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) 2013 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>
42 typedef void (*dmu_tx_hold_func_t
)(dmu_tx_t
*tx
, struct dnode
*dn
,
43 uint64_t arg1
, uint64_t arg2
);
47 dmu_tx_create_dd(dsl_dir_t
*dd
)
49 dmu_tx_t
*tx
= kmem_zalloc(sizeof (dmu_tx_t
), KM_SLEEP
);
52 tx
->tx_pool
= dd
->dd_pool
;
53 list_create(&tx
->tx_holds
, sizeof (dmu_tx_hold_t
),
54 offsetof(dmu_tx_hold_t
, txh_node
));
55 list_create(&tx
->tx_callbacks
, sizeof (dmu_tx_callback_t
),
56 offsetof(dmu_tx_callback_t
, dcb_node
));
58 refcount_create(&tx
->tx_space_written
);
59 refcount_create(&tx
->tx_space_freed
);
65 dmu_tx_create(objset_t
*os
)
67 dmu_tx_t
*tx
= dmu_tx_create_dd(os
->os_dsl_dataset
->ds_dir
);
69 tx
->tx_lastsnap_txg
= dsl_dataset_prev_snap_txg(os
->os_dsl_dataset
);
74 dmu_tx_create_assigned(struct dsl_pool
*dp
, uint64_t txg
)
76 dmu_tx_t
*tx
= dmu_tx_create_dd(NULL
);
78 ASSERT3U(txg
, <=, dp
->dp_tx
.tx_open_txg
);
87 dmu_tx_is_syncing(dmu_tx_t
*tx
)
89 return (tx
->tx_anyobj
);
93 dmu_tx_private_ok(dmu_tx_t
*tx
)
95 return (tx
->tx_anyobj
);
98 static dmu_tx_hold_t
*
99 dmu_tx_hold_object_impl(dmu_tx_t
*tx
, objset_t
*os
, uint64_t object
,
100 enum dmu_tx_hold_type type
, uint64_t arg1
, uint64_t arg2
)
106 if (object
!= DMU_NEW_OBJECT
) {
107 err
= dnode_hold(os
, object
, tx
, &dn
);
113 if (err
== 0 && tx
->tx_txg
!= 0) {
114 mutex_enter(&dn
->dn_mtx
);
116 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
117 * problem, but there's no way for it to happen (for
120 ASSERT(dn
->dn_assigned_txg
== 0);
121 dn
->dn_assigned_txg
= tx
->tx_txg
;
122 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
123 mutex_exit(&dn
->dn_mtx
);
127 txh
= kmem_zalloc(sizeof (dmu_tx_hold_t
), KM_SLEEP
);
131 txh
->txh_type
= type
;
132 txh
->txh_arg1
= arg1
;
133 txh
->txh_arg2
= arg2
;
135 list_insert_tail(&tx
->tx_holds
, txh
);
141 dmu_tx_add_new_object(dmu_tx_t
*tx
, objset_t
*os
, uint64_t object
)
144 * If we're syncing, they can manipulate any object anyhow, and
145 * the hold on the dnode_t can cause problems.
147 if (!dmu_tx_is_syncing(tx
)) {
148 (void) dmu_tx_hold_object_impl(tx
, os
,
149 object
, THT_NEWOBJECT
, 0, 0);
154 dmu_tx_check_ioerr(zio_t
*zio
, dnode_t
*dn
, int level
, uint64_t blkid
)
159 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
160 db
= dbuf_hold_level(dn
, level
, blkid
, FTAG
);
161 rw_exit(&dn
->dn_struct_rwlock
);
164 err
= dbuf_read(db
, zio
, DB_RF_CANFAIL
| DB_RF_NOPREFETCH
);
170 dmu_tx_count_twig(dmu_tx_hold_t
*txh
, dnode_t
*dn
, dmu_buf_impl_t
*db
,
171 int level
, uint64_t blkid
, boolean_t freeable
, uint64_t *history
)
173 objset_t
*os
= dn
->dn_objset
;
174 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
175 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
176 dmu_buf_impl_t
*parent
= NULL
;
180 if (level
>= dn
->dn_nlevels
|| history
[level
] == blkid
)
183 history
[level
] = blkid
;
185 space
= (level
== 0) ? dn
->dn_datablksz
: (1ULL << dn
->dn_indblkshift
);
187 if (db
== NULL
|| db
== dn
->dn_dbuf
) {
191 ASSERT(DB_DNODE(db
) == dn
);
192 ASSERT(db
->db_level
== level
);
193 ASSERT(db
->db
.db_size
== space
);
194 ASSERT(db
->db_blkid
== blkid
);
196 parent
= db
->db_parent
;
199 freeable
= (bp
&& (freeable
||
200 dsl_dataset_block_freeable(ds
, bp
, bp
->blk_birth
)));
203 txh
->txh_space_tooverwrite
+= space
;
205 txh
->txh_space_towrite
+= space
;
207 txh
->txh_space_tounref
+= bp_get_dsize(os
->os_spa
, bp
);
209 dmu_tx_count_twig(txh
, dn
, parent
, level
+ 1,
210 blkid
>> epbs
, freeable
, history
);
215 dmu_tx_count_write(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
217 dnode_t
*dn
= txh
->txh_dnode
;
218 uint64_t start
, end
, i
;
219 int min_bs
, max_bs
, min_ibs
, max_ibs
, epbs
, bits
;
225 min_bs
= SPA_MINBLOCKSHIFT
;
226 max_bs
= SPA_MAXBLOCKSHIFT
;
227 min_ibs
= DN_MIN_INDBLKSHIFT
;
228 max_ibs
= DN_MAX_INDBLKSHIFT
;
231 uint64_t history
[DN_MAX_LEVELS
];
232 int nlvls
= dn
->dn_nlevels
;
236 * For i/o error checking, read the first and last level-0
237 * blocks (if they are not aligned), and all the level-1 blocks.
239 if (dn
->dn_maxblkid
== 0) {
240 delta
= dn
->dn_datablksz
;
241 start
= (off
< dn
->dn_datablksz
) ? 0 : 1;
242 end
= (off
+len
<= dn
->dn_datablksz
) ? 0 : 1;
243 if (start
== 0 && (off
> 0 || len
< dn
->dn_datablksz
)) {
244 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
250 zio_t
*zio
= zio_root(dn
->dn_objset
->os_spa
,
251 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
253 /* first level-0 block */
254 start
= off
>> dn
->dn_datablkshift
;
255 if (P2PHASE(off
, dn
->dn_datablksz
) ||
256 len
< dn
->dn_datablksz
) {
257 err
= dmu_tx_check_ioerr(zio
, dn
, 0, start
);
262 /* last level-0 block */
263 end
= (off
+len
-1) >> dn
->dn_datablkshift
;
264 if (end
!= start
&& end
<= dn
->dn_maxblkid
&&
265 P2PHASE(off
+len
, dn
->dn_datablksz
)) {
266 err
= dmu_tx_check_ioerr(zio
, dn
, 0, end
);
273 int shft
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
274 for (i
= (start
>>shft
)+1; i
< end
>>shft
; i
++) {
275 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
284 delta
= P2NPHASE(off
, dn
->dn_datablksz
);
287 min_ibs
= max_ibs
= dn
->dn_indblkshift
;
288 if (dn
->dn_maxblkid
> 0) {
290 * The blocksize can't change,
291 * so we can make a more precise estimate.
293 ASSERT(dn
->dn_datablkshift
!= 0);
294 min_bs
= max_bs
= dn
->dn_datablkshift
;
298 * If this write is not off the end of the file
299 * we need to account for overwrites/unref.
301 if (start
<= dn
->dn_maxblkid
) {
302 for (int l
= 0; l
< DN_MAX_LEVELS
; l
++)
305 while (start
<= dn
->dn_maxblkid
) {
308 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
309 err
= dbuf_hold_impl(dn
, 0, start
, FALSE
, FTAG
, &db
);
310 rw_exit(&dn
->dn_struct_rwlock
);
313 txh
->txh_tx
->tx_err
= err
;
317 dmu_tx_count_twig(txh
, dn
, db
, 0, start
, B_FALSE
,
322 * Account for new indirects appearing
323 * before this IO gets assigned into a txg.
326 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
327 for (bits
-= epbs
* (nlvls
- 1);
328 bits
>= 0; bits
-= epbs
)
329 txh
->txh_fudge
+= 1ULL << max_ibs
;
335 delta
= dn
->dn_datablksz
;
340 * 'end' is the last thing we will access, not one past.
341 * This way we won't overflow when accessing the last byte.
343 start
= P2ALIGN(off
, 1ULL << max_bs
);
344 end
= P2ROUNDUP(off
+ len
, 1ULL << max_bs
) - 1;
345 txh
->txh_space_towrite
+= end
- start
+ 1;
350 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
353 * The object contains at most 2^(64 - min_bs) blocks,
354 * and each indirect level maps 2^epbs.
356 for (bits
= 64 - min_bs
; bits
>= 0; bits
-= epbs
) {
359 ASSERT3U(end
, >=, start
);
360 txh
->txh_space_towrite
+= (end
- start
+ 1) << max_ibs
;
363 * We also need a new blkid=0 indirect block
364 * to reference any existing file data.
366 txh
->txh_space_towrite
+= 1ULL << max_ibs
;
371 if (txh
->txh_space_towrite
+ txh
->txh_space_tooverwrite
>
376 txh
->txh_tx
->tx_err
= err
;
380 dmu_tx_count_dnode(dmu_tx_hold_t
*txh
)
382 dnode_t
*dn
= txh
->txh_dnode
;
383 dnode_t
*mdn
= DMU_META_DNODE(txh
->txh_tx
->tx_objset
);
384 uint64_t space
= mdn
->dn_datablksz
+
385 ((mdn
->dn_nlevels
-1) << mdn
->dn_indblkshift
);
387 if (dn
&& dn
->dn_dbuf
->db_blkptr
&&
388 dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
389 dn
->dn_dbuf
->db_blkptr
, dn
->dn_dbuf
->db_blkptr
->blk_birth
)) {
390 txh
->txh_space_tooverwrite
+= space
;
391 txh
->txh_space_tounref
+= space
;
393 txh
->txh_space_towrite
+= space
;
394 if (dn
&& dn
->dn_dbuf
->db_blkptr
)
395 txh
->txh_space_tounref
+= space
;
400 dmu_tx_hold_write(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, int len
)
404 ASSERT(tx
->tx_txg
== 0);
405 ASSERT(len
< DMU_MAX_ACCESS
);
406 ASSERT(len
== 0 || UINT64_MAX
- off
>= len
- 1);
408 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
409 object
, THT_WRITE
, off
, len
);
413 dmu_tx_count_write(txh
, off
, len
);
414 dmu_tx_count_dnode(txh
);
418 dmu_tx_count_free(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
420 uint64_t blkid
, nblks
, lastblk
;
421 uint64_t space
= 0, unref
= 0, skipped
= 0;
422 dnode_t
*dn
= txh
->txh_dnode
;
423 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
424 spa_t
*spa
= txh
->txh_tx
->tx_pool
->dp_spa
;
426 uint64_t l0span
= 0, nl1blks
= 0;
428 if (dn
->dn_nlevels
== 0)
432 * The struct_rwlock protects us against dn_nlevels
433 * changing, in case (against all odds) we manage to dirty &
434 * sync out the changes after we check for being dirty.
435 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
437 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
438 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
439 if (dn
->dn_maxblkid
== 0) {
440 if (off
== 0 && len
>= dn
->dn_datablksz
) {
444 rw_exit(&dn
->dn_struct_rwlock
);
448 blkid
= off
>> dn
->dn_datablkshift
;
449 nblks
= (len
+ dn
->dn_datablksz
- 1) >> dn
->dn_datablkshift
;
451 if (blkid
>= dn
->dn_maxblkid
) {
452 rw_exit(&dn
->dn_struct_rwlock
);
455 if (blkid
+ nblks
> dn
->dn_maxblkid
)
456 nblks
= dn
->dn_maxblkid
- blkid
;
459 l0span
= nblks
; /* save for later use to calc level > 1 overhead */
460 if (dn
->dn_nlevels
== 1) {
462 for (i
= 0; i
< nblks
; i
++) {
463 blkptr_t
*bp
= dn
->dn_phys
->dn_blkptr
;
464 ASSERT3U(blkid
+ i
, <, dn
->dn_nblkptr
);
466 if (dsl_dataset_block_freeable(ds
, bp
, bp
->blk_birth
)) {
467 dprintf_bp(bp
, "can free old%s", "");
468 space
+= bp_get_dsize(spa
, bp
);
470 unref
+= BP_GET_ASIZE(bp
);
476 lastblk
= blkid
+ nblks
- 1;
478 dmu_buf_impl_t
*dbuf
;
479 uint64_t ibyte
, new_blkid
;
481 int err
, i
, blkoff
, tochk
;
484 ibyte
= blkid
<< dn
->dn_datablkshift
;
485 err
= dnode_next_offset(dn
,
486 DNODE_FIND_HAVELOCK
, &ibyte
, 2, 1, 0);
487 new_blkid
= ibyte
>> dn
->dn_datablkshift
;
489 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
493 txh
->txh_tx
->tx_err
= err
;
496 if (new_blkid
> lastblk
) {
497 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
501 if (new_blkid
> blkid
) {
502 ASSERT((new_blkid
>> epbs
) > (blkid
>> epbs
));
503 skipped
+= (new_blkid
>> epbs
) - (blkid
>> epbs
) - 1;
504 nblks
-= new_blkid
- blkid
;
507 blkoff
= P2PHASE(blkid
, epb
);
508 tochk
= MIN(epb
- blkoff
, nblks
);
510 err
= dbuf_hold_impl(dn
, 1, blkid
>> epbs
, FALSE
, FTAG
, &dbuf
);
512 txh
->txh_tx
->tx_err
= err
;
516 txh
->txh_memory_tohold
+= dbuf
->db
.db_size
;
519 * We don't check memory_tohold against DMU_MAX_ACCESS because
520 * memory_tohold is an over-estimation (especially the >L1
521 * indirect blocks), so it could fail. Callers should have
522 * already verified that they will not be holding too much
526 err
= dbuf_read(dbuf
, NULL
, DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
);
528 txh
->txh_tx
->tx_err
= err
;
529 dbuf_rele(dbuf
, FTAG
);
533 bp
= dbuf
->db
.db_data
;
536 for (i
= 0; i
< tochk
; i
++) {
537 if (dsl_dataset_block_freeable(ds
, &bp
[i
],
539 dprintf_bp(&bp
[i
], "can free old%s", "");
540 space
+= bp_get_dsize(spa
, &bp
[i
]);
542 unref
+= BP_GET_ASIZE(bp
);
544 dbuf_rele(dbuf
, FTAG
);
550 rw_exit(&dn
->dn_struct_rwlock
);
553 * Add in memory requirements of higher-level indirects.
554 * This assumes a worst-possible scenario for dn_nlevels and a
555 * worst-possible distribution of l1-blocks over the region to free.
558 uint64_t blkcnt
= 1 + ((l0span
>> epbs
) >> epbs
);
561 * Here we don't use DN_MAX_LEVEL, but calculate it with the
562 * given datablkshift and indblkshift. This makes the
563 * difference between 19 and 8 on large files.
565 int maxlevel
= 2 + (DN_MAX_OFFSET_SHIFT
- dn
->dn_datablkshift
) /
566 (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
568 while (level
++ < maxlevel
) {
569 txh
->txh_memory_tohold
+= MAX(MIN(blkcnt
, nl1blks
), 1)
570 << dn
->dn_indblkshift
;
571 blkcnt
= 1 + (blkcnt
>> epbs
);
575 /* account for new level 1 indirect blocks that might show up */
577 txh
->txh_fudge
+= skipped
<< dn
->dn_indblkshift
;
578 skipped
= MIN(skipped
, DMU_MAX_DELETEBLKCNT
>> epbs
);
579 txh
->txh_memory_tohold
+= skipped
<< dn
->dn_indblkshift
;
581 txh
->txh_space_tofree
+= space
;
582 txh
->txh_space_tounref
+= unref
;
586 dmu_tx_hold_free(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, uint64_t len
)
590 uint64_t start
, end
, i
;
594 ASSERT(tx
->tx_txg
== 0);
596 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
597 object
, THT_FREE
, off
, len
);
604 dmu_tx_count_write(txh
, off
, 1);
606 if (len
!= DMU_OBJECT_END
)
607 dmu_tx_count_write(txh
, off
+len
, 1);
609 dmu_tx_count_dnode(txh
);
611 if (off
>= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
)
613 if (len
== DMU_OBJECT_END
)
614 len
= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
- off
;
617 * For i/o error checking, read the first and last level-0
618 * blocks, and all the level-1 blocks. The above count_write's
619 * have already taken care of the level-0 blocks.
621 if (dn
->dn_nlevels
> 1) {
622 shift
= dn
->dn_datablkshift
+ dn
->dn_indblkshift
-
624 start
= off
>> shift
;
625 end
= dn
->dn_datablkshift
? ((off
+len
) >> shift
) : 0;
627 zio
= zio_root(tx
->tx_pool
->dp_spa
,
628 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
629 for (i
= start
; i
<= end
; i
++) {
630 uint64_t ibyte
= i
<< shift
;
631 err
= dnode_next_offset(dn
, 0, &ibyte
, 2, 1, 0);
640 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
653 dmu_tx_count_free(txh
, off
, len
);
657 dmu_tx_hold_zap(dmu_tx_t
*tx
, uint64_t object
, int add
, const char *name
)
664 ASSERT(tx
->tx_txg
== 0);
666 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
667 object
, THT_ZAP
, add
, (uintptr_t)name
);
672 dmu_tx_count_dnode(txh
);
676 * We will be able to fit a new object's entries into one leaf
677 * block. So there will be at most 2 blocks total,
678 * including the header block.
680 dmu_tx_count_write(txh
, 0, 2 << fzap_default_block_shift
);
684 ASSERT3P(DMU_OT_BYTESWAP(dn
->dn_type
), ==, DMU_BSWAP_ZAP
);
686 if (dn
->dn_maxblkid
== 0 && !add
) {
690 * If there is only one block (i.e. this is a micro-zap)
691 * and we are not adding anything, the accounting is simple.
693 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
700 * Use max block size here, since we don't know how much
701 * the size will change between now and the dbuf dirty call.
703 bp
= &dn
->dn_phys
->dn_blkptr
[0];
704 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
706 txh
->txh_space_tooverwrite
+= SPA_MAXBLOCKSIZE
;
708 txh
->txh_space_towrite
+= SPA_MAXBLOCKSIZE
;
710 txh
->txh_space_tounref
+= SPA_MAXBLOCKSIZE
;
714 if (dn
->dn_maxblkid
> 0 && name
) {
716 * access the name in this fat-zap so that we'll check
717 * for i/o errors to the leaf blocks, etc.
719 err
= zap_lookup(dn
->dn_objset
, dn
->dn_object
, name
,
727 err
= zap_count_write(dn
->dn_objset
, dn
->dn_object
, name
, add
,
728 &txh
->txh_space_towrite
, &txh
->txh_space_tooverwrite
);
731 * If the modified blocks are scattered to the four winds,
732 * we'll have to modify an indirect twig for each.
734 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
735 for (nblocks
= dn
->dn_maxblkid
>> epbs
; nblocks
!= 0; nblocks
>>= epbs
)
736 if (dn
->dn_objset
->os_dsl_dataset
->ds_phys
->ds_prev_snap_obj
)
737 txh
->txh_space_towrite
+= 3 << dn
->dn_indblkshift
;
739 txh
->txh_space_tooverwrite
+= 3 << dn
->dn_indblkshift
;
743 dmu_tx_hold_bonus(dmu_tx_t
*tx
, uint64_t object
)
747 ASSERT(tx
->tx_txg
== 0);
749 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
750 object
, THT_BONUS
, 0, 0);
752 dmu_tx_count_dnode(txh
);
756 dmu_tx_hold_space(dmu_tx_t
*tx
, uint64_t space
)
759 ASSERT(tx
->tx_txg
== 0);
761 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
762 DMU_NEW_OBJECT
, THT_SPACE
, space
, 0);
764 txh
->txh_space_towrite
+= space
;
768 dmu_tx_holds(dmu_tx_t
*tx
, uint64_t object
)
774 * By asserting that the tx is assigned, we're counting the
775 * number of dn_tx_holds, which is the same as the number of
776 * dn_holds. Otherwise, we'd be counting dn_holds, but
777 * dn_tx_holds could be 0.
779 ASSERT(tx
->tx_txg
!= 0);
781 /* if (tx->tx_anyobj == TRUE) */
784 for (txh
= list_head(&tx
->tx_holds
); txh
;
785 txh
= list_next(&tx
->tx_holds
, txh
)) {
786 if (txh
->txh_dnode
&& txh
->txh_dnode
->dn_object
== object
)
795 dmu_tx_dirty_buf(dmu_tx_t
*tx
, dmu_buf_impl_t
*db
)
798 int match_object
= FALSE
, match_offset
= FALSE
;
803 ASSERT(tx
->tx_txg
!= 0);
804 ASSERT(tx
->tx_objset
== NULL
|| dn
->dn_objset
== tx
->tx_objset
);
805 ASSERT3U(dn
->dn_object
, ==, db
->db
.db_object
);
812 /* XXX No checking on the meta dnode for now */
813 if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
818 for (txh
= list_head(&tx
->tx_holds
); txh
;
819 txh
= list_next(&tx
->tx_holds
, txh
)) {
820 ASSERT(dn
== NULL
|| dn
->dn_assigned_txg
== tx
->tx_txg
);
821 if (txh
->txh_dnode
== dn
&& txh
->txh_type
!= THT_NEWOBJECT
)
823 if (txh
->txh_dnode
== NULL
|| txh
->txh_dnode
== dn
) {
824 int datablkshift
= dn
->dn_datablkshift
?
825 dn
->dn_datablkshift
: SPA_MAXBLOCKSHIFT
;
826 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
827 int shift
= datablkshift
+ epbs
* db
->db_level
;
828 uint64_t beginblk
= shift
>= 64 ? 0 :
829 (txh
->txh_arg1
>> shift
);
830 uint64_t endblk
= shift
>= 64 ? 0 :
831 ((txh
->txh_arg1
+ txh
->txh_arg2
- 1) >> shift
);
832 uint64_t blkid
= db
->db_blkid
;
834 /* XXX txh_arg2 better not be zero... */
836 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
837 txh
->txh_type
, beginblk
, endblk
);
839 switch (txh
->txh_type
) {
841 if (blkid
>= beginblk
&& blkid
<= endblk
)
844 * We will let this hold work for the bonus
845 * or spill buffer so that we don't need to
846 * hold it when creating a new object.
848 if (blkid
== DMU_BONUS_BLKID
||
849 blkid
== DMU_SPILL_BLKID
)
852 * They might have to increase nlevels,
853 * thus dirtying the new TLIBs. Or the
854 * might have to change the block size,
855 * thus dirying the new lvl=0 blk=0.
862 * We will dirty all the level 1 blocks in
863 * the free range and perhaps the first and
864 * last level 0 block.
866 if (blkid
>= beginblk
&& (blkid
<= endblk
||
867 txh
->txh_arg2
== DMU_OBJECT_END
))
871 if (blkid
== DMU_SPILL_BLKID
)
875 if (blkid
== DMU_BONUS_BLKID
)
885 ASSERT(!"bad txh_type");
888 if (match_object
&& match_offset
) {
894 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
895 (u_longlong_t
)db
->db
.db_object
, db
->db_level
,
896 (u_longlong_t
)db
->db_blkid
);
901 dmu_tx_try_assign(dmu_tx_t
*tx
, uint64_t txg_how
)
904 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
905 uint64_t memory
, asize
, fsize
, usize
;
906 uint64_t towrite
, tofree
, tooverwrite
, tounref
, tohold
, fudge
;
913 if (spa_suspended(spa
)) {
915 * If the user has indicated a blocking failure mode
916 * then return ERESTART which will block in dmu_tx_wait().
917 * Otherwise, return EIO so that an error can get
918 * propagated back to the VOP calls.
920 * Note that we always honor the txg_how flag regardless
921 * of the failuremode setting.
923 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_CONTINUE
&&
930 tx
->tx_txg
= txg_hold_open(tx
->tx_pool
, &tx
->tx_txgh
);
931 tx
->tx_needassign_txh
= NULL
;
934 * NB: No error returns are allowed after txg_hold_open, but
935 * before processing the dnode holds, due to the
936 * dmu_tx_unassign() logic.
939 towrite
= tofree
= tooverwrite
= tounref
= tohold
= fudge
= 0;
940 for (txh
= list_head(&tx
->tx_holds
); txh
;
941 txh
= list_next(&tx
->tx_holds
, txh
)) {
942 dnode_t
*dn
= txh
->txh_dnode
;
944 mutex_enter(&dn
->dn_mtx
);
945 if (dn
->dn_assigned_txg
== tx
->tx_txg
- 1) {
946 mutex_exit(&dn
->dn_mtx
);
947 tx
->tx_needassign_txh
= txh
;
950 if (dn
->dn_assigned_txg
== 0)
951 dn
->dn_assigned_txg
= tx
->tx_txg
;
952 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
953 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
954 mutex_exit(&dn
->dn_mtx
);
956 towrite
+= txh
->txh_space_towrite
;
957 tofree
+= txh
->txh_space_tofree
;
958 tooverwrite
+= txh
->txh_space_tooverwrite
;
959 tounref
+= txh
->txh_space_tounref
;
960 tohold
+= txh
->txh_memory_tohold
;
961 fudge
+= txh
->txh_fudge
;
965 * NB: This check must be after we've held the dnodes, so that
966 * the dmu_tx_unassign() logic will work properly
968 if (txg_how
>= TXG_INITIAL
&& txg_how
!= tx
->tx_txg
)
972 * If a snapshot has been taken since we made our estimates,
973 * assume that we won't be able to free or overwrite anything.
976 dsl_dataset_prev_snap_txg(tx
->tx_objset
->os_dsl_dataset
) >
977 tx
->tx_lastsnap_txg
) {
978 towrite
+= tooverwrite
;
979 tooverwrite
= tofree
= 0;
982 /* needed allocation: worst-case estimate of write space */
983 asize
= spa_get_asize(tx
->tx_pool
->dp_spa
, towrite
+ tooverwrite
);
984 /* freed space estimate: worst-case overwrite + free estimate */
985 fsize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tooverwrite
) + tofree
;
986 /* convert unrefd space to worst-case estimate */
987 usize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tounref
);
988 /* calculate memory footprint estimate */
989 memory
= towrite
+ tooverwrite
+ tohold
;
993 * Add in 'tohold' to account for our dirty holds on this memory
994 * XXX - the "fudge" factor is to account for skipped blocks that
995 * we missed because dnode_next_offset() misses in-core-only blocks.
997 tx
->tx_space_towrite
= asize
+
998 spa_get_asize(tx
->tx_pool
->dp_spa
, tohold
+ fudge
);
999 tx
->tx_space_tofree
= tofree
;
1000 tx
->tx_space_tooverwrite
= tooverwrite
;
1001 tx
->tx_space_tounref
= tounref
;
1004 if (tx
->tx_dir
&& asize
!= 0) {
1005 int err
= dsl_dir_tempreserve_space(tx
->tx_dir
, memory
,
1006 asize
, fsize
, usize
, &tx
->tx_tempreserve_cookie
, tx
);
1015 dmu_tx_unassign(dmu_tx_t
*tx
)
1019 if (tx
->tx_txg
== 0)
1022 txg_rele_to_quiesce(&tx
->tx_txgh
);
1024 for (txh
= list_head(&tx
->tx_holds
); txh
!= tx
->tx_needassign_txh
;
1025 txh
= list_next(&tx
->tx_holds
, txh
)) {
1026 dnode_t
*dn
= txh
->txh_dnode
;
1030 mutex_enter(&dn
->dn_mtx
);
1031 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1033 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1034 dn
->dn_assigned_txg
= 0;
1035 cv_broadcast(&dn
->dn_notxholds
);
1037 mutex_exit(&dn
->dn_mtx
);
1040 txg_rele_to_sync(&tx
->tx_txgh
);
1042 tx
->tx_lasttried_txg
= tx
->tx_txg
;
1047 * Assign tx to a transaction group. txg_how can be one of:
1049 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1050 * a new one. This should be used when you're not holding locks.
1051 * If will only fail if we're truly out of space (or over quota).
1053 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1054 * blocking, returns immediately with ERESTART. This should be used
1055 * whenever you're holding locks. On an ERESTART error, the caller
1056 * should drop locks, do a dmu_tx_wait(tx), and try again.
1058 * (3) A specific txg. Use this if you need to ensure that multiple
1059 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1060 * returns ERESTART if it can't assign you into the requested txg.
1063 dmu_tx_assign(dmu_tx_t
*tx
, uint64_t txg_how
)
1067 ASSERT(tx
->tx_txg
== 0);
1068 ASSERT(txg_how
!= 0);
1069 ASSERT(!dsl_pool_sync_context(tx
->tx_pool
));
1071 while ((err
= dmu_tx_try_assign(tx
, txg_how
)) != 0) {
1072 dmu_tx_unassign(tx
);
1074 if (err
!= ERESTART
|| txg_how
!= TXG_WAIT
)
1080 txg_rele_to_quiesce(&tx
->tx_txgh
);
1086 dmu_tx_wait(dmu_tx_t
*tx
)
1088 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
1090 ASSERT(tx
->tx_txg
== 0);
1093 * It's possible that the pool has become active after this thread
1094 * has tried to obtain a tx. If that's the case then his
1095 * tx_lasttried_txg would not have been assigned.
1097 if (spa_suspended(spa
) || tx
->tx_lasttried_txg
== 0) {
1098 txg_wait_synced(tx
->tx_pool
, spa_last_synced_txg(spa
) + 1);
1099 } else if (tx
->tx_needassign_txh
) {
1100 dnode_t
*dn
= tx
->tx_needassign_txh
->txh_dnode
;
1102 mutex_enter(&dn
->dn_mtx
);
1103 while (dn
->dn_assigned_txg
== tx
->tx_lasttried_txg
- 1)
1104 cv_wait(&dn
->dn_notxholds
, &dn
->dn_mtx
);
1105 mutex_exit(&dn
->dn_mtx
);
1106 tx
->tx_needassign_txh
= NULL
;
1108 txg_wait_open(tx
->tx_pool
, tx
->tx_lasttried_txg
+ 1);
1113 dmu_tx_willuse_space(dmu_tx_t
*tx
, int64_t delta
)
1116 if (tx
->tx_dir
== NULL
|| delta
== 0)
1120 ASSERT3U(refcount_count(&tx
->tx_space_written
) + delta
, <=,
1121 tx
->tx_space_towrite
);
1122 (void) refcount_add_many(&tx
->tx_space_written
, delta
, NULL
);
1124 (void) refcount_add_many(&tx
->tx_space_freed
, -delta
, NULL
);
1130 dmu_tx_commit(dmu_tx_t
*tx
)
1134 ASSERT(tx
->tx_txg
!= 0);
1136 while (txh
= list_head(&tx
->tx_holds
)) {
1137 dnode_t
*dn
= txh
->txh_dnode
;
1139 list_remove(&tx
->tx_holds
, txh
);
1140 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1143 mutex_enter(&dn
->dn_mtx
);
1144 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1146 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1147 dn
->dn_assigned_txg
= 0;
1148 cv_broadcast(&dn
->dn_notxholds
);
1150 mutex_exit(&dn
->dn_mtx
);
1154 if (tx
->tx_tempreserve_cookie
)
1155 dsl_dir_tempreserve_clear(tx
->tx_tempreserve_cookie
, tx
);
1157 if (!list_is_empty(&tx
->tx_callbacks
))
1158 txg_register_callbacks(&tx
->tx_txgh
, &tx
->tx_callbacks
);
1160 if (tx
->tx_anyobj
== FALSE
)
1161 txg_rele_to_sync(&tx
->tx_txgh
);
1163 list_destroy(&tx
->tx_callbacks
);
1164 list_destroy(&tx
->tx_holds
);
1166 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1167 tx
->tx_space_towrite
, refcount_count(&tx
->tx_space_written
),
1168 tx
->tx_space_tofree
, refcount_count(&tx
->tx_space_freed
));
1169 refcount_destroy_many(&tx
->tx_space_written
,
1170 refcount_count(&tx
->tx_space_written
));
1171 refcount_destroy_many(&tx
->tx_space_freed
,
1172 refcount_count(&tx
->tx_space_freed
));
1174 kmem_free(tx
, sizeof (dmu_tx_t
));
1178 dmu_tx_abort(dmu_tx_t
*tx
)
1182 ASSERT(tx
->tx_txg
== 0);
1184 while (txh
= list_head(&tx
->tx_holds
)) {
1185 dnode_t
*dn
= txh
->txh_dnode
;
1187 list_remove(&tx
->tx_holds
, txh
);
1188 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1194 * Call any registered callbacks with an error code.
1196 if (!list_is_empty(&tx
->tx_callbacks
))
1197 dmu_tx_do_callbacks(&tx
->tx_callbacks
, ECANCELED
);
1199 list_destroy(&tx
->tx_callbacks
);
1200 list_destroy(&tx
->tx_holds
);
1202 refcount_destroy_many(&tx
->tx_space_written
,
1203 refcount_count(&tx
->tx_space_written
));
1204 refcount_destroy_many(&tx
->tx_space_freed
,
1205 refcount_count(&tx
->tx_space_freed
));
1207 kmem_free(tx
, sizeof (dmu_tx_t
));
1211 dmu_tx_get_txg(dmu_tx_t
*tx
)
1213 ASSERT(tx
->tx_txg
!= 0);
1214 return (tx
->tx_txg
);
1218 dmu_tx_callback_register(dmu_tx_t
*tx
, dmu_tx_callback_func_t
*func
, void *data
)
1220 dmu_tx_callback_t
*dcb
;
1222 dcb
= kmem_alloc(sizeof (dmu_tx_callback_t
), KM_SLEEP
);
1224 dcb
->dcb_func
= func
;
1225 dcb
->dcb_data
= data
;
1227 list_insert_tail(&tx
->tx_callbacks
, dcb
);
1231 * Call all the commit callbacks on a list, with a given error code.
1234 dmu_tx_do_callbacks(list_t
*cb_list
, int error
)
1236 dmu_tx_callback_t
*dcb
;
1238 while (dcb
= list_head(cb_list
)) {
1239 list_remove(cb_list
, dcb
);
1240 dcb
->dcb_func(dcb
->dcb_data
, error
);
1241 kmem_free(dcb
, sizeof (dmu_tx_callback_t
));
1246 * Interface to hold a bunch of attributes.
1247 * used for creating new files.
1248 * attrsize is the total size of all attributes
1249 * to be added during object creation
1251 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1255 * hold necessary attribute name for attribute registration.
1256 * should be a very rare case where this is needed. If it does
1257 * happen it would only happen on the first write to the file system.
1260 dmu_tx_sa_registration_hold(sa_os_t
*sa
, dmu_tx_t
*tx
)
1264 if (!sa
->sa_need_attr_registration
)
1267 for (i
= 0; i
!= sa
->sa_num_attrs
; i
++) {
1268 if (!sa
->sa_attr_table
[i
].sa_registered
) {
1269 if (sa
->sa_reg_attr_obj
)
1270 dmu_tx_hold_zap(tx
, sa
->sa_reg_attr_obj
,
1271 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1273 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
,
1274 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1281 dmu_tx_hold_spill(dmu_tx_t
*tx
, uint64_t object
)
1286 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, object
,
1289 dn
= txh
->txh_dnode
;
1294 /* If blkptr doesn't exist then add space to towrite */
1295 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
1296 txh
->txh_space_towrite
+= SPA_MAXBLOCKSIZE
;
1300 bp
= &dn
->dn_phys
->dn_spill
;
1301 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
1303 txh
->txh_space_tooverwrite
+= SPA_MAXBLOCKSIZE
;
1305 txh
->txh_space_towrite
+= SPA_MAXBLOCKSIZE
;
1306 if (!BP_IS_HOLE(bp
))
1307 txh
->txh_space_tounref
+= SPA_MAXBLOCKSIZE
;
1312 dmu_tx_hold_sa_create(dmu_tx_t
*tx
, int attrsize
)
1314 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1316 dmu_tx_hold_bonus(tx
, DMU_NEW_OBJECT
);
1318 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1321 if (tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1322 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1324 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1325 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1326 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1327 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1330 dmu_tx_sa_registration_hold(sa
, tx
);
1332 if (attrsize
<= DN_MAX_BONUSLEN
&& !sa
->sa_force_spill
)
1335 (void) dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, DMU_NEW_OBJECT
,
1342 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1344 * variable_size is the total size of all variable sized attributes
1345 * passed to this function. It is not the total size of all
1346 * variable size attributes that *may* exist on this object.
1349 dmu_tx_hold_sa(dmu_tx_t
*tx
, sa_handle_t
*hdl
, boolean_t may_grow
)
1352 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1354 ASSERT(hdl
!= NULL
);
1356 object
= sa_handle_object(hdl
);
1358 dmu_tx_hold_bonus(tx
, object
);
1360 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1363 if (tx
->tx_objset
->os_sa
->sa_reg_attr_obj
== 0 ||
1364 tx
->tx_objset
->os_sa
->sa_layout_attr_obj
== 0) {
1365 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1366 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1367 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1368 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1371 dmu_tx_sa_registration_hold(sa
, tx
);
1373 if (may_grow
&& tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1374 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1376 if (sa
->sa_force_spill
|| may_grow
|| hdl
->sa_spill
) {
1377 ASSERT(tx
->tx_txg
== 0);
1378 dmu_tx_hold_spill(tx
, object
);
1380 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)hdl
->sa_bonus
;
1385 if (dn
->dn_have_spill
) {
1386 ASSERT(tx
->tx_txg
== 0);
1387 dmu_tx_hold_spill(tx
, object
);