xref: /freebsd/sys/contrib/openzfs/module/zfs/zil.c (revision 13ec1e3155c7e9bf037b12af186351b7fa9b9450)
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) 2011, 2018 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Integros [integros.com]
25  * Copyright (c) 2018 Datto Inc.
26  */
27 
28 /* Portions Copyright 2010 Robert Milkowski */
29 
30 #include <sys/zfs_context.h>
31 #include <sys/spa.h>
32 #include <sys/spa_impl.h>
33 #include <sys/dmu.h>
34 #include <sys/zap.h>
35 #include <sys/arc.h>
36 #include <sys/stat.h>
37 #include <sys/zil.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
43 #include <sys/metaslab.h>
44 #include <sys/trace_zfs.h>
45 #include <sys/abd.h>
46 
47 /*
48  * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
49  * calls that change the file system. Each itx has enough information to
50  * be able to replay them after a system crash, power loss, or
51  * equivalent failure mode. These are stored in memory until either:
52  *
53  *   1. they are committed to the pool by the DMU transaction group
54  *      (txg), at which point they can be discarded; or
55  *   2. they are committed to the on-disk ZIL for the dataset being
56  *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
57  *      requirement).
58  *
59  * In the event of a crash or power loss, the itxs contained by each
60  * dataset's on-disk ZIL will be replayed when that dataset is first
61  * instantiated (e.g. if the dataset is a normal filesystem, when it is
62  * first mounted).
63  *
64  * As hinted at above, there is one ZIL per dataset (both the in-memory
65  * representation, and the on-disk representation). The on-disk format
66  * consists of 3 parts:
67  *
68  * 	- a single, per-dataset, ZIL header; which points to a chain of
69  * 	- zero or more ZIL blocks; each of which contains
70  * 	- zero or more ZIL records
71  *
72  * A ZIL record holds the information necessary to replay a single
73  * system call transaction. A ZIL block can hold many ZIL records, and
74  * the blocks are chained together, similarly to a singly linked list.
75  *
76  * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
77  * block in the chain, and the ZIL header points to the first block in
78  * the chain.
79  *
80  * Note, there is not a fixed place in the pool to hold these ZIL
81  * blocks; they are dynamically allocated and freed as needed from the
82  * blocks available on the pool, though they can be preferentially
83  * allocated from a dedicated "log" vdev.
84  */
85 
86 /*
87  * This controls the amount of time that a ZIL block (lwb) will remain
88  * "open" when it isn't "full", and it has a thread waiting for it to be
89  * committed to stable storage. Please refer to the zil_commit_waiter()
90  * function (and the comments within it) for more details.
91  */
92 static int zfs_commit_timeout_pct = 5;
93 
94 /*
95  * See zil.h for more information about these fields.
96  */
97 static zil_stats_t zil_stats = {
98 	{ "zil_commit_count",			KSTAT_DATA_UINT64 },
99 	{ "zil_commit_writer_count",		KSTAT_DATA_UINT64 },
100 	{ "zil_itx_count",			KSTAT_DATA_UINT64 },
101 	{ "zil_itx_indirect_count",		KSTAT_DATA_UINT64 },
102 	{ "zil_itx_indirect_bytes",		KSTAT_DATA_UINT64 },
103 	{ "zil_itx_copied_count",		KSTAT_DATA_UINT64 },
104 	{ "zil_itx_copied_bytes",		KSTAT_DATA_UINT64 },
105 	{ "zil_itx_needcopy_count",		KSTAT_DATA_UINT64 },
106 	{ "zil_itx_needcopy_bytes",		KSTAT_DATA_UINT64 },
107 	{ "zil_itx_metaslab_normal_count",	KSTAT_DATA_UINT64 },
108 	{ "zil_itx_metaslab_normal_bytes",	KSTAT_DATA_UINT64 },
109 	{ "zil_itx_metaslab_slog_count",	KSTAT_DATA_UINT64 },
110 	{ "zil_itx_metaslab_slog_bytes",	KSTAT_DATA_UINT64 },
111 };
112 
113 static kstat_t *zil_ksp;
114 
115 /*
116  * Disable intent logging replay.  This global ZIL switch affects all pools.
117  */
118 int zil_replay_disable = 0;
119 
120 /*
121  * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
122  * the disk(s) by the ZIL after an LWB write has completed. Setting this
123  * will cause ZIL corruption on power loss if a volatile out-of-order
124  * write cache is enabled.
125  */
126 static int zil_nocacheflush = 0;
127 
128 /*
129  * Limit SLOG write size per commit executed with synchronous priority.
130  * Any writes above that will be executed with lower (asynchronous) priority
131  * to limit potential SLOG device abuse by single active ZIL writer.
132  */
133 static unsigned long zil_slog_bulk = 768 * 1024;
134 
135 static kmem_cache_t *zil_lwb_cache;
136 static kmem_cache_t *zil_zcw_cache;
137 
138 #define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
139     sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
140 
141 static int
142 zil_bp_compare(const void *x1, const void *x2)
143 {
144 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
145 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
146 
147 	int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
148 	if (likely(cmp))
149 		return (cmp);
150 
151 	return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
152 }
153 
154 static void
155 zil_bp_tree_init(zilog_t *zilog)
156 {
157 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
158 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
159 }
160 
161 static void
162 zil_bp_tree_fini(zilog_t *zilog)
163 {
164 	avl_tree_t *t = &zilog->zl_bp_tree;
165 	zil_bp_node_t *zn;
166 	void *cookie = NULL;
167 
168 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
169 		kmem_free(zn, sizeof (zil_bp_node_t));
170 
171 	avl_destroy(t);
172 }
173 
174 int
175 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
176 {
177 	avl_tree_t *t = &zilog->zl_bp_tree;
178 	const dva_t *dva;
179 	zil_bp_node_t *zn;
180 	avl_index_t where;
181 
182 	if (BP_IS_EMBEDDED(bp))
183 		return (0);
184 
185 	dva = BP_IDENTITY(bp);
186 
187 	if (avl_find(t, dva, &where) != NULL)
188 		return (SET_ERROR(EEXIST));
189 
190 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
191 	zn->zn_dva = *dva;
192 	avl_insert(t, zn, where);
193 
194 	return (0);
195 }
196 
197 static zil_header_t *
198 zil_header_in_syncing_context(zilog_t *zilog)
199 {
200 	return ((zil_header_t *)zilog->zl_header);
201 }
202 
203 static void
204 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
205 {
206 	zio_cksum_t *zc = &bp->blk_cksum;
207 
208 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0],
209 	    sizeof (zc->zc_word[ZIL_ZC_GUID_0]));
210 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1],
211 	    sizeof (zc->zc_word[ZIL_ZC_GUID_1]));
212 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
213 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
214 }
215 
216 /*
217  * Read a log block and make sure it's valid.
218  */
219 static int
220 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
221     blkptr_t *nbp, void *dst, char **end)
222 {
223 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
224 	arc_flags_t aflags = ARC_FLAG_WAIT;
225 	arc_buf_t *abuf = NULL;
226 	zbookmark_phys_t zb;
227 	int error;
228 
229 	if (zilog->zl_header->zh_claim_txg == 0)
230 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
231 
232 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
233 		zio_flags |= ZIO_FLAG_SPECULATIVE;
234 
235 	if (!decrypt)
236 		zio_flags |= ZIO_FLAG_RAW;
237 
238 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
239 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
240 
241 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
242 	    &abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
243 
244 	if (error == 0) {
245 		zio_cksum_t cksum = bp->blk_cksum;
246 
247 		/*
248 		 * Validate the checksummed log block.
249 		 *
250 		 * Sequence numbers should be... sequential.  The checksum
251 		 * verifier for the next block should be bp's checksum plus 1.
252 		 *
253 		 * Also check the log chain linkage and size used.
254 		 */
255 		cksum.zc_word[ZIL_ZC_SEQ]++;
256 
257 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
258 			zil_chain_t *zilc = abuf->b_data;
259 			char *lr = (char *)(zilc + 1);
260 			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
261 
262 			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
263 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
264 				error = SET_ERROR(ECKSUM);
265 			} else {
266 				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
267 				bcopy(lr, dst, len);
268 				*end = (char *)dst + len;
269 				*nbp = zilc->zc_next_blk;
270 			}
271 		} else {
272 			char *lr = abuf->b_data;
273 			uint64_t size = BP_GET_LSIZE(bp);
274 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
275 
276 			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
277 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
278 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
279 				error = SET_ERROR(ECKSUM);
280 			} else {
281 				ASSERT3U(zilc->zc_nused, <=,
282 				    SPA_OLD_MAXBLOCKSIZE);
283 				bcopy(lr, dst, zilc->zc_nused);
284 				*end = (char *)dst + zilc->zc_nused;
285 				*nbp = zilc->zc_next_blk;
286 			}
287 		}
288 
289 		arc_buf_destroy(abuf, &abuf);
290 	}
291 
292 	return (error);
293 }
294 
295 /*
296  * Read a TX_WRITE log data block.
297  */
298 static int
299 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
300 {
301 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
302 	const blkptr_t *bp = &lr->lr_blkptr;
303 	arc_flags_t aflags = ARC_FLAG_WAIT;
304 	arc_buf_t *abuf = NULL;
305 	zbookmark_phys_t zb;
306 	int error;
307 
308 	if (BP_IS_HOLE(bp)) {
309 		if (wbuf != NULL)
310 			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
311 		return (0);
312 	}
313 
314 	if (zilog->zl_header->zh_claim_txg == 0)
315 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
316 
317 	/*
318 	 * If we are not using the resulting data, we are just checking that
319 	 * it hasn't been corrupted so we don't need to waste CPU time
320 	 * decompressing and decrypting it.
321 	 */
322 	if (wbuf == NULL)
323 		zio_flags |= ZIO_FLAG_RAW;
324 
325 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
326 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
327 
328 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
329 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
330 
331 	if (error == 0) {
332 		if (wbuf != NULL)
333 			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
334 		arc_buf_destroy(abuf, &abuf);
335 	}
336 
337 	return (error);
338 }
339 
340 /*
341  * Parse the intent log, and call parse_func for each valid record within.
342  */
343 int
344 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
345     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
346     boolean_t decrypt)
347 {
348 	const zil_header_t *zh = zilog->zl_header;
349 	boolean_t claimed = !!zh->zh_claim_txg;
350 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
351 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
352 	uint64_t max_blk_seq = 0;
353 	uint64_t max_lr_seq = 0;
354 	uint64_t blk_count = 0;
355 	uint64_t lr_count = 0;
356 	blkptr_t blk, next_blk;
357 	char *lrbuf, *lrp;
358 	int error = 0;
359 
360 	bzero(&next_blk, sizeof (blkptr_t));
361 
362 	/*
363 	 * Old logs didn't record the maximum zh_claim_lr_seq.
364 	 */
365 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
366 		claim_lr_seq = UINT64_MAX;
367 
368 	/*
369 	 * Starting at the block pointed to by zh_log we read the log chain.
370 	 * For each block in the chain we strongly check that block to
371 	 * ensure its validity.  We stop when an invalid block is found.
372 	 * For each block pointer in the chain we call parse_blk_func().
373 	 * For each record in each valid block we call parse_lr_func().
374 	 * If the log has been claimed, stop if we encounter a sequence
375 	 * number greater than the highest claimed sequence number.
376 	 */
377 	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
378 	zil_bp_tree_init(zilog);
379 
380 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
381 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
382 		int reclen;
383 		char *end = NULL;
384 
385 		if (blk_seq > claim_blk_seq)
386 			break;
387 
388 		error = parse_blk_func(zilog, &blk, arg, txg);
389 		if (error != 0)
390 			break;
391 		ASSERT3U(max_blk_seq, <, blk_seq);
392 		max_blk_seq = blk_seq;
393 		blk_count++;
394 
395 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
396 			break;
397 
398 		error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
399 		    lrbuf, &end);
400 		if (error != 0)
401 			break;
402 
403 		for (lrp = lrbuf; lrp < end; lrp += reclen) {
404 			lr_t *lr = (lr_t *)lrp;
405 			reclen = lr->lrc_reclen;
406 			ASSERT3U(reclen, >=, sizeof (lr_t));
407 			if (lr->lrc_seq > claim_lr_seq)
408 				goto done;
409 
410 			error = parse_lr_func(zilog, lr, arg, txg);
411 			if (error != 0)
412 				goto done;
413 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
414 			max_lr_seq = lr->lrc_seq;
415 			lr_count++;
416 		}
417 	}
418 done:
419 	zilog->zl_parse_error = error;
420 	zilog->zl_parse_blk_seq = max_blk_seq;
421 	zilog->zl_parse_lr_seq = max_lr_seq;
422 	zilog->zl_parse_blk_count = blk_count;
423 	zilog->zl_parse_lr_count = lr_count;
424 
425 	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
426 	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq) ||
427 	    (decrypt && error == EIO));
428 
429 	zil_bp_tree_fini(zilog);
430 	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
431 
432 	return (error);
433 }
434 
435 static int
436 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
437     uint64_t first_txg)
438 {
439 	(void) tx;
440 	ASSERT(!BP_IS_HOLE(bp));
441 
442 	/*
443 	 * As we call this function from the context of a rewind to a
444 	 * checkpoint, each ZIL block whose txg is later than the txg
445 	 * that we rewind to is invalid. Thus, we return -1 so
446 	 * zil_parse() doesn't attempt to read it.
447 	 */
448 	if (bp->blk_birth >= first_txg)
449 		return (-1);
450 
451 	if (zil_bp_tree_add(zilog, bp) != 0)
452 		return (0);
453 
454 	zio_free(zilog->zl_spa, first_txg, bp);
455 	return (0);
456 }
457 
458 static int
459 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
460     uint64_t first_txg)
461 {
462 	(void) zilog, (void) lrc, (void) tx, (void) first_txg;
463 	return (0);
464 }
465 
466 static int
467 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
468     uint64_t first_txg)
469 {
470 	/*
471 	 * Claim log block if not already committed and not already claimed.
472 	 * If tx == NULL, just verify that the block is claimable.
473 	 */
474 	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
475 	    zil_bp_tree_add(zilog, bp) != 0)
476 		return (0);
477 
478 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
479 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
480 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
481 }
482 
483 static int
484 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
485     uint64_t first_txg)
486 {
487 	lr_write_t *lr = (lr_write_t *)lrc;
488 	int error;
489 
490 	if (lrc->lrc_txtype != TX_WRITE)
491 		return (0);
492 
493 	/*
494 	 * If the block is not readable, don't claim it.  This can happen
495 	 * in normal operation when a log block is written to disk before
496 	 * some of the dmu_sync() blocks it points to.  In this case, the
497 	 * transaction cannot have been committed to anyone (we would have
498 	 * waited for all writes to be stable first), so it is semantically
499 	 * correct to declare this the end of the log.
500 	 */
501 	if (lr->lr_blkptr.blk_birth >= first_txg) {
502 		error = zil_read_log_data(zilog, lr, NULL);
503 		if (error != 0)
504 			return (error);
505 	}
506 
507 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
508 }
509 
510 static int
511 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
512     uint64_t claim_txg)
513 {
514 	(void) claim_txg;
515 
516 	zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
517 
518 	return (0);
519 }
520 
521 static int
522 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
523     uint64_t claim_txg)
524 {
525 	lr_write_t *lr = (lr_write_t *)lrc;
526 	blkptr_t *bp = &lr->lr_blkptr;
527 
528 	/*
529 	 * If we previously claimed it, we need to free it.
530 	 */
531 	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
532 	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
533 	    !BP_IS_HOLE(bp))
534 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
535 
536 	return (0);
537 }
538 
539 static int
540 zil_lwb_vdev_compare(const void *x1, const void *x2)
541 {
542 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
543 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
544 
545 	return (TREE_CMP(v1, v2));
546 }
547 
548 static lwb_t *
549 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg,
550     boolean_t fastwrite)
551 {
552 	lwb_t *lwb;
553 
554 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
555 	lwb->lwb_zilog = zilog;
556 	lwb->lwb_blk = *bp;
557 	lwb->lwb_fastwrite = fastwrite;
558 	lwb->lwb_slog = slog;
559 	lwb->lwb_state = LWB_STATE_CLOSED;
560 	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
561 	lwb->lwb_max_txg = txg;
562 	lwb->lwb_write_zio = NULL;
563 	lwb->lwb_root_zio = NULL;
564 	lwb->lwb_tx = NULL;
565 	lwb->lwb_issued_timestamp = 0;
566 	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
567 		lwb->lwb_nused = sizeof (zil_chain_t);
568 		lwb->lwb_sz = BP_GET_LSIZE(bp);
569 	} else {
570 		lwb->lwb_nused = 0;
571 		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
572 	}
573 
574 	mutex_enter(&zilog->zl_lock);
575 	list_insert_tail(&zilog->zl_lwb_list, lwb);
576 	mutex_exit(&zilog->zl_lock);
577 
578 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
579 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
580 	VERIFY(list_is_empty(&lwb->lwb_waiters));
581 	VERIFY(list_is_empty(&lwb->lwb_itxs));
582 
583 	return (lwb);
584 }
585 
586 static void
587 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
588 {
589 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
590 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
591 	VERIFY(list_is_empty(&lwb->lwb_waiters));
592 	VERIFY(list_is_empty(&lwb->lwb_itxs));
593 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
594 	ASSERT3P(lwb->lwb_write_zio, ==, NULL);
595 	ASSERT3P(lwb->lwb_root_zio, ==, NULL);
596 	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
597 	ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
598 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
599 
600 	/*
601 	 * Clear the zilog's field to indicate this lwb is no longer
602 	 * valid, and prevent use-after-free errors.
603 	 */
604 	if (zilog->zl_last_lwb_opened == lwb)
605 		zilog->zl_last_lwb_opened = NULL;
606 
607 	kmem_cache_free(zil_lwb_cache, lwb);
608 }
609 
610 /*
611  * Called when we create in-memory log transactions so that we know
612  * to cleanup the itxs at the end of spa_sync().
613  */
614 static void
615 zilog_dirty(zilog_t *zilog, uint64_t txg)
616 {
617 	dsl_pool_t *dp = zilog->zl_dmu_pool;
618 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
619 
620 	ASSERT(spa_writeable(zilog->zl_spa));
621 
622 	if (ds->ds_is_snapshot)
623 		panic("dirtying snapshot!");
624 
625 	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
626 		/* up the hold count until we can be written out */
627 		dmu_buf_add_ref(ds->ds_dbuf, zilog);
628 
629 		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
630 	}
631 }
632 
633 /*
634  * Determine if the zil is dirty in the specified txg. Callers wanting to
635  * ensure that the dirty state does not change must hold the itxg_lock for
636  * the specified txg. Holding the lock will ensure that the zil cannot be
637  * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
638  * state.
639  */
640 static boolean_t __maybe_unused
641 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
642 {
643 	dsl_pool_t *dp = zilog->zl_dmu_pool;
644 
645 	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
646 		return (B_TRUE);
647 	return (B_FALSE);
648 }
649 
650 /*
651  * Determine if the zil is dirty. The zil is considered dirty if it has
652  * any pending itx records that have not been cleaned by zil_clean().
653  */
654 static boolean_t
655 zilog_is_dirty(zilog_t *zilog)
656 {
657 	dsl_pool_t *dp = zilog->zl_dmu_pool;
658 
659 	for (int t = 0; t < TXG_SIZE; t++) {
660 		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
661 			return (B_TRUE);
662 	}
663 	return (B_FALSE);
664 }
665 
666 /*
667  * Create an on-disk intent log.
668  */
669 static lwb_t *
670 zil_create(zilog_t *zilog)
671 {
672 	const zil_header_t *zh = zilog->zl_header;
673 	lwb_t *lwb = NULL;
674 	uint64_t txg = 0;
675 	dmu_tx_t *tx = NULL;
676 	blkptr_t blk;
677 	int error = 0;
678 	boolean_t fastwrite = FALSE;
679 	boolean_t slog = FALSE;
680 
681 	/*
682 	 * Wait for any previous destroy to complete.
683 	 */
684 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
685 
686 	ASSERT(zh->zh_claim_txg == 0);
687 	ASSERT(zh->zh_replay_seq == 0);
688 
689 	blk = zh->zh_log;
690 
691 	/*
692 	 * Allocate an initial log block if:
693 	 *    - there isn't one already
694 	 *    - the existing block is the wrong endianness
695 	 */
696 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
697 		tx = dmu_tx_create(zilog->zl_os);
698 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
699 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
700 		txg = dmu_tx_get_txg(tx);
701 
702 		if (!BP_IS_HOLE(&blk)) {
703 			zio_free(zilog->zl_spa, txg, &blk);
704 			BP_ZERO(&blk);
705 		}
706 
707 		error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
708 		    ZIL_MIN_BLKSZ, &slog);
709 		fastwrite = TRUE;
710 
711 		if (error == 0)
712 			zil_init_log_chain(zilog, &blk);
713 	}
714 
715 	/*
716 	 * Allocate a log write block (lwb) for the first log block.
717 	 */
718 	if (error == 0)
719 		lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
720 
721 	/*
722 	 * If we just allocated the first log block, commit our transaction
723 	 * and wait for zil_sync() to stuff the block pointer into zh_log.
724 	 * (zh is part of the MOS, so we cannot modify it in open context.)
725 	 */
726 	if (tx != NULL) {
727 		dmu_tx_commit(tx);
728 		txg_wait_synced(zilog->zl_dmu_pool, txg);
729 	}
730 
731 	ASSERT(error != 0 || bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
732 	IMPLY(error == 0, lwb != NULL);
733 
734 	return (lwb);
735 }
736 
737 /*
738  * In one tx, free all log blocks and clear the log header. If keep_first
739  * is set, then we're replaying a log with no content. We want to keep the
740  * first block, however, so that the first synchronous transaction doesn't
741  * require a txg_wait_synced() in zil_create(). We don't need to
742  * txg_wait_synced() here either when keep_first is set, because both
743  * zil_create() and zil_destroy() will wait for any in-progress destroys
744  * to complete.
745  */
746 void
747 zil_destroy(zilog_t *zilog, boolean_t keep_first)
748 {
749 	const zil_header_t *zh = zilog->zl_header;
750 	lwb_t *lwb;
751 	dmu_tx_t *tx;
752 	uint64_t txg;
753 
754 	/*
755 	 * Wait for any previous destroy to complete.
756 	 */
757 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
758 
759 	zilog->zl_old_header = *zh;		/* debugging aid */
760 
761 	if (BP_IS_HOLE(&zh->zh_log))
762 		return;
763 
764 	tx = dmu_tx_create(zilog->zl_os);
765 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
766 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
767 	txg = dmu_tx_get_txg(tx);
768 
769 	mutex_enter(&zilog->zl_lock);
770 
771 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
772 	zilog->zl_destroy_txg = txg;
773 	zilog->zl_keep_first = keep_first;
774 
775 	if (!list_is_empty(&zilog->zl_lwb_list)) {
776 		ASSERT(zh->zh_claim_txg == 0);
777 		VERIFY(!keep_first);
778 		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
779 			if (lwb->lwb_fastwrite)
780 				metaslab_fastwrite_unmark(zilog->zl_spa,
781 				    &lwb->lwb_blk);
782 
783 			list_remove(&zilog->zl_lwb_list, lwb);
784 			if (lwb->lwb_buf != NULL)
785 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
786 			zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
787 			zil_free_lwb(zilog, lwb);
788 		}
789 	} else if (!keep_first) {
790 		zil_destroy_sync(zilog, tx);
791 	}
792 	mutex_exit(&zilog->zl_lock);
793 
794 	dmu_tx_commit(tx);
795 }
796 
797 void
798 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
799 {
800 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
801 	(void) zil_parse(zilog, zil_free_log_block,
802 	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
803 }
804 
805 int
806 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
807 {
808 	dmu_tx_t *tx = txarg;
809 	zilog_t *zilog;
810 	uint64_t first_txg;
811 	zil_header_t *zh;
812 	objset_t *os;
813 	int error;
814 
815 	error = dmu_objset_own_obj(dp, ds->ds_object,
816 	    DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
817 	if (error != 0) {
818 		/*
819 		 * EBUSY indicates that the objset is inconsistent, in which
820 		 * case it can not have a ZIL.
821 		 */
822 		if (error != EBUSY) {
823 			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
824 			    (unsigned long long)ds->ds_object, error);
825 		}
826 
827 		return (0);
828 	}
829 
830 	zilog = dmu_objset_zil(os);
831 	zh = zil_header_in_syncing_context(zilog);
832 	ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
833 	first_txg = spa_min_claim_txg(zilog->zl_spa);
834 
835 	/*
836 	 * If the spa_log_state is not set to be cleared, check whether
837 	 * the current uberblock is a checkpoint one and if the current
838 	 * header has been claimed before moving on.
839 	 *
840 	 * If the current uberblock is a checkpointed uberblock then
841 	 * one of the following scenarios took place:
842 	 *
843 	 * 1] We are currently rewinding to the checkpoint of the pool.
844 	 * 2] We crashed in the middle of a checkpoint rewind but we
845 	 *    did manage to write the checkpointed uberblock to the
846 	 *    vdev labels, so when we tried to import the pool again
847 	 *    the checkpointed uberblock was selected from the import
848 	 *    procedure.
849 	 *
850 	 * In both cases we want to zero out all the ZIL blocks, except
851 	 * the ones that have been claimed at the time of the checkpoint
852 	 * (their zh_claim_txg != 0). The reason is that these blocks
853 	 * may be corrupted since we may have reused their locations on
854 	 * disk after we took the checkpoint.
855 	 *
856 	 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
857 	 * when we first figure out whether the current uberblock is
858 	 * checkpointed or not. Unfortunately, that would discard all
859 	 * the logs, including the ones that are claimed, and we would
860 	 * leak space.
861 	 */
862 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
863 	    (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
864 	    zh->zh_claim_txg == 0)) {
865 		if (!BP_IS_HOLE(&zh->zh_log)) {
866 			(void) zil_parse(zilog, zil_clear_log_block,
867 			    zil_noop_log_record, tx, first_txg, B_FALSE);
868 		}
869 		BP_ZERO(&zh->zh_log);
870 		if (os->os_encrypted)
871 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
872 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
873 		dmu_objset_disown(os, B_FALSE, FTAG);
874 		return (0);
875 	}
876 
877 	/*
878 	 * If we are not rewinding and opening the pool normally, then
879 	 * the min_claim_txg should be equal to the first txg of the pool.
880 	 */
881 	ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
882 
883 	/*
884 	 * Claim all log blocks if we haven't already done so, and remember
885 	 * the highest claimed sequence number.  This ensures that if we can
886 	 * read only part of the log now (e.g. due to a missing device),
887 	 * but we can read the entire log later, we will not try to replay
888 	 * or destroy beyond the last block we successfully claimed.
889 	 */
890 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
891 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
892 		(void) zil_parse(zilog, zil_claim_log_block,
893 		    zil_claim_log_record, tx, first_txg, B_FALSE);
894 		zh->zh_claim_txg = first_txg;
895 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
896 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
897 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
898 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
899 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
900 		if (os->os_encrypted)
901 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
902 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
903 	}
904 
905 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
906 	dmu_objset_disown(os, B_FALSE, FTAG);
907 	return (0);
908 }
909 
910 /*
911  * Check the log by walking the log chain.
912  * Checksum errors are ok as they indicate the end of the chain.
913  * Any other error (no device or read failure) returns an error.
914  */
915 int
916 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
917 {
918 	(void) dp;
919 	zilog_t *zilog;
920 	objset_t *os;
921 	blkptr_t *bp;
922 	int error;
923 
924 	ASSERT(tx == NULL);
925 
926 	error = dmu_objset_from_ds(ds, &os);
927 	if (error != 0) {
928 		cmn_err(CE_WARN, "can't open objset %llu, error %d",
929 		    (unsigned long long)ds->ds_object, error);
930 		return (0);
931 	}
932 
933 	zilog = dmu_objset_zil(os);
934 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
935 
936 	if (!BP_IS_HOLE(bp)) {
937 		vdev_t *vd;
938 		boolean_t valid = B_TRUE;
939 
940 		/*
941 		 * Check the first block and determine if it's on a log device
942 		 * which may have been removed or faulted prior to loading this
943 		 * pool.  If so, there's no point in checking the rest of the
944 		 * log as its content should have already been synced to the
945 		 * pool.
946 		 */
947 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
948 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
949 		if (vd->vdev_islog && vdev_is_dead(vd))
950 			valid = vdev_log_state_valid(vd);
951 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
952 
953 		if (!valid)
954 			return (0);
955 
956 		/*
957 		 * Check whether the current uberblock is checkpointed (e.g.
958 		 * we are rewinding) and whether the current header has been
959 		 * claimed or not. If it hasn't then skip verifying it. We
960 		 * do this because its ZIL blocks may be part of the pool's
961 		 * state before the rewind, which is no longer valid.
962 		 */
963 		zil_header_t *zh = zil_header_in_syncing_context(zilog);
964 		if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
965 		    zh->zh_claim_txg == 0)
966 			return (0);
967 	}
968 
969 	/*
970 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
971 	 * any blocks, but just determine whether it is possible to do so.
972 	 * In addition to checking the log chain, zil_claim_log_block()
973 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
974 	 * which will update spa_max_claim_txg.  See spa_load() for details.
975 	 */
976 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
977 	    zilog->zl_header->zh_claim_txg ? -1ULL :
978 	    spa_min_claim_txg(os->os_spa), B_FALSE);
979 
980 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
981 }
982 
983 /*
984  * When an itx is "skipped", this function is used to properly mark the
985  * waiter as "done, and signal any thread(s) waiting on it. An itx can
986  * be skipped (and not committed to an lwb) for a variety of reasons,
987  * one of them being that the itx was committed via spa_sync(), prior to
988  * it being committed to an lwb; this can happen if a thread calling
989  * zil_commit() is racing with spa_sync().
990  */
991 static void
992 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
993 {
994 	mutex_enter(&zcw->zcw_lock);
995 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
996 	zcw->zcw_done = B_TRUE;
997 	cv_broadcast(&zcw->zcw_cv);
998 	mutex_exit(&zcw->zcw_lock);
999 }
1000 
1001 /*
1002  * This function is used when the given waiter is to be linked into an
1003  * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1004  * At this point, the waiter will no longer be referenced by the itx,
1005  * and instead, will be referenced by the lwb.
1006  */
1007 static void
1008 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1009 {
1010 	/*
1011 	 * The lwb_waiters field of the lwb is protected by the zilog's
1012 	 * zl_lock, thus it must be held when calling this function.
1013 	 */
1014 	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1015 
1016 	mutex_enter(&zcw->zcw_lock);
1017 	ASSERT(!list_link_active(&zcw->zcw_node));
1018 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1019 	ASSERT3P(lwb, !=, NULL);
1020 	ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
1021 	    lwb->lwb_state == LWB_STATE_ISSUED ||
1022 	    lwb->lwb_state == LWB_STATE_WRITE_DONE);
1023 
1024 	list_insert_tail(&lwb->lwb_waiters, zcw);
1025 	zcw->zcw_lwb = lwb;
1026 	mutex_exit(&zcw->zcw_lock);
1027 }
1028 
1029 /*
1030  * This function is used when zio_alloc_zil() fails to allocate a ZIL
1031  * block, and the given waiter must be linked to the "nolwb waiters"
1032  * list inside of zil_process_commit_list().
1033  */
1034 static void
1035 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1036 {
1037 	mutex_enter(&zcw->zcw_lock);
1038 	ASSERT(!list_link_active(&zcw->zcw_node));
1039 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1040 	list_insert_tail(nolwb, zcw);
1041 	mutex_exit(&zcw->zcw_lock);
1042 }
1043 
1044 void
1045 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1046 {
1047 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1048 	avl_index_t where;
1049 	zil_vdev_node_t *zv, zvsearch;
1050 	int ndvas = BP_GET_NDVAS(bp);
1051 	int i;
1052 
1053 	if (zil_nocacheflush)
1054 		return;
1055 
1056 	mutex_enter(&lwb->lwb_vdev_lock);
1057 	for (i = 0; i < ndvas; i++) {
1058 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1059 		if (avl_find(t, &zvsearch, &where) == NULL) {
1060 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1061 			zv->zv_vdev = zvsearch.zv_vdev;
1062 			avl_insert(t, zv, where);
1063 		}
1064 	}
1065 	mutex_exit(&lwb->lwb_vdev_lock);
1066 }
1067 
1068 static void
1069 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1070 {
1071 	avl_tree_t *src = &lwb->lwb_vdev_tree;
1072 	avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1073 	void *cookie = NULL;
1074 	zil_vdev_node_t *zv;
1075 
1076 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1077 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1078 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1079 
1080 	/*
1081 	 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1082 	 * not need the protection of lwb_vdev_lock (it will only be modified
1083 	 * while holding zilog->zl_lock) as its writes and those of its
1084 	 * children have all completed.  The younger 'nlwb' may be waiting on
1085 	 * future writes to additional vdevs.
1086 	 */
1087 	mutex_enter(&nlwb->lwb_vdev_lock);
1088 	/*
1089 	 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1090 	 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1091 	 */
1092 	while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1093 		avl_index_t where;
1094 
1095 		if (avl_find(dst, zv, &where) == NULL) {
1096 			avl_insert(dst, zv, where);
1097 		} else {
1098 			kmem_free(zv, sizeof (*zv));
1099 		}
1100 	}
1101 	mutex_exit(&nlwb->lwb_vdev_lock);
1102 }
1103 
1104 void
1105 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1106 {
1107 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1108 }
1109 
1110 /*
1111  * This function is a called after all vdevs associated with a given lwb
1112  * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1113  * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1114  * all "previous" lwb's will have completed before this function is
1115  * called; i.e. this function is called for all previous lwbs before
1116  * it's called for "this" lwb (enforced via zio the dependencies
1117  * configured in zil_lwb_set_zio_dependency()).
1118  *
1119  * The intention is for this function to be called as soon as the
1120  * contents of an lwb are considered "stable" on disk, and will survive
1121  * any sudden loss of power. At this point, any threads waiting for the
1122  * lwb to reach this state are signalled, and the "waiter" structures
1123  * are marked "done".
1124  */
1125 static void
1126 zil_lwb_flush_vdevs_done(zio_t *zio)
1127 {
1128 	lwb_t *lwb = zio->io_private;
1129 	zilog_t *zilog = lwb->lwb_zilog;
1130 	dmu_tx_t *tx = lwb->lwb_tx;
1131 	zil_commit_waiter_t *zcw;
1132 	itx_t *itx;
1133 
1134 	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1135 
1136 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1137 
1138 	mutex_enter(&zilog->zl_lock);
1139 
1140 	/*
1141 	 * Ensure the lwb buffer pointer is cleared before releasing the
1142 	 * txg. If we have had an allocation failure and the txg is
1143 	 * waiting to sync then we want zil_sync() to remove the lwb so
1144 	 * that it's not picked up as the next new one in
1145 	 * zil_process_commit_list(). zil_sync() will only remove the
1146 	 * lwb if lwb_buf is null.
1147 	 */
1148 	lwb->lwb_buf = NULL;
1149 	lwb->lwb_tx = NULL;
1150 
1151 	ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1152 	zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1153 
1154 	lwb->lwb_root_zio = NULL;
1155 
1156 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1157 	lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1158 
1159 	if (zilog->zl_last_lwb_opened == lwb) {
1160 		/*
1161 		 * Remember the highest committed log sequence number
1162 		 * for ztest. We only update this value when all the log
1163 		 * writes succeeded, because ztest wants to ASSERT that
1164 		 * it got the whole log chain.
1165 		 */
1166 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1167 	}
1168 
1169 	while ((itx = list_head(&lwb->lwb_itxs)) != NULL) {
1170 		list_remove(&lwb->lwb_itxs, itx);
1171 		zil_itx_destroy(itx);
1172 	}
1173 
1174 	while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1175 		mutex_enter(&zcw->zcw_lock);
1176 
1177 		ASSERT(list_link_active(&zcw->zcw_node));
1178 		list_remove(&lwb->lwb_waiters, zcw);
1179 
1180 		ASSERT3P(zcw->zcw_lwb, ==, lwb);
1181 		zcw->zcw_lwb = NULL;
1182 		/*
1183 		 * We expect any ZIO errors from child ZIOs to have been
1184 		 * propagated "up" to this specific LWB's root ZIO, in
1185 		 * order for this error handling to work correctly. This
1186 		 * includes ZIO errors from either this LWB's write or
1187 		 * flush, as well as any errors from other dependent LWBs
1188 		 * (e.g. a root LWB ZIO that might be a child of this LWB).
1189 		 *
1190 		 * With that said, it's important to note that LWB flush
1191 		 * errors are not propagated up to the LWB root ZIO.
1192 		 * This is incorrect behavior, and results in VDEV flush
1193 		 * errors not being handled correctly here. See the
1194 		 * comment above the call to "zio_flush" for details.
1195 		 */
1196 
1197 		zcw->zcw_zio_error = zio->io_error;
1198 
1199 		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1200 		zcw->zcw_done = B_TRUE;
1201 		cv_broadcast(&zcw->zcw_cv);
1202 
1203 		mutex_exit(&zcw->zcw_lock);
1204 	}
1205 
1206 	mutex_exit(&zilog->zl_lock);
1207 
1208 	/*
1209 	 * Now that we've written this log block, we have a stable pointer
1210 	 * to the next block in the chain, so it's OK to let the txg in
1211 	 * which we allocated the next block sync.
1212 	 */
1213 	dmu_tx_commit(tx);
1214 }
1215 
1216 /*
1217  * This is called when an lwb's write zio completes. The callback's
1218  * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1219  * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1220  * in writing out this specific lwb's data, and in the case that cache
1221  * flushes have been deferred, vdevs involved in writing the data for
1222  * previous lwbs. The writes corresponding to all the vdevs in the
1223  * lwb_vdev_tree will have completed by the time this is called, due to
1224  * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1225  * which takes deferred flushes into account. The lwb will be "done"
1226  * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1227  * completion callback for the lwb's root zio.
1228  */
1229 static void
1230 zil_lwb_write_done(zio_t *zio)
1231 {
1232 	lwb_t *lwb = zio->io_private;
1233 	spa_t *spa = zio->io_spa;
1234 	zilog_t *zilog = lwb->lwb_zilog;
1235 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1236 	void *cookie = NULL;
1237 	zil_vdev_node_t *zv;
1238 	lwb_t *nlwb;
1239 
1240 	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1241 
1242 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1243 	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1244 	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1245 	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1246 	ASSERT(!BP_IS_GANG(zio->io_bp));
1247 	ASSERT(!BP_IS_HOLE(zio->io_bp));
1248 	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1249 
1250 	abd_free(zio->io_abd);
1251 
1252 	mutex_enter(&zilog->zl_lock);
1253 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1254 	lwb->lwb_state = LWB_STATE_WRITE_DONE;
1255 	lwb->lwb_write_zio = NULL;
1256 	lwb->lwb_fastwrite = FALSE;
1257 	nlwb = list_next(&zilog->zl_lwb_list, lwb);
1258 	mutex_exit(&zilog->zl_lock);
1259 
1260 	if (avl_numnodes(t) == 0)
1261 		return;
1262 
1263 	/*
1264 	 * If there was an IO error, we're not going to call zio_flush()
1265 	 * on these vdevs, so we simply empty the tree and free the
1266 	 * nodes. We avoid calling zio_flush() since there isn't any
1267 	 * good reason for doing so, after the lwb block failed to be
1268 	 * written out.
1269 	 *
1270 	 * Additionally, we don't perform any further error handling at
1271 	 * this point (e.g. setting "zcw_zio_error" appropriately), as
1272 	 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1273 	 * we expect any error seen here, to have been propagated to
1274 	 * that function).
1275 	 */
1276 	if (zio->io_error != 0) {
1277 		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1278 			kmem_free(zv, sizeof (*zv));
1279 		return;
1280 	}
1281 
1282 	/*
1283 	 * If this lwb does not have any threads waiting for it to
1284 	 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1285 	 * command to the vdevs written to by "this" lwb, and instead
1286 	 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1287 	 * command for those vdevs. Thus, we merge the vdev tree of
1288 	 * "this" lwb with the vdev tree of the "next" lwb in the list,
1289 	 * and assume the "next" lwb will handle flushing the vdevs (or
1290 	 * deferring the flush(s) again).
1291 	 *
1292 	 * This is a useful performance optimization, especially for
1293 	 * workloads with lots of async write activity and few sync
1294 	 * write and/or fsync activity, as it has the potential to
1295 	 * coalesce multiple flush commands to a vdev into one.
1296 	 */
1297 	if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) {
1298 		zil_lwb_flush_defer(lwb, nlwb);
1299 		ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1300 		return;
1301 	}
1302 
1303 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1304 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1305 		if (vd != NULL) {
1306 			/*
1307 			 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1308 			 * always used within "zio_flush". This means,
1309 			 * any errors when flushing the vdev(s), will
1310 			 * (unfortunately) not be handled correctly,
1311 			 * since these "zio_flush" errors will not be
1312 			 * propagated up to "zil_lwb_flush_vdevs_done".
1313 			 */
1314 			zio_flush(lwb->lwb_root_zio, vd);
1315 		}
1316 		kmem_free(zv, sizeof (*zv));
1317 	}
1318 }
1319 
1320 static void
1321 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1322 {
1323 	lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1324 
1325 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1326 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
1327 
1328 	/*
1329 	 * The zilog's "zl_last_lwb_opened" field is used to build the
1330 	 * lwb/zio dependency chain, which is used to preserve the
1331 	 * ordering of lwb completions that is required by the semantics
1332 	 * of the ZIL. Each new lwb zio becomes a parent of the
1333 	 * "previous" lwb zio, such that the new lwb's zio cannot
1334 	 * complete until the "previous" lwb's zio completes.
1335 	 *
1336 	 * This is required by the semantics of zil_commit(); the commit
1337 	 * waiters attached to the lwbs will be woken in the lwb zio's
1338 	 * completion callback, so this zio dependency graph ensures the
1339 	 * waiters are woken in the correct order (the same order the
1340 	 * lwbs were created).
1341 	 */
1342 	if (last_lwb_opened != NULL &&
1343 	    last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) {
1344 		ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1345 		    last_lwb_opened->lwb_state == LWB_STATE_ISSUED ||
1346 		    last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE);
1347 
1348 		ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1349 		zio_add_child(lwb->lwb_root_zio,
1350 		    last_lwb_opened->lwb_root_zio);
1351 
1352 		/*
1353 		 * If the previous lwb's write hasn't already completed,
1354 		 * we also want to order the completion of the lwb write
1355 		 * zios (above, we only order the completion of the lwb
1356 		 * root zios). This is required because of how we can
1357 		 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1358 		 *
1359 		 * When the DKIOCFLUSHWRITECACHE commands are deferred,
1360 		 * the previous lwb will rely on this lwb to flush the
1361 		 * vdevs written to by that previous lwb. Thus, we need
1362 		 * to ensure this lwb doesn't issue the flush until
1363 		 * after the previous lwb's write completes. We ensure
1364 		 * this ordering by setting the zio parent/child
1365 		 * relationship here.
1366 		 *
1367 		 * Without this relationship on the lwb's write zio,
1368 		 * it's possible for this lwb's write to complete prior
1369 		 * to the previous lwb's write completing; and thus, the
1370 		 * vdevs for the previous lwb would be flushed prior to
1371 		 * that lwb's data being written to those vdevs (the
1372 		 * vdevs are flushed in the lwb write zio's completion
1373 		 * handler, zil_lwb_write_done()).
1374 		 */
1375 		if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) {
1376 			ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1377 			    last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1378 
1379 			ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL);
1380 			zio_add_child(lwb->lwb_write_zio,
1381 			    last_lwb_opened->lwb_write_zio);
1382 		}
1383 	}
1384 }
1385 
1386 
1387 /*
1388  * This function's purpose is to "open" an lwb such that it is ready to
1389  * accept new itxs being committed to it. To do this, the lwb's zio
1390  * structures are created, and linked to the lwb. This function is
1391  * idempotent; if the passed in lwb has already been opened, this
1392  * function is essentially a no-op.
1393  */
1394 static void
1395 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1396 {
1397 	zbookmark_phys_t zb;
1398 	zio_priority_t prio;
1399 
1400 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1401 	ASSERT3P(lwb, !=, NULL);
1402 	EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1403 	EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1404 
1405 	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1406 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1407 	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1408 
1409 	/* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1410 	mutex_enter(&zilog->zl_lock);
1411 	if (lwb->lwb_root_zio == NULL) {
1412 		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1413 		    BP_GET_LSIZE(&lwb->lwb_blk));
1414 
1415 		if (!lwb->lwb_fastwrite) {
1416 			metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
1417 			lwb->lwb_fastwrite = 1;
1418 		}
1419 
1420 		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1421 			prio = ZIO_PRIORITY_SYNC_WRITE;
1422 		else
1423 			prio = ZIO_PRIORITY_ASYNC_WRITE;
1424 
1425 		lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1426 		    zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1427 		ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1428 
1429 		lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1430 		    zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1431 		    BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1432 		    prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb);
1433 		ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1434 
1435 		lwb->lwb_state = LWB_STATE_OPENED;
1436 
1437 		zil_lwb_set_zio_dependency(zilog, lwb);
1438 		zilog->zl_last_lwb_opened = lwb;
1439 	}
1440 	mutex_exit(&zilog->zl_lock);
1441 
1442 	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1443 	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1444 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1445 }
1446 
1447 /*
1448  * Define a limited set of intent log block sizes.
1449  *
1450  * These must be a multiple of 4KB. Note only the amount used (again
1451  * aligned to 4KB) actually gets written. However, we can't always just
1452  * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1453  */
1454 static const struct {
1455 	uint64_t	limit;
1456 	uint64_t	blksz;
1457 } zil_block_buckets[] = {
1458 	{ 4096,		4096 },			/* non TX_WRITE */
1459 	{ 8192 + 4096,	8192 + 4096 },		/* database */
1460 	{ 32768 + 4096,	32768 + 4096 },		/* NFS writes */
1461 	{ 65536 + 4096,	65536 + 4096 },		/* 64KB writes */
1462 	{ 131072,	131072 },		/* < 128KB writes */
1463 	{ 131072 +4096,	65536 + 4096 },		/* 128KB writes */
1464 	{ UINT64_MAX,	SPA_OLD_MAXBLOCKSIZE},	/* > 128KB writes */
1465 };
1466 
1467 /*
1468  * Maximum block size used by the ZIL.  This is picked up when the ZIL is
1469  * initialized.  Otherwise this should not be used directly; see
1470  * zl_max_block_size instead.
1471  */
1472 static int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1473 
1474 /*
1475  * Start a log block write and advance to the next log block.
1476  * Calls are serialized.
1477  */
1478 static lwb_t *
1479 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1480 {
1481 	lwb_t *nlwb = NULL;
1482 	zil_chain_t *zilc;
1483 	spa_t *spa = zilog->zl_spa;
1484 	blkptr_t *bp;
1485 	dmu_tx_t *tx;
1486 	uint64_t txg;
1487 	uint64_t zil_blksz, wsz;
1488 	int i, error;
1489 	boolean_t slog;
1490 
1491 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1492 	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1493 	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1494 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1495 
1496 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1497 		zilc = (zil_chain_t *)lwb->lwb_buf;
1498 		bp = &zilc->zc_next_blk;
1499 	} else {
1500 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1501 		bp = &zilc->zc_next_blk;
1502 	}
1503 
1504 	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1505 
1506 	/*
1507 	 * Allocate the next block and save its address in this block
1508 	 * before writing it in order to establish the log chain.
1509 	 * Note that if the allocation of nlwb synced before we wrote
1510 	 * the block that points at it (lwb), we'd leak it if we crashed.
1511 	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1512 	 * We dirty the dataset to ensure that zil_sync() will be called
1513 	 * to clean up in the event of allocation failure or I/O failure.
1514 	 */
1515 
1516 	tx = dmu_tx_create(zilog->zl_os);
1517 
1518 	/*
1519 	 * Since we are not going to create any new dirty data, and we
1520 	 * can even help with clearing the existing dirty data, we
1521 	 * should not be subject to the dirty data based delays. We
1522 	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1523 	 */
1524 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1525 
1526 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1527 	txg = dmu_tx_get_txg(tx);
1528 
1529 	lwb->lwb_tx = tx;
1530 
1531 	/*
1532 	 * Log blocks are pre-allocated. Here we select the size of the next
1533 	 * block, based on size used in the last block.
1534 	 * - first find the smallest bucket that will fit the block from a
1535 	 *   limited set of block sizes. This is because it's faster to write
1536 	 *   blocks allocated from the same metaslab as they are adjacent or
1537 	 *   close.
1538 	 * - next find the maximum from the new suggested size and an array of
1539 	 *   previous sizes. This lessens a picket fence effect of wrongly
1540 	 *   guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1541 	 *   requests.
1542 	 *
1543 	 * Note we only write what is used, but we can't just allocate
1544 	 * the maximum block size because we can exhaust the available
1545 	 * pool log space.
1546 	 */
1547 	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1548 	for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++)
1549 		continue;
1550 	zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size);
1551 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1552 	for (i = 0; i < ZIL_PREV_BLKS; i++)
1553 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1554 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1555 
1556 	BP_ZERO(bp);
1557 	error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog);
1558 	if (slog) {
1559 		ZIL_STAT_BUMP(zil_itx_metaslab_slog_count);
1560 		ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused);
1561 	} else {
1562 		ZIL_STAT_BUMP(zil_itx_metaslab_normal_count);
1563 		ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused);
1564 	}
1565 	if (error == 0) {
1566 		ASSERT3U(bp->blk_birth, ==, txg);
1567 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1568 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1569 
1570 		/*
1571 		 * Allocate a new log write block (lwb).
1572 		 */
1573 		nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
1574 	}
1575 
1576 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1577 		/* For Slim ZIL only write what is used. */
1578 		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1579 		ASSERT3U(wsz, <=, lwb->lwb_sz);
1580 		zio_shrink(lwb->lwb_write_zio, wsz);
1581 
1582 	} else {
1583 		wsz = lwb->lwb_sz;
1584 	}
1585 
1586 	zilc->zc_pad = 0;
1587 	zilc->zc_nused = lwb->lwb_nused;
1588 	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1589 
1590 	/*
1591 	 * clear unused data for security
1592 	 */
1593 	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1594 
1595 	spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1596 
1597 	zil_lwb_add_block(lwb, &lwb->lwb_blk);
1598 	lwb->lwb_issued_timestamp = gethrtime();
1599 	lwb->lwb_state = LWB_STATE_ISSUED;
1600 
1601 	zio_nowait(lwb->lwb_root_zio);
1602 	zio_nowait(lwb->lwb_write_zio);
1603 
1604 	/*
1605 	 * If there was an allocation failure then nlwb will be null which
1606 	 * forces a txg_wait_synced().
1607 	 */
1608 	return (nlwb);
1609 }
1610 
1611 /*
1612  * Maximum amount of write data that can be put into single log block.
1613  */
1614 uint64_t
1615 zil_max_log_data(zilog_t *zilog)
1616 {
1617 	return (zilog->zl_max_block_size -
1618 	    sizeof (zil_chain_t) - sizeof (lr_write_t));
1619 }
1620 
1621 /*
1622  * Maximum amount of log space we agree to waste to reduce number of
1623  * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1624  */
1625 static inline uint64_t
1626 zil_max_waste_space(zilog_t *zilog)
1627 {
1628 	return (zil_max_log_data(zilog) / 8);
1629 }
1630 
1631 /*
1632  * Maximum amount of write data for WR_COPIED.  For correctness, consumers
1633  * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1634  * maximum sized log block, because each WR_COPIED record must fit in a
1635  * single log block.  For space efficiency, we want to fit two records into a
1636  * max-sized log block.
1637  */
1638 uint64_t
1639 zil_max_copied_data(zilog_t *zilog)
1640 {
1641 	return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 -
1642 	    sizeof (lr_write_t));
1643 }
1644 
1645 static lwb_t *
1646 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1647 {
1648 	lr_t *lrcb, *lrc;
1649 	lr_write_t *lrwb, *lrw;
1650 	char *lr_buf;
1651 	uint64_t dlen, dnow, dpad, lwb_sp, reclen, txg, max_log_data;
1652 
1653 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1654 	ASSERT3P(lwb, !=, NULL);
1655 	ASSERT3P(lwb->lwb_buf, !=, NULL);
1656 
1657 	zil_lwb_write_open(zilog, lwb);
1658 
1659 	lrc = &itx->itx_lr;
1660 	lrw = (lr_write_t *)lrc;
1661 
1662 	/*
1663 	 * A commit itx doesn't represent any on-disk state; instead
1664 	 * it's simply used as a place holder on the commit list, and
1665 	 * provides a mechanism for attaching a "commit waiter" onto the
1666 	 * correct lwb (such that the waiter can be signalled upon
1667 	 * completion of that lwb). Thus, we don't process this itx's
1668 	 * log record if it's a commit itx (these itx's don't have log
1669 	 * records), and instead link the itx's waiter onto the lwb's
1670 	 * list of waiters.
1671 	 *
1672 	 * For more details, see the comment above zil_commit().
1673 	 */
1674 	if (lrc->lrc_txtype == TX_COMMIT) {
1675 		mutex_enter(&zilog->zl_lock);
1676 		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1677 		itx->itx_private = NULL;
1678 		mutex_exit(&zilog->zl_lock);
1679 		return (lwb);
1680 	}
1681 
1682 	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1683 		dlen = P2ROUNDUP_TYPED(
1684 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1685 		dpad = dlen - lrw->lr_length;
1686 	} else {
1687 		dlen = dpad = 0;
1688 	}
1689 	reclen = lrc->lrc_reclen;
1690 	zilog->zl_cur_used += (reclen + dlen);
1691 	txg = lrc->lrc_txg;
1692 
1693 	ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1694 
1695 cont:
1696 	/*
1697 	 * If this record won't fit in the current log block, start a new one.
1698 	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1699 	 */
1700 	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1701 	max_log_data = zil_max_log_data(zilog);
1702 	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1703 	    lwb_sp < zil_max_waste_space(zilog) &&
1704 	    (dlen % max_log_data == 0 ||
1705 	    lwb_sp < reclen + dlen % max_log_data))) {
1706 		lwb = zil_lwb_write_issue(zilog, lwb);
1707 		if (lwb == NULL)
1708 			return (NULL);
1709 		zil_lwb_write_open(zilog, lwb);
1710 		ASSERT(LWB_EMPTY(lwb));
1711 		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1712 
1713 		/*
1714 		 * There must be enough space in the new, empty log block to
1715 		 * hold reclen.  For WR_COPIED, we need to fit the whole
1716 		 * record in one block, and reclen is the header size + the
1717 		 * data size. For WR_NEED_COPY, we can create multiple
1718 		 * records, splitting the data into multiple blocks, so we
1719 		 * only need to fit one word of data per block; in this case
1720 		 * reclen is just the header size (no data).
1721 		 */
1722 		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1723 	}
1724 
1725 	dnow = MIN(dlen, lwb_sp - reclen);
1726 	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1727 	bcopy(lrc, lr_buf, reclen);
1728 	lrcb = (lr_t *)lr_buf;		/* Like lrc, but inside lwb. */
1729 	lrwb = (lr_write_t *)lrcb;	/* Like lrw, but inside lwb. */
1730 
1731 	ZIL_STAT_BUMP(zil_itx_count);
1732 
1733 	/*
1734 	 * If it's a write, fetch the data or get its blkptr as appropriate.
1735 	 */
1736 	if (lrc->lrc_txtype == TX_WRITE) {
1737 		if (txg > spa_freeze_txg(zilog->zl_spa))
1738 			txg_wait_synced(zilog->zl_dmu_pool, txg);
1739 		if (itx->itx_wr_state == WR_COPIED) {
1740 			ZIL_STAT_BUMP(zil_itx_copied_count);
1741 			ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length);
1742 		} else {
1743 			char *dbuf;
1744 			int error;
1745 
1746 			if (itx->itx_wr_state == WR_NEED_COPY) {
1747 				dbuf = lr_buf + reclen;
1748 				lrcb->lrc_reclen += dnow;
1749 				if (lrwb->lr_length > dnow)
1750 					lrwb->lr_length = dnow;
1751 				lrw->lr_offset += dnow;
1752 				lrw->lr_length -= dnow;
1753 				ZIL_STAT_BUMP(zil_itx_needcopy_count);
1754 				ZIL_STAT_INCR(zil_itx_needcopy_bytes, dnow);
1755 			} else {
1756 				ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
1757 				dbuf = NULL;
1758 				ZIL_STAT_BUMP(zil_itx_indirect_count);
1759 				ZIL_STAT_INCR(zil_itx_indirect_bytes,
1760 				    lrw->lr_length);
1761 			}
1762 
1763 			/*
1764 			 * We pass in the "lwb_write_zio" rather than
1765 			 * "lwb_root_zio" so that the "lwb_write_zio"
1766 			 * becomes the parent of any zio's created by
1767 			 * the "zl_get_data" callback. The vdevs are
1768 			 * flushed after the "lwb_write_zio" completes,
1769 			 * so we want to make sure that completion
1770 			 * callback waits for these additional zio's,
1771 			 * such that the vdevs used by those zio's will
1772 			 * be included in the lwb's vdev tree, and those
1773 			 * vdevs will be properly flushed. If we passed
1774 			 * in "lwb_root_zio" here, then these additional
1775 			 * vdevs may not be flushed; e.g. if these zio's
1776 			 * completed after "lwb_write_zio" completed.
1777 			 */
1778 			error = zilog->zl_get_data(itx->itx_private,
1779 			    itx->itx_gen, lrwb, dbuf, lwb,
1780 			    lwb->lwb_write_zio);
1781 			if (dbuf != NULL && error == 0 && dnow == dlen)
1782 				/* Zero any padding bytes in the last block. */
1783 				bzero((char *)dbuf + lrwb->lr_length, dpad);
1784 
1785 			if (error == EIO) {
1786 				txg_wait_synced(zilog->zl_dmu_pool, txg);
1787 				return (lwb);
1788 			}
1789 			if (error != 0) {
1790 				ASSERT(error == ENOENT || error == EEXIST ||
1791 				    error == EALREADY);
1792 				return (lwb);
1793 			}
1794 		}
1795 	}
1796 
1797 	/*
1798 	 * We're actually making an entry, so update lrc_seq to be the
1799 	 * log record sequence number.  Note that this is generally not
1800 	 * equal to the itx sequence number because not all transactions
1801 	 * are synchronous, and sometimes spa_sync() gets there first.
1802 	 */
1803 	lrcb->lrc_seq = ++zilog->zl_lr_seq;
1804 	lwb->lwb_nused += reclen + dnow;
1805 
1806 	zil_lwb_add_txg(lwb, txg);
1807 
1808 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1809 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1810 
1811 	dlen -= dnow;
1812 	if (dlen > 0) {
1813 		zilog->zl_cur_used += reclen;
1814 		goto cont;
1815 	}
1816 
1817 	return (lwb);
1818 }
1819 
1820 itx_t *
1821 zil_itx_create(uint64_t txtype, size_t olrsize)
1822 {
1823 	size_t itxsize, lrsize;
1824 	itx_t *itx;
1825 
1826 	lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t);
1827 	itxsize = offsetof(itx_t, itx_lr) + lrsize;
1828 
1829 	itx = zio_data_buf_alloc(itxsize);
1830 	itx->itx_lr.lrc_txtype = txtype;
1831 	itx->itx_lr.lrc_reclen = lrsize;
1832 	itx->itx_lr.lrc_seq = 0;	/* defensive */
1833 	bzero((char *)&itx->itx_lr + olrsize, lrsize - olrsize);
1834 	itx->itx_sync = B_TRUE;		/* default is synchronous */
1835 	itx->itx_callback = NULL;
1836 	itx->itx_callback_data = NULL;
1837 	itx->itx_size = itxsize;
1838 
1839 	return (itx);
1840 }
1841 
1842 void
1843 zil_itx_destroy(itx_t *itx)
1844 {
1845 	IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
1846 	IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
1847 
1848 	if (itx->itx_callback != NULL)
1849 		itx->itx_callback(itx->itx_callback_data);
1850 
1851 	zio_data_buf_free(itx, itx->itx_size);
1852 }
1853 
1854 /*
1855  * Free up the sync and async itxs. The itxs_t has already been detached
1856  * so no locks are needed.
1857  */
1858 static void
1859 zil_itxg_clean(void *arg)
1860 {
1861 	itx_t *itx;
1862 	list_t *list;
1863 	avl_tree_t *t;
1864 	void *cookie;
1865 	itxs_t *itxs = arg;
1866 	itx_async_node_t *ian;
1867 
1868 	list = &itxs->i_sync_list;
1869 	while ((itx = list_head(list)) != NULL) {
1870 		/*
1871 		 * In the general case, commit itxs will not be found
1872 		 * here, as they'll be committed to an lwb via
1873 		 * zil_lwb_commit(), and free'd in that function. Having
1874 		 * said that, it is still possible for commit itxs to be
1875 		 * found here, due to the following race:
1876 		 *
1877 		 *	- a thread calls zil_commit() which assigns the
1878 		 *	  commit itx to a per-txg i_sync_list
1879 		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
1880 		 *	  while the waiter is still on the i_sync_list
1881 		 *
1882 		 * There's nothing to prevent syncing the txg while the
1883 		 * waiter is on the i_sync_list. This normally doesn't
1884 		 * happen because spa_sync() is slower than zil_commit(),
1885 		 * but if zil_commit() calls txg_wait_synced() (e.g.
1886 		 * because zil_create() or zil_commit_writer_stall() is
1887 		 * called) we will hit this case.
1888 		 */
1889 		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1890 			zil_commit_waiter_skip(itx->itx_private);
1891 
1892 		list_remove(list, itx);
1893 		zil_itx_destroy(itx);
1894 	}
1895 
1896 	cookie = NULL;
1897 	t = &itxs->i_async_tree;
1898 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1899 		list = &ian->ia_list;
1900 		while ((itx = list_head(list)) != NULL) {
1901 			list_remove(list, itx);
1902 			/* commit itxs should never be on the async lists. */
1903 			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1904 			zil_itx_destroy(itx);
1905 		}
1906 		list_destroy(list);
1907 		kmem_free(ian, sizeof (itx_async_node_t));
1908 	}
1909 	avl_destroy(t);
1910 
1911 	kmem_free(itxs, sizeof (itxs_t));
1912 }
1913 
1914 static int
1915 zil_aitx_compare(const void *x1, const void *x2)
1916 {
1917 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1918 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1919 
1920 	return (TREE_CMP(o1, o2));
1921 }
1922 
1923 /*
1924  * Remove all async itx with the given oid.
1925  */
1926 void
1927 zil_remove_async(zilog_t *zilog, uint64_t oid)
1928 {
1929 	uint64_t otxg, txg;
1930 	itx_async_node_t *ian;
1931 	avl_tree_t *t;
1932 	avl_index_t where;
1933 	list_t clean_list;
1934 	itx_t *itx;
1935 
1936 	ASSERT(oid != 0);
1937 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1938 
1939 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1940 		otxg = ZILTEST_TXG;
1941 	else
1942 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1943 
1944 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1945 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1946 
1947 		mutex_enter(&itxg->itxg_lock);
1948 		if (itxg->itxg_txg != txg) {
1949 			mutex_exit(&itxg->itxg_lock);
1950 			continue;
1951 		}
1952 
1953 		/*
1954 		 * Locate the object node and append its list.
1955 		 */
1956 		t = &itxg->itxg_itxs->i_async_tree;
1957 		ian = avl_find(t, &oid, &where);
1958 		if (ian != NULL)
1959 			list_move_tail(&clean_list, &ian->ia_list);
1960 		mutex_exit(&itxg->itxg_lock);
1961 	}
1962 	while ((itx = list_head(&clean_list)) != NULL) {
1963 		list_remove(&clean_list, itx);
1964 		/* commit itxs should never be on the async lists. */
1965 		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1966 		zil_itx_destroy(itx);
1967 	}
1968 	list_destroy(&clean_list);
1969 }
1970 
1971 void
1972 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1973 {
1974 	uint64_t txg;
1975 	itxg_t *itxg;
1976 	itxs_t *itxs, *clean = NULL;
1977 
1978 	/*
1979 	 * Ensure the data of a renamed file is committed before the rename.
1980 	 */
1981 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1982 		zil_async_to_sync(zilog, itx->itx_oid);
1983 
1984 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1985 		txg = ZILTEST_TXG;
1986 	else
1987 		txg = dmu_tx_get_txg(tx);
1988 
1989 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1990 	mutex_enter(&itxg->itxg_lock);
1991 	itxs = itxg->itxg_itxs;
1992 	if (itxg->itxg_txg != txg) {
1993 		if (itxs != NULL) {
1994 			/*
1995 			 * The zil_clean callback hasn't got around to cleaning
1996 			 * this itxg. Save the itxs for release below.
1997 			 * This should be rare.
1998 			 */
1999 			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2000 			    "txg %llu", (u_longlong_t)itxg->itxg_txg);
2001 			clean = itxg->itxg_itxs;
2002 		}
2003 		itxg->itxg_txg = txg;
2004 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
2005 		    KM_SLEEP);
2006 
2007 		list_create(&itxs->i_sync_list, sizeof (itx_t),
2008 		    offsetof(itx_t, itx_node));
2009 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
2010 		    sizeof (itx_async_node_t),
2011 		    offsetof(itx_async_node_t, ia_node));
2012 	}
2013 	if (itx->itx_sync) {
2014 		list_insert_tail(&itxs->i_sync_list, itx);
2015 	} else {
2016 		avl_tree_t *t = &itxs->i_async_tree;
2017 		uint64_t foid =
2018 		    LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
2019 		itx_async_node_t *ian;
2020 		avl_index_t where;
2021 
2022 		ian = avl_find(t, &foid, &where);
2023 		if (ian == NULL) {
2024 			ian = kmem_alloc(sizeof (itx_async_node_t),
2025 			    KM_SLEEP);
2026 			list_create(&ian->ia_list, sizeof (itx_t),
2027 			    offsetof(itx_t, itx_node));
2028 			ian->ia_foid = foid;
2029 			avl_insert(t, ian, where);
2030 		}
2031 		list_insert_tail(&ian->ia_list, itx);
2032 	}
2033 
2034 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
2035 
2036 	/*
2037 	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2038 	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2039 	 * need to be careful to always dirty the ZIL using the "real"
2040 	 * TXG (not itxg_txg) even when the SPA is frozen.
2041 	 */
2042 	zilog_dirty(zilog, dmu_tx_get_txg(tx));
2043 	mutex_exit(&itxg->itxg_lock);
2044 
2045 	/* Release the old itxs now we've dropped the lock */
2046 	if (clean != NULL)
2047 		zil_itxg_clean(clean);
2048 }
2049 
2050 /*
2051  * If there are any in-memory intent log transactions which have now been
2052  * synced then start up a taskq to free them. We should only do this after we
2053  * have written out the uberblocks (i.e. txg has been committed) so that
2054  * don't inadvertently clean out in-memory log records that would be required
2055  * by zil_commit().
2056  */
2057 void
2058 zil_clean(zilog_t *zilog, uint64_t synced_txg)
2059 {
2060 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2061 	itxs_t *clean_me;
2062 
2063 	ASSERT3U(synced_txg, <, ZILTEST_TXG);
2064 
2065 	mutex_enter(&itxg->itxg_lock);
2066 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2067 		mutex_exit(&itxg->itxg_lock);
2068 		return;
2069 	}
2070 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2071 	ASSERT3U(itxg->itxg_txg, !=, 0);
2072 	clean_me = itxg->itxg_itxs;
2073 	itxg->itxg_itxs = NULL;
2074 	itxg->itxg_txg = 0;
2075 	mutex_exit(&itxg->itxg_lock);
2076 	/*
2077 	 * Preferably start a task queue to free up the old itxs but
2078 	 * if taskq_dispatch can't allocate resources to do that then
2079 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
2080 	 * created a bad performance problem.
2081 	 */
2082 	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2083 	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2084 	taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
2085 	    zil_itxg_clean, clean_me, TQ_NOSLEEP);
2086 	if (id == TASKQID_INVALID)
2087 		zil_itxg_clean(clean_me);
2088 }
2089 
2090 /*
2091  * This function will traverse the queue of itxs that need to be
2092  * committed, and move them onto the ZIL's zl_itx_commit_list.
2093  */
2094 static void
2095 zil_get_commit_list(zilog_t *zilog)
2096 {
2097 	uint64_t otxg, txg;
2098 	list_t *commit_list = &zilog->zl_itx_commit_list;
2099 
2100 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2101 
2102 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2103 		otxg = ZILTEST_TXG;
2104 	else
2105 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2106 
2107 	/*
2108 	 * This is inherently racy, since there is nothing to prevent
2109 	 * the last synced txg from changing. That's okay since we'll
2110 	 * only commit things in the future.
2111 	 */
2112 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2113 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2114 
2115 		mutex_enter(&itxg->itxg_lock);
2116 		if (itxg->itxg_txg != txg) {
2117 			mutex_exit(&itxg->itxg_lock);
2118 			continue;
2119 		}
2120 
2121 		/*
2122 		 * If we're adding itx records to the zl_itx_commit_list,
2123 		 * then the zil better be dirty in this "txg". We can assert
2124 		 * that here since we're holding the itxg_lock which will
2125 		 * prevent spa_sync from cleaning it. Once we add the itxs
2126 		 * to the zl_itx_commit_list we must commit it to disk even
2127 		 * if it's unnecessary (i.e. the txg was synced).
2128 		 */
2129 		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2130 		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2131 		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
2132 
2133 		mutex_exit(&itxg->itxg_lock);
2134 	}
2135 }
2136 
2137 /*
2138  * Move the async itxs for a specified object to commit into sync lists.
2139  */
2140 void
2141 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2142 {
2143 	uint64_t otxg, txg;
2144 	itx_async_node_t *ian;
2145 	avl_tree_t *t;
2146 	avl_index_t where;
2147 
2148 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2149 		otxg = ZILTEST_TXG;
2150 	else
2151 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2152 
2153 	/*
2154 	 * This is inherently racy, since there is nothing to prevent
2155 	 * the last synced txg from changing.
2156 	 */
2157 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2158 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2159 
2160 		mutex_enter(&itxg->itxg_lock);
2161 		if (itxg->itxg_txg != txg) {
2162 			mutex_exit(&itxg->itxg_lock);
2163 			continue;
2164 		}
2165 
2166 		/*
2167 		 * If a foid is specified then find that node and append its
2168 		 * list. Otherwise walk the tree appending all the lists
2169 		 * to the sync list. We add to the end rather than the
2170 		 * beginning to ensure the create has happened.
2171 		 */
2172 		t = &itxg->itxg_itxs->i_async_tree;
2173 		if (foid != 0) {
2174 			ian = avl_find(t, &foid, &where);
2175 			if (ian != NULL) {
2176 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2177 				    &ian->ia_list);
2178 			}
2179 		} else {
2180 			void *cookie = NULL;
2181 
2182 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2183 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2184 				    &ian->ia_list);
2185 				list_destroy(&ian->ia_list);
2186 				kmem_free(ian, sizeof (itx_async_node_t));
2187 			}
2188 		}
2189 		mutex_exit(&itxg->itxg_lock);
2190 	}
2191 }
2192 
2193 /*
2194  * This function will prune commit itxs that are at the head of the
2195  * commit list (it won't prune past the first non-commit itx), and
2196  * either: a) attach them to the last lwb that's still pending
2197  * completion, or b) skip them altogether.
2198  *
2199  * This is used as a performance optimization to prevent commit itxs
2200  * from generating new lwbs when it's unnecessary to do so.
2201  */
2202 static void
2203 zil_prune_commit_list(zilog_t *zilog)
2204 {
2205 	itx_t *itx;
2206 
2207 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2208 
2209 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2210 		lr_t *lrc = &itx->itx_lr;
2211 		if (lrc->lrc_txtype != TX_COMMIT)
2212 			break;
2213 
2214 		mutex_enter(&zilog->zl_lock);
2215 
2216 		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2217 		if (last_lwb == NULL ||
2218 		    last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2219 			/*
2220 			 * All of the itxs this waiter was waiting on
2221 			 * must have already completed (or there were
2222 			 * never any itx's for it to wait on), so it's
2223 			 * safe to skip this waiter and mark it done.
2224 			 */
2225 			zil_commit_waiter_skip(itx->itx_private);
2226 		} else {
2227 			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2228 			itx->itx_private = NULL;
2229 		}
2230 
2231 		mutex_exit(&zilog->zl_lock);
2232 
2233 		list_remove(&zilog->zl_itx_commit_list, itx);
2234 		zil_itx_destroy(itx);
2235 	}
2236 
2237 	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2238 }
2239 
2240 static void
2241 zil_commit_writer_stall(zilog_t *zilog)
2242 {
2243 	/*
2244 	 * When zio_alloc_zil() fails to allocate the next lwb block on
2245 	 * disk, we must call txg_wait_synced() to ensure all of the
2246 	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2247 	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2248 	 * to zil_process_commit_list()) will have to call zil_create(),
2249 	 * and start a new ZIL chain.
2250 	 *
2251 	 * Since zil_alloc_zil() failed, the lwb that was previously
2252 	 * issued does not have a pointer to the "next" lwb on disk.
2253 	 * Thus, if another ZIL writer thread was to allocate the "next"
2254 	 * on-disk lwb, that block could be leaked in the event of a
2255 	 * crash (because the previous lwb on-disk would not point to
2256 	 * it).
2257 	 *
2258 	 * We must hold the zilog's zl_issuer_lock while we do this, to
2259 	 * ensure no new threads enter zil_process_commit_list() until
2260 	 * all lwb's in the zl_lwb_list have been synced and freed
2261 	 * (which is achieved via the txg_wait_synced() call).
2262 	 */
2263 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2264 	txg_wait_synced(zilog->zl_dmu_pool, 0);
2265 	ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2266 }
2267 
2268 /*
2269  * This function will traverse the commit list, creating new lwbs as
2270  * needed, and committing the itxs from the commit list to these newly
2271  * created lwbs. Additionally, as a new lwb is created, the previous
2272  * lwb will be issued to the zio layer to be written to disk.
2273  */
2274 static void
2275 zil_process_commit_list(zilog_t *zilog)
2276 {
2277 	spa_t *spa = zilog->zl_spa;
2278 	list_t nolwb_itxs;
2279 	list_t nolwb_waiters;
2280 	lwb_t *lwb;
2281 	itx_t *itx;
2282 
2283 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2284 
2285 	/*
2286 	 * Return if there's nothing to commit before we dirty the fs by
2287 	 * calling zil_create().
2288 	 */
2289 	if (list_head(&zilog->zl_itx_commit_list) == NULL)
2290 		return;
2291 
2292 	list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
2293 	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2294 	    offsetof(zil_commit_waiter_t, zcw_node));
2295 
2296 	lwb = list_tail(&zilog->zl_lwb_list);
2297 	if (lwb == NULL) {
2298 		lwb = zil_create(zilog);
2299 	} else {
2300 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2301 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2302 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2303 	}
2304 
2305 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2306 		lr_t *lrc = &itx->itx_lr;
2307 		uint64_t txg = lrc->lrc_txg;
2308 
2309 		ASSERT3U(txg, !=, 0);
2310 
2311 		if (lrc->lrc_txtype == TX_COMMIT) {
2312 			DTRACE_PROBE2(zil__process__commit__itx,
2313 			    zilog_t *, zilog, itx_t *, itx);
2314 		} else {
2315 			DTRACE_PROBE2(zil__process__normal__itx,
2316 			    zilog_t *, zilog, itx_t *, itx);
2317 		}
2318 
2319 		list_remove(&zilog->zl_itx_commit_list, itx);
2320 
2321 		boolean_t synced = txg <= spa_last_synced_txg(spa);
2322 		boolean_t frozen = txg > spa_freeze_txg(spa);
2323 
2324 		/*
2325 		 * If the txg of this itx has already been synced out, then
2326 		 * we don't need to commit this itx to an lwb. This is
2327 		 * because the data of this itx will have already been
2328 		 * written to the main pool. This is inherently racy, and
2329 		 * it's still ok to commit an itx whose txg has already
2330 		 * been synced; this will result in a write that's
2331 		 * unnecessary, but will do no harm.
2332 		 *
2333 		 * With that said, we always want to commit TX_COMMIT itxs
2334 		 * to an lwb, regardless of whether or not that itx's txg
2335 		 * has been synced out. We do this to ensure any OPENED lwb
2336 		 * will always have at least one zil_commit_waiter_t linked
2337 		 * to the lwb.
2338 		 *
2339 		 * As a counter-example, if we skipped TX_COMMIT itx's
2340 		 * whose txg had already been synced, the following
2341 		 * situation could occur if we happened to be racing with
2342 		 * spa_sync:
2343 		 *
2344 		 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2345 		 *    itx's txg is 10 and the last synced txg is 9.
2346 		 * 2. spa_sync finishes syncing out txg 10.
2347 		 * 3. We move to the next itx in the list, it's a TX_COMMIT
2348 		 *    whose txg is 10, so we skip it rather than committing
2349 		 *    it to the lwb used in (1).
2350 		 *
2351 		 * If the itx that is skipped in (3) is the last TX_COMMIT
2352 		 * itx in the commit list, than it's possible for the lwb
2353 		 * used in (1) to remain in the OPENED state indefinitely.
2354 		 *
2355 		 * To prevent the above scenario from occurring, ensuring
2356 		 * that once an lwb is OPENED it will transition to ISSUED
2357 		 * and eventually DONE, we always commit TX_COMMIT itx's to
2358 		 * an lwb here, even if that itx's txg has already been
2359 		 * synced.
2360 		 *
2361 		 * Finally, if the pool is frozen, we _always_ commit the
2362 		 * itx.  The point of freezing the pool is to prevent data
2363 		 * from being written to the main pool via spa_sync, and
2364 		 * instead rely solely on the ZIL to persistently store the
2365 		 * data; i.e.  when the pool is frozen, the last synced txg
2366 		 * value can't be trusted.
2367 		 */
2368 		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2369 			if (lwb != NULL) {
2370 				lwb = zil_lwb_commit(zilog, itx, lwb);
2371 
2372 				if (lwb == NULL)
2373 					list_insert_tail(&nolwb_itxs, itx);
2374 				else
2375 					list_insert_tail(&lwb->lwb_itxs, itx);
2376 			} else {
2377 				if (lrc->lrc_txtype == TX_COMMIT) {
2378 					zil_commit_waiter_link_nolwb(
2379 					    itx->itx_private, &nolwb_waiters);
2380 				}
2381 
2382 				list_insert_tail(&nolwb_itxs, itx);
2383 			}
2384 		} else {
2385 			ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
2386 			zil_itx_destroy(itx);
2387 		}
2388 	}
2389 
2390 	if (lwb == NULL) {
2391 		/*
2392 		 * This indicates zio_alloc_zil() failed to allocate the
2393 		 * "next" lwb on-disk. When this happens, we must stall
2394 		 * the ZIL write pipeline; see the comment within
2395 		 * zil_commit_writer_stall() for more details.
2396 		 */
2397 		zil_commit_writer_stall(zilog);
2398 
2399 		/*
2400 		 * Additionally, we have to signal and mark the "nolwb"
2401 		 * waiters as "done" here, since without an lwb, we
2402 		 * can't do this via zil_lwb_flush_vdevs_done() like
2403 		 * normal.
2404 		 */
2405 		zil_commit_waiter_t *zcw;
2406 		while ((zcw = list_head(&nolwb_waiters)) != NULL) {
2407 			zil_commit_waiter_skip(zcw);
2408 			list_remove(&nolwb_waiters, zcw);
2409 		}
2410 
2411 		/*
2412 		 * And finally, we have to destroy the itx's that
2413 		 * couldn't be committed to an lwb; this will also call
2414 		 * the itx's callback if one exists for the itx.
2415 		 */
2416 		while ((itx = list_head(&nolwb_itxs)) != NULL) {
2417 			list_remove(&nolwb_itxs, itx);
2418 			zil_itx_destroy(itx);
2419 		}
2420 	} else {
2421 		ASSERT(list_is_empty(&nolwb_waiters));
2422 		ASSERT3P(lwb, !=, NULL);
2423 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2424 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2425 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2426 
2427 		/*
2428 		 * At this point, the ZIL block pointed at by the "lwb"
2429 		 * variable is in one of the following states: "closed"
2430 		 * or "open".
2431 		 *
2432 		 * If it's "closed", then no itxs have been committed to
2433 		 * it, so there's no point in issuing its zio (i.e. it's
2434 		 * "empty").
2435 		 *
2436 		 * If it's "open", then it contains one or more itxs that
2437 		 * eventually need to be committed to stable storage. In
2438 		 * this case we intentionally do not issue the lwb's zio
2439 		 * to disk yet, and instead rely on one of the following
2440 		 * two mechanisms for issuing the zio:
2441 		 *
2442 		 * 1. Ideally, there will be more ZIL activity occurring
2443 		 * on the system, such that this function will be
2444 		 * immediately called again (not necessarily by the same
2445 		 * thread) and this lwb's zio will be issued via
2446 		 * zil_lwb_commit(). This way, the lwb is guaranteed to
2447 		 * be "full" when it is issued to disk, and we'll make
2448 		 * use of the lwb's size the best we can.
2449 		 *
2450 		 * 2. If there isn't sufficient ZIL activity occurring on
2451 		 * the system, such that this lwb's zio isn't issued via
2452 		 * zil_lwb_commit(), zil_commit_waiter() will issue the
2453 		 * lwb's zio. If this occurs, the lwb is not guaranteed
2454 		 * to be "full" by the time its zio is issued, and means
2455 		 * the size of the lwb was "too large" given the amount
2456 		 * of ZIL activity occurring on the system at that time.
2457 		 *
2458 		 * We do this for a couple of reasons:
2459 		 *
2460 		 * 1. To try and reduce the number of IOPs needed to
2461 		 * write the same number of itxs. If an lwb has space
2462 		 * available in its buffer for more itxs, and more itxs
2463 		 * will be committed relatively soon (relative to the
2464 		 * latency of performing a write), then it's beneficial
2465 		 * to wait for these "next" itxs. This way, more itxs
2466 		 * can be committed to stable storage with fewer writes.
2467 		 *
2468 		 * 2. To try and use the largest lwb block size that the
2469 		 * incoming rate of itxs can support. Again, this is to
2470 		 * try and pack as many itxs into as few lwbs as
2471 		 * possible, without significantly impacting the latency
2472 		 * of each individual itx.
2473 		 */
2474 	}
2475 }
2476 
2477 /*
2478  * This function is responsible for ensuring the passed in commit waiter
2479  * (and associated commit itx) is committed to an lwb. If the waiter is
2480  * not already committed to an lwb, all itxs in the zilog's queue of
2481  * itxs will be processed. The assumption is the passed in waiter's
2482  * commit itx will found in the queue just like the other non-commit
2483  * itxs, such that when the entire queue is processed, the waiter will
2484  * have been committed to an lwb.
2485  *
2486  * The lwb associated with the passed in waiter is not guaranteed to
2487  * have been issued by the time this function completes. If the lwb is
2488  * not issued, we rely on future calls to zil_commit_writer() to issue
2489  * the lwb, or the timeout mechanism found in zil_commit_waiter().
2490  */
2491 static void
2492 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2493 {
2494 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2495 	ASSERT(spa_writeable(zilog->zl_spa));
2496 
2497 	mutex_enter(&zilog->zl_issuer_lock);
2498 
2499 	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2500 		/*
2501 		 * It's possible that, while we were waiting to acquire
2502 		 * the "zl_issuer_lock", another thread committed this
2503 		 * waiter to an lwb. If that occurs, we bail out early,
2504 		 * without processing any of the zilog's queue of itxs.
2505 		 *
2506 		 * On certain workloads and system configurations, the
2507 		 * "zl_issuer_lock" can become highly contended. In an
2508 		 * attempt to reduce this contention, we immediately drop
2509 		 * the lock if the waiter has already been processed.
2510 		 *
2511 		 * We've measured this optimization to reduce CPU spent
2512 		 * contending on this lock by up to 5%, using a system
2513 		 * with 32 CPUs, low latency storage (~50 usec writes),
2514 		 * and 1024 threads performing sync writes.
2515 		 */
2516 		goto out;
2517 	}
2518 
2519 	ZIL_STAT_BUMP(zil_commit_writer_count);
2520 
2521 	zil_get_commit_list(zilog);
2522 	zil_prune_commit_list(zilog);
2523 	zil_process_commit_list(zilog);
2524 
2525 out:
2526 	mutex_exit(&zilog->zl_issuer_lock);
2527 }
2528 
2529 static void
2530 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2531 {
2532 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2533 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2534 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2535 
2536 	lwb_t *lwb = zcw->zcw_lwb;
2537 	ASSERT3P(lwb, !=, NULL);
2538 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2539 
2540 	/*
2541 	 * If the lwb has already been issued by another thread, we can
2542 	 * immediately return since there's no work to be done (the
2543 	 * point of this function is to issue the lwb). Additionally, we
2544 	 * do this prior to acquiring the zl_issuer_lock, to avoid
2545 	 * acquiring it when it's not necessary to do so.
2546 	 */
2547 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2548 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2549 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2550 		return;
2551 
2552 	/*
2553 	 * In order to call zil_lwb_write_issue() we must hold the
2554 	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2555 	 * since we're already holding the commit waiter's "zcw_lock",
2556 	 * and those two locks are acquired in the opposite order
2557 	 * elsewhere.
2558 	 */
2559 	mutex_exit(&zcw->zcw_lock);
2560 	mutex_enter(&zilog->zl_issuer_lock);
2561 	mutex_enter(&zcw->zcw_lock);
2562 
2563 	/*
2564 	 * Since we just dropped and re-acquired the commit waiter's
2565 	 * lock, we have to re-check to see if the waiter was marked
2566 	 * "done" during that process. If the waiter was marked "done",
2567 	 * the "lwb" pointer is no longer valid (it can be free'd after
2568 	 * the waiter is marked "done"), so without this check we could
2569 	 * wind up with a use-after-free error below.
2570 	 */
2571 	if (zcw->zcw_done)
2572 		goto out;
2573 
2574 	ASSERT3P(lwb, ==, zcw->zcw_lwb);
2575 
2576 	/*
2577 	 * We've already checked this above, but since we hadn't acquired
2578 	 * the zilog's zl_issuer_lock, we have to perform this check a
2579 	 * second time while holding the lock.
2580 	 *
2581 	 * We don't need to hold the zl_lock since the lwb cannot transition
2582 	 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2583 	 * _can_ transition from ISSUED to DONE, but it's OK to race with
2584 	 * that transition since we treat the lwb the same, whether it's in
2585 	 * the ISSUED or DONE states.
2586 	 *
2587 	 * The important thing, is we treat the lwb differently depending on
2588 	 * if it's ISSUED or OPENED, and block any other threads that might
2589 	 * attempt to issue this lwb. For that reason we hold the
2590 	 * zl_issuer_lock when checking the lwb_state; we must not call
2591 	 * zil_lwb_write_issue() if the lwb had already been issued.
2592 	 *
2593 	 * See the comment above the lwb_state_t structure definition for
2594 	 * more details on the lwb states, and locking requirements.
2595 	 */
2596 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2597 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2598 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2599 		goto out;
2600 
2601 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2602 
2603 	/*
2604 	 * As described in the comments above zil_commit_waiter() and
2605 	 * zil_process_commit_list(), we need to issue this lwb's zio
2606 	 * since we've reached the commit waiter's timeout and it still
2607 	 * hasn't been issued.
2608 	 */
2609 	lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2610 
2611 	IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2612 
2613 	/*
2614 	 * Since the lwb's zio hadn't been issued by the time this thread
2615 	 * reached its timeout, we reset the zilog's "zl_cur_used" field
2616 	 * to influence the zil block size selection algorithm.
2617 	 *
2618 	 * By having to issue the lwb's zio here, it means the size of the
2619 	 * lwb was too large, given the incoming throughput of itxs.  By
2620 	 * setting "zl_cur_used" to zero, we communicate this fact to the
2621 	 * block size selection algorithm, so it can take this information
2622 	 * into account, and potentially select a smaller size for the
2623 	 * next lwb block that is allocated.
2624 	 */
2625 	zilog->zl_cur_used = 0;
2626 
2627 	if (nlwb == NULL) {
2628 		/*
2629 		 * When zil_lwb_write_issue() returns NULL, this
2630 		 * indicates zio_alloc_zil() failed to allocate the
2631 		 * "next" lwb on-disk. When this occurs, the ZIL write
2632 		 * pipeline must be stalled; see the comment within the
2633 		 * zil_commit_writer_stall() function for more details.
2634 		 *
2635 		 * We must drop the commit waiter's lock prior to
2636 		 * calling zil_commit_writer_stall() or else we can wind
2637 		 * up with the following deadlock:
2638 		 *
2639 		 * - This thread is waiting for the txg to sync while
2640 		 *   holding the waiter's lock; txg_wait_synced() is
2641 		 *   used within txg_commit_writer_stall().
2642 		 *
2643 		 * - The txg can't sync because it is waiting for this
2644 		 *   lwb's zio callback to call dmu_tx_commit().
2645 		 *
2646 		 * - The lwb's zio callback can't call dmu_tx_commit()
2647 		 *   because it's blocked trying to acquire the waiter's
2648 		 *   lock, which occurs prior to calling dmu_tx_commit()
2649 		 */
2650 		mutex_exit(&zcw->zcw_lock);
2651 		zil_commit_writer_stall(zilog);
2652 		mutex_enter(&zcw->zcw_lock);
2653 	}
2654 
2655 out:
2656 	mutex_exit(&zilog->zl_issuer_lock);
2657 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2658 }
2659 
2660 /*
2661  * This function is responsible for performing the following two tasks:
2662  *
2663  * 1. its primary responsibility is to block until the given "commit
2664  *    waiter" is considered "done".
2665  *
2666  * 2. its secondary responsibility is to issue the zio for the lwb that
2667  *    the given "commit waiter" is waiting on, if this function has
2668  *    waited "long enough" and the lwb is still in the "open" state.
2669  *
2670  * Given a sufficient amount of itxs being generated and written using
2671  * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2672  * function. If this does not occur, this secondary responsibility will
2673  * ensure the lwb is issued even if there is not other synchronous
2674  * activity on the system.
2675  *
2676  * For more details, see zil_process_commit_list(); more specifically,
2677  * the comment at the bottom of that function.
2678  */
2679 static void
2680 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2681 {
2682 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2683 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2684 	ASSERT(spa_writeable(zilog->zl_spa));
2685 
2686 	mutex_enter(&zcw->zcw_lock);
2687 
2688 	/*
2689 	 * The timeout is scaled based on the lwb latency to avoid
2690 	 * significantly impacting the latency of each individual itx.
2691 	 * For more details, see the comment at the bottom of the
2692 	 * zil_process_commit_list() function.
2693 	 */
2694 	int pct = MAX(zfs_commit_timeout_pct, 1);
2695 	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2696 	hrtime_t wakeup = gethrtime() + sleep;
2697 	boolean_t timedout = B_FALSE;
2698 
2699 	while (!zcw->zcw_done) {
2700 		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2701 
2702 		lwb_t *lwb = zcw->zcw_lwb;
2703 
2704 		/*
2705 		 * Usually, the waiter will have a non-NULL lwb field here,
2706 		 * but it's possible for it to be NULL as a result of
2707 		 * zil_commit() racing with spa_sync().
2708 		 *
2709 		 * When zil_clean() is called, it's possible for the itxg
2710 		 * list (which may be cleaned via a taskq) to contain
2711 		 * commit itxs. When this occurs, the commit waiters linked
2712 		 * off of these commit itxs will not be committed to an
2713 		 * lwb.  Additionally, these commit waiters will not be
2714 		 * marked done until zil_commit_waiter_skip() is called via
2715 		 * zil_itxg_clean().
2716 		 *
2717 		 * Thus, it's possible for this commit waiter (i.e. the
2718 		 * "zcw" variable) to be found in this "in between" state;
2719 		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2720 		 * been skipped, so it's "zcw_done" field is still B_FALSE.
2721 		 */
2722 		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2723 
2724 		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2725 			ASSERT3B(timedout, ==, B_FALSE);
2726 
2727 			/*
2728 			 * If the lwb hasn't been issued yet, then we
2729 			 * need to wait with a timeout, in case this
2730 			 * function needs to issue the lwb after the
2731 			 * timeout is reached; responsibility (2) from
2732 			 * the comment above this function.
2733 			 */
2734 			int rc = cv_timedwait_hires(&zcw->zcw_cv,
2735 			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2736 			    CALLOUT_FLAG_ABSOLUTE);
2737 
2738 			if (rc != -1 || zcw->zcw_done)
2739 				continue;
2740 
2741 			timedout = B_TRUE;
2742 			zil_commit_waiter_timeout(zilog, zcw);
2743 
2744 			if (!zcw->zcw_done) {
2745 				/*
2746 				 * If the commit waiter has already been
2747 				 * marked "done", it's possible for the
2748 				 * waiter's lwb structure to have already
2749 				 * been freed.  Thus, we can only reliably
2750 				 * make these assertions if the waiter
2751 				 * isn't done.
2752 				 */
2753 				ASSERT3P(lwb, ==, zcw->zcw_lwb);
2754 				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2755 			}
2756 		} else {
2757 			/*
2758 			 * If the lwb isn't open, then it must have already
2759 			 * been issued. In that case, there's no need to
2760 			 * use a timeout when waiting for the lwb to
2761 			 * complete.
2762 			 *
2763 			 * Additionally, if the lwb is NULL, the waiter
2764 			 * will soon be signaled and marked done via
2765 			 * zil_clean() and zil_itxg_clean(), so no timeout
2766 			 * is required.
2767 			 */
2768 
2769 			IMPLY(lwb != NULL,
2770 			    lwb->lwb_state == LWB_STATE_ISSUED ||
2771 			    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2772 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
2773 			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2774 		}
2775 	}
2776 
2777 	mutex_exit(&zcw->zcw_lock);
2778 }
2779 
2780 static zil_commit_waiter_t *
2781 zil_alloc_commit_waiter(void)
2782 {
2783 	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2784 
2785 	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2786 	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2787 	list_link_init(&zcw->zcw_node);
2788 	zcw->zcw_lwb = NULL;
2789 	zcw->zcw_done = B_FALSE;
2790 	zcw->zcw_zio_error = 0;
2791 
2792 	return (zcw);
2793 }
2794 
2795 static void
2796 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2797 {
2798 	ASSERT(!list_link_active(&zcw->zcw_node));
2799 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
2800 	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2801 	mutex_destroy(&zcw->zcw_lock);
2802 	cv_destroy(&zcw->zcw_cv);
2803 	kmem_cache_free(zil_zcw_cache, zcw);
2804 }
2805 
2806 /*
2807  * This function is used to create a TX_COMMIT itx and assign it. This
2808  * way, it will be linked into the ZIL's list of synchronous itxs, and
2809  * then later committed to an lwb (or skipped) when
2810  * zil_process_commit_list() is called.
2811  */
2812 static void
2813 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2814 {
2815 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2816 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2817 
2818 	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2819 	itx->itx_sync = B_TRUE;
2820 	itx->itx_private = zcw;
2821 
2822 	zil_itx_assign(zilog, itx, tx);
2823 
2824 	dmu_tx_commit(tx);
2825 }
2826 
2827 /*
2828  * Commit ZFS Intent Log transactions (itxs) to stable storage.
2829  *
2830  * When writing ZIL transactions to the on-disk representation of the
2831  * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2832  * itxs can be committed to a single lwb. Once a lwb is written and
2833  * committed to stable storage (i.e. the lwb is written, and vdevs have
2834  * been flushed), each itx that was committed to that lwb is also
2835  * considered to be committed to stable storage.
2836  *
2837  * When an itx is committed to an lwb, the log record (lr_t) contained
2838  * by the itx is copied into the lwb's zio buffer, and once this buffer
2839  * is written to disk, it becomes an on-disk ZIL block.
2840  *
2841  * As itxs are generated, they're inserted into the ZIL's queue of
2842  * uncommitted itxs. The semantics of zil_commit() are such that it will
2843  * block until all itxs that were in the queue when it was called, are
2844  * committed to stable storage.
2845  *
2846  * If "foid" is zero, this means all "synchronous" and "asynchronous"
2847  * itxs, for all objects in the dataset, will be committed to stable
2848  * storage prior to zil_commit() returning. If "foid" is non-zero, all
2849  * "synchronous" itxs for all objects, but only "asynchronous" itxs
2850  * that correspond to the foid passed in, will be committed to stable
2851  * storage prior to zil_commit() returning.
2852  *
2853  * Generally speaking, when zil_commit() is called, the consumer doesn't
2854  * actually care about _all_ of the uncommitted itxs. Instead, they're
2855  * simply trying to waiting for a specific itx to be committed to disk,
2856  * but the interface(s) for interacting with the ZIL don't allow such
2857  * fine-grained communication. A better interface would allow a consumer
2858  * to create and assign an itx, and then pass a reference to this itx to
2859  * zil_commit(); such that zil_commit() would return as soon as that
2860  * specific itx was committed to disk (instead of waiting for _all_
2861  * itxs to be committed).
2862  *
2863  * When a thread calls zil_commit() a special "commit itx" will be
2864  * generated, along with a corresponding "waiter" for this commit itx.
2865  * zil_commit() will wait on this waiter's CV, such that when the waiter
2866  * is marked done, and signaled, zil_commit() will return.
2867  *
2868  * This commit itx is inserted into the queue of uncommitted itxs. This
2869  * provides an easy mechanism for determining which itxs were in the
2870  * queue prior to zil_commit() having been called, and which itxs were
2871  * added after zil_commit() was called.
2872  *
2873  * The commit it is special; it doesn't have any on-disk representation.
2874  * When a commit itx is "committed" to an lwb, the waiter associated
2875  * with it is linked onto the lwb's list of waiters. Then, when that lwb
2876  * completes, each waiter on the lwb's list is marked done and signaled
2877  * -- allowing the thread waiting on the waiter to return from zil_commit().
2878  *
2879  * It's important to point out a few critical factors that allow us
2880  * to make use of the commit itxs, commit waiters, per-lwb lists of
2881  * commit waiters, and zio completion callbacks like we're doing:
2882  *
2883  *   1. The list of waiters for each lwb is traversed, and each commit
2884  *      waiter is marked "done" and signaled, in the zio completion
2885  *      callback of the lwb's zio[*].
2886  *
2887  *      * Actually, the waiters are signaled in the zio completion
2888  *        callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2889  *        that are sent to the vdevs upon completion of the lwb zio.
2890  *
2891  *   2. When the itxs are inserted into the ZIL's queue of uncommitted
2892  *      itxs, the order in which they are inserted is preserved[*]; as
2893  *      itxs are added to the queue, they are added to the tail of
2894  *      in-memory linked lists.
2895  *
2896  *      When committing the itxs to lwbs (to be written to disk), they
2897  *      are committed in the same order in which the itxs were added to
2898  *      the uncommitted queue's linked list(s); i.e. the linked list of
2899  *      itxs to commit is traversed from head to tail, and each itx is
2900  *      committed to an lwb in that order.
2901  *
2902  *      * To clarify:
2903  *
2904  *        - the order of "sync" itxs is preserved w.r.t. other
2905  *          "sync" itxs, regardless of the corresponding objects.
2906  *        - the order of "async" itxs is preserved w.r.t. other
2907  *          "async" itxs corresponding to the same object.
2908  *        - the order of "async" itxs is *not* preserved w.r.t. other
2909  *          "async" itxs corresponding to different objects.
2910  *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
2911  *          versa) is *not* preserved, even for itxs that correspond
2912  *          to the same object.
2913  *
2914  *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
2915  *      zil_get_commit_list(), and zil_process_commit_list().
2916  *
2917  *   3. The lwbs represent a linked list of blocks on disk. Thus, any
2918  *      lwb cannot be considered committed to stable storage, until its
2919  *      "previous" lwb is also committed to stable storage. This fact,
2920  *      coupled with the fact described above, means that itxs are
2921  *      committed in (roughly) the order in which they were generated.
2922  *      This is essential because itxs are dependent on prior itxs.
2923  *      Thus, we *must not* deem an itx as being committed to stable
2924  *      storage, until *all* prior itxs have also been committed to
2925  *      stable storage.
2926  *
2927  *      To enforce this ordering of lwb zio's, while still leveraging as
2928  *      much of the underlying storage performance as possible, we rely
2929  *      on two fundamental concepts:
2930  *
2931  *          1. The creation and issuance of lwb zio's is protected by
2932  *             the zilog's "zl_issuer_lock", which ensures only a single
2933  *             thread is creating and/or issuing lwb's at a time
2934  *          2. The "previous" lwb is a child of the "current" lwb
2935  *             (leveraging the zio parent-child dependency graph)
2936  *
2937  *      By relying on this parent-child zio relationship, we can have
2938  *      many lwb zio's concurrently issued to the underlying storage,
2939  *      but the order in which they complete will be the same order in
2940  *      which they were created.
2941  */
2942 void
2943 zil_commit(zilog_t *zilog, uint64_t foid)
2944 {
2945 	/*
2946 	 * We should never attempt to call zil_commit on a snapshot for
2947 	 * a couple of reasons:
2948 	 *
2949 	 * 1. A snapshot may never be modified, thus it cannot have any
2950 	 *    in-flight itxs that would have modified the dataset.
2951 	 *
2952 	 * 2. By design, when zil_commit() is called, a commit itx will
2953 	 *    be assigned to this zilog; as a result, the zilog will be
2954 	 *    dirtied. We must not dirty the zilog of a snapshot; there's
2955 	 *    checks in the code that enforce this invariant, and will
2956 	 *    cause a panic if it's not upheld.
2957 	 */
2958 	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2959 
2960 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2961 		return;
2962 
2963 	if (!spa_writeable(zilog->zl_spa)) {
2964 		/*
2965 		 * If the SPA is not writable, there should never be any
2966 		 * pending itxs waiting to be committed to disk. If that
2967 		 * weren't true, we'd skip writing those itxs out, and
2968 		 * would break the semantics of zil_commit(); thus, we're
2969 		 * verifying that truth before we return to the caller.
2970 		 */
2971 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
2972 		ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2973 		for (int i = 0; i < TXG_SIZE; i++)
2974 			ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2975 		return;
2976 	}
2977 
2978 	/*
2979 	 * If the ZIL is suspended, we don't want to dirty it by calling
2980 	 * zil_commit_itx_assign() below, nor can we write out
2981 	 * lwbs like would be done in zil_commit_write(). Thus, we
2982 	 * simply rely on txg_wait_synced() to maintain the necessary
2983 	 * semantics, and avoid calling those functions altogether.
2984 	 */
2985 	if (zilog->zl_suspend > 0) {
2986 		txg_wait_synced(zilog->zl_dmu_pool, 0);
2987 		return;
2988 	}
2989 
2990 	zil_commit_impl(zilog, foid);
2991 }
2992 
2993 void
2994 zil_commit_impl(zilog_t *zilog, uint64_t foid)
2995 {
2996 	ZIL_STAT_BUMP(zil_commit_count);
2997 
2998 	/*
2999 	 * Move the "async" itxs for the specified foid to the "sync"
3000 	 * queues, such that they will be later committed (or skipped)
3001 	 * to an lwb when zil_process_commit_list() is called.
3002 	 *
3003 	 * Since these "async" itxs must be committed prior to this
3004 	 * call to zil_commit returning, we must perform this operation
3005 	 * before we call zil_commit_itx_assign().
3006 	 */
3007 	zil_async_to_sync(zilog, foid);
3008 
3009 	/*
3010 	 * We allocate a new "waiter" structure which will initially be
3011 	 * linked to the commit itx using the itx's "itx_private" field.
3012 	 * Since the commit itx doesn't represent any on-disk state,
3013 	 * when it's committed to an lwb, rather than copying the its
3014 	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3015 	 * added to the lwb's list of waiters. Then, when the lwb is
3016 	 * committed to stable storage, each waiter in the lwb's list of
3017 	 * waiters will be marked "done", and signalled.
3018 	 *
3019 	 * We must create the waiter and assign the commit itx prior to
3020 	 * calling zil_commit_writer(), or else our specific commit itx
3021 	 * is not guaranteed to be committed to an lwb prior to calling
3022 	 * zil_commit_waiter().
3023 	 */
3024 	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
3025 	zil_commit_itx_assign(zilog, zcw);
3026 
3027 	zil_commit_writer(zilog, zcw);
3028 	zil_commit_waiter(zilog, zcw);
3029 
3030 	if (zcw->zcw_zio_error != 0) {
3031 		/*
3032 		 * If there was an error writing out the ZIL blocks that
3033 		 * this thread is waiting on, then we fallback to
3034 		 * relying on spa_sync() to write out the data this
3035 		 * thread is waiting on. Obviously this has performance
3036 		 * implications, but the expectation is for this to be
3037 		 * an exceptional case, and shouldn't occur often.
3038 		 */
3039 		DTRACE_PROBE2(zil__commit__io__error,
3040 		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
3041 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3042 	}
3043 
3044 	zil_free_commit_waiter(zcw);
3045 }
3046 
3047 /*
3048  * Called in syncing context to free committed log blocks and update log header.
3049  */
3050 void
3051 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
3052 {
3053 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
3054 	uint64_t txg = dmu_tx_get_txg(tx);
3055 	spa_t *spa = zilog->zl_spa;
3056 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
3057 	lwb_t *lwb;
3058 
3059 	/*
3060 	 * We don't zero out zl_destroy_txg, so make sure we don't try
3061 	 * to destroy it twice.
3062 	 */
3063 	if (spa_sync_pass(spa) != 1)
3064 		return;
3065 
3066 	mutex_enter(&zilog->zl_lock);
3067 
3068 	ASSERT(zilog->zl_stop_sync == 0);
3069 
3070 	if (*replayed_seq != 0) {
3071 		ASSERT(zh->zh_replay_seq < *replayed_seq);
3072 		zh->zh_replay_seq = *replayed_seq;
3073 		*replayed_seq = 0;
3074 	}
3075 
3076 	if (zilog->zl_destroy_txg == txg) {
3077 		blkptr_t blk = zh->zh_log;
3078 
3079 		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
3080 
3081 		bzero(zh, sizeof (zil_header_t));
3082 		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
3083 
3084 		if (zilog->zl_keep_first) {
3085 			/*
3086 			 * If this block was part of log chain that couldn't
3087 			 * be claimed because a device was missing during
3088 			 * zil_claim(), but that device later returns,
3089 			 * then this block could erroneously appear valid.
3090 			 * To guard against this, assign a new GUID to the new
3091 			 * log chain so it doesn't matter what blk points to.
3092 			 */
3093 			zil_init_log_chain(zilog, &blk);
3094 			zh->zh_log = blk;
3095 		}
3096 	}
3097 
3098 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
3099 		zh->zh_log = lwb->lwb_blk;
3100 		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
3101 			break;
3102 		list_remove(&zilog->zl_lwb_list, lwb);
3103 		zio_free(spa, txg, &lwb->lwb_blk);
3104 		zil_free_lwb(zilog, lwb);
3105 
3106 		/*
3107 		 * If we don't have anything left in the lwb list then
3108 		 * we've had an allocation failure and we need to zero
3109 		 * out the zil_header blkptr so that we don't end
3110 		 * up freeing the same block twice.
3111 		 */
3112 		if (list_head(&zilog->zl_lwb_list) == NULL)
3113 			BP_ZERO(&zh->zh_log);
3114 	}
3115 
3116 	/*
3117 	 * Remove fastwrite on any blocks that have been pre-allocated for
3118 	 * the next commit. This prevents fastwrite counter pollution by
3119 	 * unused, long-lived LWBs.
3120 	 */
3121 	for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
3122 		if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) {
3123 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3124 			lwb->lwb_fastwrite = 0;
3125 		}
3126 	}
3127 
3128 	mutex_exit(&zilog->zl_lock);
3129 }
3130 
3131 static int
3132 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
3133 {
3134 	(void) unused, (void) kmflag;
3135 	lwb_t *lwb = vbuf;
3136 	list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3137 	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
3138 	    offsetof(zil_commit_waiter_t, zcw_node));
3139 	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
3140 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
3141 	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
3142 	return (0);
3143 }
3144 
3145 static void
3146 zil_lwb_dest(void *vbuf, void *unused)
3147 {
3148 	(void) unused;
3149 	lwb_t *lwb = vbuf;
3150 	mutex_destroy(&lwb->lwb_vdev_lock);
3151 	avl_destroy(&lwb->lwb_vdev_tree);
3152 	list_destroy(&lwb->lwb_waiters);
3153 	list_destroy(&lwb->lwb_itxs);
3154 }
3155 
3156 void
3157 zil_init(void)
3158 {
3159 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3160 	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3161 
3162 	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3163 	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3164 
3165 	zil_ksp = kstat_create("zfs", 0, "zil", "misc",
3166 	    KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
3167 	    KSTAT_FLAG_VIRTUAL);
3168 
3169 	if (zil_ksp != NULL) {
3170 		zil_ksp->ks_data = &zil_stats;
3171 		kstat_install(zil_ksp);
3172 	}
3173 }
3174 
3175 void
3176 zil_fini(void)
3177 {
3178 	kmem_cache_destroy(zil_zcw_cache);
3179 	kmem_cache_destroy(zil_lwb_cache);
3180 
3181 	if (zil_ksp != NULL) {
3182 		kstat_delete(zil_ksp);
3183 		zil_ksp = NULL;
3184 	}
3185 }
3186 
3187 void
3188 zil_set_sync(zilog_t *zilog, uint64_t sync)
3189 {
3190 	zilog->zl_sync = sync;
3191 }
3192 
3193 void
3194 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3195 {
3196 	zilog->zl_logbias = logbias;
3197 }
3198 
3199 zilog_t *
3200 zil_alloc(objset_t *os, zil_header_t *zh_phys)
3201 {
3202 	zilog_t *zilog;
3203 
3204 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3205 
3206 	zilog->zl_header = zh_phys;
3207 	zilog->zl_os = os;
3208 	zilog->zl_spa = dmu_objset_spa(os);
3209 	zilog->zl_dmu_pool = dmu_objset_pool(os);
3210 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
3211 	zilog->zl_logbias = dmu_objset_logbias(os);
3212 	zilog->zl_sync = dmu_objset_syncprop(os);
3213 	zilog->zl_dirty_max_txg = 0;
3214 	zilog->zl_last_lwb_opened = NULL;
3215 	zilog->zl_last_lwb_latency = 0;
3216 	zilog->zl_max_block_size = zil_maxblocksize;
3217 
3218 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3219 	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3220 
3221 	for (int i = 0; i < TXG_SIZE; i++) {
3222 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3223 		    MUTEX_DEFAULT, NULL);
3224 	}
3225 
3226 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3227 	    offsetof(lwb_t, lwb_node));
3228 
3229 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3230 	    offsetof(itx_t, itx_node));
3231 
3232 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3233 
3234 	return (zilog);
3235 }
3236 
3237 void
3238 zil_free(zilog_t *zilog)
3239 {
3240 	int i;
3241 
3242 	zilog->zl_stop_sync = 1;
3243 
3244 	ASSERT0(zilog->zl_suspend);
3245 	ASSERT0(zilog->zl_suspending);
3246 
3247 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3248 	list_destroy(&zilog->zl_lwb_list);
3249 
3250 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3251 	list_destroy(&zilog->zl_itx_commit_list);
3252 
3253 	for (i = 0; i < TXG_SIZE; i++) {
3254 		/*
3255 		 * It's possible for an itx to be generated that doesn't dirty
3256 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3257 		 * callback to remove the entry. We remove those here.
3258 		 *
3259 		 * Also free up the ziltest itxs.
3260 		 */
3261 		if (zilog->zl_itxg[i].itxg_itxs)
3262 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3263 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3264 	}
3265 
3266 	mutex_destroy(&zilog->zl_issuer_lock);
3267 	mutex_destroy(&zilog->zl_lock);
3268 
3269 	cv_destroy(&zilog->zl_cv_suspend);
3270 
3271 	kmem_free(zilog, sizeof (zilog_t));
3272 }
3273 
3274 /*
3275  * Open an intent log.
3276  */
3277 zilog_t *
3278 zil_open(objset_t *os, zil_get_data_t *get_data)
3279 {
3280 	zilog_t *zilog = dmu_objset_zil(os);
3281 
3282 	ASSERT3P(zilog->zl_get_data, ==, NULL);
3283 	ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3284 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3285 
3286 	zilog->zl_get_data = get_data;
3287 
3288 	return (zilog);
3289 }
3290 
3291 /*
3292  * Close an intent log.
3293  */
3294 void
3295 zil_close(zilog_t *zilog)
3296 {
3297 	lwb_t *lwb;
3298 	uint64_t txg;
3299 
3300 	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3301 		zil_commit(zilog, 0);
3302 	} else {
3303 		ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
3304 		ASSERT0(zilog->zl_dirty_max_txg);
3305 		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3306 	}
3307 
3308 	mutex_enter(&zilog->zl_lock);
3309 	lwb = list_tail(&zilog->zl_lwb_list);
3310 	if (lwb == NULL)
3311 		txg = zilog->zl_dirty_max_txg;
3312 	else
3313 		txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
3314 	mutex_exit(&zilog->zl_lock);
3315 
3316 	/*
3317 	 * We need to use txg_wait_synced() to wait long enough for the
3318 	 * ZIL to be clean, and to wait for all pending lwbs to be
3319 	 * written out.
3320 	 */
3321 	if (txg != 0)
3322 		txg_wait_synced(zilog->zl_dmu_pool, txg);
3323 
3324 	if (zilog_is_dirty(zilog))
3325 		zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog,
3326 		    (u_longlong_t)txg);
3327 	if (txg < spa_freeze_txg(zilog->zl_spa))
3328 		VERIFY(!zilog_is_dirty(zilog));
3329 
3330 	zilog->zl_get_data = NULL;
3331 
3332 	/*
3333 	 * We should have only one lwb left on the list; remove it now.
3334 	 */
3335 	mutex_enter(&zilog->zl_lock);
3336 	lwb = list_head(&zilog->zl_lwb_list);
3337 	if (lwb != NULL) {
3338 		ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
3339 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3340 
3341 		if (lwb->lwb_fastwrite)
3342 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3343 
3344 		list_remove(&zilog->zl_lwb_list, lwb);
3345 		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3346 		zil_free_lwb(zilog, lwb);
3347 	}
3348 	mutex_exit(&zilog->zl_lock);
3349 }
3350 
3351 static char *suspend_tag = "zil suspending";
3352 
3353 /*
3354  * Suspend an intent log.  While in suspended mode, we still honor
3355  * synchronous semantics, but we rely on txg_wait_synced() to do it.
3356  * On old version pools, we suspend the log briefly when taking a
3357  * snapshot so that it will have an empty intent log.
3358  *
3359  * Long holds are not really intended to be used the way we do here --
3360  * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
3361  * could fail.  Therefore we take pains to only put a long hold if it is
3362  * actually necessary.  Fortunately, it will only be necessary if the
3363  * objset is currently mounted (or the ZVOL equivalent).  In that case it
3364  * will already have a long hold, so we are not really making things any worse.
3365  *
3366  * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3367  * zvol_state_t), and use their mechanism to prevent their hold from being
3368  * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
3369  * very little gain.
3370  *
3371  * if cookiep == NULL, this does both the suspend & resume.
3372  * Otherwise, it returns with the dataset "long held", and the cookie
3373  * should be passed into zil_resume().
3374  */
3375 int
3376 zil_suspend(const char *osname, void **cookiep)
3377 {
3378 	objset_t *os;
3379 	zilog_t *zilog;
3380 	const zil_header_t *zh;
3381 	int error;
3382 
3383 	error = dmu_objset_hold(osname, suspend_tag, &os);
3384 	if (error != 0)
3385 		return (error);
3386 	zilog = dmu_objset_zil(os);
3387 
3388 	mutex_enter(&zilog->zl_lock);
3389 	zh = zilog->zl_header;
3390 
3391 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
3392 		mutex_exit(&zilog->zl_lock);
3393 		dmu_objset_rele(os, suspend_tag);
3394 		return (SET_ERROR(EBUSY));
3395 	}
3396 
3397 	/*
3398 	 * Don't put a long hold in the cases where we can avoid it.  This
3399 	 * is when there is no cookie so we are doing a suspend & resume
3400 	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3401 	 * for the suspend because it's already suspended, or there's no ZIL.
3402 	 */
3403 	if (cookiep == NULL && !zilog->zl_suspending &&
3404 	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3405 		mutex_exit(&zilog->zl_lock);
3406 		dmu_objset_rele(os, suspend_tag);
3407 		return (0);
3408 	}
3409 
3410 	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3411 	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3412 
3413 	zilog->zl_suspend++;
3414 
3415 	if (zilog->zl_suspend > 1) {
3416 		/*
3417 		 * Someone else is already suspending it.
3418 		 * Just wait for them to finish.
3419 		 */
3420 
3421 		while (zilog->zl_suspending)
3422 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3423 		mutex_exit(&zilog->zl_lock);
3424 
3425 		if (cookiep == NULL)
3426 			zil_resume(os);
3427 		else
3428 			*cookiep = os;
3429 		return (0);
3430 	}
3431 
3432 	/*
3433 	 * If there is no pointer to an on-disk block, this ZIL must not
3434 	 * be active (e.g. filesystem not mounted), so there's nothing
3435 	 * to clean up.
3436 	 */
3437 	if (BP_IS_HOLE(&zh->zh_log)) {
3438 		ASSERT(cookiep != NULL); /* fast path already handled */
3439 
3440 		*cookiep = os;
3441 		mutex_exit(&zilog->zl_lock);
3442 		return (0);
3443 	}
3444 
3445 	/*
3446 	 * The ZIL has work to do. Ensure that the associated encryption
3447 	 * key will remain mapped while we are committing the log by
3448 	 * grabbing a reference to it. If the key isn't loaded we have no
3449 	 * choice but to return an error until the wrapping key is loaded.
3450 	 */
3451 	if (os->os_encrypted &&
3452 	    dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
3453 		zilog->zl_suspend--;
3454 		mutex_exit(&zilog->zl_lock);
3455 		dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3456 		dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3457 		return (SET_ERROR(EACCES));
3458 	}
3459 
3460 	zilog->zl_suspending = B_TRUE;
3461 	mutex_exit(&zilog->zl_lock);
3462 
3463 	/*
3464 	 * We need to use zil_commit_impl to ensure we wait for all
3465 	 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
3466 	 * to disk before proceeding. If we used zil_commit instead, it
3467 	 * would just call txg_wait_synced(), because zl_suspend is set.
3468 	 * txg_wait_synced() doesn't wait for these lwb's to be
3469 	 * LWB_STATE_FLUSH_DONE before returning.
3470 	 */
3471 	zil_commit_impl(zilog, 0);
3472 
3473 	/*
3474 	 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3475 	 * use txg_wait_synced() to ensure the data from the zilog has
3476 	 * migrated to the main pool before calling zil_destroy().
3477 	 */
3478 	txg_wait_synced(zilog->zl_dmu_pool, 0);
3479 
3480 	zil_destroy(zilog, B_FALSE);
3481 
3482 	mutex_enter(&zilog->zl_lock);
3483 	zilog->zl_suspending = B_FALSE;
3484 	cv_broadcast(&zilog->zl_cv_suspend);
3485 	mutex_exit(&zilog->zl_lock);
3486 
3487 	if (os->os_encrypted)
3488 		dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
3489 
3490 	if (cookiep == NULL)
3491 		zil_resume(os);
3492 	else
3493 		*cookiep = os;
3494 	return (0);
3495 }
3496 
3497 void
3498 zil_resume(void *cookie)
3499 {
3500 	objset_t *os = cookie;
3501 	zilog_t *zilog = dmu_objset_zil(os);
3502 
3503 	mutex_enter(&zilog->zl_lock);
3504 	ASSERT(zilog->zl_suspend != 0);
3505 	zilog->zl_suspend--;
3506 	mutex_exit(&zilog->zl_lock);
3507 	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3508 	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3509 }
3510 
3511 typedef struct zil_replay_arg {
3512 	zil_replay_func_t *const *zr_replay;
3513 	void		*zr_arg;
3514 	boolean_t	zr_byteswap;
3515 	char		*zr_lr;
3516 } zil_replay_arg_t;
3517 
3518 static int
3519 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
3520 {
3521 	char name[ZFS_MAX_DATASET_NAME_LEN];
3522 
3523 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
3524 
3525 	dmu_objset_name(zilog->zl_os, name);
3526 
3527 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3528 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3529 	    (u_longlong_t)lr->lrc_seq,
3530 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3531 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
3532 
3533 	return (error);
3534 }
3535 
3536 static int
3537 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
3538     uint64_t claim_txg)
3539 {
3540 	zil_replay_arg_t *zr = zra;
3541 	const zil_header_t *zh = zilog->zl_header;
3542 	uint64_t reclen = lr->lrc_reclen;
3543 	uint64_t txtype = lr->lrc_txtype;
3544 	int error = 0;
3545 
3546 	zilog->zl_replaying_seq = lr->lrc_seq;
3547 
3548 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
3549 		return (0);
3550 
3551 	if (lr->lrc_txg < claim_txg)		/* already committed */
3552 		return (0);
3553 
3554 	/* Strip case-insensitive bit, still present in log record */
3555 	txtype &= ~TX_CI;
3556 
3557 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
3558 		return (zil_replay_error(zilog, lr, EINVAL));
3559 
3560 	/*
3561 	 * If this record type can be logged out of order, the object
3562 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
3563 	 */
3564 	if (TX_OOO(txtype)) {
3565 		error = dmu_object_info(zilog->zl_os,
3566 		    LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
3567 		if (error == ENOENT || error == EEXIST)
3568 			return (0);
3569 	}
3570 
3571 	/*
3572 	 * Make a copy of the data so we can revise and extend it.
3573 	 */
3574 	bcopy(lr, zr->zr_lr, reclen);
3575 
3576 	/*
3577 	 * If this is a TX_WRITE with a blkptr, suck in the data.
3578 	 */
3579 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3580 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
3581 		    zr->zr_lr + reclen);
3582 		if (error != 0)
3583 			return (zil_replay_error(zilog, lr, error));
3584 	}
3585 
3586 	/*
3587 	 * The log block containing this lr may have been byteswapped
3588 	 * so that we can easily examine common fields like lrc_txtype.
3589 	 * However, the log is a mix of different record types, and only the
3590 	 * replay vectors know how to byteswap their records.  Therefore, if
3591 	 * the lr was byteswapped, undo it before invoking the replay vector.
3592 	 */
3593 	if (zr->zr_byteswap)
3594 		byteswap_uint64_array(zr->zr_lr, reclen);
3595 
3596 	/*
3597 	 * We must now do two things atomically: replay this log record,
3598 	 * and update the log header sequence number to reflect the fact that
3599 	 * we did so. At the end of each replay function the sequence number
3600 	 * is updated if we are in replay mode.
3601 	 */
3602 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3603 	if (error != 0) {
3604 		/*
3605 		 * The DMU's dnode layer doesn't see removes until the txg
3606 		 * commits, so a subsequent claim can spuriously fail with
3607 		 * EEXIST. So if we receive any error we try syncing out
3608 		 * any removes then retry the transaction.  Note that we
3609 		 * specify B_FALSE for byteswap now, so we don't do it twice.
3610 		 */
3611 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3612 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3613 		if (error != 0)
3614 			return (zil_replay_error(zilog, lr, error));
3615 	}
3616 	return (0);
3617 }
3618 
3619 static int
3620 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
3621 {
3622 	(void) bp, (void) arg, (void) claim_txg;
3623 
3624 	zilog->zl_replay_blks++;
3625 
3626 	return (0);
3627 }
3628 
3629 /*
3630  * If this dataset has a non-empty intent log, replay it and destroy it.
3631  */
3632 void
3633 zil_replay(objset_t *os, void *arg,
3634     zil_replay_func_t *const replay_func[TX_MAX_TYPE])
3635 {
3636 	zilog_t *zilog = dmu_objset_zil(os);
3637 	const zil_header_t *zh = zilog->zl_header;
3638 	zil_replay_arg_t zr;
3639 
3640 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3641 		zil_destroy(zilog, B_TRUE);
3642 		return;
3643 	}
3644 
3645 	zr.zr_replay = replay_func;
3646 	zr.zr_arg = arg;
3647 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3648 	zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3649 
3650 	/*
3651 	 * Wait for in-progress removes to sync before starting replay.
3652 	 */
3653 	txg_wait_synced(zilog->zl_dmu_pool, 0);
3654 
3655 	zilog->zl_replay = B_TRUE;
3656 	zilog->zl_replay_time = ddi_get_lbolt();
3657 	ASSERT(zilog->zl_replay_blks == 0);
3658 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3659 	    zh->zh_claim_txg, B_TRUE);
3660 	vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3661 
3662 	zil_destroy(zilog, B_FALSE);
3663 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3664 	zilog->zl_replay = B_FALSE;
3665 }
3666 
3667 boolean_t
3668 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3669 {
3670 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3671 		return (B_TRUE);
3672 
3673 	if (zilog->zl_replay) {
3674 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3675 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3676 		    zilog->zl_replaying_seq;
3677 		return (B_TRUE);
3678 	}
3679 
3680 	return (B_FALSE);
3681 }
3682 
3683 int
3684 zil_reset(const char *osname, void *arg)
3685 {
3686 	(void) arg;
3687 
3688 	int error = zil_suspend(osname, NULL);
3689 	/* EACCES means crypto key not loaded */
3690 	if ((error == EACCES) || (error == EBUSY))
3691 		return (SET_ERROR(error));
3692 	if (error != 0)
3693 		return (SET_ERROR(EEXIST));
3694 	return (0);
3695 }
3696 
3697 EXPORT_SYMBOL(zil_alloc);
3698 EXPORT_SYMBOL(zil_free);
3699 EXPORT_SYMBOL(zil_open);
3700 EXPORT_SYMBOL(zil_close);
3701 EXPORT_SYMBOL(zil_replay);
3702 EXPORT_SYMBOL(zil_replaying);
3703 EXPORT_SYMBOL(zil_destroy);
3704 EXPORT_SYMBOL(zil_destroy_sync);
3705 EXPORT_SYMBOL(zil_itx_create);
3706 EXPORT_SYMBOL(zil_itx_destroy);
3707 EXPORT_SYMBOL(zil_itx_assign);
3708 EXPORT_SYMBOL(zil_commit);
3709 EXPORT_SYMBOL(zil_claim);
3710 EXPORT_SYMBOL(zil_check_log_chain);
3711 EXPORT_SYMBOL(zil_sync);
3712 EXPORT_SYMBOL(zil_clean);
3713 EXPORT_SYMBOL(zil_suspend);
3714 EXPORT_SYMBOL(zil_resume);
3715 EXPORT_SYMBOL(zil_lwb_add_block);
3716 EXPORT_SYMBOL(zil_bp_tree_add);
3717 EXPORT_SYMBOL(zil_set_sync);
3718 EXPORT_SYMBOL(zil_set_logbias);
3719 
3720 /* BEGIN CSTYLED */
3721 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, INT, ZMOD_RW,
3722 	"ZIL block open timeout percentage");
3723 
3724 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
3725 	"Disable intent logging replay");
3726 
3727 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
3728 	"Disable ZIL cache flushes");
3729 
3730 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, ULONG, ZMOD_RW,
3731 	"Limit in bytes slog sync writes per commit");
3732 
3733 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, INT, ZMOD_RW,
3734 	"Limit in bytes of ZIL log block size");
3735 /* END CSTYLED */
3736