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.
26 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dmu_objset.h>
30 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
31 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
32 #include <sys/dsl_pool.h>
33 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/sa_impl.h>
37 #include <sys/zfs_context.h>
38 #include <sys/varargs.h>
40 typedef void (*dmu_tx_hold_func_t
)(dmu_tx_t
*tx
, struct dnode
*dn
,
41 uint64_t arg1
, uint64_t arg2
);
43 dmu_tx_stats_t dmu_tx_stats
= {
44 { "dmu_tx_assigned", KSTAT_DATA_UINT64
},
45 { "dmu_tx_delay", KSTAT_DATA_UINT64
},
46 { "dmu_tx_error", KSTAT_DATA_UINT64
},
47 { "dmu_tx_suspended", KSTAT_DATA_UINT64
},
48 { "dmu_tx_group", KSTAT_DATA_UINT64
},
49 { "dmu_tx_how", KSTAT_DATA_UINT64
},
50 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64
},
51 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64
},
52 { "dmu_tx_memory_inflight", KSTAT_DATA_UINT64
},
53 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64
},
54 { "dmu_tx_write_limit", KSTAT_DATA_UINT64
},
55 { "dmu_tx_quota", KSTAT_DATA_UINT64
},
58 static kstat_t
*dmu_tx_ksp
;
61 dmu_tx_create_dd(dsl_dir_t
*dd
)
63 dmu_tx_t
*tx
= kmem_zalloc(sizeof (dmu_tx_t
), KM_SLEEP
);
66 tx
->tx_pool
= dd
->dd_pool
;
67 list_create(&tx
->tx_holds
, sizeof (dmu_tx_hold_t
),
68 offsetof(dmu_tx_hold_t
, txh_node
));
69 list_create(&tx
->tx_callbacks
, sizeof (dmu_tx_callback_t
),
70 offsetof(dmu_tx_callback_t
, dcb_node
));
72 refcount_create(&tx
->tx_space_written
);
73 refcount_create(&tx
->tx_space_freed
);
79 dmu_tx_create(objset_t
*os
)
81 dmu_tx_t
*tx
= dmu_tx_create_dd(os
->os_dsl_dataset
->ds_dir
);
83 tx
->tx_lastsnap_txg
= dsl_dataset_prev_snap_txg(os
->os_dsl_dataset
);
88 dmu_tx_create_assigned(struct dsl_pool
*dp
, uint64_t txg
)
90 dmu_tx_t
*tx
= dmu_tx_create_dd(NULL
);
92 ASSERT3U(txg
, <=, dp
->dp_tx
.tx_open_txg
);
101 dmu_tx_is_syncing(dmu_tx_t
*tx
)
103 return (tx
->tx_anyobj
);
107 dmu_tx_private_ok(dmu_tx_t
*tx
)
109 return (tx
->tx_anyobj
);
112 static dmu_tx_hold_t
*
113 dmu_tx_hold_object_impl(dmu_tx_t
*tx
, objset_t
*os
, uint64_t object
,
114 enum dmu_tx_hold_type type
, uint64_t arg1
, uint64_t arg2
)
120 if (object
!= DMU_NEW_OBJECT
) {
121 err
= dnode_hold(os
, object
, tx
, &dn
);
127 if (err
== 0 && tx
->tx_txg
!= 0) {
128 mutex_enter(&dn
->dn_mtx
);
130 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
131 * problem, but there's no way for it to happen (for
134 ASSERT(dn
->dn_assigned_txg
== 0);
135 dn
->dn_assigned_txg
= tx
->tx_txg
;
136 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
137 mutex_exit(&dn
->dn_mtx
);
141 txh
= kmem_zalloc(sizeof (dmu_tx_hold_t
), KM_SLEEP
);
145 txh
->txh_type
= type
;
146 txh
->txh_arg1
= arg1
;
147 txh
->txh_arg2
= arg2
;
149 list_insert_tail(&tx
->tx_holds
, txh
);
155 dmu_tx_add_new_object(dmu_tx_t
*tx
, objset_t
*os
, uint64_t object
)
158 * If we're syncing, they can manipulate any object anyhow, and
159 * the hold on the dnode_t can cause problems.
161 if (!dmu_tx_is_syncing(tx
)) {
162 (void) dmu_tx_hold_object_impl(tx
, os
,
163 object
, THT_NEWOBJECT
, 0, 0);
168 dmu_tx_check_ioerr(zio_t
*zio
, dnode_t
*dn
, int level
, uint64_t blkid
)
173 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
174 db
= dbuf_hold_level(dn
, level
, blkid
, FTAG
);
175 rw_exit(&dn
->dn_struct_rwlock
);
178 err
= dbuf_read(db
, zio
, DB_RF_CANFAIL
| DB_RF_NOPREFETCH
);
184 dmu_tx_count_twig(dmu_tx_hold_t
*txh
, dnode_t
*dn
, dmu_buf_impl_t
*db
,
185 int level
, uint64_t blkid
, boolean_t freeable
, uint64_t *history
)
187 objset_t
*os
= dn
->dn_objset
;
188 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
189 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
190 dmu_buf_impl_t
*parent
= NULL
;
194 if (level
>= dn
->dn_nlevels
|| history
[level
] == blkid
)
197 history
[level
] = blkid
;
199 space
= (level
== 0) ? dn
->dn_datablksz
: (1ULL << dn
->dn_indblkshift
);
201 if (db
== NULL
|| db
== dn
->dn_dbuf
) {
205 ASSERT(DB_DNODE(db
) == dn
);
206 ASSERT(db
->db_level
== level
);
207 ASSERT(db
->db
.db_size
== space
);
208 ASSERT(db
->db_blkid
== blkid
);
210 parent
= db
->db_parent
;
213 freeable
= (bp
&& (freeable
||
214 dsl_dataset_block_freeable(ds
, bp
, bp
->blk_birth
)));
217 txh
->txh_space_tooverwrite
+= space
;
219 txh
->txh_space_towrite
+= space
;
221 txh
->txh_space_tounref
+= bp_get_dsize(os
->os_spa
, bp
);
223 dmu_tx_count_twig(txh
, dn
, parent
, level
+ 1,
224 blkid
>> epbs
, freeable
, history
);
229 dmu_tx_count_write(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
231 dnode_t
*dn
= txh
->txh_dnode
;
232 uint64_t start
, end
, i
;
233 int min_bs
, max_bs
, min_ibs
, max_ibs
, epbs
, bits
;
240 min_bs
= SPA_MINBLOCKSHIFT
;
241 max_bs
= SPA_MAXBLOCKSHIFT
;
242 min_ibs
= DN_MIN_INDBLKSHIFT
;
243 max_ibs
= DN_MAX_INDBLKSHIFT
;
246 uint64_t history
[DN_MAX_LEVELS
];
247 int nlvls
= dn
->dn_nlevels
;
251 * For i/o error checking, read the first and last level-0
252 * blocks (if they are not aligned), and all the level-1 blocks.
254 if (dn
->dn_maxblkid
== 0) {
255 delta
= dn
->dn_datablksz
;
256 start
= (off
< dn
->dn_datablksz
) ? 0 : 1;
257 end
= (off
+len
<= dn
->dn_datablksz
) ? 0 : 1;
258 if (start
== 0 && (off
> 0 || len
< dn
->dn_datablksz
)) {
259 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
265 zio_t
*zio
= zio_root(dn
->dn_objset
->os_spa
,
266 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
268 /* first level-0 block */
269 start
= off
>> dn
->dn_datablkshift
;
270 if (P2PHASE(off
, dn
->dn_datablksz
) ||
271 len
< dn
->dn_datablksz
) {
272 err
= dmu_tx_check_ioerr(zio
, dn
, 0, start
);
277 /* last level-0 block */
278 end
= (off
+len
-1) >> dn
->dn_datablkshift
;
279 if (end
!= start
&& end
<= dn
->dn_maxblkid
&&
280 P2PHASE(off
+len
, dn
->dn_datablksz
)) {
281 err
= dmu_tx_check_ioerr(zio
, dn
, 0, end
);
288 int shft
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
289 for (i
= (start
>>shft
)+1; i
< end
>>shft
; i
++) {
290 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
299 delta
= P2NPHASE(off
, dn
->dn_datablksz
);
302 if (dn
->dn_maxblkid
> 0) {
304 * The blocksize can't change,
305 * so we can make a more precise estimate.
307 ASSERT(dn
->dn_datablkshift
!= 0);
308 min_bs
= max_bs
= dn
->dn_datablkshift
;
309 min_ibs
= max_ibs
= dn
->dn_indblkshift
;
310 } else if (dn
->dn_indblkshift
> max_ibs
) {
312 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
313 * the code will still work correctly on older pools.
315 min_ibs
= max_ibs
= dn
->dn_indblkshift
;
319 * If this write is not off the end of the file
320 * we need to account for overwrites/unref.
322 if (start
<= dn
->dn_maxblkid
) {
323 for (l
= 0; l
< DN_MAX_LEVELS
; l
++)
326 while (start
<= dn
->dn_maxblkid
) {
329 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
330 err
= dbuf_hold_impl(dn
, 0, start
, FALSE
, FTAG
, &db
);
331 rw_exit(&dn
->dn_struct_rwlock
);
334 txh
->txh_tx
->tx_err
= err
;
338 dmu_tx_count_twig(txh
, dn
, db
, 0, start
, B_FALSE
,
343 * Account for new indirects appearing
344 * before this IO gets assigned into a txg.
347 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
348 for (bits
-= epbs
* (nlvls
- 1);
349 bits
>= 0; bits
-= epbs
)
350 txh
->txh_fudge
+= 1ULL << max_ibs
;
356 delta
= dn
->dn_datablksz
;
361 * 'end' is the last thing we will access, not one past.
362 * This way we won't overflow when accessing the last byte.
364 start
= P2ALIGN(off
, 1ULL << max_bs
);
365 end
= P2ROUNDUP(off
+ len
, 1ULL << max_bs
) - 1;
366 txh
->txh_space_towrite
+= end
- start
+ 1;
371 epbs
= min_ibs
- SPA_BLKPTRSHIFT
;
374 * The object contains at most 2^(64 - min_bs) blocks,
375 * and each indirect level maps 2^epbs.
377 for (bits
= 64 - min_bs
; bits
>= 0; bits
-= epbs
) {
380 ASSERT3U(end
, >=, start
);
381 txh
->txh_space_towrite
+= (end
- start
+ 1) << max_ibs
;
384 * We also need a new blkid=0 indirect block
385 * to reference any existing file data.
387 txh
->txh_space_towrite
+= 1ULL << max_ibs
;
392 if (txh
->txh_space_towrite
+ txh
->txh_space_tooverwrite
>
397 txh
->txh_tx
->tx_err
= err
;
401 dmu_tx_count_dnode(dmu_tx_hold_t
*txh
)
403 dnode_t
*dn
= txh
->txh_dnode
;
404 dnode_t
*mdn
= DMU_META_DNODE(txh
->txh_tx
->tx_objset
);
405 uint64_t space
= mdn
->dn_datablksz
+
406 ((mdn
->dn_nlevels
-1) << mdn
->dn_indblkshift
);
408 if (dn
&& dn
->dn_dbuf
->db_blkptr
&&
409 dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
410 dn
->dn_dbuf
->db_blkptr
, dn
->dn_dbuf
->db_blkptr
->blk_birth
)) {
411 txh
->txh_space_tooverwrite
+= space
;
412 txh
->txh_space_tounref
+= space
;
414 txh
->txh_space_towrite
+= space
;
415 if (dn
&& dn
->dn_dbuf
->db_blkptr
)
416 txh
->txh_space_tounref
+= space
;
421 dmu_tx_hold_write(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, int len
)
425 ASSERT(tx
->tx_txg
== 0);
426 ASSERT(len
< DMU_MAX_ACCESS
);
427 ASSERT(len
== 0 || UINT64_MAX
- off
>= len
- 1);
429 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
430 object
, THT_WRITE
, off
, len
);
434 dmu_tx_count_write(txh
, off
, len
);
435 dmu_tx_count_dnode(txh
);
439 dmu_tx_count_free(dmu_tx_hold_t
*txh
, uint64_t off
, uint64_t len
)
441 uint64_t blkid
, nblks
, lastblk
;
442 uint64_t space
= 0, unref
= 0, skipped
= 0;
443 dnode_t
*dn
= txh
->txh_dnode
;
444 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
445 spa_t
*spa
= txh
->txh_tx
->tx_pool
->dp_spa
;
448 if (dn
->dn_nlevels
== 0)
452 * The struct_rwlock protects us against dn_nlevels
453 * changing, in case (against all odds) we manage to dirty &
454 * sync out the changes after we check for being dirty.
455 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
457 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
458 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
459 if (dn
->dn_maxblkid
== 0) {
460 if (off
== 0 && len
>= dn
->dn_datablksz
) {
464 rw_exit(&dn
->dn_struct_rwlock
);
468 blkid
= off
>> dn
->dn_datablkshift
;
469 nblks
= (len
+ dn
->dn_datablksz
- 1) >> dn
->dn_datablkshift
;
471 if (blkid
>= dn
->dn_maxblkid
) {
472 rw_exit(&dn
->dn_struct_rwlock
);
475 if (blkid
+ nblks
> dn
->dn_maxblkid
)
476 nblks
= dn
->dn_maxblkid
- blkid
;
479 if (dn
->dn_nlevels
== 1) {
481 for (i
= 0; i
< nblks
; i
++) {
482 blkptr_t
*bp
= dn
->dn_phys
->dn_blkptr
;
483 ASSERT3U(blkid
+ i
, <, dn
->dn_nblkptr
);
485 if (dsl_dataset_block_freeable(ds
, bp
, bp
->blk_birth
)) {
486 dprintf_bp(bp
, "can free old%s", "");
487 space
+= bp_get_dsize(spa
, bp
);
489 unref
+= BP_GET_ASIZE(bp
);
495 * Add in memory requirements of higher-level indirects.
496 * This assumes a worst-possible scenario for dn_nlevels.
499 uint64_t blkcnt
= 1 + ((nblks
>> epbs
) >> epbs
);
500 int level
= (dn
->dn_nlevels
> 1) ? 2 : 1;
502 while (level
++ < DN_MAX_LEVELS
) {
503 txh
->txh_memory_tohold
+= blkcnt
<< dn
->dn_indblkshift
;
504 blkcnt
= 1 + (blkcnt
>> epbs
);
506 ASSERT(blkcnt
<= dn
->dn_nblkptr
);
509 lastblk
= blkid
+ nblks
- 1;
511 dmu_buf_impl_t
*dbuf
;
512 uint64_t ibyte
, new_blkid
;
514 int err
, i
, blkoff
, tochk
;
517 ibyte
= blkid
<< dn
->dn_datablkshift
;
518 err
= dnode_next_offset(dn
,
519 DNODE_FIND_HAVELOCK
, &ibyte
, 2, 1, 0);
520 new_blkid
= ibyte
>> dn
->dn_datablkshift
;
522 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
526 txh
->txh_tx
->tx_err
= err
;
529 if (new_blkid
> lastblk
) {
530 skipped
+= (lastblk
>> epbs
) - (blkid
>> epbs
) + 1;
534 if (new_blkid
> blkid
) {
535 ASSERT((new_blkid
>> epbs
) > (blkid
>> epbs
));
536 skipped
+= (new_blkid
>> epbs
) - (blkid
>> epbs
) - 1;
537 nblks
-= new_blkid
- blkid
;
540 blkoff
= P2PHASE(blkid
, epb
);
541 tochk
= MIN(epb
- blkoff
, nblks
);
543 err
= dbuf_hold_impl(dn
, 1, blkid
>> epbs
, FALSE
, FTAG
, &dbuf
);
545 txh
->txh_tx
->tx_err
= err
;
549 txh
->txh_memory_tohold
+= dbuf
->db
.db_size
;
552 * We don't check memory_tohold against DMU_MAX_ACCESS because
553 * memory_tohold is an over-estimation (especially the >L1
554 * indirect blocks), so it could fail. Callers should have
555 * already verified that they will not be holding too much
559 err
= dbuf_read(dbuf
, NULL
, DB_RF_HAVESTRUCT
| DB_RF_CANFAIL
);
561 txh
->txh_tx
->tx_err
= err
;
562 dbuf_rele(dbuf
, FTAG
);
566 bp
= dbuf
->db
.db_data
;
569 for (i
= 0; i
< tochk
; i
++) {
570 if (dsl_dataset_block_freeable(ds
, &bp
[i
],
572 dprintf_bp(&bp
[i
], "can free old%s", "");
573 space
+= bp_get_dsize(spa
, &bp
[i
]);
575 unref
+= BP_GET_ASIZE(bp
);
577 dbuf_rele(dbuf
, FTAG
);
582 rw_exit(&dn
->dn_struct_rwlock
);
584 /* account for new level 1 indirect blocks that might show up */
586 txh
->txh_fudge
+= skipped
<< dn
->dn_indblkshift
;
587 skipped
= MIN(skipped
, DMU_MAX_DELETEBLKCNT
>> epbs
);
588 txh
->txh_memory_tohold
+= skipped
<< dn
->dn_indblkshift
;
590 txh
->txh_space_tofree
+= space
;
591 txh
->txh_space_tounref
+= unref
;
595 dmu_tx_hold_free(dmu_tx_t
*tx
, uint64_t object
, uint64_t off
, uint64_t len
)
599 uint64_t start
, end
, i
;
603 ASSERT(tx
->tx_txg
== 0);
605 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
606 object
, THT_FREE
, off
, len
);
613 dmu_tx_count_write(txh
, off
, 1);
615 if (len
!= DMU_OBJECT_END
)
616 dmu_tx_count_write(txh
, off
+len
, 1);
618 dmu_tx_count_dnode(txh
);
620 if (off
>= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
)
622 if (len
== DMU_OBJECT_END
)
623 len
= (dn
->dn_maxblkid
+1) * dn
->dn_datablksz
- off
;
626 * For i/o error checking, read the first and last level-0
627 * blocks, and all the level-1 blocks. The above count_write's
628 * have already taken care of the level-0 blocks.
630 if (dn
->dn_nlevels
> 1) {
631 shift
= dn
->dn_datablkshift
+ dn
->dn_indblkshift
-
633 start
= off
>> shift
;
634 end
= dn
->dn_datablkshift
? ((off
+len
) >> shift
) : 0;
636 zio
= zio_root(tx
->tx_pool
->dp_spa
,
637 NULL
, NULL
, ZIO_FLAG_CANFAIL
);
638 for (i
= start
; i
<= end
; i
++) {
639 uint64_t ibyte
= i
<< shift
;
640 err
= dnode_next_offset(dn
, 0, &ibyte
, 2, 1, 0);
649 err
= dmu_tx_check_ioerr(zio
, dn
, 1, i
);
662 dmu_tx_count_free(txh
, off
, len
);
666 dmu_tx_hold_zap(dmu_tx_t
*tx
, uint64_t object
, int add
, const char *name
)
673 ASSERT(tx
->tx_txg
== 0);
675 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
676 object
, THT_ZAP
, add
, (uintptr_t)name
);
681 dmu_tx_count_dnode(txh
);
685 * We will be able to fit a new object's entries into one leaf
686 * block. So there will be at most 2 blocks total,
687 * including the header block.
689 dmu_tx_count_write(txh
, 0, 2 << fzap_default_block_shift
);
693 ASSERT3P(dmu_ot
[dn
->dn_type
].ot_byteswap
, ==, zap_byteswap
);
695 if (dn
->dn_maxblkid
== 0 && !add
) {
697 * If there is only one block (i.e. this is a micro-zap)
698 * and we are not adding anything, the accounting is simple.
700 err
= dmu_tx_check_ioerr(NULL
, dn
, 0, 0);
707 * Use max block size here, since we don't know how much
708 * the size will change between now and the dbuf dirty call.
710 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
711 &dn
->dn_phys
->dn_blkptr
[0],
712 dn
->dn_phys
->dn_blkptr
[0].blk_birth
)) {
713 txh
->txh_space_tooverwrite
+= SPA_MAXBLOCKSIZE
;
715 txh
->txh_space_towrite
+= SPA_MAXBLOCKSIZE
;
717 if (dn
->dn_phys
->dn_blkptr
[0].blk_birth
)
718 txh
->txh_space_tounref
+= SPA_MAXBLOCKSIZE
;
722 if (dn
->dn_maxblkid
> 0 && name
) {
724 * access the name in this fat-zap so that we'll check
725 * for i/o errors to the leaf blocks, etc.
727 err
= zap_lookup(dn
->dn_objset
, dn
->dn_object
, name
,
735 err
= zap_count_write(dn
->dn_objset
, dn
->dn_object
, name
, add
,
736 &txh
->txh_space_towrite
, &txh
->txh_space_tooverwrite
);
739 * If the modified blocks are scattered to the four winds,
740 * we'll have to modify an indirect twig for each.
742 epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
743 for (nblocks
= dn
->dn_maxblkid
>> epbs
; nblocks
!= 0; nblocks
>>= epbs
)
744 if (dn
->dn_objset
->os_dsl_dataset
->ds_phys
->ds_prev_snap_obj
)
745 txh
->txh_space_towrite
+= 3 << dn
->dn_indblkshift
;
747 txh
->txh_space_tooverwrite
+= 3 << dn
->dn_indblkshift
;
751 dmu_tx_hold_bonus(dmu_tx_t
*tx
, uint64_t object
)
755 ASSERT(tx
->tx_txg
== 0);
757 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
758 object
, THT_BONUS
, 0, 0);
760 dmu_tx_count_dnode(txh
);
764 dmu_tx_hold_space(dmu_tx_t
*tx
, uint64_t space
)
767 ASSERT(tx
->tx_txg
== 0);
769 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
,
770 DMU_NEW_OBJECT
, THT_SPACE
, space
, 0);
772 txh
->txh_space_towrite
+= space
;
776 dmu_tx_holds(dmu_tx_t
*tx
, uint64_t object
)
782 * By asserting that the tx is assigned, we're counting the
783 * number of dn_tx_holds, which is the same as the number of
784 * dn_holds. Otherwise, we'd be counting dn_holds, but
785 * dn_tx_holds could be 0.
787 ASSERT(tx
->tx_txg
!= 0);
789 /* if (tx->tx_anyobj == TRUE) */
792 for (txh
= list_head(&tx
->tx_holds
); txh
;
793 txh
= list_next(&tx
->tx_holds
, txh
)) {
794 if (txh
->txh_dnode
&& txh
->txh_dnode
->dn_object
== object
)
803 dmu_tx_dirty_buf(dmu_tx_t
*tx
, dmu_buf_impl_t
*db
)
806 int match_object
= FALSE
, match_offset
= FALSE
;
811 ASSERT(tx
->tx_txg
!= 0);
812 ASSERT(tx
->tx_objset
== NULL
|| dn
->dn_objset
== tx
->tx_objset
);
813 ASSERT3U(dn
->dn_object
, ==, db
->db
.db_object
);
820 /* XXX No checking on the meta dnode for now */
821 if (db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
826 for (txh
= list_head(&tx
->tx_holds
); txh
;
827 txh
= list_next(&tx
->tx_holds
, txh
)) {
828 ASSERT(dn
== NULL
|| dn
->dn_assigned_txg
== tx
->tx_txg
);
829 if (txh
->txh_dnode
== dn
&& txh
->txh_type
!= THT_NEWOBJECT
)
831 if (txh
->txh_dnode
== NULL
|| txh
->txh_dnode
== dn
) {
832 int datablkshift
= dn
->dn_datablkshift
?
833 dn
->dn_datablkshift
: SPA_MAXBLOCKSHIFT
;
834 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
835 int shift
= datablkshift
+ epbs
* db
->db_level
;
836 uint64_t beginblk
= shift
>= 64 ? 0 :
837 (txh
->txh_arg1
>> shift
);
838 uint64_t endblk
= shift
>= 64 ? 0 :
839 ((txh
->txh_arg1
+ txh
->txh_arg2
- 1) >> shift
);
840 uint64_t blkid
= db
->db_blkid
;
842 /* XXX txh_arg2 better not be zero... */
844 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
845 txh
->txh_type
, beginblk
, endblk
);
847 switch (txh
->txh_type
) {
849 if (blkid
>= beginblk
&& blkid
<= endblk
)
852 * We will let this hold work for the bonus
853 * or spill buffer so that we don't need to
854 * hold it when creating a new object.
856 if (blkid
== DMU_BONUS_BLKID
||
857 blkid
== DMU_SPILL_BLKID
)
860 * They might have to increase nlevels,
861 * thus dirtying the new TLIBs. Or the
862 * might have to change the block size,
863 * thus dirying the new lvl=0 blk=0.
870 * We will dirty all the level 1 blocks in
871 * the free range and perhaps the first and
872 * last level 0 block.
874 if (blkid
>= beginblk
&& (blkid
<= endblk
||
875 txh
->txh_arg2
== DMU_OBJECT_END
))
879 if (blkid
== DMU_SPILL_BLKID
)
883 if (blkid
== DMU_BONUS_BLKID
)
893 ASSERT(!"bad txh_type");
896 if (match_object
&& match_offset
) {
902 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
903 (u_longlong_t
)db
->db
.db_object
, db
->db_level
,
904 (u_longlong_t
)db
->db_blkid
);
909 dmu_tx_try_assign(dmu_tx_t
*tx
, uint64_t txg_how
)
912 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
913 uint64_t memory
, asize
, fsize
, usize
;
914 uint64_t towrite
, tofree
, tooverwrite
, tounref
, tohold
, fudge
;
916 ASSERT3U(tx
->tx_txg
, ==, 0);
919 DMU_TX_STAT_BUMP(dmu_tx_error
);
923 if (spa_suspended(spa
)) {
924 DMU_TX_STAT_BUMP(dmu_tx_suspended
);
927 * If the user has indicated a blocking failure mode
928 * then return ERESTART which will block in dmu_tx_wait().
929 * Otherwise, return EIO so that an error can get
930 * propagated back to the VOP calls.
932 * Note that we always honor the txg_how flag regardless
933 * of the failuremode setting.
935 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_CONTINUE
&&
942 tx
->tx_txg
= txg_hold_open(tx
->tx_pool
, &tx
->tx_txgh
);
943 tx
->tx_needassign_txh
= NULL
;
946 * NB: No error returns are allowed after txg_hold_open, but
947 * before processing the dnode holds, due to the
948 * dmu_tx_unassign() logic.
951 towrite
= tofree
= tooverwrite
= tounref
= tohold
= fudge
= 0;
952 for (txh
= list_head(&tx
->tx_holds
); txh
;
953 txh
= list_next(&tx
->tx_holds
, txh
)) {
954 dnode_t
*dn
= txh
->txh_dnode
;
956 mutex_enter(&dn
->dn_mtx
);
957 if (dn
->dn_assigned_txg
== tx
->tx_txg
- 1) {
958 mutex_exit(&dn
->dn_mtx
);
959 tx
->tx_needassign_txh
= txh
;
960 DMU_TX_STAT_BUMP(dmu_tx_group
);
963 if (dn
->dn_assigned_txg
== 0)
964 dn
->dn_assigned_txg
= tx
->tx_txg
;
965 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
966 (void) refcount_add(&dn
->dn_tx_holds
, tx
);
967 mutex_exit(&dn
->dn_mtx
);
969 towrite
+= txh
->txh_space_towrite
;
970 tofree
+= txh
->txh_space_tofree
;
971 tooverwrite
+= txh
->txh_space_tooverwrite
;
972 tounref
+= txh
->txh_space_tounref
;
973 tohold
+= txh
->txh_memory_tohold
;
974 fudge
+= txh
->txh_fudge
;
978 * NB: This check must be after we've held the dnodes, so that
979 * the dmu_tx_unassign() logic will work properly
981 if (txg_how
>= TXG_INITIAL
&& txg_how
!= tx
->tx_txg
) {
982 DMU_TX_STAT_BUMP(dmu_tx_how
);
987 * If a snapshot has been taken since we made our estimates,
988 * assume that we won't be able to free or overwrite anything.
991 dsl_dataset_prev_snap_txg(tx
->tx_objset
->os_dsl_dataset
) >
992 tx
->tx_lastsnap_txg
) {
993 towrite
+= tooverwrite
;
994 tooverwrite
= tofree
= 0;
997 /* needed allocation: worst-case estimate of write space */
998 asize
= spa_get_asize(tx
->tx_pool
->dp_spa
, towrite
+ tooverwrite
);
999 /* freed space estimate: worst-case overwrite + free estimate */
1000 fsize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tooverwrite
) + tofree
;
1001 /* convert unrefd space to worst-case estimate */
1002 usize
= spa_get_asize(tx
->tx_pool
->dp_spa
, tounref
);
1003 /* calculate memory footprint estimate */
1004 memory
= towrite
+ tooverwrite
+ tohold
;
1008 * Add in 'tohold' to account for our dirty holds on this memory
1009 * XXX - the "fudge" factor is to account for skipped blocks that
1010 * we missed because dnode_next_offset() misses in-core-only blocks.
1012 tx
->tx_space_towrite
= asize
+
1013 spa_get_asize(tx
->tx_pool
->dp_spa
, tohold
+ fudge
);
1014 tx
->tx_space_tofree
= tofree
;
1015 tx
->tx_space_tooverwrite
= tooverwrite
;
1016 tx
->tx_space_tounref
= tounref
;
1019 if (tx
->tx_dir
&& asize
!= 0) {
1020 int err
= dsl_dir_tempreserve_space(tx
->tx_dir
, memory
,
1021 asize
, fsize
, usize
, &tx
->tx_tempreserve_cookie
, tx
);
1026 DMU_TX_STAT_BUMP(dmu_tx_assigned
);
1032 dmu_tx_unassign(dmu_tx_t
*tx
)
1036 if (tx
->tx_txg
== 0)
1039 txg_rele_to_quiesce(&tx
->tx_txgh
);
1041 for (txh
= list_head(&tx
->tx_holds
); txh
!= tx
->tx_needassign_txh
;
1042 txh
= list_next(&tx
->tx_holds
, txh
)) {
1043 dnode_t
*dn
= txh
->txh_dnode
;
1047 mutex_enter(&dn
->dn_mtx
);
1048 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1050 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1051 dn
->dn_assigned_txg
= 0;
1052 cv_broadcast(&dn
->dn_notxholds
);
1054 mutex_exit(&dn
->dn_mtx
);
1057 txg_rele_to_sync(&tx
->tx_txgh
);
1059 tx
->tx_lasttried_txg
= tx
->tx_txg
;
1064 * Assign tx to a transaction group. txg_how can be one of:
1066 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1067 * a new one. This should be used when you're not holding locks.
1068 * If will only fail if we're truly out of space (or over quota).
1070 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1071 * blocking, returns immediately with ERESTART. This should be used
1072 * whenever you're holding locks. On an ERESTART error, the caller
1073 * should drop locks, do a dmu_tx_wait(tx), and try again.
1075 * (3) A specific txg. Use this if you need to ensure that multiple
1076 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1077 * returns ERESTART if it can't assign you into the requested txg.
1080 dmu_tx_assign(dmu_tx_t
*tx
, uint64_t txg_how
)
1084 ASSERT(tx
->tx_txg
== 0);
1085 ASSERT(txg_how
!= 0);
1086 ASSERT(!dsl_pool_sync_context(tx
->tx_pool
));
1088 while ((err
= dmu_tx_try_assign(tx
, txg_how
)) != 0) {
1089 dmu_tx_unassign(tx
);
1091 if (err
!= ERESTART
|| txg_how
!= TXG_WAIT
)
1097 txg_rele_to_quiesce(&tx
->tx_txgh
);
1103 dmu_tx_wait(dmu_tx_t
*tx
)
1105 spa_t
*spa
= tx
->tx_pool
->dp_spa
;
1107 ASSERT(tx
->tx_txg
== 0);
1110 * It's possible that the pool has become active after this thread
1111 * has tried to obtain a tx. If that's the case then his
1112 * tx_lasttried_txg would not have been assigned.
1114 if (spa_suspended(spa
) || tx
->tx_lasttried_txg
== 0) {
1115 txg_wait_synced(tx
->tx_pool
, spa_last_synced_txg(spa
) + 1);
1116 } else if (tx
->tx_needassign_txh
) {
1117 dnode_t
*dn
= tx
->tx_needassign_txh
->txh_dnode
;
1119 mutex_enter(&dn
->dn_mtx
);
1120 while (dn
->dn_assigned_txg
== tx
->tx_lasttried_txg
- 1)
1121 cv_wait(&dn
->dn_notxholds
, &dn
->dn_mtx
);
1122 mutex_exit(&dn
->dn_mtx
);
1123 tx
->tx_needassign_txh
= NULL
;
1125 txg_wait_open(tx
->tx_pool
, tx
->tx_lasttried_txg
+ 1);
1130 dmu_tx_willuse_space(dmu_tx_t
*tx
, int64_t delta
)
1133 if (tx
->tx_dir
== NULL
|| delta
== 0)
1137 ASSERT3U(refcount_count(&tx
->tx_space_written
) + delta
, <=,
1138 tx
->tx_space_towrite
);
1139 (void) refcount_add_many(&tx
->tx_space_written
, delta
, NULL
);
1141 (void) refcount_add_many(&tx
->tx_space_freed
, -delta
, NULL
);
1147 dmu_tx_commit(dmu_tx_t
*tx
)
1151 ASSERT(tx
->tx_txg
!= 0);
1153 while ((txh
= list_head(&tx
->tx_holds
))) {
1154 dnode_t
*dn
= txh
->txh_dnode
;
1156 list_remove(&tx
->tx_holds
, txh
);
1157 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1160 mutex_enter(&dn
->dn_mtx
);
1161 ASSERT3U(dn
->dn_assigned_txg
, ==, tx
->tx_txg
);
1163 if (refcount_remove(&dn
->dn_tx_holds
, tx
) == 0) {
1164 dn
->dn_assigned_txg
= 0;
1165 cv_broadcast(&dn
->dn_notxholds
);
1167 mutex_exit(&dn
->dn_mtx
);
1171 if (tx
->tx_tempreserve_cookie
)
1172 dsl_dir_tempreserve_clear(tx
->tx_tempreserve_cookie
, tx
);
1174 if (!list_is_empty(&tx
->tx_callbacks
))
1175 txg_register_callbacks(&tx
->tx_txgh
, &tx
->tx_callbacks
);
1177 if (tx
->tx_anyobj
== FALSE
)
1178 txg_rele_to_sync(&tx
->tx_txgh
);
1180 list_destroy(&tx
->tx_callbacks
);
1181 list_destroy(&tx
->tx_holds
);
1183 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1184 tx
->tx_space_towrite
, refcount_count(&tx
->tx_space_written
),
1185 tx
->tx_space_tofree
, refcount_count(&tx
->tx_space_freed
));
1186 refcount_destroy_many(&tx
->tx_space_written
,
1187 refcount_count(&tx
->tx_space_written
));
1188 refcount_destroy_many(&tx
->tx_space_freed
,
1189 refcount_count(&tx
->tx_space_freed
));
1191 kmem_free(tx
, sizeof (dmu_tx_t
));
1195 dmu_tx_abort(dmu_tx_t
*tx
)
1199 ASSERT(tx
->tx_txg
== 0);
1201 while ((txh
= list_head(&tx
->tx_holds
))) {
1202 dnode_t
*dn
= txh
->txh_dnode
;
1204 list_remove(&tx
->tx_holds
, txh
);
1205 kmem_free(txh
, sizeof (dmu_tx_hold_t
));
1211 * Call any registered callbacks with an error code.
1213 if (!list_is_empty(&tx
->tx_callbacks
))
1214 dmu_tx_do_callbacks(&tx
->tx_callbacks
, ECANCELED
);
1216 list_destroy(&tx
->tx_callbacks
);
1217 list_destroy(&tx
->tx_holds
);
1219 refcount_destroy_many(&tx
->tx_space_written
,
1220 refcount_count(&tx
->tx_space_written
));
1221 refcount_destroy_many(&tx
->tx_space_freed
,
1222 refcount_count(&tx
->tx_space_freed
));
1224 kmem_free(tx
, sizeof (dmu_tx_t
));
1228 dmu_tx_get_txg(dmu_tx_t
*tx
)
1230 ASSERT(tx
->tx_txg
!= 0);
1231 return (tx
->tx_txg
);
1235 dmu_tx_callback_register(dmu_tx_t
*tx
, dmu_tx_callback_func_t
*func
, void *data
)
1237 dmu_tx_callback_t
*dcb
;
1239 dcb
= kmem_alloc(sizeof (dmu_tx_callback_t
), KM_SLEEP
);
1241 dcb
->dcb_func
= func
;
1242 dcb
->dcb_data
= data
;
1244 list_insert_tail(&tx
->tx_callbacks
, dcb
);
1248 * Call all the commit callbacks on a list, with a given error code.
1251 dmu_tx_do_callbacks(list_t
*cb_list
, int error
)
1253 dmu_tx_callback_t
*dcb
;
1255 while ((dcb
= list_head(cb_list
))) {
1256 list_remove(cb_list
, dcb
);
1257 dcb
->dcb_func(dcb
->dcb_data
, error
);
1258 kmem_free(dcb
, sizeof (dmu_tx_callback_t
));
1263 * Interface to hold a bunch of attributes.
1264 * used for creating new files.
1265 * attrsize is the total size of all attributes
1266 * to be added during object creation
1268 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1272 * hold necessary attribute name for attribute registration.
1273 * should be a very rare case where this is needed. If it does
1274 * happen it would only happen on the first write to the file system.
1277 dmu_tx_sa_registration_hold(sa_os_t
*sa
, dmu_tx_t
*tx
)
1281 if (!sa
->sa_need_attr_registration
)
1284 for (i
= 0; i
!= sa
->sa_num_attrs
; i
++) {
1285 if (!sa
->sa_attr_table
[i
].sa_registered
) {
1286 if (sa
->sa_reg_attr_obj
)
1287 dmu_tx_hold_zap(tx
, sa
->sa_reg_attr_obj
,
1288 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1290 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
,
1291 B_TRUE
, sa
->sa_attr_table
[i
].sa_name
);
1298 dmu_tx_hold_spill(dmu_tx_t
*tx
, uint64_t object
)
1304 txh
= dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, object
,
1307 dn
= txh
->txh_dnode
;
1312 /* If blkptr doesn't exist then add space to towrite */
1313 bp
= &dn
->dn_phys
->dn_spill
;
1314 if (BP_IS_HOLE(bp
)) {
1315 txh
->txh_space_towrite
+= SPA_MAXBLOCKSIZE
;
1316 txh
->txh_space_tounref
= 0;
1318 if (dsl_dataset_block_freeable(dn
->dn_objset
->os_dsl_dataset
,
1320 txh
->txh_space_tooverwrite
+= SPA_MAXBLOCKSIZE
;
1322 txh
->txh_space_towrite
+= SPA_MAXBLOCKSIZE
;
1324 txh
->txh_space_tounref
+= SPA_MAXBLOCKSIZE
;
1329 dmu_tx_hold_sa_create(dmu_tx_t
*tx
, int attrsize
)
1331 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1333 dmu_tx_hold_bonus(tx
, DMU_NEW_OBJECT
);
1335 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1338 if (tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1339 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1341 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1342 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1343 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1344 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1347 dmu_tx_sa_registration_hold(sa
, tx
);
1349 if (attrsize
<= DN_MAX_BONUSLEN
&& !sa
->sa_force_spill
)
1352 (void) dmu_tx_hold_object_impl(tx
, tx
->tx_objset
, DMU_NEW_OBJECT
,
1359 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1361 * variable_size is the total size of all variable sized attributes
1362 * passed to this function. It is not the total size of all
1363 * variable size attributes that *may* exist on this object.
1366 dmu_tx_hold_sa(dmu_tx_t
*tx
, sa_handle_t
*hdl
, boolean_t may_grow
)
1369 sa_os_t
*sa
= tx
->tx_objset
->os_sa
;
1371 ASSERT(hdl
!= NULL
);
1373 object
= sa_handle_object(hdl
);
1375 dmu_tx_hold_bonus(tx
, object
);
1377 if (tx
->tx_objset
->os_sa
->sa_master_obj
== 0)
1380 if (tx
->tx_objset
->os_sa
->sa_reg_attr_obj
== 0 ||
1381 tx
->tx_objset
->os_sa
->sa_layout_attr_obj
== 0) {
1382 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_LAYOUTS
);
1383 dmu_tx_hold_zap(tx
, sa
->sa_master_obj
, B_TRUE
, SA_REGISTRY
);
1384 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1385 dmu_tx_hold_zap(tx
, DMU_NEW_OBJECT
, B_TRUE
, NULL
);
1388 dmu_tx_sa_registration_hold(sa
, tx
);
1390 if (may_grow
&& tx
->tx_objset
->os_sa
->sa_layout_attr_obj
)
1391 dmu_tx_hold_zap(tx
, sa
->sa_layout_attr_obj
, B_TRUE
, NULL
);
1393 if (sa
->sa_force_spill
|| may_grow
|| hdl
->sa_spill
) {
1394 ASSERT(tx
->tx_txg
== 0);
1395 dmu_tx_hold_spill(tx
, object
);
1397 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)hdl
->sa_bonus
;
1402 if (dn
->dn_have_spill
) {
1403 ASSERT(tx
->tx_txg
== 0);
1404 dmu_tx_hold_spill(tx
, object
);
1413 dmu_tx_ksp
= kstat_create("zfs", 0, "dmu_tx", "misc",
1414 KSTAT_TYPE_NAMED
, sizeof (dmu_tx_stats
) / sizeof (kstat_named_t
),
1415 KSTAT_FLAG_VIRTUAL
);
1417 if (dmu_tx_ksp
!= NULL
) {
1418 dmu_tx_ksp
->ks_data
= &dmu_tx_stats
;
1419 kstat_install(dmu_tx_ksp
);
1426 if (dmu_tx_ksp
!= NULL
) {
1427 kstat_delete(dmu_tx_ksp
);
1432 #if defined(_KERNEL) && defined(HAVE_SPL)
1433 EXPORT_SYMBOL(dmu_tx_create
);
1434 EXPORT_SYMBOL(dmu_tx_hold_write
);
1435 EXPORT_SYMBOL(dmu_tx_hold_free
);
1436 EXPORT_SYMBOL(dmu_tx_hold_zap
);
1437 EXPORT_SYMBOL(dmu_tx_hold_bonus
);
1438 EXPORT_SYMBOL(dmu_tx_abort
);
1439 EXPORT_SYMBOL(dmu_tx_assign
);
1440 EXPORT_SYMBOL(dmu_tx_wait
);
1441 EXPORT_SYMBOL(dmu_tx_commit
);
1442 EXPORT_SYMBOL(dmu_tx_get_txg
);
1443 EXPORT_SYMBOL(dmu_tx_callback_register
);
1444 EXPORT_SYMBOL(dmu_tx_do_callbacks
);
1445 EXPORT_SYMBOL(dmu_tx_hold_spill
);
1446 EXPORT_SYMBOL(dmu_tx_hold_sa_create
);
1447 EXPORT_SYMBOL(dmu_tx_hold_sa
);