]>
git.proxmox.com Git - mirror_zfs-debian.git/blob - module/zfs/dnode_sync.c
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 2008 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #pragma ident "%Z%%M% %I% %E% SMI"
28 #include <sys/zfs_context.h>
30 #include <sys/dnode.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dataset.h>
38 dnode_increase_indirection(dnode_t
*dn
, dmu_tx_t
*tx
)
41 int txgoff
= tx
->tx_txg
& TXG_MASK
;
42 int nblkptr
= dn
->dn_phys
->dn_nblkptr
;
43 int old_toplvl
= dn
->dn_phys
->dn_nlevels
- 1;
44 int new_level
= dn
->dn_next_nlevels
[txgoff
];
47 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
49 /* this dnode can't be paged out because it's dirty */
50 ASSERT(dn
->dn_phys
->dn_type
!= DMU_OT_NONE
);
51 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
52 ASSERT(new_level
> 1 && dn
->dn_phys
->dn_nlevels
> 0);
54 db
= dbuf_hold_level(dn
, dn
->dn_phys
->dn_nlevels
, 0, FTAG
);
57 dn
->dn_phys
->dn_nlevels
= new_level
;
58 dprintf("os=%p obj=%llu, increase to %d\n", dn
->dn_objset
,
59 dn
->dn_object
, dn
->dn_phys
->dn_nlevels
);
61 /* check for existing blkptrs in the dnode */
62 for (i
= 0; i
< nblkptr
; i
++)
63 if (!BP_IS_HOLE(&dn
->dn_phys
->dn_blkptr
[i
]))
66 /* transfer dnode's block pointers to new indirect block */
67 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
|DB_RF_HAVESTRUCT
);
68 ASSERT(db
->db
.db_data
);
69 ASSERT(arc_released(db
->db_buf
));
70 ASSERT3U(sizeof (blkptr_t
) * nblkptr
, <=, db
->db
.db_size
);
71 bcopy(dn
->dn_phys
->dn_blkptr
, db
->db
.db_data
,
72 sizeof (blkptr_t
) * nblkptr
);
73 arc_buf_freeze(db
->db_buf
);
76 /* set dbuf's parent pointers to new indirect buf */
77 for (i
= 0; i
< nblkptr
; i
++) {
78 dmu_buf_impl_t
*child
= dbuf_find(dn
, old_toplvl
, i
);
82 ASSERT3P(child
->db_dnode
, ==, dn
);
83 if (child
->db_parent
&& child
->db_parent
!= dn
->dn_dbuf
) {
84 ASSERT(child
->db_parent
->db_level
== db
->db_level
);
85 ASSERT(child
->db_blkptr
!=
86 &dn
->dn_phys
->dn_blkptr
[child
->db_blkid
]);
87 mutex_exit(&child
->db_mtx
);
90 ASSERT(child
->db_parent
== NULL
||
91 child
->db_parent
== dn
->dn_dbuf
);
93 child
->db_parent
= db
;
94 dbuf_add_ref(db
, child
);
96 child
->db_blkptr
= (blkptr_t
*)db
->db
.db_data
+ i
;
98 child
->db_blkptr
= NULL
;
99 dprintf_dbuf_bp(child
, child
->db_blkptr
,
100 "changed db_blkptr to new indirect %s", "");
102 mutex_exit(&child
->db_mtx
);
105 bzero(dn
->dn_phys
->dn_blkptr
, sizeof (blkptr_t
) * nblkptr
);
109 rw_exit(&dn
->dn_struct_rwlock
);
113 free_blocks(dnode_t
*dn
, blkptr_t
*bp
, int num
, dmu_tx_t
*tx
)
115 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
116 uint64_t bytesfreed
= 0;
117 int i
, blocks_freed
= 0;
119 dprintf("ds=%p obj=%llx num=%d\n", ds
, dn
->dn_object
, num
);
121 for (i
= 0; i
< num
; i
++, bp
++) {
125 bytesfreed
+= dsl_dataset_block_kill(ds
, bp
, dn
->dn_zio
, tx
);
126 ASSERT3U(bytesfreed
, <=, DN_USED_BYTES(dn
->dn_phys
));
127 bzero(bp
, sizeof (blkptr_t
));
130 dnode_diduse_space(dn
, -bytesfreed
);
131 return (blocks_freed
);
136 free_verify(dmu_buf_impl_t
*db
, uint64_t start
, uint64_t end
, dmu_tx_t
*tx
)
140 uint64_t txg
= tx
->tx_txg
;
142 epbs
= db
->db_dnode
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
143 off
= start
- (db
->db_blkid
* 1<<epbs
);
144 num
= end
- start
+ 1;
146 ASSERT3U(off
, >=, 0);
147 ASSERT3U(num
, >=, 0);
148 ASSERT3U(db
->db_level
, >, 0);
149 ASSERT3U(db
->db
.db_size
, ==, 1<<db
->db_dnode
->dn_phys
->dn_indblkshift
);
150 ASSERT3U(off
+num
, <=, db
->db
.db_size
>> SPA_BLKPTRSHIFT
);
151 ASSERT(db
->db_blkptr
!= NULL
);
153 for (i
= off
; i
< off
+num
; i
++) {
155 dmu_buf_impl_t
*child
;
156 dbuf_dirty_record_t
*dr
;
159 ASSERT(db
->db_level
== 1);
161 rw_enter(&db
->db_dnode
->dn_struct_rwlock
, RW_READER
);
162 err
= dbuf_hold_impl(db
->db_dnode
, db
->db_level
-1,
163 (db
->db_blkid
<< epbs
) + i
, TRUE
, FTAG
, &child
);
164 rw_exit(&db
->db_dnode
->dn_struct_rwlock
);
168 ASSERT(child
->db_level
== 0);
169 dr
= child
->db_last_dirty
;
170 while (dr
&& dr
->dr_txg
> txg
)
172 ASSERT(dr
== NULL
|| dr
->dr_txg
== txg
);
174 /* data_old better be zeroed */
176 buf
= dr
->dt
.dl
.dr_data
->b_data
;
177 for (j
= 0; j
< child
->db
.db_size
>> 3; j
++) {
179 panic("freed data not zero: "
180 "child=%p i=%d off=%d num=%d\n",
181 (void *)child
, i
, off
, num
);
187 * db_data better be zeroed unless it's dirty in a
190 mutex_enter(&child
->db_mtx
);
191 buf
= child
->db
.db_data
;
192 if (buf
!= NULL
&& child
->db_state
!= DB_FILL
&&
193 child
->db_last_dirty
== NULL
) {
194 for (j
= 0; j
< child
->db
.db_size
>> 3; j
++) {
196 panic("freed data not zero: "
197 "child=%p i=%d off=%d num=%d\n",
198 (void *)child
, i
, off
, num
);
202 mutex_exit(&child
->db_mtx
);
204 dbuf_rele(child
, FTAG
);
212 free_children(dmu_buf_impl_t
*db
, uint64_t blkid
, uint64_t nblks
, int trunc
,
215 dnode_t
*dn
= db
->db_dnode
;
217 dmu_buf_impl_t
*subdb
;
218 uint64_t start
, end
, dbstart
, dbend
, i
;
219 int epbs
, shift
, err
;
221 int blocks_freed
= 0;
224 * There is a small possibility that this block will not be cached:
225 * 1 - if level > 1 and there are no children with level <= 1
226 * 2 - if we didn't get a dirty hold (because this block had just
227 * finished being written -- and so had no holds), and then this
228 * block got evicted before we got here.
230 if (db
->db_state
!= DB_CACHED
)
231 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
233 arc_release(db
->db_buf
, db
);
234 bp
= (blkptr_t
*)db
->db
.db_data
;
236 epbs
= db
->db_dnode
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
237 shift
= (db
->db_level
- 1) * epbs
;
238 dbstart
= db
->db_blkid
<< epbs
;
239 start
= blkid
>> shift
;
240 if (dbstart
< start
) {
241 bp
+= start
- dbstart
;
246 dbend
= ((db
->db_blkid
+ 1) << epbs
) - 1;
247 end
= (blkid
+ nblks
- 1) >> shift
;
252 ASSERT3U(start
, <=, end
);
254 if (db
->db_level
== 1) {
255 FREE_VERIFY(db
, start
, end
, tx
);
256 blocks_freed
= free_blocks(dn
, bp
, end
-start
+1, tx
);
257 arc_buf_freeze(db
->db_buf
);
258 ASSERT(all
|| blocks_freed
== 0 || db
->db_last_dirty
);
259 return (all
? ALL
: blocks_freed
);
262 for (i
= start
; i
<= end
; i
++, bp
++) {
265 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
266 err
= dbuf_hold_impl(dn
, db
->db_level
-1, i
, TRUE
, FTAG
, &subdb
);
267 ASSERT3U(err
, ==, 0);
268 rw_exit(&dn
->dn_struct_rwlock
);
270 if (free_children(subdb
, blkid
, nblks
, trunc
, tx
) == ALL
) {
271 ASSERT3P(subdb
->db_blkptr
, ==, bp
);
272 blocks_freed
+= free_blocks(dn
, bp
, 1, tx
);
276 dbuf_rele(subdb
, FTAG
);
278 arc_buf_freeze(db
->db_buf
);
281 for (i
= start
; i
<= end
; i
++, bp
++) {
282 if (i
== start
&& blkid
!= 0)
284 else if (i
== end
&& !trunc
)
286 ASSERT3U(bp
->blk_birth
, ==, 0);
289 ASSERT(all
|| blocks_freed
== 0 || db
->db_last_dirty
);
290 return (all
? ALL
: blocks_freed
);
294 * free_range: Traverse the indicated range of the provided file
295 * and "free" all the blocks contained there.
298 dnode_sync_free_range(dnode_t
*dn
, uint64_t blkid
, uint64_t nblks
, dmu_tx_t
*tx
)
300 blkptr_t
*bp
= dn
->dn_phys
->dn_blkptr
;
302 int trunc
, start
, end
, shift
, i
, err
;
303 int dnlevel
= dn
->dn_phys
->dn_nlevels
;
305 if (blkid
> dn
->dn_phys
->dn_maxblkid
)
308 ASSERT(dn
->dn_phys
->dn_maxblkid
< UINT64_MAX
);
309 trunc
= blkid
+ nblks
> dn
->dn_phys
->dn_maxblkid
;
311 nblks
= dn
->dn_phys
->dn_maxblkid
- blkid
+ 1;
313 /* There are no indirect blocks in the object */
315 if (blkid
>= dn
->dn_phys
->dn_nblkptr
) {
316 /* this range was never made persistent */
319 ASSERT3U(blkid
+ nblks
, <=, dn
->dn_phys
->dn_nblkptr
);
320 (void) free_blocks(dn
, bp
+ blkid
, nblks
, tx
);
322 uint64_t off
= (dn
->dn_phys
->dn_maxblkid
+ 1) *
323 (dn
->dn_phys
->dn_datablkszsec
<< SPA_MINBLOCKSHIFT
);
324 dn
->dn_phys
->dn_maxblkid
= (blkid
? blkid
- 1 : 0);
325 ASSERT(off
< dn
->dn_phys
->dn_maxblkid
||
326 dn
->dn_phys
->dn_maxblkid
== 0 ||
327 dnode_next_offset(dn
, 0, &off
, 1, 1, 0) != 0);
332 shift
= (dnlevel
- 1) * (dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
333 start
= blkid
>> shift
;
334 ASSERT(start
< dn
->dn_phys
->dn_nblkptr
);
335 end
= (blkid
+ nblks
- 1) >> shift
;
337 for (i
= start
; i
<= end
; i
++, bp
++) {
340 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
341 err
= dbuf_hold_impl(dn
, dnlevel
-1, i
, TRUE
, FTAG
, &db
);
342 ASSERT3U(err
, ==, 0);
343 rw_exit(&dn
->dn_struct_rwlock
);
345 if (free_children(db
, blkid
, nblks
, trunc
, tx
) == ALL
) {
346 ASSERT3P(db
->db_blkptr
, ==, bp
);
347 (void) free_blocks(dn
, bp
, 1, tx
);
352 uint64_t off
= (dn
->dn_phys
->dn_maxblkid
+ 1) *
353 (dn
->dn_phys
->dn_datablkszsec
<< SPA_MINBLOCKSHIFT
);
354 dn
->dn_phys
->dn_maxblkid
= (blkid
? blkid
- 1 : 0);
355 ASSERT(off
< dn
->dn_phys
->dn_maxblkid
||
356 dn
->dn_phys
->dn_maxblkid
== 0 ||
357 dnode_next_offset(dn
, 0, &off
, 1, 1, 0) != 0);
362 * Try to kick all the dnodes dbufs out of the cache...
365 dnode_evict_dbufs(dnode_t
*dn
)
371 dmu_buf_impl_t
*db
, marker
;
372 int evicting
= FALSE
;
375 mutex_enter(&dn
->dn_dbufs_mtx
);
376 list_insert_tail(&dn
->dn_dbufs
, &marker
);
377 db
= list_head(&dn
->dn_dbufs
);
378 for (; db
!= &marker
; db
= list_head(&dn
->dn_dbufs
)) {
379 list_remove(&dn
->dn_dbufs
, db
);
380 list_insert_tail(&dn
->dn_dbufs
, db
);
381 ASSERT3P(db
->db_dnode
, ==, dn
);
383 mutex_enter(&db
->db_mtx
);
384 if (db
->db_state
== DB_EVICTING
) {
387 mutex_exit(&db
->db_mtx
);
388 } else if (refcount_is_zero(&db
->db_holds
)) {
390 dbuf_clear(db
); /* exits db_mtx for us */
392 mutex_exit(&db
->db_mtx
);
396 list_remove(&dn
->dn_dbufs
, &marker
);
398 * NB: we need to drop dn_dbufs_mtx between passes so
399 * that any DB_EVICTING dbufs can make progress.
400 * Ideally, we would have some cv we could wait on, but
401 * since we don't, just wait a bit to give the other
402 * thread a chance to run.
404 mutex_exit(&dn
->dn_dbufs_mtx
);
408 ASSERT(pass
< 100); /* sanity check */
411 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
412 if (dn
->dn_bonus
&& refcount_is_zero(&dn
->dn_bonus
->db_holds
)) {
413 mutex_enter(&dn
->dn_bonus
->db_mtx
);
414 dbuf_evict(dn
->dn_bonus
);
417 rw_exit(&dn
->dn_struct_rwlock
);
421 dnode_undirty_dbufs(list_t
*list
)
423 dbuf_dirty_record_t
*dr
;
425 while (dr
= list_head(list
)) {
426 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
427 uint64_t txg
= dr
->dr_txg
;
429 mutex_enter(&db
->db_mtx
);
430 /* XXX - use dbuf_undirty()? */
431 list_remove(list
, dr
);
432 ASSERT(db
->db_last_dirty
== dr
);
433 db
->db_last_dirty
= NULL
;
434 db
->db_dirtycnt
-= 1;
435 if (db
->db_level
== 0) {
436 ASSERT(db
->db_blkid
== DB_BONUS_BLKID
||
437 dr
->dt
.dl
.dr_data
== db
->db_buf
);
439 mutex_exit(&db
->db_mtx
);
441 mutex_exit(&db
->db_mtx
);
442 dnode_undirty_dbufs(&dr
->dt
.di
.dr_children
);
444 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
445 dbuf_rele(db
, (void *)(uintptr_t)txg
);
450 dnode_sync_free(dnode_t
*dn
, dmu_tx_t
*tx
)
452 int txgoff
= tx
->tx_txg
& TXG_MASK
;
454 ASSERT(dmu_tx_is_syncing(tx
));
457 * Our contents should have been freed in dnode_sync() by the
458 * free range record inserted by the caller of dnode_free().
460 ASSERT3U(DN_USED_BYTES(dn
->dn_phys
), ==, 0);
461 ASSERT(BP_IS_HOLE(dn
->dn_phys
->dn_blkptr
));
463 dnode_undirty_dbufs(&dn
->dn_dirty_records
[txgoff
]);
464 dnode_evict_dbufs(dn
);
465 ASSERT3P(list_head(&dn
->dn_dbufs
), ==, NULL
);
468 * XXX - It would be nice to assert this, but we may still
469 * have residual holds from async evictions from the arc...
471 * zfs_obj_to_path() also depends on this being
474 * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
477 /* Undirty next bits */
478 dn
->dn_next_nlevels
[txgoff
] = 0;
479 dn
->dn_next_indblkshift
[txgoff
] = 0;
480 dn
->dn_next_blksz
[txgoff
] = 0;
482 /* ASSERT(blkptrs are zero); */
483 ASSERT(dn
->dn_phys
->dn_type
!= DMU_OT_NONE
);
484 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
486 ASSERT(dn
->dn_free_txg
> 0);
487 if (dn
->dn_allocated_txg
!= dn
->dn_free_txg
)
488 dbuf_will_dirty(dn
->dn_dbuf
, tx
);
489 bzero(dn
->dn_phys
, sizeof (dnode_phys_t
));
491 mutex_enter(&dn
->dn_mtx
);
492 dn
->dn_type
= DMU_OT_NONE
;
494 dn
->dn_allocated_txg
= 0;
496 mutex_exit(&dn
->dn_mtx
);
498 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
500 dnode_rele(dn
, (void *)(uintptr_t)tx
->tx_txg
);
502 * Now that we've released our hold, the dnode may
503 * be evicted, so we musn't access it.
508 * Write out the dnode's dirty buffers.
510 * NOTE: The dnode is kept in memory by being dirty. Once the
511 * dirty bit is cleared, it may be evicted. Beware of this!
514 dnode_sync(dnode_t
*dn
, dmu_tx_t
*tx
)
517 dnode_phys_t
*dnp
= dn
->dn_phys
;
518 int txgoff
= tx
->tx_txg
& TXG_MASK
;
519 list_t
*list
= &dn
->dn_dirty_records
[txgoff
];
521 ASSERT(dmu_tx_is_syncing(tx
));
522 ASSERT(dnp
->dn_type
!= DMU_OT_NONE
|| dn
->dn_allocated_txg
);
525 ASSERT(dn
->dn_dbuf
== NULL
|| arc_released(dn
->dn_dbuf
->db_buf
));
527 mutex_enter(&dn
->dn_mtx
);
528 if (dn
->dn_allocated_txg
== tx
->tx_txg
) {
529 /* The dnode is newly allocated or reallocated */
530 if (dnp
->dn_type
== DMU_OT_NONE
) {
531 /* this is a first alloc, not a realloc */
532 /* XXX shouldn't the phys already be zeroed? */
533 bzero(dnp
, DNODE_CORE_SIZE
);
537 if (dn
->dn_nblkptr
> dnp
->dn_nblkptr
) {
538 /* zero the new blkptrs we are gaining */
539 bzero(dnp
->dn_blkptr
+ dnp
->dn_nblkptr
,
541 (dn
->dn_nblkptr
- dnp
->dn_nblkptr
));
543 dnp
->dn_type
= dn
->dn_type
;
544 dnp
->dn_bonustype
= dn
->dn_bonustype
;
545 dnp
->dn_bonuslen
= dn
->dn_bonuslen
;
546 dnp
->dn_nblkptr
= dn
->dn_nblkptr
;
549 ASSERT(dnp
->dn_nlevels
> 1 ||
550 BP_IS_HOLE(&dnp
->dn_blkptr
[0]) ||
551 BP_GET_LSIZE(&dnp
->dn_blkptr
[0]) ==
552 dnp
->dn_datablkszsec
<< SPA_MINBLOCKSHIFT
);
554 if (dn
->dn_next_blksz
[txgoff
]) {
555 ASSERT(P2PHASE(dn
->dn_next_blksz
[txgoff
],
556 SPA_MINBLOCKSIZE
) == 0);
557 ASSERT(BP_IS_HOLE(&dnp
->dn_blkptr
[0]) ||
558 dn
->dn_maxblkid
== 0 || list_head(list
) != NULL
||
559 dn
->dn_next_blksz
[txgoff
] >> SPA_MINBLOCKSHIFT
==
560 dnp
->dn_datablkszsec
);
561 dnp
->dn_datablkszsec
=
562 dn
->dn_next_blksz
[txgoff
] >> SPA_MINBLOCKSHIFT
;
563 dn
->dn_next_blksz
[txgoff
] = 0;
566 if (dn
->dn_next_bonuslen
[txgoff
]) {
567 if (dn
->dn_next_bonuslen
[txgoff
] == DN_ZERO_BONUSLEN
)
568 dnp
->dn_bonuslen
= 0;
570 dnp
->dn_bonuslen
= dn
->dn_next_bonuslen
[txgoff
];
571 ASSERT(dnp
->dn_bonuslen
<= DN_MAX_BONUSLEN
);
572 dn
->dn_next_bonuslen
[txgoff
] = 0;
575 if (dn
->dn_next_indblkshift
[txgoff
]) {
576 ASSERT(dnp
->dn_nlevels
== 1);
577 dnp
->dn_indblkshift
= dn
->dn_next_indblkshift
[txgoff
];
578 dn
->dn_next_indblkshift
[txgoff
] = 0;
582 * Just take the live (open-context) values for checksum and compress.
583 * Strictly speaking it's a future leak, but nothing bad happens if we
584 * start using the new checksum or compress algorithm a little early.
586 dnp
->dn_checksum
= dn
->dn_checksum
;
587 dnp
->dn_compress
= dn
->dn_compress
;
589 mutex_exit(&dn
->dn_mtx
);
591 /* process all the "freed" ranges in the file */
592 while (rp
= avl_last(&dn
->dn_ranges
[txgoff
])) {
593 dnode_sync_free_range(dn
, rp
->fr_blkid
, rp
->fr_nblks
, tx
);
594 /* grab the mutex so we don't race with dnode_block_freed() */
595 mutex_enter(&dn
->dn_mtx
);
596 avl_remove(&dn
->dn_ranges
[txgoff
], rp
);
597 mutex_exit(&dn
->dn_mtx
);
598 kmem_free(rp
, sizeof (free_range_t
));
601 if (dn
->dn_free_txg
> 0 && dn
->dn_free_txg
<= tx
->tx_txg
) {
602 dnode_sync_free(dn
, tx
);
606 if (dn
->dn_next_nlevels
[txgoff
]) {
607 dnode_increase_indirection(dn
, tx
);
608 dn
->dn_next_nlevels
[txgoff
] = 0;
611 dbuf_sync_list(list
, tx
);
613 if (dn
->dn_object
!= DMU_META_DNODE_OBJECT
) {
614 ASSERT3P(list_head(list
), ==, NULL
);
615 dnode_rele(dn
, (void *)(uintptr_t)tx
->tx_txg
);
619 * Although we have dropped our reference to the dnode, it
620 * can't be evicted until its written, and we haven't yet
621 * initiated the IO for the dnode's dbuf.