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, 2016 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
,
336 FALSE
, FALSE
, FTAG
, &db
);
337 rw_exit(&dn
->dn_struct_rwlock
);
340 txh
->txh_tx
->tx_err
= err
;
344 dmu_tx_count_twig(txh
, dn
, db
, 0, start
, B_FALSE
,
349 * Account for new indirects appearing
350 * before this IO gets assigned into a txg.
353 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
354 for (bits
-= epbs
* (nlvls
- 1);
355 bits
>= 0; bits
-= epbs
)
356 txh
->txh_fudge
+= 1ULL << max_ibs
;
362 delta
= dn
->dn_datablksz
;
367 * 'end' is the last thing we will access, not one past.
368 * This way we won't overflow when accessing the last byte.
370 start
= P2ALIGN(off
, 1ULL << max_bs
);
371 end
= P2ROUNDUP(off
+ len
, 1ULL << max_bs
) - 1;
372 txh
->txh_space_towrite
+= end
- start
+ 1;
377 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
380 * The object contains at most 2^(64 - min_bs) blocks,
381 * and each indirect level maps 2^epbs.
383 for (bits
= 64 - min_bs
; bits
>= 0; bits
-= epbs
) {
386 ASSERT3U(end
, >=, start
);
387 txh
->txh_space_towrite
+= (end
- start
+ 1) << max_ibs
;
390 * We also need a new blkid=0 indirect block
391 * to reference any existing file data.
393 txh
->txh_space_towrite
+= 1ULL << max_ibs
;
398 if (txh
->txh_space_towrite
+ txh
->txh_space_tooverwrite
>
400 err
= SET_ERROR(EFBIG
);
403 txh
->txh_tx
->tx_err
= err
;
407 dmu_tx_count_dnode(dmu_tx_hold_t
*txh
)
409 dnode_t
*dn
= txh
->txh_dnode
;
410 dnode_t
*mdn
= DMU_META_DNODE(txh
->txh_tx
->tx_objset
);
411 uint64_t space
= mdn
->dn_datablksz
+
412 ((mdn
->dn_nlevels
-1) << mdn
->dn_indblkshift
);
414 if (dn
&& dn
->dn_dbuf
->db_blkptr
&&
415 dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
416 dn
->dn_dbuf
->db_blkptr
, dn
->dn_dbuf
->db_blkptr
->blk_birth
)) {
417 txh
->txh_space_tooverwrite
+= space
;
418 txh
->txh_space_tounref
+= space
;
420 txh
->txh_space_towrite
+= space
;
421 if (dn
&& dn
->dn_dbuf
->db_blkptr
)
422 txh
->txh_space_tounref
+= space
;
427 dmu_tx_hold_write(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, int len
)
431 ASSERT(tx
->tx_txg
== 0);
432 ASSERT(len
<= DMU_MAX_ACCESS
);
433 ASSERT(len
== 0 || UINT64_MAX
- off
>= len
- 1);
435 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
436 object
, THT_WRITE
, off
, len
);
440 dmu_tx_count_write(txh
, off
, len
);
441 dmu_tx_count_dnode(txh
);
445 dmu_tx_count_free(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
447 uint64_t blkid
, nblks
, lastblk
;
448 uint64_t space
= 0, unref
= 0, skipped
= 0;
449 dnode_t
*dn
= txh
->txh_dnode
;
450 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
451 spa_t
*spa
= txh
->txh_tx
->tx_pool
->dp_spa
;
453 uint64_t l0span
= 0, nl1blks
= 0;
455 if (dn
->dn_nlevels
== 0)
459 * The struct_rwlock protects us against dn_nlevels
460 * changing, in case (against all odds) we manage to dirty &
461 * sync out the changes after we check for being dirty.
462 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
464 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
465 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
466 if (dn
->dn_maxblkid
== 0) {
467 if (off
== 0 && len
>= dn
->dn_datablksz
) {
471 rw_exit(&dn
->dn_struct_rwlock
);
475 blkid
= off
>> dn
->dn_datablkshift
;
476 nblks
= (len
+ dn
->dn_datablksz
- 1) >> dn
->dn_datablkshift
;
478 if (blkid
> dn
->dn_maxblkid
) {
479 rw_exit(&dn
->dn_struct_rwlock
);
482 if (blkid
+ nblks
> dn
->dn_maxblkid
)
483 nblks
= dn
->dn_maxblkid
- blkid
+ 1;
486 l0span
= nblks
; /* save for later use to calc level > 1 overhead */
487 if (dn
->dn_nlevels
== 1) {
489 for (i
= 0; i
< nblks
; i
++) {
490 blkptr_t
*bp
= dn
->dn_phys
->dn_blkptr
;
491 ASSERT3U(blkid
+ i
, <, dn
->dn_nblkptr
);
493 if (dsl_dataset_block_freeable(ds
, bp
, bp
->blk_birth
)) {
494 dprintf_bp(bp
, "can free old%s", "");
495 space
+= bp_get_dsize(spa
, bp
);
497 unref
+= BP_GET_ASIZE(bp
);
503 lastblk
= blkid
+ nblks
- 1;
505 dmu_buf_impl_t
*dbuf
;
506 uint64_t ibyte
, new_blkid
;
508 int err
, i
, blkoff
, tochk
;
511 ibyte
= blkid
<< dn
->dn_datablkshift
;
512 err
= dnode_next_offset(dn
,
513 DNODE_FIND_HAVELOCK
, &ibyte
, 2, 1, 0);
514 new_blkid
= ibyte
>> dn
->dn_datablkshift
;
516 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
520 txh
->txh_tx
->tx_err
= err
;
523 if (new_blkid
> lastblk
) {
524 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
528 if (new_blkid
> blkid
) {
529 ASSERT((new_blkid
>> epbs
) > (blkid
>> epbs
));
530 skipped
+= (new_blkid
>> epbs
) - (blkid
>> epbs
) - 1;
531 nblks
-= new_blkid
- blkid
;
534 blkoff
= P2PHASE(blkid
, epb
);
535 tochk
= MIN(epb
- blkoff
, nblks
);
537 err
= dbuf_hold_impl(dn
, 1, blkid
>> epbs
,
538 FALSE
, FALSE
, FTAG
, &dbuf
);
540 txh
->txh_tx
->tx_err
= err
;
544 txh
->txh_memory_tohold
+= dbuf
->db
.db_size
;
547 * We don't check memory_tohold against DMU_MAX_ACCESS because
548 * memory_tohold is an over-estimation (especially the >L1
549 * indirect blocks), so it could fail. Callers should have
550 * already verified that they will not be holding too much
554 err
= dbuf_read(dbuf
, NULL
, DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
);
556 txh
->txh_tx
->tx_err
= err
;
557 dbuf_rele(dbuf
, FTAG
);
561 bp
= dbuf
->db
.db_data
;
564 for (i
= 0; i
< tochk
; i
++) {
565 if (dsl_dataset_block_freeable(ds
, &bp
[i
],
567 dprintf_bp(&bp
[i
], "can free old%s", "");
568 space
+= bp_get_dsize(spa
, &bp
[i
]);
570 unref
+= BP_GET_ASIZE(bp
);
572 dbuf_rele(dbuf
, FTAG
);
578 rw_exit(&dn
->dn_struct_rwlock
);
581 * Add in memory requirements of higher-level indirects.
582 * This assumes a worst-possible scenario for dn_nlevels and a
583 * worst-possible distribution of l1-blocks over the region to free.
586 uint64_t blkcnt
= 1 + ((l0span
>> epbs
) >> epbs
);
589 * Here we don't use DN_MAX_LEVEL, but calculate it with the
590 * given datablkshift and indblkshift. This makes the
591 * difference between 19 and 8 on large files.
593 int maxlevel
= 2 + (DN_MAX_OFFSET_SHIFT
- dn
->dn_datablkshift
) /
594 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
596 while (level
++ < maxlevel
) {
597 txh
->txh_memory_tohold
+= MAX(MIN(blkcnt
, nl1blks
), 1)
598 << dn
->dn_indblkshift
;
599 blkcnt
= 1 + (blkcnt
>> epbs
);
603 /* account for new level 1 indirect blocks that might show up */
605 txh
->txh_fudge
+= skipped
<< dn
->dn_indblkshift
;
606 skipped
= MIN(skipped
, DMU_MAX_DELETEBLKCNT
>> epbs
);
607 txh
->txh_memory_tohold
+= skipped
<< dn
->dn_indblkshift
;
609 txh
->txh_space_tofree
+= space
;
610 txh
->txh_space_tounref
+= unref
;
614 * This function marks the transaction as being a "net free". The end
615 * result is that refquotas will be disabled for this transaction, and
616 * this transaction will be able to use half of the pool space overhead
617 * (see dsl_pool_adjustedsize()). Therefore this function should only
618 * be called for transactions that we expect will not cause a net increase
619 * in the amount of space used (but it's OK if that is occasionally not true).
622 dmu_tx_mark_netfree(dmu_tx_t
*tx
)
626 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
627 DMU_NEW_OBJECT
, THT_FREE
, 0, 0);
630 * Pretend that this operation will free 1GB of space. This
631 * should be large enough to cancel out the largest write.
632 * We don't want to use something like UINT64_MAX, because that would
633 * cause overflows when doing math with these values (e.g. in
634 * dmu_tx_try_assign()).
636 txh
->txh_space_tofree
= txh
->txh_space_tounref
= 1024 * 1024 * 1024;
640 dmu_tx_hold_free(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, uint64_t len
)
647 ASSERT(tx
->tx_txg
== 0);
649 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
650 object
, THT_FREE
, off
, len
);
654 dmu_tx_count_dnode(txh
);
656 if (off
>= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
)
658 if (len
== DMU_OBJECT_END
)
659 len
= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
- off
;
661 dmu_tx_count_dnode(txh
);
664 * For i/o error checking, we read the first and last level-0
665 * blocks if they are not aligned, and all the level-1 blocks.
667 * Note: dbuf_free_range() assumes that we have not instantiated
668 * any level-0 dbufs that will be completely freed. Therefore we must
669 * exercise care to not read or count the first and last blocks
670 * if they are blocksize-aligned.
672 if (dn
->dn_datablkshift
== 0) {
673 if (off
!= 0 || len
< dn
->dn_datablksz
)
674 dmu_tx_count_write(txh
, 0, dn
->dn_datablksz
);
676 /* first block will be modified if it is not aligned */
677 if (!IS_P2ALIGNED(off
, 1 << dn
->dn_datablkshift
))
678 dmu_tx_count_write(txh
, off
, 1);
679 /* last block will be modified if it is not aligned */
680 if (!IS_P2ALIGNED(off
+ len
, 1 << dn
->dn_datablkshift
))
681 dmu_tx_count_write(txh
, off
+len
, 1);
685 * Check level-1 blocks.
687 if (dn
->dn_nlevels
> 1) {
688 int shift
= dn
->dn_datablkshift
+ dn
->dn_indblkshift
-
690 uint64_t start
= off
>> shift
;
691 uint64_t end
= (off
+ len
) >> shift
;
694 ASSERT(dn
->dn_indblkshift
!= 0);
697 * dnode_reallocate() can result in an object with indirect
698 * blocks having an odd data block size. In this case,
699 * just check the single block.
701 if (dn
->dn_datablkshift
== 0)
704 zio
= zio_root(tx
->tx_pool
->dp_spa
,
705 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
706 for (i
= start
; i
<= end
; i
++) {
707 uint64_t ibyte
= i
<< shift
;
708 err
= dnode_next_offset(dn
, 0, &ibyte
, 2, 1, 0);
710 if (err
== ESRCH
|| i
> end
)
717 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
730 dmu_tx_count_free(txh
, off
, len
);
734 dmu_tx_hold_zap(dmu_tx_t
*tx
, uint64_t object
, int add
, const char *name
)
738 dsl_dataset_phys_t
*ds_phys
;
742 ASSERT(tx
->tx_txg
== 0);
744 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
745 object
, THT_ZAP
, add
, (uintptr_t)name
);
750 dmu_tx_count_dnode(txh
);
754 * We will be able to fit a new object's entries into one leaf
755 * block. So there will be at most 2 blocks total,
756 * including the header block.
758 dmu_tx_count_write(txh
, 0, 2 << fzap_default_block_shift
);
762 ASSERT3U(DMU_OT_BYTESWAP(dn
->dn_type
), ==, DMU_BSWAP_ZAP
);
764 if (dn
->dn_maxblkid
== 0 && !add
) {
768 * If there is only one block (i.e. this is a micro-zap)
769 * and we are not adding anything, the accounting is simple.
771 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
778 * Use max block size here, since we don't know how much
779 * the size will change between now and the dbuf dirty call.
781 bp
= &dn
->dn_phys
->dn_blkptr
[0];
782 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
784 txh
->txh_space_tooverwrite
+= MZAP_MAX_BLKSZ
;
786 txh
->txh_space_towrite
+= MZAP_MAX_BLKSZ
;
788 txh
->txh_space_tounref
+= MZAP_MAX_BLKSZ
;
792 if (dn
->dn_maxblkid
> 0 && name
) {
794 * access the name in this fat-zap so that we'll check
795 * for i/o errors to the leaf blocks, etc.
797 err
= zap_lookup_by_dnode(dn
, name
, 8, 0, NULL
);
804 err
= zap_count_write_by_dnode(dn
, name
, add
,
805 &txh
->txh_space_towrite
, &txh
->txh_space_tooverwrite
);
808 * If the modified blocks are scattered to the four winds,
809 * we'll have to modify an indirect twig for each.
811 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
812 ds_phys
= dsl_dataset_phys(dn
->dn_objset
->os_dsl_dataset
);
813 for (nblocks
= dn
->dn_maxblkid
>> epbs
; nblocks
!= 0; nblocks
>>= epbs
)
814 if (ds_phys
->ds_prev_snap_obj
)
815 txh
->txh_space_towrite
+= 3 << dn
->dn_indblkshift
;
817 txh
->txh_space_tooverwrite
+= 3 << dn
->dn_indblkshift
;
821 dmu_tx_hold_bonus(dmu_tx_t
*tx
, uint64_t object
)
825 ASSERT(tx
->tx_txg
== 0);
827 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
828 object
, THT_BONUS
, 0, 0);
830 dmu_tx_count_dnode(txh
);
834 dmu_tx_hold_space(dmu_tx_t
*tx
, uint64_t space
)
838 ASSERT(tx
->tx_txg
== 0);
840 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
841 DMU_NEW_OBJECT
, THT_SPACE
, space
, 0);
843 txh
->txh_space_towrite
+= space
;
847 dmu_tx_holds(dmu_tx_t
*tx
, uint64_t object
)
853 * By asserting that the tx is assigned, we're counting the
854 * number of dn_tx_holds, which is the same as the number of
855 * dn_holds. Otherwise, we'd be counting dn_holds, but
856 * dn_tx_holds could be 0.
858 ASSERT(tx
->tx_txg
!= 0);
860 /* if (tx->tx_anyobj == TRUE) */
863 for (txh
= list_head(&tx
->tx_holds
); txh
;
864 txh
= list_next(&tx
->tx_holds
, txh
)) {
865 if (txh
->txh_dnode
&& txh
->txh_dnode
->dn_object
== object
)
874 dmu_tx_dirty_buf(dmu_tx_t
*tx
, dmu_buf_impl_t
*db
)
877 int match_object
= FALSE
, match_offset
= FALSE
;
883 ASSERT(tx
->tx_txg
!= 0);
884 ASSERT(tx
->tx_objset
== NULL
|| dn
->dn_objset
== tx
->tx_objset
);
885 ASSERT3U(dn
->dn_object
, ==, db
->db
.db_object
);
892 /* XXX No checking on the meta dnode for now */
893 if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
898 for (txh
= list_head(&tx
->tx_holds
); txh
;
899 txh
= list_next(&tx
->tx_holds
, txh
)) {
900 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
901 if (txh
->txh_dnode
== dn
&& txh
->txh_type
!= THT_NEWOBJECT
)
903 if (txh
->txh_dnode
== NULL
|| txh
->txh_dnode
== dn
) {
904 int datablkshift
= dn
->dn_datablkshift
?
905 dn
->dn_datablkshift
: SPA_MAXBLOCKSHIFT
;
906 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
907 int shift
= datablkshift
+ epbs
* db
->db_level
;
908 uint64_t beginblk
= shift
>= 64 ? 0 :
909 (txh
->txh_arg1
>> shift
);
910 uint64_t endblk
= shift
>= 64 ? 0 :
911 ((txh
->txh_arg1
+ txh
->txh_arg2
- 1) >> shift
);
912 uint64_t blkid
= db
->db_blkid
;
914 /* XXX txh_arg2 better not be zero... */
916 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
917 txh
->txh_type
, beginblk
, endblk
);
919 switch (txh
->txh_type
) {
921 if (blkid
>= beginblk
&& blkid
<= endblk
)
924 * We will let this hold work for the bonus
925 * or spill buffer so that we don't need to
926 * hold it when creating a new object.
928 if (blkid
== DMU_BONUS_BLKID
||
929 blkid
== DMU_SPILL_BLKID
)
932 * They might have to increase nlevels,
933 * thus dirtying the new TLIBs. Or the
934 * might have to change the block size,
935 * thus dirying the new lvl=0 blk=0.
942 * We will dirty all the level 1 blocks in
943 * the free range and perhaps the first and
944 * last level 0 block.
946 if (blkid
>= beginblk
&& (blkid
<= endblk
||
947 txh
->txh_arg2
== DMU_OBJECT_END
))
951 if (blkid
== DMU_SPILL_BLKID
)
955 if (blkid
== DMU_BONUS_BLKID
)
965 cmn_err(CE_PANIC
, "bad txh_type %d",
969 if (match_object
&& match_offset
) {
975 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
976 (u_longlong_t
)db
->db
.db_object
, db
->db_level
,
977 (u_longlong_t
)db
->db_blkid
);
982 * If we can't do 10 iops, something is wrong. Let us go ahead
983 * and hit zfs_dirty_data_max.
985 hrtime_t zfs_delay_max_ns
= 100 * MICROSEC
; /* 100 milliseconds */
986 int zfs_delay_resolution_ns
= 100 * 1000; /* 100 microseconds */
989 * We delay transactions when we've determined that the backend storage
990 * isn't able to accommodate the rate of incoming writes.
992 * If there is already a transaction waiting, we delay relative to when
993 * that transaction finishes waiting. This way the calculated min_time
994 * is independent of the number of threads concurrently executing
997 * If we are the only waiter, wait relative to when the transaction
998 * started, rather than the current time. This credits the transaction for
999 * "time already served", e.g. reading indirect blocks.
1001 * The minimum time for a transaction to take is calculated as:
1002 * min_time = scale * (dirty - min) / (max - dirty)
1003 * min_time is then capped at zfs_delay_max_ns.
1005 * The delay has two degrees of freedom that can be adjusted via tunables.
1006 * The percentage of dirty data at which we start to delay is defined by
1007 * zfs_delay_min_dirty_percent. This should typically be at or above
1008 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
1009 * delay after writing at full speed has failed to keep up with the incoming
1010 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
1011 * speaking, this variable determines the amount of delay at the midpoint of
1015 * 10ms +-------------------------------------------------------------*+
1031 * 2ms + (midpoint) * +
1034 * | zfs_delay_scale ----------> ******** |
1035 * 0 +-------------------------------------*********----------------+
1036 * 0% <- zfs_dirty_data_max -> 100%
1038 * Note that since the delay is added to the outstanding time remaining on the
1039 * most recent transaction, the delay is effectively the inverse of IOPS.
1040 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1041 * was chosen such that small changes in the amount of accumulated dirty data
1042 * in the first 3/4 of the curve yield relatively small differences in the
1045 * The effects can be easier to understand when the amount of delay is
1046 * represented on a log scale:
1049 * 100ms +-------------------------------------------------------------++
1058 * + zfs_delay_scale ----------> ***** +
1069 * +--------------------------------------------------------------+
1070 * 0% <- zfs_dirty_data_max -> 100%
1072 * Note here that only as the amount of dirty data approaches its limit does
1073 * the delay start to increase rapidly. The goal of a properly tuned system
1074 * should be to keep the amount of dirty data out of that range by first
1075 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1076 * optimal throughput on the backend storage, and then by changing the value
1077 * of zfs_delay_scale to increase the steepness of the curve.
1080 dmu_tx_delay(dmu_tx_t
*tx
, uint64_t dirty
)
1082 dsl_pool_t
*dp
= tx
->tx_pool
;
1083 uint64_t delay_min_bytes
=
1084 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
1085 hrtime_t wakeup
, min_tx_time
, now
;
1087 if (dirty
<= delay_min_bytes
)
1091 * The caller has already waited until we are under the max.
1092 * We make them pass us the amount of dirty data so we don't
1093 * have to handle the case of it being >= the max, which could
1094 * cause a divide-by-zero if it's == the max.
1096 ASSERT3U(dirty
, <, zfs_dirty_data_max
);
1099 min_tx_time
= zfs_delay_scale
*
1100 (dirty
- delay_min_bytes
) / (zfs_dirty_data_max
- dirty
);
1101 min_tx_time
= MIN(min_tx_time
, zfs_delay_max_ns
);
1102 if (now
> tx
->tx_start
+ min_tx_time
)
1105 DTRACE_PROBE3(delay__mintime
, dmu_tx_t
*, tx
, uint64_t, dirty
,
1106 uint64_t, min_tx_time
);
1108 mutex_enter(&dp
->dp_lock
);
1109 wakeup
= MAX(tx
->tx_start
+ min_tx_time
,
1110 dp
->dp_last_wakeup
+ min_tx_time
);
1111 dp
->dp_last_wakeup
= wakeup
;
1112 mutex_exit(&dp
->dp_lock
);
1114 zfs_sleep_until(wakeup
);
1118 dmu_tx_try_assign(dmu_tx_t
*tx
, txg_how_t txg_how
)
1121 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
1122 uint64_t memory
, asize
, fsize
, usize
;
1123 uint64_t towrite
, tofree
, tooverwrite
, tounref
, tohold
, fudge
;
1125 ASSERT0(tx
->tx_txg
);
1128 DMU_TX_STAT_BUMP(dmu_tx_error
);
1129 return (tx
->tx_err
);
1132 if (spa_suspended(spa
)) {
1133 DMU_TX_STAT_BUMP(dmu_tx_suspended
);
1136 * If the user has indicated a blocking failure mode
1137 * then return ERESTART which will block in dmu_tx_wait().
1138 * Otherwise, return EIO so that an error can get
1139 * propagated back to the VOP calls.
1141 * Note that we always honor the txg_how flag regardless
1142 * of the failuremode setting.
1144 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_CONTINUE
&&
1145 txg_how
!= TXG_WAIT
)
1146 return (SET_ERROR(EIO
));
1148 return (SET_ERROR(ERESTART
));
1151 if (!tx
->tx_waited
&&
1152 dsl_pool_need_dirty_delay(tx
->tx_pool
)) {
1153 tx
->tx_wait_dirty
= B_TRUE
;
1154 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay
);
1158 tx
->tx_txg
= txg_hold_open(tx
->tx_pool
, &tx
->tx_txgh
);
1159 tx
->tx_needassign_txh
= NULL
;
1162 * NB: No error returns are allowed after txg_hold_open, but
1163 * before processing the dnode holds, due to the
1164 * dmu_tx_unassign() logic.
1167 towrite
= tofree
= tooverwrite
= tounref
= tohold
= fudge
= 0;
1168 for (txh
= list_head(&tx
->tx_holds
); txh
;
1169 txh
= list_next(&tx
->tx_holds
, txh
)) {
1170 dnode_t
*dn
= txh
->txh_dnode
;
1172 mutex_enter(&dn
->dn_mtx
);
1173 if (dn
->dn_assigned_txg
== tx
->tx_txg
- 1) {
1174 mutex_exit(&dn
->dn_mtx
);
1175 tx
->tx_needassign_txh
= txh
;
1176 DMU_TX_STAT_BUMP(dmu_tx_group
);
1177 return (SET_ERROR(ERESTART
));
1179 if (dn
->dn_assigned_txg
== 0)
1180 dn
->dn_assigned_txg
= tx
->tx_txg
;
1181 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1182 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
1183 mutex_exit(&dn
->dn_mtx
);
1185 towrite
+= txh
->txh_space_towrite
;
1186 tofree
+= txh
->txh_space_tofree
;
1187 tooverwrite
+= txh
->txh_space_tooverwrite
;
1188 tounref
+= txh
->txh_space_tounref
;
1189 tohold
+= txh
->txh_memory_tohold
;
1190 fudge
+= txh
->txh_fudge
;
1194 * If a snapshot has been taken since we made our estimates,
1195 * assume that we won't be able to free or overwrite anything.
1197 if (tx
->tx_objset
&&
1198 dsl_dataset_prev_snap_txg(tx
->tx_objset
->os_dsl_dataset
) >
1199 tx
->tx_lastsnap_txg
) {
1200 towrite
+= tooverwrite
;
1201 tooverwrite
= tofree
= 0;
1204 /* needed allocation: worst-case estimate of write space */
1205 asize
= spa_get_asize(tx
->tx_pool
->dp_spa
, towrite
+ tooverwrite
);
1206 /* freed space estimate: worst-case overwrite + free estimate */
1207 fsize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tooverwrite
) + tofree
;
1208 /* convert unrefd space to worst-case estimate */
1209 usize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tounref
);
1210 /* calculate memory footprint estimate */
1211 memory
= towrite
+ tooverwrite
+ tohold
;
1215 * Add in 'tohold' to account for our dirty holds on this memory
1216 * XXX - the "fudge" factor is to account for skipped blocks that
1217 * we missed because dnode_next_offset() misses in-core-only blocks.
1219 tx
->tx_space_towrite
= asize
+
1220 spa_get_asize(tx
->tx_pool
->dp_spa
, tohold
+ fudge
);
1221 tx
->tx_space_tofree
= tofree
;
1222 tx
->tx_space_tooverwrite
= tooverwrite
;
1223 tx
->tx_space_tounref
= tounref
;
1226 if (tx
->tx_dir
&& asize
!= 0) {
1227 int err
= dsl_dir_tempreserve_space(tx
->tx_dir
, memory
,
1228 asize
, fsize
, usize
, &tx
->tx_tempreserve_cookie
, tx
);
1233 DMU_TX_STAT_BUMP(dmu_tx_assigned
);
1239 dmu_tx_unassign(dmu_tx_t
*tx
)
1243 if (tx
->tx_txg
== 0)
1246 txg_rele_to_quiesce(&tx
->tx_txgh
);
1249 * Walk the transaction's hold list, removing the hold on the
1250 * associated dnode, and notifying waiters if the refcount drops to 0.
1252 for (txh
= list_head(&tx
->tx_holds
); txh
!= tx
->tx_needassign_txh
;
1253 txh
= list_next(&tx
->tx_holds
, txh
)) {
1254 dnode_t
*dn
= txh
->txh_dnode
;
1258 mutex_enter(&dn
->dn_mtx
);
1259 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1261 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1262 dn
->dn_assigned_txg
= 0;
1263 cv_broadcast(&dn
->dn_notxholds
);
1265 mutex_exit(&dn
->dn_mtx
);
1268 txg_rele_to_sync(&tx
->tx_txgh
);
1270 tx
->tx_lasttried_txg
= tx
->tx_txg
;
1275 * Assign tx to a transaction group. txg_how can be one of:
1277 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1278 * a new one. This should be used when you're not holding locks.
1279 * It will only fail if we're truly out of space (or over quota).
1281 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1282 * blocking, returns immediately with ERESTART. This should be used
1283 * whenever you're holding locks. On an ERESTART error, the caller
1284 * should drop locks, do a dmu_tx_wait(tx), and try again.
1286 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1287 * has already been called on behalf of this operation (though
1288 * most likely on a different tx).
1291 dmu_tx_assign(dmu_tx_t
*tx
, txg_how_t txg_how
)
1295 ASSERT(tx
->tx_txg
== 0);
1296 ASSERT(txg_how
== TXG_WAIT
|| txg_how
== TXG_NOWAIT
||
1297 txg_how
== TXG_WAITED
);
1298 ASSERT(!dsl_pool_sync_context(tx
->tx_pool
));
1300 if (txg_how
== TXG_WAITED
)
1301 tx
->tx_waited
= B_TRUE
;
1303 /* If we might wait, we must not hold the config lock. */
1304 ASSERT(txg_how
!= TXG_WAIT
|| !dsl_pool_config_held(tx
->tx_pool
));
1306 while ((err
= dmu_tx_try_assign(tx
, txg_how
)) != 0) {
1307 dmu_tx_unassign(tx
);
1309 if (err
!= ERESTART
|| txg_how
!= TXG_WAIT
)
1315 txg_rele_to_quiesce(&tx
->tx_txgh
);
1321 dmu_tx_wait(dmu_tx_t
*tx
)
1323 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
1324 dsl_pool_t
*dp
= tx
->tx_pool
;
1327 ASSERT(tx
->tx_txg
== 0);
1328 ASSERT(!dsl_pool_config_held(tx
->tx_pool
));
1330 before
= gethrtime();
1332 if (tx
->tx_wait_dirty
) {
1336 * dmu_tx_try_assign() has determined that we need to wait
1337 * because we've consumed much or all of the dirty buffer
1340 mutex_enter(&dp
->dp_lock
);
1341 if (dp
->dp_dirty_total
>= zfs_dirty_data_max
)
1342 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max
);
1343 while (dp
->dp_dirty_total
>= zfs_dirty_data_max
)
1344 cv_wait(&dp
->dp_spaceavail_cv
, &dp
->dp_lock
);
1345 dirty
= dp
->dp_dirty_total
;
1346 mutex_exit(&dp
->dp_lock
);
1348 dmu_tx_delay(tx
, dirty
);
1350 tx
->tx_wait_dirty
= B_FALSE
;
1353 * Note: setting tx_waited only has effect if the caller
1354 * used TX_WAIT. Otherwise they are going to destroy
1355 * this tx and try again. The common case, zfs_write(),
1358 tx
->tx_waited
= B_TRUE
;
1359 } else if (spa_suspended(spa
) || tx
->tx_lasttried_txg
== 0) {
1361 * If the pool is suspended we need to wait until it
1362 * is resumed. Note that it's possible that the pool
1363 * has become active after this thread has tried to
1364 * obtain a tx. If that's the case then tx_lasttried_txg
1365 * would not have been set.
1367 txg_wait_synced(dp
, spa_last_synced_txg(spa
) + 1);
1368 } else if (tx
->tx_needassign_txh
) {
1369 dnode_t
*dn
= tx
->tx_needassign_txh
->txh_dnode
;
1371 mutex_enter(&dn
->dn_mtx
);
1372 while (dn
->dn_assigned_txg
== tx
->tx_lasttried_txg
- 1)
1373 cv_wait(&dn
->dn_notxholds
, &dn
->dn_mtx
);
1374 mutex_exit(&dn
->dn_mtx
);
1375 tx
->tx_needassign_txh
= NULL
;
1378 * A dnode is assigned to the quiescing txg. Wait for its
1379 * transaction to complete.
1381 txg_wait_open(tx
->tx_pool
, tx
->tx_lasttried_txg
+ 1);
1384 spa_tx_assign_add_nsecs(spa
, gethrtime() - before
);
1388 dmu_tx_willuse_space(dmu_tx_t
*tx
, int64_t delta
)
1391 if (tx
->tx_dir
== NULL
|| delta
== 0)
1395 ASSERT3U(refcount_count(&tx
->tx_space_written
) + delta
, <=,
1396 tx
->tx_space_towrite
);
1397 (void) refcount_add_many(&tx
->tx_space_written
, delta
, NULL
);
1399 (void) refcount_add_many(&tx
->tx_space_freed
, -delta
, NULL
);
1405 dmu_tx_commit(dmu_tx_t
*tx
)
1409 ASSERT(tx
->tx_txg
!= 0);
1412 * Go through the transaction's hold list and remove holds on
1413 * associated dnodes, notifying waiters if no holds remain.
1415 while ((txh
= list_head(&tx
->tx_holds
))) {
1416 dnode_t
*dn
= txh
->txh_dnode
;
1418 list_remove(&tx
->tx_holds
, txh
);
1419 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1422 mutex_enter(&dn
->dn_mtx
);
1423 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1425 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1426 dn
->dn_assigned_txg
= 0;
1427 cv_broadcast(&dn
->dn_notxholds
);
1429 mutex_exit(&dn
->dn_mtx
);
1433 if (tx
->tx_tempreserve_cookie
)
1434 dsl_dir_tempreserve_clear(tx
->tx_tempreserve_cookie
, tx
);
1436 if (!list_is_empty(&tx
->tx_callbacks
))
1437 txg_register_callbacks(&tx
->tx_txgh
, &tx
->tx_callbacks
);
1439 if (tx
->tx_anyobj
== FALSE
)
1440 txg_rele_to_sync(&tx
->tx_txgh
);
1442 list_destroy(&tx
->tx_callbacks
);
1443 list_destroy(&tx
->tx_holds
);
1445 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1446 tx
->tx_space_towrite
, refcount_count(&tx
->tx_space_written
),
1447 tx
->tx_space_tofree
, refcount_count(&tx
->tx_space_freed
));
1448 refcount_destroy_many(&tx
->tx_space_written
,
1449 refcount_count(&tx
->tx_space_written
));
1450 refcount_destroy_many(&tx
->tx_space_freed
,
1451 refcount_count(&tx
->tx_space_freed
));
1453 kmem_free(tx
, sizeof (dmu_tx_t
));
1457 dmu_tx_abort(dmu_tx_t
*tx
)
1461 ASSERT(tx
->tx_txg
== 0);
1463 while ((txh
= list_head(&tx
->tx_holds
))) {
1464 dnode_t
*dn
= txh
->txh_dnode
;
1466 list_remove(&tx
->tx_holds
, txh
);
1467 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1473 * Call any registered callbacks with an error code.
1475 if (!list_is_empty(&tx
->tx_callbacks
))
1476 dmu_tx_do_callbacks(&tx
->tx_callbacks
, ECANCELED
);
1478 list_destroy(&tx
->tx_callbacks
);
1479 list_destroy(&tx
->tx_holds
);
1481 refcount_destroy_many(&tx
->tx_space_written
,
1482 refcount_count(&tx
->tx_space_written
));
1483 refcount_destroy_many(&tx
->tx_space_freed
,
1484 refcount_count(&tx
->tx_space_freed
));
1486 kmem_free(tx
, sizeof (dmu_tx_t
));
1490 dmu_tx_get_txg(dmu_tx_t
*tx
)
1492 ASSERT(tx
->tx_txg
!= 0);
1493 return (tx
->tx_txg
);
1497 dmu_tx_pool(dmu_tx_t
*tx
)
1499 ASSERT(tx
->tx_pool
!= NULL
);
1500 return (tx
->tx_pool
);
1504 dmu_tx_callback_register(dmu_tx_t
*tx
, dmu_tx_callback_func_t
*func
, void *data
)
1506 dmu_tx_callback_t
*dcb
;
1508 dcb
= kmem_alloc(sizeof (dmu_tx_callback_t
), KM_SLEEP
);
1510 dcb
->dcb_func
= func
;
1511 dcb
->dcb_data
= data
;
1513 list_insert_tail(&tx
->tx_callbacks
, dcb
);
1517 * Call all the commit callbacks on a list, with a given error code.
1520 dmu_tx_do_callbacks(list_t
*cb_list
, int error
)
1522 dmu_tx_callback_t
*dcb
;
1524 while ((dcb
= list_head(cb_list
))) {
1525 list_remove(cb_list
, dcb
);
1526 dcb
->dcb_func(dcb
->dcb_data
, error
);
1527 kmem_free(dcb
, sizeof (dmu_tx_callback_t
));
1532 * Interface to hold a bunch of attributes.
1533 * used for creating new files.
1534 * attrsize is the total size of all attributes
1535 * to be added during object creation
1537 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1541 * hold necessary attribute name for attribute registration.
1542 * should be a very rare case where this is needed. If it does
1543 * happen it would only happen on the first write to the file system.
1546 dmu_tx_sa_registration_hold(sa_os_t
*sa
, dmu_tx_t
*tx
)
1550 if (!sa
->sa_need_attr_registration
)
1553 for (i
= 0; i
!= sa
->sa_num_attrs
; i
++) {
1554 if (!sa
->sa_attr_table
[i
].sa_registered
) {
1555 if (sa
->sa_reg_attr_obj
)
1556 dmu_tx_hold_zap(tx
, sa
->sa_reg_attr_obj
,
1557 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1559 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
,
1560 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1567 dmu_tx_hold_spill(dmu_tx_t
*tx
, uint64_t object
)
1572 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, object
,
1577 dn
= txh
->txh_dnode
;
1582 /* If blkptr doesn't exist then add space to towrite */
1583 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
1584 txh
->txh_space_towrite
+= SPA_OLD_MAXBLOCKSIZE
;
1588 bp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
1589 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
1591 txh
->txh_space_tooverwrite
+= SPA_OLD_MAXBLOCKSIZE
;
1593 txh
->txh_space_towrite
+= SPA_OLD_MAXBLOCKSIZE
;
1594 if (!BP_IS_HOLE(bp
))
1595 txh
->txh_space_tounref
+= SPA_OLD_MAXBLOCKSIZE
;
1600 dmu_tx_hold_sa_create(dmu_tx_t
*tx
, int attrsize
)
1602 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1604 dmu_tx_hold_bonus(tx
, DMU_NEW_OBJECT
);
1606 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1609 if (tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1610 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1612 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1613 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1614 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1615 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1618 dmu_tx_sa_registration_hold(sa
, tx
);
1620 if (attrsize
<= DN_OLD_MAX_BONUSLEN
&& !sa
->sa_force_spill
)
1623 (void) dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, DMU_NEW_OBJECT
,
1630 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1632 * variable_size is the total size of all variable sized attributes
1633 * passed to this function. It is not the total size of all
1634 * variable size attributes that *may* exist on this object.
1637 dmu_tx_hold_sa(dmu_tx_t
*tx
, sa_handle_t
*hdl
, boolean_t may_grow
)
1640 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1642 ASSERT(hdl
!= NULL
);
1644 object
= sa_handle_object(hdl
);
1646 dmu_tx_hold_bonus(tx
, object
);
1648 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1651 if (tx
->tx_objset
->os_sa
->sa_reg_attr_obj
== 0 ||
1652 tx
->tx_objset
->os_sa
->sa_layout_attr_obj
== 0) {
1653 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1654 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1655 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1656 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1659 dmu_tx_sa_registration_hold(sa
, tx
);
1661 if (may_grow
&& tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1662 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1664 if (sa
->sa_force_spill
|| may_grow
|| hdl
->sa_spill
) {
1665 ASSERT(tx
->tx_txg
== 0);
1666 dmu_tx_hold_spill(tx
, object
);
1668 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)hdl
->sa_bonus
;
1673 if (dn
->dn_have_spill
) {
1674 ASSERT(tx
->tx_txg
== 0);
1675 dmu_tx_hold_spill(tx
, object
);
1684 dmu_tx_ksp
= kstat_create("zfs", 0, "dmu_tx", "misc",
1685 KSTAT_TYPE_NAMED
, sizeof (dmu_tx_stats
) / sizeof (kstat_named_t
),
1686 KSTAT_FLAG_VIRTUAL
);
1688 if (dmu_tx_ksp
!= NULL
) {
1689 dmu_tx_ksp
->ks_data
= &dmu_tx_stats
;
1690 kstat_install(dmu_tx_ksp
);
1697 if (dmu_tx_ksp
!= NULL
) {
1698 kstat_delete(dmu_tx_ksp
);
1703 #if defined(_KERNEL) && defined(HAVE_SPL)
1704 EXPORT_SYMBOL(dmu_tx_create
);
1705 EXPORT_SYMBOL(dmu_tx_hold_write
);
1706 EXPORT_SYMBOL(dmu_tx_hold_free
);
1707 EXPORT_SYMBOL(dmu_tx_hold_zap
);
1708 EXPORT_SYMBOL(dmu_tx_hold_bonus
);
1709 EXPORT_SYMBOL(dmu_tx_abort
);
1710 EXPORT_SYMBOL(dmu_tx_assign
);
1711 EXPORT_SYMBOL(dmu_tx_wait
);
1712 EXPORT_SYMBOL(dmu_tx_commit
);
1713 EXPORT_SYMBOL(dmu_tx_get_txg
);
1714 EXPORT_SYMBOL(dmu_tx_callback_register
);
1715 EXPORT_SYMBOL(dmu_tx_do_callbacks
);
1716 EXPORT_SYMBOL(dmu_tx_hold_spill
);
1717 EXPORT_SYMBOL(dmu_tx_hold_sa_create
);
1718 EXPORT_SYMBOL(dmu_tx_hold_sa
);