xref: /illumos-gate/usr/src/uts/common/fs/zfs/zil.c (revision 1007fd6fd24227460e77ce89f5ca85641a85a576)
1 /*
2  * CDDL HEADER START
3  *
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.
7  *
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.
12  *
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]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /* Portions Copyright 2010 Robert Milkowski */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/spa.h>
29 #include <sys/dmu.h>
30 #include <sys/zap.h>
31 #include <sys/arc.h>
32 #include <sys/stat.h>
33 #include <sys/resource.h>
34 #include <sys/zil.h>
35 #include <sys/zil_impl.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/vdev_impl.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/dsl_pool.h>
40 
41 /*
42  * The zfs intent log (ZIL) saves transaction records of system calls
43  * that change the file system in memory with enough information
44  * to be able to replay them. These are stored in memory until
45  * either the DMU transaction group (txg) commits them to the stable pool
46  * and they can be discarded, or they are flushed to the stable log
47  * (also in the pool) due to a fsync, O_DSYNC or other synchronous
48  * requirement. In the event of a panic or power fail then those log
49  * records (transactions) are replayed.
50  *
51  * There is one ZIL per file system. Its on-disk (pool) format consists
52  * of 3 parts:
53  *
54  * 	- ZIL header
55  * 	- ZIL blocks
56  * 	- ZIL records
57  *
58  * A log record holds a system call transaction. Log blocks can
59  * hold many log records and the blocks are chained together.
60  * Each ZIL block contains a block pointer (blkptr_t) to the next
61  * ZIL block in the chain. The ZIL header points to the first
62  * block in the chain. Note there is not a fixed place in the pool
63  * to hold blocks. They are dynamically allocated and freed as
64  * needed from the blocks available. Figure X shows the ZIL structure:
65  */
66 
67 /*
68  * This global ZIL switch affects all pools
69  */
70 int zil_replay_disable = 0;    /* disable intent logging replay */
71 
72 /*
73  * Tunable parameter for debugging or performance analysis.  Setting
74  * zfs_nocacheflush will cause corruption on power loss if a volatile
75  * out-of-order write cache is enabled.
76  */
77 boolean_t zfs_nocacheflush = B_FALSE;
78 
79 static kmem_cache_t *zil_lwb_cache;
80 
81 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
82 
83 #define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
84     sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
85 
86 
87 /*
88  * ziltest is by and large an ugly hack, but very useful in
89  * checking replay without tedious work.
90  * When running ziltest we want to keep all itx's and so maintain
91  * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
92  * We subtract TXG_CONCURRENT_STATES to allow for common code.
93  */
94 #define	ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
95 
96 static int
97 zil_bp_compare(const void *x1, const void *x2)
98 {
99 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
100 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
101 
102 	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
103 		return (-1);
104 	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
105 		return (1);
106 
107 	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
108 		return (-1);
109 	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
110 		return (1);
111 
112 	return (0);
113 }
114 
115 static void
116 zil_bp_tree_init(zilog_t *zilog)
117 {
118 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
119 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
120 }
121 
122 static void
123 zil_bp_tree_fini(zilog_t *zilog)
124 {
125 	avl_tree_t *t = &zilog->zl_bp_tree;
126 	zil_bp_node_t *zn;
127 	void *cookie = NULL;
128 
129 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
130 		kmem_free(zn, sizeof (zil_bp_node_t));
131 
132 	avl_destroy(t);
133 }
134 
135 int
136 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
137 {
138 	avl_tree_t *t = &zilog->zl_bp_tree;
139 	const dva_t *dva = BP_IDENTITY(bp);
140 	zil_bp_node_t *zn;
141 	avl_index_t where;
142 
143 	if (avl_find(t, dva, &where) != NULL)
144 		return (EEXIST);
145 
146 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
147 	zn->zn_dva = *dva;
148 	avl_insert(t, zn, where);
149 
150 	return (0);
151 }
152 
153 static zil_header_t *
154 zil_header_in_syncing_context(zilog_t *zilog)
155 {
156 	return ((zil_header_t *)zilog->zl_header);
157 }
158 
159 static void
160 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
161 {
162 	zio_cksum_t *zc = &bp->blk_cksum;
163 
164 	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
165 	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
166 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
167 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
168 }
169 
170 /*
171  * Read a log block and make sure it's valid.
172  */
173 static int
174 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
175     char **end)
176 {
177 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
178 	uint32_t aflags = ARC_WAIT;
179 	arc_buf_t *abuf = NULL;
180 	zbookmark_t zb;
181 	int error;
182 
183 	if (zilog->zl_header->zh_claim_txg == 0)
184 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
185 
186 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
187 		zio_flags |= ZIO_FLAG_SPECULATIVE;
188 
189 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
190 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
191 
192 	error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
193 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
194 
195 	if (error == 0) {
196 		zio_cksum_t cksum = bp->blk_cksum;
197 
198 		/*
199 		 * Validate the checksummed log block.
200 		 *
201 		 * Sequence numbers should be... sequential.  The checksum
202 		 * verifier for the next block should be bp's checksum plus 1.
203 		 *
204 		 * Also check the log chain linkage and size used.
205 		 */
206 		cksum.zc_word[ZIL_ZC_SEQ]++;
207 
208 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
209 			zil_chain_t *zilc = abuf->b_data;
210 			char *lr = (char *)(zilc + 1);
211 			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
212 
213 			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
214 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
215 				error = ECKSUM;
216 			} else {
217 				bcopy(lr, dst, len);
218 				*end = (char *)dst + len;
219 				*nbp = zilc->zc_next_blk;
220 			}
221 		} else {
222 			char *lr = abuf->b_data;
223 			uint64_t size = BP_GET_LSIZE(bp);
224 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
225 
226 			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
227 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
228 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
229 				error = ECKSUM;
230 			} else {
231 				bcopy(lr, dst, zilc->zc_nused);
232 				*end = (char *)dst + zilc->zc_nused;
233 				*nbp = zilc->zc_next_blk;
234 			}
235 		}
236 
237 		VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
238 	}
239 
240 	return (error);
241 }
242 
243 /*
244  * Read a TX_WRITE log data block.
245  */
246 static int
247 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
248 {
249 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
250 	const blkptr_t *bp = &lr->lr_blkptr;
251 	uint32_t aflags = ARC_WAIT;
252 	arc_buf_t *abuf = NULL;
253 	zbookmark_t zb;
254 	int error;
255 
256 	if (BP_IS_HOLE(bp)) {
257 		if (wbuf != NULL)
258 			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
259 		return (0);
260 	}
261 
262 	if (zilog->zl_header->zh_claim_txg == 0)
263 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
264 
265 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
266 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
267 
268 	error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
269 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
270 
271 	if (error == 0) {
272 		if (wbuf != NULL)
273 			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
274 		(void) arc_buf_remove_ref(abuf, &abuf);
275 	}
276 
277 	return (error);
278 }
279 
280 /*
281  * Parse the intent log, and call parse_func for each valid record within.
282  */
283 int
284 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
285     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
286 {
287 	const zil_header_t *zh = zilog->zl_header;
288 	boolean_t claimed = !!zh->zh_claim_txg;
289 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
290 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
291 	uint64_t max_blk_seq = 0;
292 	uint64_t max_lr_seq = 0;
293 	uint64_t blk_count = 0;
294 	uint64_t lr_count = 0;
295 	blkptr_t blk, next_blk;
296 	char *lrbuf, *lrp;
297 	int error = 0;
298 
299 	/*
300 	 * Old logs didn't record the maximum zh_claim_lr_seq.
301 	 */
302 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
303 		claim_lr_seq = UINT64_MAX;
304 
305 	/*
306 	 * Starting at the block pointed to by zh_log we read the log chain.
307 	 * For each block in the chain we strongly check that block to
308 	 * ensure its validity.  We stop when an invalid block is found.
309 	 * For each block pointer in the chain we call parse_blk_func().
310 	 * For each record in each valid block we call parse_lr_func().
311 	 * If the log has been claimed, stop if we encounter a sequence
312 	 * number greater than the highest claimed sequence number.
313 	 */
314 	lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
315 	zil_bp_tree_init(zilog);
316 
317 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
318 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
319 		int reclen;
320 		char *end;
321 
322 		if (blk_seq > claim_blk_seq)
323 			break;
324 		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
325 			break;
326 		ASSERT3U(max_blk_seq, <, blk_seq);
327 		max_blk_seq = blk_seq;
328 		blk_count++;
329 
330 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
331 			break;
332 
333 		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
334 		if (error)
335 			break;
336 
337 		for (lrp = lrbuf; lrp < end; lrp += reclen) {
338 			lr_t *lr = (lr_t *)lrp;
339 			reclen = lr->lrc_reclen;
340 			ASSERT3U(reclen, >=, sizeof (lr_t));
341 			if (lr->lrc_seq > claim_lr_seq)
342 				goto done;
343 			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
344 				goto done;
345 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
346 			max_lr_seq = lr->lrc_seq;
347 			lr_count++;
348 		}
349 	}
350 done:
351 	zilog->zl_parse_error = error;
352 	zilog->zl_parse_blk_seq = max_blk_seq;
353 	zilog->zl_parse_lr_seq = max_lr_seq;
354 	zilog->zl_parse_blk_count = blk_count;
355 	zilog->zl_parse_lr_count = lr_count;
356 
357 	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
358 	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
359 
360 	zil_bp_tree_fini(zilog);
361 	zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
362 
363 	return (error);
364 }
365 
366 static int
367 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
368 {
369 	/*
370 	 * Claim log block if not already committed and not already claimed.
371 	 * If tx == NULL, just verify that the block is claimable.
372 	 */
373 	if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
374 		return (0);
375 
376 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
377 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
378 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
379 }
380 
381 static int
382 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
383 {
384 	lr_write_t *lr = (lr_write_t *)lrc;
385 	int error;
386 
387 	if (lrc->lrc_txtype != TX_WRITE)
388 		return (0);
389 
390 	/*
391 	 * If the block is not readable, don't claim it.  This can happen
392 	 * in normal operation when a log block is written to disk before
393 	 * some of the dmu_sync() blocks it points to.  In this case, the
394 	 * transaction cannot have been committed to anyone (we would have
395 	 * waited for all writes to be stable first), so it is semantically
396 	 * correct to declare this the end of the log.
397 	 */
398 	if (lr->lr_blkptr.blk_birth >= first_txg &&
399 	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
400 		return (error);
401 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
402 }
403 
404 /* ARGSUSED */
405 static int
406 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
407 {
408 	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
409 
410 	return (0);
411 }
412 
413 static int
414 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
415 {
416 	lr_write_t *lr = (lr_write_t *)lrc;
417 	blkptr_t *bp = &lr->lr_blkptr;
418 
419 	/*
420 	 * If we previously claimed it, we need to free it.
421 	 */
422 	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
423 	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
424 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
425 
426 	return (0);
427 }
428 
429 static lwb_t *
430 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
431 {
432 	lwb_t *lwb;
433 
434 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
435 	lwb->lwb_zilog = zilog;
436 	lwb->lwb_blk = *bp;
437 	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
438 	lwb->lwb_max_txg = txg;
439 	lwb->lwb_zio = NULL;
440 	lwb->lwb_tx = NULL;
441 	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
442 		lwb->lwb_nused = sizeof (zil_chain_t);
443 		lwb->lwb_sz = BP_GET_LSIZE(bp);
444 	} else {
445 		lwb->lwb_nused = 0;
446 		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
447 	}
448 
449 	mutex_enter(&zilog->zl_lock);
450 	list_insert_tail(&zilog->zl_lwb_list, lwb);
451 	mutex_exit(&zilog->zl_lock);
452 
453 	return (lwb);
454 }
455 
456 /*
457  * Create an on-disk intent log.
458  */
459 static lwb_t *
460 zil_create(zilog_t *zilog)
461 {
462 	const zil_header_t *zh = zilog->zl_header;
463 	lwb_t *lwb = NULL;
464 	uint64_t txg = 0;
465 	dmu_tx_t *tx = NULL;
466 	blkptr_t blk;
467 	int error = 0;
468 
469 	/*
470 	 * Wait for any previous destroy to complete.
471 	 */
472 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
473 
474 	ASSERT(zh->zh_claim_txg == 0);
475 	ASSERT(zh->zh_replay_seq == 0);
476 
477 	blk = zh->zh_log;
478 
479 	/*
480 	 * Allocate an initial log block if:
481 	 *    - there isn't one already
482 	 *    - the existing block is the wrong endianess
483 	 */
484 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
485 		tx = dmu_tx_create(zilog->zl_os);
486 		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
487 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
488 		txg = dmu_tx_get_txg(tx);
489 
490 		if (!BP_IS_HOLE(&blk)) {
491 			zio_free_zil(zilog->zl_spa, txg, &blk);
492 			BP_ZERO(&blk);
493 		}
494 
495 		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
496 		    ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
497 
498 		if (error == 0)
499 			zil_init_log_chain(zilog, &blk);
500 	}
501 
502 	/*
503 	 * Allocate a log write buffer (lwb) for the first log block.
504 	 */
505 	if (error == 0)
506 		lwb = zil_alloc_lwb(zilog, &blk, txg);
507 
508 	/*
509 	 * If we just allocated the first log block, commit our transaction
510 	 * and wait for zil_sync() to stuff the block poiner into zh_log.
511 	 * (zh is part of the MOS, so we cannot modify it in open context.)
512 	 */
513 	if (tx != NULL) {
514 		dmu_tx_commit(tx);
515 		txg_wait_synced(zilog->zl_dmu_pool, txg);
516 	}
517 
518 	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
519 
520 	return (lwb);
521 }
522 
523 /*
524  * In one tx, free all log blocks and clear the log header.
525  * If keep_first is set, then we're replaying a log with no content.
526  * We want to keep the first block, however, so that the first
527  * synchronous transaction doesn't require a txg_wait_synced()
528  * in zil_create().  We don't need to txg_wait_synced() here either
529  * when keep_first is set, because both zil_create() and zil_destroy()
530  * will wait for any in-progress destroys to complete.
531  */
532 void
533 zil_destroy(zilog_t *zilog, boolean_t keep_first)
534 {
535 	const zil_header_t *zh = zilog->zl_header;
536 	lwb_t *lwb;
537 	dmu_tx_t *tx;
538 	uint64_t txg;
539 
540 	/*
541 	 * Wait for any previous destroy to complete.
542 	 */
543 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
544 
545 	zilog->zl_old_header = *zh;		/* debugging aid */
546 
547 	if (BP_IS_HOLE(&zh->zh_log))
548 		return;
549 
550 	tx = dmu_tx_create(zilog->zl_os);
551 	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
552 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
553 	txg = dmu_tx_get_txg(tx);
554 
555 	mutex_enter(&zilog->zl_lock);
556 
557 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
558 	zilog->zl_destroy_txg = txg;
559 	zilog->zl_keep_first = keep_first;
560 
561 	if (!list_is_empty(&zilog->zl_lwb_list)) {
562 		ASSERT(zh->zh_claim_txg == 0);
563 		ASSERT(!keep_first);
564 		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
565 			list_remove(&zilog->zl_lwb_list, lwb);
566 			if (lwb->lwb_buf != NULL)
567 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
568 			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
569 			kmem_cache_free(zil_lwb_cache, lwb);
570 		}
571 	} else if (!keep_first) {
572 		(void) zil_parse(zilog, zil_free_log_block,
573 		    zil_free_log_record, tx, zh->zh_claim_txg);
574 	}
575 	mutex_exit(&zilog->zl_lock);
576 
577 	dmu_tx_commit(tx);
578 }
579 
580 int
581 zil_claim(const char *osname, void *txarg)
582 {
583 	dmu_tx_t *tx = txarg;
584 	uint64_t first_txg = dmu_tx_get_txg(tx);
585 	zilog_t *zilog;
586 	zil_header_t *zh;
587 	objset_t *os;
588 	int error;
589 
590 	error = dmu_objset_hold(osname, FTAG, &os);
591 	if (error) {
592 		cmn_err(CE_WARN, "can't open objset for %s", osname);
593 		return (0);
594 	}
595 
596 	zilog = dmu_objset_zil(os);
597 	zh = zil_header_in_syncing_context(zilog);
598 
599 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
600 		if (!BP_IS_HOLE(&zh->zh_log))
601 			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
602 		BP_ZERO(&zh->zh_log);
603 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
604 		dmu_objset_rele(os, FTAG);
605 		return (0);
606 	}
607 
608 	/*
609 	 * Claim all log blocks if we haven't already done so, and remember
610 	 * the highest claimed sequence number.  This ensures that if we can
611 	 * read only part of the log now (e.g. due to a missing device),
612 	 * but we can read the entire log later, we will not try to replay
613 	 * or destroy beyond the last block we successfully claimed.
614 	 */
615 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
616 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
617 		(void) zil_parse(zilog, zil_claim_log_block,
618 		    zil_claim_log_record, tx, first_txg);
619 		zh->zh_claim_txg = first_txg;
620 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
621 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
622 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
623 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
624 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
625 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
626 	}
627 
628 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
629 	dmu_objset_rele(os, FTAG);
630 	return (0);
631 }
632 
633 /*
634  * Check the log by walking the log chain.
635  * Checksum errors are ok as they indicate the end of the chain.
636  * Any other error (no device or read failure) returns an error.
637  */
638 int
639 zil_check_log_chain(const char *osname, void *tx)
640 {
641 	zilog_t *zilog;
642 	objset_t *os;
643 	blkptr_t *bp;
644 	int error;
645 
646 	ASSERT(tx == NULL);
647 
648 	error = dmu_objset_hold(osname, FTAG, &os);
649 	if (error) {
650 		cmn_err(CE_WARN, "can't open objset for %s", osname);
651 		return (0);
652 	}
653 
654 	zilog = dmu_objset_zil(os);
655 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
656 
657 	/*
658 	 * Check the first block and determine if it's on a log device
659 	 * which may have been removed or faulted prior to loading this
660 	 * pool.  If so, there's no point in checking the rest of the log
661 	 * as its content should have already been synced to the pool.
662 	 */
663 	if (!BP_IS_HOLE(bp)) {
664 		vdev_t *vd;
665 		boolean_t valid = B_TRUE;
666 
667 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
668 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
669 		if (vd->vdev_islog && vdev_is_dead(vd))
670 			valid = vdev_log_state_valid(vd);
671 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
672 
673 		if (!valid) {
674 			dmu_objset_rele(os, FTAG);
675 			return (0);
676 		}
677 	}
678 
679 	/*
680 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
681 	 * any blocks, but just determine whether it is possible to do so.
682 	 * In addition to checking the log chain, zil_claim_log_block()
683 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
684 	 * which will update spa_max_claim_txg.  See spa_load() for details.
685 	 */
686 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
687 	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
688 
689 	dmu_objset_rele(os, FTAG);
690 
691 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
692 }
693 
694 static int
695 zil_vdev_compare(const void *x1, const void *x2)
696 {
697 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
698 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
699 
700 	if (v1 < v2)
701 		return (-1);
702 	if (v1 > v2)
703 		return (1);
704 
705 	return (0);
706 }
707 
708 void
709 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
710 {
711 	avl_tree_t *t = &zilog->zl_vdev_tree;
712 	avl_index_t where;
713 	zil_vdev_node_t *zv, zvsearch;
714 	int ndvas = BP_GET_NDVAS(bp);
715 	int i;
716 
717 	if (zfs_nocacheflush)
718 		return;
719 
720 	ASSERT(zilog->zl_writer);
721 
722 	/*
723 	 * Even though we're zl_writer, we still need a lock because the
724 	 * zl_get_data() callbacks may have dmu_sync() done callbacks
725 	 * that will run concurrently.
726 	 */
727 	mutex_enter(&zilog->zl_vdev_lock);
728 	for (i = 0; i < ndvas; i++) {
729 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
730 		if (avl_find(t, &zvsearch, &where) == NULL) {
731 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
732 			zv->zv_vdev = zvsearch.zv_vdev;
733 			avl_insert(t, zv, where);
734 		}
735 	}
736 	mutex_exit(&zilog->zl_vdev_lock);
737 }
738 
739 static void
740 zil_flush_vdevs(zilog_t *zilog)
741 {
742 	spa_t *spa = zilog->zl_spa;
743 	avl_tree_t *t = &zilog->zl_vdev_tree;
744 	void *cookie = NULL;
745 	zil_vdev_node_t *zv;
746 	zio_t *zio;
747 
748 	ASSERT(zilog->zl_writer);
749 
750 	/*
751 	 * We don't need zl_vdev_lock here because we're the zl_writer,
752 	 * and all zl_get_data() callbacks are done.
753 	 */
754 	if (avl_numnodes(t) == 0)
755 		return;
756 
757 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
758 
759 	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
760 
761 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
762 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
763 		if (vd != NULL)
764 			zio_flush(zio, vd);
765 		kmem_free(zv, sizeof (*zv));
766 	}
767 
768 	/*
769 	 * Wait for all the flushes to complete.  Not all devices actually
770 	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
771 	 */
772 	(void) zio_wait(zio);
773 
774 	spa_config_exit(spa, SCL_STATE, FTAG);
775 }
776 
777 /*
778  * Function called when a log block write completes
779  */
780 static void
781 zil_lwb_write_done(zio_t *zio)
782 {
783 	lwb_t *lwb = zio->io_private;
784 	zilog_t *zilog = lwb->lwb_zilog;
785 	dmu_tx_t *tx = lwb->lwb_tx;
786 
787 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
788 	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
789 	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
790 	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
791 	ASSERT(!BP_IS_GANG(zio->io_bp));
792 	ASSERT(!BP_IS_HOLE(zio->io_bp));
793 	ASSERT(zio->io_bp->blk_fill == 0);
794 
795 	/*
796 	 * Ensure the lwb buffer pointer is cleared before releasing
797 	 * the txg. If we have had an allocation failure and
798 	 * the txg is waiting to sync then we want want zil_sync()
799 	 * to remove the lwb so that it's not picked up as the next new
800 	 * one in zil_commit_writer(). zil_sync() will only remove
801 	 * the lwb if lwb_buf is null.
802 	 */
803 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
804 	mutex_enter(&zilog->zl_lock);
805 	lwb->lwb_buf = NULL;
806 	lwb->lwb_tx = NULL;
807 	mutex_exit(&zilog->zl_lock);
808 
809 	/*
810 	 * Now that we've written this log block, we have a stable pointer
811 	 * to the next block in the chain, so it's OK to let the txg in
812 	 * which we allocated the next block sync.
813 	 */
814 	dmu_tx_commit(tx);
815 }
816 
817 /*
818  * Initialize the io for a log block.
819  */
820 static void
821 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
822 {
823 	zbookmark_t zb;
824 
825 	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
826 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
827 	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
828 
829 	if (zilog->zl_root_zio == NULL) {
830 		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
831 		    ZIO_FLAG_CANFAIL);
832 	}
833 	if (lwb->lwb_zio == NULL) {
834 		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
835 		    0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
836 		    zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
837 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
838 	}
839 }
840 
841 /*
842  * Define a limited set of intent log block sizes.
843  * These must be a multiple of 4KB. Note only the amount used (again
844  * aligned to 4KB) actually gets written. However, we can't always just
845  * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
846  */
847 uint64_t zil_block_buckets[] = {
848     4096,		/* non TX_WRITE */
849     8192+4096,		/* data base */
850     32*1024 + 4096, 	/* NFS writes */
851     UINT64_MAX
852 };
853 
854 /*
855  * Use the slog as long as the logbias is 'latency' and the current commit size
856  * is less than the limit or the total list size is less than 2X the limit.
857  * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
858  */
859 uint64_t zil_slog_limit = 1024 * 1024;
860 #define	USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
861 	(((zilog)->zl_cur_used < zil_slog_limit) || \
862 	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
863 
864 /*
865  * Start a log block write and advance to the next log block.
866  * Calls are serialized.
867  */
868 static lwb_t *
869 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
870 {
871 	lwb_t *nlwb = NULL;
872 	zil_chain_t *zilc;
873 	spa_t *spa = zilog->zl_spa;
874 	blkptr_t *bp;
875 	dmu_tx_t *tx;
876 	uint64_t txg;
877 	uint64_t zil_blksz, wsz;
878 	int i, error;
879 
880 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
881 		zilc = (zil_chain_t *)lwb->lwb_buf;
882 		bp = &zilc->zc_next_blk;
883 	} else {
884 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
885 		bp = &zilc->zc_next_blk;
886 	}
887 
888 	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
889 
890 	/*
891 	 * Allocate the next block and save its address in this block
892 	 * before writing it in order to establish the log chain.
893 	 * Note that if the allocation of nlwb synced before we wrote
894 	 * the block that points at it (lwb), we'd leak it if we crashed.
895 	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
896 	 * We dirty the dataset to ensure that zil_sync() will be called
897 	 * to clean up in the event of allocation failure or I/O failure.
898 	 */
899 	tx = dmu_tx_create(zilog->zl_os);
900 	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
901 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
902 	txg = dmu_tx_get_txg(tx);
903 
904 	lwb->lwb_tx = tx;
905 
906 	/*
907 	 * Log blocks are pre-allocated. Here we select the size of the next
908 	 * block, based on size used in the last block.
909 	 * - first find the smallest bucket that will fit the block from a
910 	 *   limited set of block sizes. This is because it's faster to write
911 	 *   blocks allocated from the same metaslab as they are adjacent or
912 	 *   close.
913 	 * - next find the maximum from the new suggested size and an array of
914 	 *   previous sizes. This lessens a picket fence effect of wrongly
915 	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
916 	 *   requests.
917 	 *
918 	 * Note we only write what is used, but we can't just allocate
919 	 * the maximum block size because we can exhaust the available
920 	 * pool log space.
921 	 */
922 	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
923 	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
924 		continue;
925 	zil_blksz = zil_block_buckets[i];
926 	if (zil_blksz == UINT64_MAX)
927 		zil_blksz = SPA_MAXBLOCKSIZE;
928 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
929 	for (i = 0; i < ZIL_PREV_BLKS; i++)
930 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
931 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
932 
933 	BP_ZERO(bp);
934 	/* pass the old blkptr in order to spread log blocks across devs */
935 	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
936 	    USE_SLOG(zilog));
937 	if (!error) {
938 		ASSERT3U(bp->blk_birth, ==, txg);
939 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
940 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
941 
942 		/*
943 		 * Allocate a new log write buffer (lwb).
944 		 */
945 		nlwb = zil_alloc_lwb(zilog, bp, txg);
946 
947 		/* Record the block for later vdev flushing */
948 		zil_add_block(zilog, &lwb->lwb_blk);
949 	}
950 
951 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
952 		/* For Slim ZIL only write what is used. */
953 		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
954 		ASSERT3U(wsz, <=, lwb->lwb_sz);
955 		zio_shrink(lwb->lwb_zio, wsz);
956 
957 	} else {
958 		wsz = lwb->lwb_sz;
959 	}
960 
961 	zilc->zc_pad = 0;
962 	zilc->zc_nused = lwb->lwb_nused;
963 	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
964 
965 	/*
966 	 * clear unused data for security
967 	 */
968 	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
969 
970 	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
971 
972 	/*
973 	 * If there was an allocation failure then nlwb will be null which
974 	 * forces a txg_wait_synced().
975 	 */
976 	return (nlwb);
977 }
978 
979 static lwb_t *
980 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
981 {
982 	lr_t *lrc = &itx->itx_lr; /* common log record */
983 	lr_write_t *lrw = (lr_write_t *)lrc;
984 	char *lr_buf;
985 	uint64_t txg = lrc->lrc_txg;
986 	uint64_t reclen = lrc->lrc_reclen;
987 	uint64_t dlen = 0;
988 
989 	if (lwb == NULL)
990 		return (NULL);
991 
992 	ASSERT(lwb->lwb_buf != NULL);
993 
994 	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
995 		dlen = P2ROUNDUP_TYPED(
996 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
997 
998 	zilog->zl_cur_used += (reclen + dlen);
999 
1000 	zil_lwb_write_init(zilog, lwb);
1001 
1002 	/*
1003 	 * If this record won't fit in the current log block, start a new one.
1004 	 */
1005 	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1006 		lwb = zil_lwb_write_start(zilog, lwb);
1007 		if (lwb == NULL)
1008 			return (NULL);
1009 		zil_lwb_write_init(zilog, lwb);
1010 		ASSERT(LWB_EMPTY(lwb));
1011 		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1012 			txg_wait_synced(zilog->zl_dmu_pool, txg);
1013 			return (lwb);
1014 		}
1015 	}
1016 
1017 	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1018 	bcopy(lrc, lr_buf, reclen);
1019 	lrc = (lr_t *)lr_buf;
1020 	lrw = (lr_write_t *)lrc;
1021 
1022 	/*
1023 	 * If it's a write, fetch the data or get its blkptr as appropriate.
1024 	 */
1025 	if (lrc->lrc_txtype == TX_WRITE) {
1026 		if (txg > spa_freeze_txg(zilog->zl_spa))
1027 			txg_wait_synced(zilog->zl_dmu_pool, txg);
1028 		if (itx->itx_wr_state != WR_COPIED) {
1029 			char *dbuf;
1030 			int error;
1031 
1032 			if (dlen) {
1033 				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1034 				dbuf = lr_buf + reclen;
1035 				lrw->lr_common.lrc_reclen += dlen;
1036 			} else {
1037 				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1038 				dbuf = NULL;
1039 			}
1040 			error = zilog->zl_get_data(
1041 			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1042 			if (error == EIO) {
1043 				txg_wait_synced(zilog->zl_dmu_pool, txg);
1044 				return (lwb);
1045 			}
1046 			if (error) {
1047 				ASSERT(error == ENOENT || error == EEXIST ||
1048 				    error == EALREADY);
1049 				return (lwb);
1050 			}
1051 		}
1052 	}
1053 
1054 	/*
1055 	 * We're actually making an entry, so update lrc_seq to be the
1056 	 * log record sequence number.  Note that this is generally not
1057 	 * equal to the itx sequence number because not all transactions
1058 	 * are synchronous, and sometimes spa_sync() gets there first.
1059 	 */
1060 	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1061 	lwb->lwb_nused += reclen + dlen;
1062 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1063 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1064 	ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1065 
1066 	return (lwb);
1067 }
1068 
1069 itx_t *
1070 zil_itx_create(uint64_t txtype, size_t lrsize)
1071 {
1072 	itx_t *itx;
1073 
1074 	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1075 
1076 	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1077 	itx->itx_lr.lrc_txtype = txtype;
1078 	itx->itx_lr.lrc_reclen = lrsize;
1079 	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1080 	itx->itx_lr.lrc_seq = 0;	/* defensive */
1081 	itx->itx_sync = B_TRUE;		/* default is synchronous */
1082 
1083 	return (itx);
1084 }
1085 
1086 void
1087 zil_itx_destroy(itx_t *itx)
1088 {
1089 	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1090 }
1091 
1092 /*
1093  * Free up the sync and async itxs. The itxs_t has already been detached
1094  * so no locks are needed.
1095  */
1096 static void
1097 zil_itxg_clean(itxs_t *itxs)
1098 {
1099 	itx_t *itx;
1100 	list_t *list;
1101 	avl_tree_t *t;
1102 	void *cookie;
1103 	itx_async_node_t *ian;
1104 
1105 	list = &itxs->i_sync_list;
1106 	while ((itx = list_head(list)) != NULL) {
1107 		list_remove(list, itx);
1108 		kmem_free(itx, offsetof(itx_t, itx_lr) +
1109 		    itx->itx_lr.lrc_reclen);
1110 	}
1111 
1112 	cookie = NULL;
1113 	t = &itxs->i_async_tree;
1114 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1115 		list = &ian->ia_list;
1116 		while ((itx = list_head(list)) != NULL) {
1117 			list_remove(list, itx);
1118 			kmem_free(itx, offsetof(itx_t, itx_lr) +
1119 			    itx->itx_lr.lrc_reclen);
1120 		}
1121 		list_destroy(list);
1122 		kmem_free(ian, sizeof (itx_async_node_t));
1123 	}
1124 	avl_destroy(t);
1125 
1126 	kmem_free(itxs, sizeof (itxs_t));
1127 }
1128 
1129 static int
1130 zil_aitx_compare(const void *x1, const void *x2)
1131 {
1132 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1133 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1134 
1135 	if (o1 < o2)
1136 		return (-1);
1137 	if (o1 > o2)
1138 		return (1);
1139 
1140 	return (0);
1141 }
1142 
1143 /*
1144  * Remove all async itx with the given oid.
1145  */
1146 static void
1147 zil_remove_async(zilog_t *zilog, uint64_t oid)
1148 {
1149 	uint64_t otxg, txg;
1150 	itx_async_node_t *ian;
1151 	avl_tree_t *t;
1152 	avl_index_t where;
1153 	list_t clean_list;
1154 	itx_t *itx;
1155 
1156 	ASSERT(oid != 0);
1157 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1158 
1159 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1160 		otxg = ZILTEST_TXG;
1161 	else
1162 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1163 
1164 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1165 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1166 
1167 		mutex_enter(&itxg->itxg_lock);
1168 		if (itxg->itxg_txg != txg) {
1169 			mutex_exit(&itxg->itxg_lock);
1170 			continue;
1171 		}
1172 
1173 		/*
1174 		 * Locate the object node and append its list.
1175 		 */
1176 		t = &itxg->itxg_itxs->i_async_tree;
1177 		ian = avl_find(t, &oid, &where);
1178 		if (ian != NULL)
1179 			list_move_tail(&clean_list, &ian->ia_list);
1180 		mutex_exit(&itxg->itxg_lock);
1181 	}
1182 	while ((itx = list_head(&clean_list)) != NULL) {
1183 		list_remove(&clean_list, itx);
1184 		kmem_free(itx, offsetof(itx_t, itx_lr) +
1185 		    itx->itx_lr.lrc_reclen);
1186 	}
1187 	list_destroy(&clean_list);
1188 }
1189 
1190 void
1191 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1192 {
1193 	uint64_t txg;
1194 	itxg_t *itxg;
1195 	itxs_t *itxs, *clean = NULL;
1196 
1197 	/*
1198 	 * Object ids can be re-instantiated in the next txg so
1199 	 * remove any async transactions to avoid future leaks.
1200 	 * This can happen if a fsync occurs on the re-instantiated
1201 	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1202 	 * the new file data and flushes a write record for the old object.
1203 	 */
1204 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1205 		zil_remove_async(zilog, itx->itx_oid);
1206 
1207 	/*
1208 	 * Ensure the data of a renamed file is committed before the rename.
1209 	 */
1210 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1211 		zil_async_to_sync(zilog, itx->itx_oid);
1212 
1213 	if (spa_freeze_txg(zilog->zl_spa) !=  UINT64_MAX)
1214 		txg = ZILTEST_TXG;
1215 	else
1216 		txg = dmu_tx_get_txg(tx);
1217 
1218 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1219 	mutex_enter(&itxg->itxg_lock);
1220 	itxs = itxg->itxg_itxs;
1221 	if (itxg->itxg_txg != txg) {
1222 		if (itxs != NULL) {
1223 			/*
1224 			 * The zil_clean callback hasn't got around to cleaning
1225 			 * this itxg. Save the itxs for release below.
1226 			 * This should be rare.
1227 			 */
1228 			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1229 			itxg->itxg_sod = 0;
1230 			clean = itxg->itxg_itxs;
1231 		}
1232 		ASSERT(itxg->itxg_sod == 0);
1233 		itxg->itxg_txg = txg;
1234 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1235 
1236 		list_create(&itxs->i_sync_list, sizeof (itx_t),
1237 		    offsetof(itx_t, itx_node));
1238 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1239 		    sizeof (itx_async_node_t),
1240 		    offsetof(itx_async_node_t, ia_node));
1241 	}
1242 	if (itx->itx_sync) {
1243 		list_insert_tail(&itxs->i_sync_list, itx);
1244 		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1245 		itxg->itxg_sod += itx->itx_sod;
1246 	} else {
1247 		avl_tree_t *t = &itxs->i_async_tree;
1248 		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1249 		itx_async_node_t *ian;
1250 		avl_index_t where;
1251 
1252 		ian = avl_find(t, &foid, &where);
1253 		if (ian == NULL) {
1254 			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1255 			list_create(&ian->ia_list, sizeof (itx_t),
1256 			    offsetof(itx_t, itx_node));
1257 			ian->ia_foid = foid;
1258 			avl_insert(t, ian, where);
1259 		}
1260 		list_insert_tail(&ian->ia_list, itx);
1261 	}
1262 
1263 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1264 	mutex_exit(&itxg->itxg_lock);
1265 
1266 	/* Release the old itxs now we've dropped the lock */
1267 	if (clean != NULL)
1268 		zil_itxg_clean(clean);
1269 }
1270 
1271 /*
1272  * If there are any in-memory intent log transactions which have now been
1273  * synced then start up a taskq to free them.
1274  */
1275 void
1276 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1277 {
1278 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1279 	itxs_t *clean_me;
1280 
1281 	mutex_enter(&itxg->itxg_lock);
1282 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1283 		mutex_exit(&itxg->itxg_lock);
1284 		return;
1285 	}
1286 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1287 	ASSERT(itxg->itxg_txg != 0);
1288 	ASSERT(zilog->zl_clean_taskq != NULL);
1289 	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1290 	itxg->itxg_sod = 0;
1291 	clean_me = itxg->itxg_itxs;
1292 	itxg->itxg_itxs = NULL;
1293 	itxg->itxg_txg = 0;
1294 	mutex_exit(&itxg->itxg_lock);
1295 	/*
1296 	 * Preferably start a task queue to free up the old itxs but
1297 	 * if taskq_dispatch can't allocate resources to do that then
1298 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1299 	 * created a bad performance problem.
1300 	 */
1301 	if (taskq_dispatch(zilog->zl_clean_taskq,
1302 	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1303 		zil_itxg_clean(clean_me);
1304 }
1305 
1306 /*
1307  * Get the list of itxs to commit into zl_itx_commit_list.
1308  */
1309 static void
1310 zil_get_commit_list(zilog_t *zilog)
1311 {
1312 	uint64_t otxg, txg;
1313 	list_t *commit_list = &zilog->zl_itx_commit_list;
1314 	uint64_t push_sod = 0;
1315 
1316 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1317 		otxg = ZILTEST_TXG;
1318 	else
1319 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1320 
1321 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1322 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1323 
1324 		mutex_enter(&itxg->itxg_lock);
1325 		if (itxg->itxg_txg != txg) {
1326 			mutex_exit(&itxg->itxg_lock);
1327 			continue;
1328 		}
1329 
1330 		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1331 		push_sod += itxg->itxg_sod;
1332 		itxg->itxg_sod = 0;
1333 
1334 		mutex_exit(&itxg->itxg_lock);
1335 	}
1336 	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1337 }
1338 
1339 /*
1340  * Move the async itxs for a specified object to commit into sync lists.
1341  */
1342 static void
1343 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1344 {
1345 	uint64_t otxg, txg;
1346 	itx_async_node_t *ian;
1347 	avl_tree_t *t;
1348 	avl_index_t where;
1349 
1350 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1351 		otxg = ZILTEST_TXG;
1352 	else
1353 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1354 
1355 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1356 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1357 
1358 		mutex_enter(&itxg->itxg_lock);
1359 		if (itxg->itxg_txg != txg) {
1360 			mutex_exit(&itxg->itxg_lock);
1361 			continue;
1362 		}
1363 
1364 		/*
1365 		 * If a foid is specified then find that node and append its
1366 		 * list. Otherwise walk the tree appending all the lists
1367 		 * to the sync list. We add to the end rather than the
1368 		 * beginning to ensure the create has happened.
1369 		 */
1370 		t = &itxg->itxg_itxs->i_async_tree;
1371 		if (foid != 0) {
1372 			ian = avl_find(t, &foid, &where);
1373 			if (ian != NULL) {
1374 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1375 				    &ian->ia_list);
1376 			}
1377 		} else {
1378 			void *cookie = NULL;
1379 
1380 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1381 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1382 				    &ian->ia_list);
1383 				list_destroy(&ian->ia_list);
1384 				kmem_free(ian, sizeof (itx_async_node_t));
1385 			}
1386 		}
1387 		mutex_exit(&itxg->itxg_lock);
1388 	}
1389 }
1390 
1391 static void
1392 zil_commit_writer(zilog_t *zilog)
1393 {
1394 	uint64_t txg;
1395 	itx_t *itx;
1396 	lwb_t *lwb;
1397 	spa_t *spa = zilog->zl_spa;
1398 	int error = 0;
1399 
1400 	ASSERT(zilog->zl_root_zio == NULL);
1401 
1402 	mutex_exit(&zilog->zl_lock);
1403 
1404 	zil_get_commit_list(zilog);
1405 
1406 	/*
1407 	 * Return if there's nothing to commit before we dirty the fs by
1408 	 * calling zil_create().
1409 	 */
1410 	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1411 		mutex_enter(&zilog->zl_lock);
1412 		return;
1413 	}
1414 
1415 	if (zilog->zl_suspend) {
1416 		lwb = NULL;
1417 	} else {
1418 		lwb = list_tail(&zilog->zl_lwb_list);
1419 		if (lwb == NULL)
1420 			lwb = zil_create(zilog);
1421 	}
1422 
1423 	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1424 	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1425 		txg = itx->itx_lr.lrc_txg;
1426 		ASSERT(txg);
1427 
1428 		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1429 			lwb = zil_lwb_commit(zilog, itx, lwb);
1430 		list_remove(&zilog->zl_itx_commit_list, itx);
1431 		kmem_free(itx, offsetof(itx_t, itx_lr)
1432 		    + itx->itx_lr.lrc_reclen);
1433 	}
1434 	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1435 
1436 	/* write the last block out */
1437 	if (lwb != NULL && lwb->lwb_zio != NULL)
1438 		lwb = zil_lwb_write_start(zilog, lwb);
1439 
1440 	zilog->zl_cur_used = 0;
1441 
1442 	/*
1443 	 * Wait if necessary for the log blocks to be on stable storage.
1444 	 */
1445 	if (zilog->zl_root_zio) {
1446 		error = zio_wait(zilog->zl_root_zio);
1447 		zilog->zl_root_zio = NULL;
1448 		zil_flush_vdevs(zilog);
1449 	}
1450 
1451 	if (error || lwb == NULL)
1452 		txg_wait_synced(zilog->zl_dmu_pool, 0);
1453 
1454 	mutex_enter(&zilog->zl_lock);
1455 
1456 	/*
1457 	 * Remember the highest committed log sequence number for ztest.
1458 	 * We only update this value when all the log writes succeeded,
1459 	 * because ztest wants to ASSERT that it got the whole log chain.
1460 	 */
1461 	if (error == 0 && lwb != NULL)
1462 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1463 }
1464 
1465 /*
1466  * Commit zfs transactions to stable storage.
1467  * If foid is 0 push out all transactions, otherwise push only those
1468  * for that object or might reference that object.
1469  *
1470  * itxs are committed in batches. In a heavily stressed zil there will be
1471  * a commit writer thread who is writing out a bunch of itxs to the log
1472  * for a set of committing threads (cthreads) in the same batch as the writer.
1473  * Those cthreads are all waiting on the same cv for that batch.
1474  *
1475  * There will also be a different and growing batch of threads that are
1476  * waiting to commit (qthreads). When the committing batch completes
1477  * a transition occurs such that the cthreads exit and the qthreads become
1478  * cthreads. One of the new cthreads becomes the writer thread for the
1479  * batch. Any new threads arriving become new qthreads.
1480  *
1481  * Only 2 condition variables are needed and there's no transition
1482  * between the two cvs needed. They just flip-flop between qthreads
1483  * and cthreads.
1484  *
1485  * Using this scheme we can efficiently wakeup up only those threads
1486  * that have been committed.
1487  */
1488 void
1489 zil_commit(zilog_t *zilog, uint64_t foid)
1490 {
1491 	uint64_t mybatch;
1492 
1493 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1494 		return;
1495 
1496 	/* move the async itxs for the foid to the sync queues */
1497 	zil_async_to_sync(zilog, foid);
1498 
1499 	mutex_enter(&zilog->zl_lock);
1500 	mybatch = zilog->zl_next_batch;
1501 	while (zilog->zl_writer) {
1502 		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1503 		if (mybatch <= zilog->zl_com_batch) {
1504 			mutex_exit(&zilog->zl_lock);
1505 			return;
1506 		}
1507 	}
1508 
1509 	zilog->zl_next_batch++;
1510 	zilog->zl_writer = B_TRUE;
1511 	zil_commit_writer(zilog);
1512 	zilog->zl_com_batch = mybatch;
1513 	zilog->zl_writer = B_FALSE;
1514 	mutex_exit(&zilog->zl_lock);
1515 
1516 	/* wake up one thread to become the next writer */
1517 	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1518 
1519 	/* wake up all threads waiting for this batch to be committed */
1520 	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1521 }
1522 
1523 /*
1524  * Called in syncing context to free committed log blocks and update log header.
1525  */
1526 void
1527 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1528 {
1529 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1530 	uint64_t txg = dmu_tx_get_txg(tx);
1531 	spa_t *spa = zilog->zl_spa;
1532 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1533 	lwb_t *lwb;
1534 
1535 	/*
1536 	 * We don't zero out zl_destroy_txg, so make sure we don't try
1537 	 * to destroy it twice.
1538 	 */
1539 	if (spa_sync_pass(spa) != 1)
1540 		return;
1541 
1542 	mutex_enter(&zilog->zl_lock);
1543 
1544 	ASSERT(zilog->zl_stop_sync == 0);
1545 
1546 	if (*replayed_seq != 0) {
1547 		ASSERT(zh->zh_replay_seq < *replayed_seq);
1548 		zh->zh_replay_seq = *replayed_seq;
1549 		*replayed_seq = 0;
1550 	}
1551 
1552 	if (zilog->zl_destroy_txg == txg) {
1553 		blkptr_t blk = zh->zh_log;
1554 
1555 		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1556 
1557 		bzero(zh, sizeof (zil_header_t));
1558 		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1559 
1560 		if (zilog->zl_keep_first) {
1561 			/*
1562 			 * If this block was part of log chain that couldn't
1563 			 * be claimed because a device was missing during
1564 			 * zil_claim(), but that device later returns,
1565 			 * then this block could erroneously appear valid.
1566 			 * To guard against this, assign a new GUID to the new
1567 			 * log chain so it doesn't matter what blk points to.
1568 			 */
1569 			zil_init_log_chain(zilog, &blk);
1570 			zh->zh_log = blk;
1571 		}
1572 	}
1573 
1574 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1575 		zh->zh_log = lwb->lwb_blk;
1576 		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1577 			break;
1578 		list_remove(&zilog->zl_lwb_list, lwb);
1579 		zio_free_zil(spa, txg, &lwb->lwb_blk);
1580 		kmem_cache_free(zil_lwb_cache, lwb);
1581 
1582 		/*
1583 		 * If we don't have anything left in the lwb list then
1584 		 * we've had an allocation failure and we need to zero
1585 		 * out the zil_header blkptr so that we don't end
1586 		 * up freeing the same block twice.
1587 		 */
1588 		if (list_head(&zilog->zl_lwb_list) == NULL)
1589 			BP_ZERO(&zh->zh_log);
1590 	}
1591 	mutex_exit(&zilog->zl_lock);
1592 }
1593 
1594 void
1595 zil_init(void)
1596 {
1597 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1598 	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1599 }
1600 
1601 void
1602 zil_fini(void)
1603 {
1604 	kmem_cache_destroy(zil_lwb_cache);
1605 }
1606 
1607 void
1608 zil_set_sync(zilog_t *zilog, uint64_t sync)
1609 {
1610 	zilog->zl_sync = sync;
1611 }
1612 
1613 void
1614 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1615 {
1616 	zilog->zl_logbias = logbias;
1617 }
1618 
1619 zilog_t *
1620 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1621 {
1622 	zilog_t *zilog;
1623 
1624 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1625 
1626 	zilog->zl_header = zh_phys;
1627 	zilog->zl_os = os;
1628 	zilog->zl_spa = dmu_objset_spa(os);
1629 	zilog->zl_dmu_pool = dmu_objset_pool(os);
1630 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1631 	zilog->zl_logbias = dmu_objset_logbias(os);
1632 	zilog->zl_sync = dmu_objset_syncprop(os);
1633 	zilog->zl_next_batch = 1;
1634 
1635 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1636 
1637 	for (int i = 0; i < TXG_SIZE; i++) {
1638 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1639 		    MUTEX_DEFAULT, NULL);
1640 	}
1641 
1642 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1643 	    offsetof(lwb_t, lwb_node));
1644 
1645 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1646 	    offsetof(itx_t, itx_node));
1647 
1648 	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1649 
1650 	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1651 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1652 
1653 	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1654 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1655 	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1656 	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1657 
1658 	return (zilog);
1659 }
1660 
1661 void
1662 zil_free(zilog_t *zilog)
1663 {
1664 	lwb_t *head_lwb;
1665 
1666 	zilog->zl_stop_sync = 1;
1667 
1668 	/*
1669 	 * After zil_close() there should only be one lwb with a buffer.
1670 	 */
1671 	head_lwb = list_head(&zilog->zl_lwb_list);
1672 	if (head_lwb) {
1673 		ASSERT(head_lwb == list_tail(&zilog->zl_lwb_list));
1674 		list_remove(&zilog->zl_lwb_list, head_lwb);
1675 		zio_buf_free(head_lwb->lwb_buf, head_lwb->lwb_sz);
1676 		kmem_cache_free(zil_lwb_cache, head_lwb);
1677 	}
1678 	list_destroy(&zilog->zl_lwb_list);
1679 
1680 	avl_destroy(&zilog->zl_vdev_tree);
1681 	mutex_destroy(&zilog->zl_vdev_lock);
1682 
1683 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1684 	list_destroy(&zilog->zl_itx_commit_list);
1685 
1686 	for (int i = 0; i < TXG_SIZE; i++) {
1687 		/*
1688 		 * It's possible for an itx to be generated that doesn't dirty
1689 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1690 		 * callback to remove the entry. We remove those here.
1691 		 *
1692 		 * Also free up the ziltest itxs.
1693 		 */
1694 		if (zilog->zl_itxg[i].itxg_itxs)
1695 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1696 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1697 	}
1698 
1699 	mutex_destroy(&zilog->zl_lock);
1700 
1701 	cv_destroy(&zilog->zl_cv_writer);
1702 	cv_destroy(&zilog->zl_cv_suspend);
1703 	cv_destroy(&zilog->zl_cv_batch[0]);
1704 	cv_destroy(&zilog->zl_cv_batch[1]);
1705 
1706 	kmem_free(zilog, sizeof (zilog_t));
1707 }
1708 
1709 /*
1710  * Open an intent log.
1711  */
1712 zilog_t *
1713 zil_open(objset_t *os, zil_get_data_t *get_data)
1714 {
1715 	zilog_t *zilog = dmu_objset_zil(os);
1716 
1717 	zilog->zl_get_data = get_data;
1718 	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1719 	    2, 2, TASKQ_PREPOPULATE);
1720 
1721 	return (zilog);
1722 }
1723 
1724 /*
1725  * Close an intent log.
1726  */
1727 void
1728 zil_close(zilog_t *zilog)
1729 {
1730 	lwb_t *tail_lwb;
1731 	uint64_t txg = 0;
1732 
1733 	zil_commit(zilog, 0); /* commit all itx */
1734 
1735 	/*
1736 	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1737 	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1738 	 * callbacks have occurred that may clean the zil.  Only then can we
1739 	 * destroy the zl_clean_taskq.
1740 	 */
1741 	mutex_enter(&zilog->zl_lock);
1742 	tail_lwb = list_tail(&zilog->zl_lwb_list);
1743 	if (tail_lwb != NULL)
1744 		txg = tail_lwb->lwb_max_txg;
1745 	mutex_exit(&zilog->zl_lock);
1746 	if (txg)
1747 		txg_wait_synced(zilog->zl_dmu_pool, txg);
1748 
1749 	taskq_destroy(zilog->zl_clean_taskq);
1750 	zilog->zl_clean_taskq = NULL;
1751 	zilog->zl_get_data = NULL;
1752 }
1753 
1754 /*
1755  * Suspend an intent log.  While in suspended mode, we still honor
1756  * synchronous semantics, but we rely on txg_wait_synced() to do it.
1757  * We suspend the log briefly when taking a snapshot so that the snapshot
1758  * contains all the data it's supposed to, and has an empty intent log.
1759  */
1760 int
1761 zil_suspend(zilog_t *zilog)
1762 {
1763 	const zil_header_t *zh = zilog->zl_header;
1764 
1765 	mutex_enter(&zilog->zl_lock);
1766 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1767 		mutex_exit(&zilog->zl_lock);
1768 		return (EBUSY);
1769 	}
1770 	if (zilog->zl_suspend++ != 0) {
1771 		/*
1772 		 * Someone else already began a suspend.
1773 		 * Just wait for them to finish.
1774 		 */
1775 		while (zilog->zl_suspending)
1776 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1777 		mutex_exit(&zilog->zl_lock);
1778 		return (0);
1779 	}
1780 	zilog->zl_suspending = B_TRUE;
1781 	mutex_exit(&zilog->zl_lock);
1782 
1783 	zil_commit(zilog, 0);
1784 
1785 	zil_destroy(zilog, B_FALSE);
1786 
1787 	mutex_enter(&zilog->zl_lock);
1788 	zilog->zl_suspending = B_FALSE;
1789 	cv_broadcast(&zilog->zl_cv_suspend);
1790 	mutex_exit(&zilog->zl_lock);
1791 
1792 	return (0);
1793 }
1794 
1795 void
1796 zil_resume(zilog_t *zilog)
1797 {
1798 	mutex_enter(&zilog->zl_lock);
1799 	ASSERT(zilog->zl_suspend != 0);
1800 	zilog->zl_suspend--;
1801 	mutex_exit(&zilog->zl_lock);
1802 }
1803 
1804 typedef struct zil_replay_arg {
1805 	zil_replay_func_t **zr_replay;
1806 	void		*zr_arg;
1807 	boolean_t	zr_byteswap;
1808 	char		*zr_lr;
1809 } zil_replay_arg_t;
1810 
1811 static int
1812 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1813 {
1814 	char name[MAXNAMELEN];
1815 
1816 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
1817 
1818 	dmu_objset_name(zilog->zl_os, name);
1819 
1820 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1821 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1822 	    (u_longlong_t)lr->lrc_seq,
1823 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1824 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1825 
1826 	return (error);
1827 }
1828 
1829 static int
1830 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1831 {
1832 	zil_replay_arg_t *zr = zra;
1833 	const zil_header_t *zh = zilog->zl_header;
1834 	uint64_t reclen = lr->lrc_reclen;
1835 	uint64_t txtype = lr->lrc_txtype;
1836 	int error = 0;
1837 
1838 	zilog->zl_replaying_seq = lr->lrc_seq;
1839 
1840 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1841 		return (0);
1842 
1843 	if (lr->lrc_txg < claim_txg)		/* already committed */
1844 		return (0);
1845 
1846 	/* Strip case-insensitive bit, still present in log record */
1847 	txtype &= ~TX_CI;
1848 
1849 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
1850 		return (zil_replay_error(zilog, lr, EINVAL));
1851 
1852 	/*
1853 	 * If this record type can be logged out of order, the object
1854 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
1855 	 */
1856 	if (TX_OOO(txtype)) {
1857 		error = dmu_object_info(zilog->zl_os,
1858 		    ((lr_ooo_t *)lr)->lr_foid, NULL);
1859 		if (error == ENOENT || error == EEXIST)
1860 			return (0);
1861 	}
1862 
1863 	/*
1864 	 * Make a copy of the data so we can revise and extend it.
1865 	 */
1866 	bcopy(lr, zr->zr_lr, reclen);
1867 
1868 	/*
1869 	 * If this is a TX_WRITE with a blkptr, suck in the data.
1870 	 */
1871 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1872 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
1873 		    zr->zr_lr + reclen);
1874 		if (error)
1875 			return (zil_replay_error(zilog, lr, error));
1876 	}
1877 
1878 	/*
1879 	 * The log block containing this lr may have been byteswapped
1880 	 * so that we can easily examine common fields like lrc_txtype.
1881 	 * However, the log is a mix of different record types, and only the
1882 	 * replay vectors know how to byteswap their records.  Therefore, if
1883 	 * the lr was byteswapped, undo it before invoking the replay vector.
1884 	 */
1885 	if (zr->zr_byteswap)
1886 		byteswap_uint64_array(zr->zr_lr, reclen);
1887 
1888 	/*
1889 	 * We must now do two things atomically: replay this log record,
1890 	 * and update the log header sequence number to reflect the fact that
1891 	 * we did so. At the end of each replay function the sequence number
1892 	 * is updated if we are in replay mode.
1893 	 */
1894 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
1895 	if (error) {
1896 		/*
1897 		 * The DMU's dnode layer doesn't see removes until the txg
1898 		 * commits, so a subsequent claim can spuriously fail with
1899 		 * EEXIST. So if we receive any error we try syncing out
1900 		 * any removes then retry the transaction.  Note that we
1901 		 * specify B_FALSE for byteswap now, so we don't do it twice.
1902 		 */
1903 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1904 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
1905 		if (error)
1906 			return (zil_replay_error(zilog, lr, error));
1907 	}
1908 	return (0);
1909 }
1910 
1911 /* ARGSUSED */
1912 static int
1913 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1914 {
1915 	zilog->zl_replay_blks++;
1916 
1917 	return (0);
1918 }
1919 
1920 /*
1921  * If this dataset has a non-empty intent log, replay it and destroy it.
1922  */
1923 void
1924 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1925 {
1926 	zilog_t *zilog = dmu_objset_zil(os);
1927 	const zil_header_t *zh = zilog->zl_header;
1928 	zil_replay_arg_t zr;
1929 
1930 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1931 		zil_destroy(zilog, B_TRUE);
1932 		return;
1933 	}
1934 
1935 	zr.zr_replay = replay_func;
1936 	zr.zr_arg = arg;
1937 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1938 	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1939 
1940 	/*
1941 	 * Wait for in-progress removes to sync before starting replay.
1942 	 */
1943 	txg_wait_synced(zilog->zl_dmu_pool, 0);
1944 
1945 	zilog->zl_replay = B_TRUE;
1946 	zilog->zl_replay_time = ddi_get_lbolt();
1947 	ASSERT(zilog->zl_replay_blks == 0);
1948 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1949 	    zh->zh_claim_txg);
1950 	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
1951 
1952 	zil_destroy(zilog, B_FALSE);
1953 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1954 	zilog->zl_replay = B_FALSE;
1955 }
1956 
1957 boolean_t
1958 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
1959 {
1960 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1961 		return (B_TRUE);
1962 
1963 	if (zilog->zl_replay) {
1964 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1965 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
1966 		    zilog->zl_replaying_seq;
1967 		return (B_TRUE);
1968 	}
1969 
1970 	return (B_FALSE);
1971 }
1972 
1973 /* ARGSUSED */
1974 int
1975 zil_vdev_offline(const char *osname, void *arg)
1976 {
1977 	objset_t *os;
1978 	zilog_t *zilog;
1979 	int error;
1980 
1981 	error = dmu_objset_hold(osname, FTAG, &os);
1982 	if (error)
1983 		return (error);
1984 
1985 	zilog = dmu_objset_zil(os);
1986 	if (zil_suspend(zilog) != 0)
1987 		error = EEXIST;
1988 	else
1989 		zil_resume(zilog);
1990 	dmu_objset_rele(os, FTAG);
1991 	return (error);
1992 }
1993