (dmu_sync_arg_t
), KM_SLEEP
);
1811 dsa
->dsa_done
= done
;
1815 zio_nowait(arc_write(pio
, os
->os_spa
, txg
,
1816 bp
, dr
->dt
.dl
.dr_data
, DBUF_IS_L2CACHEABLE(db
),
1817 &zp
, dmu_sync_ready
, NULL
, NULL
, dmu_sync_done
, dsa
,
1818 ZIO_PRIORITY_SYNC_WRITE
, ZIO_FLAG_CANFAIL
, &zb
));
1824 dmu_object_set_blocksize(objset_t
*os
, uint64_t object
, uint64_t size
, int ibs
,
1830 err
= dnode_hold(os
, object
, FTAG
, &dn
);
1833 err
= dnode_set_blksz(dn
, size
, ibs
, tx
);
1834 dnode_rele(dn
, FTAG
);
1839 dmu_object_set_checksum(objset_t
*os
, uint64_t object
, uint8_t checksum
,
1845 * Send streams include each object's checksum function. This
1846 * check ensures that the receiving system can understand the
1847 * checksum function transmitted.
1849 ASSERT3U(checksum
, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS
);
1851 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1852 ASSERT3U(checksum
, <, ZIO_CHECKSUM_FUNCTIONS
);
1853 dn
->dn_checksum
= checksum
;
1854 dnode_setdirty(dn
, tx
);
1855 dnode_rele(dn
, FTAG
);
1859 dmu_object_set_compress(objset_t
*os
, uint64_t object
, uint8_t compress
,
1865 * Send streams include each object's compression function. This
1866 * check ensures that the receiving system can understand the
1867 * compression function transmitted.
1869 ASSERT3U(compress
, <, ZIO_COMPRESS_LEGACY_FUNCTIONS
);
1871 VERIFY0(dnode_hold(os
, object
, FTAG
, &dn
));
1872 dn
->dn_compress
= compress
;
1873 dnode_setdirty(dn
, tx
);
1874 dnode_rele(dn
, FTAG
);
1877 int zfs_mdcomp_disable
= 0;
1880 * When the "redundant_metadata" property is set to "most", only indirect
1881 * blocks of this level and higher will have an additional ditto block.
1883 int zfs_redundant_metadata_most_ditto_level
= 2;
1886 dmu_write_policy(objset_t
*os
, dnode_t
*dn
, int level
, int wp
, zio_prop_t
*zp
)
1888 dmu_object_type_t type
= dn
? dn
->dn_type
: DMU_OT_OBJSET
;
1889 boolean_t ismd
= (level
> 0 || DMU_OT_IS_METADATA(type
) ||
1891 enum zio_checksum checksum
= os
->os_checksum
;
1892 enum zio_compress compress
= os
->os_compress
;
1893 enum zio_checksum dedup_checksum
= os
->os_dedup_checksum
;
1894 boolean_t dedup
= B_FALSE
;
1895 boolean_t nopwrite
= B_FALSE
;
1896 boolean_t dedup_verify
= os
->os_dedup_verify
;
1897 int copies
= os
->os_copies
;
1900 * We maintain different write policies for each of the following
1903 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1904 * 3. all other level 0 blocks
1907 if (zfs_mdcomp_disable
) {
1908 compress
= ZIO_COMPRESS_EMPTY
;
1911 * XXX -- we should design a compression algorithm
1912 * that specializes in arrays of bps.
1914 compress
= zio_compress_select(os
->os_spa
,
1915 ZIO_COMPRESS_ON
, ZIO_COMPRESS_ON
);
1919 * Metadata always gets checksummed. If the data
1920 * checksum is multi-bit correctable, and it's not a
1921 * ZBT-style checksum, then it's suitable for metadata
1922 * as well. Otherwise, the metadata checksum defaults
1925 if (!(zio_checksum_table
[checksum
].ci_flags
&
1926 ZCHECKSUM_FLAG_METADATA
) ||
1927 (zio_checksum_table
[checksum
].ci_flags
&
1928 ZCHECKSUM_FLAG_EMBEDDED
))
1929 checksum
= ZIO_CHECKSUM_FLETCHER_4
;
1931 if (os
->os_redundant_metadata
== ZFS_REDUNDANT_METADATA_ALL
||
1932 (os
->os_redundant_metadata
==
1933 ZFS_REDUNDANT_METADATA_MOST
&&
1934 (level
>= zfs_redundant_metadata_most_ditto_level
||
1935 DMU_OT_IS_METADATA(type
) || (wp
& WP_SPILL
))))
1937 } else if (wp
& WP_NOFILL
) {
1941 * If we're writing preallocated blocks, we aren't actually
1942 * writing them so don't set any policy properties. These
1943 * blocks are currently only used by an external subsystem
1944 * outside of zfs (i.e. dump) and not written by the zio
1947 compress
= ZIO_COMPRESS_OFF
;
1948 checksum
= ZIO_CHECKSUM_NOPARITY
;
1950 compress
= zio_compress_select(os
->os_spa
, dn
->dn_compress
,
1953 checksum
= (dedup_checksum
== ZIO_CHECKSUM_OFF
) ?
1954 zio_checksum_select(dn
->dn_checksum
, checksum
) :
1958 * Determine dedup setting. If we are in dmu_sync(),
1959 * we won't actually dedup now because that's all
1960 * done in syncing context; but we do want to use the
1961 * dedup checkum. If the checksum is not strong
1962 * enough to ensure unique signatures, force
1965 if (dedup_checksum
!= ZIO_CHECKSUM_OFF
) {
1966 dedup
= (wp
& WP_DMU_SYNC
) ? B_FALSE
: B_TRUE
;
1967 if (!(zio_checksum_table
[checksum
].ci_flags
&
1968 ZCHECKSUM_FLAG_DEDUP
))
1969 dedup_verify
= B_TRUE
;
1973 * Enable nopwrite if we have secure enough checksum
1974 * algorithm (see comment in zio_nop_write) and
1975 * compression is enabled. We don't enable nopwrite if
1976 * dedup is enabled as the two features are mutually
1979 nopwrite
= (!dedup
&& (zio_checksum_table
[checksum
].ci_flags
&
1980 ZCHECKSUM_FLAG_NOPWRITE
) &&
1981 compress
!= ZIO_COMPRESS_OFF
&& zfs_nopwrite_enabled
);
1984 zp
->zp_checksum
= checksum
;
1985 zp
->zp_compress
= compress
;
1986 ASSERT3U(zp
->zp_compress
, !=, ZIO_COMPRESS_INHERIT
);
1988 zp
->zp_type
= (wp
& WP_SPILL
) ? dn
->dn_bonustype
: type
;
1989 zp
->zp_level
= level
;
1990 zp
->zp_copies
= MIN(copies
, spa_max_replication(os
->os_spa
));
1991 zp
->zp_dedup
= dedup
;
1992 zp
->zp_dedup_verify
= dedup
&& dedup_verify
;
1993 zp
->zp_nopwrite
= nopwrite
;
1997 dmu_offset_next(objset_t
*os
, uint64_t object
, boolean_t hole
, uint64_t *off
)
2003 * Sync any current changes before
2004 * we go trundling through the block pointers.
2006 err
= dmu_object_wait_synced(os
, object
);
2011 err
= dnode_hold(os
, object
, FTAG
, &dn
);
2016 err
= dnode_next_offset(dn
, (hole
? DNODE_FIND_HOLE
: 0), off
, 1, 1, 0);
2017 dnode_rele(dn
, FTAG
);
2023 * Given the ZFS object, if it contains any dirty nodes
2024 * this function flushes all dirty blocks to disk. This
2025 * ensures the DMU object info is updated. A more efficient
2026 * future version might just find the TXG with the maximum
2027 * ID and wait for that to be synced.
2030 dmu_object_wait_synced(objset_t
*os
, uint64_t object
)
2035 error
= dnode_hold(os
, object
, FTAG
, &dn
);
2040 for (i
= 0; i
< TXG_SIZE
; i
++) {
2041 if (list_link_active(&dn
->dn_dirty_link
[i
])) {
2045 dnode_rele(dn
, FTAG
);
2046 if (i
!= TXG_SIZE
) {
2047 txg_wait_synced(dmu_objset_pool(os
), 0);
2054 dmu_object_info_from_dnode(dnode_t
*dn
, dmu_object_info_t
*doi
)
2058 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
2059 mutex_enter(&dn
->dn_mtx
);
2063 doi
->doi_data_block_size
= dn
->dn_datablksz
;
2064 doi
->doi_metadata_block_size
= dn
->dn_indblkshift
?
2065 1ULL << dn
->dn_indblkshift
: 0;
2066 doi
->doi_type
= dn
->dn_type
;
2067 doi
->doi_bonus_type
= dn
->dn_bonustype
;
2068 doi
->doi_bonus_size
= dn
->dn_bonuslen
;
2069 doi
->doi_indirection
= dn
->dn_nlevels
;
2070 doi
->doi_checksum
= dn
->dn_checksum
;
2071 doi
->doi_compress
= dn
->dn_compress
;
2072 doi
->doi_nblkptr
= dn
->dn_nblkptr
;
2073 doi
->doi_physical_blocks_512
= (DN_USED_BYTES(dnp
) + 256) >> 9;
2074 doi
->doi_max_offset
= (dn
->dn_maxblkid
+ 1) * dn
->dn_datablksz
;
2075 doi
->doi_fill_count
= 0;
2076 for (int i
= 0; i
< dnp
->dn_nblkptr
; i
++)
2077 doi
->doi_fill_count
+= BP_GET_FILL(&dnp
->dn_blkptr
[i
]);
2079 mutex_exit(&dn
->dn_mtx
);
2080 rw_exit(&dn
->dn_struct_rwlock
);
2084 * Get information on a DMU object.
2085 * If doi is NULL, just indicates whether the object exists.
2088 dmu_object_info(objset_t
*os
, uint64_t object
, dmu_object_info_t
*doi
)
2091 int err
= dnode_hold(os
, object
, FTAG
, &dn
);
2097 dmu_object_info_from_dnode(dn
, doi
);
2099 dnode_rele(dn
, FTAG
);
2104 * As above, but faster; can be used when you have a held dbuf in hand.
2107 dmu_object_info_from_db(dmu_buf_t
*db_fake
, dmu_object_info_t
*doi
)
2109 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2112 dmu_object_info_from_dnode(DB_DNODE(db
), doi
);
2117 * Faster still when you only care about the size.
2118 * This is specifically optimized for zfs_getattr().
2121 dmu_object_size_from_db(dmu_buf_t
*db_fake
, uint32_t *blksize
,
2122 u_longlong_t
*nblk512
)
2124 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2130 *blksize
= dn
->dn_datablksz
;
2131 /* add 1 for dnode space */
2132 *nblk512
= ((DN_USED_BYTES(dn
->dn_phys
) + SPA_MINBLOCKSIZE
/2) >>
2133 SPA_MINBLOCKSHIFT
) + 1;
2138 byteswap_uint64_array(void *vbuf
, size_t size
)
2140 uint64_t *buf
= vbuf
;
2141 size_t count
= size
>> 3;
2144 ASSERT((size
& 7) == 0);
2146 for (i
= 0; i
< count
; i
++)
2147 buf
[i
] = BSWAP_64(buf
[i
]);
2151 byteswap_uint32_array(void *vbuf
, size_t size
)
2153 uint32_t *buf
= vbuf
;
2154 size_t count
= size
>> 2;
2157 ASSERT((size
& 3) == 0);
2159 for (i
= 0; i
< count
; i
++)
2160 buf
[i
] = BSWAP_32(buf
[i
]);
2164 byteswap_uint16_array(void *vbuf
, size_t size
)
2166 uint16_t *buf
= vbuf
;
2167 size_t count
= size
>> 1;
2170 ASSERT((size
& 1) == 0);
2172 for (i
= 0; i
< count
; i
++)
2173 buf
[i
] = BSWAP_16(buf
[i
]);
2178 byteswap_uint8_array(void *vbuf
, size_t size
)
2200 arc_fini(); /* arc depends on l2arc, so arc must go first */