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 (c) 2013 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/metaslab.h>
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
53 * There is one ZIL per file system. Its on-disk (pool) format consists
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
70 * See zil.h for more information about these fields.
72 zil_stats_t zil_stats
= {
73 { "zil_commit_count", KSTAT_DATA_UINT64
},
74 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
75 { "zil_itx_count", KSTAT_DATA_UINT64
},
76 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
77 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
78 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
79 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
80 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
81 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
82 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
83 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
84 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
85 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
88 static kstat_t
*zil_ksp
;
91 * Disable intent logging replay. This global ZIL switch affects all pools.
93 int zil_replay_disable
= 0;
96 * Tunable parameter for debugging or performance analysis. Setting
97 * zfs_nocacheflush will cause corruption on power loss if a volatile
98 * out-of-order write cache is enabled.
100 int zfs_nocacheflush
= 0;
102 static kmem_cache_t
*zil_lwb_cache
;
104 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
106 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
107 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
111 * ziltest is by and large an ugly hack, but very useful in
112 * checking replay without tedious work.
113 * When running ziltest we want to keep all itx's and so maintain
114 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
115 * We subtract TXG_CONCURRENT_STATES to allow for common code.
117 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
120 zil_bp_compare(const void *x1
, const void *x2
)
122 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
123 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
125 if (DVA_GET_VDEV(dva1
) < DVA_GET_VDEV(dva2
))
127 if (DVA_GET_VDEV(dva1
) > DVA_GET_VDEV(dva2
))
130 if (DVA_GET_OFFSET(dva1
) < DVA_GET_OFFSET(dva2
))
132 if (DVA_GET_OFFSET(dva1
) > DVA_GET_OFFSET(dva2
))
139 zil_bp_tree_init(zilog_t
*zilog
)
141 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
142 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
146 zil_bp_tree_fini(zilog_t
*zilog
)
148 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
152 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
153 kmem_free(zn
, sizeof (zil_bp_node_t
));
159 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
161 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
162 const dva_t
*dva
= BP_IDENTITY(bp
);
166 if (avl_find(t
, dva
, &where
) != NULL
)
167 return (SET_ERROR(EEXIST
));
169 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_PUSHPAGE
);
171 avl_insert(t
, zn
, where
);
176 static zil_header_t
*
177 zil_header_in_syncing_context(zilog_t
*zilog
)
179 return ((zil_header_t
*)zilog
->zl_header
);
183 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
185 zio_cksum_t
*zc
= &bp
->blk_cksum
;
187 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
188 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
189 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
190 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
194 * Read a log block and make sure it's valid.
197 zil_read_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, blkptr_t
*nbp
, void *dst
,
200 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
201 uint32_t aflags
= ARC_WAIT
;
202 arc_buf_t
*abuf
= NULL
;
206 if (zilog
->zl_header
->zh_claim_txg
== 0)
207 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
209 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
210 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
212 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
213 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
215 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
216 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
219 zio_cksum_t cksum
= bp
->blk_cksum
;
222 * Validate the checksummed log block.
224 * Sequence numbers should be... sequential. The checksum
225 * verifier for the next block should be bp's checksum plus 1.
227 * Also check the log chain linkage and size used.
229 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
231 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
232 zil_chain_t
*zilc
= abuf
->b_data
;
233 char *lr
= (char *)(zilc
+ 1);
234 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
236 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
237 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
238 error
= SET_ERROR(ECKSUM
);
241 *end
= (char *)dst
+ len
;
242 *nbp
= zilc
->zc_next_blk
;
245 char *lr
= abuf
->b_data
;
246 uint64_t size
= BP_GET_LSIZE(bp
);
247 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
249 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
250 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
251 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
252 error
= SET_ERROR(ECKSUM
);
254 bcopy(lr
, dst
, zilc
->zc_nused
);
255 *end
= (char *)dst
+ zilc
->zc_nused
;
256 *nbp
= zilc
->zc_next_blk
;
260 VERIFY(arc_buf_remove_ref(abuf
, &abuf
));
267 * Read a TX_WRITE log data block.
270 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
272 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
273 const blkptr_t
*bp
= &lr
->lr_blkptr
;
274 uint32_t aflags
= ARC_WAIT
;
275 arc_buf_t
*abuf
= NULL
;
279 if (BP_IS_HOLE(bp
)) {
281 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
285 if (zilog
->zl_header
->zh_claim_txg
== 0)
286 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
288 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
289 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
291 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
292 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
296 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
297 (void) arc_buf_remove_ref(abuf
, &abuf
);
304 * Parse the intent log, and call parse_func for each valid record within.
307 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
308 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
)
310 const zil_header_t
*zh
= zilog
->zl_header
;
311 boolean_t claimed
= !!zh
->zh_claim_txg
;
312 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
313 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
314 uint64_t max_blk_seq
= 0;
315 uint64_t max_lr_seq
= 0;
316 uint64_t blk_count
= 0;
317 uint64_t lr_count
= 0;
318 blkptr_t blk
, next_blk
;
322 bzero(&next_blk
, sizeof (blkptr_t
));
325 * Old logs didn't record the maximum zh_claim_lr_seq.
327 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
328 claim_lr_seq
= UINT64_MAX
;
331 * Starting at the block pointed to by zh_log we read the log chain.
332 * For each block in the chain we strongly check that block to
333 * ensure its validity. We stop when an invalid block is found.
334 * For each block pointer in the chain we call parse_blk_func().
335 * For each record in each valid block we call parse_lr_func().
336 * If the log has been claimed, stop if we encounter a sequence
337 * number greater than the highest claimed sequence number.
339 lrbuf
= zio_buf_alloc(SPA_MAXBLOCKSIZE
);
340 zil_bp_tree_init(zilog
);
342 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
343 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
347 if (blk_seq
> claim_blk_seq
)
349 if ((error
= parse_blk_func(zilog
, &blk
, arg
, txg
)) != 0)
351 ASSERT3U(max_blk_seq
, <, blk_seq
);
352 max_blk_seq
= blk_seq
;
355 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
358 error
= zil_read_log_block(zilog
, &blk
, &next_blk
, lrbuf
, &end
);
362 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
363 lr_t
*lr
= (lr_t
*)lrp
;
364 reclen
= lr
->lrc_reclen
;
365 ASSERT3U(reclen
, >=, sizeof (lr_t
));
366 if (lr
->lrc_seq
> claim_lr_seq
)
368 if ((error
= parse_lr_func(zilog
, lr
, arg
, txg
)) != 0)
370 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
371 max_lr_seq
= lr
->lrc_seq
;
376 zilog
->zl_parse_error
= error
;
377 zilog
->zl_parse_blk_seq
= max_blk_seq
;
378 zilog
->zl_parse_lr_seq
= max_lr_seq
;
379 zilog
->zl_parse_blk_count
= blk_count
;
380 zilog
->zl_parse_lr_count
= lr_count
;
382 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
383 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
));
385 zil_bp_tree_fini(zilog
);
386 zio_buf_free(lrbuf
, SPA_MAXBLOCKSIZE
);
392 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
395 * Claim log block if not already committed and not already claimed.
396 * If tx == NULL, just verify that the block is claimable.
398 if (bp
->blk_birth
< first_txg
|| zil_bp_tree_add(zilog
, bp
) != 0)
401 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
402 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
403 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
407 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
409 lr_write_t
*lr
= (lr_write_t
*)lrc
;
412 if (lrc
->lrc_txtype
!= TX_WRITE
)
416 * If the block is not readable, don't claim it. This can happen
417 * in normal operation when a log block is written to disk before
418 * some of the dmu_sync() blocks it points to. In this case, the
419 * transaction cannot have been committed to anyone (we would have
420 * waited for all writes to be stable first), so it is semantically
421 * correct to declare this the end of the log.
423 if (lr
->lr_blkptr
.blk_birth
>= first_txg
&&
424 (error
= zil_read_log_data(zilog
, lr
, NULL
)) != 0)
426 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
431 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
433 zio_free_zil(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
439 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
441 lr_write_t
*lr
= (lr_write_t
*)lrc
;
442 blkptr_t
*bp
= &lr
->lr_blkptr
;
445 * If we previously claimed it, we need to free it.
447 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
448 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0)
449 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
455 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, uint64_t txg
, boolean_t fastwrite
)
459 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_PUSHPAGE
);
460 lwb
->lwb_zilog
= zilog
;
462 lwb
->lwb_fastwrite
= fastwrite
;
463 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
464 lwb
->lwb_max_txg
= txg
;
467 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
468 lwb
->lwb_nused
= sizeof (zil_chain_t
);
469 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
472 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
475 mutex_enter(&zilog
->zl_lock
);
476 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
477 mutex_exit(&zilog
->zl_lock
);
483 * Called when we create in-memory log transactions so that we know
484 * to cleanup the itxs at the end of spa_sync().
487 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
489 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
490 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
492 if (dsl_dataset_is_snapshot(ds
))
493 panic("dirtying snapshot!");
495 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
496 /* up the hold count until we can be written out */
497 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
502 zilog_is_dirty(zilog_t
*zilog
)
504 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
507 for (t
= 0; t
< TXG_SIZE
; t
++) {
508 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
515 * Create an on-disk intent log.
518 zil_create(zilog_t
*zilog
)
520 const zil_header_t
*zh
= zilog
->zl_header
;
526 boolean_t fastwrite
= FALSE
;
529 * Wait for any previous destroy to complete.
531 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
533 ASSERT(zh
->zh_claim_txg
== 0);
534 ASSERT(zh
->zh_replay_seq
== 0);
539 * Allocate an initial log block if:
540 * - there isn't one already
541 * - the existing block is the wrong endianess
543 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
544 tx
= dmu_tx_create(zilog
->zl_os
);
545 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
546 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
547 txg
= dmu_tx_get_txg(tx
);
549 if (!BP_IS_HOLE(&blk
)) {
550 zio_free_zil(zilog
->zl_spa
, txg
, &blk
);
554 error
= zio_alloc_zil(zilog
->zl_spa
, txg
, &blk
,
555 ZIL_MIN_BLKSZ
, B_TRUE
);
559 zil_init_log_chain(zilog
, &blk
);
563 * Allocate a log write buffer (lwb) for the first log block.
566 lwb
= zil_alloc_lwb(zilog
, &blk
, txg
, fastwrite
);
569 * If we just allocated the first log block, commit our transaction
570 * and wait for zil_sync() to stuff the block poiner into zh_log.
571 * (zh is part of the MOS, so we cannot modify it in open context.)
575 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
578 ASSERT(bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
584 * In one tx, free all log blocks and clear the log header.
585 * If keep_first is set, then we're replaying a log with no content.
586 * We want to keep the first block, however, so that the first
587 * synchronous transaction doesn't require a txg_wait_synced()
588 * in zil_create(). We don't need to txg_wait_synced() here either
589 * when keep_first is set, because both zil_create() and zil_destroy()
590 * will wait for any in-progress destroys to complete.
593 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
595 const zil_header_t
*zh
= zilog
->zl_header
;
601 * Wait for any previous destroy to complete.
603 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
605 zilog
->zl_old_header
= *zh
; /* debugging aid */
607 if (BP_IS_HOLE(&zh
->zh_log
))
610 tx
= dmu_tx_create(zilog
->zl_os
);
611 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
612 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
613 txg
= dmu_tx_get_txg(tx
);
615 mutex_enter(&zilog
->zl_lock
);
617 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
618 zilog
->zl_destroy_txg
= txg
;
619 zilog
->zl_keep_first
= keep_first
;
621 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
622 ASSERT(zh
->zh_claim_txg
== 0);
624 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
625 ASSERT(lwb
->lwb_zio
== NULL
);
626 if (lwb
->lwb_fastwrite
)
627 metaslab_fastwrite_unmark(zilog
->zl_spa
,
629 list_remove(&zilog
->zl_lwb_list
, lwb
);
630 if (lwb
->lwb_buf
!= NULL
)
631 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
632 zio_free_zil(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
633 kmem_cache_free(zil_lwb_cache
, lwb
);
635 } else if (!keep_first
) {
636 zil_destroy_sync(zilog
, tx
);
638 mutex_exit(&zilog
->zl_lock
);
644 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
646 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
647 (void) zil_parse(zilog
, zil_free_log_block
,
648 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
);
652 zil_claim(const char *osname
, void *txarg
)
654 dmu_tx_t
*tx
= txarg
;
655 uint64_t first_txg
= dmu_tx_get_txg(tx
);
661 error
= dmu_objset_own(osname
, DMU_OST_ANY
, B_FALSE
, FTAG
, &os
);
663 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
667 zilog
= dmu_objset_zil(os
);
668 zh
= zil_header_in_syncing_context(zilog
);
670 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
) {
671 if (!BP_IS_HOLE(&zh
->zh_log
))
672 zio_free_zil(zilog
->zl_spa
, first_txg
, &zh
->zh_log
);
673 BP_ZERO(&zh
->zh_log
);
674 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
675 dmu_objset_disown(os
, FTAG
);
680 * Claim all log blocks if we haven't already done so, and remember
681 * the highest claimed sequence number. This ensures that if we can
682 * read only part of the log now (e.g. due to a missing device),
683 * but we can read the entire log later, we will not try to replay
684 * or destroy beyond the last block we successfully claimed.
686 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
687 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
688 (void) zil_parse(zilog
, zil_claim_log_block
,
689 zil_claim_log_record
, tx
, first_txg
);
690 zh
->zh_claim_txg
= first_txg
;
691 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
692 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
693 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
694 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
695 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
696 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
699 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
700 dmu_objset_disown(os
, FTAG
);
705 * Check the log by walking the log chain.
706 * Checksum errors are ok as they indicate the end of the chain.
707 * Any other error (no device or read failure) returns an error.
710 zil_check_log_chain(const char *osname
, void *tx
)
719 error
= dmu_objset_hold(osname
, FTAG
, &os
);
721 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
725 zilog
= dmu_objset_zil(os
);
726 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
729 * Check the first block and determine if it's on a log device
730 * which may have been removed or faulted prior to loading this
731 * pool. If so, there's no point in checking the rest of the log
732 * as its content should have already been synced to the pool.
734 if (!BP_IS_HOLE(bp
)) {
736 boolean_t valid
= B_TRUE
;
738 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
739 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
740 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
741 valid
= vdev_log_state_valid(vd
);
742 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
745 dmu_objset_rele(os
, FTAG
);
751 * Because tx == NULL, zil_claim_log_block() will not actually claim
752 * any blocks, but just determine whether it is possible to do so.
753 * In addition to checking the log chain, zil_claim_log_block()
754 * will invoke zio_claim() with a done func of spa_claim_notify(),
755 * which will update spa_max_claim_txg. See spa_load() for details.
757 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
758 zilog
->zl_header
->zh_claim_txg
? -1ULL : spa_first_txg(os
->os_spa
));
760 dmu_objset_rele(os
, FTAG
);
762 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
766 zil_vdev_compare(const void *x1
, const void *x2
)
768 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
769 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
780 zil_add_block(zilog_t
*zilog
, const blkptr_t
*bp
)
782 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
784 zil_vdev_node_t
*zv
, zvsearch
;
785 int ndvas
= BP_GET_NDVAS(bp
);
788 if (zfs_nocacheflush
)
791 ASSERT(zilog
->zl_writer
);
794 * Even though we're zl_writer, we still need a lock because the
795 * zl_get_data() callbacks may have dmu_sync() done callbacks
796 * that will run concurrently.
798 mutex_enter(&zilog
->zl_vdev_lock
);
799 for (i
= 0; i
< ndvas
; i
++) {
800 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
801 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
802 zv
= kmem_alloc(sizeof (*zv
), KM_PUSHPAGE
);
803 zv
->zv_vdev
= zvsearch
.zv_vdev
;
804 avl_insert(t
, zv
, where
);
807 mutex_exit(&zilog
->zl_vdev_lock
);
811 zil_flush_vdevs(zilog_t
*zilog
)
813 spa_t
*spa
= zilog
->zl_spa
;
814 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
819 ASSERT(zilog
->zl_writer
);
822 * We don't need zl_vdev_lock here because we're the zl_writer,
823 * and all zl_get_data() callbacks are done.
825 if (avl_numnodes(t
) == 0)
828 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
830 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
832 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
833 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
836 kmem_free(zv
, sizeof (*zv
));
840 * Wait for all the flushes to complete. Not all devices actually
841 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
843 (void) zio_wait(zio
);
845 spa_config_exit(spa
, SCL_STATE
, FTAG
);
849 * Function called when a log block write completes
852 zil_lwb_write_done(zio_t
*zio
)
854 lwb_t
*lwb
= zio
->io_private
;
855 zilog_t
*zilog
= lwb
->lwb_zilog
;
856 dmu_tx_t
*tx
= lwb
->lwb_tx
;
858 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
859 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
860 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
861 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
862 ASSERT(!BP_IS_GANG(zio
->io_bp
));
863 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
864 ASSERT(zio
->io_bp
->blk_fill
== 0);
867 * Ensure the lwb buffer pointer is cleared before releasing
868 * the txg. If we have had an allocation failure and
869 * the txg is waiting to sync then we want want zil_sync()
870 * to remove the lwb so that it's not picked up as the next new
871 * one in zil_commit_writer(). zil_sync() will only remove
872 * the lwb if lwb_buf is null.
874 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
875 mutex_enter(&zilog
->zl_lock
);
877 lwb
->lwb_fastwrite
= FALSE
;
880 mutex_exit(&zilog
->zl_lock
);
883 * Now that we've written this log block, we have a stable pointer
884 * to the next block in the chain, so it's OK to let the txg in
885 * which we allocated the next block sync.
891 * Initialize the io for a log block.
894 zil_lwb_write_init(zilog_t
*zilog
, lwb_t
*lwb
)
898 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
899 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
900 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
902 if (zilog
->zl_root_zio
== NULL
) {
903 zilog
->zl_root_zio
= zio_root(zilog
->zl_spa
, NULL
, NULL
,
907 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
908 mutex_enter(&zilog
->zl_lock
);
909 if (lwb
->lwb_zio
== NULL
) {
910 if (!lwb
->lwb_fastwrite
) {
911 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
912 lwb
->lwb_fastwrite
= 1;
914 lwb
->lwb_zio
= zio_rewrite(zilog
->zl_root_zio
, zilog
->zl_spa
,
915 0, &lwb
->lwb_blk
, lwb
->lwb_buf
, BP_GET_LSIZE(&lwb
->lwb_blk
),
916 zil_lwb_write_done
, lwb
, ZIO_PRIORITY_SYNC_WRITE
,
917 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
|
918 ZIO_FLAG_FASTWRITE
, &zb
);
920 mutex_exit(&zilog
->zl_lock
);
924 * Define a limited set of intent log block sizes.
926 * These must be a multiple of 4KB. Note only the amount used (again
927 * aligned to 4KB) actually gets written. However, we can't always just
928 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
930 uint64_t zil_block_buckets
[] = {
931 4096, /* non TX_WRITE */
932 8192+4096, /* data base */
933 32*1024 + 4096, /* NFS writes */
938 * Use the slog as long as the current commit size is less than the
939 * limit or the total list size is less than 2X the limit. Limit
940 * checking is disabled by setting zil_slog_limit to UINT64_MAX.
942 unsigned long zil_slog_limit
= 1024 * 1024;
943 #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \
944 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))
947 * Start a log block write and advance to the next log block.
948 * Calls are serialized.
951 zil_lwb_write_start(zilog_t
*zilog
, lwb_t
*lwb
)
955 spa_t
*spa
= zilog
->zl_spa
;
959 uint64_t zil_blksz
, wsz
;
963 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
964 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
965 bp
= &zilc
->zc_next_blk
;
967 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
968 bp
= &zilc
->zc_next_blk
;
971 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
974 * Allocate the next block and save its address in this block
975 * before writing it in order to establish the log chain.
976 * Note that if the allocation of nlwb synced before we wrote
977 * the block that points at it (lwb), we'd leak it if we crashed.
978 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
979 * We dirty the dataset to ensure that zil_sync() will be called
980 * to clean up in the event of allocation failure or I/O failure.
982 tx
= dmu_tx_create(zilog
->zl_os
);
983 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
984 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
985 txg
= dmu_tx_get_txg(tx
);
990 * Log blocks are pre-allocated. Here we select the size of the next
991 * block, based on size used in the last block.
992 * - first find the smallest bucket that will fit the block from a
993 * limited set of block sizes. This is because it's faster to write
994 * blocks allocated from the same metaslab as they are adjacent or
996 * - next find the maximum from the new suggested size and an array of
997 * previous sizes. This lessens a picket fence effect of wrongly
998 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1001 * Note we only write what is used, but we can't just allocate
1002 * the maximum block size because we can exhaust the available
1005 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1006 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
1008 zil_blksz
= zil_block_buckets
[i
];
1009 if (zil_blksz
== UINT64_MAX
)
1010 zil_blksz
= SPA_MAXBLOCKSIZE
;
1011 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1012 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1013 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1014 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1017 use_slog
= USE_SLOG(zilog
);
1018 error
= zio_alloc_zil(spa
, txg
, bp
, zil_blksz
,
1021 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count
);
1022 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes
, lwb
->lwb_nused
);
1024 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count
);
1025 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes
, lwb
->lwb_nused
);
1028 ASSERT3U(bp
->blk_birth
, ==, txg
);
1029 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1030 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1033 * Allocate a new log write buffer (lwb).
1035 nlwb
= zil_alloc_lwb(zilog
, bp
, txg
, TRUE
);
1037 /* Record the block for later vdev flushing */
1038 zil_add_block(zilog
, &lwb
->lwb_blk
);
1041 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1042 /* For Slim ZIL only write what is used. */
1043 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1044 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1045 zio_shrink(lwb
->lwb_zio
, wsz
);
1052 zilc
->zc_nused
= lwb
->lwb_nused
;
1053 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1056 * clear unused data for security
1058 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
1060 zio_nowait(lwb
->lwb_zio
); /* Kick off the write for the old log block */
1063 * If there was an allocation failure then nlwb will be null which
1064 * forces a txg_wait_synced().
1070 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1072 lr_t
*lrc
= &itx
->itx_lr
; /* common log record */
1073 lr_write_t
*lrw
= (lr_write_t
*)lrc
;
1075 uint64_t txg
= lrc
->lrc_txg
;
1076 uint64_t reclen
= lrc
->lrc_reclen
;
1082 ASSERT(lwb
->lwb_buf
!= NULL
);
1083 ASSERT(zilog_is_dirty(zilog
) ||
1084 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
1086 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
)
1087 dlen
= P2ROUNDUP_TYPED(
1088 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1090 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1092 zil_lwb_write_init(zilog
, lwb
);
1095 * If this record won't fit in the current log block, start a new one.
1097 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1098 lwb
= zil_lwb_write_start(zilog
, lwb
);
1101 zil_lwb_write_init(zilog
, lwb
);
1102 ASSERT(LWB_EMPTY(lwb
));
1103 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1104 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1109 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1110 bcopy(lrc
, lr_buf
, reclen
);
1111 lrc
= (lr_t
*)lr_buf
;
1112 lrw
= (lr_write_t
*)lrc
;
1114 ZIL_STAT_BUMP(zil_itx_count
);
1117 * If it's a write, fetch the data or get its blkptr as appropriate.
1119 if (lrc
->lrc_txtype
== TX_WRITE
) {
1120 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1121 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1122 if (itx
->itx_wr_state
== WR_COPIED
) {
1123 ZIL_STAT_BUMP(zil_itx_copied_count
);
1124 ZIL_STAT_INCR(zil_itx_copied_bytes
, lrw
->lr_length
);
1130 ASSERT(itx
->itx_wr_state
== WR_NEED_COPY
);
1131 dbuf
= lr_buf
+ reclen
;
1132 lrw
->lr_common
.lrc_reclen
+= dlen
;
1133 ZIL_STAT_BUMP(zil_itx_needcopy_count
);
1134 ZIL_STAT_INCR(zil_itx_needcopy_bytes
,
1137 ASSERT(itx
->itx_wr_state
== WR_INDIRECT
);
1139 ZIL_STAT_BUMP(zil_itx_indirect_count
);
1140 ZIL_STAT_INCR(zil_itx_indirect_bytes
,
1143 error
= zilog
->zl_get_data(
1144 itx
->itx_private
, lrw
, dbuf
, lwb
->lwb_zio
);
1146 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1150 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1158 * We're actually making an entry, so update lrc_seq to be the
1159 * log record sequence number. Note that this is generally not
1160 * equal to the itx sequence number because not all transactions
1161 * are synchronous, and sometimes spa_sync() gets there first.
1163 lrc
->lrc_seq
= ++zilog
->zl_lr_seq
; /* we are single threaded */
1164 lwb
->lwb_nused
+= reclen
+ dlen
;
1165 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1166 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1167 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1173 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1177 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1179 itx
= kmem_alloc(offsetof(itx_t
, itx_lr
) + lrsize
,
1180 KM_PUSHPAGE
| KM_NODEBUG
);
1181 itx
->itx_lr
.lrc_txtype
= txtype
;
1182 itx
->itx_lr
.lrc_reclen
= lrsize
;
1183 itx
->itx_sod
= lrsize
; /* if write & WR_NEED_COPY will be increased */
1184 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1185 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1186 itx
->itx_callback
= NULL
;
1187 itx
->itx_callback_data
= NULL
;
1193 zil_itx_destroy(itx_t
*itx
)
1195 kmem_free(itx
, offsetof(itx_t
, itx_lr
) + itx
->itx_lr
.lrc_reclen
);
1199 * Free up the sync and async itxs. The itxs_t has already been detached
1200 * so no locks are needed.
1203 zil_itxg_clean(itxs_t
*itxs
)
1209 itx_async_node_t
*ian
;
1211 list
= &itxs
->i_sync_list
;
1212 while ((itx
= list_head(list
)) != NULL
) {
1213 if (itx
->itx_callback
!= NULL
)
1214 itx
->itx_callback(itx
->itx_callback_data
);
1215 list_remove(list
, itx
);
1216 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1217 itx
->itx_lr
.lrc_reclen
);
1221 t
= &itxs
->i_async_tree
;
1222 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1223 list
= &ian
->ia_list
;
1224 while ((itx
= list_head(list
)) != NULL
) {
1225 if (itx
->itx_callback
!= NULL
)
1226 itx
->itx_callback(itx
->itx_callback_data
);
1227 list_remove(list
, itx
);
1228 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1229 itx
->itx_lr
.lrc_reclen
);
1232 kmem_free(ian
, sizeof (itx_async_node_t
));
1236 kmem_free(itxs
, sizeof (itxs_t
));
1240 zil_aitx_compare(const void *x1
, const void *x2
)
1242 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1243 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1254 * Remove all async itx with the given oid.
1257 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1260 itx_async_node_t
*ian
;
1267 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1269 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1272 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1274 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1275 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1277 mutex_enter(&itxg
->itxg_lock
);
1278 if (itxg
->itxg_txg
!= txg
) {
1279 mutex_exit(&itxg
->itxg_lock
);
1284 * Locate the object node and append its list.
1286 t
= &itxg
->itxg_itxs
->i_async_tree
;
1287 ian
= avl_find(t
, &oid
, &where
);
1289 list_move_tail(&clean_list
, &ian
->ia_list
);
1290 mutex_exit(&itxg
->itxg_lock
);
1292 while ((itx
= list_head(&clean_list
)) != NULL
) {
1293 if (itx
->itx_callback
!= NULL
)
1294 itx
->itx_callback(itx
->itx_callback_data
);
1295 list_remove(&clean_list
, itx
);
1296 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1297 itx
->itx_lr
.lrc_reclen
);
1299 list_destroy(&clean_list
);
1303 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1307 itxs_t
*itxs
, *clean
= NULL
;
1310 * Object ids can be re-instantiated in the next txg so
1311 * remove any async transactions to avoid future leaks.
1312 * This can happen if a fsync occurs on the re-instantiated
1313 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1314 * the new file data and flushes a write record for the old object.
1316 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_REMOVE
)
1317 zil_remove_async(zilog
, itx
->itx_oid
);
1320 * Ensure the data of a renamed file is committed before the rename.
1322 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1323 zil_async_to_sync(zilog
, itx
->itx_oid
);
1325 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1328 txg
= dmu_tx_get_txg(tx
);
1330 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1331 mutex_enter(&itxg
->itxg_lock
);
1332 itxs
= itxg
->itxg_itxs
;
1333 if (itxg
->itxg_txg
!= txg
) {
1336 * The zil_clean callback hasn't got around to cleaning
1337 * this itxg. Save the itxs for release below.
1338 * This should be rare.
1340 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1342 clean
= itxg
->itxg_itxs
;
1344 ASSERT(itxg
->itxg_sod
== 0);
1345 itxg
->itxg_txg
= txg
;
1346 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
1349 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1350 offsetof(itx_t
, itx_node
));
1351 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1352 sizeof (itx_async_node_t
),
1353 offsetof(itx_async_node_t
, ia_node
));
1355 if (itx
->itx_sync
) {
1356 list_insert_tail(&itxs
->i_sync_list
, itx
);
1357 atomic_add_64(&zilog
->zl_itx_list_sz
, itx
->itx_sod
);
1358 itxg
->itxg_sod
+= itx
->itx_sod
;
1360 avl_tree_t
*t
= &itxs
->i_async_tree
;
1361 uint64_t foid
= ((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
;
1362 itx_async_node_t
*ian
;
1365 ian
= avl_find(t
, &foid
, &where
);
1367 ian
= kmem_alloc(sizeof (itx_async_node_t
),
1369 list_create(&ian
->ia_list
, sizeof (itx_t
),
1370 offsetof(itx_t
, itx_node
));
1371 ian
->ia_foid
= foid
;
1372 avl_insert(t
, ian
, where
);
1374 list_insert_tail(&ian
->ia_list
, itx
);
1377 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1378 zilog_dirty(zilog
, txg
);
1379 mutex_exit(&itxg
->itxg_lock
);
1381 /* Release the old itxs now we've dropped the lock */
1383 zil_itxg_clean(clean
);
1387 * If there are any in-memory intent log transactions which have now been
1388 * synced then start up a taskq to free them. We should only do this after we
1389 * have written out the uberblocks (i.e. txg has been comitted) so that
1390 * don't inadvertently clean out in-memory log records that would be required
1394 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
1396 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
1399 mutex_enter(&itxg
->itxg_lock
);
1400 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
1401 mutex_exit(&itxg
->itxg_lock
);
1404 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
1405 ASSERT(itxg
->itxg_txg
!= 0);
1406 ASSERT(zilog
->zl_clean_taskq
!= NULL
);
1407 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1409 clean_me
= itxg
->itxg_itxs
;
1410 itxg
->itxg_itxs
= NULL
;
1412 mutex_exit(&itxg
->itxg_lock
);
1414 * Preferably start a task queue to free up the old itxs but
1415 * if taskq_dispatch can't allocate resources to do that then
1416 * free it in-line. This should be rare. Note, using TQ_SLEEP
1417 * created a bad performance problem.
1419 if (taskq_dispatch(zilog
->zl_clean_taskq
,
1420 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
) == 0)
1421 zil_itxg_clean(clean_me
);
1425 * Get the list of itxs to commit into zl_itx_commit_list.
1428 zil_get_commit_list(zilog_t
*zilog
)
1431 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
1432 uint64_t push_sod
= 0;
1434 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1437 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1439 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1440 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1442 mutex_enter(&itxg
->itxg_lock
);
1443 if (itxg
->itxg_txg
!= txg
) {
1444 mutex_exit(&itxg
->itxg_lock
);
1448 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
1449 push_sod
+= itxg
->itxg_sod
;
1452 mutex_exit(&itxg
->itxg_lock
);
1454 atomic_add_64(&zilog
->zl_itx_list_sz
, -push_sod
);
1458 * Move the async itxs for a specified object to commit into sync lists.
1461 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
1464 itx_async_node_t
*ian
;
1468 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1471 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1473 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1474 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1476 mutex_enter(&itxg
->itxg_lock
);
1477 if (itxg
->itxg_txg
!= txg
) {
1478 mutex_exit(&itxg
->itxg_lock
);
1483 * If a foid is specified then find that node and append its
1484 * list. Otherwise walk the tree appending all the lists
1485 * to the sync list. We add to the end rather than the
1486 * beginning to ensure the create has happened.
1488 t
= &itxg
->itxg_itxs
->i_async_tree
;
1490 ian
= avl_find(t
, &foid
, &where
);
1492 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1496 void *cookie
= NULL
;
1498 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1499 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1501 list_destroy(&ian
->ia_list
);
1502 kmem_free(ian
, sizeof (itx_async_node_t
));
1505 mutex_exit(&itxg
->itxg_lock
);
1510 zil_commit_writer(zilog_t
*zilog
)
1515 spa_t
*spa
= zilog
->zl_spa
;
1518 ASSERT(zilog
->zl_root_zio
== NULL
);
1520 mutex_exit(&zilog
->zl_lock
);
1522 zil_get_commit_list(zilog
);
1525 * Return if there's nothing to commit before we dirty the fs by
1526 * calling zil_create().
1528 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
) {
1529 mutex_enter(&zilog
->zl_lock
);
1533 if (zilog
->zl_suspend
) {
1536 lwb
= list_tail(&zilog
->zl_lwb_list
);
1538 lwb
= zil_create(zilog
);
1541 DTRACE_PROBE1(zil__cw1
, zilog_t
*, zilog
);
1542 for (itx
= list_head(&zilog
->zl_itx_commit_list
); itx
!= NULL
;
1543 itx
= list_next(&zilog
->zl_itx_commit_list
, itx
)) {
1544 txg
= itx
->itx_lr
.lrc_txg
;
1547 if (txg
> spa_last_synced_txg(spa
) || txg
> spa_freeze_txg(spa
))
1548 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
1550 DTRACE_PROBE1(zil__cw2
, zilog_t
*, zilog
);
1552 /* write the last block out */
1553 if (lwb
!= NULL
&& lwb
->lwb_zio
!= NULL
)
1554 lwb
= zil_lwb_write_start(zilog
, lwb
);
1556 zilog
->zl_cur_used
= 0;
1559 * Wait if necessary for the log blocks to be on stable storage.
1561 if (zilog
->zl_root_zio
) {
1562 error
= zio_wait(zilog
->zl_root_zio
);
1563 zilog
->zl_root_zio
= NULL
;
1564 zil_flush_vdevs(zilog
);
1567 if (error
|| lwb
== NULL
)
1568 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
1570 while ((itx
= list_head(&zilog
->zl_itx_commit_list
))) {
1571 txg
= itx
->itx_lr
.lrc_txg
;
1574 if (itx
->itx_callback
!= NULL
)
1575 itx
->itx_callback(itx
->itx_callback_data
);
1576 list_remove(&zilog
->zl_itx_commit_list
, itx
);
1577 kmem_free(itx
, offsetof(itx_t
, itx_lr
)
1578 + itx
->itx_lr
.lrc_reclen
);
1581 mutex_enter(&zilog
->zl_lock
);
1584 * Remember the highest committed log sequence number for ztest.
1585 * We only update this value when all the log writes succeeded,
1586 * because ztest wants to ASSERT that it got the whole log chain.
1588 if (error
== 0 && lwb
!= NULL
)
1589 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1593 * Commit zfs transactions to stable storage.
1594 * If foid is 0 push out all transactions, otherwise push only those
1595 * for that object or might reference that object.
1597 * itxs are committed in batches. In a heavily stressed zil there will be
1598 * a commit writer thread who is writing out a bunch of itxs to the log
1599 * for a set of committing threads (cthreads) in the same batch as the writer.
1600 * Those cthreads are all waiting on the same cv for that batch.
1602 * There will also be a different and growing batch of threads that are
1603 * waiting to commit (qthreads). When the committing batch completes
1604 * a transition occurs such that the cthreads exit and the qthreads become
1605 * cthreads. One of the new cthreads becomes the writer thread for the
1606 * batch. Any new threads arriving become new qthreads.
1608 * Only 2 condition variables are needed and there's no transition
1609 * between the two cvs needed. They just flip-flop between qthreads
1612 * Using this scheme we can efficiently wakeup up only those threads
1613 * that have been committed.
1616 zil_commit(zilog_t
*zilog
, uint64_t foid
)
1620 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
1623 ZIL_STAT_BUMP(zil_commit_count
);
1625 /* move the async itxs for the foid to the sync queues */
1626 zil_async_to_sync(zilog
, foid
);
1628 mutex_enter(&zilog
->zl_lock
);
1629 mybatch
= zilog
->zl_next_batch
;
1630 while (zilog
->zl_writer
) {
1631 cv_wait(&zilog
->zl_cv_batch
[mybatch
& 1], &zilog
->zl_lock
);
1632 if (mybatch
<= zilog
->zl_com_batch
) {
1633 mutex_exit(&zilog
->zl_lock
);
1638 zilog
->zl_next_batch
++;
1639 zilog
->zl_writer
= B_TRUE
;
1640 ZIL_STAT_BUMP(zil_commit_writer_count
);
1641 zil_commit_writer(zilog
);
1642 zilog
->zl_com_batch
= mybatch
;
1643 zilog
->zl_writer
= B_FALSE
;
1645 /* wake up one thread to become the next writer */
1646 cv_signal(&zilog
->zl_cv_batch
[(mybatch
+1) & 1]);
1648 /* wake up all threads waiting for this batch to be committed */
1649 cv_broadcast(&zilog
->zl_cv_batch
[mybatch
& 1]);
1651 mutex_exit(&zilog
->zl_lock
);
1655 * Called in syncing context to free committed log blocks and update log header.
1658 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1660 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1661 uint64_t txg
= dmu_tx_get_txg(tx
);
1662 spa_t
*spa
= zilog
->zl_spa
;
1663 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
1667 * We don't zero out zl_destroy_txg, so make sure we don't try
1668 * to destroy it twice.
1670 if (spa_sync_pass(spa
) != 1)
1673 mutex_enter(&zilog
->zl_lock
);
1675 ASSERT(zilog
->zl_stop_sync
== 0);
1677 if (*replayed_seq
!= 0) {
1678 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
1679 zh
->zh_replay_seq
= *replayed_seq
;
1683 if (zilog
->zl_destroy_txg
== txg
) {
1684 blkptr_t blk
= zh
->zh_log
;
1686 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
1688 bzero(zh
, sizeof (zil_header_t
));
1689 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
1691 if (zilog
->zl_keep_first
) {
1693 * If this block was part of log chain that couldn't
1694 * be claimed because a device was missing during
1695 * zil_claim(), but that device later returns,
1696 * then this block could erroneously appear valid.
1697 * To guard against this, assign a new GUID to the new
1698 * log chain so it doesn't matter what blk points to.
1700 zil_init_log_chain(zilog
, &blk
);
1705 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
1706 zh
->zh_log
= lwb
->lwb_blk
;
1707 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
1710 ASSERT(lwb
->lwb_zio
== NULL
);
1712 list_remove(&zilog
->zl_lwb_list
, lwb
);
1713 zio_free_zil(spa
, txg
, &lwb
->lwb_blk
);
1714 kmem_cache_free(zil_lwb_cache
, lwb
);
1717 * If we don't have anything left in the lwb list then
1718 * we've had an allocation failure and we need to zero
1719 * out the zil_header blkptr so that we don't end
1720 * up freeing the same block twice.
1722 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
1723 BP_ZERO(&zh
->zh_log
);
1727 * Remove fastwrite on any blocks that have been pre-allocated for
1728 * the next commit. This prevents fastwrite counter pollution by
1729 * unused, long-lived LWBs.
1731 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
1732 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_zio
) {
1733 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1734 lwb
->lwb_fastwrite
= 0;
1738 mutex_exit(&zilog
->zl_lock
);
1744 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
1745 sizeof (struct lwb
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
1747 zil_ksp
= kstat_create("zfs", 0, "zil", "misc",
1748 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
1749 KSTAT_FLAG_VIRTUAL
);
1751 if (zil_ksp
!= NULL
) {
1752 zil_ksp
->ks_data
= &zil_stats
;
1753 kstat_install(zil_ksp
);
1760 kmem_cache_destroy(zil_lwb_cache
);
1762 if (zil_ksp
!= NULL
) {
1763 kstat_delete(zil_ksp
);
1769 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
1771 zilog
->zl_sync
= sync
;
1775 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
1777 zilog
->zl_logbias
= logbias
;
1781 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
1786 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_PUSHPAGE
);
1788 zilog
->zl_header
= zh_phys
;
1790 zilog
->zl_spa
= dmu_objset_spa(os
);
1791 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
1792 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
1793 zilog
->zl_logbias
= dmu_objset_logbias(os
);
1794 zilog
->zl_sync
= dmu_objset_syncprop(os
);
1795 zilog
->zl_next_batch
= 1;
1797 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1799 for (i
= 0; i
< TXG_SIZE
; i
++) {
1800 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
1801 MUTEX_DEFAULT
, NULL
);
1804 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
1805 offsetof(lwb_t
, lwb_node
));
1807 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
1808 offsetof(itx_t
, itx_node
));
1810 mutex_init(&zilog
->zl_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1812 avl_create(&zilog
->zl_vdev_tree
, zil_vdev_compare
,
1813 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
1815 cv_init(&zilog
->zl_cv_writer
, NULL
, CV_DEFAULT
, NULL
);
1816 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
1817 cv_init(&zilog
->zl_cv_batch
[0], NULL
, CV_DEFAULT
, NULL
);
1818 cv_init(&zilog
->zl_cv_batch
[1], NULL
, CV_DEFAULT
, NULL
);
1824 zil_free(zilog_t
*zilog
)
1828 zilog
->zl_stop_sync
= 1;
1830 ASSERT0(zilog
->zl_suspend
);
1831 ASSERT0(zilog
->zl_suspending
);
1833 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1834 list_destroy(&zilog
->zl_lwb_list
);
1836 avl_destroy(&zilog
->zl_vdev_tree
);
1837 mutex_destroy(&zilog
->zl_vdev_lock
);
1839 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
1840 list_destroy(&zilog
->zl_itx_commit_list
);
1842 for (i
= 0; i
< TXG_SIZE
; i
++) {
1844 * It's possible for an itx to be generated that doesn't dirty
1845 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1846 * callback to remove the entry. We remove those here.
1848 * Also free up the ziltest itxs.
1850 if (zilog
->zl_itxg
[i
].itxg_itxs
)
1851 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
1852 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
1855 mutex_destroy(&zilog
->zl_lock
);
1857 cv_destroy(&zilog
->zl_cv_writer
);
1858 cv_destroy(&zilog
->zl_cv_suspend
);
1859 cv_destroy(&zilog
->zl_cv_batch
[0]);
1860 cv_destroy(&zilog
->zl_cv_batch
[1]);
1862 kmem_free(zilog
, sizeof (zilog_t
));
1866 * Open an intent log.
1869 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
1871 zilog_t
*zilog
= dmu_objset_zil(os
);
1873 ASSERT(zilog
->zl_clean_taskq
== NULL
);
1874 ASSERT(zilog
->zl_get_data
== NULL
);
1875 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1877 zilog
->zl_get_data
= get_data
;
1878 zilog
->zl_clean_taskq
= taskq_create("zil_clean", 1, minclsyspri
,
1879 2, 2, TASKQ_PREPOPULATE
);
1885 * Close an intent log.
1888 zil_close(zilog_t
*zilog
)
1893 zil_commit(zilog
, 0); /* commit all itx */
1896 * The lwb_max_txg for the stubby lwb will reflect the last activity
1897 * for the zil. After a txg_wait_synced() on the txg we know all the
1898 * callbacks have occurred that may clean the zil. Only then can we
1899 * destroy the zl_clean_taskq.
1901 mutex_enter(&zilog
->zl_lock
);
1902 lwb
= list_tail(&zilog
->zl_lwb_list
);
1904 txg
= lwb
->lwb_max_txg
;
1905 mutex_exit(&zilog
->zl_lock
);
1907 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1908 ASSERT(!zilog_is_dirty(zilog
));
1910 taskq_destroy(zilog
->zl_clean_taskq
);
1911 zilog
->zl_clean_taskq
= NULL
;
1912 zilog
->zl_get_data
= NULL
;
1915 * We should have only one LWB left on the list; remove it now.
1917 mutex_enter(&zilog
->zl_lock
);
1918 lwb
= list_head(&zilog
->zl_lwb_list
);
1920 ASSERT(lwb
== list_tail(&zilog
->zl_lwb_list
));
1921 ASSERT(lwb
->lwb_zio
== NULL
);
1922 if (lwb
->lwb_fastwrite
)
1923 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1924 list_remove(&zilog
->zl_lwb_list
, lwb
);
1925 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1926 kmem_cache_free(zil_lwb_cache
, lwb
);
1928 mutex_exit(&zilog
->zl_lock
);
1931 static char *suspend_tag
= "zil suspending";
1934 * Suspend an intent log. While in suspended mode, we still honor
1935 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1936 * On old version pools, we suspend the log briefly when taking a
1937 * snapshot so that it will have an empty intent log.
1939 * Long holds are not really intended to be used the way we do here --
1940 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1941 * could fail. Therefore we take pains to only put a long hold if it is
1942 * actually necessary. Fortunately, it will only be necessary if the
1943 * objset is currently mounted (or the ZVOL equivalent). In that case it
1944 * will already have a long hold, so we are not really making things any worse.
1946 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1947 * zvol_state_t), and use their mechanism to prevent their hold from being
1948 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1951 * if cookiep == NULL, this does both the suspend & resume.
1952 * Otherwise, it returns with the dataset "long held", and the cookie
1953 * should be passed into zil_resume().
1956 zil_suspend(const char *osname
, void **cookiep
)
1960 const zil_header_t
*zh
;
1963 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
1966 zilog
= dmu_objset_zil(os
);
1968 mutex_enter(&zilog
->zl_lock
);
1969 zh
= zilog
->zl_header
;
1971 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
1972 mutex_exit(&zilog
->zl_lock
);
1973 dmu_objset_rele(os
, suspend_tag
);
1974 return (SET_ERROR(EBUSY
));
1978 * Don't put a long hold in the cases where we can avoid it. This
1979 * is when there is no cookie so we are doing a suspend & resume
1980 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1981 * for the suspend because it's already suspended, or there's no ZIL.
1983 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
1984 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
1985 mutex_exit(&zilog
->zl_lock
);
1986 dmu_objset_rele(os
, suspend_tag
);
1990 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
1991 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
1993 zilog
->zl_suspend
++;
1995 if (zilog
->zl_suspend
> 1) {
1997 * Someone else is already suspending it.
1998 * Just wait for them to finish.
2001 while (zilog
->zl_suspending
)
2002 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
2003 mutex_exit(&zilog
->zl_lock
);
2005 if (cookiep
== NULL
)
2013 * If there is no pointer to an on-disk block, this ZIL must not
2014 * be active (e.g. filesystem not mounted), so there's nothing
2017 if (BP_IS_HOLE(&zh
->zh_log
)) {
2018 ASSERT(cookiep
!= NULL
); /* fast path already handled */
2021 mutex_exit(&zilog
->zl_lock
);
2025 zilog
->zl_suspending
= B_TRUE
;
2026 mutex_exit(&zilog
->zl_lock
);
2028 zil_commit(zilog
, 0);
2030 zil_destroy(zilog
, B_FALSE
);
2032 mutex_enter(&zilog
->zl_lock
);
2033 zilog
->zl_suspending
= B_FALSE
;
2034 cv_broadcast(&zilog
->zl_cv_suspend
);
2035 mutex_exit(&zilog
->zl_lock
);
2037 if (cookiep
== NULL
)
2045 zil_resume(void *cookie
)
2047 objset_t
*os
= cookie
;
2048 zilog_t
*zilog
= dmu_objset_zil(os
);
2050 mutex_enter(&zilog
->zl_lock
);
2051 ASSERT(zilog
->zl_suspend
!= 0);
2052 zilog
->zl_suspend
--;
2053 mutex_exit(&zilog
->zl_lock
);
2054 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
2055 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
2058 typedef struct zil_replay_arg
{
2059 zil_replay_func_t
*zr_replay
;
2061 boolean_t zr_byteswap
;
2066 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
2068 char name
[MAXNAMELEN
];
2070 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
2072 dmu_objset_name(zilog
->zl_os
, name
);
2074 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
2075 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
2076 (u_longlong_t
)lr
->lrc_seq
,
2077 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
2078 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
2084 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
2086 zil_replay_arg_t
*zr
= zra
;
2087 const zil_header_t
*zh
= zilog
->zl_header
;
2088 uint64_t reclen
= lr
->lrc_reclen
;
2089 uint64_t txtype
= lr
->lrc_txtype
;
2092 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
2094 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
2097 if (lr
->lrc_txg
< claim_txg
) /* already committed */
2100 /* Strip case-insensitive bit, still present in log record */
2103 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
2104 return (zil_replay_error(zilog
, lr
, EINVAL
));
2107 * If this record type can be logged out of order, the object
2108 * (lr_foid) may no longer exist. That's legitimate, not an error.
2110 if (TX_OOO(txtype
)) {
2111 error
= dmu_object_info(zilog
->zl_os
,
2112 ((lr_ooo_t
*)lr
)->lr_foid
, NULL
);
2113 if (error
== ENOENT
|| error
== EEXIST
)
2118 * Make a copy of the data so we can revise and extend it.
2120 bcopy(lr
, zr
->zr_lr
, reclen
);
2123 * If this is a TX_WRITE with a blkptr, suck in the data.
2125 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
2126 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
2127 zr
->zr_lr
+ reclen
);
2129 return (zil_replay_error(zilog
, lr
, error
));
2133 * The log block containing this lr may have been byteswapped
2134 * so that we can easily examine common fields like lrc_txtype.
2135 * However, the log is a mix of different record types, and only the
2136 * replay vectors know how to byteswap their records. Therefore, if
2137 * the lr was byteswapped, undo it before invoking the replay vector.
2139 if (zr
->zr_byteswap
)
2140 byteswap_uint64_array(zr
->zr_lr
, reclen
);
2143 * We must now do two things atomically: replay this log record,
2144 * and update the log header sequence number to reflect the fact that
2145 * we did so. At the end of each replay function the sequence number
2146 * is updated if we are in replay mode.
2148 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
2151 * The DMU's dnode layer doesn't see removes until the txg
2152 * commits, so a subsequent claim can spuriously fail with
2153 * EEXIST. So if we receive any error we try syncing out
2154 * any removes then retry the transaction. Note that we
2155 * specify B_FALSE for byteswap now, so we don't do it twice.
2157 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
2158 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
2160 return (zil_replay_error(zilog
, lr
, error
));
2167 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
2169 zilog
->zl_replay_blks
++;
2175 * If this dataset has a non-empty intent log, replay it and destroy it.
2178 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t replay_func
[TX_MAX_TYPE
])
2180 zilog_t
*zilog
= dmu_objset_zil(os
);
2181 const zil_header_t
*zh
= zilog
->zl_header
;
2182 zil_replay_arg_t zr
;
2184 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
2185 zil_destroy(zilog
, B_TRUE
);
2189 zr
.zr_replay
= replay_func
;
2191 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
2192 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_PUSHPAGE
);
2195 * Wait for in-progress removes to sync before starting replay.
2197 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2199 zilog
->zl_replay
= B_TRUE
;
2200 zilog
->zl_replay_time
= ddi_get_lbolt();
2201 ASSERT(zilog
->zl_replay_blks
== 0);
2202 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
2204 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
2206 zil_destroy(zilog
, B_FALSE
);
2207 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
2208 zilog
->zl_replay
= B_FALSE
;
2212 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
2214 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
2217 if (zilog
->zl_replay
) {
2218 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
2219 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
2220 zilog
->zl_replaying_seq
;
2229 zil_vdev_offline(const char *osname
, void *arg
)
2233 error
= zil_suspend(osname
, NULL
);
2235 return (SET_ERROR(EEXIST
));
2239 #if defined(_KERNEL) && defined(HAVE_SPL)
2240 module_param(zil_replay_disable
, int, 0644);
2241 MODULE_PARM_DESC(zil_replay_disable
, "Disable intent logging replay");
2243 module_param(zfs_nocacheflush
, int, 0644);
2244 MODULE_PARM_DESC(zfs_nocacheflush
, "Disable cache flushes");
2246 module_param(zil_slog_limit
, ulong
, 0644);
2247 MODULE_PARM_DESC(zil_slog_limit
, "Max commit bytes to separate log device");