xref: /freebsd/sys/contrib/openzfs/module/zfs/zil.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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 https://opensource.org/licenses/CDDL-1.0.
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 #include <sys/wmsum.h>
47 
48 /*
49  * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
50  * calls that change the file system. Each itx has enough information to
51  * be able to replay them after a system crash, power loss, or
52  * equivalent failure mode. These are stored in memory until either:
53  *
54  *   1. they are committed to the pool by the DMU transaction group
55  *      (txg), at which point they can be discarded; or
56  *   2. they are committed to the on-disk ZIL for the dataset being
57  *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
58  *      requirement).
59  *
60  * In the event of a crash or power loss, the itxs contained by each
61  * dataset's on-disk ZIL will be replayed when that dataset is first
62  * instantiated (e.g. if the dataset is a normal filesystem, when it is
63  * first mounted).
64  *
65  * As hinted at above, there is one ZIL per dataset (both the in-memory
66  * representation, and the on-disk representation). The on-disk format
67  * consists of 3 parts:
68  *
69  * 	- a single, per-dataset, ZIL header; which points to a chain of
70  * 	- zero or more ZIL blocks; each of which contains
71  * 	- zero or more ZIL records
72  *
73  * A ZIL record holds the information necessary to replay a single
74  * system call transaction. A ZIL block can hold many ZIL records, and
75  * the blocks are chained together, similarly to a singly linked list.
76  *
77  * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
78  * block in the chain, and the ZIL header points to the first block in
79  * the chain.
80  *
81  * Note, there is not a fixed place in the pool to hold these ZIL
82  * blocks; they are dynamically allocated and freed as needed from the
83  * blocks available on the pool, though they can be preferentially
84  * allocated from a dedicated "log" vdev.
85  */
86 
87 /*
88  * This controls the amount of time that a ZIL block (lwb) will remain
89  * "open" when it isn't "full", and it has a thread waiting for it to be
90  * committed to stable storage. Please refer to the zil_commit_waiter()
91  * function (and the comments within it) for more details.
92  */
93 static int zfs_commit_timeout_pct = 5;
94 
95 /*
96  * See zil.h for more information about these fields.
97  */
98 static zil_kstat_values_t zil_stats = {
99 	{ "zil_commit_count",			KSTAT_DATA_UINT64 },
100 	{ "zil_commit_writer_count",		KSTAT_DATA_UINT64 },
101 	{ "zil_itx_count",			KSTAT_DATA_UINT64 },
102 	{ "zil_itx_indirect_count",		KSTAT_DATA_UINT64 },
103 	{ "zil_itx_indirect_bytes",		KSTAT_DATA_UINT64 },
104 	{ "zil_itx_copied_count",		KSTAT_DATA_UINT64 },
105 	{ "zil_itx_copied_bytes",		KSTAT_DATA_UINT64 },
106 	{ "zil_itx_needcopy_count",		KSTAT_DATA_UINT64 },
107 	{ "zil_itx_needcopy_bytes",		KSTAT_DATA_UINT64 },
108 	{ "zil_itx_metaslab_normal_count",	KSTAT_DATA_UINT64 },
109 	{ "zil_itx_metaslab_normal_bytes",	KSTAT_DATA_UINT64 },
110 	{ "zil_itx_metaslab_slog_count",	KSTAT_DATA_UINT64 },
111 	{ "zil_itx_metaslab_slog_bytes",	KSTAT_DATA_UINT64 },
112 };
113 
114 static zil_sums_t zil_sums_global;
115 static kstat_t *zil_kstats_global;
116 
117 /*
118  * Disable intent logging replay.  This global ZIL switch affects all pools.
119  */
120 int zil_replay_disable = 0;
121 
122 /*
123  * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
124  * the disk(s) by the ZIL after an LWB write has completed. Setting this
125  * will cause ZIL corruption on power loss if a volatile out-of-order
126  * write cache is enabled.
127  */
128 static int zil_nocacheflush = 0;
129 
130 /*
131  * Limit SLOG write size per commit executed with synchronous priority.
132  * Any writes above that will be executed with lower (asynchronous) priority
133  * to limit potential SLOG device abuse by single active ZIL writer.
134  */
135 static unsigned long zil_slog_bulk = 768 * 1024;
136 
137 static kmem_cache_t *zil_lwb_cache;
138 static kmem_cache_t *zil_zcw_cache;
139 
140 #define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
141     sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
142 
143 static int
144 zil_bp_compare(const void *x1, const void *x2)
145 {
146 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
147 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
148 
149 	int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
150 	if (likely(cmp))
151 		return (cmp);
152 
153 	return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
154 }
155 
156 static void
157 zil_bp_tree_init(zilog_t *zilog)
158 {
159 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
160 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
161 }
162 
163 static void
164 zil_bp_tree_fini(zilog_t *zilog)
165 {
166 	avl_tree_t *t = &zilog->zl_bp_tree;
167 	zil_bp_node_t *zn;
168 	void *cookie = NULL;
169 
170 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
171 		kmem_free(zn, sizeof (zil_bp_node_t));
172 
173 	avl_destroy(t);
174 }
175 
176 int
177 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
178 {
179 	avl_tree_t *t = &zilog->zl_bp_tree;
180 	const dva_t *dva;
181 	zil_bp_node_t *zn;
182 	avl_index_t where;
183 
184 	if (BP_IS_EMBEDDED(bp))
185 		return (0);
186 
187 	dva = BP_IDENTITY(bp);
188 
189 	if (avl_find(t, dva, &where) != NULL)
190 		return (SET_ERROR(EEXIST));
191 
192 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
193 	zn->zn_dva = *dva;
194 	avl_insert(t, zn, where);
195 
196 	return (0);
197 }
198 
199 static zil_header_t *
200 zil_header_in_syncing_context(zilog_t *zilog)
201 {
202 	return ((zil_header_t *)zilog->zl_header);
203 }
204 
205 static void
206 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
207 {
208 	zio_cksum_t *zc = &bp->blk_cksum;
209 
210 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0],
211 	    sizeof (zc->zc_word[ZIL_ZC_GUID_0]));
212 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1],
213 	    sizeof (zc->zc_word[ZIL_ZC_GUID_1]));
214 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
215 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
216 }
217 
218 static int
219 zil_kstats_global_update(kstat_t *ksp, int rw)
220 {
221 	zil_kstat_values_t *zs = ksp->ks_data;
222 	ASSERT3P(&zil_stats, ==, zs);
223 
224 	if (rw == KSTAT_WRITE) {
225 		return (SET_ERROR(EACCES));
226 	}
227 
228 	zil_kstat_values_update(zs, &zil_sums_global);
229 
230 	return (0);
231 }
232 
233 /*
234  * Read a log block and make sure it's valid.
235  */
236 static int
237 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
238     blkptr_t *nbp, void *dst, char **end)
239 {
240 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
241 	arc_flags_t aflags = ARC_FLAG_WAIT;
242 	arc_buf_t *abuf = NULL;
243 	zbookmark_phys_t zb;
244 	int error;
245 
246 	if (zilog->zl_header->zh_claim_txg == 0)
247 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
248 
249 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
250 		zio_flags |= ZIO_FLAG_SPECULATIVE;
251 
252 	if (!decrypt)
253 		zio_flags |= ZIO_FLAG_RAW;
254 
255 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
256 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
257 
258 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
259 	    &abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
260 
261 	if (error == 0) {
262 		zio_cksum_t cksum = bp->blk_cksum;
263 
264 		/*
265 		 * Validate the checksummed log block.
266 		 *
267 		 * Sequence numbers should be... sequential.  The checksum
268 		 * verifier for the next block should be bp's checksum plus 1.
269 		 *
270 		 * Also check the log chain linkage and size used.
271 		 */
272 		cksum.zc_word[ZIL_ZC_SEQ]++;
273 
274 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
275 			zil_chain_t *zilc = abuf->b_data;
276 			char *lr = (char *)(zilc + 1);
277 			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
278 
279 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
280 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
281 				error = SET_ERROR(ECKSUM);
282 			} else {
283 				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
284 				memcpy(dst, lr, len);
285 				*end = (char *)dst + len;
286 				*nbp = zilc->zc_next_blk;
287 			}
288 		} else {
289 			char *lr = abuf->b_data;
290 			uint64_t size = BP_GET_LSIZE(bp);
291 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
292 
293 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
294 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
295 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
296 				error = SET_ERROR(ECKSUM);
297 			} else {
298 				ASSERT3U(zilc->zc_nused, <=,
299 				    SPA_OLD_MAXBLOCKSIZE);
300 				memcpy(dst, lr, zilc->zc_nused);
301 				*end = (char *)dst + zilc->zc_nused;
302 				*nbp = zilc->zc_next_blk;
303 			}
304 		}
305 
306 		arc_buf_destroy(abuf, &abuf);
307 	}
308 
309 	return (error);
310 }
311 
312 /*
313  * Read a TX_WRITE log data block.
314  */
315 static int
316 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
317 {
318 	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
319 	const blkptr_t *bp = &lr->lr_blkptr;
320 	arc_flags_t aflags = ARC_FLAG_WAIT;
321 	arc_buf_t *abuf = NULL;
322 	zbookmark_phys_t zb;
323 	int error;
324 
325 	if (BP_IS_HOLE(bp)) {
326 		if (wbuf != NULL)
327 			memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length));
328 		return (0);
329 	}
330 
331 	if (zilog->zl_header->zh_claim_txg == 0)
332 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
333 
334 	/*
335 	 * If we are not using the resulting data, we are just checking that
336 	 * it hasn't been corrupted so we don't need to waste CPU time
337 	 * decompressing and decrypting it.
338 	 */
339 	if (wbuf == NULL)
340 		zio_flags |= ZIO_FLAG_RAW;
341 
342 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
343 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
344 
345 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
346 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
347 
348 	if (error == 0) {
349 		if (wbuf != NULL)
350 			memcpy(wbuf, abuf->b_data, arc_buf_size(abuf));
351 		arc_buf_destroy(abuf, &abuf);
352 	}
353 
354 	return (error);
355 }
356 
357 void
358 zil_sums_init(zil_sums_t *zs)
359 {
360 	wmsum_init(&zs->zil_commit_count, 0);
361 	wmsum_init(&zs->zil_commit_writer_count, 0);
362 	wmsum_init(&zs->zil_itx_count, 0);
363 	wmsum_init(&zs->zil_itx_indirect_count, 0);
364 	wmsum_init(&zs->zil_itx_indirect_bytes, 0);
365 	wmsum_init(&zs->zil_itx_copied_count, 0);
366 	wmsum_init(&zs->zil_itx_copied_bytes, 0);
367 	wmsum_init(&zs->zil_itx_needcopy_count, 0);
368 	wmsum_init(&zs->zil_itx_needcopy_bytes, 0);
369 	wmsum_init(&zs->zil_itx_metaslab_normal_count, 0);
370 	wmsum_init(&zs->zil_itx_metaslab_normal_bytes, 0);
371 	wmsum_init(&zs->zil_itx_metaslab_slog_count, 0);
372 	wmsum_init(&zs->zil_itx_metaslab_slog_bytes, 0);
373 }
374 
375 void
376 zil_sums_fini(zil_sums_t *zs)
377 {
378 	wmsum_fini(&zs->zil_commit_count);
379 	wmsum_fini(&zs->zil_commit_writer_count);
380 	wmsum_fini(&zs->zil_itx_count);
381 	wmsum_fini(&zs->zil_itx_indirect_count);
382 	wmsum_fini(&zs->zil_itx_indirect_bytes);
383 	wmsum_fini(&zs->zil_itx_copied_count);
384 	wmsum_fini(&zs->zil_itx_copied_bytes);
385 	wmsum_fini(&zs->zil_itx_needcopy_count);
386 	wmsum_fini(&zs->zil_itx_needcopy_bytes);
387 	wmsum_fini(&zs->zil_itx_metaslab_normal_count);
388 	wmsum_fini(&zs->zil_itx_metaslab_normal_bytes);
389 	wmsum_fini(&zs->zil_itx_metaslab_slog_count);
390 	wmsum_fini(&zs->zil_itx_metaslab_slog_bytes);
391 }
392 
393 void
394 zil_kstat_values_update(zil_kstat_values_t *zs, zil_sums_t *zil_sums)
395 {
396 	zs->zil_commit_count.value.ui64 =
397 	    wmsum_value(&zil_sums->zil_commit_count);
398 	zs->zil_commit_writer_count.value.ui64 =
399 	    wmsum_value(&zil_sums->zil_commit_writer_count);
400 	zs->zil_itx_count.value.ui64 =
401 	    wmsum_value(&zil_sums->zil_itx_count);
402 	zs->zil_itx_indirect_count.value.ui64 =
403 	    wmsum_value(&zil_sums->zil_itx_indirect_count);
404 	zs->zil_itx_indirect_bytes.value.ui64 =
405 	    wmsum_value(&zil_sums->zil_itx_indirect_bytes);
406 	zs->zil_itx_copied_count.value.ui64 =
407 	    wmsum_value(&zil_sums->zil_itx_copied_count);
408 	zs->zil_itx_copied_bytes.value.ui64 =
409 	    wmsum_value(&zil_sums->zil_itx_copied_bytes);
410 	zs->zil_itx_needcopy_count.value.ui64 =
411 	    wmsum_value(&zil_sums->zil_itx_needcopy_count);
412 	zs->zil_itx_needcopy_bytes.value.ui64 =
413 	    wmsum_value(&zil_sums->zil_itx_needcopy_bytes);
414 	zs->zil_itx_metaslab_normal_count.value.ui64 =
415 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_count);
416 	zs->zil_itx_metaslab_normal_bytes.value.ui64 =
417 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_bytes);
418 	zs->zil_itx_metaslab_slog_count.value.ui64 =
419 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_count);
420 	zs->zil_itx_metaslab_slog_bytes.value.ui64 =
421 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_bytes);
422 }
423 
424 /*
425  * Parse the intent log, and call parse_func for each valid record within.
426  */
427 int
428 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
429     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
430     boolean_t decrypt)
431 {
432 	const zil_header_t *zh = zilog->zl_header;
433 	boolean_t claimed = !!zh->zh_claim_txg;
434 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
435 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
436 	uint64_t max_blk_seq = 0;
437 	uint64_t max_lr_seq = 0;
438 	uint64_t blk_count = 0;
439 	uint64_t lr_count = 0;
440 	blkptr_t blk, next_blk = {{{{0}}}};
441 	char *lrbuf, *lrp;
442 	int error = 0;
443 
444 	/*
445 	 * Old logs didn't record the maximum zh_claim_lr_seq.
446 	 */
447 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
448 		claim_lr_seq = UINT64_MAX;
449 
450 	/*
451 	 * Starting at the block pointed to by zh_log we read the log chain.
452 	 * For each block in the chain we strongly check that block to
453 	 * ensure its validity.  We stop when an invalid block is found.
454 	 * For each block pointer in the chain we call parse_blk_func().
455 	 * For each record in each valid block we call parse_lr_func().
456 	 * If the log has been claimed, stop if we encounter a sequence
457 	 * number greater than the highest claimed sequence number.
458 	 */
459 	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
460 	zil_bp_tree_init(zilog);
461 
462 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
463 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
464 		int reclen;
465 		char *end = NULL;
466 
467 		if (blk_seq > claim_blk_seq)
468 			break;
469 
470 		error = parse_blk_func(zilog, &blk, arg, txg);
471 		if (error != 0)
472 			break;
473 		ASSERT3U(max_blk_seq, <, blk_seq);
474 		max_blk_seq = blk_seq;
475 		blk_count++;
476 
477 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
478 			break;
479 
480 		error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
481 		    lrbuf, &end);
482 		if (error != 0)
483 			break;
484 
485 		for (lrp = lrbuf; lrp < end; lrp += reclen) {
486 			lr_t *lr = (lr_t *)lrp;
487 			reclen = lr->lrc_reclen;
488 			ASSERT3U(reclen, >=, sizeof (lr_t));
489 			if (lr->lrc_seq > claim_lr_seq)
490 				goto done;
491 
492 			error = parse_lr_func(zilog, lr, arg, txg);
493 			if (error != 0)
494 				goto done;
495 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
496 			max_lr_seq = lr->lrc_seq;
497 			lr_count++;
498 		}
499 	}
500 done:
501 	zilog->zl_parse_error = error;
502 	zilog->zl_parse_blk_seq = max_blk_seq;
503 	zilog->zl_parse_lr_seq = max_lr_seq;
504 	zilog->zl_parse_blk_count = blk_count;
505 	zilog->zl_parse_lr_count = lr_count;
506 
507 	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
508 	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq) ||
509 	    (decrypt && error == EIO));
510 
511 	zil_bp_tree_fini(zilog);
512 	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
513 
514 	return (error);
515 }
516 
517 static int
518 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
519     uint64_t first_txg)
520 {
521 	(void) tx;
522 	ASSERT(!BP_IS_HOLE(bp));
523 
524 	/*
525 	 * As we call this function from the context of a rewind to a
526 	 * checkpoint, each ZIL block whose txg is later than the txg
527 	 * that we rewind to is invalid. Thus, we return -1 so
528 	 * zil_parse() doesn't attempt to read it.
529 	 */
530 	if (bp->blk_birth >= first_txg)
531 		return (-1);
532 
533 	if (zil_bp_tree_add(zilog, bp) != 0)
534 		return (0);
535 
536 	zio_free(zilog->zl_spa, first_txg, bp);
537 	return (0);
538 }
539 
540 static int
541 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
542     uint64_t first_txg)
543 {
544 	(void) zilog, (void) lrc, (void) tx, (void) first_txg;
545 	return (0);
546 }
547 
548 static int
549 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
550     uint64_t first_txg)
551 {
552 	/*
553 	 * Claim log block if not already committed and not already claimed.
554 	 * If tx == NULL, just verify that the block is claimable.
555 	 */
556 	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
557 	    zil_bp_tree_add(zilog, bp) != 0)
558 		return (0);
559 
560 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
561 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
562 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
563 }
564 
565 static int
566 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
567     uint64_t first_txg)
568 {
569 	lr_write_t *lr = (lr_write_t *)lrc;
570 	int error;
571 
572 	if (lrc->lrc_txtype != TX_WRITE)
573 		return (0);
574 
575 	/*
576 	 * If the block is not readable, don't claim it.  This can happen
577 	 * in normal operation when a log block is written to disk before
578 	 * some of the dmu_sync() blocks it points to.  In this case, the
579 	 * transaction cannot have been committed to anyone (we would have
580 	 * waited for all writes to be stable first), so it is semantically
581 	 * correct to declare this the end of the log.
582 	 */
583 	if (lr->lr_blkptr.blk_birth >= first_txg) {
584 		error = zil_read_log_data(zilog, lr, NULL);
585 		if (error != 0)
586 			return (error);
587 	}
588 
589 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
590 }
591 
592 static int
593 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
594     uint64_t claim_txg)
595 {
596 	(void) claim_txg;
597 
598 	zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
599 
600 	return (0);
601 }
602 
603 static int
604 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
605     uint64_t claim_txg)
606 {
607 	lr_write_t *lr = (lr_write_t *)lrc;
608 	blkptr_t *bp = &lr->lr_blkptr;
609 
610 	/*
611 	 * If we previously claimed it, we need to free it.
612 	 */
613 	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
614 	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
615 	    !BP_IS_HOLE(bp))
616 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
617 
618 	return (0);
619 }
620 
621 static int
622 zil_lwb_vdev_compare(const void *x1, const void *x2)
623 {
624 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
625 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
626 
627 	return (TREE_CMP(v1, v2));
628 }
629 
630 static lwb_t *
631 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg,
632     boolean_t fastwrite)
633 {
634 	lwb_t *lwb;
635 
636 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
637 	lwb->lwb_zilog = zilog;
638 	lwb->lwb_blk = *bp;
639 	lwb->lwb_fastwrite = fastwrite;
640 	lwb->lwb_slog = slog;
641 	lwb->lwb_state = LWB_STATE_CLOSED;
642 	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
643 	lwb->lwb_max_txg = txg;
644 	lwb->lwb_write_zio = NULL;
645 	lwb->lwb_root_zio = NULL;
646 	lwb->lwb_issued_timestamp = 0;
647 	lwb->lwb_issued_txg = 0;
648 	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
649 		lwb->lwb_nused = sizeof (zil_chain_t);
650 		lwb->lwb_sz = BP_GET_LSIZE(bp);
651 	} else {
652 		lwb->lwb_nused = 0;
653 		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
654 	}
655 
656 	mutex_enter(&zilog->zl_lock);
657 	list_insert_tail(&zilog->zl_lwb_list, lwb);
658 	mutex_exit(&zilog->zl_lock);
659 
660 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
661 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
662 	VERIFY(list_is_empty(&lwb->lwb_waiters));
663 	VERIFY(list_is_empty(&lwb->lwb_itxs));
664 
665 	return (lwb);
666 }
667 
668 static void
669 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
670 {
671 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
672 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
673 	VERIFY(list_is_empty(&lwb->lwb_waiters));
674 	VERIFY(list_is_empty(&lwb->lwb_itxs));
675 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
676 	ASSERT3P(lwb->lwb_write_zio, ==, NULL);
677 	ASSERT3P(lwb->lwb_root_zio, ==, NULL);
678 	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
679 	ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
680 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
681 
682 	/*
683 	 * Clear the zilog's field to indicate this lwb is no longer
684 	 * valid, and prevent use-after-free errors.
685 	 */
686 	if (zilog->zl_last_lwb_opened == lwb)
687 		zilog->zl_last_lwb_opened = NULL;
688 
689 	kmem_cache_free(zil_lwb_cache, lwb);
690 }
691 
692 /*
693  * Called when we create in-memory log transactions so that we know
694  * to cleanup the itxs at the end of spa_sync().
695  */
696 static void
697 zilog_dirty(zilog_t *zilog, uint64_t txg)
698 {
699 	dsl_pool_t *dp = zilog->zl_dmu_pool;
700 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
701 
702 	ASSERT(spa_writeable(zilog->zl_spa));
703 
704 	if (ds->ds_is_snapshot)
705 		panic("dirtying snapshot!");
706 
707 	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
708 		/* up the hold count until we can be written out */
709 		dmu_buf_add_ref(ds->ds_dbuf, zilog);
710 
711 		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
712 	}
713 }
714 
715 /*
716  * Determine if the zil is dirty in the specified txg. Callers wanting to
717  * ensure that the dirty state does not change must hold the itxg_lock for
718  * the specified txg. Holding the lock will ensure that the zil cannot be
719  * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
720  * state.
721  */
722 static boolean_t __maybe_unused
723 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
724 {
725 	dsl_pool_t *dp = zilog->zl_dmu_pool;
726 
727 	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
728 		return (B_TRUE);
729 	return (B_FALSE);
730 }
731 
732 /*
733  * Determine if the zil is dirty. The zil is considered dirty if it has
734  * any pending itx records that have not been cleaned by zil_clean().
735  */
736 static boolean_t
737 zilog_is_dirty(zilog_t *zilog)
738 {
739 	dsl_pool_t *dp = zilog->zl_dmu_pool;
740 
741 	for (int t = 0; t < TXG_SIZE; t++) {
742 		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
743 			return (B_TRUE);
744 	}
745 	return (B_FALSE);
746 }
747 
748 /*
749  * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
750  * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
751  * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
752  * zil_commit.
753  */
754 static void
755 zil_commit_activate_saxattr_feature(zilog_t *zilog)
756 {
757 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
758 	uint64_t txg = 0;
759 	dmu_tx_t *tx = NULL;
760 
761 	if (spa_feature_is_enabled(zilog->zl_spa,
762 	    SPA_FEATURE_ZILSAXATTR) &&
763 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL &&
764 	    !dsl_dataset_feature_is_active(ds,
765 	    SPA_FEATURE_ZILSAXATTR)) {
766 		tx = dmu_tx_create(zilog->zl_os);
767 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
768 		dsl_dataset_dirty(ds, tx);
769 		txg = dmu_tx_get_txg(tx);
770 
771 		mutex_enter(&ds->ds_lock);
772 		ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
773 		    (void *)B_TRUE;
774 		mutex_exit(&ds->ds_lock);
775 		dmu_tx_commit(tx);
776 		txg_wait_synced(zilog->zl_dmu_pool, txg);
777 	}
778 }
779 
780 /*
781  * Create an on-disk intent log.
782  */
783 static lwb_t *
784 zil_create(zilog_t *zilog)
785 {
786 	const zil_header_t *zh = zilog->zl_header;
787 	lwb_t *lwb = NULL;
788 	uint64_t txg = 0;
789 	dmu_tx_t *tx = NULL;
790 	blkptr_t blk;
791 	int error = 0;
792 	boolean_t fastwrite = FALSE;
793 	boolean_t slog = FALSE;
794 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
795 
796 
797 	/*
798 	 * Wait for any previous destroy to complete.
799 	 */
800 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
801 
802 	ASSERT(zh->zh_claim_txg == 0);
803 	ASSERT(zh->zh_replay_seq == 0);
804 
805 	blk = zh->zh_log;
806 
807 	/*
808 	 * Allocate an initial log block if:
809 	 *    - there isn't one already
810 	 *    - the existing block is the wrong endianness
811 	 */
812 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
813 		tx = dmu_tx_create(zilog->zl_os);
814 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
815 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
816 		txg = dmu_tx_get_txg(tx);
817 
818 		if (!BP_IS_HOLE(&blk)) {
819 			zio_free(zilog->zl_spa, txg, &blk);
820 			BP_ZERO(&blk);
821 		}
822 
823 		error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
824 		    ZIL_MIN_BLKSZ, &slog);
825 		fastwrite = TRUE;
826 
827 		if (error == 0)
828 			zil_init_log_chain(zilog, &blk);
829 	}
830 
831 	/*
832 	 * Allocate a log write block (lwb) for the first log block.
833 	 */
834 	if (error == 0)
835 		lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
836 
837 	/*
838 	 * If we just allocated the first log block, commit our transaction
839 	 * and wait for zil_sync() to stuff the block pointer into zh_log.
840 	 * (zh is part of the MOS, so we cannot modify it in open context.)
841 	 */
842 	if (tx != NULL) {
843 		/*
844 		 * If "zilsaxattr" feature is enabled on zpool, then activate
845 		 * it now when we're creating the ZIL chain. We can't wait with
846 		 * this until we write the first xattr log record because we
847 		 * need to wait for the feature activation to sync out.
848 		 */
849 		if (spa_feature_is_enabled(zilog->zl_spa,
850 		    SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) !=
851 		    DMU_OST_ZVOL) {
852 			mutex_enter(&ds->ds_lock);
853 			ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
854 			    (void *)B_TRUE;
855 			mutex_exit(&ds->ds_lock);
856 		}
857 
858 		dmu_tx_commit(tx);
859 		txg_wait_synced(zilog->zl_dmu_pool, txg);
860 	} else {
861 		/*
862 		 * This branch covers the case where we enable the feature on a
863 		 * zpool that has existing ZIL headers.
864 		 */
865 		zil_commit_activate_saxattr_feature(zilog);
866 	}
867 	IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
868 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL,
869 	    dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR));
870 
871 	ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
872 	IMPLY(error == 0, lwb != NULL);
873 
874 	return (lwb);
875 }
876 
877 /*
878  * In one tx, free all log blocks and clear the log header. If keep_first
879  * is set, then we're replaying a log with no content. We want to keep the
880  * first block, however, so that the first synchronous transaction doesn't
881  * require a txg_wait_synced() in zil_create(). We don't need to
882  * txg_wait_synced() here either when keep_first is set, because both
883  * zil_create() and zil_destroy() will wait for any in-progress destroys
884  * to complete.
885  */
886 void
887 zil_destroy(zilog_t *zilog, boolean_t keep_first)
888 {
889 	const zil_header_t *zh = zilog->zl_header;
890 	lwb_t *lwb;
891 	dmu_tx_t *tx;
892 	uint64_t txg;
893 
894 	/*
895 	 * Wait for any previous destroy to complete.
896 	 */
897 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
898 
899 	zilog->zl_old_header = *zh;		/* debugging aid */
900 
901 	if (BP_IS_HOLE(&zh->zh_log))
902 		return;
903 
904 	tx = dmu_tx_create(zilog->zl_os);
905 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
906 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
907 	txg = dmu_tx_get_txg(tx);
908 
909 	mutex_enter(&zilog->zl_lock);
910 
911 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
912 	zilog->zl_destroy_txg = txg;
913 	zilog->zl_keep_first = keep_first;
914 
915 	if (!list_is_empty(&zilog->zl_lwb_list)) {
916 		ASSERT(zh->zh_claim_txg == 0);
917 		VERIFY(!keep_first);
918 		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
919 			if (lwb->lwb_fastwrite)
920 				metaslab_fastwrite_unmark(zilog->zl_spa,
921 				    &lwb->lwb_blk);
922 
923 			list_remove(&zilog->zl_lwb_list, lwb);
924 			if (lwb->lwb_buf != NULL)
925 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
926 			zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
927 			zil_free_lwb(zilog, lwb);
928 		}
929 	} else if (!keep_first) {
930 		zil_destroy_sync(zilog, tx);
931 	}
932 	mutex_exit(&zilog->zl_lock);
933 
934 	dmu_tx_commit(tx);
935 }
936 
937 void
938 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
939 {
940 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
941 	(void) zil_parse(zilog, zil_free_log_block,
942 	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
943 }
944 
945 int
946 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
947 {
948 	dmu_tx_t *tx = txarg;
949 	zilog_t *zilog;
950 	uint64_t first_txg;
951 	zil_header_t *zh;
952 	objset_t *os;
953 	int error;
954 
955 	error = dmu_objset_own_obj(dp, ds->ds_object,
956 	    DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
957 	if (error != 0) {
958 		/*
959 		 * EBUSY indicates that the objset is inconsistent, in which
960 		 * case it can not have a ZIL.
961 		 */
962 		if (error != EBUSY) {
963 			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
964 			    (unsigned long long)ds->ds_object, error);
965 		}
966 
967 		return (0);
968 	}
969 
970 	zilog = dmu_objset_zil(os);
971 	zh = zil_header_in_syncing_context(zilog);
972 	ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
973 	first_txg = spa_min_claim_txg(zilog->zl_spa);
974 
975 	/*
976 	 * If the spa_log_state is not set to be cleared, check whether
977 	 * the current uberblock is a checkpoint one and if the current
978 	 * header has been claimed before moving on.
979 	 *
980 	 * If the current uberblock is a checkpointed uberblock then
981 	 * one of the following scenarios took place:
982 	 *
983 	 * 1] We are currently rewinding to the checkpoint of the pool.
984 	 * 2] We crashed in the middle of a checkpoint rewind but we
985 	 *    did manage to write the checkpointed uberblock to the
986 	 *    vdev labels, so when we tried to import the pool again
987 	 *    the checkpointed uberblock was selected from the import
988 	 *    procedure.
989 	 *
990 	 * In both cases we want to zero out all the ZIL blocks, except
991 	 * the ones that have been claimed at the time of the checkpoint
992 	 * (their zh_claim_txg != 0). The reason is that these blocks
993 	 * may be corrupted since we may have reused their locations on
994 	 * disk after we took the checkpoint.
995 	 *
996 	 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
997 	 * when we first figure out whether the current uberblock is
998 	 * checkpointed or not. Unfortunately, that would discard all
999 	 * the logs, including the ones that are claimed, and we would
1000 	 * leak space.
1001 	 */
1002 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
1003 	    (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1004 	    zh->zh_claim_txg == 0)) {
1005 		if (!BP_IS_HOLE(&zh->zh_log)) {
1006 			(void) zil_parse(zilog, zil_clear_log_block,
1007 			    zil_noop_log_record, tx, first_txg, B_FALSE);
1008 		}
1009 		BP_ZERO(&zh->zh_log);
1010 		if (os->os_encrypted)
1011 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1012 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1013 		dmu_objset_disown(os, B_FALSE, FTAG);
1014 		return (0);
1015 	}
1016 
1017 	/*
1018 	 * If we are not rewinding and opening the pool normally, then
1019 	 * the min_claim_txg should be equal to the first txg of the pool.
1020 	 */
1021 	ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
1022 
1023 	/*
1024 	 * Claim all log blocks if we haven't already done so, and remember
1025 	 * the highest claimed sequence number.  This ensures that if we can
1026 	 * read only part of the log now (e.g. due to a missing device),
1027 	 * but we can read the entire log later, we will not try to replay
1028 	 * or destroy beyond the last block we successfully claimed.
1029 	 */
1030 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
1031 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
1032 		(void) zil_parse(zilog, zil_claim_log_block,
1033 		    zil_claim_log_record, tx, first_txg, B_FALSE);
1034 		zh->zh_claim_txg = first_txg;
1035 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
1036 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
1037 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
1038 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
1039 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
1040 		if (os->os_encrypted)
1041 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1042 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1043 	}
1044 
1045 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
1046 	dmu_objset_disown(os, B_FALSE, FTAG);
1047 	return (0);
1048 }
1049 
1050 /*
1051  * Check the log by walking the log chain.
1052  * Checksum errors are ok as they indicate the end of the chain.
1053  * Any other error (no device or read failure) returns an error.
1054  */
1055 int
1056 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
1057 {
1058 	(void) dp;
1059 	zilog_t *zilog;
1060 	objset_t *os;
1061 	blkptr_t *bp;
1062 	int error;
1063 
1064 	ASSERT(tx == NULL);
1065 
1066 	error = dmu_objset_from_ds(ds, &os);
1067 	if (error != 0) {
1068 		cmn_err(CE_WARN, "can't open objset %llu, error %d",
1069 		    (unsigned long long)ds->ds_object, error);
1070 		return (0);
1071 	}
1072 
1073 	zilog = dmu_objset_zil(os);
1074 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
1075 
1076 	if (!BP_IS_HOLE(bp)) {
1077 		vdev_t *vd;
1078 		boolean_t valid = B_TRUE;
1079 
1080 		/*
1081 		 * Check the first block and determine if it's on a log device
1082 		 * which may have been removed or faulted prior to loading this
1083 		 * pool.  If so, there's no point in checking the rest of the
1084 		 * log as its content should have already been synced to the
1085 		 * pool.
1086 		 */
1087 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
1088 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
1089 		if (vd->vdev_islog && vdev_is_dead(vd))
1090 			valid = vdev_log_state_valid(vd);
1091 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
1092 
1093 		if (!valid)
1094 			return (0);
1095 
1096 		/*
1097 		 * Check whether the current uberblock is checkpointed (e.g.
1098 		 * we are rewinding) and whether the current header has been
1099 		 * claimed or not. If it hasn't then skip verifying it. We
1100 		 * do this because its ZIL blocks may be part of the pool's
1101 		 * state before the rewind, which is no longer valid.
1102 		 */
1103 		zil_header_t *zh = zil_header_in_syncing_context(zilog);
1104 		if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1105 		    zh->zh_claim_txg == 0)
1106 			return (0);
1107 	}
1108 
1109 	/*
1110 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
1111 	 * any blocks, but just determine whether it is possible to do so.
1112 	 * In addition to checking the log chain, zil_claim_log_block()
1113 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
1114 	 * which will update spa_max_claim_txg.  See spa_load() for details.
1115 	 */
1116 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
1117 	    zilog->zl_header->zh_claim_txg ? -1ULL :
1118 	    spa_min_claim_txg(os->os_spa), B_FALSE);
1119 
1120 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
1121 }
1122 
1123 /*
1124  * When an itx is "skipped", this function is used to properly mark the
1125  * waiter as "done, and signal any thread(s) waiting on it. An itx can
1126  * be skipped (and not committed to an lwb) for a variety of reasons,
1127  * one of them being that the itx was committed via spa_sync(), prior to
1128  * it being committed to an lwb; this can happen if a thread calling
1129  * zil_commit() is racing with spa_sync().
1130  */
1131 static void
1132 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
1133 {
1134 	mutex_enter(&zcw->zcw_lock);
1135 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1136 	zcw->zcw_done = B_TRUE;
1137 	cv_broadcast(&zcw->zcw_cv);
1138 	mutex_exit(&zcw->zcw_lock);
1139 }
1140 
1141 /*
1142  * This function is used when the given waiter is to be linked into an
1143  * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1144  * At this point, the waiter will no longer be referenced by the itx,
1145  * and instead, will be referenced by the lwb.
1146  */
1147 static void
1148 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1149 {
1150 	/*
1151 	 * The lwb_waiters field of the lwb is protected by the zilog's
1152 	 * zl_lock, thus it must be held when calling this function.
1153 	 */
1154 	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1155 
1156 	mutex_enter(&zcw->zcw_lock);
1157 	ASSERT(!list_link_active(&zcw->zcw_node));
1158 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1159 	ASSERT3P(lwb, !=, NULL);
1160 	ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
1161 	    lwb->lwb_state == LWB_STATE_ISSUED ||
1162 	    lwb->lwb_state == LWB_STATE_WRITE_DONE);
1163 
1164 	list_insert_tail(&lwb->lwb_waiters, zcw);
1165 	zcw->zcw_lwb = lwb;
1166 	mutex_exit(&zcw->zcw_lock);
1167 }
1168 
1169 /*
1170  * This function is used when zio_alloc_zil() fails to allocate a ZIL
1171  * block, and the given waiter must be linked to the "nolwb waiters"
1172  * list inside of zil_process_commit_list().
1173  */
1174 static void
1175 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1176 {
1177 	mutex_enter(&zcw->zcw_lock);
1178 	ASSERT(!list_link_active(&zcw->zcw_node));
1179 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1180 	list_insert_tail(nolwb, zcw);
1181 	mutex_exit(&zcw->zcw_lock);
1182 }
1183 
1184 void
1185 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1186 {
1187 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1188 	avl_index_t where;
1189 	zil_vdev_node_t *zv, zvsearch;
1190 	int ndvas = BP_GET_NDVAS(bp);
1191 	int i;
1192 
1193 	if (zil_nocacheflush)
1194 		return;
1195 
1196 	mutex_enter(&lwb->lwb_vdev_lock);
1197 	for (i = 0; i < ndvas; i++) {
1198 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1199 		if (avl_find(t, &zvsearch, &where) == NULL) {
1200 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1201 			zv->zv_vdev = zvsearch.zv_vdev;
1202 			avl_insert(t, zv, where);
1203 		}
1204 	}
1205 	mutex_exit(&lwb->lwb_vdev_lock);
1206 }
1207 
1208 static void
1209 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1210 {
1211 	avl_tree_t *src = &lwb->lwb_vdev_tree;
1212 	avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1213 	void *cookie = NULL;
1214 	zil_vdev_node_t *zv;
1215 
1216 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1217 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1218 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1219 
1220 	/*
1221 	 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1222 	 * not need the protection of lwb_vdev_lock (it will only be modified
1223 	 * while holding zilog->zl_lock) as its writes and those of its
1224 	 * children have all completed.  The younger 'nlwb' may be waiting on
1225 	 * future writes to additional vdevs.
1226 	 */
1227 	mutex_enter(&nlwb->lwb_vdev_lock);
1228 	/*
1229 	 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1230 	 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1231 	 */
1232 	while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1233 		avl_index_t where;
1234 
1235 		if (avl_find(dst, zv, &where) == NULL) {
1236 			avl_insert(dst, zv, where);
1237 		} else {
1238 			kmem_free(zv, sizeof (*zv));
1239 		}
1240 	}
1241 	mutex_exit(&nlwb->lwb_vdev_lock);
1242 }
1243 
1244 void
1245 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1246 {
1247 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1248 }
1249 
1250 /*
1251  * This function is a called after all vdevs associated with a given lwb
1252  * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1253  * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1254  * all "previous" lwb's will have completed before this function is
1255  * called; i.e. this function is called for all previous lwbs before
1256  * it's called for "this" lwb (enforced via zio the dependencies
1257  * configured in zil_lwb_set_zio_dependency()).
1258  *
1259  * The intention is for this function to be called as soon as the
1260  * contents of an lwb are considered "stable" on disk, and will survive
1261  * any sudden loss of power. At this point, any threads waiting for the
1262  * lwb to reach this state are signalled, and the "waiter" structures
1263  * are marked "done".
1264  */
1265 static void
1266 zil_lwb_flush_vdevs_done(zio_t *zio)
1267 {
1268 	lwb_t *lwb = zio->io_private;
1269 	zilog_t *zilog = lwb->lwb_zilog;
1270 	zil_commit_waiter_t *zcw;
1271 	itx_t *itx;
1272 	uint64_t txg;
1273 
1274 	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1275 
1276 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1277 
1278 	mutex_enter(&zilog->zl_lock);
1279 
1280 	/*
1281 	 * If we have had an allocation failure and the txg is
1282 	 * waiting to sync then we want zil_sync() to remove the lwb so
1283 	 * that it's not picked up as the next new one in
1284 	 * zil_process_commit_list(). zil_sync() will only remove the
1285 	 * lwb if lwb_buf is null.
1286 	 */
1287 	lwb->lwb_buf = NULL;
1288 
1289 	ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1290 	zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1291 
1292 	lwb->lwb_root_zio = NULL;
1293 
1294 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1295 	lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1296 
1297 	if (zilog->zl_last_lwb_opened == lwb) {
1298 		/*
1299 		 * Remember the highest committed log sequence number
1300 		 * for ztest. We only update this value when all the log
1301 		 * writes succeeded, because ztest wants to ASSERT that
1302 		 * it got the whole log chain.
1303 		 */
1304 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1305 	}
1306 
1307 	while ((itx = list_head(&lwb->lwb_itxs)) != NULL) {
1308 		list_remove(&lwb->lwb_itxs, itx);
1309 		zil_itx_destroy(itx);
1310 	}
1311 
1312 	while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1313 		mutex_enter(&zcw->zcw_lock);
1314 
1315 		ASSERT(list_link_active(&zcw->zcw_node));
1316 		list_remove(&lwb->lwb_waiters, zcw);
1317 
1318 		ASSERT3P(zcw->zcw_lwb, ==, lwb);
1319 		zcw->zcw_lwb = NULL;
1320 		/*
1321 		 * We expect any ZIO errors from child ZIOs to have been
1322 		 * propagated "up" to this specific LWB's root ZIO, in
1323 		 * order for this error handling to work correctly. This
1324 		 * includes ZIO errors from either this LWB's write or
1325 		 * flush, as well as any errors from other dependent LWBs
1326 		 * (e.g. a root LWB ZIO that might be a child of this LWB).
1327 		 *
1328 		 * With that said, it's important to note that LWB flush
1329 		 * errors are not propagated up to the LWB root ZIO.
1330 		 * This is incorrect behavior, and results in VDEV flush
1331 		 * errors not being handled correctly here. See the
1332 		 * comment above the call to "zio_flush" for details.
1333 		 */
1334 
1335 		zcw->zcw_zio_error = zio->io_error;
1336 
1337 		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1338 		zcw->zcw_done = B_TRUE;
1339 		cv_broadcast(&zcw->zcw_cv);
1340 
1341 		mutex_exit(&zcw->zcw_lock);
1342 	}
1343 
1344 	mutex_exit(&zilog->zl_lock);
1345 
1346 	mutex_enter(&zilog->zl_lwb_io_lock);
1347 	txg = lwb->lwb_issued_txg;
1348 	ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0);
1349 	zilog->zl_lwb_inflight[txg & TXG_MASK]--;
1350 	if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0)
1351 		cv_broadcast(&zilog->zl_lwb_io_cv);
1352 	mutex_exit(&zilog->zl_lwb_io_lock);
1353 }
1354 
1355 /*
1356  * Wait for the completion of all issued write/flush of that txg provided.
1357  * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1358  */
1359 static void
1360 zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg)
1361 {
1362 	ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa));
1363 
1364 	mutex_enter(&zilog->zl_lwb_io_lock);
1365 	while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0)
1366 		cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock);
1367 	mutex_exit(&zilog->zl_lwb_io_lock);
1368 
1369 #ifdef ZFS_DEBUG
1370 	mutex_enter(&zilog->zl_lock);
1371 	mutex_enter(&zilog->zl_lwb_io_lock);
1372 	lwb_t *lwb = list_head(&zilog->zl_lwb_list);
1373 	while (lwb != NULL && lwb->lwb_max_txg <= txg) {
1374 		if (lwb->lwb_issued_txg <= txg) {
1375 			ASSERT(lwb->lwb_state != LWB_STATE_ISSUED);
1376 			ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE);
1377 			IMPLY(lwb->lwb_issued_txg > 0,
1378 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
1379 		}
1380 		IMPLY(lwb->lwb_state == LWB_STATE_FLUSH_DONE,
1381 		    lwb->lwb_buf == NULL);
1382 		lwb = list_next(&zilog->zl_lwb_list, lwb);
1383 	}
1384 	mutex_exit(&zilog->zl_lwb_io_lock);
1385 	mutex_exit(&zilog->zl_lock);
1386 #endif
1387 }
1388 
1389 /*
1390  * This is called when an lwb's write zio completes. The callback's
1391  * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1392  * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1393  * in writing out this specific lwb's data, and in the case that cache
1394  * flushes have been deferred, vdevs involved in writing the data for
1395  * previous lwbs. The writes corresponding to all the vdevs in the
1396  * lwb_vdev_tree will have completed by the time this is called, due to
1397  * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1398  * which takes deferred flushes into account. The lwb will be "done"
1399  * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1400  * completion callback for the lwb's root zio.
1401  */
1402 static void
1403 zil_lwb_write_done(zio_t *zio)
1404 {
1405 	lwb_t *lwb = zio->io_private;
1406 	spa_t *spa = zio->io_spa;
1407 	zilog_t *zilog = lwb->lwb_zilog;
1408 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1409 	void *cookie = NULL;
1410 	zil_vdev_node_t *zv;
1411 	lwb_t *nlwb;
1412 
1413 	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1414 
1415 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1416 	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1417 	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1418 	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1419 	ASSERT(!BP_IS_GANG(zio->io_bp));
1420 	ASSERT(!BP_IS_HOLE(zio->io_bp));
1421 	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1422 
1423 	abd_free(zio->io_abd);
1424 
1425 	mutex_enter(&zilog->zl_lock);
1426 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1427 	lwb->lwb_state = LWB_STATE_WRITE_DONE;
1428 	lwb->lwb_write_zio = NULL;
1429 	lwb->lwb_fastwrite = FALSE;
1430 	nlwb = list_next(&zilog->zl_lwb_list, lwb);
1431 	mutex_exit(&zilog->zl_lock);
1432 
1433 	if (avl_numnodes(t) == 0)
1434 		return;
1435 
1436 	/*
1437 	 * If there was an IO error, we're not going to call zio_flush()
1438 	 * on these vdevs, so we simply empty the tree and free the
1439 	 * nodes. We avoid calling zio_flush() since there isn't any
1440 	 * good reason for doing so, after the lwb block failed to be
1441 	 * written out.
1442 	 *
1443 	 * Additionally, we don't perform any further error handling at
1444 	 * this point (e.g. setting "zcw_zio_error" appropriately), as
1445 	 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1446 	 * we expect any error seen here, to have been propagated to
1447 	 * that function).
1448 	 */
1449 	if (zio->io_error != 0) {
1450 		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1451 			kmem_free(zv, sizeof (*zv));
1452 		return;
1453 	}
1454 
1455 	/*
1456 	 * If this lwb does not have any threads waiting for it to
1457 	 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1458 	 * command to the vdevs written to by "this" lwb, and instead
1459 	 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1460 	 * command for those vdevs. Thus, we merge the vdev tree of
1461 	 * "this" lwb with the vdev tree of the "next" lwb in the list,
1462 	 * and assume the "next" lwb will handle flushing the vdevs (or
1463 	 * deferring the flush(s) again).
1464 	 *
1465 	 * This is a useful performance optimization, especially for
1466 	 * workloads with lots of async write activity and few sync
1467 	 * write and/or fsync activity, as it has the potential to
1468 	 * coalesce multiple flush commands to a vdev into one.
1469 	 */
1470 	if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) {
1471 		zil_lwb_flush_defer(lwb, nlwb);
1472 		ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1473 		return;
1474 	}
1475 
1476 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1477 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1478 		if (vd != NULL) {
1479 			/*
1480 			 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1481 			 * always used within "zio_flush". This means,
1482 			 * any errors when flushing the vdev(s), will
1483 			 * (unfortunately) not be handled correctly,
1484 			 * since these "zio_flush" errors will not be
1485 			 * propagated up to "zil_lwb_flush_vdevs_done".
1486 			 */
1487 			zio_flush(lwb->lwb_root_zio, vd);
1488 		}
1489 		kmem_free(zv, sizeof (*zv));
1490 	}
1491 }
1492 
1493 static void
1494 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1495 {
1496 	lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1497 
1498 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1499 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
1500 
1501 	/*
1502 	 * The zilog's "zl_last_lwb_opened" field is used to build the
1503 	 * lwb/zio dependency chain, which is used to preserve the
1504 	 * ordering of lwb completions that is required by the semantics
1505 	 * of the ZIL. Each new lwb zio becomes a parent of the
1506 	 * "previous" lwb zio, such that the new lwb's zio cannot
1507 	 * complete until the "previous" lwb's zio completes.
1508 	 *
1509 	 * This is required by the semantics of zil_commit(); the commit
1510 	 * waiters attached to the lwbs will be woken in the lwb zio's
1511 	 * completion callback, so this zio dependency graph ensures the
1512 	 * waiters are woken in the correct order (the same order the
1513 	 * lwbs were created).
1514 	 */
1515 	if (last_lwb_opened != NULL &&
1516 	    last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) {
1517 		ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1518 		    last_lwb_opened->lwb_state == LWB_STATE_ISSUED ||
1519 		    last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE);
1520 
1521 		ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1522 		zio_add_child(lwb->lwb_root_zio,
1523 		    last_lwb_opened->lwb_root_zio);
1524 
1525 		/*
1526 		 * If the previous lwb's write hasn't already completed,
1527 		 * we also want to order the completion of the lwb write
1528 		 * zios (above, we only order the completion of the lwb
1529 		 * root zios). This is required because of how we can
1530 		 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1531 		 *
1532 		 * When the DKIOCFLUSHWRITECACHE commands are deferred,
1533 		 * the previous lwb will rely on this lwb to flush the
1534 		 * vdevs written to by that previous lwb. Thus, we need
1535 		 * to ensure this lwb doesn't issue the flush until
1536 		 * after the previous lwb's write completes. We ensure
1537 		 * this ordering by setting the zio parent/child
1538 		 * relationship here.
1539 		 *
1540 		 * Without this relationship on the lwb's write zio,
1541 		 * it's possible for this lwb's write to complete prior
1542 		 * to the previous lwb's write completing; and thus, the
1543 		 * vdevs for the previous lwb would be flushed prior to
1544 		 * that lwb's data being written to those vdevs (the
1545 		 * vdevs are flushed in the lwb write zio's completion
1546 		 * handler, zil_lwb_write_done()).
1547 		 */
1548 		if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) {
1549 			ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1550 			    last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1551 
1552 			ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL);
1553 			zio_add_child(lwb->lwb_write_zio,
1554 			    last_lwb_opened->lwb_write_zio);
1555 		}
1556 	}
1557 }
1558 
1559 
1560 /*
1561  * This function's purpose is to "open" an lwb such that it is ready to
1562  * accept new itxs being committed to it. To do this, the lwb's zio
1563  * structures are created, and linked to the lwb. This function is
1564  * idempotent; if the passed in lwb has already been opened, this
1565  * function is essentially a no-op.
1566  */
1567 static void
1568 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1569 {
1570 	zbookmark_phys_t zb;
1571 	zio_priority_t prio;
1572 
1573 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1574 	ASSERT3P(lwb, !=, NULL);
1575 	EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1576 	EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1577 
1578 	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1579 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1580 	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1581 
1582 	/* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1583 	mutex_enter(&zilog->zl_lock);
1584 	if (lwb->lwb_root_zio == NULL) {
1585 		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1586 		    BP_GET_LSIZE(&lwb->lwb_blk));
1587 
1588 		if (!lwb->lwb_fastwrite) {
1589 			metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
1590 			lwb->lwb_fastwrite = 1;
1591 		}
1592 
1593 		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1594 			prio = ZIO_PRIORITY_SYNC_WRITE;
1595 		else
1596 			prio = ZIO_PRIORITY_ASYNC_WRITE;
1597 
1598 		lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1599 		    zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1600 		ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1601 
1602 		lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1603 		    zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1604 		    BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1605 		    prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb);
1606 		ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1607 
1608 		lwb->lwb_state = LWB_STATE_OPENED;
1609 
1610 		zil_lwb_set_zio_dependency(zilog, lwb);
1611 		zilog->zl_last_lwb_opened = lwb;
1612 	}
1613 	mutex_exit(&zilog->zl_lock);
1614 
1615 	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1616 	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1617 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1618 }
1619 
1620 /*
1621  * Define a limited set of intent log block sizes.
1622  *
1623  * These must be a multiple of 4KB. Note only the amount used (again
1624  * aligned to 4KB) actually gets written. However, we can't always just
1625  * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1626  */
1627 static const struct {
1628 	uint64_t	limit;
1629 	uint64_t	blksz;
1630 } zil_block_buckets[] = {
1631 	{ 4096,		4096 },			/* non TX_WRITE */
1632 	{ 8192 + 4096,	8192 + 4096 },		/* database */
1633 	{ 32768 + 4096,	32768 + 4096 },		/* NFS writes */
1634 	{ 65536 + 4096,	65536 + 4096 },		/* 64KB writes */
1635 	{ 131072,	131072 },		/* < 128KB writes */
1636 	{ 131072 +4096,	65536 + 4096 },		/* 128KB writes */
1637 	{ UINT64_MAX,	SPA_OLD_MAXBLOCKSIZE},	/* > 128KB writes */
1638 };
1639 
1640 /*
1641  * Maximum block size used by the ZIL.  This is picked up when the ZIL is
1642  * initialized.  Otherwise this should not be used directly; see
1643  * zl_max_block_size instead.
1644  */
1645 static int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1646 
1647 /*
1648  * Start a log block write and advance to the next log block.
1649  * Calls are serialized.
1650  */
1651 static lwb_t *
1652 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1653 {
1654 	lwb_t *nlwb = NULL;
1655 	zil_chain_t *zilc;
1656 	spa_t *spa = zilog->zl_spa;
1657 	blkptr_t *bp;
1658 	dmu_tx_t *tx;
1659 	uint64_t txg;
1660 	uint64_t zil_blksz, wsz;
1661 	int i, error;
1662 	boolean_t slog;
1663 
1664 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1665 	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1666 	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1667 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1668 
1669 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1670 		zilc = (zil_chain_t *)lwb->lwb_buf;
1671 		bp = &zilc->zc_next_blk;
1672 	} else {
1673 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1674 		bp = &zilc->zc_next_blk;
1675 	}
1676 
1677 	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1678 
1679 	/*
1680 	 * Allocate the next block and save its address in this block
1681 	 * before writing it in order to establish the log chain.
1682 	 */
1683 
1684 	tx = dmu_tx_create(zilog->zl_os);
1685 
1686 	/*
1687 	 * Since we are not going to create any new dirty data, and we
1688 	 * can even help with clearing the existing dirty data, we
1689 	 * should not be subject to the dirty data based delays. We
1690 	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1691 	 */
1692 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1693 
1694 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1695 	txg = dmu_tx_get_txg(tx);
1696 
1697 	mutex_enter(&zilog->zl_lwb_io_lock);
1698 	lwb->lwb_issued_txg = txg;
1699 	zilog->zl_lwb_inflight[txg & TXG_MASK]++;
1700 	zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg);
1701 	mutex_exit(&zilog->zl_lwb_io_lock);
1702 
1703 	/*
1704 	 * Log blocks are pre-allocated. Here we select the size of the next
1705 	 * block, based on size used in the last block.
1706 	 * - first find the smallest bucket that will fit the block from a
1707 	 *   limited set of block sizes. This is because it's faster to write
1708 	 *   blocks allocated from the same metaslab as they are adjacent or
1709 	 *   close.
1710 	 * - next find the maximum from the new suggested size and an array of
1711 	 *   previous sizes. This lessens a picket fence effect of wrongly
1712 	 *   guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1713 	 *   requests.
1714 	 *
1715 	 * Note we only write what is used, but we can't just allocate
1716 	 * the maximum block size because we can exhaust the available
1717 	 * pool log space.
1718 	 */
1719 	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1720 	for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++)
1721 		continue;
1722 	zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size);
1723 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1724 	for (i = 0; i < ZIL_PREV_BLKS; i++)
1725 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1726 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1727 
1728 	BP_ZERO(bp);
1729 	error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog);
1730 	if (slog) {
1731 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_slog_count);
1732 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_bytes,
1733 		    lwb->lwb_nused);
1734 	} else {
1735 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_normal_count);
1736 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_bytes,
1737 		    lwb->lwb_nused);
1738 	}
1739 	if (error == 0) {
1740 		ASSERT3U(bp->blk_birth, ==, txg);
1741 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1742 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1743 
1744 		/*
1745 		 * Allocate a new log write block (lwb).
1746 		 */
1747 		nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
1748 	}
1749 
1750 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1751 		/* For Slim ZIL only write what is used. */
1752 		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1753 		ASSERT3U(wsz, <=, lwb->lwb_sz);
1754 		zio_shrink(lwb->lwb_write_zio, wsz);
1755 
1756 	} else {
1757 		wsz = lwb->lwb_sz;
1758 	}
1759 
1760 	zilc->zc_pad = 0;
1761 	zilc->zc_nused = lwb->lwb_nused;
1762 	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1763 
1764 	/*
1765 	 * clear unused data for security
1766 	 */
1767 	memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused);
1768 
1769 	spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1770 
1771 	zil_lwb_add_block(lwb, &lwb->lwb_blk);
1772 	lwb->lwb_issued_timestamp = gethrtime();
1773 	lwb->lwb_state = LWB_STATE_ISSUED;
1774 
1775 	zio_nowait(lwb->lwb_root_zio);
1776 	zio_nowait(lwb->lwb_write_zio);
1777 
1778 	dmu_tx_commit(tx);
1779 
1780 	/*
1781 	 * If there was an allocation failure then nlwb will be null which
1782 	 * forces a txg_wait_synced().
1783 	 */
1784 	return (nlwb);
1785 }
1786 
1787 /*
1788  * Maximum amount of write data that can be put into single log block.
1789  */
1790 uint64_t
1791 zil_max_log_data(zilog_t *zilog)
1792 {
1793 	return (zilog->zl_max_block_size -
1794 	    sizeof (zil_chain_t) - sizeof (lr_write_t));
1795 }
1796 
1797 /*
1798  * Maximum amount of log space we agree to waste to reduce number of
1799  * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1800  */
1801 static inline uint64_t
1802 zil_max_waste_space(zilog_t *zilog)
1803 {
1804 	return (zil_max_log_data(zilog) / 8);
1805 }
1806 
1807 /*
1808  * Maximum amount of write data for WR_COPIED.  For correctness, consumers
1809  * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1810  * maximum sized log block, because each WR_COPIED record must fit in a
1811  * single log block.  For space efficiency, we want to fit two records into a
1812  * max-sized log block.
1813  */
1814 uint64_t
1815 zil_max_copied_data(zilog_t *zilog)
1816 {
1817 	return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 -
1818 	    sizeof (lr_write_t));
1819 }
1820 
1821 static lwb_t *
1822 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1823 {
1824 	lr_t *lrcb, *lrc;
1825 	lr_write_t *lrwb, *lrw;
1826 	char *lr_buf;
1827 	uint64_t dlen, dnow, dpad, lwb_sp, reclen, txg, max_log_data;
1828 
1829 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1830 	ASSERT3P(lwb, !=, NULL);
1831 	ASSERT3P(lwb->lwb_buf, !=, NULL);
1832 
1833 	zil_lwb_write_open(zilog, lwb);
1834 
1835 	lrc = &itx->itx_lr;
1836 	lrw = (lr_write_t *)lrc;
1837 
1838 	/*
1839 	 * A commit itx doesn't represent any on-disk state; instead
1840 	 * it's simply used as a place holder on the commit list, and
1841 	 * provides a mechanism for attaching a "commit waiter" onto the
1842 	 * correct lwb (such that the waiter can be signalled upon
1843 	 * completion of that lwb). Thus, we don't process this itx's
1844 	 * log record if it's a commit itx (these itx's don't have log
1845 	 * records), and instead link the itx's waiter onto the lwb's
1846 	 * list of waiters.
1847 	 *
1848 	 * For more details, see the comment above zil_commit().
1849 	 */
1850 	if (lrc->lrc_txtype == TX_COMMIT) {
1851 		mutex_enter(&zilog->zl_lock);
1852 		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1853 		itx->itx_private = NULL;
1854 		mutex_exit(&zilog->zl_lock);
1855 		return (lwb);
1856 	}
1857 
1858 	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1859 		dlen = P2ROUNDUP_TYPED(
1860 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1861 		dpad = dlen - lrw->lr_length;
1862 	} else {
1863 		dlen = dpad = 0;
1864 	}
1865 	reclen = lrc->lrc_reclen;
1866 	zilog->zl_cur_used += (reclen + dlen);
1867 	txg = lrc->lrc_txg;
1868 
1869 	ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1870 
1871 cont:
1872 	/*
1873 	 * If this record won't fit in the current log block, start a new one.
1874 	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1875 	 */
1876 	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1877 	max_log_data = zil_max_log_data(zilog);
1878 	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1879 	    lwb_sp < zil_max_waste_space(zilog) &&
1880 	    (dlen % max_log_data == 0 ||
1881 	    lwb_sp < reclen + dlen % max_log_data))) {
1882 		lwb = zil_lwb_write_issue(zilog, lwb);
1883 		if (lwb == NULL)
1884 			return (NULL);
1885 		zil_lwb_write_open(zilog, lwb);
1886 		ASSERT(LWB_EMPTY(lwb));
1887 		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1888 
1889 		/*
1890 		 * There must be enough space in the new, empty log block to
1891 		 * hold reclen.  For WR_COPIED, we need to fit the whole
1892 		 * record in one block, and reclen is the header size + the
1893 		 * data size. For WR_NEED_COPY, we can create multiple
1894 		 * records, splitting the data into multiple blocks, so we
1895 		 * only need to fit one word of data per block; in this case
1896 		 * reclen is just the header size (no data).
1897 		 */
1898 		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1899 	}
1900 
1901 	dnow = MIN(dlen, lwb_sp - reclen);
1902 	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1903 	memcpy(lr_buf, lrc, reclen);
1904 	lrcb = (lr_t *)lr_buf;		/* Like lrc, but inside lwb. */
1905 	lrwb = (lr_write_t *)lrcb;	/* Like lrw, but inside lwb. */
1906 
1907 	ZIL_STAT_BUMP(zilog, zil_itx_count);
1908 
1909 	/*
1910 	 * If it's a write, fetch the data or get its blkptr as appropriate.
1911 	 */
1912 	if (lrc->lrc_txtype == TX_WRITE) {
1913 		if (txg > spa_freeze_txg(zilog->zl_spa))
1914 			txg_wait_synced(zilog->zl_dmu_pool, txg);
1915 		if (itx->itx_wr_state == WR_COPIED) {
1916 			ZIL_STAT_BUMP(zilog, zil_itx_copied_count);
1917 			ZIL_STAT_INCR(zilog, zil_itx_copied_bytes,
1918 			    lrw->lr_length);
1919 		} else {
1920 			char *dbuf;
1921 			int error;
1922 
1923 			if (itx->itx_wr_state == WR_NEED_COPY) {
1924 				dbuf = lr_buf + reclen;
1925 				lrcb->lrc_reclen += dnow;
1926 				if (lrwb->lr_length > dnow)
1927 					lrwb->lr_length = dnow;
1928 				lrw->lr_offset += dnow;
1929 				lrw->lr_length -= dnow;
1930 				ZIL_STAT_BUMP(zilog, zil_itx_needcopy_count);
1931 				ZIL_STAT_INCR(zilog, zil_itx_needcopy_bytes,
1932 				    dnow);
1933 			} else {
1934 				ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
1935 				dbuf = NULL;
1936 				ZIL_STAT_BUMP(zilog, zil_itx_indirect_count);
1937 				ZIL_STAT_INCR(zilog, zil_itx_indirect_bytes,
1938 				    lrw->lr_length);
1939 			}
1940 
1941 			/*
1942 			 * We pass in the "lwb_write_zio" rather than
1943 			 * "lwb_root_zio" so that the "lwb_write_zio"
1944 			 * becomes the parent of any zio's created by
1945 			 * the "zl_get_data" callback. The vdevs are
1946 			 * flushed after the "lwb_write_zio" completes,
1947 			 * so we want to make sure that completion
1948 			 * callback waits for these additional zio's,
1949 			 * such that the vdevs used by those zio's will
1950 			 * be included in the lwb's vdev tree, and those
1951 			 * vdevs will be properly flushed. If we passed
1952 			 * in "lwb_root_zio" here, then these additional
1953 			 * vdevs may not be flushed; e.g. if these zio's
1954 			 * completed after "lwb_write_zio" completed.
1955 			 */
1956 			error = zilog->zl_get_data(itx->itx_private,
1957 			    itx->itx_gen, lrwb, dbuf, lwb,
1958 			    lwb->lwb_write_zio);
1959 			if (dbuf != NULL && error == 0 && dnow == dlen)
1960 				/* Zero any padding bytes in the last block. */
1961 				memset((char *)dbuf + lrwb->lr_length, 0, dpad);
1962 
1963 			if (error == EIO) {
1964 				txg_wait_synced(zilog->zl_dmu_pool, txg);
1965 				return (lwb);
1966 			}
1967 			if (error != 0) {
1968 				ASSERT(error == ENOENT || error == EEXIST ||
1969 				    error == EALREADY);
1970 				return (lwb);
1971 			}
1972 		}
1973 	}
1974 
1975 	/*
1976 	 * We're actually making an entry, so update lrc_seq to be the
1977 	 * log record sequence number.  Note that this is generally not
1978 	 * equal to the itx sequence number because not all transactions
1979 	 * are synchronous, and sometimes spa_sync() gets there first.
1980 	 */
1981 	lrcb->lrc_seq = ++zilog->zl_lr_seq;
1982 	lwb->lwb_nused += reclen + dnow;
1983 
1984 	zil_lwb_add_txg(lwb, txg);
1985 
1986 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1987 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1988 
1989 	dlen -= dnow;
1990 	if (dlen > 0) {
1991 		zilog->zl_cur_used += reclen;
1992 		goto cont;
1993 	}
1994 
1995 	return (lwb);
1996 }
1997 
1998 itx_t *
1999 zil_itx_create(uint64_t txtype, size_t olrsize)
2000 {
2001 	size_t itxsize, lrsize;
2002 	itx_t *itx;
2003 
2004 	lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t);
2005 	itxsize = offsetof(itx_t, itx_lr) + lrsize;
2006 
2007 	itx = zio_data_buf_alloc(itxsize);
2008 	itx->itx_lr.lrc_txtype = txtype;
2009 	itx->itx_lr.lrc_reclen = lrsize;
2010 	itx->itx_lr.lrc_seq = 0;	/* defensive */
2011 	memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize);
2012 	itx->itx_sync = B_TRUE;		/* default is synchronous */
2013 	itx->itx_callback = NULL;
2014 	itx->itx_callback_data = NULL;
2015 	itx->itx_size = itxsize;
2016 
2017 	return (itx);
2018 }
2019 
2020 void
2021 zil_itx_destroy(itx_t *itx)
2022 {
2023 	IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
2024 	IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2025 
2026 	if (itx->itx_callback != NULL)
2027 		itx->itx_callback(itx->itx_callback_data);
2028 
2029 	zio_data_buf_free(itx, itx->itx_size);
2030 }
2031 
2032 /*
2033  * Free up the sync and async itxs. The itxs_t has already been detached
2034  * so no locks are needed.
2035  */
2036 static void
2037 zil_itxg_clean(void *arg)
2038 {
2039 	itx_t *itx;
2040 	list_t *list;
2041 	avl_tree_t *t;
2042 	void *cookie;
2043 	itxs_t *itxs = arg;
2044 	itx_async_node_t *ian;
2045 
2046 	list = &itxs->i_sync_list;
2047 	while ((itx = list_head(list)) != NULL) {
2048 		/*
2049 		 * In the general case, commit itxs will not be found
2050 		 * here, as they'll be committed to an lwb via
2051 		 * zil_lwb_commit(), and free'd in that function. Having
2052 		 * said that, it is still possible for commit itxs to be
2053 		 * found here, due to the following race:
2054 		 *
2055 		 *	- a thread calls zil_commit() which assigns the
2056 		 *	  commit itx to a per-txg i_sync_list
2057 		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
2058 		 *	  while the waiter is still on the i_sync_list
2059 		 *
2060 		 * There's nothing to prevent syncing the txg while the
2061 		 * waiter is on the i_sync_list. This normally doesn't
2062 		 * happen because spa_sync() is slower than zil_commit(),
2063 		 * but if zil_commit() calls txg_wait_synced() (e.g.
2064 		 * because zil_create() or zil_commit_writer_stall() is
2065 		 * called) we will hit this case.
2066 		 */
2067 		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
2068 			zil_commit_waiter_skip(itx->itx_private);
2069 
2070 		list_remove(list, itx);
2071 		zil_itx_destroy(itx);
2072 	}
2073 
2074 	cookie = NULL;
2075 	t = &itxs->i_async_tree;
2076 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2077 		list = &ian->ia_list;
2078 		while ((itx = list_head(list)) != NULL) {
2079 			list_remove(list, itx);
2080 			/* commit itxs should never be on the async lists. */
2081 			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2082 			zil_itx_destroy(itx);
2083 		}
2084 		list_destroy(list);
2085 		kmem_free(ian, sizeof (itx_async_node_t));
2086 	}
2087 	avl_destroy(t);
2088 
2089 	kmem_free(itxs, sizeof (itxs_t));
2090 }
2091 
2092 static int
2093 zil_aitx_compare(const void *x1, const void *x2)
2094 {
2095 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
2096 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
2097 
2098 	return (TREE_CMP(o1, o2));
2099 }
2100 
2101 /*
2102  * Remove all async itx with the given oid.
2103  */
2104 void
2105 zil_remove_async(zilog_t *zilog, uint64_t oid)
2106 {
2107 	uint64_t otxg, txg;
2108 	itx_async_node_t *ian;
2109 	avl_tree_t *t;
2110 	avl_index_t where;
2111 	list_t clean_list;
2112 	itx_t *itx;
2113 
2114 	ASSERT(oid != 0);
2115 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
2116 
2117 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2118 		otxg = ZILTEST_TXG;
2119 	else
2120 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2121 
2122 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2123 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2124 
2125 		mutex_enter(&itxg->itxg_lock);
2126 		if (itxg->itxg_txg != txg) {
2127 			mutex_exit(&itxg->itxg_lock);
2128 			continue;
2129 		}
2130 
2131 		/*
2132 		 * Locate the object node and append its list.
2133 		 */
2134 		t = &itxg->itxg_itxs->i_async_tree;
2135 		ian = avl_find(t, &oid, &where);
2136 		if (ian != NULL)
2137 			list_move_tail(&clean_list, &ian->ia_list);
2138 		mutex_exit(&itxg->itxg_lock);
2139 	}
2140 	while ((itx = list_head(&clean_list)) != NULL) {
2141 		list_remove(&clean_list, itx);
2142 		/* commit itxs should never be on the async lists. */
2143 		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2144 		zil_itx_destroy(itx);
2145 	}
2146 	list_destroy(&clean_list);
2147 }
2148 
2149 void
2150 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
2151 {
2152 	uint64_t txg;
2153 	itxg_t *itxg;
2154 	itxs_t *itxs, *clean = NULL;
2155 
2156 	/*
2157 	 * Ensure the data of a renamed file is committed before the rename.
2158 	 */
2159 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
2160 		zil_async_to_sync(zilog, itx->itx_oid);
2161 
2162 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
2163 		txg = ZILTEST_TXG;
2164 	else
2165 		txg = dmu_tx_get_txg(tx);
2166 
2167 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
2168 	mutex_enter(&itxg->itxg_lock);
2169 	itxs = itxg->itxg_itxs;
2170 	if (itxg->itxg_txg != txg) {
2171 		if (itxs != NULL) {
2172 			/*
2173 			 * The zil_clean callback hasn't got around to cleaning
2174 			 * this itxg. Save the itxs for release below.
2175 			 * This should be rare.
2176 			 */
2177 			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2178 			    "txg %llu", (u_longlong_t)itxg->itxg_txg);
2179 			clean = itxg->itxg_itxs;
2180 		}
2181 		itxg->itxg_txg = txg;
2182 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
2183 		    KM_SLEEP);
2184 
2185 		list_create(&itxs->i_sync_list, sizeof (itx_t),
2186 		    offsetof(itx_t, itx_node));
2187 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
2188 		    sizeof (itx_async_node_t),
2189 		    offsetof(itx_async_node_t, ia_node));
2190 	}
2191 	if (itx->itx_sync) {
2192 		list_insert_tail(&itxs->i_sync_list, itx);
2193 	} else {
2194 		avl_tree_t *t = &itxs->i_async_tree;
2195 		uint64_t foid =
2196 		    LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
2197 		itx_async_node_t *ian;
2198 		avl_index_t where;
2199 
2200 		ian = avl_find(t, &foid, &where);
2201 		if (ian == NULL) {
2202 			ian = kmem_alloc(sizeof (itx_async_node_t),
2203 			    KM_SLEEP);
2204 			list_create(&ian->ia_list, sizeof (itx_t),
2205 			    offsetof(itx_t, itx_node));
2206 			ian->ia_foid = foid;
2207 			avl_insert(t, ian, where);
2208 		}
2209 		list_insert_tail(&ian->ia_list, itx);
2210 	}
2211 
2212 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
2213 
2214 	/*
2215 	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2216 	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2217 	 * need to be careful to always dirty the ZIL using the "real"
2218 	 * TXG (not itxg_txg) even when the SPA is frozen.
2219 	 */
2220 	zilog_dirty(zilog, dmu_tx_get_txg(tx));
2221 	mutex_exit(&itxg->itxg_lock);
2222 
2223 	/* Release the old itxs now we've dropped the lock */
2224 	if (clean != NULL)
2225 		zil_itxg_clean(clean);
2226 }
2227 
2228 /*
2229  * If there are any in-memory intent log transactions which have now been
2230  * synced then start up a taskq to free them. We should only do this after we
2231  * have written out the uberblocks (i.e. txg has been committed) so that
2232  * don't inadvertently clean out in-memory log records that would be required
2233  * by zil_commit().
2234  */
2235 void
2236 zil_clean(zilog_t *zilog, uint64_t synced_txg)
2237 {
2238 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2239 	itxs_t *clean_me;
2240 
2241 	ASSERT3U(synced_txg, <, ZILTEST_TXG);
2242 
2243 	mutex_enter(&itxg->itxg_lock);
2244 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2245 		mutex_exit(&itxg->itxg_lock);
2246 		return;
2247 	}
2248 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2249 	ASSERT3U(itxg->itxg_txg, !=, 0);
2250 	clean_me = itxg->itxg_itxs;
2251 	itxg->itxg_itxs = NULL;
2252 	itxg->itxg_txg = 0;
2253 	mutex_exit(&itxg->itxg_lock);
2254 	/*
2255 	 * Preferably start a task queue to free up the old itxs but
2256 	 * if taskq_dispatch can't allocate resources to do that then
2257 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
2258 	 * created a bad performance problem.
2259 	 */
2260 	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2261 	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2262 	taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
2263 	    zil_itxg_clean, clean_me, TQ_NOSLEEP);
2264 	if (id == TASKQID_INVALID)
2265 		zil_itxg_clean(clean_me);
2266 }
2267 
2268 /*
2269  * This function will traverse the queue of itxs that need to be
2270  * committed, and move them onto the ZIL's zl_itx_commit_list.
2271  */
2272 static void
2273 zil_get_commit_list(zilog_t *zilog)
2274 {
2275 	uint64_t otxg, txg;
2276 	list_t *commit_list = &zilog->zl_itx_commit_list;
2277 
2278 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2279 
2280 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2281 		otxg = ZILTEST_TXG;
2282 	else
2283 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2284 
2285 	/*
2286 	 * This is inherently racy, since there is nothing to prevent
2287 	 * the last synced txg from changing. That's okay since we'll
2288 	 * only commit things in the future.
2289 	 */
2290 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2291 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2292 
2293 		mutex_enter(&itxg->itxg_lock);
2294 		if (itxg->itxg_txg != txg) {
2295 			mutex_exit(&itxg->itxg_lock);
2296 			continue;
2297 		}
2298 
2299 		/*
2300 		 * If we're adding itx records to the zl_itx_commit_list,
2301 		 * then the zil better be dirty in this "txg". We can assert
2302 		 * that here since we're holding the itxg_lock which will
2303 		 * prevent spa_sync from cleaning it. Once we add the itxs
2304 		 * to the zl_itx_commit_list we must commit it to disk even
2305 		 * if it's unnecessary (i.e. the txg was synced).
2306 		 */
2307 		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2308 		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2309 		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
2310 
2311 		mutex_exit(&itxg->itxg_lock);
2312 	}
2313 }
2314 
2315 /*
2316  * Move the async itxs for a specified object to commit into sync lists.
2317  */
2318 void
2319 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2320 {
2321 	uint64_t otxg, txg;
2322 	itx_async_node_t *ian;
2323 	avl_tree_t *t;
2324 	avl_index_t where;
2325 
2326 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2327 		otxg = ZILTEST_TXG;
2328 	else
2329 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2330 
2331 	/*
2332 	 * This is inherently racy, since there is nothing to prevent
2333 	 * the last synced txg from changing.
2334 	 */
2335 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2336 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2337 
2338 		mutex_enter(&itxg->itxg_lock);
2339 		if (itxg->itxg_txg != txg) {
2340 			mutex_exit(&itxg->itxg_lock);
2341 			continue;
2342 		}
2343 
2344 		/*
2345 		 * If a foid is specified then find that node and append its
2346 		 * list. Otherwise walk the tree appending all the lists
2347 		 * to the sync list. We add to the end rather than the
2348 		 * beginning to ensure the create has happened.
2349 		 */
2350 		t = &itxg->itxg_itxs->i_async_tree;
2351 		if (foid != 0) {
2352 			ian = avl_find(t, &foid, &where);
2353 			if (ian != NULL) {
2354 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2355 				    &ian->ia_list);
2356 			}
2357 		} else {
2358 			void *cookie = NULL;
2359 
2360 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2361 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2362 				    &ian->ia_list);
2363 				list_destroy(&ian->ia_list);
2364 				kmem_free(ian, sizeof (itx_async_node_t));
2365 			}
2366 		}
2367 		mutex_exit(&itxg->itxg_lock);
2368 	}
2369 }
2370 
2371 /*
2372  * This function will prune commit itxs that are at the head of the
2373  * commit list (it won't prune past the first non-commit itx), and
2374  * either: a) attach them to the last lwb that's still pending
2375  * completion, or b) skip them altogether.
2376  *
2377  * This is used as a performance optimization to prevent commit itxs
2378  * from generating new lwbs when it's unnecessary to do so.
2379  */
2380 static void
2381 zil_prune_commit_list(zilog_t *zilog)
2382 {
2383 	itx_t *itx;
2384 
2385 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2386 
2387 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2388 		lr_t *lrc = &itx->itx_lr;
2389 		if (lrc->lrc_txtype != TX_COMMIT)
2390 			break;
2391 
2392 		mutex_enter(&zilog->zl_lock);
2393 
2394 		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2395 		if (last_lwb == NULL ||
2396 		    last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2397 			/*
2398 			 * All of the itxs this waiter was waiting on
2399 			 * must have already completed (or there were
2400 			 * never any itx's for it to wait on), so it's
2401 			 * safe to skip this waiter and mark it done.
2402 			 */
2403 			zil_commit_waiter_skip(itx->itx_private);
2404 		} else {
2405 			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2406 			itx->itx_private = NULL;
2407 		}
2408 
2409 		mutex_exit(&zilog->zl_lock);
2410 
2411 		list_remove(&zilog->zl_itx_commit_list, itx);
2412 		zil_itx_destroy(itx);
2413 	}
2414 
2415 	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2416 }
2417 
2418 static void
2419 zil_commit_writer_stall(zilog_t *zilog)
2420 {
2421 	/*
2422 	 * When zio_alloc_zil() fails to allocate the next lwb block on
2423 	 * disk, we must call txg_wait_synced() to ensure all of the
2424 	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2425 	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2426 	 * to zil_process_commit_list()) will have to call zil_create(),
2427 	 * and start a new ZIL chain.
2428 	 *
2429 	 * Since zil_alloc_zil() failed, the lwb that was previously
2430 	 * issued does not have a pointer to the "next" lwb on disk.
2431 	 * Thus, if another ZIL writer thread was to allocate the "next"
2432 	 * on-disk lwb, that block could be leaked in the event of a
2433 	 * crash (because the previous lwb on-disk would not point to
2434 	 * it).
2435 	 *
2436 	 * We must hold the zilog's zl_issuer_lock while we do this, to
2437 	 * ensure no new threads enter zil_process_commit_list() until
2438 	 * all lwb's in the zl_lwb_list have been synced and freed
2439 	 * (which is achieved via the txg_wait_synced() call).
2440 	 */
2441 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2442 	txg_wait_synced(zilog->zl_dmu_pool, 0);
2443 	ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2444 }
2445 
2446 /*
2447  * This function will traverse the commit list, creating new lwbs as
2448  * needed, and committing the itxs from the commit list to these newly
2449  * created lwbs. Additionally, as a new lwb is created, the previous
2450  * lwb will be issued to the zio layer to be written to disk.
2451  */
2452 static void
2453 zil_process_commit_list(zilog_t *zilog)
2454 {
2455 	spa_t *spa = zilog->zl_spa;
2456 	list_t nolwb_itxs;
2457 	list_t nolwb_waiters;
2458 	lwb_t *lwb;
2459 	itx_t *itx;
2460 
2461 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2462 
2463 	/*
2464 	 * Return if there's nothing to commit before we dirty the fs by
2465 	 * calling zil_create().
2466 	 */
2467 	if (list_head(&zilog->zl_itx_commit_list) == NULL)
2468 		return;
2469 
2470 	list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
2471 	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2472 	    offsetof(zil_commit_waiter_t, zcw_node));
2473 
2474 	lwb = list_tail(&zilog->zl_lwb_list);
2475 	if (lwb == NULL) {
2476 		lwb = zil_create(zilog);
2477 	} else {
2478 		/*
2479 		 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2480 		 * have already been created (zl_lwb_list not empty).
2481 		 */
2482 		zil_commit_activate_saxattr_feature(zilog);
2483 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2484 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2485 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2486 	}
2487 
2488 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2489 		lr_t *lrc = &itx->itx_lr;
2490 		uint64_t txg = lrc->lrc_txg;
2491 
2492 		ASSERT3U(txg, !=, 0);
2493 
2494 		if (lrc->lrc_txtype == TX_COMMIT) {
2495 			DTRACE_PROBE2(zil__process__commit__itx,
2496 			    zilog_t *, zilog, itx_t *, itx);
2497 		} else {
2498 			DTRACE_PROBE2(zil__process__normal__itx,
2499 			    zilog_t *, zilog, itx_t *, itx);
2500 		}
2501 
2502 		list_remove(&zilog->zl_itx_commit_list, itx);
2503 
2504 		boolean_t synced = txg <= spa_last_synced_txg(spa);
2505 		boolean_t frozen = txg > spa_freeze_txg(spa);
2506 
2507 		/*
2508 		 * If the txg of this itx has already been synced out, then
2509 		 * we don't need to commit this itx to an lwb. This is
2510 		 * because the data of this itx will have already been
2511 		 * written to the main pool. This is inherently racy, and
2512 		 * it's still ok to commit an itx whose txg has already
2513 		 * been synced; this will result in a write that's
2514 		 * unnecessary, but will do no harm.
2515 		 *
2516 		 * With that said, we always want to commit TX_COMMIT itxs
2517 		 * to an lwb, regardless of whether or not that itx's txg
2518 		 * has been synced out. We do this to ensure any OPENED lwb
2519 		 * will always have at least one zil_commit_waiter_t linked
2520 		 * to the lwb.
2521 		 *
2522 		 * As a counter-example, if we skipped TX_COMMIT itx's
2523 		 * whose txg had already been synced, the following
2524 		 * situation could occur if we happened to be racing with
2525 		 * spa_sync:
2526 		 *
2527 		 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2528 		 *    itx's txg is 10 and the last synced txg is 9.
2529 		 * 2. spa_sync finishes syncing out txg 10.
2530 		 * 3. We move to the next itx in the list, it's a TX_COMMIT
2531 		 *    whose txg is 10, so we skip it rather than committing
2532 		 *    it to the lwb used in (1).
2533 		 *
2534 		 * If the itx that is skipped in (3) is the last TX_COMMIT
2535 		 * itx in the commit list, than it's possible for the lwb
2536 		 * used in (1) to remain in the OPENED state indefinitely.
2537 		 *
2538 		 * To prevent the above scenario from occurring, ensuring
2539 		 * that once an lwb is OPENED it will transition to ISSUED
2540 		 * and eventually DONE, we always commit TX_COMMIT itx's to
2541 		 * an lwb here, even if that itx's txg has already been
2542 		 * synced.
2543 		 *
2544 		 * Finally, if the pool is frozen, we _always_ commit the
2545 		 * itx.  The point of freezing the pool is to prevent data
2546 		 * from being written to the main pool via spa_sync, and
2547 		 * instead rely solely on the ZIL to persistently store the
2548 		 * data; i.e.  when the pool is frozen, the last synced txg
2549 		 * value can't be trusted.
2550 		 */
2551 		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2552 			if (lwb != NULL) {
2553 				lwb = zil_lwb_commit(zilog, itx, lwb);
2554 
2555 				if (lwb == NULL)
2556 					list_insert_tail(&nolwb_itxs, itx);
2557 				else
2558 					list_insert_tail(&lwb->lwb_itxs, itx);
2559 			} else {
2560 				if (lrc->lrc_txtype == TX_COMMIT) {
2561 					zil_commit_waiter_link_nolwb(
2562 					    itx->itx_private, &nolwb_waiters);
2563 				}
2564 
2565 				list_insert_tail(&nolwb_itxs, itx);
2566 			}
2567 		} else {
2568 			ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
2569 			zil_itx_destroy(itx);
2570 		}
2571 	}
2572 
2573 	if (lwb == NULL) {
2574 		/*
2575 		 * This indicates zio_alloc_zil() failed to allocate the
2576 		 * "next" lwb on-disk. When this happens, we must stall
2577 		 * the ZIL write pipeline; see the comment within
2578 		 * zil_commit_writer_stall() for more details.
2579 		 */
2580 		zil_commit_writer_stall(zilog);
2581 
2582 		/*
2583 		 * Additionally, we have to signal and mark the "nolwb"
2584 		 * waiters as "done" here, since without an lwb, we
2585 		 * can't do this via zil_lwb_flush_vdevs_done() like
2586 		 * normal.
2587 		 */
2588 		zil_commit_waiter_t *zcw;
2589 		while ((zcw = list_head(&nolwb_waiters)) != NULL) {
2590 			zil_commit_waiter_skip(zcw);
2591 			list_remove(&nolwb_waiters, zcw);
2592 		}
2593 
2594 		/*
2595 		 * And finally, we have to destroy the itx's that
2596 		 * couldn't be committed to an lwb; this will also call
2597 		 * the itx's callback if one exists for the itx.
2598 		 */
2599 		while ((itx = list_head(&nolwb_itxs)) != NULL) {
2600 			list_remove(&nolwb_itxs, itx);
2601 			zil_itx_destroy(itx);
2602 		}
2603 	} else {
2604 		ASSERT(list_is_empty(&nolwb_waiters));
2605 		ASSERT3P(lwb, !=, NULL);
2606 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2607 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2608 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2609 
2610 		/*
2611 		 * At this point, the ZIL block pointed at by the "lwb"
2612 		 * variable is in one of the following states: "closed"
2613 		 * or "open".
2614 		 *
2615 		 * If it's "closed", then no itxs have been committed to
2616 		 * it, so there's no point in issuing its zio (i.e. it's
2617 		 * "empty").
2618 		 *
2619 		 * If it's "open", then it contains one or more itxs that
2620 		 * eventually need to be committed to stable storage. In
2621 		 * this case we intentionally do not issue the lwb's zio
2622 		 * to disk yet, and instead rely on one of the following
2623 		 * two mechanisms for issuing the zio:
2624 		 *
2625 		 * 1. Ideally, there will be more ZIL activity occurring
2626 		 * on the system, such that this function will be
2627 		 * immediately called again (not necessarily by the same
2628 		 * thread) and this lwb's zio will be issued via
2629 		 * zil_lwb_commit(). This way, the lwb is guaranteed to
2630 		 * be "full" when it is issued to disk, and we'll make
2631 		 * use of the lwb's size the best we can.
2632 		 *
2633 		 * 2. If there isn't sufficient ZIL activity occurring on
2634 		 * the system, such that this lwb's zio isn't issued via
2635 		 * zil_lwb_commit(), zil_commit_waiter() will issue the
2636 		 * lwb's zio. If this occurs, the lwb is not guaranteed
2637 		 * to be "full" by the time its zio is issued, and means
2638 		 * the size of the lwb was "too large" given the amount
2639 		 * of ZIL activity occurring on the system at that time.
2640 		 *
2641 		 * We do this for a couple of reasons:
2642 		 *
2643 		 * 1. To try and reduce the number of IOPs needed to
2644 		 * write the same number of itxs. If an lwb has space
2645 		 * available in its buffer for more itxs, and more itxs
2646 		 * will be committed relatively soon (relative to the
2647 		 * latency of performing a write), then it's beneficial
2648 		 * to wait for these "next" itxs. This way, more itxs
2649 		 * can be committed to stable storage with fewer writes.
2650 		 *
2651 		 * 2. To try and use the largest lwb block size that the
2652 		 * incoming rate of itxs can support. Again, this is to
2653 		 * try and pack as many itxs into as few lwbs as
2654 		 * possible, without significantly impacting the latency
2655 		 * of each individual itx.
2656 		 */
2657 	}
2658 }
2659 
2660 /*
2661  * This function is responsible for ensuring the passed in commit waiter
2662  * (and associated commit itx) is committed to an lwb. If the waiter is
2663  * not already committed to an lwb, all itxs in the zilog's queue of
2664  * itxs will be processed. The assumption is the passed in waiter's
2665  * commit itx will found in the queue just like the other non-commit
2666  * itxs, such that when the entire queue is processed, the waiter will
2667  * have been committed to an lwb.
2668  *
2669  * The lwb associated with the passed in waiter is not guaranteed to
2670  * have been issued by the time this function completes. If the lwb is
2671  * not issued, we rely on future calls to zil_commit_writer() to issue
2672  * the lwb, or the timeout mechanism found in zil_commit_waiter().
2673  */
2674 static void
2675 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2676 {
2677 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2678 	ASSERT(spa_writeable(zilog->zl_spa));
2679 
2680 	mutex_enter(&zilog->zl_issuer_lock);
2681 
2682 	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2683 		/*
2684 		 * It's possible that, while we were waiting to acquire
2685 		 * the "zl_issuer_lock", another thread committed this
2686 		 * waiter to an lwb. If that occurs, we bail out early,
2687 		 * without processing any of the zilog's queue of itxs.
2688 		 *
2689 		 * On certain workloads and system configurations, the
2690 		 * "zl_issuer_lock" can become highly contended. In an
2691 		 * attempt to reduce this contention, we immediately drop
2692 		 * the lock if the waiter has already been processed.
2693 		 *
2694 		 * We've measured this optimization to reduce CPU spent
2695 		 * contending on this lock by up to 5%, using a system
2696 		 * with 32 CPUs, low latency storage (~50 usec writes),
2697 		 * and 1024 threads performing sync writes.
2698 		 */
2699 		goto out;
2700 	}
2701 
2702 	ZIL_STAT_BUMP(zilog, zil_commit_writer_count);
2703 
2704 	zil_get_commit_list(zilog);
2705 	zil_prune_commit_list(zilog);
2706 	zil_process_commit_list(zilog);
2707 
2708 out:
2709 	mutex_exit(&zilog->zl_issuer_lock);
2710 }
2711 
2712 static void
2713 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2714 {
2715 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2716 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2717 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2718 
2719 	lwb_t *lwb = zcw->zcw_lwb;
2720 	ASSERT3P(lwb, !=, NULL);
2721 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2722 
2723 	/*
2724 	 * If the lwb has already been issued by another thread, we can
2725 	 * immediately return since there's no work to be done (the
2726 	 * point of this function is to issue the lwb). Additionally, we
2727 	 * do this prior to acquiring the zl_issuer_lock, to avoid
2728 	 * acquiring it when it's not necessary to do so.
2729 	 */
2730 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2731 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2732 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2733 		return;
2734 
2735 	/*
2736 	 * In order to call zil_lwb_write_issue() we must hold the
2737 	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2738 	 * since we're already holding the commit waiter's "zcw_lock",
2739 	 * and those two locks are acquired in the opposite order
2740 	 * elsewhere.
2741 	 */
2742 	mutex_exit(&zcw->zcw_lock);
2743 	mutex_enter(&zilog->zl_issuer_lock);
2744 	mutex_enter(&zcw->zcw_lock);
2745 
2746 	/*
2747 	 * Since we just dropped and re-acquired the commit waiter's
2748 	 * lock, we have to re-check to see if the waiter was marked
2749 	 * "done" during that process. If the waiter was marked "done",
2750 	 * the "lwb" pointer is no longer valid (it can be free'd after
2751 	 * the waiter is marked "done"), so without this check we could
2752 	 * wind up with a use-after-free error below.
2753 	 */
2754 	if (zcw->zcw_done)
2755 		goto out;
2756 
2757 	ASSERT3P(lwb, ==, zcw->zcw_lwb);
2758 
2759 	/*
2760 	 * We've already checked this above, but since we hadn't acquired
2761 	 * the zilog's zl_issuer_lock, we have to perform this check a
2762 	 * second time while holding the lock.
2763 	 *
2764 	 * We don't need to hold the zl_lock since the lwb cannot transition
2765 	 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2766 	 * _can_ transition from ISSUED to DONE, but it's OK to race with
2767 	 * that transition since we treat the lwb the same, whether it's in
2768 	 * the ISSUED or DONE states.
2769 	 *
2770 	 * The important thing, is we treat the lwb differently depending on
2771 	 * if it's ISSUED or OPENED, and block any other threads that might
2772 	 * attempt to issue this lwb. For that reason we hold the
2773 	 * zl_issuer_lock when checking the lwb_state; we must not call
2774 	 * zil_lwb_write_issue() if the lwb had already been issued.
2775 	 *
2776 	 * See the comment above the lwb_state_t structure definition for
2777 	 * more details on the lwb states, and locking requirements.
2778 	 */
2779 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2780 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2781 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2782 		goto out;
2783 
2784 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2785 
2786 	/*
2787 	 * As described in the comments above zil_commit_waiter() and
2788 	 * zil_process_commit_list(), we need to issue this lwb's zio
2789 	 * since we've reached the commit waiter's timeout and it still
2790 	 * hasn't been issued.
2791 	 */
2792 	lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2793 
2794 	IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2795 
2796 	/*
2797 	 * Since the lwb's zio hadn't been issued by the time this thread
2798 	 * reached its timeout, we reset the zilog's "zl_cur_used" field
2799 	 * to influence the zil block size selection algorithm.
2800 	 *
2801 	 * By having to issue the lwb's zio here, it means the size of the
2802 	 * lwb was too large, given the incoming throughput of itxs.  By
2803 	 * setting "zl_cur_used" to zero, we communicate this fact to the
2804 	 * block size selection algorithm, so it can take this information
2805 	 * into account, and potentially select a smaller size for the
2806 	 * next lwb block that is allocated.
2807 	 */
2808 	zilog->zl_cur_used = 0;
2809 
2810 	if (nlwb == NULL) {
2811 		/*
2812 		 * When zil_lwb_write_issue() returns NULL, this
2813 		 * indicates zio_alloc_zil() failed to allocate the
2814 		 * "next" lwb on-disk. When this occurs, the ZIL write
2815 		 * pipeline must be stalled; see the comment within the
2816 		 * zil_commit_writer_stall() function for more details.
2817 		 *
2818 		 * We must drop the commit waiter's lock prior to
2819 		 * calling zil_commit_writer_stall() or else we can wind
2820 		 * up with the following deadlock:
2821 		 *
2822 		 * - This thread is waiting for the txg to sync while
2823 		 *   holding the waiter's lock; txg_wait_synced() is
2824 		 *   used within txg_commit_writer_stall().
2825 		 *
2826 		 * - The txg can't sync because it is waiting for this
2827 		 *   lwb's zio callback to call dmu_tx_commit().
2828 		 *
2829 		 * - The lwb's zio callback can't call dmu_tx_commit()
2830 		 *   because it's blocked trying to acquire the waiter's
2831 		 *   lock, which occurs prior to calling dmu_tx_commit()
2832 		 */
2833 		mutex_exit(&zcw->zcw_lock);
2834 		zil_commit_writer_stall(zilog);
2835 		mutex_enter(&zcw->zcw_lock);
2836 	}
2837 
2838 out:
2839 	mutex_exit(&zilog->zl_issuer_lock);
2840 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2841 }
2842 
2843 /*
2844  * This function is responsible for performing the following two tasks:
2845  *
2846  * 1. its primary responsibility is to block until the given "commit
2847  *    waiter" is considered "done".
2848  *
2849  * 2. its secondary responsibility is to issue the zio for the lwb that
2850  *    the given "commit waiter" is waiting on, if this function has
2851  *    waited "long enough" and the lwb is still in the "open" state.
2852  *
2853  * Given a sufficient amount of itxs being generated and written using
2854  * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2855  * function. If this does not occur, this secondary responsibility will
2856  * ensure the lwb is issued even if there is not other synchronous
2857  * activity on the system.
2858  *
2859  * For more details, see zil_process_commit_list(); more specifically,
2860  * the comment at the bottom of that function.
2861  */
2862 static void
2863 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2864 {
2865 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2866 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2867 	ASSERT(spa_writeable(zilog->zl_spa));
2868 
2869 	mutex_enter(&zcw->zcw_lock);
2870 
2871 	/*
2872 	 * The timeout is scaled based on the lwb latency to avoid
2873 	 * significantly impacting the latency of each individual itx.
2874 	 * For more details, see the comment at the bottom of the
2875 	 * zil_process_commit_list() function.
2876 	 */
2877 	int pct = MAX(zfs_commit_timeout_pct, 1);
2878 	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2879 	hrtime_t wakeup = gethrtime() + sleep;
2880 	boolean_t timedout = B_FALSE;
2881 
2882 	while (!zcw->zcw_done) {
2883 		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2884 
2885 		lwb_t *lwb = zcw->zcw_lwb;
2886 
2887 		/*
2888 		 * Usually, the waiter will have a non-NULL lwb field here,
2889 		 * but it's possible for it to be NULL as a result of
2890 		 * zil_commit() racing with spa_sync().
2891 		 *
2892 		 * When zil_clean() is called, it's possible for the itxg
2893 		 * list (which may be cleaned via a taskq) to contain
2894 		 * commit itxs. When this occurs, the commit waiters linked
2895 		 * off of these commit itxs will not be committed to an
2896 		 * lwb.  Additionally, these commit waiters will not be
2897 		 * marked done until zil_commit_waiter_skip() is called via
2898 		 * zil_itxg_clean().
2899 		 *
2900 		 * Thus, it's possible for this commit waiter (i.e. the
2901 		 * "zcw" variable) to be found in this "in between" state;
2902 		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2903 		 * been skipped, so it's "zcw_done" field is still B_FALSE.
2904 		 */
2905 		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2906 
2907 		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2908 			ASSERT3B(timedout, ==, B_FALSE);
2909 
2910 			/*
2911 			 * If the lwb hasn't been issued yet, then we
2912 			 * need to wait with a timeout, in case this
2913 			 * function needs to issue the lwb after the
2914 			 * timeout is reached; responsibility (2) from
2915 			 * the comment above this function.
2916 			 */
2917 			int rc = cv_timedwait_hires(&zcw->zcw_cv,
2918 			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2919 			    CALLOUT_FLAG_ABSOLUTE);
2920 
2921 			if (rc != -1 || zcw->zcw_done)
2922 				continue;
2923 
2924 			timedout = B_TRUE;
2925 			zil_commit_waiter_timeout(zilog, zcw);
2926 
2927 			if (!zcw->zcw_done) {
2928 				/*
2929 				 * If the commit waiter has already been
2930 				 * marked "done", it's possible for the
2931 				 * waiter's lwb structure to have already
2932 				 * been freed.  Thus, we can only reliably
2933 				 * make these assertions if the waiter
2934 				 * isn't done.
2935 				 */
2936 				ASSERT3P(lwb, ==, zcw->zcw_lwb);
2937 				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2938 			}
2939 		} else {
2940 			/*
2941 			 * If the lwb isn't open, then it must have already
2942 			 * been issued. In that case, there's no need to
2943 			 * use a timeout when waiting for the lwb to
2944 			 * complete.
2945 			 *
2946 			 * Additionally, if the lwb is NULL, the waiter
2947 			 * will soon be signaled and marked done via
2948 			 * zil_clean() and zil_itxg_clean(), so no timeout
2949 			 * is required.
2950 			 */
2951 
2952 			IMPLY(lwb != NULL,
2953 			    lwb->lwb_state == LWB_STATE_ISSUED ||
2954 			    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2955 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
2956 			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2957 		}
2958 	}
2959 
2960 	mutex_exit(&zcw->zcw_lock);
2961 }
2962 
2963 static zil_commit_waiter_t *
2964 zil_alloc_commit_waiter(void)
2965 {
2966 	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2967 
2968 	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2969 	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2970 	list_link_init(&zcw->zcw_node);
2971 	zcw->zcw_lwb = NULL;
2972 	zcw->zcw_done = B_FALSE;
2973 	zcw->zcw_zio_error = 0;
2974 
2975 	return (zcw);
2976 }
2977 
2978 static void
2979 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2980 {
2981 	ASSERT(!list_link_active(&zcw->zcw_node));
2982 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
2983 	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2984 	mutex_destroy(&zcw->zcw_lock);
2985 	cv_destroy(&zcw->zcw_cv);
2986 	kmem_cache_free(zil_zcw_cache, zcw);
2987 }
2988 
2989 /*
2990  * This function is used to create a TX_COMMIT itx and assign it. This
2991  * way, it will be linked into the ZIL's list of synchronous itxs, and
2992  * then later committed to an lwb (or skipped) when
2993  * zil_process_commit_list() is called.
2994  */
2995 static void
2996 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2997 {
2998 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2999 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
3000 
3001 	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
3002 	itx->itx_sync = B_TRUE;
3003 	itx->itx_private = zcw;
3004 
3005 	zil_itx_assign(zilog, itx, tx);
3006 
3007 	dmu_tx_commit(tx);
3008 }
3009 
3010 /*
3011  * Commit ZFS Intent Log transactions (itxs) to stable storage.
3012  *
3013  * When writing ZIL transactions to the on-disk representation of the
3014  * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3015  * itxs can be committed to a single lwb. Once a lwb is written and
3016  * committed to stable storage (i.e. the lwb is written, and vdevs have
3017  * been flushed), each itx that was committed to that lwb is also
3018  * considered to be committed to stable storage.
3019  *
3020  * When an itx is committed to an lwb, the log record (lr_t) contained
3021  * by the itx is copied into the lwb's zio buffer, and once this buffer
3022  * is written to disk, it becomes an on-disk ZIL block.
3023  *
3024  * As itxs are generated, they're inserted into the ZIL's queue of
3025  * uncommitted itxs. The semantics of zil_commit() are such that it will
3026  * block until all itxs that were in the queue when it was called, are
3027  * committed to stable storage.
3028  *
3029  * If "foid" is zero, this means all "synchronous" and "asynchronous"
3030  * itxs, for all objects in the dataset, will be committed to stable
3031  * storage prior to zil_commit() returning. If "foid" is non-zero, all
3032  * "synchronous" itxs for all objects, but only "asynchronous" itxs
3033  * that correspond to the foid passed in, will be committed to stable
3034  * storage prior to zil_commit() returning.
3035  *
3036  * Generally speaking, when zil_commit() is called, the consumer doesn't
3037  * actually care about _all_ of the uncommitted itxs. Instead, they're
3038  * simply trying to waiting for a specific itx to be committed to disk,
3039  * but the interface(s) for interacting with the ZIL don't allow such
3040  * fine-grained communication. A better interface would allow a consumer
3041  * to create and assign an itx, and then pass a reference to this itx to
3042  * zil_commit(); such that zil_commit() would return as soon as that
3043  * specific itx was committed to disk (instead of waiting for _all_
3044  * itxs to be committed).
3045  *
3046  * When a thread calls zil_commit() a special "commit itx" will be
3047  * generated, along with a corresponding "waiter" for this commit itx.
3048  * zil_commit() will wait on this waiter's CV, such that when the waiter
3049  * is marked done, and signaled, zil_commit() will return.
3050  *
3051  * This commit itx is inserted into the queue of uncommitted itxs. This
3052  * provides an easy mechanism for determining which itxs were in the
3053  * queue prior to zil_commit() having been called, and which itxs were
3054  * added after zil_commit() was called.
3055  *
3056  * The commit itx is special; it doesn't have any on-disk representation.
3057  * When a commit itx is "committed" to an lwb, the waiter associated
3058  * with it is linked onto the lwb's list of waiters. Then, when that lwb
3059  * completes, each waiter on the lwb's list is marked done and signaled
3060  * -- allowing the thread waiting on the waiter to return from zil_commit().
3061  *
3062  * It's important to point out a few critical factors that allow us
3063  * to make use of the commit itxs, commit waiters, per-lwb lists of
3064  * commit waiters, and zio completion callbacks like we're doing:
3065  *
3066  *   1. The list of waiters for each lwb is traversed, and each commit
3067  *      waiter is marked "done" and signaled, in the zio completion
3068  *      callback of the lwb's zio[*].
3069  *
3070  *      * Actually, the waiters are signaled in the zio completion
3071  *        callback of the root zio for the DKIOCFLUSHWRITECACHE commands
3072  *        that are sent to the vdevs upon completion of the lwb zio.
3073  *
3074  *   2. When the itxs are inserted into the ZIL's queue of uncommitted
3075  *      itxs, the order in which they are inserted is preserved[*]; as
3076  *      itxs are added to the queue, they are added to the tail of
3077  *      in-memory linked lists.
3078  *
3079  *      When committing the itxs to lwbs (to be written to disk), they
3080  *      are committed in the same order in which the itxs were added to
3081  *      the uncommitted queue's linked list(s); i.e. the linked list of
3082  *      itxs to commit is traversed from head to tail, and each itx is
3083  *      committed to an lwb in that order.
3084  *
3085  *      * To clarify:
3086  *
3087  *        - the order of "sync" itxs is preserved w.r.t. other
3088  *          "sync" itxs, regardless of the corresponding objects.
3089  *        - the order of "async" itxs is preserved w.r.t. other
3090  *          "async" itxs corresponding to the same object.
3091  *        - the order of "async" itxs is *not* preserved w.r.t. other
3092  *          "async" itxs corresponding to different objects.
3093  *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
3094  *          versa) is *not* preserved, even for itxs that correspond
3095  *          to the same object.
3096  *
3097  *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
3098  *      zil_get_commit_list(), and zil_process_commit_list().
3099  *
3100  *   3. The lwbs represent a linked list of blocks on disk. Thus, any
3101  *      lwb cannot be considered committed to stable storage, until its
3102  *      "previous" lwb is also committed to stable storage. This fact,
3103  *      coupled with the fact described above, means that itxs are
3104  *      committed in (roughly) the order in which they were generated.
3105  *      This is essential because itxs are dependent on prior itxs.
3106  *      Thus, we *must not* deem an itx as being committed to stable
3107  *      storage, until *all* prior itxs have also been committed to
3108  *      stable storage.
3109  *
3110  *      To enforce this ordering of lwb zio's, while still leveraging as
3111  *      much of the underlying storage performance as possible, we rely
3112  *      on two fundamental concepts:
3113  *
3114  *          1. The creation and issuance of lwb zio's is protected by
3115  *             the zilog's "zl_issuer_lock", which ensures only a single
3116  *             thread is creating and/or issuing lwb's at a time
3117  *          2. The "previous" lwb is a child of the "current" lwb
3118  *             (leveraging the zio parent-child dependency graph)
3119  *
3120  *      By relying on this parent-child zio relationship, we can have
3121  *      many lwb zio's concurrently issued to the underlying storage,
3122  *      but the order in which they complete will be the same order in
3123  *      which they were created.
3124  */
3125 void
3126 zil_commit(zilog_t *zilog, uint64_t foid)
3127 {
3128 	/*
3129 	 * We should never attempt to call zil_commit on a snapshot for
3130 	 * a couple of reasons:
3131 	 *
3132 	 * 1. A snapshot may never be modified, thus it cannot have any
3133 	 *    in-flight itxs that would have modified the dataset.
3134 	 *
3135 	 * 2. By design, when zil_commit() is called, a commit itx will
3136 	 *    be assigned to this zilog; as a result, the zilog will be
3137 	 *    dirtied. We must not dirty the zilog of a snapshot; there's
3138 	 *    checks in the code that enforce this invariant, and will
3139 	 *    cause a panic if it's not upheld.
3140 	 */
3141 	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
3142 
3143 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3144 		return;
3145 
3146 	if (!spa_writeable(zilog->zl_spa)) {
3147 		/*
3148 		 * If the SPA is not writable, there should never be any
3149 		 * pending itxs waiting to be committed to disk. If that
3150 		 * weren't true, we'd skip writing those itxs out, and
3151 		 * would break the semantics of zil_commit(); thus, we're
3152 		 * verifying that truth before we return to the caller.
3153 		 */
3154 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3155 		ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3156 		for (int i = 0; i < TXG_SIZE; i++)
3157 			ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
3158 		return;
3159 	}
3160 
3161 	/*
3162 	 * If the ZIL is suspended, we don't want to dirty it by calling
3163 	 * zil_commit_itx_assign() below, nor can we write out
3164 	 * lwbs like would be done in zil_commit_write(). Thus, we
3165 	 * simply rely on txg_wait_synced() to maintain the necessary
3166 	 * semantics, and avoid calling those functions altogether.
3167 	 */
3168 	if (zilog->zl_suspend > 0) {
3169 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3170 		return;
3171 	}
3172 
3173 	zil_commit_impl(zilog, foid);
3174 }
3175 
3176 void
3177 zil_commit_impl(zilog_t *zilog, uint64_t foid)
3178 {
3179 	ZIL_STAT_BUMP(zilog, zil_commit_count);
3180 
3181 	/*
3182 	 * Move the "async" itxs for the specified foid to the "sync"
3183 	 * queues, such that they will be later committed (or skipped)
3184 	 * to an lwb when zil_process_commit_list() is called.
3185 	 *
3186 	 * Since these "async" itxs must be committed prior to this
3187 	 * call to zil_commit returning, we must perform this operation
3188 	 * before we call zil_commit_itx_assign().
3189 	 */
3190 	zil_async_to_sync(zilog, foid);
3191 
3192 	/*
3193 	 * We allocate a new "waiter" structure which will initially be
3194 	 * linked to the commit itx using the itx's "itx_private" field.
3195 	 * Since the commit itx doesn't represent any on-disk state,
3196 	 * when it's committed to an lwb, rather than copying the its
3197 	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3198 	 * added to the lwb's list of waiters. Then, when the lwb is
3199 	 * committed to stable storage, each waiter in the lwb's list of
3200 	 * waiters will be marked "done", and signalled.
3201 	 *
3202 	 * We must create the waiter and assign the commit itx prior to
3203 	 * calling zil_commit_writer(), or else our specific commit itx
3204 	 * is not guaranteed to be committed to an lwb prior to calling
3205 	 * zil_commit_waiter().
3206 	 */
3207 	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
3208 	zil_commit_itx_assign(zilog, zcw);
3209 
3210 	zil_commit_writer(zilog, zcw);
3211 	zil_commit_waiter(zilog, zcw);
3212 
3213 	if (zcw->zcw_zio_error != 0) {
3214 		/*
3215 		 * If there was an error writing out the ZIL blocks that
3216 		 * this thread is waiting on, then we fallback to
3217 		 * relying on spa_sync() to write out the data this
3218 		 * thread is waiting on. Obviously this has performance
3219 		 * implications, but the expectation is for this to be
3220 		 * an exceptional case, and shouldn't occur often.
3221 		 */
3222 		DTRACE_PROBE2(zil__commit__io__error,
3223 		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
3224 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3225 	}
3226 
3227 	zil_free_commit_waiter(zcw);
3228 }
3229 
3230 /*
3231  * Called in syncing context to free committed log blocks and update log header.
3232  */
3233 void
3234 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
3235 {
3236 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
3237 	uint64_t txg = dmu_tx_get_txg(tx);
3238 	spa_t *spa = zilog->zl_spa;
3239 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
3240 	lwb_t *lwb;
3241 
3242 	/*
3243 	 * We don't zero out zl_destroy_txg, so make sure we don't try
3244 	 * to destroy it twice.
3245 	 */
3246 	if (spa_sync_pass(spa) != 1)
3247 		return;
3248 
3249 	zil_lwb_flush_wait_all(zilog, txg);
3250 
3251 	mutex_enter(&zilog->zl_lock);
3252 
3253 	ASSERT(zilog->zl_stop_sync == 0);
3254 
3255 	if (*replayed_seq != 0) {
3256 		ASSERT(zh->zh_replay_seq < *replayed_seq);
3257 		zh->zh_replay_seq = *replayed_seq;
3258 		*replayed_seq = 0;
3259 	}
3260 
3261 	if (zilog->zl_destroy_txg == txg) {
3262 		blkptr_t blk = zh->zh_log;
3263 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
3264 
3265 		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
3266 
3267 		memset(zh, 0, sizeof (zil_header_t));
3268 		memset(zilog->zl_replayed_seq, 0,
3269 		    sizeof (zilog->zl_replayed_seq));
3270 
3271 		if (zilog->zl_keep_first) {
3272 			/*
3273 			 * If this block was part of log chain that couldn't
3274 			 * be claimed because a device was missing during
3275 			 * zil_claim(), but that device later returns,
3276 			 * then this block could erroneously appear valid.
3277 			 * To guard against this, assign a new GUID to the new
3278 			 * log chain so it doesn't matter what blk points to.
3279 			 */
3280 			zil_init_log_chain(zilog, &blk);
3281 			zh->zh_log = blk;
3282 		} else {
3283 			/*
3284 			 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3285 			 * records. So, deactivate the feature for this dataset.
3286 			 * We activate it again when we start a new ZIL chain.
3287 			 */
3288 			if (dsl_dataset_feature_is_active(ds,
3289 			    SPA_FEATURE_ZILSAXATTR))
3290 				dsl_dataset_deactivate_feature(ds,
3291 				    SPA_FEATURE_ZILSAXATTR, tx);
3292 		}
3293 	}
3294 
3295 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
3296 		zh->zh_log = lwb->lwb_blk;
3297 		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
3298 			break;
3299 		list_remove(&zilog->zl_lwb_list, lwb);
3300 		zio_free(spa, txg, &lwb->lwb_blk);
3301 		zil_free_lwb(zilog, lwb);
3302 
3303 		/*
3304 		 * If we don't have anything left in the lwb list then
3305 		 * we've had an allocation failure and we need to zero
3306 		 * out the zil_header blkptr so that we don't end
3307 		 * up freeing the same block twice.
3308 		 */
3309 		if (list_head(&zilog->zl_lwb_list) == NULL)
3310 			BP_ZERO(&zh->zh_log);
3311 	}
3312 
3313 	/*
3314 	 * Remove fastwrite on any blocks that have been pre-allocated for
3315 	 * the next commit. This prevents fastwrite counter pollution by
3316 	 * unused, long-lived LWBs.
3317 	 */
3318 	for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
3319 		if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) {
3320 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3321 			lwb->lwb_fastwrite = 0;
3322 		}
3323 	}
3324 
3325 	mutex_exit(&zilog->zl_lock);
3326 }
3327 
3328 static int
3329 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
3330 {
3331 	(void) unused, (void) kmflag;
3332 	lwb_t *lwb = vbuf;
3333 	list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3334 	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
3335 	    offsetof(zil_commit_waiter_t, zcw_node));
3336 	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
3337 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
3338 	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
3339 	return (0);
3340 }
3341 
3342 static void
3343 zil_lwb_dest(void *vbuf, void *unused)
3344 {
3345 	(void) unused;
3346 	lwb_t *lwb = vbuf;
3347 	mutex_destroy(&lwb->lwb_vdev_lock);
3348 	avl_destroy(&lwb->lwb_vdev_tree);
3349 	list_destroy(&lwb->lwb_waiters);
3350 	list_destroy(&lwb->lwb_itxs);
3351 }
3352 
3353 void
3354 zil_init(void)
3355 {
3356 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3357 	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3358 
3359 	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3360 	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3361 
3362 	zil_sums_init(&zil_sums_global);
3363 	zil_kstats_global = kstat_create("zfs", 0, "zil", "misc",
3364 	    KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
3365 	    KSTAT_FLAG_VIRTUAL);
3366 
3367 	if (zil_kstats_global != NULL) {
3368 		zil_kstats_global->ks_data = &zil_stats;
3369 		zil_kstats_global->ks_update = zil_kstats_global_update;
3370 		zil_kstats_global->ks_private = NULL;
3371 		kstat_install(zil_kstats_global);
3372 	}
3373 }
3374 
3375 void
3376 zil_fini(void)
3377 {
3378 	kmem_cache_destroy(zil_zcw_cache);
3379 	kmem_cache_destroy(zil_lwb_cache);
3380 
3381 	if (zil_kstats_global != NULL) {
3382 		kstat_delete(zil_kstats_global);
3383 		zil_kstats_global = NULL;
3384 	}
3385 
3386 	zil_sums_fini(&zil_sums_global);
3387 }
3388 
3389 void
3390 zil_set_sync(zilog_t *zilog, uint64_t sync)
3391 {
3392 	zilog->zl_sync = sync;
3393 }
3394 
3395 void
3396 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3397 {
3398 	zilog->zl_logbias = logbias;
3399 }
3400 
3401 zilog_t *
3402 zil_alloc(objset_t *os, zil_header_t *zh_phys)
3403 {
3404 	zilog_t *zilog;
3405 
3406 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3407 
3408 	zilog->zl_header = zh_phys;
3409 	zilog->zl_os = os;
3410 	zilog->zl_spa = dmu_objset_spa(os);
3411 	zilog->zl_dmu_pool = dmu_objset_pool(os);
3412 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
3413 	zilog->zl_logbias = dmu_objset_logbias(os);
3414 	zilog->zl_sync = dmu_objset_syncprop(os);
3415 	zilog->zl_dirty_max_txg = 0;
3416 	zilog->zl_last_lwb_opened = NULL;
3417 	zilog->zl_last_lwb_latency = 0;
3418 	zilog->zl_max_block_size = zil_maxblocksize;
3419 
3420 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3421 	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3422 	mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL);
3423 
3424 	for (int i = 0; i < TXG_SIZE; i++) {
3425 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3426 		    MUTEX_DEFAULT, NULL);
3427 	}
3428 
3429 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3430 	    offsetof(lwb_t, lwb_node));
3431 
3432 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3433 	    offsetof(itx_t, itx_node));
3434 
3435 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3436 	cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL);
3437 
3438 	return (zilog);
3439 }
3440 
3441 void
3442 zil_free(zilog_t *zilog)
3443 {
3444 	int i;
3445 
3446 	zilog->zl_stop_sync = 1;
3447 
3448 	ASSERT0(zilog->zl_suspend);
3449 	ASSERT0(zilog->zl_suspending);
3450 
3451 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3452 	list_destroy(&zilog->zl_lwb_list);
3453 
3454 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3455 	list_destroy(&zilog->zl_itx_commit_list);
3456 
3457 	for (i = 0; i < TXG_SIZE; i++) {
3458 		/*
3459 		 * It's possible for an itx to be generated that doesn't dirty
3460 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3461 		 * callback to remove the entry. We remove those here.
3462 		 *
3463 		 * Also free up the ziltest itxs.
3464 		 */
3465 		if (zilog->zl_itxg[i].itxg_itxs)
3466 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3467 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3468 	}
3469 
3470 	mutex_destroy(&zilog->zl_issuer_lock);
3471 	mutex_destroy(&zilog->zl_lock);
3472 	mutex_destroy(&zilog->zl_lwb_io_lock);
3473 
3474 	cv_destroy(&zilog->zl_cv_suspend);
3475 	cv_destroy(&zilog->zl_lwb_io_cv);
3476 
3477 	kmem_free(zilog, sizeof (zilog_t));
3478 }
3479 
3480 /*
3481  * Open an intent log.
3482  */
3483 zilog_t *
3484 zil_open(objset_t *os, zil_get_data_t *get_data, zil_sums_t *zil_sums)
3485 {
3486 	zilog_t *zilog = dmu_objset_zil(os);
3487 
3488 	ASSERT3P(zilog->zl_get_data, ==, NULL);
3489 	ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3490 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3491 
3492 	zilog->zl_get_data = get_data;
3493 	zilog->zl_sums = zil_sums;
3494 
3495 	return (zilog);
3496 }
3497 
3498 /*
3499  * Close an intent log.
3500  */
3501 void
3502 zil_close(zilog_t *zilog)
3503 {
3504 	lwb_t *lwb;
3505 	uint64_t txg;
3506 
3507 	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3508 		zil_commit(zilog, 0);
3509 	} else {
3510 		ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
3511 		ASSERT0(zilog->zl_dirty_max_txg);
3512 		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3513 	}
3514 
3515 	mutex_enter(&zilog->zl_lock);
3516 	lwb = list_tail(&zilog->zl_lwb_list);
3517 	if (lwb == NULL)
3518 		txg = zilog->zl_dirty_max_txg;
3519 	else
3520 		txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
3521 	mutex_exit(&zilog->zl_lock);
3522 
3523 	/*
3524 	 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3525 	 * on the time when the dmu_tx transaction is assigned in
3526 	 * zil_lwb_write_issue().
3527 	 */
3528 	mutex_enter(&zilog->zl_lwb_io_lock);
3529 	txg = MAX(zilog->zl_lwb_max_issued_txg, txg);
3530 	mutex_exit(&zilog->zl_lwb_io_lock);
3531 
3532 	/*
3533 	 * We need to use txg_wait_synced() to wait until that txg is synced.
3534 	 * zil_sync() will guarantee all lwbs up to that txg have been
3535 	 * written out, flushed, and cleaned.
3536 	 */
3537 	if (txg != 0)
3538 		txg_wait_synced(zilog->zl_dmu_pool, txg);
3539 
3540 	if (zilog_is_dirty(zilog))
3541 		zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog,
3542 		    (u_longlong_t)txg);
3543 	if (txg < spa_freeze_txg(zilog->zl_spa))
3544 		VERIFY(!zilog_is_dirty(zilog));
3545 
3546 	zilog->zl_get_data = NULL;
3547 
3548 	/*
3549 	 * We should have only one lwb left on the list; remove it now.
3550 	 */
3551 	mutex_enter(&zilog->zl_lock);
3552 	lwb = list_head(&zilog->zl_lwb_list);
3553 	if (lwb != NULL) {
3554 		ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
3555 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3556 
3557 		if (lwb->lwb_fastwrite)
3558 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3559 
3560 		list_remove(&zilog->zl_lwb_list, lwb);
3561 		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3562 		zil_free_lwb(zilog, lwb);
3563 	}
3564 	mutex_exit(&zilog->zl_lock);
3565 }
3566 
3567 static const char *suspend_tag = "zil suspending";
3568 
3569 /*
3570  * Suspend an intent log.  While in suspended mode, we still honor
3571  * synchronous semantics, but we rely on txg_wait_synced() to do it.
3572  * On old version pools, we suspend the log briefly when taking a
3573  * snapshot so that it will have an empty intent log.
3574  *
3575  * Long holds are not really intended to be used the way we do here --
3576  * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
3577  * could fail.  Therefore we take pains to only put a long hold if it is
3578  * actually necessary.  Fortunately, it will only be necessary if the
3579  * objset is currently mounted (or the ZVOL equivalent).  In that case it
3580  * will already have a long hold, so we are not really making things any worse.
3581  *
3582  * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3583  * zvol_state_t), and use their mechanism to prevent their hold from being
3584  * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
3585  * very little gain.
3586  *
3587  * if cookiep == NULL, this does both the suspend & resume.
3588  * Otherwise, it returns with the dataset "long held", and the cookie
3589  * should be passed into zil_resume().
3590  */
3591 int
3592 zil_suspend(const char *osname, void **cookiep)
3593 {
3594 	objset_t *os;
3595 	zilog_t *zilog;
3596 	const zil_header_t *zh;
3597 	int error;
3598 
3599 	error = dmu_objset_hold(osname, suspend_tag, &os);
3600 	if (error != 0)
3601 		return (error);
3602 	zilog = dmu_objset_zil(os);
3603 
3604 	mutex_enter(&zilog->zl_lock);
3605 	zh = zilog->zl_header;
3606 
3607 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
3608 		mutex_exit(&zilog->zl_lock);
3609 		dmu_objset_rele(os, suspend_tag);
3610 		return (SET_ERROR(EBUSY));
3611 	}
3612 
3613 	/*
3614 	 * Don't put a long hold in the cases where we can avoid it.  This
3615 	 * is when there is no cookie so we are doing a suspend & resume
3616 	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3617 	 * for the suspend because it's already suspended, or there's no ZIL.
3618 	 */
3619 	if (cookiep == NULL && !zilog->zl_suspending &&
3620 	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3621 		mutex_exit(&zilog->zl_lock);
3622 		dmu_objset_rele(os, suspend_tag);
3623 		return (0);
3624 	}
3625 
3626 	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3627 	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3628 
3629 	zilog->zl_suspend++;
3630 
3631 	if (zilog->zl_suspend > 1) {
3632 		/*
3633 		 * Someone else is already suspending it.
3634 		 * Just wait for them to finish.
3635 		 */
3636 
3637 		while (zilog->zl_suspending)
3638 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3639 		mutex_exit(&zilog->zl_lock);
3640 
3641 		if (cookiep == NULL)
3642 			zil_resume(os);
3643 		else
3644 			*cookiep = os;
3645 		return (0);
3646 	}
3647 
3648 	/*
3649 	 * If there is no pointer to an on-disk block, this ZIL must not
3650 	 * be active (e.g. filesystem not mounted), so there's nothing
3651 	 * to clean up.
3652 	 */
3653 	if (BP_IS_HOLE(&zh->zh_log)) {
3654 		ASSERT(cookiep != NULL); /* fast path already handled */
3655 
3656 		*cookiep = os;
3657 		mutex_exit(&zilog->zl_lock);
3658 		return (0);
3659 	}
3660 
3661 	/*
3662 	 * The ZIL has work to do. Ensure that the associated encryption
3663 	 * key will remain mapped while we are committing the log by
3664 	 * grabbing a reference to it. If the key isn't loaded we have no
3665 	 * choice but to return an error until the wrapping key is loaded.
3666 	 */
3667 	if (os->os_encrypted &&
3668 	    dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
3669 		zilog->zl_suspend--;
3670 		mutex_exit(&zilog->zl_lock);
3671 		dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3672 		dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3673 		return (SET_ERROR(EACCES));
3674 	}
3675 
3676 	zilog->zl_suspending = B_TRUE;
3677 	mutex_exit(&zilog->zl_lock);
3678 
3679 	/*
3680 	 * We need to use zil_commit_impl to ensure we wait for all
3681 	 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
3682 	 * to disk before proceeding. If we used zil_commit instead, it
3683 	 * would just call txg_wait_synced(), because zl_suspend is set.
3684 	 * txg_wait_synced() doesn't wait for these lwb's to be
3685 	 * LWB_STATE_FLUSH_DONE before returning.
3686 	 */
3687 	zil_commit_impl(zilog, 0);
3688 
3689 	/*
3690 	 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3691 	 * use txg_wait_synced() to ensure the data from the zilog has
3692 	 * migrated to the main pool before calling zil_destroy().
3693 	 */
3694 	txg_wait_synced(zilog->zl_dmu_pool, 0);
3695 
3696 	zil_destroy(zilog, B_FALSE);
3697 
3698 	mutex_enter(&zilog->zl_lock);
3699 	zilog->zl_suspending = B_FALSE;
3700 	cv_broadcast(&zilog->zl_cv_suspend);
3701 	mutex_exit(&zilog->zl_lock);
3702 
3703 	if (os->os_encrypted)
3704 		dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
3705 
3706 	if (cookiep == NULL)
3707 		zil_resume(os);
3708 	else
3709 		*cookiep = os;
3710 	return (0);
3711 }
3712 
3713 void
3714 zil_resume(void *cookie)
3715 {
3716 	objset_t *os = cookie;
3717 	zilog_t *zilog = dmu_objset_zil(os);
3718 
3719 	mutex_enter(&zilog->zl_lock);
3720 	ASSERT(zilog->zl_suspend != 0);
3721 	zilog->zl_suspend--;
3722 	mutex_exit(&zilog->zl_lock);
3723 	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3724 	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3725 }
3726 
3727 typedef struct zil_replay_arg {
3728 	zil_replay_func_t *const *zr_replay;
3729 	void		*zr_arg;
3730 	boolean_t	zr_byteswap;
3731 	char		*zr_lr;
3732 } zil_replay_arg_t;
3733 
3734 static int
3735 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
3736 {
3737 	char name[ZFS_MAX_DATASET_NAME_LEN];
3738 
3739 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
3740 
3741 	dmu_objset_name(zilog->zl_os, name);
3742 
3743 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3744 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3745 	    (u_longlong_t)lr->lrc_seq,
3746 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3747 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
3748 
3749 	return (error);
3750 }
3751 
3752 static int
3753 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
3754     uint64_t claim_txg)
3755 {
3756 	zil_replay_arg_t *zr = zra;
3757 	const zil_header_t *zh = zilog->zl_header;
3758 	uint64_t reclen = lr->lrc_reclen;
3759 	uint64_t txtype = lr->lrc_txtype;
3760 	int error = 0;
3761 
3762 	zilog->zl_replaying_seq = lr->lrc_seq;
3763 
3764 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
3765 		return (0);
3766 
3767 	if (lr->lrc_txg < claim_txg)		/* already committed */
3768 		return (0);
3769 
3770 	/* Strip case-insensitive bit, still present in log record */
3771 	txtype &= ~TX_CI;
3772 
3773 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
3774 		return (zil_replay_error(zilog, lr, EINVAL));
3775 
3776 	/*
3777 	 * If this record type can be logged out of order, the object
3778 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
3779 	 */
3780 	if (TX_OOO(txtype)) {
3781 		error = dmu_object_info(zilog->zl_os,
3782 		    LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
3783 		if (error == ENOENT || error == EEXIST)
3784 			return (0);
3785 	}
3786 
3787 	/*
3788 	 * Make a copy of the data so we can revise and extend it.
3789 	 */
3790 	memcpy(zr->zr_lr, lr, reclen);
3791 
3792 	/*
3793 	 * If this is a TX_WRITE with a blkptr, suck in the data.
3794 	 */
3795 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3796 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
3797 		    zr->zr_lr + reclen);
3798 		if (error != 0)
3799 			return (zil_replay_error(zilog, lr, error));
3800 	}
3801 
3802 	/*
3803 	 * The log block containing this lr may have been byteswapped
3804 	 * so that we can easily examine common fields like lrc_txtype.
3805 	 * However, the log is a mix of different record types, and only the
3806 	 * replay vectors know how to byteswap their records.  Therefore, if
3807 	 * the lr was byteswapped, undo it before invoking the replay vector.
3808 	 */
3809 	if (zr->zr_byteswap)
3810 		byteswap_uint64_array(zr->zr_lr, reclen);
3811 
3812 	/*
3813 	 * We must now do two things atomically: replay this log record,
3814 	 * and update the log header sequence number to reflect the fact that
3815 	 * we did so. At the end of each replay function the sequence number
3816 	 * is updated if we are in replay mode.
3817 	 */
3818 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3819 	if (error != 0) {
3820 		/*
3821 		 * The DMU's dnode layer doesn't see removes until the txg
3822 		 * commits, so a subsequent claim can spuriously fail with
3823 		 * EEXIST. So if we receive any error we try syncing out
3824 		 * any removes then retry the transaction.  Note that we
3825 		 * specify B_FALSE for byteswap now, so we don't do it twice.
3826 		 */
3827 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3828 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3829 		if (error != 0)
3830 			return (zil_replay_error(zilog, lr, error));
3831 	}
3832 	return (0);
3833 }
3834 
3835 static int
3836 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
3837 {
3838 	(void) bp, (void) arg, (void) claim_txg;
3839 
3840 	zilog->zl_replay_blks++;
3841 
3842 	return (0);
3843 }
3844 
3845 /*
3846  * If this dataset has a non-empty intent log, replay it and destroy it.
3847  */
3848 void
3849 zil_replay(objset_t *os, void *arg,
3850     zil_replay_func_t *const replay_func[TX_MAX_TYPE])
3851 {
3852 	zilog_t *zilog = dmu_objset_zil(os);
3853 	const zil_header_t *zh = zilog->zl_header;
3854 	zil_replay_arg_t zr;
3855 
3856 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3857 		zil_destroy(zilog, B_TRUE);
3858 		return;
3859 	}
3860 
3861 	zr.zr_replay = replay_func;
3862 	zr.zr_arg = arg;
3863 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3864 	zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3865 
3866 	/*
3867 	 * Wait for in-progress removes to sync before starting replay.
3868 	 */
3869 	txg_wait_synced(zilog->zl_dmu_pool, 0);
3870 
3871 	zilog->zl_replay = B_TRUE;
3872 	zilog->zl_replay_time = ddi_get_lbolt();
3873 	ASSERT(zilog->zl_replay_blks == 0);
3874 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3875 	    zh->zh_claim_txg, B_TRUE);
3876 	vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3877 
3878 	zil_destroy(zilog, B_FALSE);
3879 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3880 	zilog->zl_replay = B_FALSE;
3881 }
3882 
3883 boolean_t
3884 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3885 {
3886 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3887 		return (B_TRUE);
3888 
3889 	if (zilog->zl_replay) {
3890 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3891 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3892 		    zilog->zl_replaying_seq;
3893 		return (B_TRUE);
3894 	}
3895 
3896 	return (B_FALSE);
3897 }
3898 
3899 int
3900 zil_reset(const char *osname, void *arg)
3901 {
3902 	(void) arg;
3903 
3904 	int error = zil_suspend(osname, NULL);
3905 	/* EACCES means crypto key not loaded */
3906 	if ((error == EACCES) || (error == EBUSY))
3907 		return (SET_ERROR(error));
3908 	if (error != 0)
3909 		return (SET_ERROR(EEXIST));
3910 	return (0);
3911 }
3912 
3913 EXPORT_SYMBOL(zil_alloc);
3914 EXPORT_SYMBOL(zil_free);
3915 EXPORT_SYMBOL(zil_open);
3916 EXPORT_SYMBOL(zil_close);
3917 EXPORT_SYMBOL(zil_replay);
3918 EXPORT_SYMBOL(zil_replaying);
3919 EXPORT_SYMBOL(zil_destroy);
3920 EXPORT_SYMBOL(zil_destroy_sync);
3921 EXPORT_SYMBOL(zil_itx_create);
3922 EXPORT_SYMBOL(zil_itx_destroy);
3923 EXPORT_SYMBOL(zil_itx_assign);
3924 EXPORT_SYMBOL(zil_commit);
3925 EXPORT_SYMBOL(zil_claim);
3926 EXPORT_SYMBOL(zil_check_log_chain);
3927 EXPORT_SYMBOL(zil_sync);
3928 EXPORT_SYMBOL(zil_clean);
3929 EXPORT_SYMBOL(zil_suspend);
3930 EXPORT_SYMBOL(zil_resume);
3931 EXPORT_SYMBOL(zil_lwb_add_block);
3932 EXPORT_SYMBOL(zil_bp_tree_add);
3933 EXPORT_SYMBOL(zil_set_sync);
3934 EXPORT_SYMBOL(zil_set_logbias);
3935 EXPORT_SYMBOL(zil_sums_init);
3936 EXPORT_SYMBOL(zil_sums_fini);
3937 EXPORT_SYMBOL(zil_kstat_values_update);
3938 
3939 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, INT, ZMOD_RW,
3940 	"ZIL block open timeout percentage");
3941 
3942 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
3943 	"Disable intent logging replay");
3944 
3945 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
3946 	"Disable ZIL cache flushes");
3947 
3948 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, ULONG, ZMOD_RW,
3949 	"Limit in bytes slog sync writes per commit");
3950 
3951 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, INT, ZMOD_RW,
3952 	"Limit in bytes of ZIL log block size");
3953