xref: /illumos-gate/usr/src/uts/common/fs/zfs/zil.c (revision 69962b5647e4a8b9b14998733b765925381b727e)
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  * Copyright (c) 2013 by Delphix. All rights reserved.
24  */
25 
26 /* Portions Copyright 2010 Robert Milkowski */
27 
28 #include <sys/zfs_context.h>
29 #include <sys/spa.h>
30 #include <sys/dmu.h>
31 #include <sys/zap.h>
32 #include <sys/arc.h>
33 #include <sys/stat.h>
34 #include <sys/resource.h>
35 #include <sys/zil.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 
42 /*
43  * The zfs intent log (ZIL) saves transaction records of system calls
44  * that change the file system in memory with enough information
45  * to be able to replay them. These are stored in memory until
46  * either the DMU transaction group (txg) commits them to the stable pool
47  * and they can be discarded, or they are flushed to the stable log
48  * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49  * requirement. In the event of a panic or power fail then those log
50  * records (transactions) are replayed.
51  *
52  * There is one ZIL per file system. Its on-disk (pool) format consists
53  * of 3 parts:
54  *
55  * 	- ZIL header
56  * 	- ZIL blocks
57  * 	- ZIL records
58  *
59  * A log record holds a system call transaction. Log blocks can
60  * hold many log records and the blocks are chained together.
61  * Each ZIL block contains a block pointer (blkptr_t) to the next
62  * ZIL block in the chain. The ZIL header points to the first
63  * block in the chain. Note there is not a fixed place in the pool
64  * to hold blocks. They are dynamically allocated and freed as
65  * needed from the blocks available. Figure X shows the ZIL structure:
66  */
67 
68 /*
69  * Disable intent logging replay.  This global ZIL switch affects all pools.
70  */
71 int zil_replay_disable = 0;
72 
73 /*
74  * Tunable parameter for debugging or performance analysis.  Setting
75  * zfs_nocacheflush will cause corruption on power loss if a volatile
76  * out-of-order write cache is enabled.
77  */
78 boolean_t zfs_nocacheflush = B_FALSE;
79 
80 static kmem_cache_t *zil_lwb_cache;
81 
82 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
83 
84 #define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
85     sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
86 
87 
88 /*
89  * ziltest is by and large an ugly hack, but very useful in
90  * checking replay without tedious work.
91  * When running ziltest we want to keep all itx's and so maintain
92  * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
93  * We subtract TXG_CONCURRENT_STATES to allow for common code.
94  */
95 #define	ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
96 
97 static int
98 zil_bp_compare(const void *x1, const void *x2)
99 {
100 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
101 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
102 
103 	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
104 		return (-1);
105 	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
106 		return (1);
107 
108 	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
109 		return (-1);
110 	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
111 		return (1);
112 
113 	return (0);
114 }
115 
116 static void
117 zil_bp_tree_init(zilog_t *zilog)
118 {
119 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
120 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
121 }
122 
123 static void
124 zil_bp_tree_fini(zilog_t *zilog)
125 {
126 	avl_tree_t *t = &zilog->zl_bp_tree;
127 	zil_bp_node_t *zn;
128 	void *cookie = NULL;
129 
130 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
131 		kmem_free(zn, sizeof (zil_bp_node_t));
132 
133 	avl_destroy(t);
134 }
135 
136 int
137 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
138 {
139 	avl_tree_t *t = &zilog->zl_bp_tree;
140 	const dva_t *dva = BP_IDENTITY(bp);
141 	zil_bp_node_t *zn;
142 	avl_index_t where;
143 
144 	if (avl_find(t, dva, &where) != NULL)
145 		return (SET_ERROR(EEXIST));
146 
147 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
148 	zn->zn_dva = *dva;
149 	avl_insert(t, zn, where);
150 
151 	return (0);
152 }
153 
154 static zil_header_t *
155 zil_header_in_syncing_context(zilog_t *zilog)
156 {
157 	return ((zil_header_t *)zilog->zl_header);
158 }
159 
160 static void
161 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
162 {
163 	zio_cksum_t *zc = &bp->blk_cksum;
164 
165 	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
166 	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
167 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
168 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
169 }
170 
171 /*
172  * Read a log block and make sure it's valid.
173  */
174 static int
175 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
176     char **end)
177 {
178 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
179 	uint32_t aflags = ARC_WAIT;
180 	arc_buf_t *abuf = NULL;
181 	zbookmark_t zb;
182 	int error;
183 
184 	if (zilog->zl_header->zh_claim_txg == 0)
185 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
186 
187 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
188 		zio_flags |= ZIO_FLAG_SPECULATIVE;
189 
190 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
191 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
192 
193 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
194 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
195 
196 	if (error == 0) {
197 		zio_cksum_t cksum = bp->blk_cksum;
198 
199 		/*
200 		 * Validate the checksummed log block.
201 		 *
202 		 * Sequence numbers should be... sequential.  The checksum
203 		 * verifier for the next block should be bp's checksum plus 1.
204 		 *
205 		 * Also check the log chain linkage and size used.
206 		 */
207 		cksum.zc_word[ZIL_ZC_SEQ]++;
208 
209 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
210 			zil_chain_t *zilc = abuf->b_data;
211 			char *lr = (char *)(zilc + 1);
212 			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
213 
214 			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
215 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
216 				error = SET_ERROR(ECKSUM);
217 			} else {
218 				bcopy(lr, dst, len);
219 				*end = (char *)dst + len;
220 				*nbp = zilc->zc_next_blk;
221 			}
222 		} else {
223 			char *lr = abuf->b_data;
224 			uint64_t size = BP_GET_LSIZE(bp);
225 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
226 
227 			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
228 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
229 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
230 				error = SET_ERROR(ECKSUM);
231 			} else {
232 				bcopy(lr, dst, zilc->zc_nused);
233 				*end = (char *)dst + zilc->zc_nused;
234 				*nbp = zilc->zc_next_blk;
235 			}
236 		}
237 
238 		VERIFY(arc_buf_remove_ref(abuf, &abuf));
239 	}
240 
241 	return (error);
242 }
243 
244 /*
245  * Read a TX_WRITE log data block.
246  */
247 static int
248 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
249 {
250 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
251 	const blkptr_t *bp = &lr->lr_blkptr;
252 	uint32_t aflags = ARC_WAIT;
253 	arc_buf_t *abuf = NULL;
254 	zbookmark_t zb;
255 	int error;
256 
257 	if (BP_IS_HOLE(bp)) {
258 		if (wbuf != NULL)
259 			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
260 		return (0);
261 	}
262 
263 	if (zilog->zl_header->zh_claim_txg == 0)
264 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
265 
266 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
267 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
268 
269 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
270 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
271 
272 	if (error == 0) {
273 		if (wbuf != NULL)
274 			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
275 		(void) arc_buf_remove_ref(abuf, &abuf);
276 	}
277 
278 	return (error);
279 }
280 
281 /*
282  * Parse the intent log, and call parse_func for each valid record within.
283  */
284 int
285 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
286     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
287 {
288 	const zil_header_t *zh = zilog->zl_header;
289 	boolean_t claimed = !!zh->zh_claim_txg;
290 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
291 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
292 	uint64_t max_blk_seq = 0;
293 	uint64_t max_lr_seq = 0;
294 	uint64_t blk_count = 0;
295 	uint64_t lr_count = 0;
296 	blkptr_t blk, next_blk;
297 	char *lrbuf, *lrp;
298 	int error = 0;
299 
300 	/*
301 	 * Old logs didn't record the maximum zh_claim_lr_seq.
302 	 */
303 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
304 		claim_lr_seq = UINT64_MAX;
305 
306 	/*
307 	 * Starting at the block pointed to by zh_log we read the log chain.
308 	 * For each block in the chain we strongly check that block to
309 	 * ensure its validity.  We stop when an invalid block is found.
310 	 * For each block pointer in the chain we call parse_blk_func().
311 	 * For each record in each valid block we call parse_lr_func().
312 	 * If the log has been claimed, stop if we encounter a sequence
313 	 * number greater than the highest claimed sequence number.
314 	 */
315 	lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
316 	zil_bp_tree_init(zilog);
317 
318 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
319 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
320 		int reclen;
321 		char *end;
322 
323 		if (blk_seq > claim_blk_seq)
324 			break;
325 		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
326 			break;
327 		ASSERT3U(max_blk_seq, <, blk_seq);
328 		max_blk_seq = blk_seq;
329 		blk_count++;
330 
331 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
332 			break;
333 
334 		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
335 		if (error != 0)
336 			break;
337 
338 		for (lrp = lrbuf; lrp < end; lrp += reclen) {
339 			lr_t *lr = (lr_t *)lrp;
340 			reclen = lr->lrc_reclen;
341 			ASSERT3U(reclen, >=, sizeof (lr_t));
342 			if (lr->lrc_seq > claim_lr_seq)
343 				goto done;
344 			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
345 				goto done;
346 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
347 			max_lr_seq = lr->lrc_seq;
348 			lr_count++;
349 		}
350 	}
351 done:
352 	zilog->zl_parse_error = error;
353 	zilog->zl_parse_blk_seq = max_blk_seq;
354 	zilog->zl_parse_lr_seq = max_lr_seq;
355 	zilog->zl_parse_blk_count = blk_count;
356 	zilog->zl_parse_lr_count = lr_count;
357 
358 	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
359 	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
360 
361 	zil_bp_tree_fini(zilog);
362 	zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
363 
364 	return (error);
365 }
366 
367 static int
368 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
369 {
370 	/*
371 	 * Claim log block if not already committed and not already claimed.
372 	 * If tx == NULL, just verify that the block is claimable.
373 	 */
374 	if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
375 		return (0);
376 
377 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
378 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
379 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
380 }
381 
382 static int
383 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
384 {
385 	lr_write_t *lr = (lr_write_t *)lrc;
386 	int error;
387 
388 	if (lrc->lrc_txtype != TX_WRITE)
389 		return (0);
390 
391 	/*
392 	 * If the block is not readable, don't claim it.  This can happen
393 	 * in normal operation when a log block is written to disk before
394 	 * some of the dmu_sync() blocks it points to.  In this case, the
395 	 * transaction cannot have been committed to anyone (we would have
396 	 * waited for all writes to be stable first), so it is semantically
397 	 * correct to declare this the end of the log.
398 	 */
399 	if (lr->lr_blkptr.blk_birth >= first_txg &&
400 	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
401 		return (error);
402 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
403 }
404 
405 /* ARGSUSED */
406 static int
407 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
408 {
409 	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
410 
411 	return (0);
412 }
413 
414 static int
415 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
416 {
417 	lr_write_t *lr = (lr_write_t *)lrc;
418 	blkptr_t *bp = &lr->lr_blkptr;
419 
420 	/*
421 	 * If we previously claimed it, we need to free it.
422 	 */
423 	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
424 	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
425 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
426 
427 	return (0);
428 }
429 
430 static lwb_t *
431 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
432 {
433 	lwb_t *lwb;
434 
435 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
436 	lwb->lwb_zilog = zilog;
437 	lwb->lwb_blk = *bp;
438 	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
439 	lwb->lwb_max_txg = txg;
440 	lwb->lwb_zio = NULL;
441 	lwb->lwb_tx = NULL;
442 	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
443 		lwb->lwb_nused = sizeof (zil_chain_t);
444 		lwb->lwb_sz = BP_GET_LSIZE(bp);
445 	} else {
446 		lwb->lwb_nused = 0;
447 		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
448 	}
449 
450 	mutex_enter(&zilog->zl_lock);
451 	list_insert_tail(&zilog->zl_lwb_list, lwb);
452 	mutex_exit(&zilog->zl_lock);
453 
454 	return (lwb);
455 }
456 
457 /*
458  * Called when we create in-memory log transactions so that we know
459  * to cleanup the itxs at the end of spa_sync().
460  */
461 void
462 zilog_dirty(zilog_t *zilog, uint64_t txg)
463 {
464 	dsl_pool_t *dp = zilog->zl_dmu_pool;
465 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
466 
467 	if (dsl_dataset_is_snapshot(ds))
468 		panic("dirtying snapshot!");
469 
470 	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
471 		/* up the hold count until we can be written out */
472 		dmu_buf_add_ref(ds->ds_dbuf, zilog);
473 	}
474 }
475 
476 boolean_t
477 zilog_is_dirty(zilog_t *zilog)
478 {
479 	dsl_pool_t *dp = zilog->zl_dmu_pool;
480 
481 	for (int t = 0; t < TXG_SIZE; t++) {
482 		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
483 			return (B_TRUE);
484 	}
485 	return (B_FALSE);
486 }
487 
488 /*
489  * Create an on-disk intent log.
490  */
491 static lwb_t *
492 zil_create(zilog_t *zilog)
493 {
494 	const zil_header_t *zh = zilog->zl_header;
495 	lwb_t *lwb = NULL;
496 	uint64_t txg = 0;
497 	dmu_tx_t *tx = NULL;
498 	blkptr_t blk;
499 	int error = 0;
500 
501 	/*
502 	 * Wait for any previous destroy to complete.
503 	 */
504 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
505 
506 	ASSERT(zh->zh_claim_txg == 0);
507 	ASSERT(zh->zh_replay_seq == 0);
508 
509 	blk = zh->zh_log;
510 
511 	/*
512 	 * Allocate an initial log block if:
513 	 *    - there isn't one already
514 	 *    - the existing block is the wrong endianess
515 	 */
516 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
517 		tx = dmu_tx_create(zilog->zl_os);
518 		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
519 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
520 		txg = dmu_tx_get_txg(tx);
521 
522 		if (!BP_IS_HOLE(&blk)) {
523 			zio_free_zil(zilog->zl_spa, txg, &blk);
524 			BP_ZERO(&blk);
525 		}
526 
527 		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
528 		    ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
529 
530 		if (error == 0)
531 			zil_init_log_chain(zilog, &blk);
532 	}
533 
534 	/*
535 	 * Allocate a log write buffer (lwb) for the first log block.
536 	 */
537 	if (error == 0)
538 		lwb = zil_alloc_lwb(zilog, &blk, txg);
539 
540 	/*
541 	 * If we just allocated the first log block, commit our transaction
542 	 * and wait for zil_sync() to stuff the block poiner into zh_log.
543 	 * (zh is part of the MOS, so we cannot modify it in open context.)
544 	 */
545 	if (tx != NULL) {
546 		dmu_tx_commit(tx);
547 		txg_wait_synced(zilog->zl_dmu_pool, txg);
548 	}
549 
550 	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
551 
552 	return (lwb);
553 }
554 
555 /*
556  * In one tx, free all log blocks and clear the log header.
557  * If keep_first is set, then we're replaying a log with no content.
558  * We want to keep the first block, however, so that the first
559  * synchronous transaction doesn't require a txg_wait_synced()
560  * in zil_create().  We don't need to txg_wait_synced() here either
561  * when keep_first is set, because both zil_create() and zil_destroy()
562  * will wait for any in-progress destroys to complete.
563  */
564 void
565 zil_destroy(zilog_t *zilog, boolean_t keep_first)
566 {
567 	const zil_header_t *zh = zilog->zl_header;
568 	lwb_t *lwb;
569 	dmu_tx_t *tx;
570 	uint64_t txg;
571 
572 	/*
573 	 * Wait for any previous destroy to complete.
574 	 */
575 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
576 
577 	zilog->zl_old_header = *zh;		/* debugging aid */
578 
579 	if (BP_IS_HOLE(&zh->zh_log))
580 		return;
581 
582 	tx = dmu_tx_create(zilog->zl_os);
583 	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
584 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
585 	txg = dmu_tx_get_txg(tx);
586 
587 	mutex_enter(&zilog->zl_lock);
588 
589 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
590 	zilog->zl_destroy_txg = txg;
591 	zilog->zl_keep_first = keep_first;
592 
593 	if (!list_is_empty(&zilog->zl_lwb_list)) {
594 		ASSERT(zh->zh_claim_txg == 0);
595 		VERIFY(!keep_first);
596 		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
597 			list_remove(&zilog->zl_lwb_list, lwb);
598 			if (lwb->lwb_buf != NULL)
599 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
600 			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
601 			kmem_cache_free(zil_lwb_cache, lwb);
602 		}
603 	} else if (!keep_first) {
604 		zil_destroy_sync(zilog, tx);
605 	}
606 	mutex_exit(&zilog->zl_lock);
607 
608 	dmu_tx_commit(tx);
609 }
610 
611 void
612 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
613 {
614 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
615 	(void) zil_parse(zilog, zil_free_log_block,
616 	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
617 }
618 
619 int
620 zil_claim(const char *osname, void *txarg)
621 {
622 	dmu_tx_t *tx = txarg;
623 	uint64_t first_txg = dmu_tx_get_txg(tx);
624 	zilog_t *zilog;
625 	zil_header_t *zh;
626 	objset_t *os;
627 	int error;
628 
629 	error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
630 	if (error != 0) {
631 		cmn_err(CE_WARN, "can't open objset for %s", osname);
632 		return (0);
633 	}
634 
635 	zilog = dmu_objset_zil(os);
636 	zh = zil_header_in_syncing_context(zilog);
637 
638 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
639 		if (!BP_IS_HOLE(&zh->zh_log))
640 			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
641 		BP_ZERO(&zh->zh_log);
642 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
643 		dmu_objset_disown(os, FTAG);
644 		return (0);
645 	}
646 
647 	/*
648 	 * Claim all log blocks if we haven't already done so, and remember
649 	 * the highest claimed sequence number.  This ensures that if we can
650 	 * read only part of the log now (e.g. due to a missing device),
651 	 * but we can read the entire log later, we will not try to replay
652 	 * or destroy beyond the last block we successfully claimed.
653 	 */
654 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
655 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
656 		(void) zil_parse(zilog, zil_claim_log_block,
657 		    zil_claim_log_record, tx, first_txg);
658 		zh->zh_claim_txg = first_txg;
659 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
660 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
661 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
662 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
663 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
664 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
665 	}
666 
667 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
668 	dmu_objset_disown(os, FTAG);
669 	return (0);
670 }
671 
672 /*
673  * Check the log by walking the log chain.
674  * Checksum errors are ok as they indicate the end of the chain.
675  * Any other error (no device or read failure) returns an error.
676  */
677 int
678 zil_check_log_chain(const char *osname, void *tx)
679 {
680 	zilog_t *zilog;
681 	objset_t *os;
682 	blkptr_t *bp;
683 	int error;
684 
685 	ASSERT(tx == NULL);
686 
687 	error = dmu_objset_hold(osname, FTAG, &os);
688 	if (error != 0) {
689 		cmn_err(CE_WARN, "can't open objset for %s", osname);
690 		return (0);
691 	}
692 
693 	zilog = dmu_objset_zil(os);
694 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
695 
696 	/*
697 	 * Check the first block and determine if it's on a log device
698 	 * which may have been removed or faulted prior to loading this
699 	 * pool.  If so, there's no point in checking the rest of the log
700 	 * as its content should have already been synced to the pool.
701 	 */
702 	if (!BP_IS_HOLE(bp)) {
703 		vdev_t *vd;
704 		boolean_t valid = B_TRUE;
705 
706 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
707 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
708 		if (vd->vdev_islog && vdev_is_dead(vd))
709 			valid = vdev_log_state_valid(vd);
710 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
711 
712 		if (!valid) {
713 			dmu_objset_rele(os, FTAG);
714 			return (0);
715 		}
716 	}
717 
718 	/*
719 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
720 	 * any blocks, but just determine whether it is possible to do so.
721 	 * In addition to checking the log chain, zil_claim_log_block()
722 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
723 	 * which will update spa_max_claim_txg.  See spa_load() for details.
724 	 */
725 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
726 	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
727 
728 	dmu_objset_rele(os, FTAG);
729 
730 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
731 }
732 
733 static int
734 zil_vdev_compare(const void *x1, const void *x2)
735 {
736 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
737 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
738 
739 	if (v1 < v2)
740 		return (-1);
741 	if (v1 > v2)
742 		return (1);
743 
744 	return (0);
745 }
746 
747 void
748 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
749 {
750 	avl_tree_t *t = &zilog->zl_vdev_tree;
751 	avl_index_t where;
752 	zil_vdev_node_t *zv, zvsearch;
753 	int ndvas = BP_GET_NDVAS(bp);
754 	int i;
755 
756 	if (zfs_nocacheflush)
757 		return;
758 
759 	ASSERT(zilog->zl_writer);
760 
761 	/*
762 	 * Even though we're zl_writer, we still need a lock because the
763 	 * zl_get_data() callbacks may have dmu_sync() done callbacks
764 	 * that will run concurrently.
765 	 */
766 	mutex_enter(&zilog->zl_vdev_lock);
767 	for (i = 0; i < ndvas; i++) {
768 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
769 		if (avl_find(t, &zvsearch, &where) == NULL) {
770 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
771 			zv->zv_vdev = zvsearch.zv_vdev;
772 			avl_insert(t, zv, where);
773 		}
774 	}
775 	mutex_exit(&zilog->zl_vdev_lock);
776 }
777 
778 static void
779 zil_flush_vdevs(zilog_t *zilog)
780 {
781 	spa_t *spa = zilog->zl_spa;
782 	avl_tree_t *t = &zilog->zl_vdev_tree;
783 	void *cookie = NULL;
784 	zil_vdev_node_t *zv;
785 	zio_t *zio;
786 
787 	ASSERT(zilog->zl_writer);
788 
789 	/*
790 	 * We don't need zl_vdev_lock here because we're the zl_writer,
791 	 * and all zl_get_data() callbacks are done.
792 	 */
793 	if (avl_numnodes(t) == 0)
794 		return;
795 
796 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
797 
798 	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
799 
800 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
801 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
802 		if (vd != NULL)
803 			zio_flush(zio, vd);
804 		kmem_free(zv, sizeof (*zv));
805 	}
806 
807 	/*
808 	 * Wait for all the flushes to complete.  Not all devices actually
809 	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
810 	 */
811 	(void) zio_wait(zio);
812 
813 	spa_config_exit(spa, SCL_STATE, FTAG);
814 }
815 
816 /*
817  * Function called when a log block write completes
818  */
819 static void
820 zil_lwb_write_done(zio_t *zio)
821 {
822 	lwb_t *lwb = zio->io_private;
823 	zilog_t *zilog = lwb->lwb_zilog;
824 	dmu_tx_t *tx = lwb->lwb_tx;
825 
826 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
827 	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
828 	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
829 	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
830 	ASSERT(!BP_IS_GANG(zio->io_bp));
831 	ASSERT(!BP_IS_HOLE(zio->io_bp));
832 	ASSERT(zio->io_bp->blk_fill == 0);
833 
834 	/*
835 	 * Ensure the lwb buffer pointer is cleared before releasing
836 	 * the txg. If we have had an allocation failure and
837 	 * the txg is waiting to sync then we want want zil_sync()
838 	 * to remove the lwb so that it's not picked up as the next new
839 	 * one in zil_commit_writer(). zil_sync() will only remove
840 	 * the lwb if lwb_buf is null.
841 	 */
842 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
843 	mutex_enter(&zilog->zl_lock);
844 	lwb->lwb_buf = NULL;
845 	lwb->lwb_tx = NULL;
846 	mutex_exit(&zilog->zl_lock);
847 
848 	/*
849 	 * Now that we've written this log block, we have a stable pointer
850 	 * to the next block in the chain, so it's OK to let the txg in
851 	 * which we allocated the next block sync.
852 	 */
853 	dmu_tx_commit(tx);
854 }
855 
856 /*
857  * Initialize the io for a log block.
858  */
859 static void
860 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
861 {
862 	zbookmark_t zb;
863 
864 	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
865 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
866 	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
867 
868 	if (zilog->zl_root_zio == NULL) {
869 		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
870 		    ZIO_FLAG_CANFAIL);
871 	}
872 	if (lwb->lwb_zio == NULL) {
873 		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
874 		    0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
875 		    zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
876 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
877 	}
878 }
879 
880 /*
881  * Define a limited set of intent log block sizes.
882  *
883  * These must be a multiple of 4KB. Note only the amount used (again
884  * aligned to 4KB) actually gets written. However, we can't always just
885  * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
886  */
887 uint64_t zil_block_buckets[] = {
888     4096,		/* non TX_WRITE */
889     8192+4096,		/* data base */
890     32*1024 + 4096, 	/* NFS writes */
891     UINT64_MAX
892 };
893 
894 /*
895  * Use the slog as long as the logbias is 'latency' and the current commit size
896  * is less than the limit or the total list size is less than 2X the limit.
897  * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
898  */
899 uint64_t zil_slog_limit = 1024 * 1024;
900 #define	USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
901 	(((zilog)->zl_cur_used < zil_slog_limit) || \
902 	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
903 
904 /*
905  * Start a log block write and advance to the next log block.
906  * Calls are serialized.
907  */
908 static lwb_t *
909 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
910 {
911 	lwb_t *nlwb = NULL;
912 	zil_chain_t *zilc;
913 	spa_t *spa = zilog->zl_spa;
914 	blkptr_t *bp;
915 	dmu_tx_t *tx;
916 	uint64_t txg;
917 	uint64_t zil_blksz, wsz;
918 	int i, error;
919 
920 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
921 		zilc = (zil_chain_t *)lwb->lwb_buf;
922 		bp = &zilc->zc_next_blk;
923 	} else {
924 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
925 		bp = &zilc->zc_next_blk;
926 	}
927 
928 	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
929 
930 	/*
931 	 * Allocate the next block and save its address in this block
932 	 * before writing it in order to establish the log chain.
933 	 * Note that if the allocation of nlwb synced before we wrote
934 	 * the block that points at it (lwb), we'd leak it if we crashed.
935 	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
936 	 * We dirty the dataset to ensure that zil_sync() will be called
937 	 * to clean up in the event of allocation failure or I/O failure.
938 	 */
939 	tx = dmu_tx_create(zilog->zl_os);
940 	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
941 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
942 	txg = dmu_tx_get_txg(tx);
943 
944 	lwb->lwb_tx = tx;
945 
946 	/*
947 	 * Log blocks are pre-allocated. Here we select the size of the next
948 	 * block, based on size used in the last block.
949 	 * - first find the smallest bucket that will fit the block from a
950 	 *   limited set of block sizes. This is because it's faster to write
951 	 *   blocks allocated from the same metaslab as they are adjacent or
952 	 *   close.
953 	 * - next find the maximum from the new suggested size and an array of
954 	 *   previous sizes. This lessens a picket fence effect of wrongly
955 	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
956 	 *   requests.
957 	 *
958 	 * Note we only write what is used, but we can't just allocate
959 	 * the maximum block size because we can exhaust the available
960 	 * pool log space.
961 	 */
962 	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
963 	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
964 		continue;
965 	zil_blksz = zil_block_buckets[i];
966 	if (zil_blksz == UINT64_MAX)
967 		zil_blksz = SPA_MAXBLOCKSIZE;
968 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
969 	for (i = 0; i < ZIL_PREV_BLKS; i++)
970 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
971 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
972 
973 	BP_ZERO(bp);
974 	/* pass the old blkptr in order to spread log blocks across devs */
975 	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
976 	    USE_SLOG(zilog));
977 	if (error == 0) {
978 		ASSERT3U(bp->blk_birth, ==, txg);
979 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
980 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
981 
982 		/*
983 		 * Allocate a new log write buffer (lwb).
984 		 */
985 		nlwb = zil_alloc_lwb(zilog, bp, txg);
986 
987 		/* Record the block for later vdev flushing */
988 		zil_add_block(zilog, &lwb->lwb_blk);
989 	}
990 
991 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
992 		/* For Slim ZIL only write what is used. */
993 		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
994 		ASSERT3U(wsz, <=, lwb->lwb_sz);
995 		zio_shrink(lwb->lwb_zio, wsz);
996 
997 	} else {
998 		wsz = lwb->lwb_sz;
999 	}
1000 
1001 	zilc->zc_pad = 0;
1002 	zilc->zc_nused = lwb->lwb_nused;
1003 	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1004 
1005 	/*
1006 	 * clear unused data for security
1007 	 */
1008 	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1009 
1010 	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1011 
1012 	/*
1013 	 * If there was an allocation failure then nlwb will be null which
1014 	 * forces a txg_wait_synced().
1015 	 */
1016 	return (nlwb);
1017 }
1018 
1019 static lwb_t *
1020 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1021 {
1022 	lr_t *lrc = &itx->itx_lr; /* common log record */
1023 	lr_write_t *lrw = (lr_write_t *)lrc;
1024 	char *lr_buf;
1025 	uint64_t txg = lrc->lrc_txg;
1026 	uint64_t reclen = lrc->lrc_reclen;
1027 	uint64_t dlen = 0;
1028 
1029 	if (lwb == NULL)
1030 		return (NULL);
1031 
1032 	ASSERT(lwb->lwb_buf != NULL);
1033 	ASSERT(zilog_is_dirty(zilog) ||
1034 	    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1035 
1036 	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1037 		dlen = P2ROUNDUP_TYPED(
1038 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1039 
1040 	zilog->zl_cur_used += (reclen + dlen);
1041 
1042 	zil_lwb_write_init(zilog, lwb);
1043 
1044 	/*
1045 	 * If this record won't fit in the current log block, start a new one.
1046 	 */
1047 	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1048 		lwb = zil_lwb_write_start(zilog, lwb);
1049 		if (lwb == NULL)
1050 			return (NULL);
1051 		zil_lwb_write_init(zilog, lwb);
1052 		ASSERT(LWB_EMPTY(lwb));
1053 		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1054 			txg_wait_synced(zilog->zl_dmu_pool, txg);
1055 			return (lwb);
1056 		}
1057 	}
1058 
1059 	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1060 	bcopy(lrc, lr_buf, reclen);
1061 	lrc = (lr_t *)lr_buf;
1062 	lrw = (lr_write_t *)lrc;
1063 
1064 	/*
1065 	 * If it's a write, fetch the data or get its blkptr as appropriate.
1066 	 */
1067 	if (lrc->lrc_txtype == TX_WRITE) {
1068 		if (txg > spa_freeze_txg(zilog->zl_spa))
1069 			txg_wait_synced(zilog->zl_dmu_pool, txg);
1070 		if (itx->itx_wr_state != WR_COPIED) {
1071 			char *dbuf;
1072 			int error;
1073 
1074 			if (dlen) {
1075 				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1076 				dbuf = lr_buf + reclen;
1077 				lrw->lr_common.lrc_reclen += dlen;
1078 			} else {
1079 				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1080 				dbuf = NULL;
1081 			}
1082 			error = zilog->zl_get_data(
1083 			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1084 			if (error == EIO) {
1085 				txg_wait_synced(zilog->zl_dmu_pool, txg);
1086 				return (lwb);
1087 			}
1088 			if (error != 0) {
1089 				ASSERT(error == ENOENT || error == EEXIST ||
1090 				    error == EALREADY);
1091 				return (lwb);
1092 			}
1093 		}
1094 	}
1095 
1096 	/*
1097 	 * We're actually making an entry, so update lrc_seq to be the
1098 	 * log record sequence number.  Note that this is generally not
1099 	 * equal to the itx sequence number because not all transactions
1100 	 * are synchronous, and sometimes spa_sync() gets there first.
1101 	 */
1102 	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1103 	lwb->lwb_nused += reclen + dlen;
1104 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1105 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1106 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1107 
1108 	return (lwb);
1109 }
1110 
1111 itx_t *
1112 zil_itx_create(uint64_t txtype, size_t lrsize)
1113 {
1114 	itx_t *itx;
1115 
1116 	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1117 
1118 	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1119 	itx->itx_lr.lrc_txtype = txtype;
1120 	itx->itx_lr.lrc_reclen = lrsize;
1121 	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1122 	itx->itx_lr.lrc_seq = 0;	/* defensive */
1123 	itx->itx_sync = B_TRUE;		/* default is synchronous */
1124 
1125 	return (itx);
1126 }
1127 
1128 void
1129 zil_itx_destroy(itx_t *itx)
1130 {
1131 	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1132 }
1133 
1134 /*
1135  * Free up the sync and async itxs. The itxs_t has already been detached
1136  * so no locks are needed.
1137  */
1138 static void
1139 zil_itxg_clean(itxs_t *itxs)
1140 {
1141 	itx_t *itx;
1142 	list_t *list;
1143 	avl_tree_t *t;
1144 	void *cookie;
1145 	itx_async_node_t *ian;
1146 
1147 	list = &itxs->i_sync_list;
1148 	while ((itx = list_head(list)) != NULL) {
1149 		list_remove(list, itx);
1150 		kmem_free(itx, offsetof(itx_t, itx_lr) +
1151 		    itx->itx_lr.lrc_reclen);
1152 	}
1153 
1154 	cookie = NULL;
1155 	t = &itxs->i_async_tree;
1156 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1157 		list = &ian->ia_list;
1158 		while ((itx = list_head(list)) != NULL) {
1159 			list_remove(list, itx);
1160 			kmem_free(itx, offsetof(itx_t, itx_lr) +
1161 			    itx->itx_lr.lrc_reclen);
1162 		}
1163 		list_destroy(list);
1164 		kmem_free(ian, sizeof (itx_async_node_t));
1165 	}
1166 	avl_destroy(t);
1167 
1168 	kmem_free(itxs, sizeof (itxs_t));
1169 }
1170 
1171 static int
1172 zil_aitx_compare(const void *x1, const void *x2)
1173 {
1174 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1175 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1176 
1177 	if (o1 < o2)
1178 		return (-1);
1179 	if (o1 > o2)
1180 		return (1);
1181 
1182 	return (0);
1183 }
1184 
1185 /*
1186  * Remove all async itx with the given oid.
1187  */
1188 static void
1189 zil_remove_async(zilog_t *zilog, uint64_t oid)
1190 {
1191 	uint64_t otxg, txg;
1192 	itx_async_node_t *ian;
1193 	avl_tree_t *t;
1194 	avl_index_t where;
1195 	list_t clean_list;
1196 	itx_t *itx;
1197 
1198 	ASSERT(oid != 0);
1199 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1200 
1201 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1202 		otxg = ZILTEST_TXG;
1203 	else
1204 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1205 
1206 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1207 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1208 
1209 		mutex_enter(&itxg->itxg_lock);
1210 		if (itxg->itxg_txg != txg) {
1211 			mutex_exit(&itxg->itxg_lock);
1212 			continue;
1213 		}
1214 
1215 		/*
1216 		 * Locate the object node and append its list.
1217 		 */
1218 		t = &itxg->itxg_itxs->i_async_tree;
1219 		ian = avl_find(t, &oid, &where);
1220 		if (ian != NULL)
1221 			list_move_tail(&clean_list, &ian->ia_list);
1222 		mutex_exit(&itxg->itxg_lock);
1223 	}
1224 	while ((itx = list_head(&clean_list)) != NULL) {
1225 		list_remove(&clean_list, itx);
1226 		kmem_free(itx, offsetof(itx_t, itx_lr) +
1227 		    itx->itx_lr.lrc_reclen);
1228 	}
1229 	list_destroy(&clean_list);
1230 }
1231 
1232 void
1233 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1234 {
1235 	uint64_t txg;
1236 	itxg_t *itxg;
1237 	itxs_t *itxs, *clean = NULL;
1238 
1239 	/*
1240 	 * Object ids can be re-instantiated in the next txg so
1241 	 * remove any async transactions to avoid future leaks.
1242 	 * This can happen if a fsync occurs on the re-instantiated
1243 	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1244 	 * the new file data and flushes a write record for the old object.
1245 	 */
1246 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1247 		zil_remove_async(zilog, itx->itx_oid);
1248 
1249 	/*
1250 	 * Ensure the data of a renamed file is committed before the rename.
1251 	 */
1252 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1253 		zil_async_to_sync(zilog, itx->itx_oid);
1254 
1255 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1256 		txg = ZILTEST_TXG;
1257 	else
1258 		txg = dmu_tx_get_txg(tx);
1259 
1260 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1261 	mutex_enter(&itxg->itxg_lock);
1262 	itxs = itxg->itxg_itxs;
1263 	if (itxg->itxg_txg != txg) {
1264 		if (itxs != NULL) {
1265 			/*
1266 			 * The zil_clean callback hasn't got around to cleaning
1267 			 * this itxg. Save the itxs for release below.
1268 			 * This should be rare.
1269 			 */
1270 			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1271 			itxg->itxg_sod = 0;
1272 			clean = itxg->itxg_itxs;
1273 		}
1274 		ASSERT(itxg->itxg_sod == 0);
1275 		itxg->itxg_txg = txg;
1276 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1277 
1278 		list_create(&itxs->i_sync_list, sizeof (itx_t),
1279 		    offsetof(itx_t, itx_node));
1280 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1281 		    sizeof (itx_async_node_t),
1282 		    offsetof(itx_async_node_t, ia_node));
1283 	}
1284 	if (itx->itx_sync) {
1285 		list_insert_tail(&itxs->i_sync_list, itx);
1286 		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1287 		itxg->itxg_sod += itx->itx_sod;
1288 	} else {
1289 		avl_tree_t *t = &itxs->i_async_tree;
1290 		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1291 		itx_async_node_t *ian;
1292 		avl_index_t where;
1293 
1294 		ian = avl_find(t, &foid, &where);
1295 		if (ian == NULL) {
1296 			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1297 			list_create(&ian->ia_list, sizeof (itx_t),
1298 			    offsetof(itx_t, itx_node));
1299 			ian->ia_foid = foid;
1300 			avl_insert(t, ian, where);
1301 		}
1302 		list_insert_tail(&ian->ia_list, itx);
1303 	}
1304 
1305 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1306 	zilog_dirty(zilog, txg);
1307 	mutex_exit(&itxg->itxg_lock);
1308 
1309 	/* Release the old itxs now we've dropped the lock */
1310 	if (clean != NULL)
1311 		zil_itxg_clean(clean);
1312 }
1313 
1314 /*
1315  * If there are any in-memory intent log transactions which have now been
1316  * synced then start up a taskq to free them. We should only do this after we
1317  * have written out the uberblocks (i.e. txg has been comitted) so that
1318  * don't inadvertently clean out in-memory log records that would be required
1319  * by zil_commit().
1320  */
1321 void
1322 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1323 {
1324 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1325 	itxs_t *clean_me;
1326 
1327 	mutex_enter(&itxg->itxg_lock);
1328 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1329 		mutex_exit(&itxg->itxg_lock);
1330 		return;
1331 	}
1332 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1333 	ASSERT(itxg->itxg_txg != 0);
1334 	ASSERT(zilog->zl_clean_taskq != NULL);
1335 	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1336 	itxg->itxg_sod = 0;
1337 	clean_me = itxg->itxg_itxs;
1338 	itxg->itxg_itxs = NULL;
1339 	itxg->itxg_txg = 0;
1340 	mutex_exit(&itxg->itxg_lock);
1341 	/*
1342 	 * Preferably start a task queue to free up the old itxs but
1343 	 * if taskq_dispatch can't allocate resources to do that then
1344 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1345 	 * created a bad performance problem.
1346 	 */
1347 	if (taskq_dispatch(zilog->zl_clean_taskq,
1348 	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1349 		zil_itxg_clean(clean_me);
1350 }
1351 
1352 /*
1353  * Get the list of itxs to commit into zl_itx_commit_list.
1354  */
1355 static void
1356 zil_get_commit_list(zilog_t *zilog)
1357 {
1358 	uint64_t otxg, txg;
1359 	list_t *commit_list = &zilog->zl_itx_commit_list;
1360 	uint64_t push_sod = 0;
1361 
1362 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1363 		otxg = ZILTEST_TXG;
1364 	else
1365 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1366 
1367 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1368 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1369 
1370 		mutex_enter(&itxg->itxg_lock);
1371 		if (itxg->itxg_txg != txg) {
1372 			mutex_exit(&itxg->itxg_lock);
1373 			continue;
1374 		}
1375 
1376 		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1377 		push_sod += itxg->itxg_sod;
1378 		itxg->itxg_sod = 0;
1379 
1380 		mutex_exit(&itxg->itxg_lock);
1381 	}
1382 	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1383 }
1384 
1385 /*
1386  * Move the async itxs for a specified object to commit into sync lists.
1387  */
1388 static void
1389 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1390 {
1391 	uint64_t otxg, txg;
1392 	itx_async_node_t *ian;
1393 	avl_tree_t *t;
1394 	avl_index_t where;
1395 
1396 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1397 		otxg = ZILTEST_TXG;
1398 	else
1399 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1400 
1401 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1402 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1403 
1404 		mutex_enter(&itxg->itxg_lock);
1405 		if (itxg->itxg_txg != txg) {
1406 			mutex_exit(&itxg->itxg_lock);
1407 			continue;
1408 		}
1409 
1410 		/*
1411 		 * If a foid is specified then find that node and append its
1412 		 * list. Otherwise walk the tree appending all the lists
1413 		 * to the sync list. We add to the end rather than the
1414 		 * beginning to ensure the create has happened.
1415 		 */
1416 		t = &itxg->itxg_itxs->i_async_tree;
1417 		if (foid != 0) {
1418 			ian = avl_find(t, &foid, &where);
1419 			if (ian != NULL) {
1420 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1421 				    &ian->ia_list);
1422 			}
1423 		} else {
1424 			void *cookie = NULL;
1425 
1426 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1427 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1428 				    &ian->ia_list);
1429 				list_destroy(&ian->ia_list);
1430 				kmem_free(ian, sizeof (itx_async_node_t));
1431 			}
1432 		}
1433 		mutex_exit(&itxg->itxg_lock);
1434 	}
1435 }
1436 
1437 static void
1438 zil_commit_writer(zilog_t *zilog)
1439 {
1440 	uint64_t txg;
1441 	itx_t *itx;
1442 	lwb_t *lwb;
1443 	spa_t *spa = zilog->zl_spa;
1444 	int error = 0;
1445 
1446 	ASSERT(zilog->zl_root_zio == NULL);
1447 
1448 	mutex_exit(&zilog->zl_lock);
1449 
1450 	zil_get_commit_list(zilog);
1451 
1452 	/*
1453 	 * Return if there's nothing to commit before we dirty the fs by
1454 	 * calling zil_create().
1455 	 */
1456 	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1457 		mutex_enter(&zilog->zl_lock);
1458 		return;
1459 	}
1460 
1461 	if (zilog->zl_suspend) {
1462 		lwb = NULL;
1463 	} else {
1464 		lwb = list_tail(&zilog->zl_lwb_list);
1465 		if (lwb == NULL)
1466 			lwb = zil_create(zilog);
1467 	}
1468 
1469 	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1470 	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1471 		txg = itx->itx_lr.lrc_txg;
1472 		ASSERT(txg);
1473 
1474 		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1475 			lwb = zil_lwb_commit(zilog, itx, lwb);
1476 		list_remove(&zilog->zl_itx_commit_list, itx);
1477 		kmem_free(itx, offsetof(itx_t, itx_lr)
1478 		    + itx->itx_lr.lrc_reclen);
1479 	}
1480 	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1481 
1482 	/* write the last block out */
1483 	if (lwb != NULL && lwb->lwb_zio != NULL)
1484 		lwb = zil_lwb_write_start(zilog, lwb);
1485 
1486 	zilog->zl_cur_used = 0;
1487 
1488 	/*
1489 	 * Wait if necessary for the log blocks to be on stable storage.
1490 	 */
1491 	if (zilog->zl_root_zio) {
1492 		error = zio_wait(zilog->zl_root_zio);
1493 		zilog->zl_root_zio = NULL;
1494 		zil_flush_vdevs(zilog);
1495 	}
1496 
1497 	if (error || lwb == NULL)
1498 		txg_wait_synced(zilog->zl_dmu_pool, 0);
1499 
1500 	mutex_enter(&zilog->zl_lock);
1501 
1502 	/*
1503 	 * Remember the highest committed log sequence number for ztest.
1504 	 * We only update this value when all the log writes succeeded,
1505 	 * because ztest wants to ASSERT that it got the whole log chain.
1506 	 */
1507 	if (error == 0 && lwb != NULL)
1508 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1509 }
1510 
1511 /*
1512  * Commit zfs transactions to stable storage.
1513  * If foid is 0 push out all transactions, otherwise push only those
1514  * for that object or might reference that object.
1515  *
1516  * itxs are committed in batches. In a heavily stressed zil there will be
1517  * a commit writer thread who is writing out a bunch of itxs to the log
1518  * for a set of committing threads (cthreads) in the same batch as the writer.
1519  * Those cthreads are all waiting on the same cv for that batch.
1520  *
1521  * There will also be a different and growing batch of threads that are
1522  * waiting to commit (qthreads). When the committing batch completes
1523  * a transition occurs such that the cthreads exit and the qthreads become
1524  * cthreads. One of the new cthreads becomes the writer thread for the
1525  * batch. Any new threads arriving become new qthreads.
1526  *
1527  * Only 2 condition variables are needed and there's no transition
1528  * between the two cvs needed. They just flip-flop between qthreads
1529  * and cthreads.
1530  *
1531  * Using this scheme we can efficiently wakeup up only those threads
1532  * that have been committed.
1533  */
1534 void
1535 zil_commit(zilog_t *zilog, uint64_t foid)
1536 {
1537 	uint64_t mybatch;
1538 
1539 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1540 		return;
1541 
1542 	/* move the async itxs for the foid to the sync queues */
1543 	zil_async_to_sync(zilog, foid);
1544 
1545 	mutex_enter(&zilog->zl_lock);
1546 	mybatch = zilog->zl_next_batch;
1547 	while (zilog->zl_writer) {
1548 		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1549 		if (mybatch <= zilog->zl_com_batch) {
1550 			mutex_exit(&zilog->zl_lock);
1551 			return;
1552 		}
1553 	}
1554 
1555 	zilog->zl_next_batch++;
1556 	zilog->zl_writer = B_TRUE;
1557 	zil_commit_writer(zilog);
1558 	zilog->zl_com_batch = mybatch;
1559 	zilog->zl_writer = B_FALSE;
1560 	mutex_exit(&zilog->zl_lock);
1561 
1562 	/* wake up one thread to become the next writer */
1563 	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1564 
1565 	/* wake up all threads waiting for this batch to be committed */
1566 	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1567 }
1568 
1569 /*
1570  * Called in syncing context to free committed log blocks and update log header.
1571  */
1572 void
1573 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1574 {
1575 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1576 	uint64_t txg = dmu_tx_get_txg(tx);
1577 	spa_t *spa = zilog->zl_spa;
1578 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1579 	lwb_t *lwb;
1580 
1581 	/*
1582 	 * We don't zero out zl_destroy_txg, so make sure we don't try
1583 	 * to destroy it twice.
1584 	 */
1585 	if (spa_sync_pass(spa) != 1)
1586 		return;
1587 
1588 	mutex_enter(&zilog->zl_lock);
1589 
1590 	ASSERT(zilog->zl_stop_sync == 0);
1591 
1592 	if (*replayed_seq != 0) {
1593 		ASSERT(zh->zh_replay_seq < *replayed_seq);
1594 		zh->zh_replay_seq = *replayed_seq;
1595 		*replayed_seq = 0;
1596 	}
1597 
1598 	if (zilog->zl_destroy_txg == txg) {
1599 		blkptr_t blk = zh->zh_log;
1600 
1601 		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1602 
1603 		bzero(zh, sizeof (zil_header_t));
1604 		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1605 
1606 		if (zilog->zl_keep_first) {
1607 			/*
1608 			 * If this block was part of log chain that couldn't
1609 			 * be claimed because a device was missing during
1610 			 * zil_claim(), but that device later returns,
1611 			 * then this block could erroneously appear valid.
1612 			 * To guard against this, assign a new GUID to the new
1613 			 * log chain so it doesn't matter what blk points to.
1614 			 */
1615 			zil_init_log_chain(zilog, &blk);
1616 			zh->zh_log = blk;
1617 		}
1618 	}
1619 
1620 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1621 		zh->zh_log = lwb->lwb_blk;
1622 		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1623 			break;
1624 		list_remove(&zilog->zl_lwb_list, lwb);
1625 		zio_free_zil(spa, txg, &lwb->lwb_blk);
1626 		kmem_cache_free(zil_lwb_cache, lwb);
1627 
1628 		/*
1629 		 * If we don't have anything left in the lwb list then
1630 		 * we've had an allocation failure and we need to zero
1631 		 * out the zil_header blkptr so that we don't end
1632 		 * up freeing the same block twice.
1633 		 */
1634 		if (list_head(&zilog->zl_lwb_list) == NULL)
1635 			BP_ZERO(&zh->zh_log);
1636 	}
1637 	mutex_exit(&zilog->zl_lock);
1638 }
1639 
1640 void
1641 zil_init(void)
1642 {
1643 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1644 	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1645 }
1646 
1647 void
1648 zil_fini(void)
1649 {
1650 	kmem_cache_destroy(zil_lwb_cache);
1651 }
1652 
1653 void
1654 zil_set_sync(zilog_t *zilog, uint64_t sync)
1655 {
1656 	zilog->zl_sync = sync;
1657 }
1658 
1659 void
1660 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1661 {
1662 	zilog->zl_logbias = logbias;
1663 }
1664 
1665 zilog_t *
1666 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1667 {
1668 	zilog_t *zilog;
1669 
1670 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1671 
1672 	zilog->zl_header = zh_phys;
1673 	zilog->zl_os = os;
1674 	zilog->zl_spa = dmu_objset_spa(os);
1675 	zilog->zl_dmu_pool = dmu_objset_pool(os);
1676 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1677 	zilog->zl_logbias = dmu_objset_logbias(os);
1678 	zilog->zl_sync = dmu_objset_syncprop(os);
1679 	zilog->zl_next_batch = 1;
1680 
1681 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1682 
1683 	for (int i = 0; i < TXG_SIZE; i++) {
1684 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1685 		    MUTEX_DEFAULT, NULL);
1686 	}
1687 
1688 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1689 	    offsetof(lwb_t, lwb_node));
1690 
1691 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1692 	    offsetof(itx_t, itx_node));
1693 
1694 	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1695 
1696 	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1697 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1698 
1699 	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1700 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1701 	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1702 	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1703 
1704 	return (zilog);
1705 }
1706 
1707 void
1708 zil_free(zilog_t *zilog)
1709 {
1710 	zilog->zl_stop_sync = 1;
1711 
1712 	ASSERT0(zilog->zl_suspend);
1713 	ASSERT0(zilog->zl_suspending);
1714 
1715 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1716 	list_destroy(&zilog->zl_lwb_list);
1717 
1718 	avl_destroy(&zilog->zl_vdev_tree);
1719 	mutex_destroy(&zilog->zl_vdev_lock);
1720 
1721 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1722 	list_destroy(&zilog->zl_itx_commit_list);
1723 
1724 	for (int i = 0; i < TXG_SIZE; i++) {
1725 		/*
1726 		 * It's possible for an itx to be generated that doesn't dirty
1727 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1728 		 * callback to remove the entry. We remove those here.
1729 		 *
1730 		 * Also free up the ziltest itxs.
1731 		 */
1732 		if (zilog->zl_itxg[i].itxg_itxs)
1733 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1734 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1735 	}
1736 
1737 	mutex_destroy(&zilog->zl_lock);
1738 
1739 	cv_destroy(&zilog->zl_cv_writer);
1740 	cv_destroy(&zilog->zl_cv_suspend);
1741 	cv_destroy(&zilog->zl_cv_batch[0]);
1742 	cv_destroy(&zilog->zl_cv_batch[1]);
1743 
1744 	kmem_free(zilog, sizeof (zilog_t));
1745 }
1746 
1747 /*
1748  * Open an intent log.
1749  */
1750 zilog_t *
1751 zil_open(objset_t *os, zil_get_data_t *get_data)
1752 {
1753 	zilog_t *zilog = dmu_objset_zil(os);
1754 
1755 	ASSERT(zilog->zl_clean_taskq == NULL);
1756 	ASSERT(zilog->zl_get_data == NULL);
1757 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1758 
1759 	zilog->zl_get_data = get_data;
1760 	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1761 	    2, 2, TASKQ_PREPOPULATE);
1762 
1763 	return (zilog);
1764 }
1765 
1766 /*
1767  * Close an intent log.
1768  */
1769 void
1770 zil_close(zilog_t *zilog)
1771 {
1772 	lwb_t *lwb;
1773 	uint64_t txg = 0;
1774 
1775 	zil_commit(zilog, 0); /* commit all itx */
1776 
1777 	/*
1778 	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1779 	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1780 	 * callbacks have occurred that may clean the zil.  Only then can we
1781 	 * destroy the zl_clean_taskq.
1782 	 */
1783 	mutex_enter(&zilog->zl_lock);
1784 	lwb = list_tail(&zilog->zl_lwb_list);
1785 	if (lwb != NULL)
1786 		txg = lwb->lwb_max_txg;
1787 	mutex_exit(&zilog->zl_lock);
1788 	if (txg)
1789 		txg_wait_synced(zilog->zl_dmu_pool, txg);
1790 	ASSERT(!zilog_is_dirty(zilog));
1791 
1792 	taskq_destroy(zilog->zl_clean_taskq);
1793 	zilog->zl_clean_taskq = NULL;
1794 	zilog->zl_get_data = NULL;
1795 
1796 	/*
1797 	 * We should have only one LWB left on the list; remove it now.
1798 	 */
1799 	mutex_enter(&zilog->zl_lock);
1800 	lwb = list_head(&zilog->zl_lwb_list);
1801 	if (lwb != NULL) {
1802 		ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1803 		list_remove(&zilog->zl_lwb_list, lwb);
1804 		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1805 		kmem_cache_free(zil_lwb_cache, lwb);
1806 	}
1807 	mutex_exit(&zilog->zl_lock);
1808 }
1809 
1810 static char *suspend_tag = "zil suspending";
1811 
1812 /*
1813  * Suspend an intent log.  While in suspended mode, we still honor
1814  * synchronous semantics, but we rely on txg_wait_synced() to do it.
1815  * On old version pools, we suspend the log briefly when taking a
1816  * snapshot so that it will have an empty intent log.
1817  *
1818  * Long holds are not really intended to be used the way we do here --
1819  * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
1820  * could fail.  Therefore we take pains to only put a long hold if it is
1821  * actually necessary.  Fortunately, it will only be necessary if the
1822  * objset is currently mounted (or the ZVOL equivalent).  In that case it
1823  * will already have a long hold, so we are not really making things any worse.
1824  *
1825  * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1826  * zvol_state_t), and use their mechanism to prevent their hold from being
1827  * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
1828  * very little gain.
1829  *
1830  * if cookiep == NULL, this does both the suspend & resume.
1831  * Otherwise, it returns with the dataset "long held", and the cookie
1832  * should be passed into zil_resume().
1833  */
1834 int
1835 zil_suspend(const char *osname, void **cookiep)
1836 {
1837 	objset_t *os;
1838 	zilog_t *zilog;
1839 	const zil_header_t *zh;
1840 	int error;
1841 
1842 	error = dmu_objset_hold(osname, suspend_tag, &os);
1843 	if (error != 0)
1844 		return (error);
1845 	zilog = dmu_objset_zil(os);
1846 
1847 	mutex_enter(&zilog->zl_lock);
1848 	zh = zilog->zl_header;
1849 
1850 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1851 		mutex_exit(&zilog->zl_lock);
1852 		dmu_objset_rele(os, suspend_tag);
1853 		return (SET_ERROR(EBUSY));
1854 	}
1855 
1856 	/*
1857 	 * Don't put a long hold in the cases where we can avoid it.  This
1858 	 * is when there is no cookie so we are doing a suspend & resume
1859 	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1860 	 * for the suspend because it's already suspended, or there's no ZIL.
1861 	 */
1862 	if (cookiep == NULL && !zilog->zl_suspending &&
1863 	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1864 		mutex_exit(&zilog->zl_lock);
1865 		dmu_objset_rele(os, suspend_tag);
1866 		return (0);
1867 	}
1868 
1869 	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1870 	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1871 
1872 	zilog->zl_suspend++;
1873 
1874 	if (zilog->zl_suspend > 1) {
1875 		/*
1876 		 * Someone else is already suspending it.
1877 		 * Just wait for them to finish.
1878 		 */
1879 
1880 		while (zilog->zl_suspending)
1881 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1882 		mutex_exit(&zilog->zl_lock);
1883 
1884 		if (cookiep == NULL)
1885 			zil_resume(os);
1886 		else
1887 			*cookiep = os;
1888 		return (0);
1889 	}
1890 
1891 	/*
1892 	 * If there is no pointer to an on-disk block, this ZIL must not
1893 	 * be active (e.g. filesystem not mounted), so there's nothing
1894 	 * to clean up.
1895 	 */
1896 	if (BP_IS_HOLE(&zh->zh_log)) {
1897 		ASSERT(cookiep != NULL); /* fast path already handled */
1898 
1899 		*cookiep = os;
1900 		mutex_exit(&zilog->zl_lock);
1901 		return (0);
1902 	}
1903 
1904 	zilog->zl_suspending = B_TRUE;
1905 	mutex_exit(&zilog->zl_lock);
1906 
1907 	zil_commit(zilog, 0);
1908 
1909 	zil_destroy(zilog, B_FALSE);
1910 
1911 	mutex_enter(&zilog->zl_lock);
1912 	zilog->zl_suspending = B_FALSE;
1913 	cv_broadcast(&zilog->zl_cv_suspend);
1914 	mutex_exit(&zilog->zl_lock);
1915 
1916 	if (cookiep == NULL)
1917 		zil_resume(os);
1918 	else
1919 		*cookiep = os;
1920 	return (0);
1921 }
1922 
1923 void
1924 zil_resume(void *cookie)
1925 {
1926 	objset_t *os = cookie;
1927 	zilog_t *zilog = dmu_objset_zil(os);
1928 
1929 	mutex_enter(&zilog->zl_lock);
1930 	ASSERT(zilog->zl_suspend != 0);
1931 	zilog->zl_suspend--;
1932 	mutex_exit(&zilog->zl_lock);
1933 	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1934 	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1935 }
1936 
1937 typedef struct zil_replay_arg {
1938 	zil_replay_func_t **zr_replay;
1939 	void		*zr_arg;
1940 	boolean_t	zr_byteswap;
1941 	char		*zr_lr;
1942 } zil_replay_arg_t;
1943 
1944 static int
1945 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1946 {
1947 	char name[MAXNAMELEN];
1948 
1949 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
1950 
1951 	dmu_objset_name(zilog->zl_os, name);
1952 
1953 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1954 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1955 	    (u_longlong_t)lr->lrc_seq,
1956 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1957 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1958 
1959 	return (error);
1960 }
1961 
1962 static int
1963 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1964 {
1965 	zil_replay_arg_t *zr = zra;
1966 	const zil_header_t *zh = zilog->zl_header;
1967 	uint64_t reclen = lr->lrc_reclen;
1968 	uint64_t txtype = lr->lrc_txtype;
1969 	int error = 0;
1970 
1971 	zilog->zl_replaying_seq = lr->lrc_seq;
1972 
1973 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1974 		return (0);
1975 
1976 	if (lr->lrc_txg < claim_txg)		/* already committed */
1977 		return (0);
1978 
1979 	/* Strip case-insensitive bit, still present in log record */
1980 	txtype &= ~TX_CI;
1981 
1982 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
1983 		return (zil_replay_error(zilog, lr, EINVAL));
1984 
1985 	/*
1986 	 * If this record type can be logged out of order, the object
1987 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
1988 	 */
1989 	if (TX_OOO(txtype)) {
1990 		error = dmu_object_info(zilog->zl_os,
1991 		    ((lr_ooo_t *)lr)->lr_foid, NULL);
1992 		if (error == ENOENT || error == EEXIST)
1993 			return (0);
1994 	}
1995 
1996 	/*
1997 	 * Make a copy of the data so we can revise and extend it.
1998 	 */
1999 	bcopy(lr, zr->zr_lr, reclen);
2000 
2001 	/*
2002 	 * If this is a TX_WRITE with a blkptr, suck in the data.
2003 	 */
2004 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2005 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
2006 		    zr->zr_lr + reclen);
2007 		if (error != 0)
2008 			return (zil_replay_error(zilog, lr, error));
2009 	}
2010 
2011 	/*
2012 	 * The log block containing this lr may have been byteswapped
2013 	 * so that we can easily examine common fields like lrc_txtype.
2014 	 * However, the log is a mix of different record types, and only the
2015 	 * replay vectors know how to byteswap their records.  Therefore, if
2016 	 * the lr was byteswapped, undo it before invoking the replay vector.
2017 	 */
2018 	if (zr->zr_byteswap)
2019 		byteswap_uint64_array(zr->zr_lr, reclen);
2020 
2021 	/*
2022 	 * We must now do two things atomically: replay this log record,
2023 	 * and update the log header sequence number to reflect the fact that
2024 	 * we did so. At the end of each replay function the sequence number
2025 	 * is updated if we are in replay mode.
2026 	 */
2027 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2028 	if (error != 0) {
2029 		/*
2030 		 * The DMU's dnode layer doesn't see removes until the txg
2031 		 * commits, so a subsequent claim can spuriously fail with
2032 		 * EEXIST. So if we receive any error we try syncing out
2033 		 * any removes then retry the transaction.  Note that we
2034 		 * specify B_FALSE for byteswap now, so we don't do it twice.
2035 		 */
2036 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2037 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2038 		if (error != 0)
2039 			return (zil_replay_error(zilog, lr, error));
2040 	}
2041 	return (0);
2042 }
2043 
2044 /* ARGSUSED */
2045 static int
2046 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2047 {
2048 	zilog->zl_replay_blks++;
2049 
2050 	return (0);
2051 }
2052 
2053 /*
2054  * If this dataset has a non-empty intent log, replay it and destroy it.
2055  */
2056 void
2057 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2058 {
2059 	zilog_t *zilog = dmu_objset_zil(os);
2060 	const zil_header_t *zh = zilog->zl_header;
2061 	zil_replay_arg_t zr;
2062 
2063 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2064 		zil_destroy(zilog, B_TRUE);
2065 		return;
2066 	}
2067 
2068 	zr.zr_replay = replay_func;
2069 	zr.zr_arg = arg;
2070 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2071 	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2072 
2073 	/*
2074 	 * Wait for in-progress removes to sync before starting replay.
2075 	 */
2076 	txg_wait_synced(zilog->zl_dmu_pool, 0);
2077 
2078 	zilog->zl_replay = B_TRUE;
2079 	zilog->zl_replay_time = ddi_get_lbolt();
2080 	ASSERT(zilog->zl_replay_blks == 0);
2081 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2082 	    zh->zh_claim_txg);
2083 	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2084 
2085 	zil_destroy(zilog, B_FALSE);
2086 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2087 	zilog->zl_replay = B_FALSE;
2088 }
2089 
2090 boolean_t
2091 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2092 {
2093 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2094 		return (B_TRUE);
2095 
2096 	if (zilog->zl_replay) {
2097 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2098 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2099 		    zilog->zl_replaying_seq;
2100 		return (B_TRUE);
2101 	}
2102 
2103 	return (B_FALSE);
2104 }
2105 
2106 /* ARGSUSED */
2107 int
2108 zil_vdev_offline(const char *osname, void *arg)
2109 {
2110 	int error;
2111 
2112 	error = zil_suspend(osname, NULL);
2113 	if (error != 0)
2114 		return (SET_ERROR(EEXIST));
2115 	return (0);
2116 }
2117