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 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
28 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
41 #include <sys/trace_dmu.h>
43 typedef void (*dmu_tx_hold_func_t
)(dmu_tx_t
*tx
, struct dnode
*dn
,
44 uint64_t arg1
, uint64_t arg2
);
46 dmu_tx_stats_t dmu_tx_stats
= {
47 { "dmu_tx_assigned", KSTAT_DATA_UINT64
},
48 { "dmu_tx_delay", KSTAT_DATA_UINT64
},
49 { "dmu_tx_error", KSTAT_DATA_UINT64
},
50 { "dmu_tx_suspended", KSTAT_DATA_UINT64
},
51 { "dmu_tx_group", KSTAT_DATA_UINT64
},
52 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64
},
53 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64
},
54 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64
},
55 { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64
},
56 { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64
},
57 { "dmu_tx_quota", KSTAT_DATA_UINT64
},
60 static kstat_t
*dmu_tx_ksp
;
63 dmu_tx_create_dd(dsl_dir_t
*dd
)
65 dmu_tx_t
*tx
= kmem_zalloc(sizeof (dmu_tx_t
), KM_SLEEP
);
68 tx
->tx_pool
= dd
->dd_pool
;
69 list_create(&tx
->tx_holds
, sizeof (dmu_tx_hold_t
),
70 offsetof(dmu_tx_hold_t
, txh_node
));
71 list_create(&tx
->tx_callbacks
, sizeof (dmu_tx_callback_t
),
72 offsetof(dmu_tx_callback_t
, dcb_node
));
73 tx
->tx_start
= gethrtime();
75 refcount_create(&tx
->tx_space_written
);
76 refcount_create(&tx
->tx_space_freed
);
82 dmu_tx_create(objset_t
*os
)
84 dmu_tx_t
*tx
= dmu_tx_create_dd(os
->os_dsl_dataset
->ds_dir
);
86 tx
->tx_lastsnap_txg
= dsl_dataset_prev_snap_txg(os
->os_dsl_dataset
);
91 dmu_tx_create_assigned(struct dsl_pool
*dp
, uint64_t txg
)
93 dmu_tx_t
*tx
= dmu_tx_create_dd(NULL
);
95 ASSERT3U(txg
, <=, dp
->dp_tx
.tx_open_txg
);
104 dmu_tx_is_syncing(dmu_tx_t
*tx
)
106 return (tx
->tx_anyobj
);
110 dmu_tx_private_ok(dmu_tx_t
*tx
)
112 return (tx
->tx_anyobj
);
115 static dmu_tx_hold_t
*
116 dmu_tx_hold_object_impl(dmu_tx_t
*tx
, objset_t
*os
, uint64_t object
,
117 enum dmu_tx_hold_type type
, uint64_t arg1
, uint64_t arg2
)
123 if (object
!= DMU_NEW_OBJECT
) {
124 err
= dnode_hold(os
, object
, tx
, &dn
);
130 if (err
== 0 && tx
->tx_txg
!= 0) {
131 mutex_enter(&dn
->dn_mtx
);
133 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
134 * problem, but there's no way for it to happen (for
137 ASSERT(dn
->dn_assigned_txg
== 0);
138 dn
->dn_assigned_txg
= tx
->tx_txg
;
139 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
140 mutex_exit(&dn
->dn_mtx
);
144 txh
= kmem_zalloc(sizeof (dmu_tx_hold_t
), KM_SLEEP
);
148 txh
->txh_type
= type
;
149 txh
->txh_arg1
= arg1
;
150 txh
->txh_arg2
= arg2
;
152 list_insert_tail(&tx
->tx_holds
, txh
);
158 dmu_tx_add_new_object(dmu_tx_t
*tx
, objset_t
*os
, uint64_t object
)
161 * If we're syncing, they can manipulate any object anyhow, and
162 * the hold on the dnode_t can cause problems.
164 if (!dmu_tx_is_syncing(tx
)) {
165 (void) dmu_tx_hold_object_impl(tx
, os
,
166 object
, THT_NEWOBJECT
, 0, 0);
171 dmu_tx_check_ioerr(zio_t
*zio
, dnode_t
*dn
, int level
, uint64_t blkid
)
176 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
177 db
= dbuf_hold_level(dn
, level
, blkid
, FTAG
);
178 rw_exit(&dn
->dn_struct_rwlock
);
180 return (SET_ERROR(EIO
));
181 err
= dbuf_read(db
, zio
, DB_RF_CANFAIL
| DB_RF_NOPREFETCH
);
187 dmu_tx_count_twig(dmu_tx_hold_t
*txh
, dnode_t
*dn
, dmu_buf_impl_t
*db
,
188 int level
, uint64_t blkid
, boolean_t freeable
, uint64_t *history
)
190 objset_t
*os
= dn
->dn_objset
;
191 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
192 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
193 dmu_buf_impl_t
*parent
= NULL
;
197 if (level
>= dn
->dn_nlevels
|| history
[level
] == blkid
)
200 history
[level
] = blkid
;
202 space
= (level
== 0) ? dn
->dn_datablksz
: (1ULL << dn
->dn_indblkshift
);
204 if (db
== NULL
|| db
== dn
->dn_dbuf
) {
208 ASSERT(DB_DNODE(db
) == dn
);
209 ASSERT(db
->db_level
== level
);
210 ASSERT(db
->db
.db_size
== space
);
211 ASSERT(db
->db_blkid
== blkid
);
213 parent
= db
->db_parent
;
216 freeable
= (bp
&& (freeable
||
217 dsl_dataset_block_freeable(ds
, bp
, bp
->blk_birth
)));
220 txh
->txh_space_tooverwrite
+= space
;
222 txh
->txh_space_towrite
+= space
;
224 txh
->txh_space_tounref
+= bp_get_dsize(os
->os_spa
, bp
);
226 dmu_tx_count_twig(txh
, dn
, parent
, level
+ 1,
227 blkid
>> epbs
, freeable
, history
);
232 dmu_tx_count_write(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
234 dnode_t
*dn
= txh
->txh_dnode
;
235 uint64_t start
, end
, i
;
236 int min_bs
, max_bs
, min_ibs
, max_ibs
, epbs
, bits
;
243 min_bs
= SPA_MINBLOCKSHIFT
;
244 max_bs
= highbit64(txh
->txh_tx
->tx_objset
->os_recordsize
) - 1;
245 min_ibs
= DN_MIN_INDBLKSHIFT
;
246 max_ibs
= DN_MAX_INDBLKSHIFT
;
249 uint64_t history
[DN_MAX_LEVELS
];
250 int nlvls
= dn
->dn_nlevels
;
254 * For i/o error checking, read the first and last level-0
255 * blocks (if they are not aligned), and all the level-1 blocks.
257 if (dn
->dn_maxblkid
== 0) {
258 delta
= dn
->dn_datablksz
;
259 start
= (off
< dn
->dn_datablksz
) ? 0 : 1;
260 end
= (off
+len
<= dn
->dn_datablksz
) ? 0 : 1;
261 if (start
== 0 && (off
> 0 || len
< dn
->dn_datablksz
)) {
262 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
268 zio_t
*zio
= zio_root(dn
->dn_objset
->os_spa
,
269 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
271 /* first level-0 block */
272 start
= off
>> dn
->dn_datablkshift
;
273 if (P2PHASE(off
, dn
->dn_datablksz
) ||
274 len
< dn
->dn_datablksz
) {
275 err
= dmu_tx_check_ioerr(zio
, dn
, 0, start
);
280 /* last level-0 block */
281 end
= (off
+len
-1) >> dn
->dn_datablkshift
;
282 if (end
!= start
&& end
<= dn
->dn_maxblkid
&&
283 P2PHASE(off
+len
, dn
->dn_datablksz
)) {
284 err
= dmu_tx_check_ioerr(zio
, dn
, 0, end
);
291 int shft
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
292 for (i
= (start
>>shft
)+1; i
< end
>>shft
; i
++) {
293 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
302 delta
= P2NPHASE(off
, dn
->dn_datablksz
);
305 min_ibs
= max_ibs
= dn
->dn_indblkshift
;
306 if (dn
->dn_maxblkid
> 0) {
308 * The blocksize can't change,
309 * so we can make a more precise estimate.
311 ASSERT(dn
->dn_datablkshift
!= 0);
312 min_bs
= max_bs
= dn
->dn_datablkshift
;
315 * The blocksize can increase up to the recordsize,
316 * or if it is already more than the recordsize,
317 * up to the next power of 2.
319 min_bs
= highbit64(dn
->dn_datablksz
- 1);
320 max_bs
= MAX(max_bs
, highbit64(dn
->dn_datablksz
- 1));
324 * If this write is not off the end of the file
325 * we need to account for overwrites/unref.
327 if (start
<= dn
->dn_maxblkid
) {
328 for (l
= 0; l
< DN_MAX_LEVELS
; l
++)
331 while (start
<= dn
->dn_maxblkid
) {
334 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
335 err
= dbuf_hold_impl(dn
, 0, start
, FALSE
, FTAG
, &db
);
336 rw_exit(&dn
->dn_struct_rwlock
);
339 txh
->txh_tx
->tx_err
= err
;
343 dmu_tx_count_twig(txh
, dn
, db
, 0, start
, B_FALSE
,
348 * Account for new indirects appearing
349 * before this IO gets assigned into a txg.
352 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
353 for (bits
-= epbs
* (nlvls
- 1);
354 bits
>= 0; bits
-= epbs
)
355 txh
->txh_fudge
+= 1ULL << max_ibs
;
361 delta
= dn
->dn_datablksz
;
366 * 'end' is the last thing we will access, not one past.
367 * This way we won't overflow when accessing the last byte.
369 start
= P2ALIGN(off
, 1ULL << max_bs
);
370 end
= P2ROUNDUP(off
+ len
, 1ULL << max_bs
) - 1;
371 txh
->txh_space_towrite
+= end
- start
+ 1;
376 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
379 * The object contains at most 2^(64 - min_bs) blocks,
380 * and each indirect level maps 2^epbs.
382 for (bits
= 64 - min_bs
; bits
>= 0; bits
-= epbs
) {
385 ASSERT3U(end
, >=, start
);
386 txh
->txh_space_towrite
+= (end
- start
+ 1) << max_ibs
;
389 * We also need a new blkid=0 indirect block
390 * to reference any existing file data.
392 txh
->txh_space_towrite
+= 1ULL << max_ibs
;
397 if (txh
->txh_space_towrite
+ txh
->txh_space_tooverwrite
>
399 err
= SET_ERROR(EFBIG
);
402 txh
->txh_tx
->tx_err
= err
;
406 dmu_tx_count_dnode(dmu_tx_hold_t
*txh
)
408 dnode_t
*dn
= txh
->txh_dnode
;
409 dnode_t
*mdn
= DMU_META_DNODE(txh
->txh_tx
->tx_objset
);
410 uint64_t space
= mdn
->dn_datablksz
+
411 ((mdn
->dn_nlevels
-1) << mdn
->dn_indblkshift
);
413 if (dn
&& dn
->dn_dbuf
->db_blkptr
&&
414 dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
415 dn
->dn_dbuf
->db_blkptr
, dn
->dn_dbuf
->db_blkptr
->blk_birth
)) {
416 txh
->txh_space_tooverwrite
+= space
;
417 txh
->txh_space_tounref
+= space
;
419 txh
->txh_space_towrite
+= space
;
420 if (dn
&& dn
->dn_dbuf
->db_blkptr
)
421 txh
->txh_space_tounref
+= space
;
426 dmu_tx_hold_write(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, int len
)
430 ASSERT(tx
->tx_txg
== 0);
431 ASSERT(len
<= DMU_MAX_ACCESS
);
432 ASSERT(len
== 0 || UINT64_MAX
- off
>= len
- 1);
434 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
435 object
, THT_WRITE
, off
, len
);
439 dmu_tx_count_write(txh
, off
, len
);
440 dmu_tx_count_dnode(txh
);
444 dmu_tx_count_free(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
446 uint64_t blkid
, nblks
, lastblk
;
447 uint64_t space
= 0, unref
= 0, skipped
= 0;
448 dnode_t
*dn
= txh
->txh_dnode
;
449 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
450 spa_t
*spa
= txh
->txh_tx
->tx_pool
->dp_spa
;
452 uint64_t l0span
= 0, nl1blks
= 0;
454 if (dn
->dn_nlevels
== 0)
458 * The struct_rwlock protects us against dn_nlevels
459 * changing, in case (against all odds) we manage to dirty &
460 * sync out the changes after we check for being dirty.
461 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
463 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
464 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
465 if (dn
->dn_maxblkid
== 0) {
466 if (off
== 0 && len
>= dn
->dn_datablksz
) {
470 rw_exit(&dn
->dn_struct_rwlock
);
474 blkid
= off
>> dn
->dn_datablkshift
;
475 nblks
= (len
+ dn
->dn_datablksz
- 1) >> dn
->dn_datablkshift
;
477 if (blkid
> dn
->dn_maxblkid
) {
478 rw_exit(&dn
->dn_struct_rwlock
);
481 if (blkid
+ nblks
> dn
->dn_maxblkid
)
482 nblks
= dn
->dn_maxblkid
- blkid
+ 1;
485 l0span
= nblks
; /* save for later use to calc level > 1 overhead */
486 if (dn
->dn_nlevels
== 1) {
488 for (i
= 0; i
< nblks
; i
++) {
489 blkptr_t
*bp
= dn
->dn_phys
->dn_blkptr
;
490 ASSERT3U(blkid
+ i
, <, dn
->dn_nblkptr
);
492 if (dsl_dataset_block_freeable(ds
, bp
, bp
->blk_birth
)) {
493 dprintf_bp(bp
, "can free old%s", "");
494 space
+= bp_get_dsize(spa
, bp
);
496 unref
+= BP_GET_ASIZE(bp
);
502 lastblk
= blkid
+ nblks
- 1;
504 dmu_buf_impl_t
*dbuf
;
505 uint64_t ibyte
, new_blkid
;
507 int err
, i
, blkoff
, tochk
;
510 ibyte
= blkid
<< dn
->dn_datablkshift
;
511 err
= dnode_next_offset(dn
,
512 DNODE_FIND_HAVELOCK
, &ibyte
, 2, 1, 0);
513 new_blkid
= ibyte
>> dn
->dn_datablkshift
;
515 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
519 txh
->txh_tx
->tx_err
= err
;
522 if (new_blkid
> lastblk
) {
523 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
527 if (new_blkid
> blkid
) {
528 ASSERT((new_blkid
>> epbs
) > (blkid
>> epbs
));
529 skipped
+= (new_blkid
>> epbs
) - (blkid
>> epbs
) - 1;
530 nblks
-= new_blkid
- blkid
;
533 blkoff
= P2PHASE(blkid
, epb
);
534 tochk
= MIN(epb
- blkoff
, nblks
);
536 err
= dbuf_hold_impl(dn
, 1, blkid
>> epbs
, FALSE
, FTAG
, &dbuf
);
538 txh
->txh_tx
->tx_err
= err
;
542 txh
->txh_memory_tohold
+= dbuf
->db
.db_size
;
545 * We don't check memory_tohold against DMU_MAX_ACCESS because
546 * memory_tohold is an over-estimation (especially the >L1
547 * indirect blocks), so it could fail. Callers should have
548 * already verified that they will not be holding too much
552 err
= dbuf_read(dbuf
, NULL
, DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
);
554 txh
->txh_tx
->tx_err
= err
;
555 dbuf_rele(dbuf
, FTAG
);
559 bp
= dbuf
->db
.db_data
;
562 for (i
= 0; i
< tochk
; i
++) {
563 if (dsl_dataset_block_freeable(ds
, &bp
[i
],
565 dprintf_bp(&bp
[i
], "can free old%s", "");
566 space
+= bp_get_dsize(spa
, &bp
[i
]);
568 unref
+= BP_GET_ASIZE(bp
);
570 dbuf_rele(dbuf
, FTAG
);
576 rw_exit(&dn
->dn_struct_rwlock
);
579 * Add in memory requirements of higher-level indirects.
580 * This assumes a worst-possible scenario for dn_nlevels and a
581 * worst-possible distribution of l1-blocks over the region to free.
584 uint64_t blkcnt
= 1 + ((l0span
>> epbs
) >> epbs
);
587 * Here we don't use DN_MAX_LEVEL, but calculate it with the
588 * given datablkshift and indblkshift. This makes the
589 * difference between 19 and 8 on large files.
591 int maxlevel
= 2 + (DN_MAX_OFFSET_SHIFT
- dn
->dn_datablkshift
) /
592 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
594 while (level
++ < maxlevel
) {
595 txh
->txh_memory_tohold
+= MAX(MIN(blkcnt
, nl1blks
), 1)
596 << dn
->dn_indblkshift
;
597 blkcnt
= 1 + (blkcnt
>> epbs
);
601 /* account for new level 1 indirect blocks that might show up */
603 txh
->txh_fudge
+= skipped
<< dn
->dn_indblkshift
;
604 skipped
= MIN(skipped
, DMU_MAX_DELETEBLKCNT
>> epbs
);
605 txh
->txh_memory_tohold
+= skipped
<< dn
->dn_indblkshift
;
607 txh
->txh_space_tofree
+= space
;
608 txh
->txh_space_tounref
+= unref
;
612 dmu_tx_hold_free(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, uint64_t len
)
619 ASSERT(tx
->tx_txg
== 0);
621 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
622 object
, THT_FREE
, off
, len
);
626 dmu_tx_count_dnode(txh
);
628 if (off
>= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
)
630 if (len
== DMU_OBJECT_END
)
631 len
= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
- off
;
633 dmu_tx_count_dnode(txh
);
636 * For i/o error checking, we read the first and last level-0
637 * blocks if they are not aligned, and all the level-1 blocks.
639 * Note: dbuf_free_range() assumes that we have not instantiated
640 * any level-0 dbufs that will be completely freed. Therefore we must
641 * exercise care to not read or count the first and last blocks
642 * if they are blocksize-aligned.
644 if (dn
->dn_datablkshift
== 0) {
645 if (off
!= 0 || len
< dn
->dn_datablksz
)
646 dmu_tx_count_write(txh
, 0, dn
->dn_datablksz
);
648 /* first block will be modified if it is not aligned */
649 if (!IS_P2ALIGNED(off
, 1 << dn
->dn_datablkshift
))
650 dmu_tx_count_write(txh
, off
, 1);
651 /* last block will be modified if it is not aligned */
652 if (!IS_P2ALIGNED(off
+ len
, 1 << dn
->dn_datablkshift
))
653 dmu_tx_count_write(txh
, off
+len
, 1);
657 * Check level-1 blocks.
659 if (dn
->dn_nlevels
> 1) {
660 int shift
= dn
->dn_datablkshift
+ dn
->dn_indblkshift
-
662 uint64_t start
= off
>> shift
;
663 uint64_t end
= (off
+ len
) >> shift
;
666 ASSERT(dn
->dn_indblkshift
!= 0);
669 * dnode_reallocate() can result in an object with indirect
670 * blocks having an odd data block size. In this case,
671 * just check the single block.
673 if (dn
->dn_datablkshift
== 0)
676 zio
= zio_root(tx
->tx_pool
->dp_spa
,
677 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
678 for (i
= start
; i
<= end
; i
++) {
679 uint64_t ibyte
= i
<< shift
;
680 err
= dnode_next_offset(dn
, 0, &ibyte
, 2, 1, 0);
682 if (err
== ESRCH
|| i
> end
)
689 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
702 dmu_tx_count_free(txh
, off
, len
);
706 dmu_tx_hold_zap(dmu_tx_t
*tx
, uint64_t object
, int add
, const char *name
)
710 dsl_dataset_phys_t
*ds_phys
;
714 ASSERT(tx
->tx_txg
== 0);
716 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
717 object
, THT_ZAP
, add
, (uintptr_t)name
);
722 dmu_tx_count_dnode(txh
);
726 * We will be able to fit a new object's entries into one leaf
727 * block. So there will be at most 2 blocks total,
728 * including the header block.
730 dmu_tx_count_write(txh
, 0, 2 << fzap_default_block_shift
);
734 ASSERT3U(DMU_OT_BYTESWAP(dn
->dn_type
), ==, DMU_BSWAP_ZAP
);
736 if (dn
->dn_maxblkid
== 0 && !add
) {
740 * If there is only one block (i.e. this is a micro-zap)
741 * and we are not adding anything, the accounting is simple.
743 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
750 * Use max block size here, since we don't know how much
751 * the size will change between now and the dbuf dirty call.
753 bp
= &dn
->dn_phys
->dn_blkptr
[0];
754 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
756 txh
->txh_space_tooverwrite
+= MZAP_MAX_BLKSZ
;
758 txh
->txh_space_towrite
+= MZAP_MAX_BLKSZ
;
760 txh
->txh_space_tounref
+= MZAP_MAX_BLKSZ
;
764 if (dn
->dn_maxblkid
> 0 && name
) {
766 * access the name in this fat-zap so that we'll check
767 * for i/o errors to the leaf blocks, etc.
769 err
= zap_lookup(dn
->dn_objset
, dn
->dn_object
, name
,
777 err
= zap_count_write(dn
->dn_objset
, dn
->dn_object
, name
, add
,
778 &txh
->txh_space_towrite
, &txh
->txh_space_tooverwrite
);
781 * If the modified blocks are scattered to the four winds,
782 * we'll have to modify an indirect twig for each.
784 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
785 ds_phys
= dsl_dataset_phys(dn
->dn_objset
->os_dsl_dataset
);
786 for (nblocks
= dn
->dn_maxblkid
>> epbs
; nblocks
!= 0; nblocks
>>= epbs
)
787 if (ds_phys
->ds_prev_snap_obj
)
788 txh
->txh_space_towrite
+= 3 << dn
->dn_indblkshift
;
790 txh
->txh_space_tooverwrite
+= 3 << dn
->dn_indblkshift
;
794 dmu_tx_hold_bonus(dmu_tx_t
*tx
, uint64_t object
)
798 ASSERT(tx
->tx_txg
== 0);
800 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
801 object
, THT_BONUS
, 0, 0);
803 dmu_tx_count_dnode(txh
);
807 dmu_tx_hold_space(dmu_tx_t
*tx
, uint64_t space
)
811 ASSERT(tx
->tx_txg
== 0);
813 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
814 DMU_NEW_OBJECT
, THT_SPACE
, space
, 0);
816 txh
->txh_space_towrite
+= space
;
820 dmu_tx_holds(dmu_tx_t
*tx
, uint64_t object
)
826 * By asserting that the tx is assigned, we're counting the
827 * number of dn_tx_holds, which is the same as the number of
828 * dn_holds. Otherwise, we'd be counting dn_holds, but
829 * dn_tx_holds could be 0.
831 ASSERT(tx
->tx_txg
!= 0);
833 /* if (tx->tx_anyobj == TRUE) */
836 for (txh
= list_head(&tx
->tx_holds
); txh
;
837 txh
= list_next(&tx
->tx_holds
, txh
)) {
838 if (txh
->txh_dnode
&& txh
->txh_dnode
->dn_object
== object
)
847 dmu_tx_dirty_buf(dmu_tx_t
*tx
, dmu_buf_impl_t
*db
)
850 int match_object
= FALSE
, match_offset
= FALSE
;
856 ASSERT(tx
->tx_txg
!= 0);
857 ASSERT(tx
->tx_objset
== NULL
|| dn
->dn_objset
== tx
->tx_objset
);
858 ASSERT3U(dn
->dn_object
, ==, db
->db
.db_object
);
865 /* XXX No checking on the meta dnode for now */
866 if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
871 for (txh
= list_head(&tx
->tx_holds
); txh
;
872 txh
= list_next(&tx
->tx_holds
, txh
)) {
873 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
874 if (txh
->txh_dnode
== dn
&& txh
->txh_type
!= THT_NEWOBJECT
)
876 if (txh
->txh_dnode
== NULL
|| txh
->txh_dnode
== dn
) {
877 int datablkshift
= dn
->dn_datablkshift
?
878 dn
->dn_datablkshift
: SPA_MAXBLOCKSHIFT
;
879 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
880 int shift
= datablkshift
+ epbs
* db
->db_level
;
881 uint64_t beginblk
= shift
>= 64 ? 0 :
882 (txh
->txh_arg1
>> shift
);
883 uint64_t endblk
= shift
>= 64 ? 0 :
884 ((txh
->txh_arg1
+ txh
->txh_arg2
- 1) >> shift
);
885 uint64_t blkid
= db
->db_blkid
;
887 /* XXX txh_arg2 better not be zero... */
889 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
890 txh
->txh_type
, beginblk
, endblk
);
892 switch (txh
->txh_type
) {
894 if (blkid
>= beginblk
&& blkid
<= endblk
)
897 * We will let this hold work for the bonus
898 * or spill buffer so that we don't need to
899 * hold it when creating a new object.
901 if (blkid
== DMU_BONUS_BLKID
||
902 blkid
== DMU_SPILL_BLKID
)
905 * They might have to increase nlevels,
906 * thus dirtying the new TLIBs. Or the
907 * might have to change the block size,
908 * thus dirying the new lvl=0 blk=0.
915 * We will dirty all the level 1 blocks in
916 * the free range and perhaps the first and
917 * last level 0 block.
919 if (blkid
>= beginblk
&& (blkid
<= endblk
||
920 txh
->txh_arg2
== DMU_OBJECT_END
))
924 if (blkid
== DMU_SPILL_BLKID
)
928 if (blkid
== DMU_BONUS_BLKID
)
938 cmn_err(CE_PANIC
, "bad txh_type %d",
942 if (match_object
&& match_offset
) {
948 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
949 (u_longlong_t
)db
->db
.db_object
, db
->db_level
,
950 (u_longlong_t
)db
->db_blkid
);
955 * If we can't do 10 iops, something is wrong. Let us go ahead
956 * and hit zfs_dirty_data_max.
958 hrtime_t zfs_delay_max_ns
= 100 * MICROSEC
; /* 100 milliseconds */
959 int zfs_delay_resolution_ns
= 100 * 1000; /* 100 microseconds */
962 * We delay transactions when we've determined that the backend storage
963 * isn't able to accommodate the rate of incoming writes.
965 * If there is already a transaction waiting, we delay relative to when
966 * that transaction finishes waiting. This way the calculated min_time
967 * is independent of the number of threads concurrently executing
970 * If we are the only waiter, wait relative to when the transaction
971 * started, rather than the current time. This credits the transaction for
972 * "time already served", e.g. reading indirect blocks.
974 * The minimum time for a transaction to take is calculated as:
975 * min_time = scale * (dirty - min) / (max - dirty)
976 * min_time is then capped at zfs_delay_max_ns.
978 * The delay has two degrees of freedom that can be adjusted via tunables.
979 * The percentage of dirty data at which we start to delay is defined by
980 * zfs_delay_min_dirty_percent. This should typically be at or above
981 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
982 * delay after writing at full speed has failed to keep up with the incoming
983 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
984 * speaking, this variable determines the amount of delay at the midpoint of
988 * 10ms +-------------------------------------------------------------*+
1004 * 2ms + (midpoint) * +
1007 * | zfs_delay_scale ----------> ******** |
1008 * 0 +-------------------------------------*********----------------+
1009 * 0% <- zfs_dirty_data_max -> 100%
1011 * Note that since the delay is added to the outstanding time remaining on the
1012 * most recent transaction, the delay is effectively the inverse of IOPS.
1013 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1014 * was chosen such that small changes in the amount of accumulated dirty data
1015 * in the first 3/4 of the curve yield relatively small differences in the
1018 * The effects can be easier to understand when the amount of delay is
1019 * represented on a log scale:
1022 * 100ms +-------------------------------------------------------------++
1031 * + zfs_delay_scale ----------> ***** +
1042 * +--------------------------------------------------------------+
1043 * 0% <- zfs_dirty_data_max -> 100%
1045 * Note here that only as the amount of dirty data approaches its limit does
1046 * the delay start to increase rapidly. The goal of a properly tuned system
1047 * should be to keep the amount of dirty data out of that range by first
1048 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1049 * optimal throughput on the backend storage, and then by changing the value
1050 * of zfs_delay_scale to increase the steepness of the curve.
1053 dmu_tx_delay(dmu_tx_t
*tx
, uint64_t dirty
)
1055 dsl_pool_t
*dp
= tx
->tx_pool
;
1056 uint64_t delay_min_bytes
=
1057 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
1058 hrtime_t wakeup
, min_tx_time
, now
;
1060 if (dirty
<= delay_min_bytes
)
1064 * The caller has already waited until we are under the max.
1065 * We make them pass us the amount of dirty data so we don't
1066 * have to handle the case of it being >= the max, which could
1067 * cause a divide-by-zero if it's == the max.
1069 ASSERT3U(dirty
, <, zfs_dirty_data_max
);
1072 min_tx_time
= zfs_delay_scale
*
1073 (dirty
- delay_min_bytes
) / (zfs_dirty_data_max
- dirty
);
1074 min_tx_time
= MIN(min_tx_time
, zfs_delay_max_ns
);
1075 if (now
> tx
->tx_start
+ min_tx_time
)
1078 DTRACE_PROBE3(delay__mintime
, dmu_tx_t
*, tx
, uint64_t, dirty
,
1079 uint64_t, min_tx_time
);
1081 mutex_enter(&dp
->dp_lock
);
1082 wakeup
= MAX(tx
->tx_start
+ min_tx_time
,
1083 dp
->dp_last_wakeup
+ min_tx_time
);
1084 dp
->dp_last_wakeup
= wakeup
;
1085 mutex_exit(&dp
->dp_lock
);
1087 zfs_sleep_until(wakeup
);
1091 dmu_tx_try_assign(dmu_tx_t
*tx
, txg_how_t txg_how
)
1094 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
1095 uint64_t memory
, asize
, fsize
, usize
;
1096 uint64_t towrite
, tofree
, tooverwrite
, tounref
, tohold
, fudge
;
1098 ASSERT0(tx
->tx_txg
);
1101 DMU_TX_STAT_BUMP(dmu_tx_error
);
1102 return (tx
->tx_err
);
1105 if (spa_suspended(spa
)) {
1106 DMU_TX_STAT_BUMP(dmu_tx_suspended
);
1109 * If the user has indicated a blocking failure mode
1110 * then return ERESTART which will block in dmu_tx_wait().
1111 * Otherwise, return EIO so that an error can get
1112 * propagated back to the VOP calls.
1114 * Note that we always honor the txg_how flag regardless
1115 * of the failuremode setting.
1117 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_CONTINUE
&&
1118 txg_how
!= TXG_WAIT
)
1119 return (SET_ERROR(EIO
));
1121 return (SET_ERROR(ERESTART
));
1124 if (!tx
->tx_waited
&&
1125 dsl_pool_need_dirty_delay(tx
->tx_pool
)) {
1126 tx
->tx_wait_dirty
= B_TRUE
;
1127 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay
);
1131 tx
->tx_txg
= txg_hold_open(tx
->tx_pool
, &tx
->tx_txgh
);
1132 tx
->tx_needassign_txh
= NULL
;
1135 * NB: No error returns are allowed after txg_hold_open, but
1136 * before processing the dnode holds, due to the
1137 * dmu_tx_unassign() logic.
1140 towrite
= tofree
= tooverwrite
= tounref
= tohold
= fudge
= 0;
1141 for (txh
= list_head(&tx
->tx_holds
); txh
;
1142 txh
= list_next(&tx
->tx_holds
, txh
)) {
1143 dnode_t
*dn
= txh
->txh_dnode
;
1145 mutex_enter(&dn
->dn_mtx
);
1146 if (dn
->dn_assigned_txg
== tx
->tx_txg
- 1) {
1147 mutex_exit(&dn
->dn_mtx
);
1148 tx
->tx_needassign_txh
= txh
;
1149 DMU_TX_STAT_BUMP(dmu_tx_group
);
1150 return (SET_ERROR(ERESTART
));
1152 if (dn
->dn_assigned_txg
== 0)
1153 dn
->dn_assigned_txg
= tx
->tx_txg
;
1154 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1155 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
1156 mutex_exit(&dn
->dn_mtx
);
1158 towrite
+= txh
->txh_space_towrite
;
1159 tofree
+= txh
->txh_space_tofree
;
1160 tooverwrite
+= txh
->txh_space_tooverwrite
;
1161 tounref
+= txh
->txh_space_tounref
;
1162 tohold
+= txh
->txh_memory_tohold
;
1163 fudge
+= txh
->txh_fudge
;
1167 * If a snapshot has been taken since we made our estimates,
1168 * assume that we won't be able to free or overwrite anything.
1170 if (tx
->tx_objset
&&
1171 dsl_dataset_prev_snap_txg(tx
->tx_objset
->os_dsl_dataset
) >
1172 tx
->tx_lastsnap_txg
) {
1173 towrite
+= tooverwrite
;
1174 tooverwrite
= tofree
= 0;
1177 /* needed allocation: worst-case estimate of write space */
1178 asize
= spa_get_asize(tx
->tx_pool
->dp_spa
, towrite
+ tooverwrite
);
1179 /* freed space estimate: worst-case overwrite + free estimate */
1180 fsize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tooverwrite
) + tofree
;
1181 /* convert unrefd space to worst-case estimate */
1182 usize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tounref
);
1183 /* calculate memory footprint estimate */
1184 memory
= towrite
+ tooverwrite
+ tohold
;
1188 * Add in 'tohold' to account for our dirty holds on this memory
1189 * XXX - the "fudge" factor is to account for skipped blocks that
1190 * we missed because dnode_next_offset() misses in-core-only blocks.
1192 tx
->tx_space_towrite
= asize
+
1193 spa_get_asize(tx
->tx_pool
->dp_spa
, tohold
+ fudge
);
1194 tx
->tx_space_tofree
= tofree
;
1195 tx
->tx_space_tooverwrite
= tooverwrite
;
1196 tx
->tx_space_tounref
= tounref
;
1199 if (tx
->tx_dir
&& asize
!= 0) {
1200 int err
= dsl_dir_tempreserve_space(tx
->tx_dir
, memory
,
1201 asize
, fsize
, usize
, &tx
->tx_tempreserve_cookie
, tx
);
1206 DMU_TX_STAT_BUMP(dmu_tx_assigned
);
1212 dmu_tx_unassign(dmu_tx_t
*tx
)
1216 if (tx
->tx_txg
== 0)
1219 txg_rele_to_quiesce(&tx
->tx_txgh
);
1222 * Walk the transaction's hold list, removing the hold on the
1223 * associated dnode, and notifying waiters if the refcount drops to 0.
1225 for (txh
= list_head(&tx
->tx_holds
); txh
!= tx
->tx_needassign_txh
;
1226 txh
= list_next(&tx
->tx_holds
, txh
)) {
1227 dnode_t
*dn
= txh
->txh_dnode
;
1231 mutex_enter(&dn
->dn_mtx
);
1232 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1234 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1235 dn
->dn_assigned_txg
= 0;
1236 cv_broadcast(&dn
->dn_notxholds
);
1238 mutex_exit(&dn
->dn_mtx
);
1241 txg_rele_to_sync(&tx
->tx_txgh
);
1243 tx
->tx_lasttried_txg
= tx
->tx_txg
;
1248 * Assign tx to a transaction group. txg_how can be one of:
1250 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1251 * a new one. This should be used when you're not holding locks.
1252 * It will only fail if we're truly out of space (or over quota).
1254 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1255 * blocking, returns immediately with ERESTART. This should be used
1256 * whenever you're holding locks. On an ERESTART error, the caller
1257 * should drop locks, do a dmu_tx_wait(tx), and try again.
1259 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1260 * has already been called on behalf of this operation (though
1261 * most likely on a different tx).
1264 dmu_tx_assign(dmu_tx_t
*tx
, txg_how_t txg_how
)
1268 ASSERT(tx
->tx_txg
== 0);
1269 ASSERT(txg_how
== TXG_WAIT
|| txg_how
== TXG_NOWAIT
||
1270 txg_how
== TXG_WAITED
);
1271 ASSERT(!dsl_pool_sync_context(tx
->tx_pool
));
1273 if (txg_how
== TXG_WAITED
)
1274 tx
->tx_waited
= B_TRUE
;
1276 /* If we might wait, we must not hold the config lock. */
1277 ASSERT(txg_how
!= TXG_WAIT
|| !dsl_pool_config_held(tx
->tx_pool
));
1279 while ((err
= dmu_tx_try_assign(tx
, txg_how
)) != 0) {
1280 dmu_tx_unassign(tx
);
1282 if (err
!= ERESTART
|| txg_how
!= TXG_WAIT
)
1288 txg_rele_to_quiesce(&tx
->tx_txgh
);
1294 dmu_tx_wait(dmu_tx_t
*tx
)
1296 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
1297 dsl_pool_t
*dp
= tx
->tx_pool
;
1300 ASSERT(tx
->tx_txg
== 0);
1301 ASSERT(!dsl_pool_config_held(tx
->tx_pool
));
1303 before
= gethrtime();
1305 if (tx
->tx_wait_dirty
) {
1309 * dmu_tx_try_assign() has determined that we need to wait
1310 * because we've consumed much or all of the dirty buffer
1313 mutex_enter(&dp
->dp_lock
);
1314 if (dp
->dp_dirty_total
>= zfs_dirty_data_max
)
1315 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max
);
1316 while (dp
->dp_dirty_total
>= zfs_dirty_data_max
)
1317 cv_wait(&dp
->dp_spaceavail_cv
, &dp
->dp_lock
);
1318 dirty
= dp
->dp_dirty_total
;
1319 mutex_exit(&dp
->dp_lock
);
1321 dmu_tx_delay(tx
, dirty
);
1323 tx
->tx_wait_dirty
= B_FALSE
;
1326 * Note: setting tx_waited only has effect if the caller
1327 * used TX_WAIT. Otherwise they are going to destroy
1328 * this tx and try again. The common case, zfs_write(),
1331 tx
->tx_waited
= B_TRUE
;
1332 } else if (spa_suspended(spa
) || tx
->tx_lasttried_txg
== 0) {
1334 * If the pool is suspended we need to wait until it
1335 * is resumed. Note that it's possible that the pool
1336 * has become active after this thread has tried to
1337 * obtain a tx. If that's the case then tx_lasttried_txg
1338 * would not have been set.
1340 txg_wait_synced(dp
, spa_last_synced_txg(spa
) + 1);
1341 } else if (tx
->tx_needassign_txh
) {
1342 dnode_t
*dn
= tx
->tx_needassign_txh
->txh_dnode
;
1344 mutex_enter(&dn
->dn_mtx
);
1345 while (dn
->dn_assigned_txg
== tx
->tx_lasttried_txg
- 1)
1346 cv_wait(&dn
->dn_notxholds
, &dn
->dn_mtx
);
1347 mutex_exit(&dn
->dn_mtx
);
1348 tx
->tx_needassign_txh
= NULL
;
1351 * A dnode is assigned to the quiescing txg. Wait for its
1352 * transaction to complete.
1354 txg_wait_open(tx
->tx_pool
, tx
->tx_lasttried_txg
+ 1);
1357 spa_tx_assign_add_nsecs(spa
, gethrtime() - before
);
1361 dmu_tx_willuse_space(dmu_tx_t
*tx
, int64_t delta
)
1364 if (tx
->tx_dir
== NULL
|| delta
== 0)
1368 ASSERT3U(refcount_count(&tx
->tx_space_written
) + delta
, <=,
1369 tx
->tx_space_towrite
);
1370 (void) refcount_add_many(&tx
->tx_space_written
, delta
, NULL
);
1372 (void) refcount_add_many(&tx
->tx_space_freed
, -delta
, NULL
);
1378 dmu_tx_commit(dmu_tx_t
*tx
)
1382 ASSERT(tx
->tx_txg
!= 0);
1385 * Go through the transaction's hold list and remove holds on
1386 * associated dnodes, notifying waiters if no holds remain.
1388 while ((txh
= list_head(&tx
->tx_holds
))) {
1389 dnode_t
*dn
= txh
->txh_dnode
;
1391 list_remove(&tx
->tx_holds
, txh
);
1392 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1395 mutex_enter(&dn
->dn_mtx
);
1396 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1398 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1399 dn
->dn_assigned_txg
= 0;
1400 cv_broadcast(&dn
->dn_notxholds
);
1402 mutex_exit(&dn
->dn_mtx
);
1406 if (tx
->tx_tempreserve_cookie
)
1407 dsl_dir_tempreserve_clear(tx
->tx_tempreserve_cookie
, tx
);
1409 if (!list_is_empty(&tx
->tx_callbacks
))
1410 txg_register_callbacks(&tx
->tx_txgh
, &tx
->tx_callbacks
);
1412 if (tx
->tx_anyobj
== FALSE
)
1413 txg_rele_to_sync(&tx
->tx_txgh
);
1415 list_destroy(&tx
->tx_callbacks
);
1416 list_destroy(&tx
->tx_holds
);
1418 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1419 tx
->tx_space_towrite
, refcount_count(&tx
->tx_space_written
),
1420 tx
->tx_space_tofree
, refcount_count(&tx
->tx_space_freed
));
1421 refcount_destroy_many(&tx
->tx_space_written
,
1422 refcount_count(&tx
->tx_space_written
));
1423 refcount_destroy_many(&tx
->tx_space_freed
,
1424 refcount_count(&tx
->tx_space_freed
));
1426 kmem_free(tx
, sizeof (dmu_tx_t
));
1430 dmu_tx_abort(dmu_tx_t
*tx
)
1434 ASSERT(tx
->tx_txg
== 0);
1436 while ((txh
= list_head(&tx
->tx_holds
))) {
1437 dnode_t
*dn
= txh
->txh_dnode
;
1439 list_remove(&tx
->tx_holds
, txh
);
1440 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1446 * Call any registered callbacks with an error code.
1448 if (!list_is_empty(&tx
->tx_callbacks
))
1449 dmu_tx_do_callbacks(&tx
->tx_callbacks
, ECANCELED
);
1451 list_destroy(&tx
->tx_callbacks
);
1452 list_destroy(&tx
->tx_holds
);
1454 refcount_destroy_many(&tx
->tx_space_written
,
1455 refcount_count(&tx
->tx_space_written
));
1456 refcount_destroy_many(&tx
->tx_space_freed
,
1457 refcount_count(&tx
->tx_space_freed
));
1459 kmem_free(tx
, sizeof (dmu_tx_t
));
1463 dmu_tx_get_txg(dmu_tx_t
*tx
)
1465 ASSERT(tx
->tx_txg
!= 0);
1466 return (tx
->tx_txg
);
1470 dmu_tx_pool(dmu_tx_t
*tx
)
1472 ASSERT(tx
->tx_pool
!= NULL
);
1473 return (tx
->tx_pool
);
1477 dmu_tx_callback_register(dmu_tx_t
*tx
, dmu_tx_callback_func_t
*func
, void *data
)
1479 dmu_tx_callback_t
*dcb
;
1481 dcb
= kmem_alloc(sizeof (dmu_tx_callback_t
), KM_SLEEP
);
1483 dcb
->dcb_func
= func
;
1484 dcb
->dcb_data
= data
;
1486 list_insert_tail(&tx
->tx_callbacks
, dcb
);
1490 * Call all the commit callbacks on a list, with a given error code.
1493 dmu_tx_do_callbacks(list_t
*cb_list
, int error
)
1495 dmu_tx_callback_t
*dcb
;
1497 while ((dcb
= list_head(cb_list
))) {
1498 list_remove(cb_list
, dcb
);
1499 dcb
->dcb_func(dcb
->dcb_data
, error
);
1500 kmem_free(dcb
, sizeof (dmu_tx_callback_t
));
1505 * Interface to hold a bunch of attributes.
1506 * used for creating new files.
1507 * attrsize is the total size of all attributes
1508 * to be added during object creation
1510 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1514 * hold necessary attribute name for attribute registration.
1515 * should be a very rare case where this is needed. If it does
1516 * happen it would only happen on the first write to the file system.
1519 dmu_tx_sa_registration_hold(sa_os_t
*sa
, dmu_tx_t
*tx
)
1523 if (!sa
->sa_need_attr_registration
)
1526 for (i
= 0; i
!= sa
->sa_num_attrs
; i
++) {
1527 if (!sa
->sa_attr_table
[i
].sa_registered
) {
1528 if (sa
->sa_reg_attr_obj
)
1529 dmu_tx_hold_zap(tx
, sa
->sa_reg_attr_obj
,
1530 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1532 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
,
1533 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1540 dmu_tx_hold_spill(dmu_tx_t
*tx
, uint64_t object
)
1545 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, object
,
1550 dn
= txh
->txh_dnode
;
1555 /* If blkptr doesn't exist then add space to towrite */
1556 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
1557 txh
->txh_space_towrite
+= SPA_OLD_MAXBLOCKSIZE
;
1561 bp
= &dn
->dn_phys
->dn_spill
;
1562 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
1564 txh
->txh_space_tooverwrite
+= SPA_OLD_MAXBLOCKSIZE
;
1566 txh
->txh_space_towrite
+= SPA_OLD_MAXBLOCKSIZE
;
1567 if (!BP_IS_HOLE(bp
))
1568 txh
->txh_space_tounref
+= SPA_OLD_MAXBLOCKSIZE
;
1573 dmu_tx_hold_sa_create(dmu_tx_t
*tx
, int attrsize
)
1575 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1577 dmu_tx_hold_bonus(tx
, DMU_NEW_OBJECT
);
1579 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1582 if (tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1583 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1585 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1586 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1587 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1588 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1591 dmu_tx_sa_registration_hold(sa
, tx
);
1593 if (attrsize
<= DN_MAX_BONUSLEN
&& !sa
->sa_force_spill
)
1596 (void) dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, DMU_NEW_OBJECT
,
1603 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1605 * variable_size is the total size of all variable sized attributes
1606 * passed to this function. It is not the total size of all
1607 * variable size attributes that *may* exist on this object.
1610 dmu_tx_hold_sa(dmu_tx_t
*tx
, sa_handle_t
*hdl
, boolean_t may_grow
)
1613 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1615 ASSERT(hdl
!= NULL
);
1617 object
= sa_handle_object(hdl
);
1619 dmu_tx_hold_bonus(tx
, object
);
1621 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1624 if (tx
->tx_objset
->os_sa
->sa_reg_attr_obj
== 0 ||
1625 tx
->tx_objset
->os_sa
->sa_layout_attr_obj
== 0) {
1626 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1627 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1628 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1629 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1632 dmu_tx_sa_registration_hold(sa
, tx
);
1634 if (may_grow
&& tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1635 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1637 if (sa
->sa_force_spill
|| may_grow
|| hdl
->sa_spill
) {
1638 ASSERT(tx
->tx_txg
== 0);
1639 dmu_tx_hold_spill(tx
, object
);
1641 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)hdl
->sa_bonus
;
1646 if (dn
->dn_have_spill
) {
1647 ASSERT(tx
->tx_txg
== 0);
1648 dmu_tx_hold_spill(tx
, object
);
1657 dmu_tx_ksp
= kstat_create("zfs", 0, "dmu_tx", "misc",
1658 KSTAT_TYPE_NAMED
, sizeof (dmu_tx_stats
) / sizeof (kstat_named_t
),
1659 KSTAT_FLAG_VIRTUAL
);
1661 if (dmu_tx_ksp
!= NULL
) {
1662 dmu_tx_ksp
->ks_data
= &dmu_tx_stats
;
1663 kstat_install(dmu_tx_ksp
);
1670 if (dmu_tx_ksp
!= NULL
) {
1671 kstat_delete(dmu_tx_ksp
);
1676 #if defined(_KERNEL) && defined(HAVE_SPL)
1677 EXPORT_SYMBOL(dmu_tx_create
);
1678 EXPORT_SYMBOL(dmu_tx_hold_write
);
1679 EXPORT_SYMBOL(dmu_tx_hold_free
);
1680 EXPORT_SYMBOL(dmu_tx_hold_zap
);
1681 EXPORT_SYMBOL(dmu_tx_hold_bonus
);
1682 EXPORT_SYMBOL(dmu_tx_abort
);
1683 EXPORT_SYMBOL(dmu_tx_assign
);
1684 EXPORT_SYMBOL(dmu_tx_wait
);
1685 EXPORT_SYMBOL(dmu_tx_commit
);
1686 EXPORT_SYMBOL(dmu_tx_get_txg
);
1687 EXPORT_SYMBOL(dmu_tx_callback_register
);
1688 EXPORT_SYMBOL(dmu_tx_do_callbacks
);
1689 EXPORT_SYMBOL(dmu_tx_hold_spill
);
1690 EXPORT_SYMBOL(dmu_tx_hold_sa_create
);
1691 EXPORT_SYMBOL(dmu_tx_hold_sa
);