xref: /freebsd/sys/contrib/openzfs/module/zfs/zil.c (revision 04dacd5691fb3de5fddd0868836c5a7ff9888370)
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/brt.h>
47 #include <sys/wmsum.h>
48 
49 /*
50  * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
51  * calls that change the file system. Each itx has enough information to
52  * be able to replay them after a system crash, power loss, or
53  * equivalent failure mode. These are stored in memory until either:
54  *
55  *   1. they are committed to the pool by the DMU transaction group
56  *      (txg), at which point they can be discarded; or
57  *   2. they are committed to the on-disk ZIL for the dataset being
58  *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
59  *      requirement).
60  *
61  * In the event of a crash or power loss, the itxs contained by each
62  * dataset's on-disk ZIL will be replayed when that dataset is first
63  * instantiated (e.g. if the dataset is a normal filesystem, when it is
64  * first mounted).
65  *
66  * As hinted at above, there is one ZIL per dataset (both the in-memory
67  * representation, and the on-disk representation). The on-disk format
68  * consists of 3 parts:
69  *
70  * 	- a single, per-dataset, ZIL header; which points to a chain of
71  * 	- zero or more ZIL blocks; each of which contains
72  * 	- zero or more ZIL records
73  *
74  * A ZIL record holds the information necessary to replay a single
75  * system call transaction. A ZIL block can hold many ZIL records, and
76  * the blocks are chained together, similarly to a singly linked list.
77  *
78  * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
79  * block in the chain, and the ZIL header points to the first block in
80  * the chain.
81  *
82  * Note, there is not a fixed place in the pool to hold these ZIL
83  * blocks; they are dynamically allocated and freed as needed from the
84  * blocks available on the pool, though they can be preferentially
85  * allocated from a dedicated "log" vdev.
86  */
87 
88 /*
89  * This controls the amount of time that a ZIL block (lwb) will remain
90  * "open" when it isn't "full", and it has a thread waiting for it to be
91  * committed to stable storage. Please refer to the zil_commit_waiter()
92  * function (and the comments within it) for more details.
93  */
94 static uint_t zfs_commit_timeout_pct = 5;
95 
96 /*
97  * Minimal time we care to delay commit waiting for more ZIL records.
98  * At least FreeBSD kernel can't sleep for less than 2us at its best.
99  * So requests to sleep for less then 5us is a waste of CPU time with
100  * a risk of significant log latency increase due to oversleep.
101  */
102 static uint64_t zil_min_commit_timeout = 5000;
103 
104 /*
105  * See zil.h for more information about these fields.
106  */
107 static zil_kstat_values_t zil_stats = {
108 	{ "zil_commit_count",			KSTAT_DATA_UINT64 },
109 	{ "zil_commit_writer_count",		KSTAT_DATA_UINT64 },
110 	{ "zil_itx_count",			KSTAT_DATA_UINT64 },
111 	{ "zil_itx_indirect_count",		KSTAT_DATA_UINT64 },
112 	{ "zil_itx_indirect_bytes",		KSTAT_DATA_UINT64 },
113 	{ "zil_itx_copied_count",		KSTAT_DATA_UINT64 },
114 	{ "zil_itx_copied_bytes",		KSTAT_DATA_UINT64 },
115 	{ "zil_itx_needcopy_count",		KSTAT_DATA_UINT64 },
116 	{ "zil_itx_needcopy_bytes",		KSTAT_DATA_UINT64 },
117 	{ "zil_itx_metaslab_normal_count",	KSTAT_DATA_UINT64 },
118 	{ "zil_itx_metaslab_normal_bytes",	KSTAT_DATA_UINT64 },
119 	{ "zil_itx_metaslab_slog_count",	KSTAT_DATA_UINT64 },
120 	{ "zil_itx_metaslab_slog_bytes",	KSTAT_DATA_UINT64 },
121 };
122 
123 static zil_sums_t zil_sums_global;
124 static kstat_t *zil_kstats_global;
125 
126 /*
127  * Disable intent logging replay.  This global ZIL switch affects all pools.
128  */
129 int zil_replay_disable = 0;
130 
131 /*
132  * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
133  * the disk(s) by the ZIL after an LWB write has completed. Setting this
134  * will cause ZIL corruption on power loss if a volatile out-of-order
135  * write cache is enabled.
136  */
137 static int zil_nocacheflush = 0;
138 
139 /*
140  * Limit SLOG write size per commit executed with synchronous priority.
141  * Any writes above that will be executed with lower (asynchronous) priority
142  * to limit potential SLOG device abuse by single active ZIL writer.
143  */
144 static uint64_t zil_slog_bulk = 768 * 1024;
145 
146 static kmem_cache_t *zil_lwb_cache;
147 static kmem_cache_t *zil_zcw_cache;
148 
149 #define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
150     sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
151 
152 static int
153 zil_bp_compare(const void *x1, const void *x2)
154 {
155 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
156 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
157 
158 	int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
159 	if (likely(cmp))
160 		return (cmp);
161 
162 	return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
163 }
164 
165 static void
166 zil_bp_tree_init(zilog_t *zilog)
167 {
168 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
169 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
170 }
171 
172 static void
173 zil_bp_tree_fini(zilog_t *zilog)
174 {
175 	avl_tree_t *t = &zilog->zl_bp_tree;
176 	zil_bp_node_t *zn;
177 	void *cookie = NULL;
178 
179 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
180 		kmem_free(zn, sizeof (zil_bp_node_t));
181 
182 	avl_destroy(t);
183 }
184 
185 int
186 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
187 {
188 	avl_tree_t *t = &zilog->zl_bp_tree;
189 	const dva_t *dva;
190 	zil_bp_node_t *zn;
191 	avl_index_t where;
192 
193 	if (BP_IS_EMBEDDED(bp))
194 		return (0);
195 
196 	dva = BP_IDENTITY(bp);
197 
198 	if (avl_find(t, dva, &where) != NULL)
199 		return (SET_ERROR(EEXIST));
200 
201 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
202 	zn->zn_dva = *dva;
203 	avl_insert(t, zn, where);
204 
205 	return (0);
206 }
207 
208 static zil_header_t *
209 zil_header_in_syncing_context(zilog_t *zilog)
210 {
211 	return ((zil_header_t *)zilog->zl_header);
212 }
213 
214 static void
215 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
216 {
217 	zio_cksum_t *zc = &bp->blk_cksum;
218 
219 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0],
220 	    sizeof (zc->zc_word[ZIL_ZC_GUID_0]));
221 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1],
222 	    sizeof (zc->zc_word[ZIL_ZC_GUID_1]));
223 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
224 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
225 }
226 
227 static int
228 zil_kstats_global_update(kstat_t *ksp, int rw)
229 {
230 	zil_kstat_values_t *zs = ksp->ks_data;
231 	ASSERT3P(&zil_stats, ==, zs);
232 
233 	if (rw == KSTAT_WRITE) {
234 		return (SET_ERROR(EACCES));
235 	}
236 
237 	zil_kstat_values_update(zs, &zil_sums_global);
238 
239 	return (0);
240 }
241 
242 /*
243  * Read a log block and make sure it's valid.
244  */
245 static int
246 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
247     blkptr_t *nbp, void *dst, char **end)
248 {
249 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
250 	arc_flags_t aflags = ARC_FLAG_WAIT;
251 	arc_buf_t *abuf = NULL;
252 	zbookmark_phys_t zb;
253 	int error;
254 
255 	if (zilog->zl_header->zh_claim_txg == 0)
256 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
257 
258 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
259 		zio_flags |= ZIO_FLAG_SPECULATIVE;
260 
261 	if (!decrypt)
262 		zio_flags |= ZIO_FLAG_RAW;
263 
264 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
265 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
266 
267 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
268 	    &abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
269 
270 	if (error == 0) {
271 		zio_cksum_t cksum = bp->blk_cksum;
272 
273 		/*
274 		 * Validate the checksummed log block.
275 		 *
276 		 * Sequence numbers should be... sequential.  The checksum
277 		 * verifier for the next block should be bp's checksum plus 1.
278 		 *
279 		 * Also check the log chain linkage and size used.
280 		 */
281 		cksum.zc_word[ZIL_ZC_SEQ]++;
282 
283 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
284 			zil_chain_t *zilc = abuf->b_data;
285 			char *lr = (char *)(zilc + 1);
286 			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
287 
288 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
289 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
290 				error = SET_ERROR(ECKSUM);
291 			} else {
292 				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
293 				memcpy(dst, lr, len);
294 				*end = (char *)dst + len;
295 				*nbp = zilc->zc_next_blk;
296 			}
297 		} else {
298 			char *lr = abuf->b_data;
299 			uint64_t size = BP_GET_LSIZE(bp);
300 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
301 
302 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
303 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
304 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
305 				error = SET_ERROR(ECKSUM);
306 			} else {
307 				ASSERT3U(zilc->zc_nused, <=,
308 				    SPA_OLD_MAXBLOCKSIZE);
309 				memcpy(dst, lr, zilc->zc_nused);
310 				*end = (char *)dst + zilc->zc_nused;
311 				*nbp = zilc->zc_next_blk;
312 			}
313 		}
314 
315 		arc_buf_destroy(abuf, &abuf);
316 	}
317 
318 	return (error);
319 }
320 
321 /*
322  * Read a TX_WRITE log data block.
323  */
324 static int
325 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
326 {
327 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
328 	const blkptr_t *bp = &lr->lr_blkptr;
329 	arc_flags_t aflags = ARC_FLAG_WAIT;
330 	arc_buf_t *abuf = NULL;
331 	zbookmark_phys_t zb;
332 	int error;
333 
334 	if (BP_IS_HOLE(bp)) {
335 		if (wbuf != NULL)
336 			memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length));
337 		return (0);
338 	}
339 
340 	if (zilog->zl_header->zh_claim_txg == 0)
341 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
342 
343 	/*
344 	 * If we are not using the resulting data, we are just checking that
345 	 * it hasn't been corrupted so we don't need to waste CPU time
346 	 * decompressing and decrypting it.
347 	 */
348 	if (wbuf == NULL)
349 		zio_flags |= ZIO_FLAG_RAW;
350 
351 	ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
352 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
353 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
354 
355 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
356 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
357 
358 	if (error == 0) {
359 		if (wbuf != NULL)
360 			memcpy(wbuf, abuf->b_data, arc_buf_size(abuf));
361 		arc_buf_destroy(abuf, &abuf);
362 	}
363 
364 	return (error);
365 }
366 
367 void
368 zil_sums_init(zil_sums_t *zs)
369 {
370 	wmsum_init(&zs->zil_commit_count, 0);
371 	wmsum_init(&zs->zil_commit_writer_count, 0);
372 	wmsum_init(&zs->zil_itx_count, 0);
373 	wmsum_init(&zs->zil_itx_indirect_count, 0);
374 	wmsum_init(&zs->zil_itx_indirect_bytes, 0);
375 	wmsum_init(&zs->zil_itx_copied_count, 0);
376 	wmsum_init(&zs->zil_itx_copied_bytes, 0);
377 	wmsum_init(&zs->zil_itx_needcopy_count, 0);
378 	wmsum_init(&zs->zil_itx_needcopy_bytes, 0);
379 	wmsum_init(&zs->zil_itx_metaslab_normal_count, 0);
380 	wmsum_init(&zs->zil_itx_metaslab_normal_bytes, 0);
381 	wmsum_init(&zs->zil_itx_metaslab_slog_count, 0);
382 	wmsum_init(&zs->zil_itx_metaslab_slog_bytes, 0);
383 }
384 
385 void
386 zil_sums_fini(zil_sums_t *zs)
387 {
388 	wmsum_fini(&zs->zil_commit_count);
389 	wmsum_fini(&zs->zil_commit_writer_count);
390 	wmsum_fini(&zs->zil_itx_count);
391 	wmsum_fini(&zs->zil_itx_indirect_count);
392 	wmsum_fini(&zs->zil_itx_indirect_bytes);
393 	wmsum_fini(&zs->zil_itx_copied_count);
394 	wmsum_fini(&zs->zil_itx_copied_bytes);
395 	wmsum_fini(&zs->zil_itx_needcopy_count);
396 	wmsum_fini(&zs->zil_itx_needcopy_bytes);
397 	wmsum_fini(&zs->zil_itx_metaslab_normal_count);
398 	wmsum_fini(&zs->zil_itx_metaslab_normal_bytes);
399 	wmsum_fini(&zs->zil_itx_metaslab_slog_count);
400 	wmsum_fini(&zs->zil_itx_metaslab_slog_bytes);
401 }
402 
403 void
404 zil_kstat_values_update(zil_kstat_values_t *zs, zil_sums_t *zil_sums)
405 {
406 	zs->zil_commit_count.value.ui64 =
407 	    wmsum_value(&zil_sums->zil_commit_count);
408 	zs->zil_commit_writer_count.value.ui64 =
409 	    wmsum_value(&zil_sums->zil_commit_writer_count);
410 	zs->zil_itx_count.value.ui64 =
411 	    wmsum_value(&zil_sums->zil_itx_count);
412 	zs->zil_itx_indirect_count.value.ui64 =
413 	    wmsum_value(&zil_sums->zil_itx_indirect_count);
414 	zs->zil_itx_indirect_bytes.value.ui64 =
415 	    wmsum_value(&zil_sums->zil_itx_indirect_bytes);
416 	zs->zil_itx_copied_count.value.ui64 =
417 	    wmsum_value(&zil_sums->zil_itx_copied_count);
418 	zs->zil_itx_copied_bytes.value.ui64 =
419 	    wmsum_value(&zil_sums->zil_itx_copied_bytes);
420 	zs->zil_itx_needcopy_count.value.ui64 =
421 	    wmsum_value(&zil_sums->zil_itx_needcopy_count);
422 	zs->zil_itx_needcopy_bytes.value.ui64 =
423 	    wmsum_value(&zil_sums->zil_itx_needcopy_bytes);
424 	zs->zil_itx_metaslab_normal_count.value.ui64 =
425 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_count);
426 	zs->zil_itx_metaslab_normal_bytes.value.ui64 =
427 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_bytes);
428 	zs->zil_itx_metaslab_slog_count.value.ui64 =
429 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_count);
430 	zs->zil_itx_metaslab_slog_bytes.value.ui64 =
431 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_bytes);
432 }
433 
434 /*
435  * Parse the intent log, and call parse_func for each valid record within.
436  */
437 int
438 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
439     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
440     boolean_t decrypt)
441 {
442 	const zil_header_t *zh = zilog->zl_header;
443 	boolean_t claimed = !!zh->zh_claim_txg;
444 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
445 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
446 	uint64_t max_blk_seq = 0;
447 	uint64_t max_lr_seq = 0;
448 	uint64_t blk_count = 0;
449 	uint64_t lr_count = 0;
450 	blkptr_t blk, next_blk = {{{{0}}}};
451 	char *lrbuf, *lrp;
452 	int error = 0;
453 
454 	/*
455 	 * Old logs didn't record the maximum zh_claim_lr_seq.
456 	 */
457 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
458 		claim_lr_seq = UINT64_MAX;
459 
460 	/*
461 	 * Starting at the block pointed to by zh_log we read the log chain.
462 	 * For each block in the chain we strongly check that block to
463 	 * ensure its validity.  We stop when an invalid block is found.
464 	 * For each block pointer in the chain we call parse_blk_func().
465 	 * For each record in each valid block we call parse_lr_func().
466 	 * If the log has been claimed, stop if we encounter a sequence
467 	 * number greater than the highest claimed sequence number.
468 	 */
469 	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
470 	zil_bp_tree_init(zilog);
471 
472 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
473 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
474 		int reclen;
475 		char *end = NULL;
476 
477 		if (blk_seq > claim_blk_seq)
478 			break;
479 
480 		error = parse_blk_func(zilog, &blk, arg, txg);
481 		if (error != 0)
482 			break;
483 		ASSERT3U(max_blk_seq, <, blk_seq);
484 		max_blk_seq = blk_seq;
485 		blk_count++;
486 
487 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
488 			break;
489 
490 		error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
491 		    lrbuf, &end);
492 		if (error != 0) {
493 			if (claimed) {
494 				char name[ZFS_MAX_DATASET_NAME_LEN];
495 
496 				dmu_objset_name(zilog->zl_os, name);
497 
498 				cmn_err(CE_WARN, "ZFS read log block error %d, "
499 				    "dataset %s, seq 0x%llx\n", error, name,
500 				    (u_longlong_t)blk_seq);
501 			}
502 			break;
503 		}
504 
505 		for (lrp = lrbuf; lrp < end; lrp += reclen) {
506 			lr_t *lr = (lr_t *)lrp;
507 			reclen = lr->lrc_reclen;
508 			ASSERT3U(reclen, >=, sizeof (lr_t));
509 			if (lr->lrc_seq > claim_lr_seq)
510 				goto done;
511 
512 			error = parse_lr_func(zilog, lr, arg, txg);
513 			if (error != 0)
514 				goto done;
515 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
516 			max_lr_seq = lr->lrc_seq;
517 			lr_count++;
518 		}
519 	}
520 done:
521 	zilog->zl_parse_error = error;
522 	zilog->zl_parse_blk_seq = max_blk_seq;
523 	zilog->zl_parse_lr_seq = max_lr_seq;
524 	zilog->zl_parse_blk_count = blk_count;
525 	zilog->zl_parse_lr_count = lr_count;
526 
527 	zil_bp_tree_fini(zilog);
528 	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
529 
530 	return (error);
531 }
532 
533 static int
534 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
535     uint64_t first_txg)
536 {
537 	(void) tx;
538 	ASSERT(!BP_IS_HOLE(bp));
539 
540 	/*
541 	 * As we call this function from the context of a rewind to a
542 	 * checkpoint, each ZIL block whose txg is later than the txg
543 	 * that we rewind to is invalid. Thus, we return -1 so
544 	 * zil_parse() doesn't attempt to read it.
545 	 */
546 	if (bp->blk_birth >= first_txg)
547 		return (-1);
548 
549 	if (zil_bp_tree_add(zilog, bp) != 0)
550 		return (0);
551 
552 	zio_free(zilog->zl_spa, first_txg, bp);
553 	return (0);
554 }
555 
556 static int
557 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
558     uint64_t first_txg)
559 {
560 	(void) zilog, (void) lrc, (void) tx, (void) first_txg;
561 	return (0);
562 }
563 
564 static int
565 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
566     uint64_t first_txg)
567 {
568 	/*
569 	 * Claim log block if not already committed and not already claimed.
570 	 * If tx == NULL, just verify that the block is claimable.
571 	 */
572 	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
573 	    zil_bp_tree_add(zilog, bp) != 0)
574 		return (0);
575 
576 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
577 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
578 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
579 }
580 
581 static int
582 zil_claim_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t first_txg)
583 {
584 	lr_write_t *lr = (lr_write_t *)lrc;
585 	int error;
586 
587 	ASSERT(lrc->lrc_txtype == TX_WRITE);
588 
589 	/*
590 	 * If the block is not readable, don't claim it.  This can happen
591 	 * in normal operation when a log block is written to disk before
592 	 * some of the dmu_sync() blocks it points to.  In this case, the
593 	 * transaction cannot have been committed to anyone (we would have
594 	 * waited for all writes to be stable first), so it is semantically
595 	 * correct to declare this the end of the log.
596 	 */
597 	if (lr->lr_blkptr.blk_birth >= first_txg) {
598 		error = zil_read_log_data(zilog, lr, NULL);
599 		if (error != 0)
600 			return (error);
601 	}
602 
603 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
604 }
605 
606 static int
607 zil_claim_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
608 {
609 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
610 	const blkptr_t *bp;
611 	spa_t *spa;
612 	uint_t ii;
613 
614 	ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE);
615 
616 	if (tx == NULL) {
617 		return (0);
618 	}
619 
620 	/*
621 	 * XXX: Do we need to byteswap lr?
622 	 */
623 
624 	spa = zilog->zl_spa;
625 
626 	for (ii = 0; ii < lr->lr_nbps; ii++) {
627 		bp = &lr->lr_bps[ii];
628 
629 		/*
630 		 * When data in embedded into BP there is no need to create
631 		 * BRT entry as there is no data block. Just copy the BP as
632 		 * it contains the data.
633 		 */
634 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
635 			brt_pending_add(spa, bp, tx);
636 		}
637 	}
638 
639 	return (0);
640 }
641 
642 static int
643 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
644     uint64_t first_txg)
645 {
646 
647 	switch (lrc->lrc_txtype) {
648 	case TX_WRITE:
649 		return (zil_claim_write(zilog, lrc, tx, first_txg));
650 	case TX_CLONE_RANGE:
651 		return (zil_claim_clone_range(zilog, lrc, tx));
652 	default:
653 		return (0);
654 	}
655 }
656 
657 static int
658 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
659     uint64_t claim_txg)
660 {
661 	(void) claim_txg;
662 
663 	zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
664 
665 	return (0);
666 }
667 
668 static int
669 zil_free_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t claim_txg)
670 {
671 	lr_write_t *lr = (lr_write_t *)lrc;
672 	blkptr_t *bp = &lr->lr_blkptr;
673 
674 	ASSERT(lrc->lrc_txtype == TX_WRITE);
675 
676 	/*
677 	 * If we previously claimed it, we need to free it.
678 	 */
679 	if (bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
680 	    !BP_IS_HOLE(bp)) {
681 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
682 	}
683 
684 	return (0);
685 }
686 
687 static int
688 zil_free_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
689 {
690 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
691 	const blkptr_t *bp;
692 	spa_t *spa;
693 	uint_t ii;
694 
695 	ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE);
696 
697 	if (tx == NULL) {
698 		return (0);
699 	}
700 
701 	spa = zilog->zl_spa;
702 
703 	for (ii = 0; ii < lr->lr_nbps; ii++) {
704 		bp = &lr->lr_bps[ii];
705 
706 		if (!BP_IS_HOLE(bp)) {
707 			zio_free(spa, dmu_tx_get_txg(tx), bp);
708 		}
709 	}
710 
711 	return (0);
712 }
713 
714 static int
715 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
716     uint64_t claim_txg)
717 {
718 
719 	if (claim_txg == 0) {
720 		return (0);
721 	}
722 
723 	switch (lrc->lrc_txtype) {
724 	case TX_WRITE:
725 		return (zil_free_write(zilog, lrc, tx, claim_txg));
726 	case TX_CLONE_RANGE:
727 		return (zil_free_clone_range(zilog, lrc, tx));
728 	default:
729 		return (0);
730 	}
731 }
732 
733 static int
734 zil_lwb_vdev_compare(const void *x1, const void *x2)
735 {
736 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
737 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
738 
739 	return (TREE_CMP(v1, v2));
740 }
741 
742 static lwb_t *
743 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg,
744     boolean_t fastwrite)
745 {
746 	lwb_t *lwb;
747 
748 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
749 	lwb->lwb_zilog = zilog;
750 	lwb->lwb_blk = *bp;
751 	lwb->lwb_fastwrite = fastwrite;
752 	lwb->lwb_slog = slog;
753 	lwb->lwb_state = LWB_STATE_CLOSED;
754 	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
755 	lwb->lwb_max_txg = txg;
756 	lwb->lwb_write_zio = NULL;
757 	lwb->lwb_root_zio = NULL;
758 	lwb->lwb_issued_timestamp = 0;
759 	lwb->lwb_issued_txg = 0;
760 	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
761 		lwb->lwb_nused = sizeof (zil_chain_t);
762 		lwb->lwb_sz = BP_GET_LSIZE(bp);
763 	} else {
764 		lwb->lwb_nused = 0;
765 		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
766 	}
767 
768 	mutex_enter(&zilog->zl_lock);
769 	list_insert_tail(&zilog->zl_lwb_list, lwb);
770 	mutex_exit(&zilog->zl_lock);
771 
772 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
773 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
774 	VERIFY(list_is_empty(&lwb->lwb_waiters));
775 	VERIFY(list_is_empty(&lwb->lwb_itxs));
776 
777 	return (lwb);
778 }
779 
780 static void
781 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
782 {
783 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
784 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
785 	VERIFY(list_is_empty(&lwb->lwb_waiters));
786 	VERIFY(list_is_empty(&lwb->lwb_itxs));
787 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
788 	ASSERT3P(lwb->lwb_write_zio, ==, NULL);
789 	ASSERT3P(lwb->lwb_root_zio, ==, NULL);
790 	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
791 	ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
792 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
793 
794 	/*
795 	 * Clear the zilog's field to indicate this lwb is no longer
796 	 * valid, and prevent use-after-free errors.
797 	 */
798 	if (zilog->zl_last_lwb_opened == lwb)
799 		zilog->zl_last_lwb_opened = NULL;
800 
801 	kmem_cache_free(zil_lwb_cache, lwb);
802 }
803 
804 /*
805  * Called when we create in-memory log transactions so that we know
806  * to cleanup the itxs at the end of spa_sync().
807  */
808 static void
809 zilog_dirty(zilog_t *zilog, uint64_t txg)
810 {
811 	dsl_pool_t *dp = zilog->zl_dmu_pool;
812 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
813 
814 	ASSERT(spa_writeable(zilog->zl_spa));
815 
816 	if (ds->ds_is_snapshot)
817 		panic("dirtying snapshot!");
818 
819 	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
820 		/* up the hold count until we can be written out */
821 		dmu_buf_add_ref(ds->ds_dbuf, zilog);
822 
823 		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
824 	}
825 }
826 
827 /*
828  * Determine if the zil is dirty in the specified txg. Callers wanting to
829  * ensure that the dirty state does not change must hold the itxg_lock for
830  * the specified txg. Holding the lock will ensure that the zil cannot be
831  * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
832  * state.
833  */
834 static boolean_t __maybe_unused
835 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
836 {
837 	dsl_pool_t *dp = zilog->zl_dmu_pool;
838 
839 	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
840 		return (B_TRUE);
841 	return (B_FALSE);
842 }
843 
844 /*
845  * Determine if the zil is dirty. The zil is considered dirty if it has
846  * any pending itx records that have not been cleaned by zil_clean().
847  */
848 static boolean_t
849 zilog_is_dirty(zilog_t *zilog)
850 {
851 	dsl_pool_t *dp = zilog->zl_dmu_pool;
852 
853 	for (int t = 0; t < TXG_SIZE; t++) {
854 		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
855 			return (B_TRUE);
856 	}
857 	return (B_FALSE);
858 }
859 
860 /*
861  * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
862  * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
863  * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
864  * zil_commit.
865  */
866 static void
867 zil_commit_activate_saxattr_feature(zilog_t *zilog)
868 {
869 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
870 	uint64_t txg = 0;
871 	dmu_tx_t *tx = NULL;
872 
873 	if (spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
874 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL &&
875 	    !dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)) {
876 		tx = dmu_tx_create(zilog->zl_os);
877 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
878 		dsl_dataset_dirty(ds, tx);
879 		txg = dmu_tx_get_txg(tx);
880 
881 		mutex_enter(&ds->ds_lock);
882 		ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
883 		    (void *)B_TRUE;
884 		mutex_exit(&ds->ds_lock);
885 		dmu_tx_commit(tx);
886 		txg_wait_synced(zilog->zl_dmu_pool, txg);
887 	}
888 }
889 
890 /*
891  * Create an on-disk intent log.
892  */
893 static lwb_t *
894 zil_create(zilog_t *zilog)
895 {
896 	const zil_header_t *zh = zilog->zl_header;
897 	lwb_t *lwb = NULL;
898 	uint64_t txg = 0;
899 	dmu_tx_t *tx = NULL;
900 	blkptr_t blk;
901 	int error = 0;
902 	boolean_t fastwrite = FALSE;
903 	boolean_t slog = FALSE;
904 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
905 
906 
907 	/*
908 	 * Wait for any previous destroy to complete.
909 	 */
910 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
911 
912 	ASSERT(zh->zh_claim_txg == 0);
913 	ASSERT(zh->zh_replay_seq == 0);
914 
915 	blk = zh->zh_log;
916 
917 	/*
918 	 * Allocate an initial log block if:
919 	 *    - there isn't one already
920 	 *    - the existing block is the wrong endianness
921 	 */
922 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
923 		tx = dmu_tx_create(zilog->zl_os);
924 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
925 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
926 		txg = dmu_tx_get_txg(tx);
927 
928 		if (!BP_IS_HOLE(&blk)) {
929 			zio_free(zilog->zl_spa, txg, &blk);
930 			BP_ZERO(&blk);
931 		}
932 
933 		error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
934 		    ZIL_MIN_BLKSZ, &slog);
935 		fastwrite = TRUE;
936 
937 		if (error == 0)
938 			zil_init_log_chain(zilog, &blk);
939 	}
940 
941 	/*
942 	 * Allocate a log write block (lwb) for the first log block.
943 	 */
944 	if (error == 0)
945 		lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
946 
947 	/*
948 	 * If we just allocated the first log block, commit our transaction
949 	 * and wait for zil_sync() to stuff the block pointer into zh_log.
950 	 * (zh is part of the MOS, so we cannot modify it in open context.)
951 	 */
952 	if (tx != NULL) {
953 		/*
954 		 * If "zilsaxattr" feature is enabled on zpool, then activate
955 		 * it now when we're creating the ZIL chain. We can't wait with
956 		 * this until we write the first xattr log record because we
957 		 * need to wait for the feature activation to sync out.
958 		 */
959 		if (spa_feature_is_enabled(zilog->zl_spa,
960 		    SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) !=
961 		    DMU_OST_ZVOL) {
962 			mutex_enter(&ds->ds_lock);
963 			ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
964 			    (void *)B_TRUE;
965 			mutex_exit(&ds->ds_lock);
966 		}
967 
968 		dmu_tx_commit(tx);
969 		txg_wait_synced(zilog->zl_dmu_pool, txg);
970 	} else {
971 		/*
972 		 * This branch covers the case where we enable the feature on a
973 		 * zpool that has existing ZIL headers.
974 		 */
975 		zil_commit_activate_saxattr_feature(zilog);
976 	}
977 	IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
978 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL,
979 	    dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR));
980 
981 	ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
982 	IMPLY(error == 0, lwb != NULL);
983 
984 	return (lwb);
985 }
986 
987 /*
988  * In one tx, free all log blocks and clear the log header. If keep_first
989  * is set, then we're replaying a log with no content. We want to keep the
990  * first block, however, so that the first synchronous transaction doesn't
991  * require a txg_wait_synced() in zil_create(). We don't need to
992  * txg_wait_synced() here either when keep_first is set, because both
993  * zil_create() and zil_destroy() will wait for any in-progress destroys
994  * to complete.
995  * Return B_TRUE if there were any entries to replay.
996  */
997 boolean_t
998 zil_destroy(zilog_t *zilog, boolean_t keep_first)
999 {
1000 	const zil_header_t *zh = zilog->zl_header;
1001 	lwb_t *lwb;
1002 	dmu_tx_t *tx;
1003 	uint64_t txg;
1004 
1005 	/*
1006 	 * Wait for any previous destroy to complete.
1007 	 */
1008 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1009 
1010 	zilog->zl_old_header = *zh;		/* debugging aid */
1011 
1012 	if (BP_IS_HOLE(&zh->zh_log))
1013 		return (B_FALSE);
1014 
1015 	tx = dmu_tx_create(zilog->zl_os);
1016 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1017 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1018 	txg = dmu_tx_get_txg(tx);
1019 
1020 	mutex_enter(&zilog->zl_lock);
1021 
1022 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
1023 	zilog->zl_destroy_txg = txg;
1024 	zilog->zl_keep_first = keep_first;
1025 
1026 	if (!list_is_empty(&zilog->zl_lwb_list)) {
1027 		ASSERT(zh->zh_claim_txg == 0);
1028 		VERIFY(!keep_first);
1029 		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1030 			if (lwb->lwb_fastwrite)
1031 				metaslab_fastwrite_unmark(zilog->zl_spa,
1032 				    &lwb->lwb_blk);
1033 
1034 			list_remove(&zilog->zl_lwb_list, lwb);
1035 			if (lwb->lwb_buf != NULL)
1036 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1037 			zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
1038 			zil_free_lwb(zilog, lwb);
1039 		}
1040 	} else if (!keep_first) {
1041 		zil_destroy_sync(zilog, tx);
1042 	}
1043 	mutex_exit(&zilog->zl_lock);
1044 
1045 	dmu_tx_commit(tx);
1046 
1047 	return (B_TRUE);
1048 }
1049 
1050 void
1051 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
1052 {
1053 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1054 	(void) zil_parse(zilog, zil_free_log_block,
1055 	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
1056 }
1057 
1058 int
1059 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
1060 {
1061 	dmu_tx_t *tx = txarg;
1062 	zilog_t *zilog;
1063 	uint64_t first_txg;
1064 	zil_header_t *zh;
1065 	objset_t *os;
1066 	int error;
1067 
1068 	error = dmu_objset_own_obj(dp, ds->ds_object,
1069 	    DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
1070 	if (error != 0) {
1071 		/*
1072 		 * EBUSY indicates that the objset is inconsistent, in which
1073 		 * case it can not have a ZIL.
1074 		 */
1075 		if (error != EBUSY) {
1076 			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
1077 			    (unsigned long long)ds->ds_object, error);
1078 		}
1079 
1080 		return (0);
1081 	}
1082 
1083 	zilog = dmu_objset_zil(os);
1084 	zh = zil_header_in_syncing_context(zilog);
1085 	ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
1086 	first_txg = spa_min_claim_txg(zilog->zl_spa);
1087 
1088 	/*
1089 	 * If the spa_log_state is not set to be cleared, check whether
1090 	 * the current uberblock is a checkpoint one and if the current
1091 	 * header has been claimed before moving on.
1092 	 *
1093 	 * If the current uberblock is a checkpointed uberblock then
1094 	 * one of the following scenarios took place:
1095 	 *
1096 	 * 1] We are currently rewinding to the checkpoint of the pool.
1097 	 * 2] We crashed in the middle of a checkpoint rewind but we
1098 	 *    did manage to write the checkpointed uberblock to the
1099 	 *    vdev labels, so when we tried to import the pool again
1100 	 *    the checkpointed uberblock was selected from the import
1101 	 *    procedure.
1102 	 *
1103 	 * In both cases we want to zero out all the ZIL blocks, except
1104 	 * the ones that have been claimed at the time of the checkpoint
1105 	 * (their zh_claim_txg != 0). The reason is that these blocks
1106 	 * may be corrupted since we may have reused their locations on
1107 	 * disk after we took the checkpoint.
1108 	 *
1109 	 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
1110 	 * when we first figure out whether the current uberblock is
1111 	 * checkpointed or not. Unfortunately, that would discard all
1112 	 * the logs, including the ones that are claimed, and we would
1113 	 * leak space.
1114 	 */
1115 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
1116 	    (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1117 	    zh->zh_claim_txg == 0)) {
1118 		if (!BP_IS_HOLE(&zh->zh_log)) {
1119 			(void) zil_parse(zilog, zil_clear_log_block,
1120 			    zil_noop_log_record, tx, first_txg, B_FALSE);
1121 		}
1122 		BP_ZERO(&zh->zh_log);
1123 		if (os->os_encrypted)
1124 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1125 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1126 		dmu_objset_disown(os, B_FALSE, FTAG);
1127 		return (0);
1128 	}
1129 
1130 	/*
1131 	 * If we are not rewinding and opening the pool normally, then
1132 	 * the min_claim_txg should be equal to the first txg of the pool.
1133 	 */
1134 	ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
1135 
1136 	/*
1137 	 * Claim all log blocks if we haven't already done so, and remember
1138 	 * the highest claimed sequence number.  This ensures that if we can
1139 	 * read only part of the log now (e.g. due to a missing device),
1140 	 * but we can read the entire log later, we will not try to replay
1141 	 * or destroy beyond the last block we successfully claimed.
1142 	 */
1143 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
1144 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
1145 		(void) zil_parse(zilog, zil_claim_log_block,
1146 		    zil_claim_log_record, tx, first_txg, B_FALSE);
1147 		zh->zh_claim_txg = first_txg;
1148 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
1149 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
1150 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
1151 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
1152 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
1153 		if (os->os_encrypted)
1154 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1155 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1156 	}
1157 
1158 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
1159 	dmu_objset_disown(os, B_FALSE, FTAG);
1160 	return (0);
1161 }
1162 
1163 /*
1164  * Check the log by walking the log chain.
1165  * Checksum errors are ok as they indicate the end of the chain.
1166  * Any other error (no device or read failure) returns an error.
1167  */
1168 int
1169 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
1170 {
1171 	(void) dp;
1172 	zilog_t *zilog;
1173 	objset_t *os;
1174 	blkptr_t *bp;
1175 	int error;
1176 
1177 	ASSERT(tx == NULL);
1178 
1179 	error = dmu_objset_from_ds(ds, &os);
1180 	if (error != 0) {
1181 		cmn_err(CE_WARN, "can't open objset %llu, error %d",
1182 		    (unsigned long long)ds->ds_object, error);
1183 		return (0);
1184 	}
1185 
1186 	zilog = dmu_objset_zil(os);
1187 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
1188 
1189 	if (!BP_IS_HOLE(bp)) {
1190 		vdev_t *vd;
1191 		boolean_t valid = B_TRUE;
1192 
1193 		/*
1194 		 * Check the first block and determine if it's on a log device
1195 		 * which may have been removed or faulted prior to loading this
1196 		 * pool.  If so, there's no point in checking the rest of the
1197 		 * log as its content should have already been synced to the
1198 		 * pool.
1199 		 */
1200 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
1201 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
1202 		if (vd->vdev_islog && vdev_is_dead(vd))
1203 			valid = vdev_log_state_valid(vd);
1204 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
1205 
1206 		if (!valid)
1207 			return (0);
1208 
1209 		/*
1210 		 * Check whether the current uberblock is checkpointed (e.g.
1211 		 * we are rewinding) and whether the current header has been
1212 		 * claimed or not. If it hasn't then skip verifying it. We
1213 		 * do this because its ZIL blocks may be part of the pool's
1214 		 * state before the rewind, which is no longer valid.
1215 		 */
1216 		zil_header_t *zh = zil_header_in_syncing_context(zilog);
1217 		if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1218 		    zh->zh_claim_txg == 0)
1219 			return (0);
1220 	}
1221 
1222 	/*
1223 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
1224 	 * any blocks, but just determine whether it is possible to do so.
1225 	 * In addition to checking the log chain, zil_claim_log_block()
1226 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
1227 	 * which will update spa_max_claim_txg.  See spa_load() for details.
1228 	 */
1229 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
1230 	    zilog->zl_header->zh_claim_txg ? -1ULL :
1231 	    spa_min_claim_txg(os->os_spa), B_FALSE);
1232 
1233 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
1234 }
1235 
1236 /*
1237  * When an itx is "skipped", this function is used to properly mark the
1238  * waiter as "done, and signal any thread(s) waiting on it. An itx can
1239  * be skipped (and not committed to an lwb) for a variety of reasons,
1240  * one of them being that the itx was committed via spa_sync(), prior to
1241  * it being committed to an lwb; this can happen if a thread calling
1242  * zil_commit() is racing with spa_sync().
1243  */
1244 static void
1245 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
1246 {
1247 	mutex_enter(&zcw->zcw_lock);
1248 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1249 	zcw->zcw_done = B_TRUE;
1250 	cv_broadcast(&zcw->zcw_cv);
1251 	mutex_exit(&zcw->zcw_lock);
1252 }
1253 
1254 /*
1255  * This function is used when the given waiter is to be linked into an
1256  * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1257  * At this point, the waiter will no longer be referenced by the itx,
1258  * and instead, will be referenced by the lwb.
1259  */
1260 static void
1261 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1262 {
1263 	/*
1264 	 * The lwb_waiters field of the lwb is protected by the zilog's
1265 	 * zl_lock, thus it must be held when calling this function.
1266 	 */
1267 	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1268 
1269 	mutex_enter(&zcw->zcw_lock);
1270 	ASSERT(!list_link_active(&zcw->zcw_node));
1271 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1272 	ASSERT3P(lwb, !=, NULL);
1273 	ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
1274 	    lwb->lwb_state == LWB_STATE_ISSUED ||
1275 	    lwb->lwb_state == LWB_STATE_WRITE_DONE);
1276 
1277 	list_insert_tail(&lwb->lwb_waiters, zcw);
1278 	zcw->zcw_lwb = lwb;
1279 	mutex_exit(&zcw->zcw_lock);
1280 }
1281 
1282 /*
1283  * This function is used when zio_alloc_zil() fails to allocate a ZIL
1284  * block, and the given waiter must be linked to the "nolwb waiters"
1285  * list inside of zil_process_commit_list().
1286  */
1287 static void
1288 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1289 {
1290 	mutex_enter(&zcw->zcw_lock);
1291 	ASSERT(!list_link_active(&zcw->zcw_node));
1292 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1293 	list_insert_tail(nolwb, zcw);
1294 	mutex_exit(&zcw->zcw_lock);
1295 }
1296 
1297 void
1298 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1299 {
1300 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1301 	avl_index_t where;
1302 	zil_vdev_node_t *zv, zvsearch;
1303 	int ndvas = BP_GET_NDVAS(bp);
1304 	int i;
1305 
1306 	if (zil_nocacheflush)
1307 		return;
1308 
1309 	mutex_enter(&lwb->lwb_vdev_lock);
1310 	for (i = 0; i < ndvas; i++) {
1311 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1312 		if (avl_find(t, &zvsearch, &where) == NULL) {
1313 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1314 			zv->zv_vdev = zvsearch.zv_vdev;
1315 			avl_insert(t, zv, where);
1316 		}
1317 	}
1318 	mutex_exit(&lwb->lwb_vdev_lock);
1319 }
1320 
1321 static void
1322 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1323 {
1324 	avl_tree_t *src = &lwb->lwb_vdev_tree;
1325 	avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1326 	void *cookie = NULL;
1327 	zil_vdev_node_t *zv;
1328 
1329 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1330 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1331 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1332 
1333 	/*
1334 	 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1335 	 * not need the protection of lwb_vdev_lock (it will only be modified
1336 	 * while holding zilog->zl_lock) as its writes and those of its
1337 	 * children have all completed.  The younger 'nlwb' may be waiting on
1338 	 * future writes to additional vdevs.
1339 	 */
1340 	mutex_enter(&nlwb->lwb_vdev_lock);
1341 	/*
1342 	 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1343 	 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1344 	 */
1345 	while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1346 		avl_index_t where;
1347 
1348 		if (avl_find(dst, zv, &where) == NULL) {
1349 			avl_insert(dst, zv, where);
1350 		} else {
1351 			kmem_free(zv, sizeof (*zv));
1352 		}
1353 	}
1354 	mutex_exit(&nlwb->lwb_vdev_lock);
1355 }
1356 
1357 void
1358 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1359 {
1360 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1361 }
1362 
1363 /*
1364  * This function is a called after all vdevs associated with a given lwb
1365  * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1366  * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1367  * all "previous" lwb's will have completed before this function is
1368  * called; i.e. this function is called for all previous lwbs before
1369  * it's called for "this" lwb (enforced via zio the dependencies
1370  * configured in zil_lwb_set_zio_dependency()).
1371  *
1372  * The intention is for this function to be called as soon as the
1373  * contents of an lwb are considered "stable" on disk, and will survive
1374  * any sudden loss of power. At this point, any threads waiting for the
1375  * lwb to reach this state are signalled, and the "waiter" structures
1376  * are marked "done".
1377  */
1378 static void
1379 zil_lwb_flush_vdevs_done(zio_t *zio)
1380 {
1381 	lwb_t *lwb = zio->io_private;
1382 	zilog_t *zilog = lwb->lwb_zilog;
1383 	zil_commit_waiter_t *zcw;
1384 	itx_t *itx;
1385 	uint64_t txg;
1386 
1387 	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1388 
1389 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1390 
1391 	mutex_enter(&zilog->zl_lock);
1392 
1393 	/*
1394 	 * If we have had an allocation failure and the txg is
1395 	 * waiting to sync then we want zil_sync() to remove the lwb so
1396 	 * that it's not picked up as the next new one in
1397 	 * zil_process_commit_list(). zil_sync() will only remove the
1398 	 * lwb if lwb_buf is null.
1399 	 */
1400 	lwb->lwb_buf = NULL;
1401 
1402 	ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1403 	zilog->zl_last_lwb_latency = (zilog->zl_last_lwb_latency * 3 +
1404 	    gethrtime() - lwb->lwb_issued_timestamp) / 4;
1405 
1406 	lwb->lwb_root_zio = NULL;
1407 
1408 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1409 	lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1410 
1411 	if (zilog->zl_last_lwb_opened == lwb) {
1412 		/*
1413 		 * Remember the highest committed log sequence number
1414 		 * for ztest. We only update this value when all the log
1415 		 * writes succeeded, because ztest wants to ASSERT that
1416 		 * it got the whole log chain.
1417 		 */
1418 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1419 	}
1420 
1421 	while ((itx = list_head(&lwb->lwb_itxs)) != NULL) {
1422 		list_remove(&lwb->lwb_itxs, itx);
1423 		zil_itx_destroy(itx);
1424 	}
1425 
1426 	while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1427 		mutex_enter(&zcw->zcw_lock);
1428 
1429 		ASSERT(list_link_active(&zcw->zcw_node));
1430 		list_remove(&lwb->lwb_waiters, zcw);
1431 
1432 		ASSERT3P(zcw->zcw_lwb, ==, lwb);
1433 		zcw->zcw_lwb = NULL;
1434 		/*
1435 		 * We expect any ZIO errors from child ZIOs to have been
1436 		 * propagated "up" to this specific LWB's root ZIO, in
1437 		 * order for this error handling to work correctly. This
1438 		 * includes ZIO errors from either this LWB's write or
1439 		 * flush, as well as any errors from other dependent LWBs
1440 		 * (e.g. a root LWB ZIO that might be a child of this LWB).
1441 		 *
1442 		 * With that said, it's important to note that LWB flush
1443 		 * errors are not propagated up to the LWB root ZIO.
1444 		 * This is incorrect behavior, and results in VDEV flush
1445 		 * errors not being handled correctly here. See the
1446 		 * comment above the call to "zio_flush" for details.
1447 		 */
1448 
1449 		zcw->zcw_zio_error = zio->io_error;
1450 
1451 		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1452 		zcw->zcw_done = B_TRUE;
1453 		cv_broadcast(&zcw->zcw_cv);
1454 
1455 		mutex_exit(&zcw->zcw_lock);
1456 	}
1457 
1458 	mutex_exit(&zilog->zl_lock);
1459 
1460 	mutex_enter(&zilog->zl_lwb_io_lock);
1461 	txg = lwb->lwb_issued_txg;
1462 	ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0);
1463 	zilog->zl_lwb_inflight[txg & TXG_MASK]--;
1464 	if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0)
1465 		cv_broadcast(&zilog->zl_lwb_io_cv);
1466 	mutex_exit(&zilog->zl_lwb_io_lock);
1467 }
1468 
1469 /*
1470  * Wait for the completion of all issued write/flush of that txg provided.
1471  * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1472  */
1473 static void
1474 zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg)
1475 {
1476 	ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa));
1477 
1478 	mutex_enter(&zilog->zl_lwb_io_lock);
1479 	while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0)
1480 		cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock);
1481 	mutex_exit(&zilog->zl_lwb_io_lock);
1482 
1483 #ifdef ZFS_DEBUG
1484 	mutex_enter(&zilog->zl_lock);
1485 	mutex_enter(&zilog->zl_lwb_io_lock);
1486 	lwb_t *lwb = list_head(&zilog->zl_lwb_list);
1487 	while (lwb != NULL && lwb->lwb_max_txg <= txg) {
1488 		if (lwb->lwb_issued_txg <= txg) {
1489 			ASSERT(lwb->lwb_state != LWB_STATE_ISSUED);
1490 			ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE);
1491 			IMPLY(lwb->lwb_issued_txg > 0,
1492 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
1493 		}
1494 		IMPLY(lwb->lwb_state == LWB_STATE_FLUSH_DONE,
1495 		    lwb->lwb_buf == NULL);
1496 		lwb = list_next(&zilog->zl_lwb_list, lwb);
1497 	}
1498 	mutex_exit(&zilog->zl_lwb_io_lock);
1499 	mutex_exit(&zilog->zl_lock);
1500 #endif
1501 }
1502 
1503 /*
1504  * This is called when an lwb's write zio completes. The callback's
1505  * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1506  * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1507  * in writing out this specific lwb's data, and in the case that cache
1508  * flushes have been deferred, vdevs involved in writing the data for
1509  * previous lwbs. The writes corresponding to all the vdevs in the
1510  * lwb_vdev_tree will have completed by the time this is called, due to
1511  * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1512  * which takes deferred flushes into account. The lwb will be "done"
1513  * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1514  * completion callback for the lwb's root zio.
1515  */
1516 static void
1517 zil_lwb_write_done(zio_t *zio)
1518 {
1519 	lwb_t *lwb = zio->io_private;
1520 	spa_t *spa = zio->io_spa;
1521 	zilog_t *zilog = lwb->lwb_zilog;
1522 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1523 	void *cookie = NULL;
1524 	zil_vdev_node_t *zv;
1525 	lwb_t *nlwb;
1526 
1527 	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1528 
1529 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1530 	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1531 	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1532 	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1533 	ASSERT(!BP_IS_GANG(zio->io_bp));
1534 	ASSERT(!BP_IS_HOLE(zio->io_bp));
1535 	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1536 
1537 	abd_free(zio->io_abd);
1538 
1539 	mutex_enter(&zilog->zl_lock);
1540 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1541 	lwb->lwb_state = LWB_STATE_WRITE_DONE;
1542 	lwb->lwb_write_zio = NULL;
1543 	lwb->lwb_fastwrite = FALSE;
1544 	nlwb = list_next(&zilog->zl_lwb_list, lwb);
1545 	mutex_exit(&zilog->zl_lock);
1546 
1547 	if (avl_numnodes(t) == 0)
1548 		return;
1549 
1550 	/*
1551 	 * If there was an IO error, we're not going to call zio_flush()
1552 	 * on these vdevs, so we simply empty the tree and free the
1553 	 * nodes. We avoid calling zio_flush() since there isn't any
1554 	 * good reason for doing so, after the lwb block failed to be
1555 	 * written out.
1556 	 *
1557 	 * Additionally, we don't perform any further error handling at
1558 	 * this point (e.g. setting "zcw_zio_error" appropriately), as
1559 	 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1560 	 * we expect any error seen here, to have been propagated to
1561 	 * that function).
1562 	 */
1563 	if (zio->io_error != 0) {
1564 		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1565 			kmem_free(zv, sizeof (*zv));
1566 		return;
1567 	}
1568 
1569 	/*
1570 	 * If this lwb does not have any threads waiting for it to
1571 	 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1572 	 * command to the vdevs written to by "this" lwb, and instead
1573 	 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1574 	 * command for those vdevs. Thus, we merge the vdev tree of
1575 	 * "this" lwb with the vdev tree of the "next" lwb in the list,
1576 	 * and assume the "next" lwb will handle flushing the vdevs (or
1577 	 * deferring the flush(s) again).
1578 	 *
1579 	 * This is a useful performance optimization, especially for
1580 	 * workloads with lots of async write activity and few sync
1581 	 * write and/or fsync activity, as it has the potential to
1582 	 * coalesce multiple flush commands to a vdev into one.
1583 	 */
1584 	if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) {
1585 		zil_lwb_flush_defer(lwb, nlwb);
1586 		ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1587 		return;
1588 	}
1589 
1590 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1591 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1592 		if (vd != NULL) {
1593 			/*
1594 			 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1595 			 * always used within "zio_flush". This means,
1596 			 * any errors when flushing the vdev(s), will
1597 			 * (unfortunately) not be handled correctly,
1598 			 * since these "zio_flush" errors will not be
1599 			 * propagated up to "zil_lwb_flush_vdevs_done".
1600 			 */
1601 			zio_flush(lwb->lwb_root_zio, vd);
1602 		}
1603 		kmem_free(zv, sizeof (*zv));
1604 	}
1605 }
1606 
1607 static void
1608 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1609 {
1610 	lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1611 
1612 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1613 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
1614 
1615 	/*
1616 	 * The zilog's "zl_last_lwb_opened" field is used to build the
1617 	 * lwb/zio dependency chain, which is used to preserve the
1618 	 * ordering of lwb completions that is required by the semantics
1619 	 * of the ZIL. Each new lwb zio becomes a parent of the
1620 	 * "previous" lwb zio, such that the new lwb's zio cannot
1621 	 * complete until the "previous" lwb's zio completes.
1622 	 *
1623 	 * This is required by the semantics of zil_commit(); the commit
1624 	 * waiters attached to the lwbs will be woken in the lwb zio's
1625 	 * completion callback, so this zio dependency graph ensures the
1626 	 * waiters are woken in the correct order (the same order the
1627 	 * lwbs were created).
1628 	 */
1629 	if (last_lwb_opened != NULL &&
1630 	    last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) {
1631 		ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1632 		    last_lwb_opened->lwb_state == LWB_STATE_ISSUED ||
1633 		    last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE);
1634 
1635 		ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1636 		zio_add_child(lwb->lwb_root_zio,
1637 		    last_lwb_opened->lwb_root_zio);
1638 
1639 		/*
1640 		 * If the previous lwb's write hasn't already completed,
1641 		 * we also want to order the completion of the lwb write
1642 		 * zios (above, we only order the completion of the lwb
1643 		 * root zios). This is required because of how we can
1644 		 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1645 		 *
1646 		 * When the DKIOCFLUSHWRITECACHE commands are deferred,
1647 		 * the previous lwb will rely on this lwb to flush the
1648 		 * vdevs written to by that previous lwb. Thus, we need
1649 		 * to ensure this lwb doesn't issue the flush until
1650 		 * after the previous lwb's write completes. We ensure
1651 		 * this ordering by setting the zio parent/child
1652 		 * relationship here.
1653 		 *
1654 		 * Without this relationship on the lwb's write zio,
1655 		 * it's possible for this lwb's write to complete prior
1656 		 * to the previous lwb's write completing; and thus, the
1657 		 * vdevs for the previous lwb would be flushed prior to
1658 		 * that lwb's data being written to those vdevs (the
1659 		 * vdevs are flushed in the lwb write zio's completion
1660 		 * handler, zil_lwb_write_done()).
1661 		 */
1662 		if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) {
1663 			ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1664 			    last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1665 
1666 			ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL);
1667 			zio_add_child(lwb->lwb_write_zio,
1668 			    last_lwb_opened->lwb_write_zio);
1669 		}
1670 	}
1671 }
1672 
1673 
1674 /*
1675  * This function's purpose is to "open" an lwb such that it is ready to
1676  * accept new itxs being committed to it. To do this, the lwb's zio
1677  * structures are created, and linked to the lwb. This function is
1678  * idempotent; if the passed in lwb has already been opened, this
1679  * function is essentially a no-op.
1680  */
1681 static void
1682 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1683 {
1684 	zbookmark_phys_t zb;
1685 	zio_priority_t prio;
1686 
1687 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1688 	ASSERT3P(lwb, !=, NULL);
1689 	EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1690 	EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1691 
1692 	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1693 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1694 	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1695 
1696 	/* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1697 	mutex_enter(&zilog->zl_lock);
1698 	if (lwb->lwb_root_zio == NULL) {
1699 		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1700 		    BP_GET_LSIZE(&lwb->lwb_blk));
1701 
1702 		if (!lwb->lwb_fastwrite) {
1703 			metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
1704 			lwb->lwb_fastwrite = 1;
1705 		}
1706 
1707 		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1708 			prio = ZIO_PRIORITY_SYNC_WRITE;
1709 		else
1710 			prio = ZIO_PRIORITY_ASYNC_WRITE;
1711 
1712 		lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1713 		    zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1714 		ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1715 
1716 		lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1717 		    zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1718 		    BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1719 		    prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb);
1720 		ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1721 
1722 		lwb->lwb_state = LWB_STATE_OPENED;
1723 
1724 		zil_lwb_set_zio_dependency(zilog, lwb);
1725 		zilog->zl_last_lwb_opened = lwb;
1726 	}
1727 	mutex_exit(&zilog->zl_lock);
1728 
1729 	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1730 	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1731 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1732 }
1733 
1734 /*
1735  * Define a limited set of intent log block sizes.
1736  *
1737  * These must be a multiple of 4KB. Note only the amount used (again
1738  * aligned to 4KB) actually gets written. However, we can't always just
1739  * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1740  */
1741 static const struct {
1742 	uint64_t	limit;
1743 	uint64_t	blksz;
1744 } zil_block_buckets[] = {
1745 	{ 4096,		4096 },			/* non TX_WRITE */
1746 	{ 8192 + 4096,	8192 + 4096 },		/* database */
1747 	{ 32768 + 4096,	32768 + 4096 },		/* NFS writes */
1748 	{ 65536 + 4096,	65536 + 4096 },		/* 64KB writes */
1749 	{ 131072,	131072 },		/* < 128KB writes */
1750 	{ 131072 +4096,	65536 + 4096 },		/* 128KB writes */
1751 	{ UINT64_MAX,	SPA_OLD_MAXBLOCKSIZE},	/* > 128KB writes */
1752 };
1753 
1754 /*
1755  * Maximum block size used by the ZIL.  This is picked up when the ZIL is
1756  * initialized.  Otherwise this should not be used directly; see
1757  * zl_max_block_size instead.
1758  */
1759 static uint_t zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1760 
1761 /*
1762  * Start a log block write and advance to the next log block.
1763  * Calls are serialized.
1764  */
1765 static lwb_t *
1766 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1767 {
1768 	lwb_t *nlwb = NULL;
1769 	zil_chain_t *zilc;
1770 	spa_t *spa = zilog->zl_spa;
1771 	blkptr_t *bp;
1772 	dmu_tx_t *tx;
1773 	uint64_t txg;
1774 	uint64_t zil_blksz, wsz;
1775 	int i, error;
1776 	boolean_t slog;
1777 
1778 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1779 	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1780 	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1781 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1782 
1783 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1784 		zilc = (zil_chain_t *)lwb->lwb_buf;
1785 		bp = &zilc->zc_next_blk;
1786 	} else {
1787 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1788 		bp = &zilc->zc_next_blk;
1789 	}
1790 
1791 	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1792 
1793 	/*
1794 	 * Allocate the next block and save its address in this block
1795 	 * before writing it in order to establish the log chain.
1796 	 */
1797 
1798 	tx = dmu_tx_create(zilog->zl_os);
1799 
1800 	/*
1801 	 * Since we are not going to create any new dirty data, and we
1802 	 * can even help with clearing the existing dirty data, we
1803 	 * should not be subject to the dirty data based delays. We
1804 	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1805 	 */
1806 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1807 
1808 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1809 	txg = dmu_tx_get_txg(tx);
1810 
1811 	mutex_enter(&zilog->zl_lwb_io_lock);
1812 	lwb->lwb_issued_txg = txg;
1813 	zilog->zl_lwb_inflight[txg & TXG_MASK]++;
1814 	zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg);
1815 	mutex_exit(&zilog->zl_lwb_io_lock);
1816 
1817 	/*
1818 	 * Log blocks are pre-allocated. Here we select the size of the next
1819 	 * block, based on size used in the last block.
1820 	 * - first find the smallest bucket that will fit the block from a
1821 	 *   limited set of block sizes. This is because it's faster to write
1822 	 *   blocks allocated from the same metaslab as they are adjacent or
1823 	 *   close.
1824 	 * - next find the maximum from the new suggested size and an array of
1825 	 *   previous sizes. This lessens a picket fence effect of wrongly
1826 	 *   guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1827 	 *   requests.
1828 	 *
1829 	 * Note we only write what is used, but we can't just allocate
1830 	 * the maximum block size because we can exhaust the available
1831 	 * pool log space.
1832 	 */
1833 	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1834 	for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++)
1835 		continue;
1836 	zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size);
1837 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1838 	for (i = 0; i < ZIL_PREV_BLKS; i++)
1839 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1840 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1841 
1842 	BP_ZERO(bp);
1843 	error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog);
1844 	if (slog) {
1845 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_slog_count);
1846 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_bytes,
1847 		    lwb->lwb_nused);
1848 	} else {
1849 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_normal_count);
1850 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_bytes,
1851 		    lwb->lwb_nused);
1852 	}
1853 	if (error == 0) {
1854 		ASSERT3U(bp->blk_birth, ==, txg);
1855 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1856 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1857 
1858 		/*
1859 		 * Allocate a new log write block (lwb).
1860 		 */
1861 		nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
1862 	}
1863 
1864 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1865 		/* For Slim ZIL only write what is used. */
1866 		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1867 		ASSERT3U(wsz, <=, lwb->lwb_sz);
1868 		zio_shrink(lwb->lwb_write_zio, wsz);
1869 
1870 	} else {
1871 		wsz = lwb->lwb_sz;
1872 	}
1873 
1874 	zilc->zc_pad = 0;
1875 	zilc->zc_nused = lwb->lwb_nused;
1876 	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1877 
1878 	/*
1879 	 * clear unused data for security
1880 	 */
1881 	memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused);
1882 
1883 	spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1884 
1885 	zil_lwb_add_block(lwb, &lwb->lwb_blk);
1886 	lwb->lwb_issued_timestamp = gethrtime();
1887 	lwb->lwb_state = LWB_STATE_ISSUED;
1888 
1889 	zio_nowait(lwb->lwb_root_zio);
1890 	zio_nowait(lwb->lwb_write_zio);
1891 
1892 	dmu_tx_commit(tx);
1893 
1894 	/*
1895 	 * If there was an allocation failure then nlwb will be null which
1896 	 * forces a txg_wait_synced().
1897 	 */
1898 	return (nlwb);
1899 }
1900 
1901 /*
1902  * Maximum amount of data that can be put into single log block.
1903  */
1904 uint64_t
1905 zil_max_log_data(zilog_t *zilog, size_t hdrsize)
1906 {
1907 	return (zilog->zl_max_block_size - sizeof (zil_chain_t) - hdrsize);
1908 }
1909 
1910 /*
1911  * Maximum amount of log space we agree to waste to reduce number of
1912  * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1913  */
1914 static inline uint64_t
1915 zil_max_waste_space(zilog_t *zilog)
1916 {
1917 	return (zil_max_log_data(zilog, sizeof (lr_write_t)) / 8);
1918 }
1919 
1920 /*
1921  * Maximum amount of write data for WR_COPIED.  For correctness, consumers
1922  * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1923  * maximum sized log block, because each WR_COPIED record must fit in a
1924  * single log block.  For space efficiency, we want to fit two records into a
1925  * max-sized log block.
1926  */
1927 uint64_t
1928 zil_max_copied_data(zilog_t *zilog)
1929 {
1930 	return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 -
1931 	    sizeof (lr_write_t));
1932 }
1933 
1934 static lwb_t *
1935 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1936 {
1937 	lr_t *lrcb, *lrc;
1938 	lr_write_t *lrwb, *lrw;
1939 	char *lr_buf;
1940 	uint64_t dlen, dnow, dpad, lwb_sp, reclen, txg, max_log_data;
1941 
1942 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1943 	ASSERT3P(lwb, !=, NULL);
1944 	ASSERT3P(lwb->lwb_buf, !=, NULL);
1945 
1946 	zil_lwb_write_open(zilog, lwb);
1947 
1948 	lrc = &itx->itx_lr;
1949 	lrw = (lr_write_t *)lrc;
1950 
1951 	/*
1952 	 * A commit itx doesn't represent any on-disk state; instead
1953 	 * it's simply used as a place holder on the commit list, and
1954 	 * provides a mechanism for attaching a "commit waiter" onto the
1955 	 * correct lwb (such that the waiter can be signalled upon
1956 	 * completion of that lwb). Thus, we don't process this itx's
1957 	 * log record if it's a commit itx (these itx's don't have log
1958 	 * records), and instead link the itx's waiter onto the lwb's
1959 	 * list of waiters.
1960 	 *
1961 	 * For more details, see the comment above zil_commit().
1962 	 */
1963 	if (lrc->lrc_txtype == TX_COMMIT) {
1964 		mutex_enter(&zilog->zl_lock);
1965 		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1966 		itx->itx_private = NULL;
1967 		mutex_exit(&zilog->zl_lock);
1968 		return (lwb);
1969 	}
1970 
1971 	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1972 		dlen = P2ROUNDUP_TYPED(
1973 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1974 		dpad = dlen - lrw->lr_length;
1975 	} else {
1976 		dlen = dpad = 0;
1977 	}
1978 	reclen = lrc->lrc_reclen;
1979 	zilog->zl_cur_used += (reclen + dlen);
1980 	txg = lrc->lrc_txg;
1981 
1982 	ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1983 
1984 cont:
1985 	/*
1986 	 * If this record won't fit in the current log block, start a new one.
1987 	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1988 	 */
1989 	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1990 	max_log_data = zil_max_log_data(zilog, sizeof (lr_write_t));
1991 	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1992 	    lwb_sp < zil_max_waste_space(zilog) &&
1993 	    (dlen % max_log_data == 0 ||
1994 	    lwb_sp < reclen + dlen % max_log_data))) {
1995 		lwb = zil_lwb_write_issue(zilog, lwb);
1996 		if (lwb == NULL)
1997 			return (NULL);
1998 		zil_lwb_write_open(zilog, lwb);
1999 		ASSERT(LWB_EMPTY(lwb));
2000 		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
2001 
2002 		/*
2003 		 * There must be enough space in the new, empty log block to
2004 		 * hold reclen.  For WR_COPIED, we need to fit the whole
2005 		 * record in one block, and reclen is the header size + the
2006 		 * data size. For WR_NEED_COPY, we can create multiple
2007 		 * records, splitting the data into multiple blocks, so we
2008 		 * only need to fit one word of data per block; in this case
2009 		 * reclen is just the header size (no data).
2010 		 */
2011 		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
2012 	}
2013 
2014 	dnow = MIN(dlen, lwb_sp - reclen);
2015 	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
2016 	memcpy(lr_buf, lrc, reclen);
2017 	lrcb = (lr_t *)lr_buf;		/* Like lrc, but inside lwb. */
2018 	lrwb = (lr_write_t *)lrcb;	/* Like lrw, but inside lwb. */
2019 
2020 	ZIL_STAT_BUMP(zilog, zil_itx_count);
2021 
2022 	/*
2023 	 * If it's a write, fetch the data or get its blkptr as appropriate.
2024 	 */
2025 	if (lrc->lrc_txtype == TX_WRITE) {
2026 		if (txg > spa_freeze_txg(zilog->zl_spa))
2027 			txg_wait_synced(zilog->zl_dmu_pool, txg);
2028 		if (itx->itx_wr_state == WR_COPIED) {
2029 			ZIL_STAT_BUMP(zilog, zil_itx_copied_count);
2030 			ZIL_STAT_INCR(zilog, zil_itx_copied_bytes,
2031 			    lrw->lr_length);
2032 		} else {
2033 			char *dbuf;
2034 			int error;
2035 
2036 			if (itx->itx_wr_state == WR_NEED_COPY) {
2037 				dbuf = lr_buf + reclen;
2038 				lrcb->lrc_reclen += dnow;
2039 				if (lrwb->lr_length > dnow)
2040 					lrwb->lr_length = dnow;
2041 				lrw->lr_offset += dnow;
2042 				lrw->lr_length -= dnow;
2043 				ZIL_STAT_BUMP(zilog, zil_itx_needcopy_count);
2044 				ZIL_STAT_INCR(zilog, zil_itx_needcopy_bytes,
2045 				    dnow);
2046 			} else {
2047 				ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
2048 				dbuf = NULL;
2049 				ZIL_STAT_BUMP(zilog, zil_itx_indirect_count);
2050 				ZIL_STAT_INCR(zilog, zil_itx_indirect_bytes,
2051 				    lrw->lr_length);
2052 			}
2053 
2054 			/*
2055 			 * We pass in the "lwb_write_zio" rather than
2056 			 * "lwb_root_zio" so that the "lwb_write_zio"
2057 			 * becomes the parent of any zio's created by
2058 			 * the "zl_get_data" callback. The vdevs are
2059 			 * flushed after the "lwb_write_zio" completes,
2060 			 * so we want to make sure that completion
2061 			 * callback waits for these additional zio's,
2062 			 * such that the vdevs used by those zio's will
2063 			 * be included in the lwb's vdev tree, and those
2064 			 * vdevs will be properly flushed. If we passed
2065 			 * in "lwb_root_zio" here, then these additional
2066 			 * vdevs may not be flushed; e.g. if these zio's
2067 			 * completed after "lwb_write_zio" completed.
2068 			 */
2069 			error = zilog->zl_get_data(itx->itx_private,
2070 			    itx->itx_gen, lrwb, dbuf, lwb,
2071 			    lwb->lwb_write_zio);
2072 			if (dbuf != NULL && error == 0 && dnow == dlen)
2073 				/* Zero any padding bytes in the last block. */
2074 				memset((char *)dbuf + lrwb->lr_length, 0, dpad);
2075 
2076 			/*
2077 			 * Typically, the only return values we should see from
2078 			 * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or
2079 			 *  EALREADY. However, it is also possible to see other
2080 			 *  error values such as ENOSPC or EINVAL from
2081 			 *  dmu_read() -> dnode_hold() -> dnode_hold_impl() or
2082 			 *  ENXIO as well as a multitude of others from the
2083 			 *  block layer through dmu_buf_hold() -> dbuf_read()
2084 			 *  -> zio_wait(), as well as through dmu_read() ->
2085 			 *  dnode_hold() -> dnode_hold_impl() -> dbuf_read() ->
2086 			 *  zio_wait(). When these errors happen, we can assume
2087 			 *  that neither an immediate write nor an indirect
2088 			 *  write occurred, so we need to fall back to
2089 			 *  txg_wait_synced(). This is unusual, so we print to
2090 			 *  dmesg whenever one of these errors occurs.
2091 			 */
2092 			switch (error) {
2093 			case 0:
2094 				break;
2095 			default:
2096 				cmn_err(CE_WARN, "zil_lwb_commit() received "
2097 				    "unexpected error %d from ->zl_get_data()"
2098 				    ". Falling back to txg_wait_synced().",
2099 				    error);
2100 				zfs_fallthrough;
2101 			case EIO:
2102 				txg_wait_synced(zilog->zl_dmu_pool, txg);
2103 				zfs_fallthrough;
2104 			case ENOENT:
2105 				zfs_fallthrough;
2106 			case EEXIST:
2107 				zfs_fallthrough;
2108 			case EALREADY:
2109 				return (lwb);
2110 			}
2111 		}
2112 	}
2113 
2114 	/*
2115 	 * We're actually making an entry, so update lrc_seq to be the
2116 	 * log record sequence number.  Note that this is generally not
2117 	 * equal to the itx sequence number because not all transactions
2118 	 * are synchronous, and sometimes spa_sync() gets there first.
2119 	 */
2120 	lrcb->lrc_seq = ++zilog->zl_lr_seq;
2121 	lwb->lwb_nused += reclen + dnow;
2122 
2123 	zil_lwb_add_txg(lwb, txg);
2124 
2125 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
2126 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
2127 
2128 	dlen -= dnow;
2129 	if (dlen > 0) {
2130 		zilog->zl_cur_used += reclen;
2131 		goto cont;
2132 	}
2133 
2134 	return (lwb);
2135 }
2136 
2137 itx_t *
2138 zil_itx_create(uint64_t txtype, size_t olrsize)
2139 {
2140 	size_t itxsize, lrsize;
2141 	itx_t *itx;
2142 
2143 	lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t);
2144 	itxsize = offsetof(itx_t, itx_lr) + lrsize;
2145 
2146 	itx = zio_data_buf_alloc(itxsize);
2147 	itx->itx_lr.lrc_txtype = txtype;
2148 	itx->itx_lr.lrc_reclen = lrsize;
2149 	itx->itx_lr.lrc_seq = 0;	/* defensive */
2150 	memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize);
2151 	itx->itx_sync = B_TRUE;		/* default is synchronous */
2152 	itx->itx_callback = NULL;
2153 	itx->itx_callback_data = NULL;
2154 	itx->itx_size = itxsize;
2155 
2156 	return (itx);
2157 }
2158 
2159 void
2160 zil_itx_destroy(itx_t *itx)
2161 {
2162 	IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
2163 	IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2164 
2165 	if (itx->itx_callback != NULL)
2166 		itx->itx_callback(itx->itx_callback_data);
2167 
2168 	zio_data_buf_free(itx, itx->itx_size);
2169 }
2170 
2171 /*
2172  * Free up the sync and async itxs. The itxs_t has already been detached
2173  * so no locks are needed.
2174  */
2175 static void
2176 zil_itxg_clean(void *arg)
2177 {
2178 	itx_t *itx;
2179 	list_t *list;
2180 	avl_tree_t *t;
2181 	void *cookie;
2182 	itxs_t *itxs = arg;
2183 	itx_async_node_t *ian;
2184 
2185 	list = &itxs->i_sync_list;
2186 	while ((itx = list_head(list)) != NULL) {
2187 		/*
2188 		 * In the general case, commit itxs will not be found
2189 		 * here, as they'll be committed to an lwb via
2190 		 * zil_lwb_commit(), and free'd in that function. Having
2191 		 * said that, it is still possible for commit itxs to be
2192 		 * found here, due to the following race:
2193 		 *
2194 		 *	- a thread calls zil_commit() which assigns the
2195 		 *	  commit itx to a per-txg i_sync_list
2196 		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
2197 		 *	  while the waiter is still on the i_sync_list
2198 		 *
2199 		 * There's nothing to prevent syncing the txg while the
2200 		 * waiter is on the i_sync_list. This normally doesn't
2201 		 * happen because spa_sync() is slower than zil_commit(),
2202 		 * but if zil_commit() calls txg_wait_synced() (e.g.
2203 		 * because zil_create() or zil_commit_writer_stall() is
2204 		 * called) we will hit this case.
2205 		 */
2206 		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
2207 			zil_commit_waiter_skip(itx->itx_private);
2208 
2209 		list_remove(list, itx);
2210 		zil_itx_destroy(itx);
2211 	}
2212 
2213 	cookie = NULL;
2214 	t = &itxs->i_async_tree;
2215 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2216 		list = &ian->ia_list;
2217 		while ((itx = list_head(list)) != NULL) {
2218 			list_remove(list, itx);
2219 			/* commit itxs should never be on the async lists. */
2220 			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2221 			zil_itx_destroy(itx);
2222 		}
2223 		list_destroy(list);
2224 		kmem_free(ian, sizeof (itx_async_node_t));
2225 	}
2226 	avl_destroy(t);
2227 
2228 	kmem_free(itxs, sizeof (itxs_t));
2229 }
2230 
2231 static int
2232 zil_aitx_compare(const void *x1, const void *x2)
2233 {
2234 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
2235 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
2236 
2237 	return (TREE_CMP(o1, o2));
2238 }
2239 
2240 /*
2241  * Remove all async itx with the given oid.
2242  */
2243 void
2244 zil_remove_async(zilog_t *zilog, uint64_t oid)
2245 {
2246 	uint64_t otxg, txg;
2247 	itx_async_node_t *ian;
2248 	avl_tree_t *t;
2249 	avl_index_t where;
2250 	list_t clean_list;
2251 	itx_t *itx;
2252 
2253 	ASSERT(oid != 0);
2254 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
2255 
2256 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2257 		otxg = ZILTEST_TXG;
2258 	else
2259 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2260 
2261 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2262 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2263 
2264 		mutex_enter(&itxg->itxg_lock);
2265 		if (itxg->itxg_txg != txg) {
2266 			mutex_exit(&itxg->itxg_lock);
2267 			continue;
2268 		}
2269 
2270 		/*
2271 		 * Locate the object node and append its list.
2272 		 */
2273 		t = &itxg->itxg_itxs->i_async_tree;
2274 		ian = avl_find(t, &oid, &where);
2275 		if (ian != NULL)
2276 			list_move_tail(&clean_list, &ian->ia_list);
2277 		mutex_exit(&itxg->itxg_lock);
2278 	}
2279 	while ((itx = list_head(&clean_list)) != NULL) {
2280 		list_remove(&clean_list, itx);
2281 		/* commit itxs should never be on the async lists. */
2282 		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2283 		zil_itx_destroy(itx);
2284 	}
2285 	list_destroy(&clean_list);
2286 }
2287 
2288 void
2289 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
2290 {
2291 	uint64_t txg;
2292 	itxg_t *itxg;
2293 	itxs_t *itxs, *clean = NULL;
2294 
2295 	/*
2296 	 * Ensure the data of a renamed file is committed before the rename.
2297 	 */
2298 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
2299 		zil_async_to_sync(zilog, itx->itx_oid);
2300 
2301 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
2302 		txg = ZILTEST_TXG;
2303 	else
2304 		txg = dmu_tx_get_txg(tx);
2305 
2306 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
2307 	mutex_enter(&itxg->itxg_lock);
2308 	itxs = itxg->itxg_itxs;
2309 	if (itxg->itxg_txg != txg) {
2310 		if (itxs != NULL) {
2311 			/*
2312 			 * The zil_clean callback hasn't got around to cleaning
2313 			 * this itxg. Save the itxs for release below.
2314 			 * This should be rare.
2315 			 */
2316 			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2317 			    "txg %llu", (u_longlong_t)itxg->itxg_txg);
2318 			clean = itxg->itxg_itxs;
2319 		}
2320 		itxg->itxg_txg = txg;
2321 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
2322 		    KM_SLEEP);
2323 
2324 		list_create(&itxs->i_sync_list, sizeof (itx_t),
2325 		    offsetof(itx_t, itx_node));
2326 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
2327 		    sizeof (itx_async_node_t),
2328 		    offsetof(itx_async_node_t, ia_node));
2329 	}
2330 	if (itx->itx_sync) {
2331 		list_insert_tail(&itxs->i_sync_list, itx);
2332 	} else {
2333 		avl_tree_t *t = &itxs->i_async_tree;
2334 		uint64_t foid =
2335 		    LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
2336 		itx_async_node_t *ian;
2337 		avl_index_t where;
2338 
2339 		ian = avl_find(t, &foid, &where);
2340 		if (ian == NULL) {
2341 			ian = kmem_alloc(sizeof (itx_async_node_t),
2342 			    KM_SLEEP);
2343 			list_create(&ian->ia_list, sizeof (itx_t),
2344 			    offsetof(itx_t, itx_node));
2345 			ian->ia_foid = foid;
2346 			avl_insert(t, ian, where);
2347 		}
2348 		list_insert_tail(&ian->ia_list, itx);
2349 	}
2350 
2351 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
2352 
2353 	/*
2354 	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2355 	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2356 	 * need to be careful to always dirty the ZIL using the "real"
2357 	 * TXG (not itxg_txg) even when the SPA is frozen.
2358 	 */
2359 	zilog_dirty(zilog, dmu_tx_get_txg(tx));
2360 	mutex_exit(&itxg->itxg_lock);
2361 
2362 	/* Release the old itxs now we've dropped the lock */
2363 	if (clean != NULL)
2364 		zil_itxg_clean(clean);
2365 }
2366 
2367 /*
2368  * If there are any in-memory intent log transactions which have now been
2369  * synced then start up a taskq to free them. We should only do this after we
2370  * have written out the uberblocks (i.e. txg has been committed) so that
2371  * don't inadvertently clean out in-memory log records that would be required
2372  * by zil_commit().
2373  */
2374 void
2375 zil_clean(zilog_t *zilog, uint64_t synced_txg)
2376 {
2377 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2378 	itxs_t *clean_me;
2379 
2380 	ASSERT3U(synced_txg, <, ZILTEST_TXG);
2381 
2382 	mutex_enter(&itxg->itxg_lock);
2383 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2384 		mutex_exit(&itxg->itxg_lock);
2385 		return;
2386 	}
2387 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2388 	ASSERT3U(itxg->itxg_txg, !=, 0);
2389 	clean_me = itxg->itxg_itxs;
2390 	itxg->itxg_itxs = NULL;
2391 	itxg->itxg_txg = 0;
2392 	mutex_exit(&itxg->itxg_lock);
2393 	/*
2394 	 * Preferably start a task queue to free up the old itxs but
2395 	 * if taskq_dispatch can't allocate resources to do that then
2396 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
2397 	 * created a bad performance problem.
2398 	 */
2399 	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2400 	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2401 	taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
2402 	    zil_itxg_clean, clean_me, TQ_NOSLEEP);
2403 	if (id == TASKQID_INVALID)
2404 		zil_itxg_clean(clean_me);
2405 }
2406 
2407 /*
2408  * This function will traverse the queue of itxs that need to be
2409  * committed, and move them onto the ZIL's zl_itx_commit_list.
2410  */
2411 static void
2412 zil_get_commit_list(zilog_t *zilog)
2413 {
2414 	uint64_t otxg, txg;
2415 	list_t *commit_list = &zilog->zl_itx_commit_list;
2416 
2417 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2418 
2419 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2420 		otxg = ZILTEST_TXG;
2421 	else
2422 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2423 
2424 	/*
2425 	 * This is inherently racy, since there is nothing to prevent
2426 	 * the last synced txg from changing. That's okay since we'll
2427 	 * only commit things in the future.
2428 	 */
2429 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2430 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2431 
2432 		mutex_enter(&itxg->itxg_lock);
2433 		if (itxg->itxg_txg != txg) {
2434 			mutex_exit(&itxg->itxg_lock);
2435 			continue;
2436 		}
2437 
2438 		/*
2439 		 * If we're adding itx records to the zl_itx_commit_list,
2440 		 * then the zil better be dirty in this "txg". We can assert
2441 		 * that here since we're holding the itxg_lock which will
2442 		 * prevent spa_sync from cleaning it. Once we add the itxs
2443 		 * to the zl_itx_commit_list we must commit it to disk even
2444 		 * if it's unnecessary (i.e. the txg was synced).
2445 		 */
2446 		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2447 		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2448 		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
2449 
2450 		mutex_exit(&itxg->itxg_lock);
2451 	}
2452 }
2453 
2454 /*
2455  * Move the async itxs for a specified object to commit into sync lists.
2456  */
2457 void
2458 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2459 {
2460 	uint64_t otxg, txg;
2461 	itx_async_node_t *ian;
2462 	avl_tree_t *t;
2463 	avl_index_t where;
2464 
2465 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2466 		otxg = ZILTEST_TXG;
2467 	else
2468 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2469 
2470 	/*
2471 	 * This is inherently racy, since there is nothing to prevent
2472 	 * the last synced txg from changing.
2473 	 */
2474 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2475 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2476 
2477 		mutex_enter(&itxg->itxg_lock);
2478 		if (itxg->itxg_txg != txg) {
2479 			mutex_exit(&itxg->itxg_lock);
2480 			continue;
2481 		}
2482 
2483 		/*
2484 		 * If a foid is specified then find that node and append its
2485 		 * list. Otherwise walk the tree appending all the lists
2486 		 * to the sync list. We add to the end rather than the
2487 		 * beginning to ensure the create has happened.
2488 		 */
2489 		t = &itxg->itxg_itxs->i_async_tree;
2490 		if (foid != 0) {
2491 			ian = avl_find(t, &foid, &where);
2492 			if (ian != NULL) {
2493 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2494 				    &ian->ia_list);
2495 			}
2496 		} else {
2497 			void *cookie = NULL;
2498 
2499 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2500 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2501 				    &ian->ia_list);
2502 				list_destroy(&ian->ia_list);
2503 				kmem_free(ian, sizeof (itx_async_node_t));
2504 			}
2505 		}
2506 		mutex_exit(&itxg->itxg_lock);
2507 	}
2508 }
2509 
2510 /*
2511  * This function will prune commit itxs that are at the head of the
2512  * commit list (it won't prune past the first non-commit itx), and
2513  * either: a) attach them to the last lwb that's still pending
2514  * completion, or b) skip them altogether.
2515  *
2516  * This is used as a performance optimization to prevent commit itxs
2517  * from generating new lwbs when it's unnecessary to do so.
2518  */
2519 static void
2520 zil_prune_commit_list(zilog_t *zilog)
2521 {
2522 	itx_t *itx;
2523 
2524 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2525 
2526 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2527 		lr_t *lrc = &itx->itx_lr;
2528 		if (lrc->lrc_txtype != TX_COMMIT)
2529 			break;
2530 
2531 		mutex_enter(&zilog->zl_lock);
2532 
2533 		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2534 		if (last_lwb == NULL ||
2535 		    last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2536 			/*
2537 			 * All of the itxs this waiter was waiting on
2538 			 * must have already completed (or there were
2539 			 * never any itx's for it to wait on), so it's
2540 			 * safe to skip this waiter and mark it done.
2541 			 */
2542 			zil_commit_waiter_skip(itx->itx_private);
2543 		} else {
2544 			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2545 			itx->itx_private = NULL;
2546 		}
2547 
2548 		mutex_exit(&zilog->zl_lock);
2549 
2550 		list_remove(&zilog->zl_itx_commit_list, itx);
2551 		zil_itx_destroy(itx);
2552 	}
2553 
2554 	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2555 }
2556 
2557 static void
2558 zil_commit_writer_stall(zilog_t *zilog)
2559 {
2560 	/*
2561 	 * When zio_alloc_zil() fails to allocate the next lwb block on
2562 	 * disk, we must call txg_wait_synced() to ensure all of the
2563 	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2564 	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2565 	 * to zil_process_commit_list()) will have to call zil_create(),
2566 	 * and start a new ZIL chain.
2567 	 *
2568 	 * Since zil_alloc_zil() failed, the lwb that was previously
2569 	 * issued does not have a pointer to the "next" lwb on disk.
2570 	 * Thus, if another ZIL writer thread was to allocate the "next"
2571 	 * on-disk lwb, that block could be leaked in the event of a
2572 	 * crash (because the previous lwb on-disk would not point to
2573 	 * it).
2574 	 *
2575 	 * We must hold the zilog's zl_issuer_lock while we do this, to
2576 	 * ensure no new threads enter zil_process_commit_list() until
2577 	 * all lwb's in the zl_lwb_list have been synced and freed
2578 	 * (which is achieved via the txg_wait_synced() call).
2579 	 */
2580 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2581 	txg_wait_synced(zilog->zl_dmu_pool, 0);
2582 	ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2583 }
2584 
2585 /*
2586  * This function will traverse the commit list, creating new lwbs as
2587  * needed, and committing the itxs from the commit list to these newly
2588  * created lwbs. Additionally, as a new lwb is created, the previous
2589  * lwb will be issued to the zio layer to be written to disk.
2590  */
2591 static void
2592 zil_process_commit_list(zilog_t *zilog)
2593 {
2594 	spa_t *spa = zilog->zl_spa;
2595 	list_t nolwb_itxs;
2596 	list_t nolwb_waiters;
2597 	lwb_t *lwb, *plwb;
2598 	itx_t *itx;
2599 	boolean_t first = B_TRUE;
2600 
2601 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2602 
2603 	/*
2604 	 * Return if there's nothing to commit before we dirty the fs by
2605 	 * calling zil_create().
2606 	 */
2607 	if (list_head(&zilog->zl_itx_commit_list) == NULL)
2608 		return;
2609 
2610 	list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
2611 	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2612 	    offsetof(zil_commit_waiter_t, zcw_node));
2613 
2614 	lwb = list_tail(&zilog->zl_lwb_list);
2615 	if (lwb == NULL) {
2616 		lwb = zil_create(zilog);
2617 	} else {
2618 		/*
2619 		 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2620 		 * have already been created (zl_lwb_list not empty).
2621 		 */
2622 		zil_commit_activate_saxattr_feature(zilog);
2623 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2624 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2625 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2626 		first = (lwb->lwb_state != LWB_STATE_OPENED) &&
2627 		    ((plwb = list_prev(&zilog->zl_lwb_list, lwb)) == NULL ||
2628 		    plwb->lwb_state == LWB_STATE_FLUSH_DONE);
2629 	}
2630 
2631 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2632 		lr_t *lrc = &itx->itx_lr;
2633 		uint64_t txg = lrc->lrc_txg;
2634 
2635 		ASSERT3U(txg, !=, 0);
2636 
2637 		if (lrc->lrc_txtype == TX_COMMIT) {
2638 			DTRACE_PROBE2(zil__process__commit__itx,
2639 			    zilog_t *, zilog, itx_t *, itx);
2640 		} else {
2641 			DTRACE_PROBE2(zil__process__normal__itx,
2642 			    zilog_t *, zilog, itx_t *, itx);
2643 		}
2644 
2645 		list_remove(&zilog->zl_itx_commit_list, itx);
2646 
2647 		boolean_t synced = txg <= spa_last_synced_txg(spa);
2648 		boolean_t frozen = txg > spa_freeze_txg(spa);
2649 
2650 		/*
2651 		 * If the txg of this itx has already been synced out, then
2652 		 * we don't need to commit this itx to an lwb. This is
2653 		 * because the data of this itx will have already been
2654 		 * written to the main pool. This is inherently racy, and
2655 		 * it's still ok to commit an itx whose txg has already
2656 		 * been synced; this will result in a write that's
2657 		 * unnecessary, but will do no harm.
2658 		 *
2659 		 * With that said, we always want to commit TX_COMMIT itxs
2660 		 * to an lwb, regardless of whether or not that itx's txg
2661 		 * has been synced out. We do this to ensure any OPENED lwb
2662 		 * will always have at least one zil_commit_waiter_t linked
2663 		 * to the lwb.
2664 		 *
2665 		 * As a counter-example, if we skipped TX_COMMIT itx's
2666 		 * whose txg had already been synced, the following
2667 		 * situation could occur if we happened to be racing with
2668 		 * spa_sync:
2669 		 *
2670 		 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2671 		 *    itx's txg is 10 and the last synced txg is 9.
2672 		 * 2. spa_sync finishes syncing out txg 10.
2673 		 * 3. We move to the next itx in the list, it's a TX_COMMIT
2674 		 *    whose txg is 10, so we skip it rather than committing
2675 		 *    it to the lwb used in (1).
2676 		 *
2677 		 * If the itx that is skipped in (3) is the last TX_COMMIT
2678 		 * itx in the commit list, than it's possible for the lwb
2679 		 * used in (1) to remain in the OPENED state indefinitely.
2680 		 *
2681 		 * To prevent the above scenario from occurring, ensuring
2682 		 * that once an lwb is OPENED it will transition to ISSUED
2683 		 * and eventually DONE, we always commit TX_COMMIT itx's to
2684 		 * an lwb here, even if that itx's txg has already been
2685 		 * synced.
2686 		 *
2687 		 * Finally, if the pool is frozen, we _always_ commit the
2688 		 * itx.  The point of freezing the pool is to prevent data
2689 		 * from being written to the main pool via spa_sync, and
2690 		 * instead rely solely on the ZIL to persistently store the
2691 		 * data; i.e.  when the pool is frozen, the last synced txg
2692 		 * value can't be trusted.
2693 		 */
2694 		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2695 			if (lwb != NULL) {
2696 				lwb = zil_lwb_commit(zilog, itx, lwb);
2697 
2698 				if (lwb == NULL)
2699 					list_insert_tail(&nolwb_itxs, itx);
2700 				else
2701 					list_insert_tail(&lwb->lwb_itxs, itx);
2702 			} else {
2703 				if (lrc->lrc_txtype == TX_COMMIT) {
2704 					zil_commit_waiter_link_nolwb(
2705 					    itx->itx_private, &nolwb_waiters);
2706 				}
2707 
2708 				list_insert_tail(&nolwb_itxs, itx);
2709 			}
2710 		} else {
2711 			ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
2712 			zil_itx_destroy(itx);
2713 		}
2714 	}
2715 
2716 	if (lwb == NULL) {
2717 		/*
2718 		 * This indicates zio_alloc_zil() failed to allocate the
2719 		 * "next" lwb on-disk. When this happens, we must stall
2720 		 * the ZIL write pipeline; see the comment within
2721 		 * zil_commit_writer_stall() for more details.
2722 		 */
2723 		zil_commit_writer_stall(zilog);
2724 
2725 		/*
2726 		 * Additionally, we have to signal and mark the "nolwb"
2727 		 * waiters as "done" here, since without an lwb, we
2728 		 * can't do this via zil_lwb_flush_vdevs_done() like
2729 		 * normal.
2730 		 */
2731 		zil_commit_waiter_t *zcw;
2732 		while ((zcw = list_head(&nolwb_waiters)) != NULL) {
2733 			zil_commit_waiter_skip(zcw);
2734 			list_remove(&nolwb_waiters, zcw);
2735 		}
2736 
2737 		/*
2738 		 * And finally, we have to destroy the itx's that
2739 		 * couldn't be committed to an lwb; this will also call
2740 		 * the itx's callback if one exists for the itx.
2741 		 */
2742 		while ((itx = list_head(&nolwb_itxs)) != NULL) {
2743 			list_remove(&nolwb_itxs, itx);
2744 			zil_itx_destroy(itx);
2745 		}
2746 	} else {
2747 		ASSERT(list_is_empty(&nolwb_waiters));
2748 		ASSERT3P(lwb, !=, NULL);
2749 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2750 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2751 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2752 
2753 		/*
2754 		 * At this point, the ZIL block pointed at by the "lwb"
2755 		 * variable is in one of the following states: "closed"
2756 		 * or "open".
2757 		 *
2758 		 * If it's "closed", then no itxs have been committed to
2759 		 * it, so there's no point in issuing its zio (i.e. it's
2760 		 * "empty").
2761 		 *
2762 		 * If it's "open", then it contains one or more itxs that
2763 		 * eventually need to be committed to stable storage. In
2764 		 * this case we intentionally do not issue the lwb's zio
2765 		 * to disk yet, and instead rely on one of the following
2766 		 * two mechanisms for issuing the zio:
2767 		 *
2768 		 * 1. Ideally, there will be more ZIL activity occurring
2769 		 * on the system, such that this function will be
2770 		 * immediately called again (not necessarily by the same
2771 		 * thread) and this lwb's zio will be issued via
2772 		 * zil_lwb_commit(). This way, the lwb is guaranteed to
2773 		 * be "full" when it is issued to disk, and we'll make
2774 		 * use of the lwb's size the best we can.
2775 		 *
2776 		 * 2. If there isn't sufficient ZIL activity occurring on
2777 		 * the system, such that this lwb's zio isn't issued via
2778 		 * zil_lwb_commit(), zil_commit_waiter() will issue the
2779 		 * lwb's zio. If this occurs, the lwb is not guaranteed
2780 		 * to be "full" by the time its zio is issued, and means
2781 		 * the size of the lwb was "too large" given the amount
2782 		 * of ZIL activity occurring on the system at that time.
2783 		 *
2784 		 * We do this for a couple of reasons:
2785 		 *
2786 		 * 1. To try and reduce the number of IOPs needed to
2787 		 * write the same number of itxs. If an lwb has space
2788 		 * available in its buffer for more itxs, and more itxs
2789 		 * will be committed relatively soon (relative to the
2790 		 * latency of performing a write), then it's beneficial
2791 		 * to wait for these "next" itxs. This way, more itxs
2792 		 * can be committed to stable storage with fewer writes.
2793 		 *
2794 		 * 2. To try and use the largest lwb block size that the
2795 		 * incoming rate of itxs can support. Again, this is to
2796 		 * try and pack as many itxs into as few lwbs as
2797 		 * possible, without significantly impacting the latency
2798 		 * of each individual itx.
2799 		 *
2800 		 * If we had no already running or open LWBs, it can be
2801 		 * the workload is single-threaded.  And if the ZIL write
2802 		 * latency is very small or if the LWB is almost full, it
2803 		 * may be cheaper to bypass the delay.
2804 		 */
2805 		if (lwb->lwb_state == LWB_STATE_OPENED && first) {
2806 			hrtime_t sleep = zilog->zl_last_lwb_latency *
2807 			    zfs_commit_timeout_pct / 100;
2808 			if (sleep < zil_min_commit_timeout ||
2809 			    lwb->lwb_sz - lwb->lwb_nused < lwb->lwb_sz / 8) {
2810 				lwb = zil_lwb_write_issue(zilog, lwb);
2811 				zilog->zl_cur_used = 0;
2812 				if (lwb == NULL)
2813 					zil_commit_writer_stall(zilog);
2814 			}
2815 		}
2816 	}
2817 }
2818 
2819 /*
2820  * This function is responsible for ensuring the passed in commit waiter
2821  * (and associated commit itx) is committed to an lwb. If the waiter is
2822  * not already committed to an lwb, all itxs in the zilog's queue of
2823  * itxs will be processed. The assumption is the passed in waiter's
2824  * commit itx will found in the queue just like the other non-commit
2825  * itxs, such that when the entire queue is processed, the waiter will
2826  * have been committed to an lwb.
2827  *
2828  * The lwb associated with the passed in waiter is not guaranteed to
2829  * have been issued by the time this function completes. If the lwb is
2830  * not issued, we rely on future calls to zil_commit_writer() to issue
2831  * the lwb, or the timeout mechanism found in zil_commit_waiter().
2832  */
2833 static void
2834 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2835 {
2836 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2837 	ASSERT(spa_writeable(zilog->zl_spa));
2838 
2839 	mutex_enter(&zilog->zl_issuer_lock);
2840 
2841 	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2842 		/*
2843 		 * It's possible that, while we were waiting to acquire
2844 		 * the "zl_issuer_lock", another thread committed this
2845 		 * waiter to an lwb. If that occurs, we bail out early,
2846 		 * without processing any of the zilog's queue of itxs.
2847 		 *
2848 		 * On certain workloads and system configurations, the
2849 		 * "zl_issuer_lock" can become highly contended. In an
2850 		 * attempt to reduce this contention, we immediately drop
2851 		 * the lock if the waiter has already been processed.
2852 		 *
2853 		 * We've measured this optimization to reduce CPU spent
2854 		 * contending on this lock by up to 5%, using a system
2855 		 * with 32 CPUs, low latency storage (~50 usec writes),
2856 		 * and 1024 threads performing sync writes.
2857 		 */
2858 		goto out;
2859 	}
2860 
2861 	ZIL_STAT_BUMP(zilog, zil_commit_writer_count);
2862 
2863 	zil_get_commit_list(zilog);
2864 	zil_prune_commit_list(zilog);
2865 	zil_process_commit_list(zilog);
2866 
2867 out:
2868 	mutex_exit(&zilog->zl_issuer_lock);
2869 }
2870 
2871 static void
2872 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2873 {
2874 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2875 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2876 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2877 
2878 	lwb_t *lwb = zcw->zcw_lwb;
2879 	ASSERT3P(lwb, !=, NULL);
2880 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2881 
2882 	/*
2883 	 * If the lwb has already been issued by another thread, we can
2884 	 * immediately return since there's no work to be done (the
2885 	 * point of this function is to issue the lwb). Additionally, we
2886 	 * do this prior to acquiring the zl_issuer_lock, to avoid
2887 	 * acquiring it when it's not necessary to do so.
2888 	 */
2889 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2890 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2891 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2892 		return;
2893 
2894 	/*
2895 	 * In order to call zil_lwb_write_issue() we must hold the
2896 	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2897 	 * since we're already holding the commit waiter's "zcw_lock",
2898 	 * and those two locks are acquired in the opposite order
2899 	 * elsewhere.
2900 	 */
2901 	mutex_exit(&zcw->zcw_lock);
2902 	mutex_enter(&zilog->zl_issuer_lock);
2903 	mutex_enter(&zcw->zcw_lock);
2904 
2905 	/*
2906 	 * Since we just dropped and re-acquired the commit waiter's
2907 	 * lock, we have to re-check to see if the waiter was marked
2908 	 * "done" during that process. If the waiter was marked "done",
2909 	 * the "lwb" pointer is no longer valid (it can be free'd after
2910 	 * the waiter is marked "done"), so without this check we could
2911 	 * wind up with a use-after-free error below.
2912 	 */
2913 	if (zcw->zcw_done)
2914 		goto out;
2915 
2916 	ASSERT3P(lwb, ==, zcw->zcw_lwb);
2917 
2918 	/*
2919 	 * We've already checked this above, but since we hadn't acquired
2920 	 * the zilog's zl_issuer_lock, we have to perform this check a
2921 	 * second time while holding the lock.
2922 	 *
2923 	 * We don't need to hold the zl_lock since the lwb cannot transition
2924 	 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2925 	 * _can_ transition from ISSUED to DONE, but it's OK to race with
2926 	 * that transition since we treat the lwb the same, whether it's in
2927 	 * the ISSUED or DONE states.
2928 	 *
2929 	 * The important thing, is we treat the lwb differently depending on
2930 	 * if it's ISSUED or OPENED, and block any other threads that might
2931 	 * attempt to issue this lwb. For that reason we hold the
2932 	 * zl_issuer_lock when checking the lwb_state; we must not call
2933 	 * zil_lwb_write_issue() if the lwb had already been issued.
2934 	 *
2935 	 * See the comment above the lwb_state_t structure definition for
2936 	 * more details on the lwb states, and locking requirements.
2937 	 */
2938 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2939 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2940 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2941 		goto out;
2942 
2943 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2944 
2945 	/*
2946 	 * As described in the comments above zil_commit_waiter() and
2947 	 * zil_process_commit_list(), we need to issue this lwb's zio
2948 	 * since we've reached the commit waiter's timeout and it still
2949 	 * hasn't been issued.
2950 	 */
2951 	lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2952 
2953 	IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2954 
2955 	/*
2956 	 * Since the lwb's zio hadn't been issued by the time this thread
2957 	 * reached its timeout, we reset the zilog's "zl_cur_used" field
2958 	 * to influence the zil block size selection algorithm.
2959 	 *
2960 	 * By having to issue the lwb's zio here, it means the size of the
2961 	 * lwb was too large, given the incoming throughput of itxs.  By
2962 	 * setting "zl_cur_used" to zero, we communicate this fact to the
2963 	 * block size selection algorithm, so it can take this information
2964 	 * into account, and potentially select a smaller size for the
2965 	 * next lwb block that is allocated.
2966 	 */
2967 	zilog->zl_cur_used = 0;
2968 
2969 	if (nlwb == NULL) {
2970 		/*
2971 		 * When zil_lwb_write_issue() returns NULL, this
2972 		 * indicates zio_alloc_zil() failed to allocate the
2973 		 * "next" lwb on-disk. When this occurs, the ZIL write
2974 		 * pipeline must be stalled; see the comment within the
2975 		 * zil_commit_writer_stall() function for more details.
2976 		 *
2977 		 * We must drop the commit waiter's lock prior to
2978 		 * calling zil_commit_writer_stall() or else we can wind
2979 		 * up with the following deadlock:
2980 		 *
2981 		 * - This thread is waiting for the txg to sync while
2982 		 *   holding the waiter's lock; txg_wait_synced() is
2983 		 *   used within txg_commit_writer_stall().
2984 		 *
2985 		 * - The txg can't sync because it is waiting for this
2986 		 *   lwb's zio callback to call dmu_tx_commit().
2987 		 *
2988 		 * - The lwb's zio callback can't call dmu_tx_commit()
2989 		 *   because it's blocked trying to acquire the waiter's
2990 		 *   lock, which occurs prior to calling dmu_tx_commit()
2991 		 */
2992 		mutex_exit(&zcw->zcw_lock);
2993 		zil_commit_writer_stall(zilog);
2994 		mutex_enter(&zcw->zcw_lock);
2995 	}
2996 
2997 out:
2998 	mutex_exit(&zilog->zl_issuer_lock);
2999 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3000 }
3001 
3002 /*
3003  * This function is responsible for performing the following two tasks:
3004  *
3005  * 1. its primary responsibility is to block until the given "commit
3006  *    waiter" is considered "done".
3007  *
3008  * 2. its secondary responsibility is to issue the zio for the lwb that
3009  *    the given "commit waiter" is waiting on, if this function has
3010  *    waited "long enough" and the lwb is still in the "open" state.
3011  *
3012  * Given a sufficient amount of itxs being generated and written using
3013  * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
3014  * function. If this does not occur, this secondary responsibility will
3015  * ensure the lwb is issued even if there is not other synchronous
3016  * activity on the system.
3017  *
3018  * For more details, see zil_process_commit_list(); more specifically,
3019  * the comment at the bottom of that function.
3020  */
3021 static void
3022 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
3023 {
3024 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
3025 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3026 	ASSERT(spa_writeable(zilog->zl_spa));
3027 
3028 	mutex_enter(&zcw->zcw_lock);
3029 
3030 	/*
3031 	 * The timeout is scaled based on the lwb latency to avoid
3032 	 * significantly impacting the latency of each individual itx.
3033 	 * For more details, see the comment at the bottom of the
3034 	 * zil_process_commit_list() function.
3035 	 */
3036 	int pct = MAX(zfs_commit_timeout_pct, 1);
3037 	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
3038 	hrtime_t wakeup = gethrtime() + sleep;
3039 	boolean_t timedout = B_FALSE;
3040 
3041 	while (!zcw->zcw_done) {
3042 		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3043 
3044 		lwb_t *lwb = zcw->zcw_lwb;
3045 
3046 		/*
3047 		 * Usually, the waiter will have a non-NULL lwb field here,
3048 		 * but it's possible for it to be NULL as a result of
3049 		 * zil_commit() racing with spa_sync().
3050 		 *
3051 		 * When zil_clean() is called, it's possible for the itxg
3052 		 * list (which may be cleaned via a taskq) to contain
3053 		 * commit itxs. When this occurs, the commit waiters linked
3054 		 * off of these commit itxs will not be committed to an
3055 		 * lwb.  Additionally, these commit waiters will not be
3056 		 * marked done until zil_commit_waiter_skip() is called via
3057 		 * zil_itxg_clean().
3058 		 *
3059 		 * Thus, it's possible for this commit waiter (i.e. the
3060 		 * "zcw" variable) to be found in this "in between" state;
3061 		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
3062 		 * been skipped, so it's "zcw_done" field is still B_FALSE.
3063 		 */
3064 		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
3065 
3066 		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
3067 			ASSERT3B(timedout, ==, B_FALSE);
3068 
3069 			/*
3070 			 * If the lwb hasn't been issued yet, then we
3071 			 * need to wait with a timeout, in case this
3072 			 * function needs to issue the lwb after the
3073 			 * timeout is reached; responsibility (2) from
3074 			 * the comment above this function.
3075 			 */
3076 			int rc = cv_timedwait_hires(&zcw->zcw_cv,
3077 			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
3078 			    CALLOUT_FLAG_ABSOLUTE);
3079 
3080 			if (rc != -1 || zcw->zcw_done)
3081 				continue;
3082 
3083 			timedout = B_TRUE;
3084 			zil_commit_waiter_timeout(zilog, zcw);
3085 
3086 			if (!zcw->zcw_done) {
3087 				/*
3088 				 * If the commit waiter has already been
3089 				 * marked "done", it's possible for the
3090 				 * waiter's lwb structure to have already
3091 				 * been freed.  Thus, we can only reliably
3092 				 * make these assertions if the waiter
3093 				 * isn't done.
3094 				 */
3095 				ASSERT3P(lwb, ==, zcw->zcw_lwb);
3096 				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
3097 			}
3098 		} else {
3099 			/*
3100 			 * If the lwb isn't open, then it must have already
3101 			 * been issued. In that case, there's no need to
3102 			 * use a timeout when waiting for the lwb to
3103 			 * complete.
3104 			 *
3105 			 * Additionally, if the lwb is NULL, the waiter
3106 			 * will soon be signaled and marked done via
3107 			 * zil_clean() and zil_itxg_clean(), so no timeout
3108 			 * is required.
3109 			 */
3110 
3111 			IMPLY(lwb != NULL,
3112 			    lwb->lwb_state == LWB_STATE_ISSUED ||
3113 			    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
3114 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
3115 			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
3116 		}
3117 	}
3118 
3119 	mutex_exit(&zcw->zcw_lock);
3120 }
3121 
3122 static zil_commit_waiter_t *
3123 zil_alloc_commit_waiter(void)
3124 {
3125 	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
3126 
3127 	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
3128 	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
3129 	list_link_init(&zcw->zcw_node);
3130 	zcw->zcw_lwb = NULL;
3131 	zcw->zcw_done = B_FALSE;
3132 	zcw->zcw_zio_error = 0;
3133 
3134 	return (zcw);
3135 }
3136 
3137 static void
3138 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
3139 {
3140 	ASSERT(!list_link_active(&zcw->zcw_node));
3141 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
3142 	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
3143 	mutex_destroy(&zcw->zcw_lock);
3144 	cv_destroy(&zcw->zcw_cv);
3145 	kmem_cache_free(zil_zcw_cache, zcw);
3146 }
3147 
3148 /*
3149  * This function is used to create a TX_COMMIT itx and assign it. This
3150  * way, it will be linked into the ZIL's list of synchronous itxs, and
3151  * then later committed to an lwb (or skipped) when
3152  * zil_process_commit_list() is called.
3153  */
3154 static void
3155 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
3156 {
3157 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
3158 
3159 	/*
3160 	 * Since we are not going to create any new dirty data, and we
3161 	 * can even help with clearing the existing dirty data, we
3162 	 * should not be subject to the dirty data based delays. We
3163 	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
3164 	 */
3165 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
3166 
3167 	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
3168 	itx->itx_sync = B_TRUE;
3169 	itx->itx_private = zcw;
3170 
3171 	zil_itx_assign(zilog, itx, tx);
3172 
3173 	dmu_tx_commit(tx);
3174 }
3175 
3176 /*
3177  * Commit ZFS Intent Log transactions (itxs) to stable storage.
3178  *
3179  * When writing ZIL transactions to the on-disk representation of the
3180  * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3181  * itxs can be committed to a single lwb. Once a lwb is written and
3182  * committed to stable storage (i.e. the lwb is written, and vdevs have
3183  * been flushed), each itx that was committed to that lwb is also
3184  * considered to be committed to stable storage.
3185  *
3186  * When an itx is committed to an lwb, the log record (lr_t) contained
3187  * by the itx is copied into the lwb's zio buffer, and once this buffer
3188  * is written to disk, it becomes an on-disk ZIL block.
3189  *
3190  * As itxs are generated, they're inserted into the ZIL's queue of
3191  * uncommitted itxs. The semantics of zil_commit() are such that it will
3192  * block until all itxs that were in the queue when it was called, are
3193  * committed to stable storage.
3194  *
3195  * If "foid" is zero, this means all "synchronous" and "asynchronous"
3196  * itxs, for all objects in the dataset, will be committed to stable
3197  * storage prior to zil_commit() returning. If "foid" is non-zero, all
3198  * "synchronous" itxs for all objects, but only "asynchronous" itxs
3199  * that correspond to the foid passed in, will be committed to stable
3200  * storage prior to zil_commit() returning.
3201  *
3202  * Generally speaking, when zil_commit() is called, the consumer doesn't
3203  * actually care about _all_ of the uncommitted itxs. Instead, they're
3204  * simply trying to waiting for a specific itx to be committed to disk,
3205  * but the interface(s) for interacting with the ZIL don't allow such
3206  * fine-grained communication. A better interface would allow a consumer
3207  * to create and assign an itx, and then pass a reference to this itx to
3208  * zil_commit(); such that zil_commit() would return as soon as that
3209  * specific itx was committed to disk (instead of waiting for _all_
3210  * itxs to be committed).
3211  *
3212  * When a thread calls zil_commit() a special "commit itx" will be
3213  * generated, along with a corresponding "waiter" for this commit itx.
3214  * zil_commit() will wait on this waiter's CV, such that when the waiter
3215  * is marked done, and signaled, zil_commit() will return.
3216  *
3217  * This commit itx is inserted into the queue of uncommitted itxs. This
3218  * provides an easy mechanism for determining which itxs were in the
3219  * queue prior to zil_commit() having been called, and which itxs were
3220  * added after zil_commit() was called.
3221  *
3222  * The commit itx is special; it doesn't have any on-disk representation.
3223  * When a commit itx is "committed" to an lwb, the waiter associated
3224  * with it is linked onto the lwb's list of waiters. Then, when that lwb
3225  * completes, each waiter on the lwb's list is marked done and signaled
3226  * -- allowing the thread waiting on the waiter to return from zil_commit().
3227  *
3228  * It's important to point out a few critical factors that allow us
3229  * to make use of the commit itxs, commit waiters, per-lwb lists of
3230  * commit waiters, and zio completion callbacks like we're doing:
3231  *
3232  *   1. The list of waiters for each lwb is traversed, and each commit
3233  *      waiter is marked "done" and signaled, in the zio completion
3234  *      callback of the lwb's zio[*].
3235  *
3236  *      * Actually, the waiters are signaled in the zio completion
3237  *        callback of the root zio for the DKIOCFLUSHWRITECACHE commands
3238  *        that are sent to the vdevs upon completion of the lwb zio.
3239  *
3240  *   2. When the itxs are inserted into the ZIL's queue of uncommitted
3241  *      itxs, the order in which they are inserted is preserved[*]; as
3242  *      itxs are added to the queue, they are added to the tail of
3243  *      in-memory linked lists.
3244  *
3245  *      When committing the itxs to lwbs (to be written to disk), they
3246  *      are committed in the same order in which the itxs were added to
3247  *      the uncommitted queue's linked list(s); i.e. the linked list of
3248  *      itxs to commit is traversed from head to tail, and each itx is
3249  *      committed to an lwb in that order.
3250  *
3251  *      * To clarify:
3252  *
3253  *        - the order of "sync" itxs is preserved w.r.t. other
3254  *          "sync" itxs, regardless of the corresponding objects.
3255  *        - the order of "async" itxs is preserved w.r.t. other
3256  *          "async" itxs corresponding to the same object.
3257  *        - the order of "async" itxs is *not* preserved w.r.t. other
3258  *          "async" itxs corresponding to different objects.
3259  *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
3260  *          versa) is *not* preserved, even for itxs that correspond
3261  *          to the same object.
3262  *
3263  *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
3264  *      zil_get_commit_list(), and zil_process_commit_list().
3265  *
3266  *   3. The lwbs represent a linked list of blocks on disk. Thus, any
3267  *      lwb cannot be considered committed to stable storage, until its
3268  *      "previous" lwb is also committed to stable storage. This fact,
3269  *      coupled with the fact described above, means that itxs are
3270  *      committed in (roughly) the order in which they were generated.
3271  *      This is essential because itxs are dependent on prior itxs.
3272  *      Thus, we *must not* deem an itx as being committed to stable
3273  *      storage, until *all* prior itxs have also been committed to
3274  *      stable storage.
3275  *
3276  *      To enforce this ordering of lwb zio's, while still leveraging as
3277  *      much of the underlying storage performance as possible, we rely
3278  *      on two fundamental concepts:
3279  *
3280  *          1. The creation and issuance of lwb zio's is protected by
3281  *             the zilog's "zl_issuer_lock", which ensures only a single
3282  *             thread is creating and/or issuing lwb's at a time
3283  *          2. The "previous" lwb is a child of the "current" lwb
3284  *             (leveraging the zio parent-child dependency graph)
3285  *
3286  *      By relying on this parent-child zio relationship, we can have
3287  *      many lwb zio's concurrently issued to the underlying storage,
3288  *      but the order in which they complete will be the same order in
3289  *      which they were created.
3290  */
3291 void
3292 zil_commit(zilog_t *zilog, uint64_t foid)
3293 {
3294 	/*
3295 	 * We should never attempt to call zil_commit on a snapshot for
3296 	 * a couple of reasons:
3297 	 *
3298 	 * 1. A snapshot may never be modified, thus it cannot have any
3299 	 *    in-flight itxs that would have modified the dataset.
3300 	 *
3301 	 * 2. By design, when zil_commit() is called, a commit itx will
3302 	 *    be assigned to this zilog; as a result, the zilog will be
3303 	 *    dirtied. We must not dirty the zilog of a snapshot; there's
3304 	 *    checks in the code that enforce this invariant, and will
3305 	 *    cause a panic if it's not upheld.
3306 	 */
3307 	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
3308 
3309 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3310 		return;
3311 
3312 	if (!spa_writeable(zilog->zl_spa)) {
3313 		/*
3314 		 * If the SPA is not writable, there should never be any
3315 		 * pending itxs waiting to be committed to disk. If that
3316 		 * weren't true, we'd skip writing those itxs out, and
3317 		 * would break the semantics of zil_commit(); thus, we're
3318 		 * verifying that truth before we return to the caller.
3319 		 */
3320 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3321 		ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3322 		for (int i = 0; i < TXG_SIZE; i++)
3323 			ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
3324 		return;
3325 	}
3326 
3327 	/*
3328 	 * If the ZIL is suspended, we don't want to dirty it by calling
3329 	 * zil_commit_itx_assign() below, nor can we write out
3330 	 * lwbs like would be done in zil_commit_write(). Thus, we
3331 	 * simply rely on txg_wait_synced() to maintain the necessary
3332 	 * semantics, and avoid calling those functions altogether.
3333 	 */
3334 	if (zilog->zl_suspend > 0) {
3335 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3336 		return;
3337 	}
3338 
3339 	zil_commit_impl(zilog, foid);
3340 }
3341 
3342 void
3343 zil_commit_impl(zilog_t *zilog, uint64_t foid)
3344 {
3345 	ZIL_STAT_BUMP(zilog, zil_commit_count);
3346 
3347 	/*
3348 	 * Move the "async" itxs for the specified foid to the "sync"
3349 	 * queues, such that they will be later committed (or skipped)
3350 	 * to an lwb when zil_process_commit_list() is called.
3351 	 *
3352 	 * Since these "async" itxs must be committed prior to this
3353 	 * call to zil_commit returning, we must perform this operation
3354 	 * before we call zil_commit_itx_assign().
3355 	 */
3356 	zil_async_to_sync(zilog, foid);
3357 
3358 	/*
3359 	 * We allocate a new "waiter" structure which will initially be
3360 	 * linked to the commit itx using the itx's "itx_private" field.
3361 	 * Since the commit itx doesn't represent any on-disk state,
3362 	 * when it's committed to an lwb, rather than copying the its
3363 	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3364 	 * added to the lwb's list of waiters. Then, when the lwb is
3365 	 * committed to stable storage, each waiter in the lwb's list of
3366 	 * waiters will be marked "done", and signalled.
3367 	 *
3368 	 * We must create the waiter and assign the commit itx prior to
3369 	 * calling zil_commit_writer(), or else our specific commit itx
3370 	 * is not guaranteed to be committed to an lwb prior to calling
3371 	 * zil_commit_waiter().
3372 	 */
3373 	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
3374 	zil_commit_itx_assign(zilog, zcw);
3375 
3376 	zil_commit_writer(zilog, zcw);
3377 	zil_commit_waiter(zilog, zcw);
3378 
3379 	if (zcw->zcw_zio_error != 0) {
3380 		/*
3381 		 * If there was an error writing out the ZIL blocks that
3382 		 * this thread is waiting on, then we fallback to
3383 		 * relying on spa_sync() to write out the data this
3384 		 * thread is waiting on. Obviously this has performance
3385 		 * implications, but the expectation is for this to be
3386 		 * an exceptional case, and shouldn't occur often.
3387 		 */
3388 		DTRACE_PROBE2(zil__commit__io__error,
3389 		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
3390 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3391 	}
3392 
3393 	zil_free_commit_waiter(zcw);
3394 }
3395 
3396 /*
3397  * Called in syncing context to free committed log blocks and update log header.
3398  */
3399 void
3400 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
3401 {
3402 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
3403 	uint64_t txg = dmu_tx_get_txg(tx);
3404 	spa_t *spa = zilog->zl_spa;
3405 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
3406 	lwb_t *lwb;
3407 
3408 	/*
3409 	 * We don't zero out zl_destroy_txg, so make sure we don't try
3410 	 * to destroy it twice.
3411 	 */
3412 	if (spa_sync_pass(spa) != 1)
3413 		return;
3414 
3415 	zil_lwb_flush_wait_all(zilog, txg);
3416 
3417 	mutex_enter(&zilog->zl_lock);
3418 
3419 	ASSERT(zilog->zl_stop_sync == 0);
3420 
3421 	if (*replayed_seq != 0) {
3422 		ASSERT(zh->zh_replay_seq < *replayed_seq);
3423 		zh->zh_replay_seq = *replayed_seq;
3424 		*replayed_seq = 0;
3425 	}
3426 
3427 	if (zilog->zl_destroy_txg == txg) {
3428 		blkptr_t blk = zh->zh_log;
3429 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
3430 
3431 		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
3432 
3433 		memset(zh, 0, sizeof (zil_header_t));
3434 		memset(zilog->zl_replayed_seq, 0,
3435 		    sizeof (zilog->zl_replayed_seq));
3436 
3437 		if (zilog->zl_keep_first) {
3438 			/*
3439 			 * If this block was part of log chain that couldn't
3440 			 * be claimed because a device was missing during
3441 			 * zil_claim(), but that device later returns,
3442 			 * then this block could erroneously appear valid.
3443 			 * To guard against this, assign a new GUID to the new
3444 			 * log chain so it doesn't matter what blk points to.
3445 			 */
3446 			zil_init_log_chain(zilog, &blk);
3447 			zh->zh_log = blk;
3448 		} else {
3449 			/*
3450 			 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3451 			 * records. So, deactivate the feature for this dataset.
3452 			 * We activate it again when we start a new ZIL chain.
3453 			 */
3454 			if (dsl_dataset_feature_is_active(ds,
3455 			    SPA_FEATURE_ZILSAXATTR))
3456 				dsl_dataset_deactivate_feature(ds,
3457 				    SPA_FEATURE_ZILSAXATTR, tx);
3458 		}
3459 	}
3460 
3461 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
3462 		zh->zh_log = lwb->lwb_blk;
3463 		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
3464 			break;
3465 		list_remove(&zilog->zl_lwb_list, lwb);
3466 		zio_free(spa, txg, &lwb->lwb_blk);
3467 		zil_free_lwb(zilog, lwb);
3468 
3469 		/*
3470 		 * If we don't have anything left in the lwb list then
3471 		 * we've had an allocation failure and we need to zero
3472 		 * out the zil_header blkptr so that we don't end
3473 		 * up freeing the same block twice.
3474 		 */
3475 		if (list_head(&zilog->zl_lwb_list) == NULL)
3476 			BP_ZERO(&zh->zh_log);
3477 	}
3478 
3479 	/*
3480 	 * Remove fastwrite on any blocks that have been pre-allocated for
3481 	 * the next commit. This prevents fastwrite counter pollution by
3482 	 * unused, long-lived LWBs.
3483 	 */
3484 	for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
3485 		if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) {
3486 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3487 			lwb->lwb_fastwrite = 0;
3488 		}
3489 	}
3490 
3491 	mutex_exit(&zilog->zl_lock);
3492 }
3493 
3494 static int
3495 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
3496 {
3497 	(void) unused, (void) kmflag;
3498 	lwb_t *lwb = vbuf;
3499 	list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3500 	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
3501 	    offsetof(zil_commit_waiter_t, zcw_node));
3502 	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
3503 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
3504 	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
3505 	return (0);
3506 }
3507 
3508 static void
3509 zil_lwb_dest(void *vbuf, void *unused)
3510 {
3511 	(void) unused;
3512 	lwb_t *lwb = vbuf;
3513 	mutex_destroy(&lwb->lwb_vdev_lock);
3514 	avl_destroy(&lwb->lwb_vdev_tree);
3515 	list_destroy(&lwb->lwb_waiters);
3516 	list_destroy(&lwb->lwb_itxs);
3517 }
3518 
3519 void
3520 zil_init(void)
3521 {
3522 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3523 	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3524 
3525 	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3526 	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3527 
3528 	zil_sums_init(&zil_sums_global);
3529 	zil_kstats_global = kstat_create("zfs", 0, "zil", "misc",
3530 	    KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
3531 	    KSTAT_FLAG_VIRTUAL);
3532 
3533 	if (zil_kstats_global != NULL) {
3534 		zil_kstats_global->ks_data = &zil_stats;
3535 		zil_kstats_global->ks_update = zil_kstats_global_update;
3536 		zil_kstats_global->ks_private = NULL;
3537 		kstat_install(zil_kstats_global);
3538 	}
3539 }
3540 
3541 void
3542 zil_fini(void)
3543 {
3544 	kmem_cache_destroy(zil_zcw_cache);
3545 	kmem_cache_destroy(zil_lwb_cache);
3546 
3547 	if (zil_kstats_global != NULL) {
3548 		kstat_delete(zil_kstats_global);
3549 		zil_kstats_global = NULL;
3550 	}
3551 
3552 	zil_sums_fini(&zil_sums_global);
3553 }
3554 
3555 void
3556 zil_set_sync(zilog_t *zilog, uint64_t sync)
3557 {
3558 	zilog->zl_sync = sync;
3559 }
3560 
3561 void
3562 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3563 {
3564 	zilog->zl_logbias = logbias;
3565 }
3566 
3567 zilog_t *
3568 zil_alloc(objset_t *os, zil_header_t *zh_phys)
3569 {
3570 	zilog_t *zilog;
3571 
3572 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3573 
3574 	zilog->zl_header = zh_phys;
3575 	zilog->zl_os = os;
3576 	zilog->zl_spa = dmu_objset_spa(os);
3577 	zilog->zl_dmu_pool = dmu_objset_pool(os);
3578 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
3579 	zilog->zl_logbias = dmu_objset_logbias(os);
3580 	zilog->zl_sync = dmu_objset_syncprop(os);
3581 	zilog->zl_dirty_max_txg = 0;
3582 	zilog->zl_last_lwb_opened = NULL;
3583 	zilog->zl_last_lwb_latency = 0;
3584 	zilog->zl_max_block_size = zil_maxblocksize;
3585 
3586 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3587 	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3588 	mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL);
3589 
3590 	for (int i = 0; i < TXG_SIZE; i++) {
3591 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3592 		    MUTEX_DEFAULT, NULL);
3593 	}
3594 
3595 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3596 	    offsetof(lwb_t, lwb_node));
3597 
3598 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3599 	    offsetof(itx_t, itx_node));
3600 
3601 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3602 	cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL);
3603 
3604 	return (zilog);
3605 }
3606 
3607 void
3608 zil_free(zilog_t *zilog)
3609 {
3610 	int i;
3611 
3612 	zilog->zl_stop_sync = 1;
3613 
3614 	ASSERT0(zilog->zl_suspend);
3615 	ASSERT0(zilog->zl_suspending);
3616 
3617 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3618 	list_destroy(&zilog->zl_lwb_list);
3619 
3620 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3621 	list_destroy(&zilog->zl_itx_commit_list);
3622 
3623 	for (i = 0; i < TXG_SIZE; i++) {
3624 		/*
3625 		 * It's possible for an itx to be generated that doesn't dirty
3626 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3627 		 * callback to remove the entry. We remove those here.
3628 		 *
3629 		 * Also free up the ziltest itxs.
3630 		 */
3631 		if (zilog->zl_itxg[i].itxg_itxs)
3632 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3633 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3634 	}
3635 
3636 	mutex_destroy(&zilog->zl_issuer_lock);
3637 	mutex_destroy(&zilog->zl_lock);
3638 	mutex_destroy(&zilog->zl_lwb_io_lock);
3639 
3640 	cv_destroy(&zilog->zl_cv_suspend);
3641 	cv_destroy(&zilog->zl_lwb_io_cv);
3642 
3643 	kmem_free(zilog, sizeof (zilog_t));
3644 }
3645 
3646 /*
3647  * Open an intent log.
3648  */
3649 zilog_t *
3650 zil_open(objset_t *os, zil_get_data_t *get_data, zil_sums_t *zil_sums)
3651 {
3652 	zilog_t *zilog = dmu_objset_zil(os);
3653 
3654 	ASSERT3P(zilog->zl_get_data, ==, NULL);
3655 	ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3656 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3657 
3658 	zilog->zl_get_data = get_data;
3659 	zilog->zl_sums = zil_sums;
3660 
3661 	return (zilog);
3662 }
3663 
3664 /*
3665  * Close an intent log.
3666  */
3667 void
3668 zil_close(zilog_t *zilog)
3669 {
3670 	lwb_t *lwb;
3671 	uint64_t txg;
3672 
3673 	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3674 		zil_commit(zilog, 0);
3675 	} else {
3676 		ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
3677 		ASSERT0(zilog->zl_dirty_max_txg);
3678 		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3679 	}
3680 
3681 	mutex_enter(&zilog->zl_lock);
3682 	lwb = list_tail(&zilog->zl_lwb_list);
3683 	if (lwb == NULL)
3684 		txg = zilog->zl_dirty_max_txg;
3685 	else
3686 		txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
3687 	mutex_exit(&zilog->zl_lock);
3688 
3689 	/*
3690 	 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3691 	 * on the time when the dmu_tx transaction is assigned in
3692 	 * zil_lwb_write_issue().
3693 	 */
3694 	mutex_enter(&zilog->zl_lwb_io_lock);
3695 	txg = MAX(zilog->zl_lwb_max_issued_txg, txg);
3696 	mutex_exit(&zilog->zl_lwb_io_lock);
3697 
3698 	/*
3699 	 * We need to use txg_wait_synced() to wait until that txg is synced.
3700 	 * zil_sync() will guarantee all lwbs up to that txg have been
3701 	 * written out, flushed, and cleaned.
3702 	 */
3703 	if (txg != 0)
3704 		txg_wait_synced(zilog->zl_dmu_pool, txg);
3705 
3706 	if (zilog_is_dirty(zilog))
3707 		zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog,
3708 		    (u_longlong_t)txg);
3709 	if (txg < spa_freeze_txg(zilog->zl_spa))
3710 		VERIFY(!zilog_is_dirty(zilog));
3711 
3712 	zilog->zl_get_data = NULL;
3713 
3714 	/*
3715 	 * We should have only one lwb left on the list; remove it now.
3716 	 */
3717 	mutex_enter(&zilog->zl_lock);
3718 	lwb = list_head(&zilog->zl_lwb_list);
3719 	if (lwb != NULL) {
3720 		ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
3721 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3722 
3723 		if (lwb->lwb_fastwrite)
3724 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3725 
3726 		list_remove(&zilog->zl_lwb_list, lwb);
3727 		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3728 		zil_free_lwb(zilog, lwb);
3729 	}
3730 	mutex_exit(&zilog->zl_lock);
3731 }
3732 
3733 static const char *suspend_tag = "zil suspending";
3734 
3735 /*
3736  * Suspend an intent log.  While in suspended mode, we still honor
3737  * synchronous semantics, but we rely on txg_wait_synced() to do it.
3738  * On old version pools, we suspend the log briefly when taking a
3739  * snapshot so that it will have an empty intent log.
3740  *
3741  * Long holds are not really intended to be used the way we do here --
3742  * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
3743  * could fail.  Therefore we take pains to only put a long hold if it is
3744  * actually necessary.  Fortunately, it will only be necessary if the
3745  * objset is currently mounted (or the ZVOL equivalent).  In that case it
3746  * will already have a long hold, so we are not really making things any worse.
3747  *
3748  * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3749  * zvol_state_t), and use their mechanism to prevent their hold from being
3750  * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
3751  * very little gain.
3752  *
3753  * if cookiep == NULL, this does both the suspend & resume.
3754  * Otherwise, it returns with the dataset "long held", and the cookie
3755  * should be passed into zil_resume().
3756  */
3757 int
3758 zil_suspend(const char *osname, void **cookiep)
3759 {
3760 	objset_t *os;
3761 	zilog_t *zilog;
3762 	const zil_header_t *zh;
3763 	int error;
3764 
3765 	error = dmu_objset_hold(osname, suspend_tag, &os);
3766 	if (error != 0)
3767 		return (error);
3768 	zilog = dmu_objset_zil(os);
3769 
3770 	mutex_enter(&zilog->zl_lock);
3771 	zh = zilog->zl_header;
3772 
3773 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
3774 		mutex_exit(&zilog->zl_lock);
3775 		dmu_objset_rele(os, suspend_tag);
3776 		return (SET_ERROR(EBUSY));
3777 	}
3778 
3779 	/*
3780 	 * Don't put a long hold in the cases where we can avoid it.  This
3781 	 * is when there is no cookie so we are doing a suspend & resume
3782 	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3783 	 * for the suspend because it's already suspended, or there's no ZIL.
3784 	 */
3785 	if (cookiep == NULL && !zilog->zl_suspending &&
3786 	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3787 		mutex_exit(&zilog->zl_lock);
3788 		dmu_objset_rele(os, suspend_tag);
3789 		return (0);
3790 	}
3791 
3792 	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3793 	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3794 
3795 	zilog->zl_suspend++;
3796 
3797 	if (zilog->zl_suspend > 1) {
3798 		/*
3799 		 * Someone else is already suspending it.
3800 		 * Just wait for them to finish.
3801 		 */
3802 
3803 		while (zilog->zl_suspending)
3804 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3805 		mutex_exit(&zilog->zl_lock);
3806 
3807 		if (cookiep == NULL)
3808 			zil_resume(os);
3809 		else
3810 			*cookiep = os;
3811 		return (0);
3812 	}
3813 
3814 	/*
3815 	 * If there is no pointer to an on-disk block, this ZIL must not
3816 	 * be active (e.g. filesystem not mounted), so there's nothing
3817 	 * to clean up.
3818 	 */
3819 	if (BP_IS_HOLE(&zh->zh_log)) {
3820 		ASSERT(cookiep != NULL); /* fast path already handled */
3821 
3822 		*cookiep = os;
3823 		mutex_exit(&zilog->zl_lock);
3824 		return (0);
3825 	}
3826 
3827 	/*
3828 	 * The ZIL has work to do. Ensure that the associated encryption
3829 	 * key will remain mapped while we are committing the log by
3830 	 * grabbing a reference to it. If the key isn't loaded we have no
3831 	 * choice but to return an error until the wrapping key is loaded.
3832 	 */
3833 	if (os->os_encrypted &&
3834 	    dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
3835 		zilog->zl_suspend--;
3836 		mutex_exit(&zilog->zl_lock);
3837 		dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3838 		dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3839 		return (SET_ERROR(EACCES));
3840 	}
3841 
3842 	zilog->zl_suspending = B_TRUE;
3843 	mutex_exit(&zilog->zl_lock);
3844 
3845 	/*
3846 	 * We need to use zil_commit_impl to ensure we wait for all
3847 	 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
3848 	 * to disk before proceeding. If we used zil_commit instead, it
3849 	 * would just call txg_wait_synced(), because zl_suspend is set.
3850 	 * txg_wait_synced() doesn't wait for these lwb's to be
3851 	 * LWB_STATE_FLUSH_DONE before returning.
3852 	 */
3853 	zil_commit_impl(zilog, 0);
3854 
3855 	/*
3856 	 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3857 	 * use txg_wait_synced() to ensure the data from the zilog has
3858 	 * migrated to the main pool before calling zil_destroy().
3859 	 */
3860 	txg_wait_synced(zilog->zl_dmu_pool, 0);
3861 
3862 	zil_destroy(zilog, B_FALSE);
3863 
3864 	mutex_enter(&zilog->zl_lock);
3865 	zilog->zl_suspending = B_FALSE;
3866 	cv_broadcast(&zilog->zl_cv_suspend);
3867 	mutex_exit(&zilog->zl_lock);
3868 
3869 	if (os->os_encrypted)
3870 		dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
3871 
3872 	if (cookiep == NULL)
3873 		zil_resume(os);
3874 	else
3875 		*cookiep = os;
3876 	return (0);
3877 }
3878 
3879 void
3880 zil_resume(void *cookie)
3881 {
3882 	objset_t *os = cookie;
3883 	zilog_t *zilog = dmu_objset_zil(os);
3884 
3885 	mutex_enter(&zilog->zl_lock);
3886 	ASSERT(zilog->zl_suspend != 0);
3887 	zilog->zl_suspend--;
3888 	mutex_exit(&zilog->zl_lock);
3889 	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3890 	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3891 }
3892 
3893 typedef struct zil_replay_arg {
3894 	zil_replay_func_t *const *zr_replay;
3895 	void		*zr_arg;
3896 	boolean_t	zr_byteswap;
3897 	char		*zr_lr;
3898 } zil_replay_arg_t;
3899 
3900 static int
3901 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
3902 {
3903 	char name[ZFS_MAX_DATASET_NAME_LEN];
3904 
3905 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
3906 
3907 	dmu_objset_name(zilog->zl_os, name);
3908 
3909 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3910 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3911 	    (u_longlong_t)lr->lrc_seq,
3912 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3913 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
3914 
3915 	return (error);
3916 }
3917 
3918 static int
3919 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
3920     uint64_t claim_txg)
3921 {
3922 	zil_replay_arg_t *zr = zra;
3923 	const zil_header_t *zh = zilog->zl_header;
3924 	uint64_t reclen = lr->lrc_reclen;
3925 	uint64_t txtype = lr->lrc_txtype;
3926 	int error = 0;
3927 
3928 	zilog->zl_replaying_seq = lr->lrc_seq;
3929 
3930 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
3931 		return (0);
3932 
3933 	if (lr->lrc_txg < claim_txg)		/* already committed */
3934 		return (0);
3935 
3936 	/* Strip case-insensitive bit, still present in log record */
3937 	txtype &= ~TX_CI;
3938 
3939 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
3940 		return (zil_replay_error(zilog, lr, EINVAL));
3941 
3942 	/*
3943 	 * If this record type can be logged out of order, the object
3944 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
3945 	 */
3946 	if (TX_OOO(txtype)) {
3947 		error = dmu_object_info(zilog->zl_os,
3948 		    LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
3949 		if (error == ENOENT || error == EEXIST)
3950 			return (0);
3951 	}
3952 
3953 	/*
3954 	 * Make a copy of the data so we can revise and extend it.
3955 	 */
3956 	memcpy(zr->zr_lr, lr, reclen);
3957 
3958 	/*
3959 	 * If this is a TX_WRITE with a blkptr, suck in the data.
3960 	 */
3961 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3962 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
3963 		    zr->zr_lr + reclen);
3964 		if (error != 0)
3965 			return (zil_replay_error(zilog, lr, error));
3966 	}
3967 
3968 	/*
3969 	 * The log block containing this lr may have been byteswapped
3970 	 * so that we can easily examine common fields like lrc_txtype.
3971 	 * However, the log is a mix of different record types, and only the
3972 	 * replay vectors know how to byteswap their records.  Therefore, if
3973 	 * the lr was byteswapped, undo it before invoking the replay vector.
3974 	 */
3975 	if (zr->zr_byteswap)
3976 		byteswap_uint64_array(zr->zr_lr, reclen);
3977 
3978 	/*
3979 	 * We must now do two things atomically: replay this log record,
3980 	 * and update the log header sequence number to reflect the fact that
3981 	 * we did so. At the end of each replay function the sequence number
3982 	 * is updated if we are in replay mode.
3983 	 */
3984 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3985 	if (error != 0) {
3986 		/*
3987 		 * The DMU's dnode layer doesn't see removes until the txg
3988 		 * commits, so a subsequent claim can spuriously fail with
3989 		 * EEXIST. So if we receive any error we try syncing out
3990 		 * any removes then retry the transaction.  Note that we
3991 		 * specify B_FALSE for byteswap now, so we don't do it twice.
3992 		 */
3993 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3994 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3995 		if (error != 0)
3996 			return (zil_replay_error(zilog, lr, error));
3997 	}
3998 	return (0);
3999 }
4000 
4001 static int
4002 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
4003 {
4004 	(void) bp, (void) arg, (void) claim_txg;
4005 
4006 	zilog->zl_replay_blks++;
4007 
4008 	return (0);
4009 }
4010 
4011 /*
4012  * If this dataset has a non-empty intent log, replay it and destroy it.
4013  * Return B_TRUE if there were any entries to replay.
4014  */
4015 boolean_t
4016 zil_replay(objset_t *os, void *arg,
4017     zil_replay_func_t *const replay_func[TX_MAX_TYPE])
4018 {
4019 	zilog_t *zilog = dmu_objset_zil(os);
4020 	const zil_header_t *zh = zilog->zl_header;
4021 	zil_replay_arg_t zr;
4022 
4023 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
4024 		return (zil_destroy(zilog, B_TRUE));
4025 	}
4026 
4027 	zr.zr_replay = replay_func;
4028 	zr.zr_arg = arg;
4029 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
4030 	zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
4031 
4032 	/*
4033 	 * Wait for in-progress removes to sync before starting replay.
4034 	 */
4035 	txg_wait_synced(zilog->zl_dmu_pool, 0);
4036 
4037 	zilog->zl_replay = B_TRUE;
4038 	zilog->zl_replay_time = ddi_get_lbolt();
4039 	ASSERT(zilog->zl_replay_blks == 0);
4040 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
4041 	    zh->zh_claim_txg, B_TRUE);
4042 	vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
4043 
4044 	zil_destroy(zilog, B_FALSE);
4045 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
4046 	zilog->zl_replay = B_FALSE;
4047 
4048 	return (B_TRUE);
4049 }
4050 
4051 boolean_t
4052 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
4053 {
4054 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
4055 		return (B_TRUE);
4056 
4057 	if (zilog->zl_replay) {
4058 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
4059 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
4060 		    zilog->zl_replaying_seq;
4061 		return (B_TRUE);
4062 	}
4063 
4064 	return (B_FALSE);
4065 }
4066 
4067 int
4068 zil_reset(const char *osname, void *arg)
4069 {
4070 	(void) arg;
4071 
4072 	int error = zil_suspend(osname, NULL);
4073 	/* EACCES means crypto key not loaded */
4074 	if ((error == EACCES) || (error == EBUSY))
4075 		return (SET_ERROR(error));
4076 	if (error != 0)
4077 		return (SET_ERROR(EEXIST));
4078 	return (0);
4079 }
4080 
4081 EXPORT_SYMBOL(zil_alloc);
4082 EXPORT_SYMBOL(zil_free);
4083 EXPORT_SYMBOL(zil_open);
4084 EXPORT_SYMBOL(zil_close);
4085 EXPORT_SYMBOL(zil_replay);
4086 EXPORT_SYMBOL(zil_replaying);
4087 EXPORT_SYMBOL(zil_destroy);
4088 EXPORT_SYMBOL(zil_destroy_sync);
4089 EXPORT_SYMBOL(zil_itx_create);
4090 EXPORT_SYMBOL(zil_itx_destroy);
4091 EXPORT_SYMBOL(zil_itx_assign);
4092 EXPORT_SYMBOL(zil_commit);
4093 EXPORT_SYMBOL(zil_claim);
4094 EXPORT_SYMBOL(zil_check_log_chain);
4095 EXPORT_SYMBOL(zil_sync);
4096 EXPORT_SYMBOL(zil_clean);
4097 EXPORT_SYMBOL(zil_suspend);
4098 EXPORT_SYMBOL(zil_resume);
4099 EXPORT_SYMBOL(zil_lwb_add_block);
4100 EXPORT_SYMBOL(zil_bp_tree_add);
4101 EXPORT_SYMBOL(zil_set_sync);
4102 EXPORT_SYMBOL(zil_set_logbias);
4103 EXPORT_SYMBOL(zil_sums_init);
4104 EXPORT_SYMBOL(zil_sums_fini);
4105 EXPORT_SYMBOL(zil_kstat_values_update);
4106 
4107 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, UINT, ZMOD_RW,
4108 	"ZIL block open timeout percentage");
4109 
4110 ZFS_MODULE_PARAM(zfs_zil, zil_, min_commit_timeout, U64, ZMOD_RW,
4111 	"Minimum delay we care for ZIL block commit");
4112 
4113 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
4114 	"Disable intent logging replay");
4115 
4116 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
4117 	"Disable ZIL cache flushes");
4118 
4119 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, U64, ZMOD_RW,
4120 	"Limit in bytes slog sync writes per commit");
4121 
4122 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, UINT, ZMOD_RW,
4123 	"Limit in bytes of ZIL log block size");
4124