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, 2014 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>
42 #include <sys/trace_zil.h>
45 * The zfs intent log (ZIL) saves transaction records of system calls
46 * that change the file system in memory with enough information
47 * to be able to replay them. These are stored in memory until
48 * either the DMU transaction group (txg) commits them to the stable pool
49 * and they can be discarded, or they are flushed to the stable log
50 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
51 * requirement. In the event of a panic or power fail then those log
52 * records (transactions) are replayed.
54 * There is one ZIL per file system. Its on-disk (pool) format consists
61 * A log record holds a system call transaction. Log blocks can
62 * hold many log records and the blocks are chained together.
63 * Each ZIL block contains a block pointer (blkptr_t) to the next
64 * ZIL block in the chain. The ZIL header points to the first
65 * block in the chain. Note there is not a fixed place in the pool
66 * to hold blocks. They are dynamically allocated and freed as
67 * needed from the blocks available. Figure X shows the ZIL structure:
71 * See zil.h for more information about these fields.
73 zil_stats_t zil_stats
= {
74 { "zil_commit_count", KSTAT_DATA_UINT64
},
75 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
76 { "zil_itx_count", KSTAT_DATA_UINT64
},
77 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
78 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
79 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
80 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
81 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
82 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
83 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
84 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
85 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
86 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
89 static kstat_t
*zil_ksp
;
92 * Disable intent logging replay. This global ZIL switch affects all pools.
94 int zil_replay_disable
= 0;
97 * Tunable parameter for debugging or performance analysis. Setting
98 * zfs_nocacheflush will cause corruption on power loss if a volatile
99 * out-of-order write cache is enabled.
101 int zfs_nocacheflush
= 0;
103 static kmem_cache_t
*zil_lwb_cache
;
105 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
107 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
108 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
112 * ziltest is by and large an ugly hack, but very useful in
113 * checking replay without tedious work.
114 * When running ziltest we want to keep all itx's and so maintain
115 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
116 * We subtract TXG_CONCURRENT_STATES to allow for common code.
118 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
121 zil_bp_compare(const void *x1
, const void *x2
)
123 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
124 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
126 if (DVA_GET_VDEV(dva1
) < DVA_GET_VDEV(dva2
))
128 if (DVA_GET_VDEV(dva1
) > DVA_GET_VDEV(dva2
))
131 if (DVA_GET_OFFSET(dva1
) < DVA_GET_OFFSET(dva2
))
133 if (DVA_GET_OFFSET(dva1
) > DVA_GET_OFFSET(dva2
))
140 zil_bp_tree_init(zilog_t
*zilog
)
142 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
143 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
147 zil_bp_tree_fini(zilog_t
*zilog
)
149 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
153 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
154 kmem_free(zn
, sizeof (zil_bp_node_t
));
160 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
162 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
167 if (BP_IS_EMBEDDED(bp
))
170 dva
= BP_IDENTITY(bp
);
172 if (avl_find(t
, dva
, &where
) != NULL
)
173 return (SET_ERROR(EEXIST
));
175 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
177 avl_insert(t
, zn
, where
);
182 static zil_header_t
*
183 zil_header_in_syncing_context(zilog_t
*zilog
)
185 return ((zil_header_t
*)zilog
->zl_header
);
189 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
191 zio_cksum_t
*zc
= &bp
->blk_cksum
;
193 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
194 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
195 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
196 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
200 * Read a log block and make sure it's valid.
203 zil_read_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, blkptr_t
*nbp
, void *dst
,
206 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
207 uint32_t aflags
= ARC_WAIT
;
208 arc_buf_t
*abuf
= NULL
;
212 if (zilog
->zl_header
->zh_claim_txg
== 0)
213 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
215 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
216 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
218 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
219 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
221 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
222 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
225 zio_cksum_t cksum
= bp
->blk_cksum
;
228 * Validate the checksummed log block.
230 * Sequence numbers should be... sequential. The checksum
231 * verifier for the next block should be bp's checksum plus 1.
233 * Also check the log chain linkage and size used.
235 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
237 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
238 zil_chain_t
*zilc
= abuf
->b_data
;
239 char *lr
= (char *)(zilc
+ 1);
240 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
242 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
243 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
244 error
= SET_ERROR(ECKSUM
);
247 *end
= (char *)dst
+ len
;
248 *nbp
= zilc
->zc_next_blk
;
251 char *lr
= abuf
->b_data
;
252 uint64_t size
= BP_GET_LSIZE(bp
);
253 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
255 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
256 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
257 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
258 error
= SET_ERROR(ECKSUM
);
260 bcopy(lr
, dst
, zilc
->zc_nused
);
261 *end
= (char *)dst
+ zilc
->zc_nused
;
262 *nbp
= zilc
->zc_next_blk
;
266 VERIFY(arc_buf_remove_ref(abuf
, &abuf
));
273 * Read a TX_WRITE log data block.
276 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
278 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
279 const blkptr_t
*bp
= &lr
->lr_blkptr
;
280 uint32_t aflags
= ARC_WAIT
;
281 arc_buf_t
*abuf
= NULL
;
285 if (BP_IS_HOLE(bp
)) {
287 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
291 if (zilog
->zl_header
->zh_claim_txg
== 0)
292 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
294 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
295 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
297 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
298 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
302 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
303 (void) arc_buf_remove_ref(abuf
, &abuf
);
310 * Parse the intent log, and call parse_func for each valid record within.
313 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
314 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
)
316 const zil_header_t
*zh
= zilog
->zl_header
;
317 boolean_t claimed
= !!zh
->zh_claim_txg
;
318 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
319 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
320 uint64_t max_blk_seq
= 0;
321 uint64_t max_lr_seq
= 0;
322 uint64_t blk_count
= 0;
323 uint64_t lr_count
= 0;
324 blkptr_t blk
, next_blk
;
328 bzero(&next_blk
, sizeof (blkptr_t
));
331 * Old logs didn't record the maximum zh_claim_lr_seq.
333 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
334 claim_lr_seq
= UINT64_MAX
;
337 * Starting at the block pointed to by zh_log we read the log chain.
338 * For each block in the chain we strongly check that block to
339 * ensure its validity. We stop when an invalid block is found.
340 * For each block pointer in the chain we call parse_blk_func().
341 * For each record in each valid block we call parse_lr_func().
342 * If the log has been claimed, stop if we encounter a sequence
343 * number greater than the highest claimed sequence number.
345 lrbuf
= zio_buf_alloc(SPA_MAXBLOCKSIZE
);
346 zil_bp_tree_init(zilog
);
348 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
349 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
353 if (blk_seq
> claim_blk_seq
)
355 if ((error
= parse_blk_func(zilog
, &blk
, arg
, txg
)) != 0)
357 ASSERT3U(max_blk_seq
, <, blk_seq
);
358 max_blk_seq
= blk_seq
;
361 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
364 error
= zil_read_log_block(zilog
, &blk
, &next_blk
, lrbuf
, &end
);
368 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
369 lr_t
*lr
= (lr_t
*)lrp
;
370 reclen
= lr
->lrc_reclen
;
371 ASSERT3U(reclen
, >=, sizeof (lr_t
));
372 if (lr
->lrc_seq
> claim_lr_seq
)
374 if ((error
= parse_lr_func(zilog
, lr
, arg
, txg
)) != 0)
376 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
377 max_lr_seq
= lr
->lrc_seq
;
382 zilog
->zl_parse_error
= error
;
383 zilog
->zl_parse_blk_seq
= max_blk_seq
;
384 zilog
->zl_parse_lr_seq
= max_lr_seq
;
385 zilog
->zl_parse_blk_count
= blk_count
;
386 zilog
->zl_parse_lr_count
= lr_count
;
388 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
389 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
));
391 zil_bp_tree_fini(zilog
);
392 zio_buf_free(lrbuf
, SPA_MAXBLOCKSIZE
);
398 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
401 * Claim log block if not already committed and not already claimed.
402 * If tx == NULL, just verify that the block is claimable.
404 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
405 zil_bp_tree_add(zilog
, bp
) != 0)
408 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
409 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
410 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
414 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
416 lr_write_t
*lr
= (lr_write_t
*)lrc
;
419 if (lrc
->lrc_txtype
!= TX_WRITE
)
423 * If the block is not readable, don't claim it. This can happen
424 * in normal operation when a log block is written to disk before
425 * some of the dmu_sync() blocks it points to. In this case, the
426 * transaction cannot have been committed to anyone (we would have
427 * waited for all writes to be stable first), so it is semantically
428 * correct to declare this the end of the log.
430 if (lr
->lr_blkptr
.blk_birth
>= first_txg
&&
431 (error
= zil_read_log_data(zilog
, lr
, NULL
)) != 0)
433 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
438 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
440 zio_free_zil(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
446 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
448 lr_write_t
*lr
= (lr_write_t
*)lrc
;
449 blkptr_t
*bp
= &lr
->lr_blkptr
;
452 * If we previously claimed it, we need to free it.
454 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
455 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
457 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
463 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, uint64_t txg
, boolean_t fastwrite
)
467 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
468 lwb
->lwb_zilog
= zilog
;
470 lwb
->lwb_fastwrite
= fastwrite
;
471 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
472 lwb
->lwb_max_txg
= txg
;
475 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
476 lwb
->lwb_nused
= sizeof (zil_chain_t
);
477 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
480 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
483 mutex_enter(&zilog
->zl_lock
);
484 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
485 mutex_exit(&zilog
->zl_lock
);
491 * Called when we create in-memory log transactions so that we know
492 * to cleanup the itxs at the end of spa_sync().
495 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
497 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
498 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
500 if (ds
->ds_is_snapshot
)
501 panic("dirtying snapshot!");
503 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
504 /* up the hold count until we can be written out */
505 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
510 zilog_is_dirty(zilog_t
*zilog
)
512 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
515 for (t
= 0; t
< TXG_SIZE
; t
++) {
516 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
523 * Create an on-disk intent log.
526 zil_create(zilog_t
*zilog
)
528 const zil_header_t
*zh
= zilog
->zl_header
;
534 boolean_t fastwrite
= FALSE
;
537 * Wait for any previous destroy to complete.
539 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
541 ASSERT(zh
->zh_claim_txg
== 0);
542 ASSERT(zh
->zh_replay_seq
== 0);
547 * Allocate an initial log block if:
548 * - there isn't one already
549 * - the existing block is the wrong endianess
551 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
552 tx
= dmu_tx_create(zilog
->zl_os
);
553 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
554 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
555 txg
= dmu_tx_get_txg(tx
);
557 if (!BP_IS_HOLE(&blk
)) {
558 zio_free_zil(zilog
->zl_spa
, txg
, &blk
);
562 error
= zio_alloc_zil(zilog
->zl_spa
, txg
, &blk
,
563 ZIL_MIN_BLKSZ
, B_TRUE
);
567 zil_init_log_chain(zilog
, &blk
);
571 * Allocate a log write buffer (lwb) for the first log block.
574 lwb
= zil_alloc_lwb(zilog
, &blk
, txg
, fastwrite
);
577 * If we just allocated the first log block, commit our transaction
578 * and wait for zil_sync() to stuff the block poiner into zh_log.
579 * (zh is part of the MOS, so we cannot modify it in open context.)
583 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
586 ASSERT(bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
592 * In one tx, free all log blocks and clear the log header.
593 * If keep_first is set, then we're replaying a log with no content.
594 * We want to keep the first block, however, so that the first
595 * synchronous transaction doesn't require a txg_wait_synced()
596 * in zil_create(). We don't need to txg_wait_synced() here either
597 * when keep_first is set, because both zil_create() and zil_destroy()
598 * will wait for any in-progress destroys to complete.
601 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
603 const zil_header_t
*zh
= zilog
->zl_header
;
609 * Wait for any previous destroy to complete.
611 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
613 zilog
->zl_old_header
= *zh
; /* debugging aid */
615 if (BP_IS_HOLE(&zh
->zh_log
))
618 tx
= dmu_tx_create(zilog
->zl_os
);
619 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
620 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
621 txg
= dmu_tx_get_txg(tx
);
623 mutex_enter(&zilog
->zl_lock
);
625 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
626 zilog
->zl_destroy_txg
= txg
;
627 zilog
->zl_keep_first
= keep_first
;
629 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
630 ASSERT(zh
->zh_claim_txg
== 0);
632 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
633 ASSERT(lwb
->lwb_zio
== NULL
);
634 if (lwb
->lwb_fastwrite
)
635 metaslab_fastwrite_unmark(zilog
->zl_spa
,
637 list_remove(&zilog
->zl_lwb_list
, lwb
);
638 if (lwb
->lwb_buf
!= NULL
)
639 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
640 zio_free_zil(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
641 kmem_cache_free(zil_lwb_cache
, lwb
);
643 } else if (!keep_first
) {
644 zil_destroy_sync(zilog
, tx
);
646 mutex_exit(&zilog
->zl_lock
);
652 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
654 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
655 (void) zil_parse(zilog
, zil_free_log_block
,
656 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
);
660 zil_claim(const char *osname
, void *txarg
)
662 dmu_tx_t
*tx
= txarg
;
663 uint64_t first_txg
= dmu_tx_get_txg(tx
);
669 error
= dmu_objset_own(osname
, DMU_OST_ANY
, B_FALSE
, FTAG
, &os
);
672 * EBUSY indicates that the objset is inconsistent, in which
673 * case it can not have a ZIL.
675 if (error
!= EBUSY
) {
676 cmn_err(CE_WARN
, "can't open objset for %s, error %u",
683 zilog
= dmu_objset_zil(os
);
684 zh
= zil_header_in_syncing_context(zilog
);
686 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
) {
687 if (!BP_IS_HOLE(&zh
->zh_log
))
688 zio_free_zil(zilog
->zl_spa
, first_txg
, &zh
->zh_log
);
689 BP_ZERO(&zh
->zh_log
);
690 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
691 dmu_objset_disown(os
, FTAG
);
696 * Claim all log blocks if we haven't already done so, and remember
697 * the highest claimed sequence number. This ensures that if we can
698 * read only part of the log now (e.g. due to a missing device),
699 * but we can read the entire log later, we will not try to replay
700 * or destroy beyond the last block we successfully claimed.
702 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
703 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
704 (void) zil_parse(zilog
, zil_claim_log_block
,
705 zil_claim_log_record
, tx
, first_txg
);
706 zh
->zh_claim_txg
= first_txg
;
707 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
708 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
709 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
710 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
711 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
712 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
715 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
716 dmu_objset_disown(os
, FTAG
);
721 * Check the log by walking the log chain.
722 * Checksum errors are ok as they indicate the end of the chain.
723 * Any other error (no device or read failure) returns an error.
726 zil_check_log_chain(const char *osname
, void *tx
)
735 error
= dmu_objset_hold(osname
, FTAG
, &os
);
737 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
741 zilog
= dmu_objset_zil(os
);
742 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
745 * Check the first block and determine if it's on a log device
746 * which may have been removed or faulted prior to loading this
747 * pool. If so, there's no point in checking the rest of the log
748 * as its content should have already been synced to the pool.
750 if (!BP_IS_HOLE(bp
)) {
752 boolean_t valid
= B_TRUE
;
754 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
755 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
756 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
757 valid
= vdev_log_state_valid(vd
);
758 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
761 dmu_objset_rele(os
, FTAG
);
767 * Because tx == NULL, zil_claim_log_block() will not actually claim
768 * any blocks, but just determine whether it is possible to do so.
769 * In addition to checking the log chain, zil_claim_log_block()
770 * will invoke zio_claim() with a done func of spa_claim_notify(),
771 * which will update spa_max_claim_txg. See spa_load() for details.
773 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
774 zilog
->zl_header
->zh_claim_txg
? -1ULL : spa_first_txg(os
->os_spa
));
776 dmu_objset_rele(os
, FTAG
);
778 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
782 zil_vdev_compare(const void *x1
, const void *x2
)
784 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
785 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
796 zil_add_block(zilog_t
*zilog
, const blkptr_t
*bp
)
798 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
800 zil_vdev_node_t
*zv
, zvsearch
;
801 int ndvas
= BP_GET_NDVAS(bp
);
804 if (zfs_nocacheflush
)
807 ASSERT(zilog
->zl_writer
);
810 * Even though we're zl_writer, we still need a lock because the
811 * zl_get_data() callbacks may have dmu_sync() done callbacks
812 * that will run concurrently.
814 mutex_enter(&zilog
->zl_vdev_lock
);
815 for (i
= 0; i
< ndvas
; i
++) {
816 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
817 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
818 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
819 zv
->zv_vdev
= zvsearch
.zv_vdev
;
820 avl_insert(t
, zv
, where
);
823 mutex_exit(&zilog
->zl_vdev_lock
);
827 zil_flush_vdevs(zilog_t
*zilog
)
829 spa_t
*spa
= zilog
->zl_spa
;
830 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
835 ASSERT(zilog
->zl_writer
);
838 * We don't need zl_vdev_lock here because we're the zl_writer,
839 * and all zl_get_data() callbacks are done.
841 if (avl_numnodes(t
) == 0)
844 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
846 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
848 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
849 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
852 kmem_free(zv
, sizeof (*zv
));
856 * Wait for all the flushes to complete. Not all devices actually
857 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
859 (void) zio_wait(zio
);
861 spa_config_exit(spa
, SCL_STATE
, FTAG
);
865 * Function called when a log block write completes
868 zil_lwb_write_done(zio_t
*zio
)
870 lwb_t
*lwb
= zio
->io_private
;
871 zilog_t
*zilog
= lwb
->lwb_zilog
;
872 dmu_tx_t
*tx
= lwb
->lwb_tx
;
874 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
875 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
876 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
877 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
878 ASSERT(!BP_IS_GANG(zio
->io_bp
));
879 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
880 ASSERT(BP_GET_FILL(zio
->io_bp
) == 0);
883 * Ensure the lwb buffer pointer is cleared before releasing
884 * the txg. If we have had an allocation failure and
885 * the txg is waiting to sync then we want want zil_sync()
886 * to remove the lwb so that it's not picked up as the next new
887 * one in zil_commit_writer(). zil_sync() will only remove
888 * the lwb if lwb_buf is null.
890 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
891 mutex_enter(&zilog
->zl_lock
);
893 lwb
->lwb_fastwrite
= FALSE
;
896 mutex_exit(&zilog
->zl_lock
);
899 * Now that we've written this log block, we have a stable pointer
900 * to the next block in the chain, so it's OK to let the txg in
901 * which we allocated the next block sync.
907 * Initialize the io for a log block.
910 zil_lwb_write_init(zilog_t
*zilog
, lwb_t
*lwb
)
914 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
915 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
916 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
918 if (zilog
->zl_root_zio
== NULL
) {
919 zilog
->zl_root_zio
= zio_root(zilog
->zl_spa
, NULL
, NULL
,
923 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
924 mutex_enter(&zilog
->zl_lock
);
925 if (lwb
->lwb_zio
== NULL
) {
926 if (!lwb
->lwb_fastwrite
) {
927 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
928 lwb
->lwb_fastwrite
= 1;
930 lwb
->lwb_zio
= zio_rewrite(zilog
->zl_root_zio
, zilog
->zl_spa
,
931 0, &lwb
->lwb_blk
, lwb
->lwb_buf
, BP_GET_LSIZE(&lwb
->lwb_blk
),
932 zil_lwb_write_done
, lwb
, ZIO_PRIORITY_SYNC_WRITE
,
933 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
|
934 ZIO_FLAG_FASTWRITE
, &zb
);
936 mutex_exit(&zilog
->zl_lock
);
940 * Define a limited set of intent log block sizes.
942 * These must be a multiple of 4KB. Note only the amount used (again
943 * aligned to 4KB) actually gets written. However, we can't always just
944 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
946 uint64_t zil_block_buckets
[] = {
947 4096, /* non TX_WRITE */
948 8192+4096, /* data base */
949 32*1024 + 4096, /* NFS writes */
954 * Use the slog as long as the current commit size is less than the
955 * limit or the total list size is less than 2X the limit. Limit
956 * checking is disabled by setting zil_slog_limit to UINT64_MAX.
958 unsigned long zil_slog_limit
= 1024 * 1024;
959 #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \
960 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))
963 * Start a log block write and advance to the next log block.
964 * Calls are serialized.
967 zil_lwb_write_start(zilog_t
*zilog
, lwb_t
*lwb
)
971 spa_t
*spa
= zilog
->zl_spa
;
975 uint64_t zil_blksz
, wsz
;
979 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
980 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
981 bp
= &zilc
->zc_next_blk
;
983 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
984 bp
= &zilc
->zc_next_blk
;
987 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
990 * Allocate the next block and save its address in this block
991 * before writing it in order to establish the log chain.
992 * Note that if the allocation of nlwb synced before we wrote
993 * the block that points at it (lwb), we'd leak it if we crashed.
994 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
995 * We dirty the dataset to ensure that zil_sync() will be called
996 * to clean up in the event of allocation failure or I/O failure.
998 tx
= dmu_tx_create(zilog
->zl_os
);
999 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
1000 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1001 txg
= dmu_tx_get_txg(tx
);
1006 * Log blocks are pre-allocated. Here we select the size of the next
1007 * block, based on size used in the last block.
1008 * - first find the smallest bucket that will fit the block from a
1009 * limited set of block sizes. This is because it's faster to write
1010 * blocks allocated from the same metaslab as they are adjacent or
1012 * - next find the maximum from the new suggested size and an array of
1013 * previous sizes. This lessens a picket fence effect of wrongly
1014 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1017 * Note we only write what is used, but we can't just allocate
1018 * the maximum block size because we can exhaust the available
1021 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1022 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
1024 zil_blksz
= zil_block_buckets
[i
];
1025 if (zil_blksz
== UINT64_MAX
)
1026 zil_blksz
= SPA_MAXBLOCKSIZE
;
1027 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1028 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1029 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1030 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1033 use_slog
= USE_SLOG(zilog
);
1034 error
= zio_alloc_zil(spa
, txg
, bp
, zil_blksz
,
1037 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count
);
1038 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes
, lwb
->lwb_nused
);
1040 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count
);
1041 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes
, lwb
->lwb_nused
);
1044 ASSERT3U(bp
->blk_birth
, ==, txg
);
1045 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1046 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1049 * Allocate a new log write buffer (lwb).
1051 nlwb
= zil_alloc_lwb(zilog
, bp
, txg
, TRUE
);
1053 /* Record the block for later vdev flushing */
1054 zil_add_block(zilog
, &lwb
->lwb_blk
);
1057 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1058 /* For Slim ZIL only write what is used. */
1059 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1060 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1061 zio_shrink(lwb
->lwb_zio
, wsz
);
1068 zilc
->zc_nused
= lwb
->lwb_nused
;
1069 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1072 * clear unused data for security
1074 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
1076 zio_nowait(lwb
->lwb_zio
); /* Kick off the write for the old log block */
1079 * If there was an allocation failure then nlwb will be null which
1080 * forces a txg_wait_synced().
1086 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1088 lr_t
*lrc
= &itx
->itx_lr
; /* common log record */
1089 lr_write_t
*lrw
= (lr_write_t
*)lrc
;
1091 uint64_t txg
= lrc
->lrc_txg
;
1092 uint64_t reclen
= lrc
->lrc_reclen
;
1098 ASSERT(lwb
->lwb_buf
!= NULL
);
1099 ASSERT(zilog_is_dirty(zilog
) ||
1100 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
1102 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
)
1103 dlen
= P2ROUNDUP_TYPED(
1104 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1106 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1108 zil_lwb_write_init(zilog
, lwb
);
1111 * If this record won't fit in the current log block, start a new one.
1113 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1114 lwb
= zil_lwb_write_start(zilog
, lwb
);
1117 zil_lwb_write_init(zilog
, lwb
);
1118 ASSERT(LWB_EMPTY(lwb
));
1119 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1120 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1125 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1126 bcopy(lrc
, lr_buf
, reclen
);
1127 lrc
= (lr_t
*)lr_buf
;
1128 lrw
= (lr_write_t
*)lrc
;
1130 ZIL_STAT_BUMP(zil_itx_count
);
1133 * If it's a write, fetch the data or get its blkptr as appropriate.
1135 if (lrc
->lrc_txtype
== TX_WRITE
) {
1136 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1137 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1138 if (itx
->itx_wr_state
== WR_COPIED
) {
1139 ZIL_STAT_BUMP(zil_itx_copied_count
);
1140 ZIL_STAT_INCR(zil_itx_copied_bytes
, lrw
->lr_length
);
1146 ASSERT(itx
->itx_wr_state
== WR_NEED_COPY
);
1147 dbuf
= lr_buf
+ reclen
;
1148 lrw
->lr_common
.lrc_reclen
+= dlen
;
1149 ZIL_STAT_BUMP(zil_itx_needcopy_count
);
1150 ZIL_STAT_INCR(zil_itx_needcopy_bytes
,
1153 ASSERT(itx
->itx_wr_state
== WR_INDIRECT
);
1155 ZIL_STAT_BUMP(zil_itx_indirect_count
);
1156 ZIL_STAT_INCR(zil_itx_indirect_bytes
,
1159 error
= zilog
->zl_get_data(
1160 itx
->itx_private
, lrw
, dbuf
, lwb
->lwb_zio
);
1162 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1166 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1174 * We're actually making an entry, so update lrc_seq to be the
1175 * log record sequence number. Note that this is generally not
1176 * equal to the itx sequence number because not all transactions
1177 * are synchronous, and sometimes spa_sync() gets there first.
1179 lrc
->lrc_seq
= ++zilog
->zl_lr_seq
; /* we are single threaded */
1180 lwb
->lwb_nused
+= reclen
+ dlen
;
1181 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1182 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1183 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1189 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1193 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1195 itx
= zio_data_buf_alloc(offsetof(itx_t
, itx_lr
) + lrsize
);
1196 itx
->itx_lr
.lrc_txtype
= txtype
;
1197 itx
->itx_lr
.lrc_reclen
= lrsize
;
1198 itx
->itx_sod
= lrsize
; /* if write & WR_NEED_COPY will be increased */
1199 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1200 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1201 itx
->itx_callback
= NULL
;
1202 itx
->itx_callback_data
= NULL
;
1208 zil_itx_destroy(itx_t
*itx
)
1210 zio_data_buf_free(itx
, offsetof(itx_t
, itx_lr
)+itx
->itx_lr
.lrc_reclen
);
1214 * Free up the sync and async itxs. The itxs_t has already been detached
1215 * so no locks are needed.
1218 zil_itxg_clean(itxs_t
*itxs
)
1224 itx_async_node_t
*ian
;
1226 list
= &itxs
->i_sync_list
;
1227 while ((itx
= list_head(list
)) != NULL
) {
1228 if (itx
->itx_callback
!= NULL
)
1229 itx
->itx_callback(itx
->itx_callback_data
);
1230 list_remove(list
, itx
);
1231 zil_itx_destroy(itx
);
1235 t
= &itxs
->i_async_tree
;
1236 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1237 list
= &ian
->ia_list
;
1238 while ((itx
= list_head(list
)) != NULL
) {
1239 if (itx
->itx_callback
!= NULL
)
1240 itx
->itx_callback(itx
->itx_callback_data
);
1241 list_remove(list
, itx
);
1242 zil_itx_destroy(itx
);
1245 kmem_free(ian
, sizeof (itx_async_node_t
));
1249 kmem_free(itxs
, sizeof (itxs_t
));
1253 zil_aitx_compare(const void *x1
, const void *x2
)
1255 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1256 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1267 * Remove all async itx with the given oid.
1270 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1273 itx_async_node_t
*ian
;
1280 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1282 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1285 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1287 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1288 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1290 mutex_enter(&itxg
->itxg_lock
);
1291 if (itxg
->itxg_txg
!= txg
) {
1292 mutex_exit(&itxg
->itxg_lock
);
1297 * Locate the object node and append its list.
1299 t
= &itxg
->itxg_itxs
->i_async_tree
;
1300 ian
= avl_find(t
, &oid
, &where
);
1302 list_move_tail(&clean_list
, &ian
->ia_list
);
1303 mutex_exit(&itxg
->itxg_lock
);
1305 while ((itx
= list_head(&clean_list
)) != NULL
) {
1306 if (itx
->itx_callback
!= NULL
)
1307 itx
->itx_callback(itx
->itx_callback_data
);
1308 list_remove(&clean_list
, itx
);
1309 zil_itx_destroy(itx
);
1311 list_destroy(&clean_list
);
1315 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1319 itxs_t
*itxs
, *clean
= NULL
;
1322 * Object ids can be re-instantiated in the next txg so
1323 * remove any async transactions to avoid future leaks.
1324 * This can happen if a fsync occurs on the re-instantiated
1325 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1326 * the new file data and flushes a write record for the old object.
1328 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_REMOVE
)
1329 zil_remove_async(zilog
, itx
->itx_oid
);
1332 * Ensure the data of a renamed file is committed before the rename.
1334 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1335 zil_async_to_sync(zilog
, itx
->itx_oid
);
1337 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1340 txg
= dmu_tx_get_txg(tx
);
1342 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1343 mutex_enter(&itxg
->itxg_lock
);
1344 itxs
= itxg
->itxg_itxs
;
1345 if (itxg
->itxg_txg
!= txg
) {
1348 * The zil_clean callback hasn't got around to cleaning
1349 * this itxg. Save the itxs for release below.
1350 * This should be rare.
1352 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1354 clean
= itxg
->itxg_itxs
;
1356 ASSERT(itxg
->itxg_sod
== 0);
1357 itxg
->itxg_txg
= txg
;
1358 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
1361 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1362 offsetof(itx_t
, itx_node
));
1363 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1364 sizeof (itx_async_node_t
),
1365 offsetof(itx_async_node_t
, ia_node
));
1367 if (itx
->itx_sync
) {
1368 list_insert_tail(&itxs
->i_sync_list
, itx
);
1369 atomic_add_64(&zilog
->zl_itx_list_sz
, itx
->itx_sod
);
1370 itxg
->itxg_sod
+= itx
->itx_sod
;
1372 avl_tree_t
*t
= &itxs
->i_async_tree
;
1373 uint64_t foid
= ((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
;
1374 itx_async_node_t
*ian
;
1377 ian
= avl_find(t
, &foid
, &where
);
1379 ian
= kmem_alloc(sizeof (itx_async_node_t
),
1381 list_create(&ian
->ia_list
, sizeof (itx_t
),
1382 offsetof(itx_t
, itx_node
));
1383 ian
->ia_foid
= foid
;
1384 avl_insert(t
, ian
, where
);
1386 list_insert_tail(&ian
->ia_list
, itx
);
1389 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1390 zilog_dirty(zilog
, txg
);
1391 mutex_exit(&itxg
->itxg_lock
);
1393 /* Release the old itxs now we've dropped the lock */
1395 zil_itxg_clean(clean
);
1399 * If there are any in-memory intent log transactions which have now been
1400 * synced then start up a taskq to free them. We should only do this after we
1401 * have written out the uberblocks (i.e. txg has been comitted) so that
1402 * don't inadvertently clean out in-memory log records that would be required
1406 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
1408 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
1411 mutex_enter(&itxg
->itxg_lock
);
1412 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
1413 mutex_exit(&itxg
->itxg_lock
);
1416 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
1417 ASSERT(itxg
->itxg_txg
!= 0);
1418 ASSERT(zilog
->zl_clean_taskq
!= NULL
);
1419 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1421 clean_me
= itxg
->itxg_itxs
;
1422 itxg
->itxg_itxs
= NULL
;
1424 mutex_exit(&itxg
->itxg_lock
);
1426 * Preferably start a task queue to free up the old itxs but
1427 * if taskq_dispatch can't allocate resources to do that then
1428 * free it in-line. This should be rare. Note, using TQ_SLEEP
1429 * created a bad performance problem.
1431 if (taskq_dispatch(zilog
->zl_clean_taskq
,
1432 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
) == 0)
1433 zil_itxg_clean(clean_me
);
1437 * Get the list of itxs to commit into zl_itx_commit_list.
1440 zil_get_commit_list(zilog_t
*zilog
)
1443 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
1444 uint64_t push_sod
= 0;
1446 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1449 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1451 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1452 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1454 mutex_enter(&itxg
->itxg_lock
);
1455 if (itxg
->itxg_txg
!= txg
) {
1456 mutex_exit(&itxg
->itxg_lock
);
1460 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
1461 push_sod
+= itxg
->itxg_sod
;
1464 mutex_exit(&itxg
->itxg_lock
);
1466 atomic_add_64(&zilog
->zl_itx_list_sz
, -push_sod
);
1470 * Move the async itxs for a specified object to commit into sync lists.
1473 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
1476 itx_async_node_t
*ian
;
1480 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1483 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1485 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1486 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1488 mutex_enter(&itxg
->itxg_lock
);
1489 if (itxg
->itxg_txg
!= txg
) {
1490 mutex_exit(&itxg
->itxg_lock
);
1495 * If a foid is specified then find that node and append its
1496 * list. Otherwise walk the tree appending all the lists
1497 * to the sync list. We add to the end rather than the
1498 * beginning to ensure the create has happened.
1500 t
= &itxg
->itxg_itxs
->i_async_tree
;
1502 ian
= avl_find(t
, &foid
, &where
);
1504 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1508 void *cookie
= NULL
;
1510 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1511 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1513 list_destroy(&ian
->ia_list
);
1514 kmem_free(ian
, sizeof (itx_async_node_t
));
1517 mutex_exit(&itxg
->itxg_lock
);
1522 zil_commit_writer(zilog_t
*zilog
)
1527 spa_t
*spa
= zilog
->zl_spa
;
1530 ASSERT(zilog
->zl_root_zio
== NULL
);
1532 mutex_exit(&zilog
->zl_lock
);
1534 zil_get_commit_list(zilog
);
1537 * Return if there's nothing to commit before we dirty the fs by
1538 * calling zil_create().
1540 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
) {
1541 mutex_enter(&zilog
->zl_lock
);
1545 if (zilog
->zl_suspend
) {
1548 lwb
= list_tail(&zilog
->zl_lwb_list
);
1550 lwb
= zil_create(zilog
);
1553 DTRACE_PROBE1(zil__cw1
, zilog_t
*, zilog
);
1554 for (itx
= list_head(&zilog
->zl_itx_commit_list
); itx
!= NULL
;
1555 itx
= list_next(&zilog
->zl_itx_commit_list
, itx
)) {
1556 txg
= itx
->itx_lr
.lrc_txg
;
1559 if (txg
> spa_last_synced_txg(spa
) || txg
> spa_freeze_txg(spa
))
1560 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
1562 DTRACE_PROBE1(zil__cw2
, zilog_t
*, zilog
);
1564 /* write the last block out */
1565 if (lwb
!= NULL
&& lwb
->lwb_zio
!= NULL
)
1566 lwb
= zil_lwb_write_start(zilog
, lwb
);
1568 zilog
->zl_cur_used
= 0;
1571 * Wait if necessary for the log blocks to be on stable storage.
1573 if (zilog
->zl_root_zio
) {
1574 error
= zio_wait(zilog
->zl_root_zio
);
1575 zilog
->zl_root_zio
= NULL
;
1576 zil_flush_vdevs(zilog
);
1579 if (error
|| lwb
== NULL
)
1580 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
1582 while ((itx
= list_head(&zilog
->zl_itx_commit_list
))) {
1583 txg
= itx
->itx_lr
.lrc_txg
;
1586 if (itx
->itx_callback
!= NULL
)
1587 itx
->itx_callback(itx
->itx_callback_data
);
1588 list_remove(&zilog
->zl_itx_commit_list
, itx
);
1589 zil_itx_destroy(itx
);
1592 mutex_enter(&zilog
->zl_lock
);
1595 * Remember the highest committed log sequence number for ztest.
1596 * We only update this value when all the log writes succeeded,
1597 * because ztest wants to ASSERT that it got the whole log chain.
1599 if (error
== 0 && lwb
!= NULL
)
1600 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1604 * Commit zfs transactions to stable storage.
1605 * If foid is 0 push out all transactions, otherwise push only those
1606 * for that object or might reference that object.
1608 * itxs are committed in batches. In a heavily stressed zil there will be
1609 * a commit writer thread who is writing out a bunch of itxs to the log
1610 * for a set of committing threads (cthreads) in the same batch as the writer.
1611 * Those cthreads are all waiting on the same cv for that batch.
1613 * There will also be a different and growing batch of threads that are
1614 * waiting to commit (qthreads). When the committing batch completes
1615 * a transition occurs such that the cthreads exit and the qthreads become
1616 * cthreads. One of the new cthreads becomes the writer thread for the
1617 * batch. Any new threads arriving become new qthreads.
1619 * Only 2 condition variables are needed and there's no transition
1620 * between the two cvs needed. They just flip-flop between qthreads
1623 * Using this scheme we can efficiently wakeup up only those threads
1624 * that have been committed.
1627 zil_commit(zilog_t
*zilog
, uint64_t foid
)
1631 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
1634 ZIL_STAT_BUMP(zil_commit_count
);
1636 /* move the async itxs for the foid to the sync queues */
1637 zil_async_to_sync(zilog
, foid
);
1639 mutex_enter(&zilog
->zl_lock
);
1640 mybatch
= zilog
->zl_next_batch
;
1641 while (zilog
->zl_writer
) {
1642 cv_wait(&zilog
->zl_cv_batch
[mybatch
& 1], &zilog
->zl_lock
);
1643 if (mybatch
<= zilog
->zl_com_batch
) {
1644 mutex_exit(&zilog
->zl_lock
);
1649 zilog
->zl_next_batch
++;
1650 zilog
->zl_writer
= B_TRUE
;
1651 ZIL_STAT_BUMP(zil_commit_writer_count
);
1652 zil_commit_writer(zilog
);
1653 zilog
->zl_com_batch
= mybatch
;
1654 zilog
->zl_writer
= B_FALSE
;
1656 /* wake up one thread to become the next writer */
1657 cv_signal(&zilog
->zl_cv_batch
[(mybatch
+1) & 1]);
1659 /* wake up all threads waiting for this batch to be committed */
1660 cv_broadcast(&zilog
->zl_cv_batch
[mybatch
& 1]);
1662 mutex_exit(&zilog
->zl_lock
);
1666 * Called in syncing context to free committed log blocks and update log header.
1669 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1671 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1672 uint64_t txg
= dmu_tx_get_txg(tx
);
1673 spa_t
*spa
= zilog
->zl_spa
;
1674 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
1678 * We don't zero out zl_destroy_txg, so make sure we don't try
1679 * to destroy it twice.
1681 if (spa_sync_pass(spa
) != 1)
1684 mutex_enter(&zilog
->zl_lock
);
1686 ASSERT(zilog
->zl_stop_sync
== 0);
1688 if (*replayed_seq
!= 0) {
1689 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
1690 zh
->zh_replay_seq
= *replayed_seq
;
1694 if (zilog
->zl_destroy_txg
== txg
) {
1695 blkptr_t blk
= zh
->zh_log
;
1697 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
1699 bzero(zh
, sizeof (zil_header_t
));
1700 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
1702 if (zilog
->zl_keep_first
) {
1704 * If this block was part of log chain that couldn't
1705 * be claimed because a device was missing during
1706 * zil_claim(), but that device later returns,
1707 * then this block could erroneously appear valid.
1708 * To guard against this, assign a new GUID to the new
1709 * log chain so it doesn't matter what blk points to.
1711 zil_init_log_chain(zilog
, &blk
);
1716 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
1717 zh
->zh_log
= lwb
->lwb_blk
;
1718 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
1721 ASSERT(lwb
->lwb_zio
== NULL
);
1723 list_remove(&zilog
->zl_lwb_list
, lwb
);
1724 zio_free_zil(spa
, txg
, &lwb
->lwb_blk
);
1725 kmem_cache_free(zil_lwb_cache
, lwb
);
1728 * If we don't have anything left in the lwb list then
1729 * we've had an allocation failure and we need to zero
1730 * out the zil_header blkptr so that we don't end
1731 * up freeing the same block twice.
1733 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
1734 BP_ZERO(&zh
->zh_log
);
1738 * Remove fastwrite on any blocks that have been pre-allocated for
1739 * the next commit. This prevents fastwrite counter pollution by
1740 * unused, long-lived LWBs.
1742 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
1743 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_zio
) {
1744 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1745 lwb
->lwb_fastwrite
= 0;
1749 mutex_exit(&zilog
->zl_lock
);
1755 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
1756 sizeof (struct lwb
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
1758 zil_ksp
= kstat_create("zfs", 0, "zil", "misc",
1759 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
1760 KSTAT_FLAG_VIRTUAL
);
1762 if (zil_ksp
!= NULL
) {
1763 zil_ksp
->ks_data
= &zil_stats
;
1764 kstat_install(zil_ksp
);
1771 kmem_cache_destroy(zil_lwb_cache
);
1773 if (zil_ksp
!= NULL
) {
1774 kstat_delete(zil_ksp
);
1780 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
1782 zilog
->zl_sync
= sync
;
1786 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
1788 zilog
->zl_logbias
= logbias
;
1792 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
1797 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
1799 zilog
->zl_header
= zh_phys
;
1801 zilog
->zl_spa
= dmu_objset_spa(os
);
1802 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
1803 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
1804 zilog
->zl_logbias
= dmu_objset_logbias(os
);
1805 zilog
->zl_sync
= dmu_objset_syncprop(os
);
1806 zilog
->zl_next_batch
= 1;
1808 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1810 for (i
= 0; i
< TXG_SIZE
; i
++) {
1811 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
1812 MUTEX_DEFAULT
, NULL
);
1815 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
1816 offsetof(lwb_t
, lwb_node
));
1818 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
1819 offsetof(itx_t
, itx_node
));
1821 mutex_init(&zilog
->zl_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1823 avl_create(&zilog
->zl_vdev_tree
, zil_vdev_compare
,
1824 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
1826 cv_init(&zilog
->zl_cv_writer
, NULL
, CV_DEFAULT
, NULL
);
1827 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
1828 cv_init(&zilog
->zl_cv_batch
[0], NULL
, CV_DEFAULT
, NULL
);
1829 cv_init(&zilog
->zl_cv_batch
[1], NULL
, CV_DEFAULT
, NULL
);
1835 zil_free(zilog_t
*zilog
)
1839 zilog
->zl_stop_sync
= 1;
1841 ASSERT0(zilog
->zl_suspend
);
1842 ASSERT0(zilog
->zl_suspending
);
1844 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1845 list_destroy(&zilog
->zl_lwb_list
);
1847 avl_destroy(&zilog
->zl_vdev_tree
);
1848 mutex_destroy(&zilog
->zl_vdev_lock
);
1850 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
1851 list_destroy(&zilog
->zl_itx_commit_list
);
1853 for (i
= 0; i
< TXG_SIZE
; i
++) {
1855 * It's possible for an itx to be generated that doesn't dirty
1856 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1857 * callback to remove the entry. We remove those here.
1859 * Also free up the ziltest itxs.
1861 if (zilog
->zl_itxg
[i
].itxg_itxs
)
1862 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
1863 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
1866 mutex_destroy(&zilog
->zl_lock
);
1868 cv_destroy(&zilog
->zl_cv_writer
);
1869 cv_destroy(&zilog
->zl_cv_suspend
);
1870 cv_destroy(&zilog
->zl_cv_batch
[0]);
1871 cv_destroy(&zilog
->zl_cv_batch
[1]);
1873 kmem_free(zilog
, sizeof (zilog_t
));
1877 * Open an intent log.
1880 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
1882 zilog_t
*zilog
= dmu_objset_zil(os
);
1884 ASSERT(zilog
->zl_clean_taskq
== NULL
);
1885 ASSERT(zilog
->zl_get_data
== NULL
);
1886 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1888 zilog
->zl_get_data
= get_data
;
1889 zilog
->zl_clean_taskq
= taskq_create("zil_clean", 1, minclsyspri
,
1890 2, 2, TASKQ_PREPOPULATE
);
1896 * Close an intent log.
1899 zil_close(zilog_t
*zilog
)
1904 zil_commit(zilog
, 0); /* commit all itx */
1907 * The lwb_max_txg for the stubby lwb will reflect the last activity
1908 * for the zil. After a txg_wait_synced() on the txg we know all the
1909 * callbacks have occurred that may clean the zil. Only then can we
1910 * destroy the zl_clean_taskq.
1912 mutex_enter(&zilog
->zl_lock
);
1913 lwb
= list_tail(&zilog
->zl_lwb_list
);
1915 txg
= lwb
->lwb_max_txg
;
1916 mutex_exit(&zilog
->zl_lock
);
1918 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1919 ASSERT(!zilog_is_dirty(zilog
));
1921 taskq_destroy(zilog
->zl_clean_taskq
);
1922 zilog
->zl_clean_taskq
= NULL
;
1923 zilog
->zl_get_data
= NULL
;
1926 * We should have only one LWB left on the list; remove it now.
1928 mutex_enter(&zilog
->zl_lock
);
1929 lwb
= list_head(&zilog
->zl_lwb_list
);
1931 ASSERT(lwb
== list_tail(&zilog
->zl_lwb_list
));
1932 ASSERT(lwb
->lwb_zio
== NULL
);
1933 if (lwb
->lwb_fastwrite
)
1934 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1935 list_remove(&zilog
->zl_lwb_list
, lwb
);
1936 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1937 kmem_cache_free(zil_lwb_cache
, lwb
);
1939 mutex_exit(&zilog
->zl_lock
);
1942 static char *suspend_tag
= "zil suspending";
1945 * Suspend an intent log. While in suspended mode, we still honor
1946 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1947 * On old version pools, we suspend the log briefly when taking a
1948 * snapshot so that it will have an empty intent log.
1950 * Long holds are not really intended to be used the way we do here --
1951 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1952 * could fail. Therefore we take pains to only put a long hold if it is
1953 * actually necessary. Fortunately, it will only be necessary if the
1954 * objset is currently mounted (or the ZVOL equivalent). In that case it
1955 * will already have a long hold, so we are not really making things any worse.
1957 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1958 * zvol_state_t), and use their mechanism to prevent their hold from being
1959 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1962 * if cookiep == NULL, this does both the suspend & resume.
1963 * Otherwise, it returns with the dataset "long held", and the cookie
1964 * should be passed into zil_resume().
1967 zil_suspend(const char *osname
, void **cookiep
)
1971 const zil_header_t
*zh
;
1974 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
1977 zilog
= dmu_objset_zil(os
);
1979 mutex_enter(&zilog
->zl_lock
);
1980 zh
= zilog
->zl_header
;
1982 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
1983 mutex_exit(&zilog
->zl_lock
);
1984 dmu_objset_rele(os
, suspend_tag
);
1985 return (SET_ERROR(EBUSY
));
1989 * Don't put a long hold in the cases where we can avoid it. This
1990 * is when there is no cookie so we are doing a suspend & resume
1991 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1992 * for the suspend because it's already suspended, or there's no ZIL.
1994 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
1995 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
1996 mutex_exit(&zilog
->zl_lock
);
1997 dmu_objset_rele(os
, suspend_tag
);
2001 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
2002 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
2004 zilog
->zl_suspend
++;
2006 if (zilog
->zl_suspend
> 1) {
2008 * Someone else is already suspending it.
2009 * Just wait for them to finish.
2012 while (zilog
->zl_suspending
)
2013 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
2014 mutex_exit(&zilog
->zl_lock
);
2016 if (cookiep
== NULL
)
2024 * If there is no pointer to an on-disk block, this ZIL must not
2025 * be active (e.g. filesystem not mounted), so there's nothing
2028 if (BP_IS_HOLE(&zh
->zh_log
)) {
2029 ASSERT(cookiep
!= NULL
); /* fast path already handled */
2032 mutex_exit(&zilog
->zl_lock
);
2036 zilog
->zl_suspending
= B_TRUE
;
2037 mutex_exit(&zilog
->zl_lock
);
2039 zil_commit(zilog
, 0);
2041 zil_destroy(zilog
, B_FALSE
);
2043 mutex_enter(&zilog
->zl_lock
);
2044 zilog
->zl_suspending
= B_FALSE
;
2045 cv_broadcast(&zilog
->zl_cv_suspend
);
2046 mutex_exit(&zilog
->zl_lock
);
2048 if (cookiep
== NULL
)
2056 zil_resume(void *cookie
)
2058 objset_t
*os
= cookie
;
2059 zilog_t
*zilog
= dmu_objset_zil(os
);
2061 mutex_enter(&zilog
->zl_lock
);
2062 ASSERT(zilog
->zl_suspend
!= 0);
2063 zilog
->zl_suspend
--;
2064 mutex_exit(&zilog
->zl_lock
);
2065 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
2066 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
2069 typedef struct zil_replay_arg
{
2070 zil_replay_func_t
*zr_replay
;
2072 boolean_t zr_byteswap
;
2077 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
2079 char name
[MAXNAMELEN
];
2081 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
2083 dmu_objset_name(zilog
->zl_os
, name
);
2085 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
2086 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
2087 (u_longlong_t
)lr
->lrc_seq
,
2088 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
2089 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
2095 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
2097 zil_replay_arg_t
*zr
= zra
;
2098 const zil_header_t
*zh
= zilog
->zl_header
;
2099 uint64_t reclen
= lr
->lrc_reclen
;
2100 uint64_t txtype
= lr
->lrc_txtype
;
2103 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
2105 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
2108 if (lr
->lrc_txg
< claim_txg
) /* already committed */
2111 /* Strip case-insensitive bit, still present in log record */
2114 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
2115 return (zil_replay_error(zilog
, lr
, EINVAL
));
2118 * If this record type can be logged out of order, the object
2119 * (lr_foid) may no longer exist. That's legitimate, not an error.
2121 if (TX_OOO(txtype
)) {
2122 error
= dmu_object_info(zilog
->zl_os
,
2123 ((lr_ooo_t
*)lr
)->lr_foid
, NULL
);
2124 if (error
== ENOENT
|| error
== EEXIST
)
2129 * Make a copy of the data so we can revise and extend it.
2131 bcopy(lr
, zr
->zr_lr
, reclen
);
2134 * If this is a TX_WRITE with a blkptr, suck in the data.
2136 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
2137 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
2138 zr
->zr_lr
+ reclen
);
2140 return (zil_replay_error(zilog
, lr
, error
));
2144 * The log block containing this lr may have been byteswapped
2145 * so that we can easily examine common fields like lrc_txtype.
2146 * However, the log is a mix of different record types, and only the
2147 * replay vectors know how to byteswap their records. Therefore, if
2148 * the lr was byteswapped, undo it before invoking the replay vector.
2150 if (zr
->zr_byteswap
)
2151 byteswap_uint64_array(zr
->zr_lr
, reclen
);
2154 * We must now do two things atomically: replay this log record,
2155 * and update the log header sequence number to reflect the fact that
2156 * we did so. At the end of each replay function the sequence number
2157 * is updated if we are in replay mode.
2159 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
2162 * The DMU's dnode layer doesn't see removes until the txg
2163 * commits, so a subsequent claim can spuriously fail with
2164 * EEXIST. So if we receive any error we try syncing out
2165 * any removes then retry the transaction. Note that we
2166 * specify B_FALSE for byteswap now, so we don't do it twice.
2168 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
2169 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
2171 return (zil_replay_error(zilog
, lr
, error
));
2178 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
2180 zilog
->zl_replay_blks
++;
2186 * If this dataset has a non-empty intent log, replay it and destroy it.
2189 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t replay_func
[TX_MAX_TYPE
])
2191 zilog_t
*zilog
= dmu_objset_zil(os
);
2192 const zil_header_t
*zh
= zilog
->zl_header
;
2193 zil_replay_arg_t zr
;
2195 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
2196 zil_destroy(zilog
, B_TRUE
);
2200 zr
.zr_replay
= replay_func
;
2202 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
2203 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
2206 * Wait for in-progress removes to sync before starting replay.
2208 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2210 zilog
->zl_replay
= B_TRUE
;
2211 zilog
->zl_replay_time
= ddi_get_lbolt();
2212 ASSERT(zilog
->zl_replay_blks
== 0);
2213 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
2215 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
2217 zil_destroy(zilog
, B_FALSE
);
2218 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
2219 zilog
->zl_replay
= B_FALSE
;
2223 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
2225 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
2228 if (zilog
->zl_replay
) {
2229 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
2230 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
2231 zilog
->zl_replaying_seq
;
2240 zil_vdev_offline(const char *osname
, void *arg
)
2244 error
= zil_suspend(osname
, NULL
);
2246 return (SET_ERROR(EEXIST
));
2250 #if defined(_KERNEL) && defined(HAVE_SPL)
2251 EXPORT_SYMBOL(zil_alloc
);
2252 EXPORT_SYMBOL(zil_free
);
2253 EXPORT_SYMBOL(zil_open
);
2254 EXPORT_SYMBOL(zil_close
);
2255 EXPORT_SYMBOL(zil_replay
);
2256 EXPORT_SYMBOL(zil_replaying
);
2257 EXPORT_SYMBOL(zil_destroy
);
2258 EXPORT_SYMBOL(zil_destroy_sync
);
2259 EXPORT_SYMBOL(zil_itx_create
);
2260 EXPORT_SYMBOL(zil_itx_destroy
);
2261 EXPORT_SYMBOL(zil_itx_assign
);
2262 EXPORT_SYMBOL(zil_commit
);
2263 EXPORT_SYMBOL(zil_vdev_offline
);
2264 EXPORT_SYMBOL(zil_claim
);
2265 EXPORT_SYMBOL(zil_check_log_chain
);
2266 EXPORT_SYMBOL(zil_sync
);
2267 EXPORT_SYMBOL(zil_clean
);
2268 EXPORT_SYMBOL(zil_suspend
);
2269 EXPORT_SYMBOL(zil_resume
);
2270 EXPORT_SYMBOL(zil_add_block
);
2271 EXPORT_SYMBOL(zil_bp_tree_add
);
2272 EXPORT_SYMBOL(zil_set_sync
);
2273 EXPORT_SYMBOL(zil_set_logbias
);
2275 module_param(zil_replay_disable
, int, 0644);
2276 MODULE_PARM_DESC(zil_replay_disable
, "Disable intent logging replay");
2278 module_param(zfs_nocacheflush
, int, 0644);
2279 MODULE_PARM_DESC(zfs_nocacheflush
, "Disable cache flushes");
2281 module_param(zil_slog_limit
, ulong
, 0644);
2282 MODULE_PARM_DESC(zil_slog_limit
, "Max commit bytes to separate log device");