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) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
27 /* Portions Copyright 2010 Robert Milkowski */
29 #include <sys/zfs_context.h>
35 #include <sys/resource.h>
37 #include <sys/zil_impl.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/vdev_impl.h>
40 #include <sys/dmu_tx.h>
41 #include <sys/dsl_pool.h>
42 #include <sys/metaslab.h>
43 #include <sys/trace_zil.h>
47 * The zfs intent log (ZIL) saves transaction records of system calls
48 * that change the file system in memory with enough information
49 * to be able to replay them. These are stored in memory until
50 * either the DMU transaction group (txg) commits them to the stable pool
51 * and they can be discarded, or they are flushed to the stable log
52 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
53 * requirement. In the event of a panic or power fail then those log
54 * records (transactions) are replayed.
56 * There is one ZIL per file system. Its on-disk (pool) format consists
63 * A log record holds a system call transaction. Log blocks can
64 * hold many log records and the blocks are chained together.
65 * Each ZIL block contains a block pointer (blkptr_t) to the next
66 * ZIL block in the chain. The ZIL header points to the first
67 * block in the chain. Note there is not a fixed place in the pool
68 * to hold blocks. They are dynamically allocated and freed as
69 * needed from the blocks available. Figure X shows the ZIL structure:
73 * See zil.h for more information about these fields.
75 zil_stats_t zil_stats
= {
76 { "zil_commit_count", KSTAT_DATA_UINT64
},
77 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
78 { "zil_itx_count", KSTAT_DATA_UINT64
},
79 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
80 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
81 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
82 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
83 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
84 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
85 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
86 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
87 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
88 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
91 static kstat_t
*zil_ksp
;
94 * Disable intent logging replay. This global ZIL switch affects all pools.
96 int zil_replay_disable
= 0;
99 * Tunable parameter for debugging or performance analysis. Setting
100 * zfs_nocacheflush will cause corruption on power loss if a volatile
101 * out-of-order write cache is enabled.
103 int zfs_nocacheflush
= 0;
106 * Limit SLOG write size per commit executed with synchronous priority.
107 * Any writes above that will be executed with lower (asynchronous) priority
108 * to limit potential SLOG device abuse by single active ZIL writer.
110 unsigned long zil_slog_bulk
= 768 * 1024;
112 static kmem_cache_t
*zil_lwb_cache
;
114 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
116 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
117 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
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 int cmp
= AVL_CMP(DVA_GET_VDEV(dva1
), DVA_GET_VDEV(dva2
));
129 return (AVL_CMP(DVA_GET_OFFSET(dva1
), DVA_GET_OFFSET(dva2
)));
133 zil_bp_tree_init(zilog_t
*zilog
)
135 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
136 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
140 zil_bp_tree_fini(zilog_t
*zilog
)
142 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
146 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
147 kmem_free(zn
, sizeof (zil_bp_node_t
));
153 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
155 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
160 if (BP_IS_EMBEDDED(bp
))
163 dva
= BP_IDENTITY(bp
);
165 if (avl_find(t
, dva
, &where
) != NULL
)
166 return (SET_ERROR(EEXIST
));
168 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
170 avl_insert(t
, zn
, where
);
175 static zil_header_t
*
176 zil_header_in_syncing_context(zilog_t
*zilog
)
178 return ((zil_header_t
*)zilog
->zl_header
);
182 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
184 zio_cksum_t
*zc
= &bp
->blk_cksum
;
186 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
187 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
188 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
189 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
193 * Read a log block and make sure it's valid.
196 zil_read_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, blkptr_t
*nbp
, void *dst
,
199 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
200 arc_flags_t aflags
= ARC_FLAG_WAIT
;
201 arc_buf_t
*abuf
= NULL
;
205 if (zilog
->zl_header
->zh_claim_txg
== 0)
206 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
208 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
209 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
211 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
212 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
214 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
215 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
218 zio_cksum_t cksum
= bp
->blk_cksum
;
221 * Validate the checksummed log block.
223 * Sequence numbers should be... sequential. The checksum
224 * verifier for the next block should be bp's checksum plus 1.
226 * Also check the log chain linkage and size used.
228 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
230 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
231 zil_chain_t
*zilc
= abuf
->b_data
;
232 char *lr
= (char *)(zilc
+ 1);
233 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
235 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
236 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
237 error
= SET_ERROR(ECKSUM
);
239 ASSERT3U(len
, <=, SPA_OLD_MAXBLOCKSIZE
);
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 ASSERT3U(zilc
->zc_nused
, <=,
255 SPA_OLD_MAXBLOCKSIZE
);
256 bcopy(lr
, dst
, zilc
->zc_nused
);
257 *end
= (char *)dst
+ zilc
->zc_nused
;
258 *nbp
= zilc
->zc_next_blk
;
262 arc_buf_destroy(abuf
, &abuf
);
269 * Read a TX_WRITE log data block.
272 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
274 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
275 const blkptr_t
*bp
= &lr
->lr_blkptr
;
276 arc_flags_t aflags
= ARC_FLAG_WAIT
;
277 arc_buf_t
*abuf
= NULL
;
281 if (BP_IS_HOLE(bp
)) {
283 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
287 if (zilog
->zl_header
->zh_claim_txg
== 0)
288 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
290 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
291 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
293 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
294 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
298 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
299 arc_buf_destroy(abuf
, &abuf
);
306 * Parse the intent log, and call parse_func for each valid record within.
309 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
310 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
)
312 const zil_header_t
*zh
= zilog
->zl_header
;
313 boolean_t claimed
= !!zh
->zh_claim_txg
;
314 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
315 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
316 uint64_t max_blk_seq
= 0;
317 uint64_t max_lr_seq
= 0;
318 uint64_t blk_count
= 0;
319 uint64_t lr_count
= 0;
320 blkptr_t blk
, next_blk
;
324 bzero(&next_blk
, sizeof (blkptr_t
));
327 * Old logs didn't record the maximum zh_claim_lr_seq.
329 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
330 claim_lr_seq
= UINT64_MAX
;
333 * Starting at the block pointed to by zh_log we read the log chain.
334 * For each block in the chain we strongly check that block to
335 * ensure its validity. We stop when an invalid block is found.
336 * For each block pointer in the chain we call parse_blk_func().
337 * For each record in each valid block we call parse_lr_func().
338 * If the log has been claimed, stop if we encounter a sequence
339 * number greater than the highest claimed sequence number.
341 lrbuf
= zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE
);
342 zil_bp_tree_init(zilog
);
344 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
345 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
349 if (blk_seq
> claim_blk_seq
)
351 if ((error
= parse_blk_func(zilog
, &blk
, arg
, txg
)) != 0)
353 ASSERT3U(max_blk_seq
, <, blk_seq
);
354 max_blk_seq
= blk_seq
;
357 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
360 error
= zil_read_log_block(zilog
, &blk
, &next_blk
, lrbuf
, &end
);
364 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
365 lr_t
*lr
= (lr_t
*)lrp
;
366 reclen
= lr
->lrc_reclen
;
367 ASSERT3U(reclen
, >=, sizeof (lr_t
));
368 if (lr
->lrc_seq
> claim_lr_seq
)
370 if ((error
= parse_lr_func(zilog
, lr
, arg
, txg
)) != 0)
372 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
373 max_lr_seq
= lr
->lrc_seq
;
378 zilog
->zl_parse_error
= error
;
379 zilog
->zl_parse_blk_seq
= max_blk_seq
;
380 zilog
->zl_parse_lr_seq
= max_lr_seq
;
381 zilog
->zl_parse_blk_count
= blk_count
;
382 zilog
->zl_parse_lr_count
= lr_count
;
384 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
385 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
));
387 zil_bp_tree_fini(zilog
);
388 zio_buf_free(lrbuf
, SPA_OLD_MAXBLOCKSIZE
);
394 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
397 * Claim log block if not already committed and not already claimed.
398 * If tx == NULL, just verify that the block is claimable.
400 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
401 zil_bp_tree_add(zilog
, bp
) != 0)
404 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
405 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
406 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
410 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
412 lr_write_t
*lr
= (lr_write_t
*)lrc
;
415 if (lrc
->lrc_txtype
!= TX_WRITE
)
419 * If the block is not readable, don't claim it. This can happen
420 * in normal operation when a log block is written to disk before
421 * some of the dmu_sync() blocks it points to. In this case, the
422 * transaction cannot have been committed to anyone (we would have
423 * waited for all writes to be stable first), so it is semantically
424 * correct to declare this the end of the log.
426 if (lr
->lr_blkptr
.blk_birth
>= first_txg
&&
427 (error
= zil_read_log_data(zilog
, lr
, NULL
)) != 0)
429 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
434 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
436 zio_free_zil(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
442 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
444 lr_write_t
*lr
= (lr_write_t
*)lrc
;
445 blkptr_t
*bp
= &lr
->lr_blkptr
;
448 * If we previously claimed it, we need to free it.
450 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
451 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
453 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
459 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, boolean_t slog
, uint64_t txg
,
464 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
465 lwb
->lwb_zilog
= zilog
;
467 lwb
->lwb_fastwrite
= fastwrite
;
468 lwb
->lwb_slog
= slog
;
469 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
470 lwb
->lwb_max_txg
= txg
;
473 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
474 lwb
->lwb_nused
= sizeof (zil_chain_t
);
475 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
478 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
481 mutex_enter(&zilog
->zl_lock
);
482 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
483 mutex_exit(&zilog
->zl_lock
);
489 * Called when we create in-memory log transactions so that we know
490 * to cleanup the itxs at the end of spa_sync().
493 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
495 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
496 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
498 if (ds
->ds_is_snapshot
)
499 panic("dirtying snapshot!");
501 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
502 /* up the hold count until we can be written out */
503 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
508 * Determine if the zil is dirty in the specified txg. Callers wanting to
509 * ensure that the dirty state does not change must hold the itxg_lock for
510 * the specified txg. Holding the lock will ensure that the zil cannot be
511 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
515 zilog_is_dirty_in_txg(zilog_t
*zilog
, uint64_t txg
)
517 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
519 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, txg
& TXG_MASK
))
525 * Determine if the zil is dirty. The zil is considered dirty if it has
526 * any pending itx records that have not been cleaned by zil_clean().
529 zilog_is_dirty(zilog_t
*zilog
)
531 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
534 for (t
= 0; t
< TXG_SIZE
; t
++) {
535 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
542 * Create an on-disk intent log.
545 zil_create(zilog_t
*zilog
)
547 const zil_header_t
*zh
= zilog
->zl_header
;
553 boolean_t fastwrite
= FALSE
;
554 boolean_t slog
= FALSE
;
557 * Wait for any previous destroy to complete.
559 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
561 ASSERT(zh
->zh_claim_txg
== 0);
562 ASSERT(zh
->zh_replay_seq
== 0);
567 * Allocate an initial log block if:
568 * - there isn't one already
569 * - the existing block is the wrong endianness
571 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
572 tx
= dmu_tx_create(zilog
->zl_os
);
573 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
574 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
575 txg
= dmu_tx_get_txg(tx
);
577 if (!BP_IS_HOLE(&blk
)) {
578 zio_free_zil(zilog
->zl_spa
, txg
, &blk
);
582 error
= zio_alloc_zil(zilog
->zl_spa
, txg
, &blk
,
583 ZIL_MIN_BLKSZ
, &slog
);
587 zil_init_log_chain(zilog
, &blk
);
591 * Allocate a log write buffer (lwb) for the first log block.
594 lwb
= zil_alloc_lwb(zilog
, &blk
, slog
, txg
, fastwrite
);
597 * If we just allocated the first log block, commit our transaction
598 * and wait for zil_sync() to stuff the block poiner into zh_log.
599 * (zh is part of the MOS, so we cannot modify it in open context.)
603 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
606 ASSERT(bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
612 * In one tx, free all log blocks and clear the log header.
613 * If keep_first is set, then we're replaying a log with no content.
614 * We want to keep the first block, however, so that the first
615 * synchronous transaction doesn't require a txg_wait_synced()
616 * in zil_create(). We don't need to txg_wait_synced() here either
617 * when keep_first is set, because both zil_create() and zil_destroy()
618 * will wait for any in-progress destroys to complete.
621 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
623 const zil_header_t
*zh
= zilog
->zl_header
;
629 * Wait for any previous destroy to complete.
631 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
633 zilog
->zl_old_header
= *zh
; /* debugging aid */
635 if (BP_IS_HOLE(&zh
->zh_log
))
638 tx
= dmu_tx_create(zilog
->zl_os
);
639 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
640 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
641 txg
= dmu_tx_get_txg(tx
);
643 mutex_enter(&zilog
->zl_lock
);
645 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
646 zilog
->zl_destroy_txg
= txg
;
647 zilog
->zl_keep_first
= keep_first
;
649 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
650 ASSERT(zh
->zh_claim_txg
== 0);
652 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
653 ASSERT(lwb
->lwb_zio
== NULL
);
654 if (lwb
->lwb_fastwrite
)
655 metaslab_fastwrite_unmark(zilog
->zl_spa
,
657 list_remove(&zilog
->zl_lwb_list
, lwb
);
658 if (lwb
->lwb_buf
!= NULL
)
659 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
660 zio_free_zil(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
661 kmem_cache_free(zil_lwb_cache
, lwb
);
663 } else if (!keep_first
) {
664 zil_destroy_sync(zilog
, tx
);
666 mutex_exit(&zilog
->zl_lock
);
672 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
674 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
675 (void) zil_parse(zilog
, zil_free_log_block
,
676 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
);
680 zil_claim(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *txarg
)
682 dmu_tx_t
*tx
= txarg
;
683 uint64_t first_txg
= dmu_tx_get_txg(tx
);
689 error
= dmu_objset_own_obj(dp
, ds
->ds_object
,
690 DMU_OST_ANY
, B_FALSE
, FTAG
, &os
);
693 * EBUSY indicates that the objset is inconsistent, in which
694 * case it can not have a ZIL.
696 if (error
!= EBUSY
) {
697 cmn_err(CE_WARN
, "can't open objset for %llu, error %u",
698 (unsigned long long)ds
->ds_object
, error
);
704 zilog
= dmu_objset_zil(os
);
705 zh
= zil_header_in_syncing_context(zilog
);
707 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
) {
708 if (!BP_IS_HOLE(&zh
->zh_log
))
709 zio_free_zil(zilog
->zl_spa
, first_txg
, &zh
->zh_log
);
710 BP_ZERO(&zh
->zh_log
);
711 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
712 dmu_objset_disown(os
, FTAG
);
717 * Claim all log blocks if we haven't already done so, and remember
718 * the highest claimed sequence number. This ensures that if we can
719 * read only part of the log now (e.g. due to a missing device),
720 * but we can read the entire log later, we will not try to replay
721 * or destroy beyond the last block we successfully claimed.
723 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
724 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
725 (void) zil_parse(zilog
, zil_claim_log_block
,
726 zil_claim_log_record
, tx
, first_txg
);
727 zh
->zh_claim_txg
= first_txg
;
728 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
729 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
730 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
731 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
732 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
733 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
736 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
737 dmu_objset_disown(os
, FTAG
);
742 * Check the log by walking the log chain.
743 * Checksum errors are ok as they indicate the end of the chain.
744 * Any other error (no device or read failure) returns an error.
748 zil_check_log_chain(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *tx
)
757 error
= dmu_objset_from_ds(ds
, &os
);
759 cmn_err(CE_WARN
, "can't open objset %llu, error %d",
760 (unsigned long long)ds
->ds_object
, error
);
764 zilog
= dmu_objset_zil(os
);
765 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
768 * Check the first block and determine if it's on a log device
769 * which may have been removed or faulted prior to loading this
770 * pool. If so, there's no point in checking the rest of the log
771 * as its content should have already been synced to the pool.
773 if (!BP_IS_HOLE(bp
)) {
775 boolean_t valid
= B_TRUE
;
777 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
778 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
779 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
780 valid
= vdev_log_state_valid(vd
);
781 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
788 * Because tx == NULL, zil_claim_log_block() will not actually claim
789 * any blocks, but just determine whether it is possible to do so.
790 * In addition to checking the log chain, zil_claim_log_block()
791 * will invoke zio_claim() with a done func of spa_claim_notify(),
792 * which will update spa_max_claim_txg. See spa_load() for details.
794 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
795 zilog
->zl_header
->zh_claim_txg
? -1ULL : spa_first_txg(os
->os_spa
));
797 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
801 zil_vdev_compare(const void *x1
, const void *x2
)
803 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
804 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
806 return (AVL_CMP(v1
, v2
));
810 zil_add_block(zilog_t
*zilog
, const blkptr_t
*bp
)
812 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
814 zil_vdev_node_t
*zv
, zvsearch
;
815 int ndvas
= BP_GET_NDVAS(bp
);
818 if (zfs_nocacheflush
)
821 ASSERT(zilog
->zl_writer
);
824 * Even though we're zl_writer, we still need a lock because the
825 * zl_get_data() callbacks may have dmu_sync() done callbacks
826 * that will run concurrently.
828 mutex_enter(&zilog
->zl_vdev_lock
);
829 for (i
= 0; i
< ndvas
; i
++) {
830 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
831 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
832 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
833 zv
->zv_vdev
= zvsearch
.zv_vdev
;
834 avl_insert(t
, zv
, where
);
837 mutex_exit(&zilog
->zl_vdev_lock
);
841 zil_flush_vdevs(zilog_t
*zilog
)
843 spa_t
*spa
= zilog
->zl_spa
;
844 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
849 ASSERT(zilog
->zl_writer
);
852 * We don't need zl_vdev_lock here because we're the zl_writer,
853 * and all zl_get_data() callbacks are done.
855 if (avl_numnodes(t
) == 0)
858 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
860 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
862 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
863 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
866 kmem_free(zv
, sizeof (*zv
));
870 * Wait for all the flushes to complete. Not all devices actually
871 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
873 (void) zio_wait(zio
);
875 spa_config_exit(spa
, SCL_STATE
, FTAG
);
879 * Function called when a log block write completes
882 zil_lwb_write_done(zio_t
*zio
)
884 lwb_t
*lwb
= zio
->io_private
;
885 zilog_t
*zilog
= lwb
->lwb_zilog
;
886 dmu_tx_t
*tx
= lwb
->lwb_tx
;
888 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
889 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
890 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
891 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
892 ASSERT(!BP_IS_GANG(zio
->io_bp
));
893 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
894 ASSERT(BP_GET_FILL(zio
->io_bp
) == 0);
897 * Ensure the lwb buffer pointer is cleared before releasing
898 * the txg. If we have had an allocation failure and
899 * the txg is waiting to sync then we want want zil_sync()
900 * to remove the lwb so that it's not picked up as the next new
901 * one in zil_commit_writer(). zil_sync() will only remove
902 * the lwb if lwb_buf is null.
904 abd_put(zio
->io_abd
);
905 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
906 mutex_enter(&zilog
->zl_lock
);
908 lwb
->lwb_fastwrite
= FALSE
;
911 mutex_exit(&zilog
->zl_lock
);
914 * Now that we've written this log block, we have a stable pointer
915 * to the next block in the chain, so it's OK to let the txg in
916 * which we allocated the next block sync.
922 * Initialize the io for a log block.
925 zil_lwb_write_init(zilog_t
*zilog
, lwb_t
*lwb
)
930 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
931 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
932 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
934 if (zilog
->zl_root_zio
== NULL
) {
935 zilog
->zl_root_zio
= zio_root(zilog
->zl_spa
, NULL
, NULL
,
939 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
940 mutex_enter(&zilog
->zl_lock
);
941 if (lwb
->lwb_zio
== NULL
) {
942 abd_t
*lwb_abd
= abd_get_from_buf(lwb
->lwb_buf
,
943 BP_GET_LSIZE(&lwb
->lwb_blk
));
944 if (!lwb
->lwb_fastwrite
) {
945 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
946 lwb
->lwb_fastwrite
= 1;
948 if (!lwb
->lwb_slog
|| zilog
->zl_cur_used
<= zil_slog_bulk
)
949 prio
= ZIO_PRIORITY_SYNC_WRITE
;
951 prio
= ZIO_PRIORITY_ASYNC_WRITE
;
952 lwb
->lwb_zio
= zio_rewrite(zilog
->zl_root_zio
, zilog
->zl_spa
,
953 0, &lwb
->lwb_blk
, lwb_abd
, BP_GET_LSIZE(&lwb
->lwb_blk
),
954 zil_lwb_write_done
, lwb
, prio
,
955 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
|
956 ZIO_FLAG_FASTWRITE
, &zb
);
958 mutex_exit(&zilog
->zl_lock
);
962 * Define a limited set of intent log block sizes.
964 * These must be a multiple of 4KB. Note only the amount used (again
965 * aligned to 4KB) actually gets written. However, we can't always just
966 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
968 uint64_t zil_block_buckets
[] = {
969 4096, /* non TX_WRITE */
970 8192+4096, /* data base */
971 32*1024 + 4096, /* NFS writes */
976 * Start a log block write and advance to the next log block.
977 * Calls are serialized.
980 zil_lwb_write_start(zilog_t
*zilog
, lwb_t
*lwb
)
984 spa_t
*spa
= zilog
->zl_spa
;
988 uint64_t zil_blksz
, wsz
;
992 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
993 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
994 bp
= &zilc
->zc_next_blk
;
996 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
997 bp
= &zilc
->zc_next_blk
;
1000 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
1003 * Allocate the next block and save its address in this block
1004 * before writing it in order to establish the log chain.
1005 * Note that if the allocation of nlwb synced before we wrote
1006 * the block that points at it (lwb), we'd leak it if we crashed.
1007 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1008 * We dirty the dataset to ensure that zil_sync() will be called
1009 * to clean up in the event of allocation failure or I/O failure.
1011 tx
= dmu_tx_create(zilog
->zl_os
);
1012 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
1013 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1014 txg
= dmu_tx_get_txg(tx
);
1019 * Log blocks are pre-allocated. Here we select the size of the next
1020 * block, based on size used in the last block.
1021 * - first find the smallest bucket that will fit the block from a
1022 * limited set of block sizes. This is because it's faster to write
1023 * blocks allocated from the same metaslab as they are adjacent or
1025 * - next find the maximum from the new suggested size and an array of
1026 * previous sizes. This lessens a picket fence effect of wrongly
1027 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1030 * Note we only write what is used, but we can't just allocate
1031 * the maximum block size because we can exhaust the available
1034 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1035 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
1037 zil_blksz
= zil_block_buckets
[i
];
1038 if (zil_blksz
== UINT64_MAX
)
1039 zil_blksz
= SPA_OLD_MAXBLOCKSIZE
;
1040 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1041 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1042 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1043 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1046 error
= zio_alloc_zil(spa
, txg
, bp
, zil_blksz
, &slog
);
1048 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count
);
1049 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes
, lwb
->lwb_nused
);
1051 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count
);
1052 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes
, lwb
->lwb_nused
);
1055 ASSERT3U(bp
->blk_birth
, ==, txg
);
1056 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1057 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1060 * Allocate a new log write buffer (lwb).
1062 nlwb
= zil_alloc_lwb(zilog
, bp
, slog
, txg
, TRUE
);
1064 /* Record the block for later vdev flushing */
1065 zil_add_block(zilog
, &lwb
->lwb_blk
);
1068 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1069 /* For Slim ZIL only write what is used. */
1070 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1071 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1072 zio_shrink(lwb
->lwb_zio
, wsz
);
1079 zilc
->zc_nused
= lwb
->lwb_nused
;
1080 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1083 * clear unused data for security
1085 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
1087 zio_nowait(lwb
->lwb_zio
); /* Kick off the write for the old log block */
1090 * If there was an allocation failure then nlwb will be null which
1091 * forces a txg_wait_synced().
1097 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1100 lr_write_t
*lrwb
, *lrw
;
1102 uint64_t dlen
, dnow
, lwb_sp
, reclen
, txg
;
1107 ASSERT(lwb
->lwb_buf
!= NULL
);
1109 lrc
= &itx
->itx_lr
; /* Common log record inside itx. */
1110 lrw
= (lr_write_t
*)lrc
; /* Write log record inside itx. */
1111 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
) {
1112 dlen
= P2ROUNDUP_TYPED(
1113 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1117 reclen
= lrc
->lrc_reclen
;
1118 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1121 zil_lwb_write_init(zilog
, lwb
);
1125 * If this record won't fit in the current log block, start a new one.
1126 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1128 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1129 if (reclen
> lwb_sp
|| (reclen
+ dlen
> lwb_sp
&&
1130 lwb_sp
< ZIL_MAX_WASTE_SPACE
&& (dlen
% ZIL_MAX_LOG_DATA
== 0 ||
1131 lwb_sp
< reclen
+ dlen
% ZIL_MAX_LOG_DATA
))) {
1132 lwb
= zil_lwb_write_start(zilog
, lwb
);
1135 zil_lwb_write_init(zilog
, lwb
);
1136 ASSERT(LWB_EMPTY(lwb
));
1137 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1138 ASSERT3U(reclen
+ MIN(dlen
, sizeof (uint64_t)), <=, lwb_sp
);
1141 dnow
= MIN(dlen
, lwb_sp
- reclen
);
1142 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1143 bcopy(lrc
, lr_buf
, reclen
);
1144 lrcb
= (lr_t
*)lr_buf
; /* Like lrc, but inside lwb. */
1145 lrwb
= (lr_write_t
*)lrcb
; /* Like lrw, but inside lwb. */
1147 ZIL_STAT_BUMP(zil_itx_count
);
1150 * If it's a write, fetch the data or get its blkptr as appropriate.
1152 if (lrc
->lrc_txtype
== TX_WRITE
) {
1153 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1154 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1155 if (itx
->itx_wr_state
== WR_COPIED
) {
1156 ZIL_STAT_BUMP(zil_itx_copied_count
);
1157 ZIL_STAT_INCR(zil_itx_copied_bytes
, lrw
->lr_length
);
1162 if (itx
->itx_wr_state
== WR_NEED_COPY
) {
1163 dbuf
= lr_buf
+ reclen
;
1164 lrcb
->lrc_reclen
+= dnow
;
1165 if (lrwb
->lr_length
> dnow
)
1166 lrwb
->lr_length
= dnow
;
1167 lrw
->lr_offset
+= dnow
;
1168 lrw
->lr_length
-= dnow
;
1169 ZIL_STAT_BUMP(zil_itx_needcopy_count
);
1170 ZIL_STAT_INCR(zil_itx_needcopy_bytes
,
1173 ASSERT(itx
->itx_wr_state
== WR_INDIRECT
);
1175 ZIL_STAT_BUMP(zil_itx_indirect_count
);
1176 ZIL_STAT_INCR(zil_itx_indirect_bytes
,
1179 error
= zilog
->zl_get_data(
1180 itx
->itx_private
, lrwb
, dbuf
, lwb
->lwb_zio
);
1182 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1186 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1194 * We're actually making an entry, so update lrc_seq to be the
1195 * log record sequence number. Note that this is generally not
1196 * equal to the itx sequence number because not all transactions
1197 * are synchronous, and sometimes spa_sync() gets there first.
1199 lrcb
->lrc_seq
= ++zilog
->zl_lr_seq
; /* we are single threaded */
1200 lwb
->lwb_nused
+= reclen
+ dnow
;
1201 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1202 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1203 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1207 zilog
->zl_cur_used
+= reclen
;
1215 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1220 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1221 itxsize
= offsetof(itx_t
, itx_lr
) + lrsize
;
1223 itx
= zio_data_buf_alloc(itxsize
);
1224 itx
->itx_lr
.lrc_txtype
= txtype
;
1225 itx
->itx_lr
.lrc_reclen
= lrsize
;
1226 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1227 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1228 itx
->itx_callback
= NULL
;
1229 itx
->itx_callback_data
= NULL
;
1230 itx
->itx_size
= itxsize
;
1236 zil_itx_destroy(itx_t
*itx
)
1238 zio_data_buf_free(itx
, itx
->itx_size
);
1242 * Free up the sync and async itxs. The itxs_t has already been detached
1243 * so no locks are needed.
1246 zil_itxg_clean(itxs_t
*itxs
)
1252 itx_async_node_t
*ian
;
1254 list
= &itxs
->i_sync_list
;
1255 while ((itx
= list_head(list
)) != NULL
) {
1256 if (itx
->itx_callback
!= NULL
)
1257 itx
->itx_callback(itx
->itx_callback_data
);
1258 list_remove(list
, itx
);
1259 zil_itx_destroy(itx
);
1263 t
= &itxs
->i_async_tree
;
1264 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1265 list
= &ian
->ia_list
;
1266 while ((itx
= list_head(list
)) != NULL
) {
1267 if (itx
->itx_callback
!= NULL
)
1268 itx
->itx_callback(itx
->itx_callback_data
);
1269 list_remove(list
, itx
);
1270 zil_itx_destroy(itx
);
1273 kmem_free(ian
, sizeof (itx_async_node_t
));
1277 kmem_free(itxs
, sizeof (itxs_t
));
1281 zil_aitx_compare(const void *x1
, const void *x2
)
1283 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1284 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1286 return (AVL_CMP(o1
, o2
));
1290 * Remove all async itx with the given oid.
1293 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1296 itx_async_node_t
*ian
;
1303 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1305 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1308 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1310 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1311 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1313 mutex_enter(&itxg
->itxg_lock
);
1314 if (itxg
->itxg_txg
!= txg
) {
1315 mutex_exit(&itxg
->itxg_lock
);
1320 * Locate the object node and append its list.
1322 t
= &itxg
->itxg_itxs
->i_async_tree
;
1323 ian
= avl_find(t
, &oid
, &where
);
1325 list_move_tail(&clean_list
, &ian
->ia_list
);
1326 mutex_exit(&itxg
->itxg_lock
);
1328 while ((itx
= list_head(&clean_list
)) != NULL
) {
1329 if (itx
->itx_callback
!= NULL
)
1330 itx
->itx_callback(itx
->itx_callback_data
);
1331 list_remove(&clean_list
, itx
);
1332 zil_itx_destroy(itx
);
1334 list_destroy(&clean_list
);
1338 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1342 itxs_t
*itxs
, *clean
= NULL
;
1345 * Object ids can be re-instantiated in the next txg so
1346 * remove any async transactions to avoid future leaks.
1347 * This can happen if a fsync occurs on the re-instantiated
1348 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1349 * the new file data and flushes a write record for the old object.
1351 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_REMOVE
)
1352 zil_remove_async(zilog
, itx
->itx_oid
);
1355 * Ensure the data of a renamed file is committed before the rename.
1357 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1358 zil_async_to_sync(zilog
, itx
->itx_oid
);
1360 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1363 txg
= dmu_tx_get_txg(tx
);
1365 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1366 mutex_enter(&itxg
->itxg_lock
);
1367 itxs
= itxg
->itxg_itxs
;
1368 if (itxg
->itxg_txg
!= txg
) {
1371 * The zil_clean callback hasn't got around to cleaning
1372 * this itxg. Save the itxs for release below.
1373 * This should be rare.
1375 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1376 "txg %llu", itxg
->itxg_txg
);
1377 clean
= itxg
->itxg_itxs
;
1379 itxg
->itxg_txg
= txg
;
1380 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
1383 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1384 offsetof(itx_t
, itx_node
));
1385 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1386 sizeof (itx_async_node_t
),
1387 offsetof(itx_async_node_t
, ia_node
));
1389 if (itx
->itx_sync
) {
1390 list_insert_tail(&itxs
->i_sync_list
, itx
);
1392 avl_tree_t
*t
= &itxs
->i_async_tree
;
1394 LR_FOID_GET_OBJ(((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
);
1395 itx_async_node_t
*ian
;
1398 ian
= avl_find(t
, &foid
, &where
);
1400 ian
= kmem_alloc(sizeof (itx_async_node_t
),
1402 list_create(&ian
->ia_list
, sizeof (itx_t
),
1403 offsetof(itx_t
, itx_node
));
1404 ian
->ia_foid
= foid
;
1405 avl_insert(t
, ian
, where
);
1407 list_insert_tail(&ian
->ia_list
, itx
);
1410 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1411 zilog_dirty(zilog
, txg
);
1412 mutex_exit(&itxg
->itxg_lock
);
1414 /* Release the old itxs now we've dropped the lock */
1416 zil_itxg_clean(clean
);
1420 * If there are any in-memory intent log transactions which have now been
1421 * synced then start up a taskq to free them. We should only do this after we
1422 * have written out the uberblocks (i.e. txg has been comitted) so that
1423 * don't inadvertently clean out in-memory log records that would be required
1427 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
1429 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
1432 mutex_enter(&itxg
->itxg_lock
);
1433 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
1434 mutex_exit(&itxg
->itxg_lock
);
1437 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
1438 ASSERT(itxg
->itxg_txg
!= 0);
1439 ASSERT(zilog
->zl_clean_taskq
!= NULL
);
1440 clean_me
= itxg
->itxg_itxs
;
1441 itxg
->itxg_itxs
= NULL
;
1443 mutex_exit(&itxg
->itxg_lock
);
1445 * Preferably start a task queue to free up the old itxs but
1446 * if taskq_dispatch can't allocate resources to do that then
1447 * free it in-line. This should be rare. Note, using TQ_SLEEP
1448 * created a bad performance problem.
1450 if (taskq_dispatch(zilog
->zl_clean_taskq
,
1451 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
) == 0)
1452 zil_itxg_clean(clean_me
);
1456 * Get the list of itxs to commit into zl_itx_commit_list.
1459 zil_get_commit_list(zilog_t
*zilog
)
1462 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
1464 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1467 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1470 * This is inherently racy, since there is nothing to prevent
1471 * the last synced txg from changing. That's okay since we'll
1472 * only commit things in the future.
1474 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1475 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1477 mutex_enter(&itxg
->itxg_lock
);
1478 if (itxg
->itxg_txg
!= txg
) {
1479 mutex_exit(&itxg
->itxg_lock
);
1484 * If we're adding itx records to the zl_itx_commit_list,
1485 * then the zil better be dirty in this "txg". We can assert
1486 * that here since we're holding the itxg_lock which will
1487 * prevent spa_sync from cleaning it. Once we add the itxs
1488 * to the zl_itx_commit_list we must commit it to disk even
1489 * if it's unnecessary (i.e. the txg was synced).
1491 ASSERT(zilog_is_dirty_in_txg(zilog
, txg
) ||
1492 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
1493 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
1495 mutex_exit(&itxg
->itxg_lock
);
1500 * Move the async itxs for a specified object to commit into sync lists.
1503 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
1506 itx_async_node_t
*ian
;
1510 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1513 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1516 * This is inherently racy, since there is nothing to prevent
1517 * the last synced txg from changing.
1519 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1520 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1522 mutex_enter(&itxg
->itxg_lock
);
1523 if (itxg
->itxg_txg
!= txg
) {
1524 mutex_exit(&itxg
->itxg_lock
);
1529 * If a foid is specified then find that node and append its
1530 * list. Otherwise walk the tree appending all the lists
1531 * to the sync list. We add to the end rather than the
1532 * beginning to ensure the create has happened.
1534 t
= &itxg
->itxg_itxs
->i_async_tree
;
1536 ian
= avl_find(t
, &foid
, &where
);
1538 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1542 void *cookie
= NULL
;
1544 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1545 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1547 list_destroy(&ian
->ia_list
);
1548 kmem_free(ian
, sizeof (itx_async_node_t
));
1551 mutex_exit(&itxg
->itxg_lock
);
1556 zil_commit_writer(zilog_t
*zilog
)
1561 spa_t
*spa
= zilog
->zl_spa
;
1564 ASSERT(zilog
->zl_root_zio
== NULL
);
1566 mutex_exit(&zilog
->zl_lock
);
1568 zil_get_commit_list(zilog
);
1571 * Return if there's nothing to commit before we dirty the fs by
1572 * calling zil_create().
1574 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
) {
1575 mutex_enter(&zilog
->zl_lock
);
1579 if (zilog
->zl_suspend
) {
1582 lwb
= list_tail(&zilog
->zl_lwb_list
);
1584 lwb
= zil_create(zilog
);
1587 DTRACE_PROBE1(zil__cw1
, zilog_t
*, zilog
);
1588 for (itx
= list_head(&zilog
->zl_itx_commit_list
); itx
!= NULL
;
1589 itx
= list_next(&zilog
->zl_itx_commit_list
, itx
)) {
1590 txg
= itx
->itx_lr
.lrc_txg
;
1591 ASSERT3U(txg
, !=, 0);
1594 * This is inherently racy and may result in us writing
1595 * out a log block for a txg that was just synced. This is
1596 * ok since we'll end cleaning up that log block the next
1597 * time we call zil_sync().
1599 if (txg
> spa_last_synced_txg(spa
) || txg
> spa_freeze_txg(spa
))
1600 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
1602 DTRACE_PROBE1(zil__cw2
, zilog_t
*, zilog
);
1604 /* write the last block out */
1605 if (lwb
!= NULL
&& lwb
->lwb_zio
!= NULL
)
1606 lwb
= zil_lwb_write_start(zilog
, lwb
);
1608 zilog
->zl_cur_used
= 0;
1611 * Wait if necessary for the log blocks to be on stable storage.
1613 if (zilog
->zl_root_zio
) {
1614 error
= zio_wait(zilog
->zl_root_zio
);
1615 zilog
->zl_root_zio
= NULL
;
1616 zil_flush_vdevs(zilog
);
1619 if (error
|| lwb
== NULL
)
1620 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
1622 while ((itx
= list_head(&zilog
->zl_itx_commit_list
))) {
1623 txg
= itx
->itx_lr
.lrc_txg
;
1626 if (itx
->itx_callback
!= NULL
)
1627 itx
->itx_callback(itx
->itx_callback_data
);
1628 list_remove(&zilog
->zl_itx_commit_list
, itx
);
1629 zil_itx_destroy(itx
);
1632 mutex_enter(&zilog
->zl_lock
);
1635 * Remember the highest committed log sequence number for ztest.
1636 * We only update this value when all the log writes succeeded,
1637 * because ztest wants to ASSERT that it got the whole log chain.
1639 if (error
== 0 && lwb
!= NULL
)
1640 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1644 * Commit zfs transactions to stable storage.
1645 * If foid is 0 push out all transactions, otherwise push only those
1646 * for that object or might reference that object.
1648 * itxs are committed in batches. In a heavily stressed zil there will be
1649 * a commit writer thread who is writing out a bunch of itxs to the log
1650 * for a set of committing threads (cthreads) in the same batch as the writer.
1651 * Those cthreads are all waiting on the same cv for that batch.
1653 * There will also be a different and growing batch of threads that are
1654 * waiting to commit (qthreads). When the committing batch completes
1655 * a transition occurs such that the cthreads exit and the qthreads become
1656 * cthreads. One of the new cthreads becomes the writer thread for the
1657 * batch. Any new threads arriving become new qthreads.
1659 * Only 2 condition variables are needed and there's no transition
1660 * between the two cvs needed. They just flip-flop between qthreads
1663 * Using this scheme we can efficiently wakeup up only those threads
1664 * that have been committed.
1667 zil_commit(zilog_t
*zilog
, uint64_t foid
)
1671 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
1674 ZIL_STAT_BUMP(zil_commit_count
);
1676 /* move the async itxs for the foid to the sync queues */
1677 zil_async_to_sync(zilog
, foid
);
1679 mutex_enter(&zilog
->zl_lock
);
1680 mybatch
= zilog
->zl_next_batch
;
1681 while (zilog
->zl_writer
) {
1682 cv_wait(&zilog
->zl_cv_batch
[mybatch
& 1], &zilog
->zl_lock
);
1683 if (mybatch
<= zilog
->zl_com_batch
) {
1684 mutex_exit(&zilog
->zl_lock
);
1689 zilog
->zl_next_batch
++;
1690 zilog
->zl_writer
= B_TRUE
;
1691 ZIL_STAT_BUMP(zil_commit_writer_count
);
1692 zil_commit_writer(zilog
);
1693 zilog
->zl_com_batch
= mybatch
;
1694 zilog
->zl_writer
= B_FALSE
;
1696 /* wake up one thread to become the next writer */
1697 cv_signal(&zilog
->zl_cv_batch
[(mybatch
+1) & 1]);
1699 /* wake up all threads waiting for this batch to be committed */
1700 cv_broadcast(&zilog
->zl_cv_batch
[mybatch
& 1]);
1702 mutex_exit(&zilog
->zl_lock
);
1706 * Called in syncing context to free committed log blocks and update log header.
1709 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1711 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1712 uint64_t txg
= dmu_tx_get_txg(tx
);
1713 spa_t
*spa
= zilog
->zl_spa
;
1714 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
1718 * We don't zero out zl_destroy_txg, so make sure we don't try
1719 * to destroy it twice.
1721 if (spa_sync_pass(spa
) != 1)
1724 mutex_enter(&zilog
->zl_lock
);
1726 ASSERT(zilog
->zl_stop_sync
== 0);
1728 if (*replayed_seq
!= 0) {
1729 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
1730 zh
->zh_replay_seq
= *replayed_seq
;
1734 if (zilog
->zl_destroy_txg
== txg
) {
1735 blkptr_t blk
= zh
->zh_log
;
1737 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
1739 bzero(zh
, sizeof (zil_header_t
));
1740 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
1742 if (zilog
->zl_keep_first
) {
1744 * If this block was part of log chain that couldn't
1745 * be claimed because a device was missing during
1746 * zil_claim(), but that device later returns,
1747 * then this block could erroneously appear valid.
1748 * To guard against this, assign a new GUID to the new
1749 * log chain so it doesn't matter what blk points to.
1751 zil_init_log_chain(zilog
, &blk
);
1756 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
1757 zh
->zh_log
= lwb
->lwb_blk
;
1758 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
1761 ASSERT(lwb
->lwb_zio
== NULL
);
1763 list_remove(&zilog
->zl_lwb_list
, lwb
);
1764 zio_free_zil(spa
, txg
, &lwb
->lwb_blk
);
1765 kmem_cache_free(zil_lwb_cache
, lwb
);
1768 * If we don't have anything left in the lwb list then
1769 * we've had an allocation failure and we need to zero
1770 * out the zil_header blkptr so that we don't end
1771 * up freeing the same block twice.
1773 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
1774 BP_ZERO(&zh
->zh_log
);
1778 * Remove fastwrite on any blocks that have been pre-allocated for
1779 * the next commit. This prevents fastwrite counter pollution by
1780 * unused, long-lived LWBs.
1782 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
1783 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_zio
) {
1784 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1785 lwb
->lwb_fastwrite
= 0;
1789 mutex_exit(&zilog
->zl_lock
);
1795 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
1796 sizeof (struct lwb
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
1798 zil_ksp
= kstat_create("zfs", 0, "zil", "misc",
1799 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
1800 KSTAT_FLAG_VIRTUAL
);
1802 if (zil_ksp
!= NULL
) {
1803 zil_ksp
->ks_data
= &zil_stats
;
1804 kstat_install(zil_ksp
);
1811 kmem_cache_destroy(zil_lwb_cache
);
1813 if (zil_ksp
!= NULL
) {
1814 kstat_delete(zil_ksp
);
1820 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
1822 zilog
->zl_sync
= sync
;
1826 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
1828 zilog
->zl_logbias
= logbias
;
1832 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
1837 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
1839 zilog
->zl_header
= zh_phys
;
1841 zilog
->zl_spa
= dmu_objset_spa(os
);
1842 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
1843 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
1844 zilog
->zl_logbias
= dmu_objset_logbias(os
);
1845 zilog
->zl_sync
= dmu_objset_syncprop(os
);
1846 zilog
->zl_next_batch
= 1;
1848 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1850 for (i
= 0; i
< TXG_SIZE
; i
++) {
1851 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
1852 MUTEX_DEFAULT
, NULL
);
1855 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
1856 offsetof(lwb_t
, lwb_node
));
1858 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
1859 offsetof(itx_t
, itx_node
));
1861 mutex_init(&zilog
->zl_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1863 avl_create(&zilog
->zl_vdev_tree
, zil_vdev_compare
,
1864 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
1866 cv_init(&zilog
->zl_cv_writer
, NULL
, CV_DEFAULT
, NULL
);
1867 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
1868 cv_init(&zilog
->zl_cv_batch
[0], NULL
, CV_DEFAULT
, NULL
);
1869 cv_init(&zilog
->zl_cv_batch
[1], NULL
, CV_DEFAULT
, NULL
);
1875 zil_free(zilog_t
*zilog
)
1879 zilog
->zl_stop_sync
= 1;
1881 ASSERT0(zilog
->zl_suspend
);
1882 ASSERT0(zilog
->zl_suspending
);
1884 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1885 list_destroy(&zilog
->zl_lwb_list
);
1887 avl_destroy(&zilog
->zl_vdev_tree
);
1888 mutex_destroy(&zilog
->zl_vdev_lock
);
1890 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
1891 list_destroy(&zilog
->zl_itx_commit_list
);
1893 for (i
= 0; i
< TXG_SIZE
; i
++) {
1895 * It's possible for an itx to be generated that doesn't dirty
1896 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1897 * callback to remove the entry. We remove those here.
1899 * Also free up the ziltest itxs.
1901 if (zilog
->zl_itxg
[i
].itxg_itxs
)
1902 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
1903 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
1906 mutex_destroy(&zilog
->zl_lock
);
1908 cv_destroy(&zilog
->zl_cv_writer
);
1909 cv_destroy(&zilog
->zl_cv_suspend
);
1910 cv_destroy(&zilog
->zl_cv_batch
[0]);
1911 cv_destroy(&zilog
->zl_cv_batch
[1]);
1913 kmem_free(zilog
, sizeof (zilog_t
));
1917 * Open an intent log.
1920 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
1922 zilog_t
*zilog
= dmu_objset_zil(os
);
1924 ASSERT(zilog
->zl_clean_taskq
== NULL
);
1925 ASSERT(zilog
->zl_get_data
== NULL
);
1926 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1928 zilog
->zl_get_data
= get_data
;
1929 zilog
->zl_clean_taskq
= taskq_create("zil_clean", 1, defclsyspri
,
1930 2, 2, TASKQ_PREPOPULATE
);
1936 * Close an intent log.
1939 zil_close(zilog_t
*zilog
)
1944 zil_commit(zilog
, 0); /* commit all itx */
1947 * The lwb_max_txg for the stubby lwb will reflect the last activity
1948 * for the zil. After a txg_wait_synced() on the txg we know all the
1949 * callbacks have occurred that may clean the zil. Only then can we
1950 * destroy the zl_clean_taskq.
1952 mutex_enter(&zilog
->zl_lock
);
1953 lwb
= list_tail(&zilog
->zl_lwb_list
);
1955 txg
= lwb
->lwb_max_txg
;
1956 mutex_exit(&zilog
->zl_lock
);
1958 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1960 if (zilog_is_dirty(zilog
))
1961 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog
, txg
);
1962 if (txg
< spa_freeze_txg(zilog
->zl_spa
))
1963 VERIFY(!zilog_is_dirty(zilog
));
1965 taskq_destroy(zilog
->zl_clean_taskq
);
1966 zilog
->zl_clean_taskq
= NULL
;
1967 zilog
->zl_get_data
= NULL
;
1970 * We should have only one LWB left on the list; remove it now.
1972 mutex_enter(&zilog
->zl_lock
);
1973 lwb
= list_head(&zilog
->zl_lwb_list
);
1975 ASSERT(lwb
== list_tail(&zilog
->zl_lwb_list
));
1976 ASSERT(lwb
->lwb_zio
== NULL
);
1977 if (lwb
->lwb_fastwrite
)
1978 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1979 list_remove(&zilog
->zl_lwb_list
, lwb
);
1980 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1981 kmem_cache_free(zil_lwb_cache
, lwb
);
1983 mutex_exit(&zilog
->zl_lock
);
1986 static char *suspend_tag
= "zil suspending";
1989 * Suspend an intent log. While in suspended mode, we still honor
1990 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1991 * On old version pools, we suspend the log briefly when taking a
1992 * snapshot so that it will have an empty intent log.
1994 * Long holds are not really intended to be used the way we do here --
1995 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1996 * could fail. Therefore we take pains to only put a long hold if it is
1997 * actually necessary. Fortunately, it will only be necessary if the
1998 * objset is currently mounted (or the ZVOL equivalent). In that case it
1999 * will already have a long hold, so we are not really making things any worse.
2001 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
2002 * zvol_state_t), and use their mechanism to prevent their hold from being
2003 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
2006 * if cookiep == NULL, this does both the suspend & resume.
2007 * Otherwise, it returns with the dataset "long held", and the cookie
2008 * should be passed into zil_resume().
2011 zil_suspend(const char *osname
, void **cookiep
)
2015 const zil_header_t
*zh
;
2018 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
2021 zilog
= dmu_objset_zil(os
);
2023 mutex_enter(&zilog
->zl_lock
);
2024 zh
= zilog
->zl_header
;
2026 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
2027 mutex_exit(&zilog
->zl_lock
);
2028 dmu_objset_rele(os
, suspend_tag
);
2029 return (SET_ERROR(EBUSY
));
2033 * Don't put a long hold in the cases where we can avoid it. This
2034 * is when there is no cookie so we are doing a suspend & resume
2035 * (i.e. called from zil_vdev_offline()), and there's nothing to do
2036 * for the suspend because it's already suspended, or there's no ZIL.
2038 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
2039 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
2040 mutex_exit(&zilog
->zl_lock
);
2041 dmu_objset_rele(os
, suspend_tag
);
2045 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
2046 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
2048 zilog
->zl_suspend
++;
2050 if (zilog
->zl_suspend
> 1) {
2052 * Someone else is already suspending it.
2053 * Just wait for them to finish.
2056 while (zilog
->zl_suspending
)
2057 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
2058 mutex_exit(&zilog
->zl_lock
);
2060 if (cookiep
== NULL
)
2068 * If there is no pointer to an on-disk block, this ZIL must not
2069 * be active (e.g. filesystem not mounted), so there's nothing
2072 if (BP_IS_HOLE(&zh
->zh_log
)) {
2073 ASSERT(cookiep
!= NULL
); /* fast path already handled */
2076 mutex_exit(&zilog
->zl_lock
);
2080 zilog
->zl_suspending
= B_TRUE
;
2081 mutex_exit(&zilog
->zl_lock
);
2083 zil_commit(zilog
, 0);
2085 zil_destroy(zilog
, B_FALSE
);
2087 mutex_enter(&zilog
->zl_lock
);
2088 zilog
->zl_suspending
= B_FALSE
;
2089 cv_broadcast(&zilog
->zl_cv_suspend
);
2090 mutex_exit(&zilog
->zl_lock
);
2092 if (cookiep
== NULL
)
2100 zil_resume(void *cookie
)
2102 objset_t
*os
= cookie
;
2103 zilog_t
*zilog
= dmu_objset_zil(os
);
2105 mutex_enter(&zilog
->zl_lock
);
2106 ASSERT(zilog
->zl_suspend
!= 0);
2107 zilog
->zl_suspend
--;
2108 mutex_exit(&zilog
->zl_lock
);
2109 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
2110 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
2113 typedef struct zil_replay_arg
{
2114 zil_replay_func_t
*zr_replay
;
2116 boolean_t zr_byteswap
;
2121 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
2123 char name
[ZFS_MAX_DATASET_NAME_LEN
];
2125 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
2127 dmu_objset_name(zilog
->zl_os
, name
);
2129 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
2130 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
2131 (u_longlong_t
)lr
->lrc_seq
,
2132 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
2133 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
2139 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
2141 zil_replay_arg_t
*zr
= zra
;
2142 const zil_header_t
*zh
= zilog
->zl_header
;
2143 uint64_t reclen
= lr
->lrc_reclen
;
2144 uint64_t txtype
= lr
->lrc_txtype
;
2147 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
2149 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
2152 if (lr
->lrc_txg
< claim_txg
) /* already committed */
2155 /* Strip case-insensitive bit, still present in log record */
2158 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
2159 return (zil_replay_error(zilog
, lr
, EINVAL
));
2162 * If this record type can be logged out of order, the object
2163 * (lr_foid) may no longer exist. That's legitimate, not an error.
2165 if (TX_OOO(txtype
)) {
2166 error
= dmu_object_info(zilog
->zl_os
,
2167 LR_FOID_GET_OBJ(((lr_ooo_t
*)lr
)->lr_foid
), NULL
);
2168 if (error
== ENOENT
|| error
== EEXIST
)
2173 * Make a copy of the data so we can revise and extend it.
2175 bcopy(lr
, zr
->zr_lr
, reclen
);
2178 * If this is a TX_WRITE with a blkptr, suck in the data.
2180 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
2181 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
2182 zr
->zr_lr
+ reclen
);
2184 return (zil_replay_error(zilog
, lr
, error
));
2188 * The log block containing this lr may have been byteswapped
2189 * so that we can easily examine common fields like lrc_txtype.
2190 * However, the log is a mix of different record types, and only the
2191 * replay vectors know how to byteswap their records. Therefore, if
2192 * the lr was byteswapped, undo it before invoking the replay vector.
2194 if (zr
->zr_byteswap
)
2195 byteswap_uint64_array(zr
->zr_lr
, reclen
);
2198 * We must now do two things atomically: replay this log record,
2199 * and update the log header sequence number to reflect the fact that
2200 * we did so. At the end of each replay function the sequence number
2201 * is updated if we are in replay mode.
2203 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
2206 * The DMU's dnode layer doesn't see removes until the txg
2207 * commits, so a subsequent claim can spuriously fail with
2208 * EEXIST. So if we receive any error we try syncing out
2209 * any removes then retry the transaction. Note that we
2210 * specify B_FALSE for byteswap now, so we don't do it twice.
2212 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
2213 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
2215 return (zil_replay_error(zilog
, lr
, error
));
2222 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
2224 zilog
->zl_replay_blks
++;
2230 * If this dataset has a non-empty intent log, replay it and destroy it.
2233 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t replay_func
[TX_MAX_TYPE
])
2235 zilog_t
*zilog
= dmu_objset_zil(os
);
2236 const zil_header_t
*zh
= zilog
->zl_header
;
2237 zil_replay_arg_t zr
;
2239 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
2240 zil_destroy(zilog
, B_TRUE
);
2244 zr
.zr_replay
= replay_func
;
2246 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
2247 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
2250 * Wait for in-progress removes to sync before starting replay.
2252 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2254 zilog
->zl_replay
= B_TRUE
;
2255 zilog
->zl_replay_time
= ddi_get_lbolt();
2256 ASSERT(zilog
->zl_replay_blks
== 0);
2257 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
2259 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
2261 zil_destroy(zilog
, B_FALSE
);
2262 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
2263 zilog
->zl_replay
= B_FALSE
;
2267 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
2269 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
2272 if (zilog
->zl_replay
) {
2273 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
2274 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
2275 zilog
->zl_replaying_seq
;
2284 zil_vdev_offline(const char *osname
, void *arg
)
2288 error
= zil_suspend(osname
, NULL
);
2290 return (SET_ERROR(EEXIST
));
2294 #if defined(_KERNEL) && defined(HAVE_SPL)
2295 EXPORT_SYMBOL(zil_alloc
);
2296 EXPORT_SYMBOL(zil_free
);
2297 EXPORT_SYMBOL(zil_open
);
2298 EXPORT_SYMBOL(zil_close
);
2299 EXPORT_SYMBOL(zil_replay
);
2300 EXPORT_SYMBOL(zil_replaying
);
2301 EXPORT_SYMBOL(zil_destroy
);
2302 EXPORT_SYMBOL(zil_destroy_sync
);
2303 EXPORT_SYMBOL(zil_itx_create
);
2304 EXPORT_SYMBOL(zil_itx_destroy
);
2305 EXPORT_SYMBOL(zil_itx_assign
);
2306 EXPORT_SYMBOL(zil_commit
);
2307 EXPORT_SYMBOL(zil_vdev_offline
);
2308 EXPORT_SYMBOL(zil_claim
);
2309 EXPORT_SYMBOL(zil_check_log_chain
);
2310 EXPORT_SYMBOL(zil_sync
);
2311 EXPORT_SYMBOL(zil_clean
);
2312 EXPORT_SYMBOL(zil_suspend
);
2313 EXPORT_SYMBOL(zil_resume
);
2314 EXPORT_SYMBOL(zil_add_block
);
2315 EXPORT_SYMBOL(zil_bp_tree_add
);
2316 EXPORT_SYMBOL(zil_set_sync
);
2317 EXPORT_SYMBOL(zil_set_logbias
);
2320 module_param(zil_replay_disable
, int, 0644);
2321 MODULE_PARM_DESC(zil_replay_disable
, "Disable intent logging replay");
2323 module_param(zfs_nocacheflush
, int, 0644);
2324 MODULE_PARM_DESC(zfs_nocacheflush
, "Disable cache flushes");
2326 module_param(zil_slog_bulk
, ulong
, 0644);
2327 MODULE_PARM_DESC(zil_slog_bulk
, "Limit in bytes slog sync writes per commit");