xref: /freebsd/sys/contrib/openzfs/module/zfs/zil.c (revision d5f55356a2fbf8222fb236fe509821e12f1ea456)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25  * Copyright (c) 2014 Integros [integros.com]
26  * Copyright (c) 2018 Datto Inc.
27  * Copyright (c) 2025, Klara, Inc.
28  */
29 
30 /* Portions Copyright 2010 Robert Milkowski */
31 
32 #include <sys/zfs_context.h>
33 #include <sys/spa.h>
34 #include <sys/spa_impl.h>
35 #include <sys/dmu.h>
36 #include <sys/zap.h>
37 #include <sys/arc.h>
38 #include <sys/stat.h>
39 #include <sys/zil.h>
40 #include <sys/zil_impl.h>
41 #include <sys/dsl_dataset.h>
42 #include <sys/vdev_impl.h>
43 #include <sys/dmu_tx.h>
44 #include <sys/dsl_pool.h>
45 #include <sys/metaslab.h>
46 #include <sys/trace_zfs.h>
47 #include <sys/abd.h>
48 #include <sys/brt.h>
49 #include <sys/wmsum.h>
50 
51 /*
52  * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
53  * calls that change the file system. Each itx has enough information to
54  * be able to replay them after a system crash, power loss, or
55  * equivalent failure mode. These are stored in memory until either:
56  *
57  *   1. they are committed to the pool by the DMU transaction group
58  *      (txg), at which point they can be discarded; or
59  *   2. they are committed to the on-disk ZIL for the dataset being
60  *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
61  *      requirement).
62  *
63  * In the event of a crash or power loss, the itxs contained by each
64  * dataset's on-disk ZIL will be replayed when that dataset is first
65  * instantiated (e.g. if the dataset is a normal filesystem, when it is
66  * first mounted).
67  *
68  * As hinted at above, there is one ZIL per dataset (both the in-memory
69  * representation, and the on-disk representation). The on-disk format
70  * consists of 3 parts:
71  *
72  * 	- a single, per-dataset, ZIL header; which points to a chain of
73  * 	- zero or more ZIL blocks; each of which contains
74  * 	- zero or more ZIL records
75  *
76  * A ZIL record holds the information necessary to replay a single
77  * system call transaction. A ZIL block can hold many ZIL records, and
78  * the blocks are chained together, similarly to a singly linked list.
79  *
80  * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
81  * block in the chain, and the ZIL header points to the first block in
82  * the chain.
83  *
84  * Note, there is not a fixed place in the pool to hold these ZIL
85  * blocks; they are dynamically allocated and freed as needed from the
86  * blocks available on the pool, though they can be preferentially
87  * allocated from a dedicated "log" vdev.
88  */
89 
90 /*
91  * This controls the amount of time that a ZIL block (lwb) will remain
92  * "open" when it isn't "full", and it has a thread waiting for it to be
93  * committed to stable storage. Please refer to the zil_commit_waiter()
94  * function (and the comments within it) for more details.
95  */
96 static uint_t zfs_commit_timeout_pct = 10;
97 
98 /*
99  * See zil.h for more information about these fields.
100  */
101 static zil_kstat_values_t zil_stats = {
102 	{ "zil_commit_count",			KSTAT_DATA_UINT64 },
103 	{ "zil_commit_writer_count",		KSTAT_DATA_UINT64 },
104 	{ "zil_commit_error_count",		KSTAT_DATA_UINT64 },
105 	{ "zil_commit_stall_count",		KSTAT_DATA_UINT64 },
106 	{ "zil_commit_suspend_count",		KSTAT_DATA_UINT64 },
107 	{ "zil_commit_crash_count",		KSTAT_DATA_UINT64 },
108 	{ "zil_itx_count",			KSTAT_DATA_UINT64 },
109 	{ "zil_itx_indirect_count",		KSTAT_DATA_UINT64 },
110 	{ "zil_itx_indirect_bytes",		KSTAT_DATA_UINT64 },
111 	{ "zil_itx_copied_count",		KSTAT_DATA_UINT64 },
112 	{ "zil_itx_copied_bytes",		KSTAT_DATA_UINT64 },
113 	{ "zil_itx_needcopy_count",		KSTAT_DATA_UINT64 },
114 	{ "zil_itx_needcopy_bytes",		KSTAT_DATA_UINT64 },
115 	{ "zil_itx_metaslab_normal_count",	KSTAT_DATA_UINT64 },
116 	{ "zil_itx_metaslab_normal_bytes",	KSTAT_DATA_UINT64 },
117 	{ "zil_itx_metaslab_normal_write",	KSTAT_DATA_UINT64 },
118 	{ "zil_itx_metaslab_normal_alloc",	KSTAT_DATA_UINT64 },
119 	{ "zil_itx_metaslab_slog_count",	KSTAT_DATA_UINT64 },
120 	{ "zil_itx_metaslab_slog_bytes",	KSTAT_DATA_UINT64 },
121 	{ "zil_itx_metaslab_slog_write",	KSTAT_DATA_UINT64 },
122 	{ "zil_itx_metaslab_slog_alloc",	KSTAT_DATA_UINT64 },
123 };
124 
125 static zil_sums_t zil_sums_global;
126 static kstat_t *zil_kstats_global;
127 
128 /*
129  * Disable intent logging replay.  This global ZIL switch affects all pools.
130  */
131 int zil_replay_disable = 0;
132 
133 /*
134  * Disable the flush commands that are normally sent to the disk(s) by the ZIL
135  * after an LWB write has completed. Setting this will cause ZIL corruption on
136  * power loss if a volatile out-of-order write cache is enabled.
137  */
138 static int zil_nocacheflush = 0;
139 
140 /*
141  * Limit SLOG write size per commit executed with synchronous priority.
142  * Any writes above that will be executed with lower (asynchronous) priority
143  * to limit potential SLOG device abuse by single active ZIL writer.
144  */
145 static uint64_t zil_slog_bulk = 64 * 1024 * 1024;
146 
147 static kmem_cache_t *zil_lwb_cache;
148 static kmem_cache_t *zil_zcw_cache;
149 
150 static int zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx);
151 static itx_t *zil_itx_clone(itx_t *oitx);
152 static uint64_t zil_max_waste_space(zilog_t *zilog);
153 
154 static int
155 zil_bp_compare(const void *x1, const void *x2)
156 {
157 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
158 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
159 
160 	int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
161 	if (likely(cmp))
162 		return (cmp);
163 
164 	return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
165 }
166 
167 static void
168 zil_bp_tree_init(zilog_t *zilog)
169 {
170 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
171 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
172 }
173 
174 static void
175 zil_bp_tree_fini(zilog_t *zilog)
176 {
177 	avl_tree_t *t = &zilog->zl_bp_tree;
178 	zil_bp_node_t *zn;
179 	void *cookie = NULL;
180 
181 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
182 		kmem_free(zn, sizeof (zil_bp_node_t));
183 
184 	avl_destroy(t);
185 }
186 
187 int
188 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
189 {
190 	avl_tree_t *t = &zilog->zl_bp_tree;
191 	const dva_t *dva;
192 	zil_bp_node_t *zn;
193 	avl_index_t where;
194 
195 	if (BP_IS_EMBEDDED(bp))
196 		return (0);
197 
198 	dva = BP_IDENTITY(bp);
199 
200 	if (avl_find(t, dva, &where) != NULL)
201 		return (SET_ERROR(EEXIST));
202 
203 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
204 	zn->zn_dva = *dva;
205 	avl_insert(t, zn, where);
206 
207 	return (0);
208 }
209 
210 static zil_header_t *
211 zil_header_in_syncing_context(zilog_t *zilog)
212 {
213 	return ((zil_header_t *)zilog->zl_header);
214 }
215 
216 static void
217 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
218 {
219 	zio_cksum_t *zc = &bp->blk_cksum;
220 
221 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0],
222 	    sizeof (zc->zc_word[ZIL_ZC_GUID_0]));
223 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1],
224 	    sizeof (zc->zc_word[ZIL_ZC_GUID_1]));
225 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
226 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
227 }
228 
229 static int
230 zil_kstats_global_update(kstat_t *ksp, int rw)
231 {
232 	zil_kstat_values_t *zs = ksp->ks_data;
233 	ASSERT3P(&zil_stats, ==, zs);
234 
235 	if (rw == KSTAT_WRITE) {
236 		return (SET_ERROR(EACCES));
237 	}
238 
239 	zil_kstat_values_update(zs, &zil_sums_global);
240 
241 	return (0);
242 }
243 
244 /*
245  * Read a log block and make sure it's valid.
246  */
247 static int
248 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
249     blkptr_t *nbp, char **begin, char **end, arc_buf_t **abuf)
250 {
251 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
252 	arc_flags_t aflags = ARC_FLAG_WAIT;
253 	zbookmark_phys_t zb;
254 	int error;
255 
256 	if (zilog->zl_header->zh_claim_txg == 0)
257 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
258 
259 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
260 		zio_flags |= ZIO_FLAG_SPECULATIVE;
261 
262 	if (!decrypt)
263 		zio_flags |= ZIO_FLAG_RAW;
264 
265 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
266 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
267 
268 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
269 	    abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
270 
271 	if (error == 0) {
272 		zio_cksum_t cksum = bp->blk_cksum;
273 
274 		/*
275 		 * Validate the checksummed log block.
276 		 *
277 		 * Sequence numbers should be... sequential.  The checksum
278 		 * verifier for the next block should be bp's checksum plus 1.
279 		 *
280 		 * Also check the log chain linkage and size used.
281 		 */
282 		cksum.zc_word[ZIL_ZC_SEQ]++;
283 
284 		uint64_t size = BP_GET_LSIZE(bp);
285 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
286 			zil_chain_t *zilc = (*abuf)->b_data;
287 			char *lr = (char *)(zilc + 1);
288 
289 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
290 			    sizeof (cksum)) ||
291 			    zilc->zc_nused < sizeof (*zilc) ||
292 			    zilc->zc_nused > size) {
293 				error = SET_ERROR(ECKSUM);
294 			} else {
295 				*begin = lr;
296 				*end = lr + zilc->zc_nused - sizeof (*zilc);
297 				*nbp = zilc->zc_next_blk;
298 			}
299 		} else {
300 			char *lr = (*abuf)->b_data;
301 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
302 
303 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
304 			    sizeof (cksum)) ||
305 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
306 				error = SET_ERROR(ECKSUM);
307 			} else {
308 				*begin = lr;
309 				*end = lr + zilc->zc_nused;
310 				*nbp = zilc->zc_next_blk;
311 			}
312 		}
313 	}
314 
315 	return (error);
316 }
317 
318 /*
319  * Read a TX_WRITE log data block.
320  */
321 static int
322 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
323 {
324 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
325 	const blkptr_t *bp = &lr->lr_blkptr;
326 	arc_flags_t aflags = ARC_FLAG_WAIT;
327 	arc_buf_t *abuf = NULL;
328 	zbookmark_phys_t zb;
329 	int error;
330 
331 	if (BP_IS_HOLE(bp)) {
332 		if (wbuf != NULL)
333 			memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length));
334 		return (0);
335 	}
336 
337 	if (zilog->zl_header->zh_claim_txg == 0)
338 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
339 
340 	/*
341 	 * If we are not using the resulting data, we are just checking that
342 	 * it hasn't been corrupted so we don't need to waste CPU time
343 	 * decompressing and decrypting it.
344 	 */
345 	if (wbuf == NULL)
346 		zio_flags |= ZIO_FLAG_RAW;
347 
348 	ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
349 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
350 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
351 
352 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
353 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
354 
355 	if (error == 0) {
356 		if (wbuf != NULL)
357 			memcpy(wbuf, abuf->b_data, arc_buf_size(abuf));
358 		arc_buf_destroy(abuf, &abuf);
359 	}
360 
361 	return (error);
362 }
363 
364 void
365 zil_sums_init(zil_sums_t *zs)
366 {
367 	wmsum_init(&zs->zil_commit_count, 0);
368 	wmsum_init(&zs->zil_commit_writer_count, 0);
369 	wmsum_init(&zs->zil_commit_error_count, 0);
370 	wmsum_init(&zs->zil_commit_stall_count, 0);
371 	wmsum_init(&zs->zil_commit_suspend_count, 0);
372 	wmsum_init(&zs->zil_commit_crash_count, 0);
373 	wmsum_init(&zs->zil_itx_count, 0);
374 	wmsum_init(&zs->zil_itx_indirect_count, 0);
375 	wmsum_init(&zs->zil_itx_indirect_bytes, 0);
376 	wmsum_init(&zs->zil_itx_copied_count, 0);
377 	wmsum_init(&zs->zil_itx_copied_bytes, 0);
378 	wmsum_init(&zs->zil_itx_needcopy_count, 0);
379 	wmsum_init(&zs->zil_itx_needcopy_bytes, 0);
380 	wmsum_init(&zs->zil_itx_metaslab_normal_count, 0);
381 	wmsum_init(&zs->zil_itx_metaslab_normal_bytes, 0);
382 	wmsum_init(&zs->zil_itx_metaslab_normal_write, 0);
383 	wmsum_init(&zs->zil_itx_metaslab_normal_alloc, 0);
384 	wmsum_init(&zs->zil_itx_metaslab_slog_count, 0);
385 	wmsum_init(&zs->zil_itx_metaslab_slog_bytes, 0);
386 	wmsum_init(&zs->zil_itx_metaslab_slog_write, 0);
387 	wmsum_init(&zs->zil_itx_metaslab_slog_alloc, 0);
388 }
389 
390 void
391 zil_sums_fini(zil_sums_t *zs)
392 {
393 	wmsum_fini(&zs->zil_commit_count);
394 	wmsum_fini(&zs->zil_commit_writer_count);
395 	wmsum_fini(&zs->zil_commit_error_count);
396 	wmsum_fini(&zs->zil_commit_stall_count);
397 	wmsum_fini(&zs->zil_commit_suspend_count);
398 	wmsum_fini(&zs->zil_commit_crash_count);
399 	wmsum_fini(&zs->zil_itx_count);
400 	wmsum_fini(&zs->zil_itx_indirect_count);
401 	wmsum_fini(&zs->zil_itx_indirect_bytes);
402 	wmsum_fini(&zs->zil_itx_copied_count);
403 	wmsum_fini(&zs->zil_itx_copied_bytes);
404 	wmsum_fini(&zs->zil_itx_needcopy_count);
405 	wmsum_fini(&zs->zil_itx_needcopy_bytes);
406 	wmsum_fini(&zs->zil_itx_metaslab_normal_count);
407 	wmsum_fini(&zs->zil_itx_metaslab_normal_bytes);
408 	wmsum_fini(&zs->zil_itx_metaslab_normal_write);
409 	wmsum_fini(&zs->zil_itx_metaslab_normal_alloc);
410 	wmsum_fini(&zs->zil_itx_metaslab_slog_count);
411 	wmsum_fini(&zs->zil_itx_metaslab_slog_bytes);
412 	wmsum_fini(&zs->zil_itx_metaslab_slog_write);
413 	wmsum_fini(&zs->zil_itx_metaslab_slog_alloc);
414 }
415 
416 void
417 zil_kstat_values_update(zil_kstat_values_t *zs, zil_sums_t *zil_sums)
418 {
419 	zs->zil_commit_count.value.ui64 =
420 	    wmsum_value(&zil_sums->zil_commit_count);
421 	zs->zil_commit_writer_count.value.ui64 =
422 	    wmsum_value(&zil_sums->zil_commit_writer_count);
423 	zs->zil_commit_error_count.value.ui64 =
424 	    wmsum_value(&zil_sums->zil_commit_error_count);
425 	zs->zil_commit_stall_count.value.ui64 =
426 	    wmsum_value(&zil_sums->zil_commit_stall_count);
427 	zs->zil_commit_suspend_count.value.ui64 =
428 	    wmsum_value(&zil_sums->zil_commit_suspend_count);
429 	zs->zil_commit_crash_count.value.ui64 =
430 	    wmsum_value(&zil_sums->zil_commit_crash_count);
431 	zs->zil_itx_count.value.ui64 =
432 	    wmsum_value(&zil_sums->zil_itx_count);
433 	zs->zil_itx_indirect_count.value.ui64 =
434 	    wmsum_value(&zil_sums->zil_itx_indirect_count);
435 	zs->zil_itx_indirect_bytes.value.ui64 =
436 	    wmsum_value(&zil_sums->zil_itx_indirect_bytes);
437 	zs->zil_itx_copied_count.value.ui64 =
438 	    wmsum_value(&zil_sums->zil_itx_copied_count);
439 	zs->zil_itx_copied_bytes.value.ui64 =
440 	    wmsum_value(&zil_sums->zil_itx_copied_bytes);
441 	zs->zil_itx_needcopy_count.value.ui64 =
442 	    wmsum_value(&zil_sums->zil_itx_needcopy_count);
443 	zs->zil_itx_needcopy_bytes.value.ui64 =
444 	    wmsum_value(&zil_sums->zil_itx_needcopy_bytes);
445 	zs->zil_itx_metaslab_normal_count.value.ui64 =
446 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_count);
447 	zs->zil_itx_metaslab_normal_bytes.value.ui64 =
448 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_bytes);
449 	zs->zil_itx_metaslab_normal_write.value.ui64 =
450 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_write);
451 	zs->zil_itx_metaslab_normal_alloc.value.ui64 =
452 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_alloc);
453 	zs->zil_itx_metaslab_slog_count.value.ui64 =
454 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_count);
455 	zs->zil_itx_metaslab_slog_bytes.value.ui64 =
456 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_bytes);
457 	zs->zil_itx_metaslab_slog_write.value.ui64 =
458 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_write);
459 	zs->zil_itx_metaslab_slog_alloc.value.ui64 =
460 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_alloc);
461 }
462 
463 /*
464  * Parse the intent log, and call parse_func for each valid record within.
465  */
466 int
467 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
468     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
469     boolean_t decrypt)
470 {
471 	const zil_header_t *zh = zilog->zl_header;
472 	boolean_t claimed = !!zh->zh_claim_txg;
473 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
474 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
475 	uint64_t max_blk_seq = 0;
476 	uint64_t max_lr_seq = 0;
477 	uint64_t blk_count = 0;
478 	uint64_t lr_count = 0;
479 	blkptr_t blk, next_blk = {{{{0}}}};
480 	int error = 0;
481 
482 	/*
483 	 * Old logs didn't record the maximum zh_claim_lr_seq.
484 	 */
485 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
486 		claim_lr_seq = UINT64_MAX;
487 
488 	/*
489 	 * Starting at the block pointed to by zh_log we read the log chain.
490 	 * For each block in the chain we strongly check that block to
491 	 * ensure its validity.  We stop when an invalid block is found.
492 	 * For each block pointer in the chain we call parse_blk_func().
493 	 * For each record in each valid block we call parse_lr_func().
494 	 * If the log has been claimed, stop if we encounter a sequence
495 	 * number greater than the highest claimed sequence number.
496 	 */
497 	zil_bp_tree_init(zilog);
498 
499 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
500 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
501 		int reclen;
502 		char *lrp, *end;
503 		arc_buf_t *abuf = NULL;
504 
505 		if (blk_seq > claim_blk_seq)
506 			break;
507 
508 		error = parse_blk_func(zilog, &blk, arg, txg);
509 		if (error != 0)
510 			break;
511 		ASSERT3U(max_blk_seq, <, blk_seq);
512 		max_blk_seq = blk_seq;
513 		blk_count++;
514 
515 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
516 			break;
517 
518 		error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
519 		    &lrp, &end, &abuf);
520 		if (error != 0) {
521 			if (abuf)
522 				arc_buf_destroy(abuf, &abuf);
523 			if (claimed) {
524 				char name[ZFS_MAX_DATASET_NAME_LEN];
525 
526 				dmu_objset_name(zilog->zl_os, name);
527 
528 				cmn_err(CE_WARN, "ZFS read log block error %d, "
529 				    "dataset %s, seq 0x%llx\n", error, name,
530 				    (u_longlong_t)blk_seq);
531 			}
532 			break;
533 		}
534 
535 		for (; lrp < end; lrp += reclen) {
536 			lr_t *lr = (lr_t *)lrp;
537 
538 			/*
539 			 * Are the remaining bytes large enough to hold an
540 			 * log record?
541 			 */
542 			if ((char *)(lr + 1) > end) {
543 				cmn_err(CE_WARN, "zil_parse: lr_t overrun");
544 				error = SET_ERROR(ECKSUM);
545 				arc_buf_destroy(abuf, &abuf);
546 				goto done;
547 			}
548 			reclen = lr->lrc_reclen;
549 			if (reclen < sizeof (lr_t) || reclen > end - lrp) {
550 				cmn_err(CE_WARN,
551 				    "zil_parse: lr_t has an invalid reclen");
552 				error = SET_ERROR(ECKSUM);
553 				arc_buf_destroy(abuf, &abuf);
554 				goto done;
555 			}
556 
557 			if (lr->lrc_seq > claim_lr_seq) {
558 				arc_buf_destroy(abuf, &abuf);
559 				goto done;
560 			}
561 
562 			error = parse_lr_func(zilog, lr, arg, txg);
563 			if (error != 0) {
564 				arc_buf_destroy(abuf, &abuf);
565 				goto done;
566 			}
567 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
568 			max_lr_seq = lr->lrc_seq;
569 			lr_count++;
570 		}
571 		arc_buf_destroy(abuf, &abuf);
572 	}
573 done:
574 	zilog->zl_parse_error = error;
575 	zilog->zl_parse_blk_seq = max_blk_seq;
576 	zilog->zl_parse_lr_seq = max_lr_seq;
577 	zilog->zl_parse_blk_count = blk_count;
578 	zilog->zl_parse_lr_count = lr_count;
579 
580 	zil_bp_tree_fini(zilog);
581 
582 	return (error);
583 }
584 
585 static int
586 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
587     uint64_t first_txg)
588 {
589 	(void) tx;
590 	ASSERT(!BP_IS_HOLE(bp));
591 
592 	/*
593 	 * As we call this function from the context of a rewind to a
594 	 * checkpoint, each ZIL block whose txg is later than the txg
595 	 * that we rewind to is invalid. Thus, we return -1 so
596 	 * zil_parse() doesn't attempt to read it.
597 	 */
598 	if (BP_GET_BIRTH(bp) >= first_txg)
599 		return (-1);
600 
601 	if (zil_bp_tree_add(zilog, bp) != 0)
602 		return (0);
603 
604 	zio_free(zilog->zl_spa, first_txg, bp);
605 	return (0);
606 }
607 
608 static int
609 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
610     uint64_t first_txg)
611 {
612 	(void) zilog, (void) lrc, (void) tx, (void) first_txg;
613 	return (0);
614 }
615 
616 static int
617 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
618     uint64_t first_txg)
619 {
620 	/*
621 	 * Claim log block if not already committed and not already claimed.
622 	 * If tx == NULL, just verify that the block is claimable.
623 	 */
624 	if (BP_IS_HOLE(bp) || BP_GET_BIRTH(bp) < first_txg ||
625 	    zil_bp_tree_add(zilog, bp) != 0)
626 		return (0);
627 
628 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
629 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
630 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
631 }
632 
633 static int
634 zil_claim_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t first_txg)
635 {
636 	lr_write_t *lr = (lr_write_t *)lrc;
637 	int error;
638 
639 	ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
640 
641 	/*
642 	 * If the block is not readable, don't claim it.  This can happen
643 	 * in normal operation when a log block is written to disk before
644 	 * some of the dmu_sync() blocks it points to.  In this case, the
645 	 * transaction cannot have been committed to anyone (we would have
646 	 * waited for all writes to be stable first), so it is semantically
647 	 * correct to declare this the end of the log.
648 	 */
649 	if (BP_GET_BIRTH(&lr->lr_blkptr) >= first_txg) {
650 		error = zil_read_log_data(zilog, lr, NULL);
651 		if (error != 0)
652 			return (error);
653 	}
654 
655 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
656 }
657 
658 static int
659 zil_claim_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx,
660     uint64_t first_txg)
661 {
662 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
663 	const blkptr_t *bp;
664 	spa_t *spa = zilog->zl_spa;
665 	uint_t ii;
666 
667 	ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
668 	ASSERT3U(lrc->lrc_reclen, >=, offsetof(lr_clone_range_t,
669 	    lr_bps[lr->lr_nbps]));
670 
671 	if (tx == NULL) {
672 		return (0);
673 	}
674 
675 	/*
676 	 * XXX: Do we need to byteswap lr?
677 	 */
678 
679 	for (ii = 0; ii < lr->lr_nbps; ii++) {
680 		bp = &lr->lr_bps[ii];
681 
682 		/*
683 		 * When data is embedded into the BP there is no need to create
684 		 * BRT entry as there is no data block.  Just copy the BP as it
685 		 * contains the data.
686 		 */
687 		if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
688 			continue;
689 
690 		/*
691 		 * We can not handle block pointers from the future, since they
692 		 * are not yet allocated.  It should not normally happen, but
693 		 * just in case lets be safe and just stop here now instead of
694 		 * corrupting the pool.
695 		 */
696 		if (BP_GET_PHYSICAL_BIRTH(bp) >= first_txg)
697 			return (SET_ERROR(ENOENT));
698 
699 		/*
700 		 * Assert the block is really allocated before we reference it.
701 		 */
702 		metaslab_check_free(spa, bp);
703 	}
704 
705 	for (ii = 0; ii < lr->lr_nbps; ii++) {
706 		bp = &lr->lr_bps[ii];
707 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp))
708 			brt_pending_add(spa, bp, tx);
709 	}
710 
711 	return (0);
712 }
713 
714 static int
715 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
716     uint64_t first_txg)
717 {
718 
719 	switch (lrc->lrc_txtype) {
720 	case TX_WRITE:
721 		return (zil_claim_write(zilog, lrc, tx, first_txg));
722 	case TX_CLONE_RANGE:
723 		return (zil_claim_clone_range(zilog, lrc, tx, first_txg));
724 	default:
725 		return (0);
726 	}
727 }
728 
729 static int
730 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
731     uint64_t claim_txg)
732 {
733 	(void) claim_txg;
734 
735 	zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
736 
737 	return (0);
738 }
739 
740 static int
741 zil_free_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t claim_txg)
742 {
743 	lr_write_t *lr = (lr_write_t *)lrc;
744 	blkptr_t *bp = &lr->lr_blkptr;
745 
746 	ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
747 
748 	/*
749 	 * If we previously claimed it, we need to free it.
750 	 */
751 	if (BP_GET_BIRTH(bp) >= claim_txg &&
752 	    zil_bp_tree_add(zilog, bp) == 0 && !BP_IS_HOLE(bp)) {
753 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
754 	}
755 
756 	return (0);
757 }
758 
759 static int
760 zil_free_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
761 {
762 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
763 	const blkptr_t *bp;
764 	spa_t *spa;
765 	uint_t ii;
766 
767 	ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
768 	ASSERT3U(lrc->lrc_reclen, >=, offsetof(lr_clone_range_t,
769 	    lr_bps[lr->lr_nbps]));
770 
771 	if (tx == NULL) {
772 		return (0);
773 	}
774 
775 	spa = zilog->zl_spa;
776 
777 	for (ii = 0; ii < lr->lr_nbps; ii++) {
778 		bp = &lr->lr_bps[ii];
779 
780 		if (!BP_IS_HOLE(bp)) {
781 			zio_free(spa, dmu_tx_get_txg(tx), bp);
782 		}
783 	}
784 
785 	return (0);
786 }
787 
788 static int
789 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
790     uint64_t claim_txg)
791 {
792 
793 	if (claim_txg == 0) {
794 		return (0);
795 	}
796 
797 	switch (lrc->lrc_txtype) {
798 	case TX_WRITE:
799 		return (zil_free_write(zilog, lrc, tx, claim_txg));
800 	case TX_CLONE_RANGE:
801 		return (zil_free_clone_range(zilog, lrc, tx));
802 	default:
803 		return (0);
804 	}
805 }
806 
807 static int
808 zil_lwb_vdev_compare(const void *x1, const void *x2)
809 {
810 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
811 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
812 
813 	return (TREE_CMP(v1, v2));
814 }
815 
816 /*
817  * Allocate a new lwb.  We may already have a block pointer for it, in which
818  * case we get size and version from there.  Or we may not yet, in which case
819  * we choose them here and later make the block allocation match.
820  */
821 static lwb_t *
822 zil_alloc_lwb(zilog_t *zilog, int sz, blkptr_t *bp, boolean_t slog,
823     uint64_t txg, lwb_state_t state)
824 {
825 	lwb_t *lwb;
826 
827 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
828 	lwb->lwb_zilog = zilog;
829 	if (bp) {
830 		lwb->lwb_blk = *bp;
831 		lwb->lwb_slim = (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2);
832 		sz = BP_GET_LSIZE(bp);
833 	} else {
834 		BP_ZERO(&lwb->lwb_blk);
835 		lwb->lwb_slim = (spa_version(zilog->zl_spa) >=
836 		    SPA_VERSION_SLIM_ZIL);
837 	}
838 	lwb->lwb_slog = slog;
839 	lwb->lwb_error = 0;
840 	if (lwb->lwb_slim) {
841 		lwb->lwb_nmax = sz;
842 		lwb->lwb_nused = lwb->lwb_nfilled = sizeof (zil_chain_t);
843 	} else {
844 		lwb->lwb_nmax = sz - sizeof (zil_chain_t);
845 		lwb->lwb_nused = lwb->lwb_nfilled = 0;
846 	}
847 	lwb->lwb_sz = sz;
848 	lwb->lwb_state = state;
849 	lwb->lwb_buf = zio_buf_alloc(sz);
850 	lwb->lwb_child_zio = NULL;
851 	lwb->lwb_write_zio = NULL;
852 	lwb->lwb_root_zio = NULL;
853 	lwb->lwb_issued_timestamp = 0;
854 	lwb->lwb_issued_txg = 0;
855 	lwb->lwb_alloc_txg = txg;
856 	lwb->lwb_max_txg = 0;
857 
858 	mutex_enter(&zilog->zl_lock);
859 	list_insert_tail(&zilog->zl_lwb_list, lwb);
860 	if (state != LWB_STATE_NEW)
861 		zilog->zl_last_lwb_opened = lwb;
862 	mutex_exit(&zilog->zl_lock);
863 
864 	return (lwb);
865 }
866 
867 static void
868 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
869 {
870 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
871 	ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
872 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
873 	ASSERT0P(lwb->lwb_child_zio);
874 	ASSERT0P(lwb->lwb_write_zio);
875 	ASSERT0P(lwb->lwb_root_zio);
876 	ASSERT3U(lwb->lwb_alloc_txg, <=, spa_syncing_txg(zilog->zl_spa));
877 	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
878 	VERIFY(list_is_empty(&lwb->lwb_itxs));
879 	VERIFY(list_is_empty(&lwb->lwb_waiters));
880 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
881 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
882 
883 	/*
884 	 * Clear the zilog's field to indicate this lwb is no longer
885 	 * valid, and prevent use-after-free errors.
886 	 */
887 	if (zilog->zl_last_lwb_opened == lwb)
888 		zilog->zl_last_lwb_opened = NULL;
889 
890 	kmem_cache_free(zil_lwb_cache, lwb);
891 }
892 
893 /*
894  * Called when we create in-memory log transactions so that we know
895  * to cleanup the itxs at the end of spa_sync().
896  */
897 static void
898 zilog_dirty(zilog_t *zilog, uint64_t txg)
899 {
900 	dsl_pool_t *dp = zilog->zl_dmu_pool;
901 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
902 
903 	ASSERT(spa_writeable(zilog->zl_spa));
904 
905 	if (ds->ds_is_snapshot)
906 		panic("dirtying snapshot!");
907 
908 	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
909 		/* up the hold count until we can be written out */
910 		dmu_buf_add_ref(ds->ds_dbuf, zilog);
911 
912 		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
913 	}
914 }
915 
916 /*
917  * Determine if the zil is dirty in the specified txg. Callers wanting to
918  * ensure that the dirty state does not change must hold the itxg_lock for
919  * the specified txg. Holding the lock will ensure that the zil cannot be
920  * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
921  * state.
922  */
923 static boolean_t __maybe_unused
924 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
925 {
926 	dsl_pool_t *dp = zilog->zl_dmu_pool;
927 
928 	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
929 		return (B_TRUE);
930 	return (B_FALSE);
931 }
932 
933 /*
934  * Determine if the zil is dirty. The zil is considered dirty if it has
935  * any pending itx records that have not been cleaned by zil_clean().
936  */
937 static boolean_t
938 zilog_is_dirty(zilog_t *zilog)
939 {
940 	dsl_pool_t *dp = zilog->zl_dmu_pool;
941 
942 	for (int t = 0; t < TXG_SIZE; t++) {
943 		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
944 			return (B_TRUE);
945 	}
946 	return (B_FALSE);
947 }
948 
949 /*
950  * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
951  * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
952  * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
953  * zil_commit.
954  */
955 static void
956 zil_commit_activate_saxattr_feature(zilog_t *zilog)
957 {
958 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
959 	uint64_t txg = 0;
960 	dmu_tx_t *tx = NULL;
961 
962 	if (spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
963 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL &&
964 	    !dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)) {
965 		tx = dmu_tx_create(zilog->zl_os);
966 		VERIFY0(dmu_tx_assign(tx, DMU_TX_WAIT | DMU_TX_SUSPEND));
967 		dsl_dataset_dirty(ds, tx);
968 		txg = dmu_tx_get_txg(tx);
969 
970 		mutex_enter(&ds->ds_lock);
971 		ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
972 		    (void *)B_TRUE;
973 		mutex_exit(&ds->ds_lock);
974 		dmu_tx_commit(tx);
975 		txg_wait_synced(zilog->zl_dmu_pool, txg);
976 	}
977 }
978 
979 /*
980  * Create an on-disk intent log.
981  */
982 static lwb_t *
983 zil_create(zilog_t *zilog)
984 {
985 	const zil_header_t *zh = zilog->zl_header;
986 	lwb_t *lwb = NULL;
987 	uint64_t txg = 0;
988 	dmu_tx_t *tx = NULL;
989 	blkptr_t blk;
990 	int error = 0;
991 	boolean_t slog = FALSE;
992 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
993 
994 
995 	/*
996 	 * Wait for any previous destroy to complete.
997 	 */
998 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
999 
1000 	ASSERT0(zh->zh_claim_txg);
1001 	ASSERT0(zh->zh_replay_seq);
1002 
1003 	blk = zh->zh_log;
1004 
1005 	/*
1006 	 * Allocate an initial log block if:
1007 	 *    - there isn't one already
1008 	 *    - the existing block is the wrong endianness
1009 	 */
1010 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
1011 		tx = dmu_tx_create(zilog->zl_os);
1012 		VERIFY0(dmu_tx_assign(tx, DMU_TX_WAIT | DMU_TX_SUSPEND));
1013 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1014 		txg = dmu_tx_get_txg(tx);
1015 
1016 		if (!BP_IS_HOLE(&blk)) {
1017 			zio_free(zilog->zl_spa, txg, &blk);
1018 			BP_ZERO(&blk);
1019 		}
1020 
1021 		error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
1022 		    ZIL_MIN_BLKSZ, &slog);
1023 		if (error == 0)
1024 			zil_init_log_chain(zilog, &blk);
1025 	}
1026 
1027 	/*
1028 	 * Allocate a log write block (lwb) for the first log block.
1029 	 */
1030 	if (error == 0)
1031 		lwb = zil_alloc_lwb(zilog, 0, &blk, slog, txg, LWB_STATE_NEW);
1032 
1033 	/*
1034 	 * If we just allocated the first log block, commit our transaction
1035 	 * and wait for zil_sync() to stuff the block pointer into zh_log.
1036 	 * (zh is part of the MOS, so we cannot modify it in open context.)
1037 	 */
1038 	if (tx != NULL) {
1039 		/*
1040 		 * If "zilsaxattr" feature is enabled on zpool, then activate
1041 		 * it now when we're creating the ZIL chain. We can't wait with
1042 		 * this until we write the first xattr log record because we
1043 		 * need to wait for the feature activation to sync out.
1044 		 */
1045 		if (spa_feature_is_enabled(zilog->zl_spa,
1046 		    SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) !=
1047 		    DMU_OST_ZVOL) {
1048 			mutex_enter(&ds->ds_lock);
1049 			ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
1050 			    (void *)B_TRUE;
1051 			mutex_exit(&ds->ds_lock);
1052 		}
1053 
1054 		dmu_tx_commit(tx);
1055 		txg_wait_synced(zilog->zl_dmu_pool, txg);
1056 	} else {
1057 		/*
1058 		 * This branch covers the case where we enable the feature on a
1059 		 * zpool that has existing ZIL headers.
1060 		 */
1061 		zil_commit_activate_saxattr_feature(zilog);
1062 	}
1063 	IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
1064 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL,
1065 	    dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR));
1066 
1067 	ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
1068 	IMPLY(error == 0, lwb != NULL);
1069 
1070 	return (lwb);
1071 }
1072 
1073 /*
1074  * In one tx, free all log blocks and clear the log header. If keep_first
1075  * is set, then we're replaying a log with no content. We want to keep the
1076  * first block, however, so that the first synchronous transaction doesn't
1077  * require a txg_wait_synced() in zil_create(). We don't need to
1078  * txg_wait_synced() here either when keep_first is set, because both
1079  * zil_create() and zil_destroy() will wait for any in-progress destroys
1080  * to complete.
1081  * Return B_TRUE if there were any entries to replay.
1082  */
1083 boolean_t
1084 zil_destroy(zilog_t *zilog, boolean_t keep_first)
1085 {
1086 	const zil_header_t *zh = zilog->zl_header;
1087 	lwb_t *lwb;
1088 	dmu_tx_t *tx;
1089 	uint64_t txg;
1090 
1091 	/*
1092 	 * Wait for any previous destroy to complete.
1093 	 */
1094 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1095 
1096 	zilog->zl_old_header = *zh;		/* debugging aid */
1097 
1098 	if (BP_IS_HOLE(&zh->zh_log))
1099 		return (B_FALSE);
1100 
1101 	tx = dmu_tx_create(zilog->zl_os);
1102 	VERIFY0(dmu_tx_assign(tx, DMU_TX_WAIT | DMU_TX_SUSPEND));
1103 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1104 	txg = dmu_tx_get_txg(tx);
1105 
1106 	mutex_enter(&zilog->zl_lock);
1107 
1108 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
1109 	zilog->zl_destroy_txg = txg;
1110 	zilog->zl_keep_first = keep_first;
1111 
1112 	if (!list_is_empty(&zilog->zl_lwb_list)) {
1113 		ASSERT0(zh->zh_claim_txg);
1114 		VERIFY(!keep_first);
1115 		while ((lwb = list_remove_head(&zilog->zl_lwb_list)) != NULL) {
1116 			if (lwb->lwb_buf != NULL)
1117 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1118 			if (!BP_IS_HOLE(&lwb->lwb_blk))
1119 				zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
1120 			zil_free_lwb(zilog, lwb);
1121 		}
1122 	} else if (!keep_first) {
1123 		zil_destroy_sync(zilog, tx);
1124 	}
1125 	mutex_exit(&zilog->zl_lock);
1126 
1127 	dmu_tx_commit(tx);
1128 
1129 	return (B_TRUE);
1130 }
1131 
1132 void
1133 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
1134 {
1135 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1136 	(void) zil_parse(zilog, zil_free_log_block,
1137 	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
1138 }
1139 
1140 int
1141 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
1142 {
1143 	dmu_tx_t *tx = txarg;
1144 	zilog_t *zilog;
1145 	uint64_t first_txg;
1146 	zil_header_t *zh;
1147 	objset_t *os;
1148 	int error;
1149 
1150 	error = dmu_objset_own_obj(dp, ds->ds_object,
1151 	    DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
1152 	if (error != 0) {
1153 		/*
1154 		 * EBUSY indicates that the objset is inconsistent, in which
1155 		 * case it can not have a ZIL.
1156 		 */
1157 		if (error != EBUSY) {
1158 			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
1159 			    (unsigned long long)ds->ds_object, error);
1160 		}
1161 
1162 		return (0);
1163 	}
1164 
1165 	zilog = dmu_objset_zil(os);
1166 	zh = zil_header_in_syncing_context(zilog);
1167 	ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
1168 	first_txg = spa_min_claim_txg(zilog->zl_spa);
1169 
1170 	/*
1171 	 * If the spa_log_state is not set to be cleared, check whether
1172 	 * the current uberblock is a checkpoint one and if the current
1173 	 * header has been claimed before moving on.
1174 	 *
1175 	 * If the current uberblock is a checkpointed uberblock then
1176 	 * one of the following scenarios took place:
1177 	 *
1178 	 * 1] We are currently rewinding to the checkpoint of the pool.
1179 	 * 2] We crashed in the middle of a checkpoint rewind but we
1180 	 *    did manage to write the checkpointed uberblock to the
1181 	 *    vdev labels, so when we tried to import the pool again
1182 	 *    the checkpointed uberblock was selected from the import
1183 	 *    procedure.
1184 	 *
1185 	 * In both cases we want to zero out all the ZIL blocks, except
1186 	 * the ones that have been claimed at the time of the checkpoint
1187 	 * (their zh_claim_txg != 0). The reason is that these blocks
1188 	 * may be corrupted since we may have reused their locations on
1189 	 * disk after we took the checkpoint.
1190 	 *
1191 	 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
1192 	 * when we first figure out whether the current uberblock is
1193 	 * checkpointed or not. Unfortunately, that would discard all
1194 	 * the logs, including the ones that are claimed, and we would
1195 	 * leak space.
1196 	 */
1197 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
1198 	    (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1199 	    zh->zh_claim_txg == 0)) {
1200 		if (!BP_IS_HOLE(&zh->zh_log)) {
1201 			(void) zil_parse(zilog, zil_clear_log_block,
1202 			    zil_noop_log_record, tx, first_txg, B_FALSE);
1203 		}
1204 		BP_ZERO(&zh->zh_log);
1205 		if (os->os_encrypted)
1206 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1207 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1208 		dmu_objset_disown(os, B_FALSE, FTAG);
1209 		return (0);
1210 	}
1211 
1212 	/*
1213 	 * If we are not rewinding and opening the pool normally, then
1214 	 * the min_claim_txg should be equal to the first txg of the pool.
1215 	 */
1216 	ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
1217 
1218 	/*
1219 	 * Claim all log blocks if we haven't already done so, and remember
1220 	 * the highest claimed sequence number.  This ensures that if we can
1221 	 * read only part of the log now (e.g. due to a missing device),
1222 	 * but we can read the entire log later, we will not try to replay
1223 	 * or destroy beyond the last block we successfully claimed.
1224 	 */
1225 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
1226 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
1227 		(void) zil_parse(zilog, zil_claim_log_block,
1228 		    zil_claim_log_record, tx, first_txg, B_FALSE);
1229 		zh->zh_claim_txg = first_txg;
1230 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
1231 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
1232 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
1233 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
1234 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
1235 		if (os->os_encrypted)
1236 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1237 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1238 	}
1239 
1240 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
1241 	dmu_objset_disown(os, B_FALSE, FTAG);
1242 	return (0);
1243 }
1244 
1245 /*
1246  * Check the log by walking the log chain.
1247  * Checksum errors are ok as they indicate the end of the chain.
1248  * Any other error (no device or read failure) returns an error.
1249  */
1250 int
1251 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
1252 {
1253 	(void) dp;
1254 	zilog_t *zilog;
1255 	objset_t *os;
1256 	blkptr_t *bp;
1257 	int error;
1258 
1259 	ASSERT0P(tx);
1260 
1261 	error = dmu_objset_from_ds(ds, &os);
1262 	if (error != 0) {
1263 		cmn_err(CE_WARN, "can't open objset %llu, error %d",
1264 		    (unsigned long long)ds->ds_object, error);
1265 		return (0);
1266 	}
1267 
1268 	zilog = dmu_objset_zil(os);
1269 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
1270 
1271 	if (!BP_IS_HOLE(bp)) {
1272 		vdev_t *vd;
1273 		boolean_t valid = B_TRUE;
1274 
1275 		/*
1276 		 * Check the first block and determine if it's on a log device
1277 		 * which may have been removed or faulted prior to loading this
1278 		 * pool.  If so, there's no point in checking the rest of the
1279 		 * log as its content should have already been synced to the
1280 		 * pool.
1281 		 */
1282 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
1283 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
1284 		if (vd->vdev_islog && vdev_is_dead(vd))
1285 			valid = vdev_log_state_valid(vd);
1286 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
1287 
1288 		if (!valid)
1289 			return (0);
1290 
1291 		/*
1292 		 * Check whether the current uberblock is checkpointed (e.g.
1293 		 * we are rewinding) and whether the current header has been
1294 		 * claimed or not. If it hasn't then skip verifying it. We
1295 		 * do this because its ZIL blocks may be part of the pool's
1296 		 * state before the rewind, which is no longer valid.
1297 		 */
1298 		zil_header_t *zh = zil_header_in_syncing_context(zilog);
1299 		if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1300 		    zh->zh_claim_txg == 0)
1301 			return (0);
1302 	}
1303 
1304 	/*
1305 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
1306 	 * any blocks, but just determine whether it is possible to do so.
1307 	 * In addition to checking the log chain, zil_claim_log_block()
1308 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
1309 	 * which will update spa_max_claim_txg.  See spa_load() for details.
1310 	 */
1311 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
1312 	    zilog->zl_header->zh_claim_txg ? -1ULL :
1313 	    spa_min_claim_txg(os->os_spa), B_FALSE);
1314 
1315 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
1316 }
1317 
1318 /*
1319  * When an itx is "skipped", this function is used to properly mark the
1320  * waiter as "done, and signal any thread(s) waiting on it. An itx can
1321  * be skipped (and not committed to an lwb) for a variety of reasons,
1322  * one of them being that the itx was committed via spa_sync(), prior to
1323  * it being committed to an lwb; this can happen if a thread calling
1324  * zil_commit() is racing with spa_sync().
1325  */
1326 static void
1327 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
1328 {
1329 	mutex_enter(&zcw->zcw_lock);
1330 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1331 	zcw->zcw_done = B_TRUE;
1332 	cv_broadcast(&zcw->zcw_cv);
1333 	mutex_exit(&zcw->zcw_lock);
1334 }
1335 
1336 /*
1337  * This function is used when the given waiter is to be linked into an
1338  * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1339  * At this point, the waiter will no longer be referenced by the itx,
1340  * and instead, will be referenced by the lwb.
1341  */
1342 static void
1343 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1344 {
1345 	/*
1346 	 * The lwb_waiters field of the lwb is protected by the zilog's
1347 	 * zl_issuer_lock while the lwb is open and zl_lock otherwise.
1348 	 * zl_issuer_lock also protects leaving the open state.
1349 	 * zcw_lwb setting is protected by zl_issuer_lock and state !=
1350 	 * flush_done, which transition is protected by zl_lock.
1351 	 */
1352 	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_issuer_lock));
1353 	IMPLY(lwb->lwb_state != LWB_STATE_OPENED,
1354 	    MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1355 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW);
1356 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1357 
1358 	ASSERT(!list_link_active(&zcw->zcw_node));
1359 	list_insert_tail(&lwb->lwb_waiters, zcw);
1360 	ASSERT0P(zcw->zcw_lwb);
1361 	zcw->zcw_lwb = lwb;
1362 }
1363 
1364 /*
1365  * This function is used when zio_alloc_zil() fails to allocate a ZIL
1366  * block, and the given waiter must be linked to the "nolwb waiters"
1367  * list inside of zil_process_commit_list().
1368  */
1369 static void
1370 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1371 {
1372 	ASSERT(!list_link_active(&zcw->zcw_node));
1373 	list_insert_tail(nolwb, zcw);
1374 	ASSERT0P(zcw->zcw_lwb);
1375 }
1376 
1377 void
1378 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1379 {
1380 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1381 	avl_index_t where;
1382 	zil_vdev_node_t *zv, zvsearch;
1383 	int ndvas = BP_GET_NDVAS(bp);
1384 	int i;
1385 
1386 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1387 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1388 
1389 	if (zil_nocacheflush)
1390 		return;
1391 
1392 	mutex_enter(&lwb->lwb_vdev_lock);
1393 	for (i = 0; i < ndvas; i++) {
1394 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1395 		if (avl_find(t, &zvsearch, &where) == NULL) {
1396 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1397 			zv->zv_vdev = zvsearch.zv_vdev;
1398 			avl_insert(t, zv, where);
1399 		}
1400 	}
1401 	mutex_exit(&lwb->lwb_vdev_lock);
1402 }
1403 
1404 static void
1405 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1406 {
1407 	avl_tree_t *src = &lwb->lwb_vdev_tree;
1408 	avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1409 	void *cookie = NULL;
1410 	zil_vdev_node_t *zv;
1411 
1412 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1413 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1414 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1415 
1416 	/*
1417 	 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1418 	 * not need the protection of lwb_vdev_lock (it will only be modified
1419 	 * while holding zilog->zl_lock) as its writes and those of its
1420 	 * children have all completed.  The younger 'nlwb' may be waiting on
1421 	 * future writes to additional vdevs.
1422 	 */
1423 	mutex_enter(&nlwb->lwb_vdev_lock);
1424 	/*
1425 	 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1426 	 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1427 	 */
1428 	while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1429 		avl_index_t where;
1430 
1431 		if (avl_find(dst, zv, &where) == NULL) {
1432 			avl_insert(dst, zv, where);
1433 		} else {
1434 			kmem_free(zv, sizeof (*zv));
1435 		}
1436 	}
1437 	mutex_exit(&nlwb->lwb_vdev_lock);
1438 }
1439 
1440 void
1441 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1442 {
1443 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1444 }
1445 
1446 /*
1447  * This function is a called after all vdevs associated with a given lwb write
1448  * have completed their flush command; or as soon as the lwb write completes,
1449  * if "zil_nocacheflush" is set. Further, all "previous" lwb's will have
1450  * completed before this function is called; i.e. this function is called for
1451  * all previous lwbs before it's called for "this" lwb (enforced via zio the
1452  * dependencies configured in zil_lwb_set_zio_dependency()).
1453  *
1454  * The intention is for this function to be called as soon as the contents of
1455  * an lwb are considered "stable" on disk, and will survive any sudden loss of
1456  * power. At this point, any threads waiting for the lwb to reach this state
1457  * are signalled, and the "waiter" structures are marked "done".
1458  */
1459 static void
1460 zil_lwb_flush_vdevs_done(zio_t *zio)
1461 {
1462 	lwb_t *lwb = zio->io_private;
1463 	zilog_t *zilog = lwb->lwb_zilog;
1464 	zil_commit_waiter_t *zcw;
1465 	itx_t *itx;
1466 
1467 	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1468 
1469 	hrtime_t t = gethrtime() - lwb->lwb_issued_timestamp;
1470 
1471 	mutex_enter(&zilog->zl_lock);
1472 
1473 	zilog->zl_last_lwb_latency = (zilog->zl_last_lwb_latency * 7 + t) / 8;
1474 
1475 	lwb->lwb_root_zio = NULL;
1476 
1477 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1478 	lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1479 
1480 	if (zilog->zl_last_lwb_opened == lwb) {
1481 		/*
1482 		 * Remember the highest committed log sequence number
1483 		 * for ztest. We only update this value when all the log
1484 		 * writes succeeded, because ztest wants to ASSERT that
1485 		 * it got the whole log chain.
1486 		 */
1487 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1488 	}
1489 
1490 	while ((itx = list_remove_head(&lwb->lwb_itxs)) != NULL)
1491 		zil_itx_destroy(itx, 0);
1492 
1493 	while ((zcw = list_remove_head(&lwb->lwb_waiters)) != NULL) {
1494 		mutex_enter(&zcw->zcw_lock);
1495 
1496 		ASSERT3P(zcw->zcw_lwb, ==, lwb);
1497 		zcw->zcw_lwb = NULL;
1498 		/*
1499 		 * We expect any ZIO errors from child ZIOs to have been
1500 		 * propagated "up" to this specific LWB's root ZIO, in
1501 		 * order for this error handling to work correctly. This
1502 		 * includes ZIO errors from either this LWB's write or
1503 		 * flush, as well as any errors from other dependent LWBs
1504 		 * (e.g. a root LWB ZIO that might be a child of this LWB).
1505 		 *
1506 		 * With that said, it's important to note that LWB flush
1507 		 * errors are not propagated up to the LWB root ZIO.
1508 		 * This is incorrect behavior, and results in VDEV flush
1509 		 * errors not being handled correctly here. See the
1510 		 * comment above the call to "zio_flush" for details.
1511 		 */
1512 
1513 		zcw->zcw_zio_error = zio->io_error;
1514 
1515 		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1516 		zcw->zcw_done = B_TRUE;
1517 		cv_broadcast(&zcw->zcw_cv);
1518 
1519 		mutex_exit(&zcw->zcw_lock);
1520 	}
1521 
1522 	uint64_t txg = lwb->lwb_issued_txg;
1523 
1524 	/* Once we drop the lock, lwb may be freed by zil_sync(). */
1525 	mutex_exit(&zilog->zl_lock);
1526 
1527 	mutex_enter(&zilog->zl_lwb_io_lock);
1528 	ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0);
1529 	zilog->zl_lwb_inflight[txg & TXG_MASK]--;
1530 	if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0)
1531 		cv_broadcast(&zilog->zl_lwb_io_cv);
1532 	mutex_exit(&zilog->zl_lwb_io_lock);
1533 }
1534 
1535 /*
1536  * Wait for the completion of all issued write/flush of that txg provided.
1537  * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1538  */
1539 static void
1540 zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg)
1541 {
1542 	ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa));
1543 
1544 	mutex_enter(&zilog->zl_lwb_io_lock);
1545 	while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0)
1546 		cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock);
1547 	mutex_exit(&zilog->zl_lwb_io_lock);
1548 
1549 #ifdef ZFS_DEBUG
1550 	mutex_enter(&zilog->zl_lock);
1551 	mutex_enter(&zilog->zl_lwb_io_lock);
1552 	lwb_t *lwb = list_head(&zilog->zl_lwb_list);
1553 	while (lwb != NULL) {
1554 		if (lwb->lwb_issued_txg <= txg) {
1555 			ASSERT(lwb->lwb_state != LWB_STATE_ISSUED);
1556 			ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE);
1557 			IMPLY(lwb->lwb_issued_txg > 0,
1558 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
1559 		}
1560 		IMPLY(lwb->lwb_state == LWB_STATE_WRITE_DONE ||
1561 		    lwb->lwb_state == LWB_STATE_FLUSH_DONE,
1562 		    lwb->lwb_buf == NULL);
1563 		lwb = list_next(&zilog->zl_lwb_list, lwb);
1564 	}
1565 	mutex_exit(&zilog->zl_lwb_io_lock);
1566 	mutex_exit(&zilog->zl_lock);
1567 #endif
1568 }
1569 
1570 /*
1571  * This is called when an lwb's write zio completes. The callback's purpose is
1572  * to issue the flush commands for the vdevs in the lwb's lwb_vdev_tree. The
1573  * tree will contain the vdevs involved in writing out this specific lwb's
1574  * data, and in the case that cache flushes have been deferred, vdevs involved
1575  * in writing the data for previous lwbs. The writes corresponding to all the
1576  * vdevs in the lwb_vdev_tree will have completed by the time this is called,
1577  * due to the zio dependencies configured in zil_lwb_set_zio_dependency(),
1578  * which takes deferred flushes into account. The lwb will be "done" once
1579  * zil_lwb_flush_vdevs_done() is called, which occurs in the zio completion
1580  * callback for the lwb's root zio.
1581  */
1582 static void
1583 zil_lwb_write_done(zio_t *zio)
1584 {
1585 	lwb_t *lwb = zio->io_private;
1586 	spa_t *spa = zio->io_spa;
1587 	zilog_t *zilog = lwb->lwb_zilog;
1588 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1589 	void *cookie = NULL;
1590 	zil_vdev_node_t *zv;
1591 	lwb_t *nlwb;
1592 
1593 	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1594 
1595 	abd_free(zio->io_abd);
1596 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1597 	lwb->lwb_buf = NULL;
1598 
1599 	mutex_enter(&zilog->zl_lock);
1600 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1601 	lwb->lwb_state = LWB_STATE_WRITE_DONE;
1602 	lwb->lwb_child_zio = NULL;
1603 	lwb->lwb_write_zio = NULL;
1604 
1605 	/*
1606 	 * If nlwb is not yet issued, zil_lwb_set_zio_dependency() is not
1607 	 * called for it yet, and when it will be, it won't be able to make
1608 	 * its write ZIO a parent this ZIO.  In such case we can not defer
1609 	 * our flushes or below may be a race between the done callbacks.
1610 	 */
1611 	nlwb = list_next(&zilog->zl_lwb_list, lwb);
1612 	if (nlwb && nlwb->lwb_state != LWB_STATE_ISSUED)
1613 		nlwb = NULL;
1614 	mutex_exit(&zilog->zl_lock);
1615 
1616 	if (avl_numnodes(t) == 0)
1617 		return;
1618 
1619 	/*
1620 	 * If there was an IO error, we're not going to call zio_flush()
1621 	 * on these vdevs, so we simply empty the tree and free the
1622 	 * nodes. We avoid calling zio_flush() since there isn't any
1623 	 * good reason for doing so, after the lwb block failed to be
1624 	 * written out.
1625 	 *
1626 	 * Additionally, we don't perform any further error handling at
1627 	 * this point (e.g. setting "zcw_zio_error" appropriately), as
1628 	 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1629 	 * we expect any error seen here, to have been propagated to
1630 	 * that function).
1631 	 */
1632 	if (zio->io_error != 0) {
1633 		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1634 			kmem_free(zv, sizeof (*zv));
1635 		return;
1636 	}
1637 
1638 	/*
1639 	 * If this lwb does not have any threads waiting for it to complete, we
1640 	 * want to defer issuing the flush command to the vdevs written to by
1641 	 * "this" lwb, and instead rely on the "next" lwb to handle the flush
1642 	 * command for those vdevs. Thus, we merge the vdev tree of "this" lwb
1643 	 * with the vdev tree of the "next" lwb in the list, and assume the
1644 	 * "next" lwb will handle flushing the vdevs (or deferring the flush(s)
1645 	 * again).
1646 	 *
1647 	 * This is a useful performance optimization, especially for workloads
1648 	 * with lots of async write activity and few sync write and/or fsync
1649 	 * activity, as it has the potential to coalesce multiple flush
1650 	 * commands to a vdev into one.
1651 	 */
1652 	if (list_is_empty(&lwb->lwb_waiters) && nlwb != NULL) {
1653 		zil_lwb_flush_defer(lwb, nlwb);
1654 		ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1655 		return;
1656 	}
1657 
1658 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1659 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1660 		if (vd != NULL) {
1661 			/*
1662 			 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1663 			 * always used within "zio_flush". This means,
1664 			 * any errors when flushing the vdev(s), will
1665 			 * (unfortunately) not be handled correctly,
1666 			 * since these "zio_flush" errors will not be
1667 			 * propagated up to "zil_lwb_flush_vdevs_done".
1668 			 */
1669 			zio_flush(lwb->lwb_root_zio, vd);
1670 		}
1671 		kmem_free(zv, sizeof (*zv));
1672 	}
1673 }
1674 
1675 /*
1676  * Build the zio dependency chain, which is used to preserve the ordering of
1677  * lwb completions that is required by the semantics of the ZIL. Each new lwb
1678  * zio becomes a parent of the previous lwb zio, such that the new lwb's zio
1679  * cannot complete until the previous lwb's zio completes.
1680  *
1681  * This is required by the semantics of zil_commit(): the commit waiters
1682  * attached to the lwbs will be woken in the lwb zio's completion callback,
1683  * so this zio dependency graph ensures the waiters are woken in the correct
1684  * order (the same order the lwbs were created).
1685  */
1686 static void
1687 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1688 {
1689 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
1690 
1691 	lwb_t *prev_lwb = list_prev(&zilog->zl_lwb_list, lwb);
1692 	if (prev_lwb == NULL ||
1693 	    prev_lwb->lwb_state == LWB_STATE_FLUSH_DONE)
1694 		return;
1695 
1696 	/*
1697 	 * If the previous lwb's write hasn't already completed, we also want
1698 	 * to order the completion of the lwb write zios (above, we only order
1699 	 * the completion of the lwb root zios). This is required because of
1700 	 * how we can defer the flush commands for any lwb without waiters.
1701 	 *
1702 	 * When the flush commands are deferred, the previous lwb will rely on
1703 	 * this lwb to flush the vdevs written to by that previous lwb. Thus,
1704 	 * we need to ensure this lwb doesn't issue the flush until after the
1705 	 * previous lwb's write completes. We ensure this ordering by setting
1706 	 * the zio parent/child relationship here.
1707 	 *
1708 	 * Without this relationship on the lwb's write zio, it's possible for
1709 	 * this lwb's write to complete prior to the previous lwb's write
1710 	 * completing; and thus, the vdevs for the previous lwb would be
1711 	 * flushed prior to that lwb's data being written to those vdevs (the
1712 	 * vdevs are flushed in the lwb write zio's completion handler,
1713 	 * zil_lwb_write_done()).
1714 	 */
1715 	if (prev_lwb->lwb_state == LWB_STATE_ISSUED) {
1716 		ASSERT3P(prev_lwb->lwb_write_zio, !=, NULL);
1717 		if (list_is_empty(&prev_lwb->lwb_waiters)) {
1718 			zio_add_child(lwb->lwb_write_zio,
1719 			    prev_lwb->lwb_write_zio);
1720 		}
1721 	} else {
1722 		ASSERT3S(prev_lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1723 	}
1724 
1725 	ASSERT3P(prev_lwb->lwb_root_zio, !=, NULL);
1726 	zio_add_child(lwb->lwb_root_zio, prev_lwb->lwb_root_zio);
1727 }
1728 
1729 
1730 /*
1731  * This function's purpose is to "open" an lwb such that it is ready to
1732  * accept new itxs being committed to it. This function is idempotent; if
1733  * the passed in lwb has already been opened, it is essentially a no-op.
1734  */
1735 static void
1736 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1737 {
1738 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1739 
1740 	if (lwb->lwb_state != LWB_STATE_NEW) {
1741 		ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1742 		return;
1743 	}
1744 
1745 	mutex_enter(&zilog->zl_lock);
1746 	lwb->lwb_state = LWB_STATE_OPENED;
1747 	zilog->zl_last_lwb_opened = lwb;
1748 	mutex_exit(&zilog->zl_lock);
1749 }
1750 
1751 /*
1752  * Maximum block size used by the ZIL.  This is picked up when the ZIL is
1753  * initialized.  Otherwise this should not be used directly; see
1754  * zl_max_block_size instead.
1755  */
1756 static uint_t zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1757 
1758 /*
1759  * Plan splitting of the provided burst size between several blocks.
1760  */
1761 static uint_t
1762 zil_lwb_plan(zilog_t *zilog, uint64_t size, uint_t *minsize)
1763 {
1764 	uint_t md = zilog->zl_max_block_size - sizeof (zil_chain_t);
1765 
1766 	if (size <= md) {
1767 		/*
1768 		 * Small bursts are written as-is in one block.
1769 		 */
1770 		*minsize = size;
1771 		return (size);
1772 	} else if (size > 8 * md) {
1773 		/*
1774 		 * Big bursts use maximum blocks.  The first block size
1775 		 * is hard to predict, but it does not really matter.
1776 		 */
1777 		*minsize = 0;
1778 		return (md);
1779 	}
1780 
1781 	/*
1782 	 * Medium bursts try to divide evenly to better utilize several SLOG
1783 	 * VDEVs.  The first block size we predict assuming the worst case of
1784 	 * maxing out others.  Fall back to using maximum blocks if due to
1785 	 * large records or wasted space we can not predict anything better.
1786 	 */
1787 	uint_t s = size;
1788 	uint_t n = DIV_ROUND_UP(s, md - sizeof (lr_write_t));
1789 	uint_t chunk = DIV_ROUND_UP(s, n);
1790 	uint_t waste = zil_max_waste_space(zilog);
1791 	waste = MAX(waste, zilog->zl_cur_max);
1792 	if (chunk <= md - waste) {
1793 		*minsize = MAX(s - (md - waste) * (n - 1), waste);
1794 		return (chunk);
1795 	} else {
1796 		*minsize = 0;
1797 		return (md);
1798 	}
1799 }
1800 
1801 /*
1802  * Try to predict next block size based on previous history.  Make prediction
1803  * sufficient for 7 of 8 previous bursts.  Don't try to save if the saving is
1804  * less then 50%, extra writes may cost more, but we don't want single spike
1805  * to badly affect our predictions.
1806  */
1807 static uint_t
1808 zil_lwb_predict(zilog_t *zilog)
1809 {
1810 	uint_t m, o;
1811 
1812 	/* If we are in the middle of a burst, take it into account also. */
1813 	if (zilog->zl_cur_size > 0) {
1814 		o = zil_lwb_plan(zilog, zilog->zl_cur_size, &m);
1815 	} else {
1816 		o = UINT_MAX;
1817 		m = 0;
1818 	}
1819 
1820 	/* Find minimum optimal size.  We don't need to go below that. */
1821 	for (int i = 0; i < ZIL_BURSTS; i++)
1822 		o = MIN(o, zilog->zl_prev_opt[i]);
1823 
1824 	/* Find two biggest minimal first block sizes above the optimal. */
1825 	uint_t m1 = MAX(m, o), m2 = o;
1826 	for (int i = 0; i < ZIL_BURSTS; i++) {
1827 		m = zilog->zl_prev_min[i];
1828 		if (m >= m1) {
1829 			m2 = m1;
1830 			m1 = m;
1831 		} else if (m > m2) {
1832 			m2 = m;
1833 		}
1834 	}
1835 
1836 	/*
1837 	 * If second minimum size gives 50% saving -- use it.  It may cost us
1838 	 * one additional write later, but the space saving is just too big.
1839 	 */
1840 	return ((m1 < m2 * 2) ? m1 : m2);
1841 }
1842 
1843 /*
1844  * Close the log block for being issued and allocate the next one.
1845  * Has to be called under zl_issuer_lock to chain more lwbs.
1846  */
1847 static lwb_t *
1848 zil_lwb_write_close(zilog_t *zilog, lwb_t *lwb, lwb_state_t state)
1849 {
1850 	uint64_t blksz, plan, plan2;
1851 
1852 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1853 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1854 	lwb->lwb_state = LWB_STATE_CLOSED;
1855 
1856 	/*
1857 	 * If there was an allocation failure then returned NULL will trigger
1858 	 * zil_commit_writer_stall() at the caller.  This is inherently racy,
1859 	 * since allocation may not have happened yet.
1860 	 */
1861 	if (lwb->lwb_error != 0)
1862 		return (NULL);
1863 
1864 	/*
1865 	 * Log blocks are pre-allocated.  Here we select the size of the next
1866 	 * block, based on what's left of this burst and the previous history.
1867 	 * While we try to only write used part of the block, we can't just
1868 	 * always allocate the maximum block size because we can exhaust all
1869 	 * available pool log space, so we try to be reasonable.
1870 	 */
1871 	if (zilog->zl_cur_left > 0) {
1872 		/*
1873 		 * We are in the middle of a burst and know how much is left.
1874 		 * But if workload is multi-threaded there may be more soon.
1875 		 * Try to predict what can it be and plan for the worst case.
1876 		 */
1877 		uint_t m;
1878 		plan = zil_lwb_plan(zilog, zilog->zl_cur_left, &m);
1879 		if (zilog->zl_parallel) {
1880 			plan2 = zil_lwb_plan(zilog, zilog->zl_cur_left +
1881 			    zil_lwb_predict(zilog), &m);
1882 			if (plan < plan2)
1883 				plan = plan2;
1884 		}
1885 	} else {
1886 		/*
1887 		 * The previous burst is done and we can only predict what
1888 		 * will come next.
1889 		 */
1890 		plan = zil_lwb_predict(zilog);
1891 	}
1892 	blksz = plan + sizeof (zil_chain_t);
1893 	blksz = P2ROUNDUP_TYPED(blksz, ZIL_MIN_BLKSZ, uint64_t);
1894 	blksz = MIN(blksz, zilog->zl_max_block_size);
1895 	DTRACE_PROBE3(zil__block__size, zilog_t *, zilog, uint64_t, blksz,
1896 	    uint64_t, plan);
1897 
1898 	return (zil_alloc_lwb(zilog, blksz, NULL, 0, 0, state));
1899 }
1900 
1901 /*
1902  * Finalize previously closed block and issue the write zio.
1903  */
1904 static int
1905 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1906 {
1907 	spa_t *spa = zilog->zl_spa;
1908 	zil_chain_t *zilc;
1909 	boolean_t slog;
1910 	zbookmark_phys_t zb;
1911 	zio_priority_t prio;
1912 	int error;
1913 
1914 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED);
1915 
1916 	/* Actually fill the lwb with the data. */
1917 	for (itx_t *itx = list_head(&lwb->lwb_itxs); itx;
1918 	    itx = list_next(&lwb->lwb_itxs, itx)) {
1919 		error = zil_lwb_commit(zilog, lwb, itx);
1920 		if (error != 0) {
1921 			ASSERT3U(error, ==, ESHUTDOWN);
1922 			return (error);
1923 		}
1924 	}
1925 	lwb->lwb_nused = lwb->lwb_nfilled;
1926 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_nmax);
1927 
1928 	lwb->lwb_root_zio = zio_root(spa, zil_lwb_flush_vdevs_done, lwb,
1929 	    ZIO_FLAG_CANFAIL);
1930 
1931 	/*
1932 	 * The lwb is now ready to be issued, but it can be only if it already
1933 	 * got its block pointer allocated or the allocation has failed.
1934 	 * Otherwise leave it as-is, relying on some other thread to issue it
1935 	 * after allocating its block pointer via calling zil_lwb_write_issue()
1936 	 * for the previous lwb(s) in the chain.
1937 	 */
1938 	mutex_enter(&zilog->zl_lock);
1939 	lwb->lwb_state = LWB_STATE_READY;
1940 	if (BP_IS_HOLE(&lwb->lwb_blk) && lwb->lwb_error == 0) {
1941 		mutex_exit(&zilog->zl_lock);
1942 		return (0);
1943 	}
1944 	mutex_exit(&zilog->zl_lock);
1945 
1946 next_lwb:
1947 	if (lwb->lwb_slim)
1948 		zilc = (zil_chain_t *)lwb->lwb_buf;
1949 	else
1950 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_nmax);
1951 	int wsz = lwb->lwb_sz;
1952 	if (lwb->lwb_error == 0) {
1953 		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, lwb->lwb_sz);
1954 		if (!lwb->lwb_slog || zilog->zl_cur_size <= zil_slog_bulk)
1955 			prio = ZIO_PRIORITY_SYNC_WRITE;
1956 		else
1957 			prio = ZIO_PRIORITY_ASYNC_WRITE;
1958 		SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1959 		    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1960 		    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1961 		lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, spa, 0,
1962 		    &lwb->lwb_blk, lwb_abd, lwb->lwb_sz, zil_lwb_write_done,
1963 		    lwb, prio, ZIO_FLAG_CANFAIL, &zb);
1964 		zil_lwb_add_block(lwb, &lwb->lwb_blk);
1965 
1966 		if (lwb->lwb_slim) {
1967 			/* For Slim ZIL only write what is used. */
1968 			wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ,
1969 			    int);
1970 			ASSERT3S(wsz, <=, lwb->lwb_sz);
1971 			zio_shrink(lwb->lwb_write_zio, wsz);
1972 			wsz = lwb->lwb_write_zio->io_size;
1973 		}
1974 		memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused);
1975 		zilc->zc_pad = 0;
1976 		zilc->zc_nused = lwb->lwb_nused;
1977 		zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1978 	} else {
1979 		/*
1980 		 * We can't write the lwb if there was an allocation failure,
1981 		 * so create a null zio instead just to maintain dependencies.
1982 		 */
1983 		lwb->lwb_write_zio = zio_null(lwb->lwb_root_zio, spa, NULL,
1984 		    zil_lwb_write_done, lwb, ZIO_FLAG_CANFAIL);
1985 		lwb->lwb_write_zio->io_error = lwb->lwb_error;
1986 	}
1987 	if (lwb->lwb_child_zio)
1988 		zio_add_child(lwb->lwb_write_zio, lwb->lwb_child_zio);
1989 
1990 	/*
1991 	 * Open transaction to allocate the next block pointer.
1992 	 */
1993 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1994 	VERIFY0(dmu_tx_assign(tx,
1995 	    DMU_TX_WAIT | DMU_TX_NOTHROTTLE | DMU_TX_SUSPEND));
1996 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1997 	uint64_t txg = dmu_tx_get_txg(tx);
1998 
1999 	/*
2000 	 * Allocate next the block pointer unless we are already in error.
2001 	 */
2002 	lwb_t *nlwb = list_next(&zilog->zl_lwb_list, lwb);
2003 	blkptr_t *bp = &zilc->zc_next_blk;
2004 	BP_ZERO(bp);
2005 	error = lwb->lwb_error;
2006 	if (error == 0) {
2007 		error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, nlwb->lwb_sz,
2008 		    &slog);
2009 	}
2010 	if (error == 0) {
2011 		ASSERT3U(BP_GET_BIRTH(bp), ==, txg);
2012 		BP_SET_CHECKSUM(bp, nlwb->lwb_slim ? ZIO_CHECKSUM_ZILOG2 :
2013 		    ZIO_CHECKSUM_ZILOG);
2014 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
2015 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
2016 	}
2017 
2018 	/*
2019 	 * Reduce TXG open time by incrementing inflight counter and committing
2020 	 * the transaciton.  zil_sync() will wait for it to return to zero.
2021 	 */
2022 	mutex_enter(&zilog->zl_lwb_io_lock);
2023 	lwb->lwb_issued_txg = txg;
2024 	zilog->zl_lwb_inflight[txg & TXG_MASK]++;
2025 	zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg);
2026 	mutex_exit(&zilog->zl_lwb_io_lock);
2027 	dmu_tx_commit(tx);
2028 
2029 	spa_config_enter(spa, SCL_STATE, lwb, RW_READER);
2030 
2031 	/*
2032 	 * We've completed all potentially blocking operations.  Update the
2033 	 * nlwb and allow it proceed without possible lock order reversals.
2034 	 */
2035 	mutex_enter(&zilog->zl_lock);
2036 	zil_lwb_set_zio_dependency(zilog, lwb);
2037 	lwb->lwb_state = LWB_STATE_ISSUED;
2038 
2039 	if (nlwb) {
2040 		nlwb->lwb_blk = *bp;
2041 		nlwb->lwb_error = error;
2042 		nlwb->lwb_slog = slog;
2043 		nlwb->lwb_alloc_txg = txg;
2044 		if (nlwb->lwb_state != LWB_STATE_READY)
2045 			nlwb = NULL;
2046 	}
2047 	mutex_exit(&zilog->zl_lock);
2048 
2049 	if (lwb->lwb_slog) {
2050 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_slog_count);
2051 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_bytes,
2052 		    lwb->lwb_nused);
2053 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_write,
2054 		    wsz);
2055 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_alloc,
2056 		    BP_GET_LSIZE(&lwb->lwb_blk));
2057 	} else {
2058 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_normal_count);
2059 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_bytes,
2060 		    lwb->lwb_nused);
2061 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_write,
2062 		    wsz);
2063 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_alloc,
2064 		    BP_GET_LSIZE(&lwb->lwb_blk));
2065 	}
2066 	lwb->lwb_issued_timestamp = gethrtime();
2067 	if (lwb->lwb_child_zio)
2068 		zio_nowait(lwb->lwb_child_zio);
2069 	zio_nowait(lwb->lwb_write_zio);
2070 	zio_nowait(lwb->lwb_root_zio);
2071 
2072 	/*
2073 	 * If nlwb was ready when we gave it the block pointer,
2074 	 * it is on us to issue it and possibly following ones.
2075 	 */
2076 	lwb = nlwb;
2077 	if (lwb)
2078 		goto next_lwb;
2079 
2080 	return (0);
2081 }
2082 
2083 /*
2084  * Maximum amount of data that can be put into single log block.
2085  */
2086 uint64_t
2087 zil_max_log_data(zilog_t *zilog, size_t hdrsize)
2088 {
2089 	return (zilog->zl_max_block_size - sizeof (zil_chain_t) - hdrsize);
2090 }
2091 
2092 /*
2093  * Maximum amount of log space we agree to waste to reduce number of
2094  * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~6%).
2095  */
2096 static inline uint64_t
2097 zil_max_waste_space(zilog_t *zilog)
2098 {
2099 	return (zil_max_log_data(zilog, sizeof (lr_write_t)) / 16);
2100 }
2101 
2102 /*
2103  * Maximum amount of write data for WR_COPIED.  For correctness, consumers
2104  * must fall back to WR_NEED_COPY if we can't fit the entire record into one
2105  * maximum sized log block, because each WR_COPIED record must fit in a
2106  * single log block.  Below that it is a tradeoff of additional memory copy
2107  * and possibly worse log space efficiency vs additional range lock/unlock.
2108  */
2109 static uint_t zil_maxcopied = 7680;
2110 
2111 /*
2112  * Largest write size to store the data directly into ZIL.
2113  */
2114 uint_t zfs_immediate_write_sz = 32768;
2115 
2116 /*
2117  * When enabled and blocks go to normal vdev, treat special vdevs as SLOG,
2118  * writing data to ZIL (WR_COPIED/WR_NEED_COPY).  Disabling this forces the
2119  * indirect writes (WR_INDIRECT) to preserve special vdev throughput and
2120  * endurance, likely at the cost of normal vdev latency.
2121  */
2122 int zil_special_is_slog = 1;
2123 
2124 uint64_t
2125 zil_max_copied_data(zilog_t *zilog)
2126 {
2127 	uint64_t max_data = zil_max_log_data(zilog, sizeof (lr_write_t));
2128 	return (MIN(max_data, zil_maxcopied));
2129 }
2130 
2131 /*
2132  * Determine the appropriate write state for ZIL transactions based on
2133  * pool configuration, data placement, write size, and logbias settings.
2134  */
2135 itx_wr_state_t
2136 zil_write_state(zilog_t *zilog, uint64_t size, uint32_t blocksize,
2137     boolean_t o_direct, boolean_t commit)
2138 {
2139 	if (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT || o_direct)
2140 		return (WR_INDIRECT);
2141 
2142 	/*
2143 	 * Don't use indirect for too small writes to reduce overhead.
2144 	 * Don't use indirect if written less than a half of a block if
2145 	 * we are going to commit it immediately, since next write might
2146 	 * rewrite the same block again, causing inflation.  If commit
2147 	 * is not planned, then next writes might coalesce, and so the
2148 	 * indirect may be perfect.
2149 	 */
2150 	boolean_t indirect = (size >= zfs_immediate_write_sz &&
2151 	    (size >= blocksize / 2 || !commit));
2152 
2153 	if (spa_has_slogs(zilog->zl_spa)) {
2154 		/* Dedicated slogs: never use indirect */
2155 		indirect = B_FALSE;
2156 	} else if (spa_has_special(zilog->zl_spa)) {
2157 		/* Special vdevs: only when beneficial */
2158 		boolean_t on_special = (blocksize <=
2159 		    zilog->zl_os->os_zpl_special_smallblock);
2160 		indirect &= (on_special || !zil_special_is_slog);
2161 	}
2162 
2163 	if (indirect)
2164 		return (WR_INDIRECT);
2165 	else if (commit)
2166 		return (WR_COPIED);
2167 	else
2168 		return (WR_NEED_COPY);
2169 }
2170 
2171 static uint64_t
2172 zil_itx_record_size(itx_t *itx)
2173 {
2174 	lr_t *lr = &itx->itx_lr;
2175 
2176 	if (lr->lrc_txtype == TX_COMMIT)
2177 		return (0);
2178 	ASSERT3U(lr->lrc_reclen, >=, sizeof (lr_t));
2179 	return (lr->lrc_reclen);
2180 }
2181 
2182 static uint64_t
2183 zil_itx_data_size(itx_t *itx)
2184 {
2185 	lr_t *lr = &itx->itx_lr;
2186 	lr_write_t *lrw = (lr_write_t *)lr;
2187 
2188 	if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
2189 		ASSERT3U(lr->lrc_reclen, ==, sizeof (lr_write_t));
2190 		return (P2ROUNDUP_TYPED(lrw->lr_length, sizeof (uint64_t),
2191 		    uint64_t));
2192 	}
2193 	return (0);
2194 }
2195 
2196 static uint64_t
2197 zil_itx_full_size(itx_t *itx)
2198 {
2199 	lr_t *lr = &itx->itx_lr;
2200 
2201 	if (lr->lrc_txtype == TX_COMMIT)
2202 		return (0);
2203 	ASSERT3U(lr->lrc_reclen, >=, sizeof (lr_t));
2204 	return (lr->lrc_reclen + zil_itx_data_size(itx));
2205 }
2206 
2207 /*
2208  * Estimate space needed in the lwb for the itx.  Allocate more lwbs or
2209  * split the itx as needed, but don't touch the actual transaction data.
2210  * Has to be called under zl_issuer_lock to call zil_lwb_write_close()
2211  * to chain more lwbs.
2212  */
2213 static lwb_t *
2214 zil_lwb_assign(zilog_t *zilog, lwb_t *lwb, itx_t *itx, list_t *ilwbs)
2215 {
2216 	itx_t *citx;
2217 	lr_t *lr, *clr;
2218 	lr_write_t *lrw;
2219 	uint64_t dlen, dnow, lwb_sp, reclen, max_log_data;
2220 
2221 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2222 	ASSERT3P(lwb, !=, NULL);
2223 	ASSERT3P(lwb->lwb_buf, !=, NULL);
2224 
2225 	zil_lwb_write_open(zilog, lwb);
2226 
2227 	lr = &itx->itx_lr;
2228 	lrw = (lr_write_t *)lr;
2229 
2230 	/*
2231 	 * A commit itx doesn't represent any on-disk state; instead
2232 	 * it's simply used as a place holder on the commit list, and
2233 	 * provides a mechanism for attaching a "commit waiter" onto the
2234 	 * correct lwb (such that the waiter can be signalled upon
2235 	 * completion of that lwb). Thus, we don't process this itx's
2236 	 * log record if it's a commit itx (these itx's don't have log
2237 	 * records), and instead link the itx's waiter onto the lwb's
2238 	 * list of waiters.
2239 	 *
2240 	 * For more details, see the comment above zil_commit().
2241 	 */
2242 	if (lr->lrc_txtype == TX_COMMIT) {
2243 		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
2244 		list_insert_tail(&lwb->lwb_itxs, itx);
2245 		return (lwb);
2246 	}
2247 
2248 	reclen = lr->lrc_reclen;
2249 	ASSERT3U(reclen, >=, sizeof (lr_t));
2250 	ASSERT3U(reclen, <=, zil_max_log_data(zilog, 0));
2251 	dlen = zil_itx_data_size(itx);
2252 
2253 cont:
2254 	/*
2255 	 * If this record won't fit in the current log block, start a new one.
2256 	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
2257 	 */
2258 	lwb_sp = lwb->lwb_nmax - lwb->lwb_nused;
2259 	max_log_data = zil_max_log_data(zilog, sizeof (lr_write_t));
2260 	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
2261 	    lwb_sp < zil_max_waste_space(zilog) &&
2262 	    (dlen % max_log_data == 0 ||
2263 	    lwb_sp < reclen + dlen % max_log_data))) {
2264 		list_insert_tail(ilwbs, lwb);
2265 		lwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_OPENED);
2266 		if (lwb == NULL)
2267 			return (NULL);
2268 		lwb_sp = lwb->lwb_nmax - lwb->lwb_nused;
2269 	}
2270 
2271 	/*
2272 	 * There must be enough space in the log block to hold reclen.
2273 	 * For WR_COPIED, we need to fit the whole record in one block,
2274 	 * and reclen is the write record header size + the data size.
2275 	 * For WR_NEED_COPY, we can create multiple records, splitting
2276 	 * the data into multiple blocks, so we only need to fit one
2277 	 * word of data per block; in this case reclen is just the header
2278 	 * size (no data).
2279 	 */
2280 	ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
2281 
2282 	dnow = MIN(dlen, lwb_sp - reclen);
2283 	if (dlen > dnow) {
2284 		ASSERT3U(lr->lrc_txtype, ==, TX_WRITE);
2285 		ASSERT3U(itx->itx_wr_state, ==, WR_NEED_COPY);
2286 		citx = zil_itx_clone(itx);
2287 		clr = &citx->itx_lr;
2288 		lr_write_t *clrw = (lr_write_t *)clr;
2289 		clrw->lr_length = dnow;
2290 		lrw->lr_offset += dnow;
2291 		lrw->lr_length -= dnow;
2292 		zilog->zl_cur_left -= dnow;
2293 	} else {
2294 		citx = itx;
2295 		clr = lr;
2296 	}
2297 
2298 	/*
2299 	 * We're actually making an entry, so update lrc_seq to be the
2300 	 * log record sequence number.  Note that this is generally not
2301 	 * equal to the itx sequence number because not all transactions
2302 	 * are synchronous, and sometimes spa_sync() gets there first.
2303 	 */
2304 	clr->lrc_seq = ++zilog->zl_lr_seq;
2305 
2306 	lwb->lwb_nused += reclen + dnow;
2307 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_nmax);
2308 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
2309 
2310 	zil_lwb_add_txg(lwb, lr->lrc_txg);
2311 	list_insert_tail(&lwb->lwb_itxs, citx);
2312 
2313 	dlen -= dnow;
2314 	if (dlen > 0)
2315 		goto cont;
2316 
2317 	if (lr->lrc_txtype == TX_WRITE &&
2318 	    lr->lrc_txg > spa_freeze_txg(zilog->zl_spa))
2319 		txg_wait_synced(zilog->zl_dmu_pool, lr->lrc_txg);
2320 
2321 	return (lwb);
2322 }
2323 
2324 static void zil_crash(zilog_t *zilog);
2325 
2326 /*
2327  * Fill the actual transaction data into the lwb, following zil_lwb_assign().
2328  * Does not require locking.
2329  */
2330 static int
2331 zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx)
2332 {
2333 	lr_t *lr, *lrb;
2334 	lr_write_t *lrw, *lrwb;
2335 	char *lr_buf;
2336 	uint64_t dlen, reclen;
2337 
2338 	lr = &itx->itx_lr;
2339 	lrw = (lr_write_t *)lr;
2340 
2341 	if (lr->lrc_txtype == TX_COMMIT)
2342 		return (0);
2343 
2344 	reclen = lr->lrc_reclen;
2345 	dlen = zil_itx_data_size(itx);
2346 	ASSERT3U(reclen + dlen, <=, lwb->lwb_nused - lwb->lwb_nfilled);
2347 
2348 	lr_buf = lwb->lwb_buf + lwb->lwb_nfilled;
2349 	memcpy(lr_buf, lr, reclen);
2350 	lrb = (lr_t *)lr_buf;		/* Like lr, but inside lwb. */
2351 	lrwb = (lr_write_t *)lrb;	/* Like lrw, but inside lwb. */
2352 
2353 	ZIL_STAT_BUMP(zilog, zil_itx_count);
2354 
2355 	/*
2356 	 * If it's a write, fetch the data or get its blkptr as appropriate.
2357 	 */
2358 	if (lr->lrc_txtype == TX_WRITE) {
2359 		if (itx->itx_wr_state == WR_COPIED) {
2360 			ZIL_STAT_BUMP(zilog, zil_itx_copied_count);
2361 			ZIL_STAT_INCR(zilog, zil_itx_copied_bytes,
2362 			    lrw->lr_length);
2363 		} else {
2364 			char *dbuf;
2365 			int error;
2366 
2367 			if (itx->itx_wr_state == WR_NEED_COPY) {
2368 				dbuf = lr_buf + reclen;
2369 				lrb->lrc_reclen += dlen;
2370 				ZIL_STAT_BUMP(zilog, zil_itx_needcopy_count);
2371 				ZIL_STAT_INCR(zilog, zil_itx_needcopy_bytes,
2372 				    dlen);
2373 			} else {
2374 				ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
2375 				dbuf = NULL;
2376 				ZIL_STAT_BUMP(zilog, zil_itx_indirect_count);
2377 				ZIL_STAT_INCR(zilog, zil_itx_indirect_bytes,
2378 				    lrw->lr_length);
2379 				if (lwb->lwb_child_zio == NULL) {
2380 					lwb->lwb_child_zio = zio_null(NULL,
2381 					    zilog->zl_spa, NULL, NULL, NULL,
2382 					    ZIO_FLAG_CANFAIL);
2383 				}
2384 			}
2385 
2386 			/*
2387 			 * The "lwb_child_zio" we pass in will become a child of
2388 			 * "lwb_write_zio", when one is created, so one will be
2389 			 * a parent of any zio's created by the "zl_get_data".
2390 			 * This way "lwb_write_zio" will first wait for children
2391 			 * block pointers before own writing, and then for their
2392 			 * writing completion before the vdev cache flushing.
2393 			 */
2394 			error = zilog->zl_get_data(itx->itx_private,
2395 			    itx->itx_gen, lrwb, dbuf, lwb,
2396 			    lwb->lwb_child_zio);
2397 			if (dbuf != NULL && error == 0) {
2398 				/* Zero any padding bytes in the last block. */
2399 				memset((char *)dbuf + lrwb->lr_length, 0,
2400 				    dlen - lrwb->lr_length);
2401 			}
2402 
2403 			/*
2404 			 * Typically, the only return values we should see from
2405 			 * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or
2406 			 *  EALREADY. However, it is also possible to see other
2407 			 *  error values such as ENOSPC or EINVAL from
2408 			 *  dmu_read() -> dnode_hold() -> dnode_hold_impl() or
2409 			 *  ENXIO as well as a multitude of others from the
2410 			 *  block layer through dmu_buf_hold() -> dbuf_read()
2411 			 *  -> zio_wait(), as well as through dmu_read() ->
2412 			 *  dnode_hold() -> dnode_hold_impl() -> dbuf_read() ->
2413 			 *  zio_wait(). When these errors happen, we can assume
2414 			 *  that neither an immediate write nor an indirect
2415 			 *  write occurred, so we need to fall back to
2416 			 *  txg_wait_synced(). This is unusual, so we print to
2417 			 *  dmesg whenever one of these errors occurs.
2418 			 */
2419 			switch (error) {
2420 			case 0:
2421 				break;
2422 			default:
2423 				cmn_err(CE_WARN, "zil_lwb_commit() received "
2424 				    "unexpected error %d from ->zl_get_data()"
2425 				    ". Falling back to txg_wait_synced().",
2426 				    error);
2427 				zfs_fallthrough;
2428 			case EIO: {
2429 				int error = txg_wait_synced_flags(
2430 				    zilog->zl_dmu_pool,
2431 				    lr->lrc_txg, TXG_WAIT_SUSPEND);
2432 				if (error != 0) {
2433 					ASSERT3U(error, ==, ESHUTDOWN);
2434 					/*
2435 					 * zil_lwb_commit() is called from a
2436 					 * loop over a list of itxs at the
2437 					 * top of zil_lwb_write_issue(), which
2438 					 * itself is called from a loop over a
2439 					 * list of lwbs in various places.
2440 					 * zil_crash() will free those itxs
2441 					 * and sometimes the lwbs, so they
2442 					 * are invalid when zil_crash() returns.
2443 					 * Callers must pretty much abort
2444 					 * immediately.
2445 					 */
2446 					zil_crash(zilog);
2447 					return (error);
2448 				}
2449 				zfs_fallthrough;
2450 			}
2451 			case ENOENT:
2452 				zfs_fallthrough;
2453 			case EEXIST:
2454 				zfs_fallthrough;
2455 			case EALREADY:
2456 				return (0);
2457 			}
2458 		}
2459 	}
2460 
2461 	lwb->lwb_nfilled += reclen + dlen;
2462 	ASSERT3S(lwb->lwb_nfilled, <=, lwb->lwb_nused);
2463 	ASSERT0(P2PHASE(lwb->lwb_nfilled, sizeof (uint64_t)));
2464 
2465 	return (0);
2466 }
2467 
2468 itx_t *
2469 zil_itx_create(uint64_t txtype, size_t olrsize)
2470 {
2471 	size_t itxsize, lrsize;
2472 	itx_t *itx;
2473 
2474 	ASSERT3U(olrsize, >=, sizeof (lr_t));
2475 	lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t);
2476 	ASSERT3U(lrsize, >=, olrsize);
2477 	itxsize = offsetof(itx_t, itx_lr) + lrsize;
2478 
2479 	itx = zio_data_buf_alloc(itxsize);
2480 	itx->itx_lr.lrc_txtype = txtype;
2481 	itx->itx_lr.lrc_reclen = lrsize;
2482 	itx->itx_lr.lrc_seq = 0;	/* defensive */
2483 	memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize);
2484 	itx->itx_sync = B_TRUE;		/* default is synchronous */
2485 	itx->itx_callback = NULL;
2486 	itx->itx_callback_data = NULL;
2487 	itx->itx_size = itxsize;
2488 
2489 	return (itx);
2490 }
2491 
2492 static itx_t *
2493 zil_itx_clone(itx_t *oitx)
2494 {
2495 	ASSERT3U(oitx->itx_size, >=, sizeof (itx_t));
2496 	ASSERT3U(oitx->itx_size, ==,
2497 	    offsetof(itx_t, itx_lr) + oitx->itx_lr.lrc_reclen);
2498 
2499 	itx_t *itx = zio_data_buf_alloc(oitx->itx_size);
2500 	memcpy(itx, oitx, oitx->itx_size);
2501 	itx->itx_callback = NULL;
2502 	itx->itx_callback_data = NULL;
2503 	return (itx);
2504 }
2505 
2506 void
2507 zil_itx_destroy(itx_t *itx, int err)
2508 {
2509 	ASSERT3U(itx->itx_size, >=, sizeof (itx_t));
2510 	ASSERT3U(itx->itx_lr.lrc_reclen, ==,
2511 	    itx->itx_size - offsetof(itx_t, itx_lr));
2512 	IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
2513 	IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2514 
2515 	if (itx->itx_callback != NULL)
2516 		itx->itx_callback(itx->itx_callback_data, err);
2517 
2518 	zio_data_buf_free(itx, itx->itx_size);
2519 }
2520 
2521 /*
2522  * Free up the sync and async itxs. The itxs_t has already been detached
2523  * so no locks are needed.
2524  */
2525 static void
2526 zil_itxg_clean(void *arg)
2527 {
2528 	itx_t *itx;
2529 	list_t *list;
2530 	avl_tree_t *t;
2531 	void *cookie;
2532 	itxs_t *itxs = arg;
2533 	itx_async_node_t *ian;
2534 
2535 	list = &itxs->i_sync_list;
2536 	while ((itx = list_remove_head(list)) != NULL) {
2537 		/*
2538 		 * In the general case, commit itxs will not be found
2539 		 * here, as they'll be committed to an lwb via
2540 		 * zil_lwb_assign(), and free'd in that function. Having
2541 		 * said that, it is still possible for commit itxs to be
2542 		 * found here, due to the following race:
2543 		 *
2544 		 *	- a thread calls zil_commit() which assigns the
2545 		 *	  commit itx to a per-txg i_sync_list
2546 		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
2547 		 *	  while the waiter is still on the i_sync_list
2548 		 *
2549 		 * There's nothing to prevent syncing the txg while the
2550 		 * waiter is on the i_sync_list. This normally doesn't
2551 		 * happen because spa_sync() is slower than zil_commit(),
2552 		 * but if zil_commit() calls txg_wait_synced() (e.g.
2553 		 * because zil_create() or zil_commit_writer_stall() is
2554 		 * called) we will hit this case.
2555 		 */
2556 		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
2557 			zil_commit_waiter_skip(itx->itx_private);
2558 
2559 		zil_itx_destroy(itx, 0);
2560 	}
2561 
2562 	cookie = NULL;
2563 	t = &itxs->i_async_tree;
2564 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2565 		list = &ian->ia_list;
2566 		while ((itx = list_remove_head(list)) != NULL) {
2567 			/* commit itxs should never be on the async lists. */
2568 			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2569 			zil_itx_destroy(itx, 0);
2570 		}
2571 		list_destroy(list);
2572 		kmem_free(ian, sizeof (itx_async_node_t));
2573 	}
2574 	avl_destroy(t);
2575 
2576 	kmem_free(itxs, sizeof (itxs_t));
2577 }
2578 
2579 static int
2580 zil_aitx_compare(const void *x1, const void *x2)
2581 {
2582 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
2583 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
2584 
2585 	return (TREE_CMP(o1, o2));
2586 }
2587 
2588 /*
2589  * Remove all async itx with the given oid.
2590  */
2591 void
2592 zil_remove_async(zilog_t *zilog, uint64_t oid)
2593 {
2594 	uint64_t otxg, txg;
2595 	itx_async_node_t *ian, ian_search;
2596 	avl_tree_t *t;
2597 	avl_index_t where;
2598 	list_t clean_list;
2599 	itx_t *itx;
2600 
2601 	ASSERT(oid != 0);
2602 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
2603 
2604 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2605 		otxg = ZILTEST_TXG;
2606 	else
2607 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2608 
2609 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2610 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2611 
2612 		mutex_enter(&itxg->itxg_lock);
2613 		if (itxg->itxg_txg != txg) {
2614 			mutex_exit(&itxg->itxg_lock);
2615 			continue;
2616 		}
2617 
2618 		/*
2619 		 * Locate the object node and append its list.
2620 		 */
2621 		t = &itxg->itxg_itxs->i_async_tree;
2622 		ian_search.ia_foid = oid;
2623 		ian = avl_find(t, &ian_search, &where);
2624 		if (ian != NULL)
2625 			list_move_tail(&clean_list, &ian->ia_list);
2626 		mutex_exit(&itxg->itxg_lock);
2627 	}
2628 	while ((itx = list_remove_head(&clean_list)) != NULL) {
2629 		/* commit itxs should never be on the async lists. */
2630 		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2631 		zil_itx_destroy(itx, 0);
2632 	}
2633 	list_destroy(&clean_list);
2634 }
2635 
2636 void
2637 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
2638 {
2639 	uint64_t txg;
2640 	itxg_t *itxg;
2641 	itxs_t *itxs, *clean = NULL;
2642 
2643 	/*
2644 	 * Ensure the data of a renamed file is committed before the rename.
2645 	 */
2646 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
2647 		zil_async_to_sync(zilog, itx->itx_oid);
2648 
2649 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
2650 		txg = ZILTEST_TXG;
2651 	else
2652 		txg = dmu_tx_get_txg(tx);
2653 
2654 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
2655 	mutex_enter(&itxg->itxg_lock);
2656 	itxs = itxg->itxg_itxs;
2657 	if (itxg->itxg_txg != txg) {
2658 		if (itxs != NULL) {
2659 			/*
2660 			 * The zil_clean callback hasn't got around to cleaning
2661 			 * this itxg. Save the itxs for release below.
2662 			 * This should be rare.
2663 			 */
2664 			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2665 			    "txg %llu", (u_longlong_t)itxg->itxg_txg);
2666 			clean = itxg->itxg_itxs;
2667 		}
2668 		itxg->itxg_txg = txg;
2669 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
2670 		    KM_SLEEP);
2671 
2672 		list_create(&itxs->i_sync_list, sizeof (itx_t),
2673 		    offsetof(itx_t, itx_node));
2674 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
2675 		    sizeof (itx_async_node_t),
2676 		    offsetof(itx_async_node_t, ia_node));
2677 	}
2678 	if (itx->itx_sync) {
2679 		list_insert_tail(&itxs->i_sync_list, itx);
2680 	} else {
2681 		avl_tree_t *t = &itxs->i_async_tree;
2682 		uint64_t foid =
2683 		    LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
2684 		itx_async_node_t *ian;
2685 		avl_index_t where;
2686 
2687 		ian = avl_find(t, &foid, &where);
2688 		if (ian == NULL) {
2689 			ian = kmem_alloc(sizeof (itx_async_node_t),
2690 			    KM_SLEEP);
2691 			list_create(&ian->ia_list, sizeof (itx_t),
2692 			    offsetof(itx_t, itx_node));
2693 			ian->ia_foid = foid;
2694 			avl_insert(t, ian, where);
2695 		}
2696 		list_insert_tail(&ian->ia_list, itx);
2697 	}
2698 
2699 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
2700 
2701 	/*
2702 	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2703 	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2704 	 * need to be careful to always dirty the ZIL using the "real"
2705 	 * TXG (not itxg_txg) even when the SPA is frozen.
2706 	 */
2707 	zilog_dirty(zilog, dmu_tx_get_txg(tx));
2708 	mutex_exit(&itxg->itxg_lock);
2709 
2710 	/* Release the old itxs now we've dropped the lock */
2711 	if (clean != NULL)
2712 		zil_itxg_clean(clean);
2713 }
2714 
2715 /*
2716  * Post-crash cleanup. This is called from zil_clean() because it needs to
2717  * do cleanup after every txg until the ZIL is restarted, and zilog_dirty()
2718  * can arrange that easily, unlike zil_sync() which is more complicated to
2719  * get a call to without actual dirty data.
2720  */
2721 static void
2722 zil_crash_clean(zilog_t *zilog, uint64_t synced_txg)
2723 {
2724 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
2725 	ASSERT3U(zilog->zl_restart_txg, >, 0);
2726 
2727 	/* Clean up anything on the crash list from earlier txgs */
2728 	lwb_t *lwb;
2729 	while ((lwb = list_head(&zilog->zl_lwb_crash_list)) != NULL) {
2730 		if (lwb->lwb_alloc_txg >= synced_txg ||
2731 		    lwb->lwb_max_txg >= synced_txg) {
2732 			/*
2733 			 * This lwb was allocated or updated on this txg, or
2734 			 * in the future. We stop processing here, to avoid
2735 			 * the strange situation of freeing a ZIL block on
2736 			 * on the same or earlier txg than what it was
2737 			 * allocated for.
2738 			 *
2739 			 * We'll take care of it on the next txg.
2740 			 */
2741 			break;
2742 		}
2743 
2744 		/* This LWB is from the past, so we can clean it up now. */
2745 		list_remove(&zilog->zl_lwb_crash_list, lwb);
2746 		if (lwb->lwb_buf != NULL)
2747 			zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
2748 		if (!BP_IS_HOLE(&lwb->lwb_blk))
2749 			/*
2750 			 * Free on the next txg, since zil_clean() is called
2751 			 * once synced_txg has already been completed.
2752 			 */
2753 			zio_free(zilog->zl_spa, synced_txg+1, &lwb->lwb_blk);
2754 		zil_free_lwb(zilog, lwb);
2755 	}
2756 
2757 	if (zilog->zl_restart_txg > synced_txg) {
2758 		/*
2759 		 * Not reached the restart txg yet, so mark the ZIL dirty for
2760 		 * the next txg and we'll consider it all again then.
2761 		 */
2762 		zilog_dirty(zilog, synced_txg+1);
2763 		return;
2764 	}
2765 
2766 	/*
2767 	 * Reached the restart txg, so we can allow new calls to zil_commit().
2768 	 * All ZIL txgs have long past so there should be no IO waiting.
2769 	 */
2770 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
2771 	ASSERT(list_is_empty(&zilog->zl_lwb_crash_list));
2772 
2773 	zilog->zl_restart_txg = 0;
2774 }
2775 
2776 /*
2777  * If there are any in-memory intent log transactions which have now been
2778  * synced then start up a taskq to free them. We should only do this after we
2779  * have written out the uberblocks (i.e. txg has been committed) so that
2780  * don't inadvertently clean out in-memory log records that would be required
2781  * by zil_commit().
2782  */
2783 void
2784 zil_clean(zilog_t *zilog, uint64_t synced_txg)
2785 {
2786 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2787 	itxs_t *clean_me;
2788 
2789 	ASSERT3U(synced_txg, <, ZILTEST_TXG);
2790 
2791 	/* Do cleanup and restart after crash. */
2792 	if (zilog->zl_restart_txg > 0) {
2793 		mutex_enter(&zilog->zl_lock);
2794 		/* Make sure we didn't lose a race. */
2795 		if (zilog->zl_restart_txg > 0)
2796 			zil_crash_clean(zilog, synced_txg);
2797 		mutex_exit(&zilog->zl_lock);
2798 	}
2799 
2800 	mutex_enter(&itxg->itxg_lock);
2801 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2802 		mutex_exit(&itxg->itxg_lock);
2803 		return;
2804 	}
2805 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2806 	ASSERT3U(itxg->itxg_txg, !=, 0);
2807 	clean_me = itxg->itxg_itxs;
2808 	itxg->itxg_itxs = NULL;
2809 	itxg->itxg_txg = 0;
2810 	mutex_exit(&itxg->itxg_lock);
2811 	/*
2812 	 * Preferably start a task queue to free up the old itxs but
2813 	 * if taskq_dispatch can't allocate resources to do that then
2814 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
2815 	 * created a bad performance problem.
2816 	 */
2817 	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2818 	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2819 	taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
2820 	    zil_itxg_clean, clean_me, TQ_NOSLEEP);
2821 	if (id == TASKQID_INVALID)
2822 		zil_itxg_clean(clean_me);
2823 }
2824 
2825 /*
2826  * This function will traverse the queue of itxs that need to be
2827  * committed, and move them onto the ZIL's zl_itx_commit_list.
2828  */
2829 static uint64_t
2830 zil_get_commit_list(zilog_t *zilog)
2831 {
2832 	uint64_t otxg, txg, wtxg = 0;
2833 	list_t *commit_list = &zilog->zl_itx_commit_list;
2834 
2835 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2836 
2837 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2838 		otxg = ZILTEST_TXG;
2839 	else
2840 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2841 
2842 	/*
2843 	 * This is inherently racy, since there is nothing to prevent
2844 	 * the last synced txg from changing. That's okay since we'll
2845 	 * only commit things in the future.
2846 	 */
2847 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2848 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2849 
2850 		mutex_enter(&itxg->itxg_lock);
2851 		if (itxg->itxg_txg != txg) {
2852 			mutex_exit(&itxg->itxg_lock);
2853 			continue;
2854 		}
2855 
2856 		/*
2857 		 * If we're adding itx records to the zl_itx_commit_list,
2858 		 * then the zil better be dirty in this "txg". We can assert
2859 		 * that here since we're holding the itxg_lock which will
2860 		 * prevent spa_sync from cleaning it. Once we add the itxs
2861 		 * to the zl_itx_commit_list we must commit it to disk even
2862 		 * if it's unnecessary (i.e. the txg was synced).
2863 		 */
2864 		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2865 		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2866 		list_t *sync_list = &itxg->itxg_itxs->i_sync_list;
2867 		itx_t *itx = NULL;
2868 		if (unlikely(zilog->zl_suspend > 0)) {
2869 			/*
2870 			 * ZIL was just suspended, but we lost the race.
2871 			 * Allow all earlier itxs to be committed, but ask
2872 			 * caller to do txg_wait_synced(txg) for any new.
2873 			 */
2874 			if (!list_is_empty(sync_list))
2875 				wtxg = MAX(wtxg, txg);
2876 		} else {
2877 			itx = list_head(sync_list);
2878 			list_move_tail(commit_list, sync_list);
2879 		}
2880 
2881 		mutex_exit(&itxg->itxg_lock);
2882 
2883 		while (itx != NULL) {
2884 			uint64_t s = zil_itx_full_size(itx);
2885 			zilog->zl_cur_size += s;
2886 			zilog->zl_cur_left += s;
2887 			s = zil_itx_record_size(itx);
2888 			zilog->zl_cur_max = MAX(zilog->zl_cur_max, s);
2889 			itx = list_next(commit_list, itx);
2890 		}
2891 	}
2892 	return (wtxg);
2893 }
2894 
2895 /*
2896  * Move the async itxs for a specified object to commit into sync lists.
2897  */
2898 void
2899 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2900 {
2901 	uint64_t otxg, txg;
2902 	itx_async_node_t *ian, ian_search;
2903 	avl_tree_t *t;
2904 	avl_index_t where;
2905 
2906 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2907 		otxg = ZILTEST_TXG;
2908 	else
2909 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2910 
2911 	/*
2912 	 * This is inherently racy, since there is nothing to prevent
2913 	 * the last synced txg from changing.
2914 	 */
2915 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2916 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2917 
2918 		mutex_enter(&itxg->itxg_lock);
2919 		if (itxg->itxg_txg != txg) {
2920 			mutex_exit(&itxg->itxg_lock);
2921 			continue;
2922 		}
2923 
2924 		/*
2925 		 * If a foid is specified then find that node and append its
2926 		 * list. Otherwise walk the tree appending all the lists
2927 		 * to the sync list. We add to the end rather than the
2928 		 * beginning to ensure the create has happened.
2929 		 */
2930 		t = &itxg->itxg_itxs->i_async_tree;
2931 		if (foid != 0) {
2932 			ian_search.ia_foid = foid;
2933 			ian = avl_find(t, &ian_search, &where);
2934 			if (ian != NULL) {
2935 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2936 				    &ian->ia_list);
2937 			}
2938 		} else {
2939 			void *cookie = NULL;
2940 
2941 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2942 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2943 				    &ian->ia_list);
2944 				list_destroy(&ian->ia_list);
2945 				kmem_free(ian, sizeof (itx_async_node_t));
2946 			}
2947 		}
2948 		mutex_exit(&itxg->itxg_lock);
2949 	}
2950 }
2951 
2952 /*
2953  * This function will prune commit itxs that are at the head of the
2954  * commit list (it won't prune past the first non-commit itx), and
2955  * either: a) attach them to the last lwb that's still pending
2956  * completion, or b) skip them altogether.
2957  *
2958  * This is used as a performance optimization to prevent commit itxs
2959  * from generating new lwbs when it's unnecessary to do so.
2960  */
2961 static void
2962 zil_prune_commit_list(zilog_t *zilog)
2963 {
2964 	itx_t *itx;
2965 
2966 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2967 
2968 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2969 		lr_t *lrc = &itx->itx_lr;
2970 		if (lrc->lrc_txtype != TX_COMMIT)
2971 			break;
2972 
2973 		mutex_enter(&zilog->zl_lock);
2974 
2975 		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2976 		if (last_lwb == NULL ||
2977 		    last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2978 			/*
2979 			 * All of the itxs this waiter was waiting on
2980 			 * must have already completed (or there were
2981 			 * never any itx's for it to wait on), so it's
2982 			 * safe to skip this waiter and mark it done.
2983 			 */
2984 			zil_commit_waiter_skip(itx->itx_private);
2985 		} else {
2986 			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2987 		}
2988 
2989 		mutex_exit(&zilog->zl_lock);
2990 
2991 		list_remove(&zilog->zl_itx_commit_list, itx);
2992 		zil_itx_destroy(itx, 0);
2993 	}
2994 
2995 	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2996 }
2997 
2998 static int
2999 zil_commit_writer_stall(zilog_t *zilog)
3000 {
3001 	/*
3002 	 * When zio_alloc_zil() fails to allocate the next lwb block on
3003 	 * disk, we must call txg_wait_synced() to ensure all of the
3004 	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
3005 	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
3006 	 * to zil_process_commit_list()) will have to call zil_create(),
3007 	 * and start a new ZIL chain.
3008 	 *
3009 	 * Since zil_alloc_zil() failed, the lwb that was previously
3010 	 * issued does not have a pointer to the "next" lwb on disk.
3011 	 * Thus, if another ZIL writer thread was to allocate the "next"
3012 	 * on-disk lwb, that block could be leaked in the event of a
3013 	 * crash (because the previous lwb on-disk would not point to
3014 	 * it).
3015 	 *
3016 	 * We must hold the zilog's zl_issuer_lock while we do this, to
3017 	 * ensure no new threads enter zil_process_commit_list() until
3018 	 * all lwb's in the zl_lwb_list have been synced and freed
3019 	 * (which is achieved via the txg_wait_synced() call).
3020 	 */
3021 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
3022 	ZIL_STAT_BUMP(zilog, zil_commit_stall_count);
3023 
3024 	int err = txg_wait_synced_flags(zilog->zl_dmu_pool, 0,
3025 	    TXG_WAIT_SUSPEND);
3026 	if (err != 0) {
3027 		ASSERT3U(err, ==, ESHUTDOWN);
3028 		zil_crash(zilog);
3029 	}
3030 
3031 	/*
3032 	 * Either zil_sync() has been called to wait for and clean up any
3033 	 * in-flight LWBs, or zil_crash() has emptied out the list and arranged
3034 	 * for them to be cleaned up later.
3035 	 */
3036 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3037 
3038 	return (err);
3039 }
3040 
3041 static void
3042 zil_burst_done(zilog_t *zilog)
3043 {
3044 	if (!list_is_empty(&zilog->zl_itx_commit_list) ||
3045 	    zilog->zl_cur_size == 0)
3046 		return;
3047 
3048 	if (zilog->zl_parallel)
3049 		zilog->zl_parallel--;
3050 
3051 	uint_t r = (zilog->zl_prev_rotor + 1) & (ZIL_BURSTS - 1);
3052 	zilog->zl_prev_rotor = r;
3053 	zilog->zl_prev_opt[r] = zil_lwb_plan(zilog, zilog->zl_cur_size,
3054 	    &zilog->zl_prev_min[r]);
3055 
3056 	zilog->zl_cur_size = 0;
3057 	zilog->zl_cur_max = 0;
3058 	zilog->zl_cur_left = 0;
3059 }
3060 
3061 /*
3062  * This function will traverse the commit list, creating new lwbs as
3063  * needed, and committing the itxs from the commit list to these newly
3064  * created lwbs. Additionally, as a new lwb is created, the previous
3065  * lwb will be issued to the zio layer to be written to disk.
3066  */
3067 static void
3068 zil_process_commit_list(zilog_t *zilog, zil_commit_waiter_t *zcw, list_t *ilwbs)
3069 {
3070 	spa_t *spa = zilog->zl_spa;
3071 	list_t nolwb_itxs;
3072 	list_t nolwb_waiters;
3073 	lwb_t *lwb, *plwb;
3074 	itx_t *itx;
3075 
3076 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
3077 
3078 	lwb = list_tail(&zilog->zl_lwb_list);
3079 	if (lwb == NULL) {
3080 		/*
3081 		 * Return if there's nothing to commit before we dirty the fs.
3082 		 */
3083 		if (list_is_empty(&zilog->zl_itx_commit_list))
3084 			return;
3085 
3086 		lwb = zil_create(zilog);
3087 	} else {
3088 		/*
3089 		 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
3090 		 * have already been created (zl_lwb_list not empty).
3091 		 */
3092 		zil_commit_activate_saxattr_feature(zilog);
3093 		ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
3094 		    lwb->lwb_state == LWB_STATE_OPENED);
3095 
3096 		/*
3097 		 * If the lwb is still opened, it means the workload is really
3098 		 * multi-threaded and we won the chance of write aggregation.
3099 		 * If it is not opened yet, but previous lwb is still not
3100 		 * flushed, it still means the workload is multi-threaded, but
3101 		 * there was too much time between the commits to aggregate, so
3102 		 * we try aggregation next times, but without too much hopes.
3103 		 */
3104 		if (lwb->lwb_state == LWB_STATE_OPENED) {
3105 			zilog->zl_parallel = ZIL_BURSTS;
3106 		} else if ((plwb = list_prev(&zilog->zl_lwb_list, lwb))
3107 		    != NULL && plwb->lwb_state != LWB_STATE_FLUSH_DONE) {
3108 			zilog->zl_parallel = MAX(zilog->zl_parallel,
3109 			    ZIL_BURSTS / 2);
3110 		}
3111 	}
3112 
3113 	list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3114 	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
3115 	    offsetof(zil_commit_waiter_t, zcw_node));
3116 
3117 	while ((itx = list_remove_head(&zilog->zl_itx_commit_list)) != NULL) {
3118 		lr_t *lrc = &itx->itx_lr;
3119 		uint64_t txg = lrc->lrc_txg;
3120 
3121 		ASSERT3U(txg, !=, 0);
3122 
3123 		if (lrc->lrc_txtype == TX_COMMIT) {
3124 			DTRACE_PROBE2(zil__process__commit__itx,
3125 			    zilog_t *, zilog, itx_t *, itx);
3126 		} else {
3127 			DTRACE_PROBE2(zil__process__normal__itx,
3128 			    zilog_t *, zilog, itx_t *, itx);
3129 		}
3130 
3131 		boolean_t synced = txg <= spa_last_synced_txg(spa);
3132 		boolean_t frozen = txg > spa_freeze_txg(spa);
3133 
3134 		/*
3135 		 * If the txg of this itx has already been synced out, then
3136 		 * we don't need to commit this itx to an lwb. This is
3137 		 * because the data of this itx will have already been
3138 		 * written to the main pool. This is inherently racy, and
3139 		 * it's still ok to commit an itx whose txg has already
3140 		 * been synced; this will result in a write that's
3141 		 * unnecessary, but will do no harm.
3142 		 *
3143 		 * With that said, we always want to commit TX_COMMIT itxs
3144 		 * to an lwb, regardless of whether or not that itx's txg
3145 		 * has been synced out. We do this to ensure any OPENED lwb
3146 		 * will always have at least one zil_commit_waiter_t linked
3147 		 * to the lwb.
3148 		 *
3149 		 * As a counter-example, if we skipped TX_COMMIT itx's
3150 		 * whose txg had already been synced, the following
3151 		 * situation could occur if we happened to be racing with
3152 		 * spa_sync:
3153 		 *
3154 		 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
3155 		 *    itx's txg is 10 and the last synced txg is 9.
3156 		 * 2. spa_sync finishes syncing out txg 10.
3157 		 * 3. We move to the next itx in the list, it's a TX_COMMIT
3158 		 *    whose txg is 10, so we skip it rather than committing
3159 		 *    it to the lwb used in (1).
3160 		 *
3161 		 * If the itx that is skipped in (3) is the last TX_COMMIT
3162 		 * itx in the commit list, than it's possible for the lwb
3163 		 * used in (1) to remain in the OPENED state indefinitely.
3164 		 *
3165 		 * To prevent the above scenario from occurring, ensuring
3166 		 * that once an lwb is OPENED it will transition to ISSUED
3167 		 * and eventually DONE, we always commit TX_COMMIT itx's to
3168 		 * an lwb here, even if that itx's txg has already been
3169 		 * synced.
3170 		 *
3171 		 * Finally, if the pool is frozen, we _always_ commit the
3172 		 * itx.  The point of freezing the pool is to prevent data
3173 		 * from being written to the main pool via spa_sync, and
3174 		 * instead rely solely on the ZIL to persistently store the
3175 		 * data; i.e.  when the pool is frozen, the last synced txg
3176 		 * value can't be trusted.
3177 		 */
3178 		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
3179 			if (lwb != NULL) {
3180 				lwb = zil_lwb_assign(zilog, lwb, itx, ilwbs);
3181 				if (lwb == NULL) {
3182 					list_insert_tail(&nolwb_itxs, itx);
3183 				} else if ((zcw->zcw_lwb != NULL &&
3184 				    zcw->zcw_lwb != lwb) || zcw->zcw_done) {
3185 					/*
3186 					 * Our lwb is done, leave the rest of
3187 					 * itx list to somebody else who care.
3188 					 */
3189 					zilog->zl_parallel = ZIL_BURSTS;
3190 					zilog->zl_cur_left -=
3191 					    zil_itx_full_size(itx);
3192 					break;
3193 				}
3194 			} else {
3195 				if (lrc->lrc_txtype == TX_COMMIT) {
3196 					zil_commit_waiter_link_nolwb(
3197 					    itx->itx_private, &nolwb_waiters);
3198 				}
3199 				list_insert_tail(&nolwb_itxs, itx);
3200 			}
3201 			zilog->zl_cur_left -= zil_itx_full_size(itx);
3202 		} else {
3203 			ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
3204 			zilog->zl_cur_left -= zil_itx_full_size(itx);
3205 			zil_itx_destroy(itx, 0);
3206 		}
3207 	}
3208 
3209 	if (lwb == NULL) {
3210 		/*
3211 		 * This indicates zio_alloc_zil() failed to allocate the
3212 		 * "next" lwb on-disk. When this happens, we must stall
3213 		 * the ZIL write pipeline; see the comment within
3214 		 * zil_commit_writer_stall() for more details.
3215 		 */
3216 		int err = 0;
3217 		while ((lwb = list_remove_head(ilwbs)) != NULL) {
3218 			err = zil_lwb_write_issue(zilog, lwb);
3219 			if (err != 0)
3220 				break;
3221 		}
3222 		if (err == 0)
3223 			err = zil_commit_writer_stall(zilog);
3224 
3225 		/*
3226 		 * Additionally, we have to signal and mark the "nolwb"
3227 		 * waiters as "done" here, since without an lwb, we
3228 		 * can't do this via zil_lwb_flush_vdevs_done() like
3229 		 * normal.
3230 		 */
3231 		zil_commit_waiter_t *zcw;
3232 		while ((zcw = list_remove_head(&nolwb_waiters)) != NULL)
3233 			zil_commit_waiter_skip(zcw);
3234 
3235 		/*
3236 		 * And finally, we have to destroy the itx's that
3237 		 * couldn't be committed to an lwb; this will also call
3238 		 * the itx's callback if one exists for the itx.
3239 		 */
3240 		while ((itx = list_remove_head(&nolwb_itxs)) != NULL)
3241 			zil_itx_destroy(itx, 0);
3242 	} else {
3243 		ASSERT(list_is_empty(&nolwb_waiters));
3244 		ASSERT3P(lwb, !=, NULL);
3245 		ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
3246 		    lwb->lwb_state == LWB_STATE_OPENED);
3247 
3248 		/*
3249 		 * At this point, the ZIL block pointed at by the "lwb"
3250 		 * variable is in "new" or "opened" state.
3251 		 *
3252 		 * If it's "new", then no itxs have been committed to it, so
3253 		 * there's no point in issuing its zio (i.e. it's "empty").
3254 		 *
3255 		 * If it's "opened", then it contains one or more itxs that
3256 		 * eventually need to be committed to stable storage. In
3257 		 * this case we intentionally do not issue the lwb's zio
3258 		 * to disk yet, and instead rely on one of the following
3259 		 * two mechanisms for issuing the zio:
3260 		 *
3261 		 * 1. Ideally, there will be more ZIL activity occurring on
3262 		 * the system, such that this function will be immediately
3263 		 * called again by different thread and this lwb will be
3264 		 * closed by zil_lwb_assign().  This way, the lwb will be
3265 		 * "full" when it is issued to disk, and we'll make use of
3266 		 * the lwb's size the best we can.
3267 		 *
3268 		 * 2. If there isn't sufficient ZIL activity occurring on
3269 		 * the system, zil_commit_waiter() will close it and issue
3270 		 * the zio.  If this occurs, the lwb is not guaranteed
3271 		 * to be "full" by the time its zio is issued, and means
3272 		 * the size of the lwb was "too large" given the amount
3273 		 * of ZIL activity occurring on the system at that time.
3274 		 *
3275 		 * We do this for a couple of reasons:
3276 		 *
3277 		 * 1. To try and reduce the number of IOPs needed to
3278 		 * write the same number of itxs. If an lwb has space
3279 		 * available in its buffer for more itxs, and more itxs
3280 		 * will be committed relatively soon (relative to the
3281 		 * latency of performing a write), then it's beneficial
3282 		 * to wait for these "next" itxs. This way, more itxs
3283 		 * can be committed to stable storage with fewer writes.
3284 		 *
3285 		 * 2. To try and use the largest lwb block size that the
3286 		 * incoming rate of itxs can support. Again, this is to
3287 		 * try and pack as many itxs into as few lwbs as
3288 		 * possible, without significantly impacting the latency
3289 		 * of each individual itx.
3290 		 */
3291 		if (lwb->lwb_state == LWB_STATE_OPENED &&
3292 		    (!zilog->zl_parallel || zilog->zl_suspend > 0)) {
3293 			zil_burst_done(zilog);
3294 			list_insert_tail(ilwbs, lwb);
3295 			lwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_NEW);
3296 			if (lwb == NULL) {
3297 				int err = 0;
3298 				while ((lwb =
3299 				    list_remove_head(ilwbs)) != NULL) {
3300 					err = zil_lwb_write_issue(zilog, lwb);
3301 					if (err != 0)
3302 						break;
3303 				}
3304 				if (err == 0)
3305 					zil_commit_writer_stall(zilog);
3306 			}
3307 		}
3308 	}
3309 }
3310 
3311 /*
3312  * This function is responsible for ensuring the passed in commit waiter
3313  * (and associated commit itx) is committed to an lwb. If the waiter is
3314  * not already committed to an lwb, all itxs in the zilog's queue of
3315  * itxs will be processed. The assumption is the passed in waiter's
3316  * commit itx will found in the queue just like the other non-commit
3317  * itxs, such that when the entire queue is processed, the waiter will
3318  * have been committed to an lwb.
3319  *
3320  * The lwb associated with the passed in waiter is not guaranteed to
3321  * have been issued by the time this function completes. If the lwb is
3322  * not issued, we rely on future calls to zil_commit_writer() to issue
3323  * the lwb, or the timeout mechanism found in zil_commit_waiter().
3324  */
3325 static uint64_t
3326 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
3327 {
3328 	list_t ilwbs;
3329 	lwb_t *lwb;
3330 	uint64_t wtxg = 0;
3331 
3332 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
3333 	ASSERT(spa_writeable(zilog->zl_spa));
3334 
3335 	list_create(&ilwbs, sizeof (lwb_t), offsetof(lwb_t, lwb_issue_node));
3336 	mutex_enter(&zilog->zl_issuer_lock);
3337 
3338 	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
3339 		/*
3340 		 * It's possible that, while we were waiting to acquire
3341 		 * the "zl_issuer_lock", another thread committed this
3342 		 * waiter to an lwb. If that occurs, we bail out early,
3343 		 * without processing any of the zilog's queue of itxs.
3344 		 *
3345 		 * On certain workloads and system configurations, the
3346 		 * "zl_issuer_lock" can become highly contended. In an
3347 		 * attempt to reduce this contention, we immediately drop
3348 		 * the lock if the waiter has already been processed.
3349 		 *
3350 		 * We've measured this optimization to reduce CPU spent
3351 		 * contending on this lock by up to 5%, using a system
3352 		 * with 32 CPUs, low latency storage (~50 usec writes),
3353 		 * and 1024 threads performing sync writes.
3354 		 */
3355 		goto out;
3356 	}
3357 
3358 	ZIL_STAT_BUMP(zilog, zil_commit_writer_count);
3359 
3360 	wtxg = zil_get_commit_list(zilog);
3361 	zil_prune_commit_list(zilog);
3362 	zil_process_commit_list(zilog, zcw, &ilwbs);
3363 
3364 	/*
3365 	 * If the ZIL failed somewhere inside zil_process_commit_list(), it's
3366 	 * will be because a fallback to txg_wait_sync_flags() happened at some
3367 	 * point (eg zil_commit_writer_stall()). All cases should issue and
3368 	 * empty ilwbs, so there will be nothing to in the issue loop below.
3369 	 * That's why we don't have to plumb the error value back from
3370 	 * zil_process_commit_list(), and don't have to skip it.
3371 	 */
3372 	IMPLY(zilog->zl_restart_txg > 0, list_is_empty(&ilwbs));
3373 
3374 out:
3375 	mutex_exit(&zilog->zl_issuer_lock);
3376 	int err = 0;
3377 	while ((lwb = list_remove_head(&ilwbs)) != NULL) {
3378 		if (err == 0)
3379 			err = zil_lwb_write_issue(zilog, lwb);
3380 	}
3381 	list_destroy(&ilwbs);
3382 	return (wtxg);
3383 }
3384 
3385 static void
3386 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
3387 {
3388 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3389 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3390 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
3391 
3392 	lwb_t *lwb = zcw->zcw_lwb;
3393 	ASSERT3P(lwb, !=, NULL);
3394 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW);
3395 
3396 	/*
3397 	 * If the lwb has already been issued by another thread, we can
3398 	 * immediately return since there's no work to be done (the
3399 	 * point of this function is to issue the lwb). Additionally, we
3400 	 * do this prior to acquiring the zl_issuer_lock, to avoid
3401 	 * acquiring it when it's not necessary to do so.
3402 	 */
3403 	if (lwb->lwb_state != LWB_STATE_OPENED)
3404 		return;
3405 
3406 	/*
3407 	 * In order to call zil_lwb_write_close() we must hold the
3408 	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
3409 	 * since we're already holding the commit waiter's "zcw_lock",
3410 	 * and those two locks are acquired in the opposite order
3411 	 * elsewhere.
3412 	 */
3413 	mutex_exit(&zcw->zcw_lock);
3414 	mutex_enter(&zilog->zl_issuer_lock);
3415 	mutex_enter(&zcw->zcw_lock);
3416 
3417 	/*
3418 	 * Since we just dropped and re-acquired the commit waiter's
3419 	 * lock, we have to re-check to see if the waiter was marked
3420 	 * "done" during that process. If the waiter was marked "done",
3421 	 * the "lwb" pointer is no longer valid (it can be free'd after
3422 	 * the waiter is marked "done"), so without this check we could
3423 	 * wind up with a use-after-free error below.
3424 	 */
3425 	if (zcw->zcw_done) {
3426 		mutex_exit(&zilog->zl_issuer_lock);
3427 		return;
3428 	}
3429 
3430 	ASSERT3P(lwb, ==, zcw->zcw_lwb);
3431 
3432 	/*
3433 	 * We've already checked this above, but since we hadn't acquired
3434 	 * the zilog's zl_issuer_lock, we have to perform this check a
3435 	 * second time while holding the lock.
3436 	 *
3437 	 * We don't need to hold the zl_lock since the lwb cannot transition
3438 	 * from OPENED to CLOSED while we hold the zl_issuer_lock. The lwb
3439 	 * _can_ transition from CLOSED to DONE, but it's OK to race with
3440 	 * that transition since we treat the lwb the same, whether it's in
3441 	 * the CLOSED, ISSUED or DONE states.
3442 	 *
3443 	 * The important thing, is we treat the lwb differently depending on
3444 	 * if it's OPENED or CLOSED, and block any other threads that might
3445 	 * attempt to close/issue this lwb. For that reason we hold the
3446 	 * zl_issuer_lock when checking the lwb_state; we must not call
3447 	 * zil_lwb_write_close() if the lwb had already been closed/issued.
3448 	 *
3449 	 * See the comment above the lwb_state_t structure definition for
3450 	 * more details on the lwb states, and locking requirements.
3451 	 */
3452 	if (lwb->lwb_state != LWB_STATE_OPENED) {
3453 		mutex_exit(&zilog->zl_issuer_lock);
3454 		return;
3455 	}
3456 
3457 	/*
3458 	 * We do not need zcw_lock once we hold zl_issuer_lock and know lwb
3459 	 * is still open.  But we have to drop it to avoid a deadlock in case
3460 	 * callback of zio issued by zil_lwb_write_issue() try to get it,
3461 	 * while zil_lwb_write_issue() is blocked on attempt to issue next
3462 	 * lwb it found in LWB_STATE_READY state.
3463 	 */
3464 	mutex_exit(&zcw->zcw_lock);
3465 
3466 	/*
3467 	 * As described in the comments above zil_commit_waiter() and
3468 	 * zil_process_commit_list(), we need to issue this lwb's zio
3469 	 * since we've reached the commit waiter's timeout and it still
3470 	 * hasn't been issued.
3471 	 */
3472 	zil_burst_done(zilog);
3473 	lwb_t *nlwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_NEW);
3474 
3475 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED);
3476 
3477 	if (nlwb == NULL) {
3478 		/*
3479 		 * When zil_lwb_write_close() returns NULL, this
3480 		 * indicates zio_alloc_zil() failed to allocate the
3481 		 * "next" lwb on-disk. When this occurs, the ZIL write
3482 		 * pipeline must be stalled; see the comment within the
3483 		 * zil_commit_writer_stall() function for more details.
3484 		 */
3485 		zil_lwb_write_issue(zilog, lwb);
3486 		zil_commit_writer_stall(zilog);
3487 		mutex_exit(&zilog->zl_issuer_lock);
3488 	} else {
3489 		mutex_exit(&zilog->zl_issuer_lock);
3490 		zil_lwb_write_issue(zilog, lwb);
3491 	}
3492 	mutex_enter(&zcw->zcw_lock);
3493 }
3494 
3495 /*
3496  * This function is responsible for performing the following two tasks:
3497  *
3498  * 1. its primary responsibility is to block until the given "commit
3499  *    waiter" is considered "done".
3500  *
3501  * 2. its secondary responsibility is to issue the zio for the lwb that
3502  *    the given "commit waiter" is waiting on, if this function has
3503  *    waited "long enough" and the lwb is still in the "open" state.
3504  *
3505  * Given a sufficient amount of itxs being generated and written using
3506  * the ZIL, the lwb's zio will be issued via the zil_lwb_assign()
3507  * function. If this does not occur, this secondary responsibility will
3508  * ensure the lwb is issued even if there is not other synchronous
3509  * activity on the system.
3510  *
3511  * For more details, see zil_process_commit_list(); more specifically,
3512  * the comment at the bottom of that function.
3513  */
3514 static void
3515 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
3516 {
3517 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
3518 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3519 	ASSERT(spa_writeable(zilog->zl_spa));
3520 
3521 	mutex_enter(&zcw->zcw_lock);
3522 
3523 	/*
3524 	 * The timeout is scaled based on the lwb latency to avoid
3525 	 * significantly impacting the latency of each individual itx.
3526 	 * For more details, see the comment at the bottom of the
3527 	 * zil_process_commit_list() function.
3528 	 */
3529 	int pct = MAX(zfs_commit_timeout_pct, 1);
3530 	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
3531 	hrtime_t wakeup = gethrtime() + sleep;
3532 	boolean_t timedout = B_FALSE;
3533 
3534 	while (!zcw->zcw_done) {
3535 		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3536 
3537 		lwb_t *lwb = zcw->zcw_lwb;
3538 
3539 		/*
3540 		 * Usually, the waiter will have a non-NULL lwb field here,
3541 		 * but it's possible for it to be NULL as a result of
3542 		 * zil_commit() racing with spa_sync().
3543 		 *
3544 		 * When zil_clean() is called, it's possible for the itxg
3545 		 * list (which may be cleaned via a taskq) to contain
3546 		 * commit itxs. When this occurs, the commit waiters linked
3547 		 * off of these commit itxs will not be committed to an
3548 		 * lwb.  Additionally, these commit waiters will not be
3549 		 * marked done until zil_commit_waiter_skip() is called via
3550 		 * zil_itxg_clean().
3551 		 *
3552 		 * Thus, it's possible for this commit waiter (i.e. the
3553 		 * "zcw" variable) to be found in this "in between" state;
3554 		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
3555 		 * been skipped, so it's "zcw_done" field is still B_FALSE.
3556 		 */
3557 		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_NEW);
3558 
3559 		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
3560 			ASSERT3B(timedout, ==, B_FALSE);
3561 
3562 			/*
3563 			 * If the lwb hasn't been issued yet, then we
3564 			 * need to wait with a timeout, in case this
3565 			 * function needs to issue the lwb after the
3566 			 * timeout is reached; responsibility (2) from
3567 			 * the comment above this function.
3568 			 */
3569 			int rc = cv_timedwait_hires(&zcw->zcw_cv,
3570 			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
3571 			    CALLOUT_FLAG_ABSOLUTE);
3572 
3573 			if (rc != -1 || zcw->zcw_done)
3574 				continue;
3575 
3576 			timedout = B_TRUE;
3577 			zil_commit_waiter_timeout(zilog, zcw);
3578 
3579 			if (!zcw->zcw_done) {
3580 				/*
3581 				 * If the commit waiter has already been
3582 				 * marked "done", it's possible for the
3583 				 * waiter's lwb structure to have already
3584 				 * been freed.  Thus, we can only reliably
3585 				 * make these assertions if the waiter
3586 				 * isn't done.
3587 				 */
3588 				ASSERT3P(lwb, ==, zcw->zcw_lwb);
3589 				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
3590 			}
3591 		} else {
3592 			/*
3593 			 * If the lwb isn't open, then it must have already
3594 			 * been issued. In that case, there's no need to
3595 			 * use a timeout when waiting for the lwb to
3596 			 * complete.
3597 			 *
3598 			 * Additionally, if the lwb is NULL, the waiter
3599 			 * will soon be signaled and marked done via
3600 			 * zil_clean() and zil_itxg_clean(), so no timeout
3601 			 * is required.
3602 			 */
3603 
3604 			IMPLY(lwb != NULL,
3605 			    lwb->lwb_state == LWB_STATE_CLOSED ||
3606 			    lwb->lwb_state == LWB_STATE_READY ||
3607 			    lwb->lwb_state == LWB_STATE_ISSUED ||
3608 			    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
3609 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
3610 			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
3611 		}
3612 	}
3613 
3614 	mutex_exit(&zcw->zcw_lock);
3615 }
3616 
3617 static zil_commit_waiter_t *
3618 zil_alloc_commit_waiter(void)
3619 {
3620 	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
3621 
3622 	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
3623 	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
3624 	list_link_init(&zcw->zcw_node);
3625 	zcw->zcw_lwb = NULL;
3626 	zcw->zcw_done = B_FALSE;
3627 	zcw->zcw_zio_error = 0;
3628 
3629 	return (zcw);
3630 }
3631 
3632 static void
3633 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
3634 {
3635 	ASSERT(!list_link_active(&zcw->zcw_node));
3636 	ASSERT0P(zcw->zcw_lwb);
3637 	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
3638 	mutex_destroy(&zcw->zcw_lock);
3639 	cv_destroy(&zcw->zcw_cv);
3640 	kmem_cache_free(zil_zcw_cache, zcw);
3641 }
3642 
3643 /*
3644  * This function is used to create a TX_COMMIT itx and assign it. This
3645  * way, it will be linked into the ZIL's list of synchronous itxs, and
3646  * then later committed to an lwb (or skipped) when
3647  * zil_process_commit_list() is called.
3648  */
3649 static void
3650 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
3651 {
3652 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
3653 
3654 	/*
3655 	 * Since we are not going to create any new dirty data, and we
3656 	 * can even help with clearing the existing dirty data, we
3657 	 * should not be subject to the dirty data based delays. We
3658 	 * use DMU_TX_NOTHROTTLE to bypass the delay mechanism.
3659 	 */
3660 	VERIFY0(dmu_tx_assign(tx,
3661 	    DMU_TX_WAIT | DMU_TX_NOTHROTTLE | DMU_TX_SUSPEND));
3662 
3663 	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
3664 	itx->itx_sync = B_TRUE;
3665 	itx->itx_private = zcw;
3666 
3667 	zil_itx_assign(zilog, itx, tx);
3668 
3669 	dmu_tx_commit(tx);
3670 }
3671 
3672 /*
3673  * Crash the ZIL. This is something like suspending, but abandons the ZIL
3674  * without further IO until the wanted txg completes. No effort is made to
3675  * close the on-disk chain or do any other on-disk work, as the pool may
3676  * have suspended. zil_sync() will handle cleanup as normal and restart the
3677  * ZIL once enough txgs have passed.
3678  */
3679 static void
3680 zil_crash(zilog_t *zilog)
3681 {
3682 	mutex_enter(&zilog->zl_lock);
3683 
3684 	uint64_t txg = spa_syncing_txg(zilog->zl_spa);
3685 	uint64_t restart_txg =
3686 	    spa_syncing_txg(zilog->zl_spa) + TXG_CONCURRENT_STATES;
3687 
3688 	if (zilog->zl_restart_txg > 0) {
3689 		/*
3690 		 * If the ZIL is already crashed, it's almost certainly because
3691 		 * we lost a race involving multiple callers from
3692 		 * zil_commit_impl().
3693 		 */
3694 
3695 		/*
3696 		 * This sanity check is to support my understanding that in the
3697 		 * event of multiple callers to zil_crash(), only one of them
3698 		 * can possibly be in the codepath to issue lwbs; the rest
3699 		 * should be calling from zil_commit_impl() after their waiters
3700 		 * have completed. As I understand it, a second thread trying
3701 		 * to issue will eventually wait on zl_issuer_lock, and then
3702 		 * have no work to do and leave.
3703 		 *
3704 		 * If more lwbs had been created an issued between zil_crash()
3705 		 * calls, then we probably just need to take those too, add
3706 		 * them to the crash list and clean them up, but it complicates
3707 		 * this function and I don't think it can happend.
3708 		 */
3709 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3710 
3711 		mutex_exit(&zilog->zl_lock);
3712 		return;
3713 	}
3714 
3715 	zilog->zl_restart_txg = restart_txg;
3716 
3717 	/*
3718 	 * Capture any live LWBs. Depending on the state of the pool they may
3719 	 * represent in-flight IO that won't return for some time, and we want
3720 	 * to make sure they don't get in the way of normal ZIL operation.
3721 	 */
3722 	ASSERT(list_is_empty(&zilog->zl_lwb_crash_list));
3723 	list_move_tail(&zilog->zl_lwb_crash_list, &zilog->zl_lwb_list);
3724 
3725 	/*
3726 	 * Run through the LWB list; erroring all itxes and signalling error
3727 	 * to all waiters.
3728 	 */
3729 	for (lwb_t *lwb = list_head(&zilog->zl_lwb_crash_list); lwb != NULL;
3730 	    lwb = list_next(&zilog->zl_lwb_crash_list, lwb)) {
3731 		itx_t *itx;
3732 		while ((itx = list_remove_head(&lwb->lwb_itxs)) != NULL)
3733 			zil_itx_destroy(itx, EIO);
3734 
3735 		zil_commit_waiter_t *zcw;
3736 		while ((zcw = list_remove_head(&lwb->lwb_waiters)) != NULL) {
3737 			mutex_enter(&zcw->zcw_lock);
3738 			zcw->zcw_lwb = NULL;
3739 			zcw->zcw_zio_error = EIO;
3740 			zcw->zcw_done = B_TRUE;
3741 			cv_broadcast(&zcw->zcw_cv);
3742 			mutex_exit(&zcw->zcw_lock);
3743 		}
3744 	}
3745 
3746 	/*
3747 	 * Zero the ZIL header bp after the ZIL restarts. We'll free it in
3748 	 * zil_clean() when we clean up the lwbs.
3749 	 */
3750 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
3751 	BP_ZERO(&zh->zh_log);
3752 
3753 	/*
3754 	 * Mark this ZIL dirty on the next txg, so that zil_clean() will be
3755 	 * called for cleanup.
3756 	 */
3757 	zilog_dirty(zilog, txg+1);
3758 
3759 	mutex_exit(&zilog->zl_lock);
3760 }
3761 
3762 /*
3763  * Commit ZFS Intent Log transactions (itxs) to stable storage.
3764  *
3765  * When writing ZIL transactions to the on-disk representation of the
3766  * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3767  * itxs can be committed to a single lwb. Once a lwb is written and
3768  * committed to stable storage (i.e. the lwb is written, and vdevs have
3769  * been flushed), each itx that was committed to that lwb is also
3770  * considered to be committed to stable storage.
3771  *
3772  * When an itx is committed to an lwb, the log record (lr_t) contained
3773  * by the itx is copied into the lwb's zio buffer, and once this buffer
3774  * is written to disk, it becomes an on-disk ZIL block.
3775  *
3776  * As itxs are generated, they're inserted into the ZIL's queue of
3777  * uncommitted itxs. The semantics of zil_commit() are such that it will
3778  * block until all itxs that were in the queue when it was called, are
3779  * committed to stable storage.
3780  *
3781  * If "foid" is zero, this means all "synchronous" and "asynchronous"
3782  * itxs, for all objects in the dataset, will be committed to stable
3783  * storage prior to zil_commit() returning. If "foid" is non-zero, all
3784  * "synchronous" itxs for all objects, but only "asynchronous" itxs
3785  * that correspond to the foid passed in, will be committed to stable
3786  * storage prior to zil_commit() returning.
3787  *
3788  * Generally speaking, when zil_commit() is called, the consumer doesn't
3789  * actually care about _all_ of the uncommitted itxs. Instead, they're
3790  * simply trying to waiting for a specific itx to be committed to disk,
3791  * but the interface(s) for interacting with the ZIL don't allow such
3792  * fine-grained communication. A better interface would allow a consumer
3793  * to create and assign an itx, and then pass a reference to this itx to
3794  * zil_commit(); such that zil_commit() would return as soon as that
3795  * specific itx was committed to disk (instead of waiting for _all_
3796  * itxs to be committed).
3797  *
3798  * When a thread calls zil_commit() a special "commit itx" will be
3799  * generated, along with a corresponding "waiter" for this commit itx.
3800  * zil_commit() will wait on this waiter's CV, such that when the waiter
3801  * is marked done, and signaled, zil_commit() will return.
3802  *
3803  * This commit itx is inserted into the queue of uncommitted itxs. This
3804  * provides an easy mechanism for determining which itxs were in the
3805  * queue prior to zil_commit() having been called, and which itxs were
3806  * added after zil_commit() was called.
3807  *
3808  * The commit itx is special; it doesn't have any on-disk representation.
3809  * When a commit itx is "committed" to an lwb, the waiter associated
3810  * with it is linked onto the lwb's list of waiters. Then, when that lwb
3811  * completes, each waiter on the lwb's list is marked done and signaled
3812  * -- allowing the thread waiting on the waiter to return from zil_commit().
3813  *
3814  * It's important to point out a few critical factors that allow us
3815  * to make use of the commit itxs, commit waiters, per-lwb lists of
3816  * commit waiters, and zio completion callbacks like we're doing:
3817  *
3818  *   1. The list of waiters for each lwb is traversed, and each commit
3819  *      waiter is marked "done" and signaled, in the zio completion
3820  *      callback of the lwb's zio[*].
3821  *
3822  *      * Actually, the waiters are signaled in the zio completion
3823  *        callback of the root zio for the flush commands that are sent to
3824  *        the vdevs upon completion of the lwb zio.
3825  *
3826  *   2. When the itxs are inserted into the ZIL's queue of uncommitted
3827  *      itxs, the order in which they are inserted is preserved[*]; as
3828  *      itxs are added to the queue, they are added to the tail of
3829  *      in-memory linked lists.
3830  *
3831  *      When committing the itxs to lwbs (to be written to disk), they
3832  *      are committed in the same order in which the itxs were added to
3833  *      the uncommitted queue's linked list(s); i.e. the linked list of
3834  *      itxs to commit is traversed from head to tail, and each itx is
3835  *      committed to an lwb in that order.
3836  *
3837  *      * To clarify:
3838  *
3839  *        - the order of "sync" itxs is preserved w.r.t. other
3840  *          "sync" itxs, regardless of the corresponding objects.
3841  *        - the order of "async" itxs is preserved w.r.t. other
3842  *          "async" itxs corresponding to the same object.
3843  *        - the order of "async" itxs is *not* preserved w.r.t. other
3844  *          "async" itxs corresponding to different objects.
3845  *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
3846  *          versa) is *not* preserved, even for itxs that correspond
3847  *          to the same object.
3848  *
3849  *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
3850  *      zil_get_commit_list(), and zil_process_commit_list().
3851  *
3852  *   3. The lwbs represent a linked list of blocks on disk. Thus, any
3853  *      lwb cannot be considered committed to stable storage, until its
3854  *      "previous" lwb is also committed to stable storage. This fact,
3855  *      coupled with the fact described above, means that itxs are
3856  *      committed in (roughly) the order in which they were generated.
3857  *      This is essential because itxs are dependent on prior itxs.
3858  *      Thus, we *must not* deem an itx as being committed to stable
3859  *      storage, until *all* prior itxs have also been committed to
3860  *      stable storage.
3861  *
3862  *      To enforce this ordering of lwb zio's, while still leveraging as
3863  *      much of the underlying storage performance as possible, we rely
3864  *      on two fundamental concepts:
3865  *
3866  *          1. The creation and issuance of lwb zio's is protected by
3867  *             the zilog's "zl_issuer_lock", which ensures only a single
3868  *             thread is creating and/or issuing lwb's at a time
3869  *          2. The "previous" lwb is a child of the "current" lwb
3870  *             (leveraging the zio parent-child dependency graph)
3871  *
3872  *      By relying on this parent-child zio relationship, we can have
3873  *      many lwb zio's concurrently issued to the underlying storage,
3874  *      but the order in which they complete will be the same order in
3875  *      which they were created.
3876  */
3877 static int zil_commit_impl(zilog_t *zilog, uint64_t foid);
3878 
3879 int
3880 zil_commit(zilog_t *zilog, uint64_t foid)
3881 {
3882 	return (zil_commit_flags(zilog, foid, ZIL_COMMIT_FAILMODE));
3883 }
3884 
3885 int
3886 zil_commit_flags(zilog_t *zilog, uint64_t foid, zil_commit_flag_t flags)
3887 {
3888 	/*
3889 	 * We should never attempt to call zil_commit on a snapshot for
3890 	 * a couple of reasons:
3891 	 *
3892 	 * 1. A snapshot may never be modified, thus it cannot have any
3893 	 *    in-flight itxs that would have modified the dataset.
3894 	 *
3895 	 * 2. By design, when zil_commit() is called, a commit itx will
3896 	 *    be assigned to this zilog; as a result, the zilog will be
3897 	 *    dirtied. We must not dirty the zilog of a snapshot; there's
3898 	 *    checks in the code that enforce this invariant, and will
3899 	 *    cause a panic if it's not upheld.
3900 	 */
3901 	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
3902 
3903 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3904 		return (0);
3905 
3906 	if (!spa_writeable(zilog->zl_spa)) {
3907 		/*
3908 		 * If the SPA is not writable, there should never be any
3909 		 * pending itxs waiting to be committed to disk. If that
3910 		 * weren't true, we'd skip writing those itxs out, and
3911 		 * would break the semantics of zil_commit(); thus, we're
3912 		 * verifying that truth before we return to the caller.
3913 		 */
3914 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3915 		ASSERT0P(zilog->zl_last_lwb_opened);
3916 		for (int i = 0; i < TXG_SIZE; i++)
3917 			ASSERT0P(zilog->zl_itxg[i].itxg_itxs);
3918 		return (0);
3919 	}
3920 
3921 	int err = 0;
3922 
3923 	/*
3924 	 * If the ZIL crashed, bypass it entirely, and rely on txg_wait_sync()
3925 	 * to get the data out to disk.
3926 	 */
3927 	if (zilog->zl_restart_txg > 0) {
3928 		ZIL_STAT_BUMP(zilog, zil_commit_crash_count);
3929 		err = txg_wait_synced_flags(zilog->zl_dmu_pool, 0,
3930 		    TXG_WAIT_SUSPEND);
3931 		goto out;
3932 	}
3933 
3934 	/*
3935 	 * If the ZIL is suspended, we don't want to dirty it by calling
3936 	 * zil_commit_itx_assign() below, nor can we write out
3937 	 * lwbs like would be done in zil_commit_write(). Thus, we
3938 	 * simply rely on txg_wait_synced() to maintain the necessary
3939 	 * semantics, and avoid calling those functions altogether.
3940 	 */
3941 	if (zilog->zl_suspend > 0) {
3942 		ZIL_STAT_BUMP(zilog, zil_commit_suspend_count);
3943 		err = txg_wait_synced_flags(zilog->zl_dmu_pool, 0,
3944 		    TXG_WAIT_SUSPEND);
3945 		if (err != 0) {
3946 			ASSERT3U(err, ==, ESHUTDOWN);
3947 			zil_crash(zilog);
3948 		}
3949 		goto out;
3950 	}
3951 
3952 	err = zil_commit_impl(zilog, foid);
3953 
3954 out:
3955 	if (err == 0)
3956 		return (0);
3957 
3958 	/*
3959 	 * The ZIL write failed and the pool is suspended. There's nothing else
3960 	 * we can do except return or block.
3961 	 */
3962 	ASSERT3U(err, ==, ESHUTDOWN);
3963 
3964 	/*
3965 	 * Return error if failmode=continue or caller will handle directly.
3966 	 */
3967 	if (!(flags & ZIL_COMMIT_FAILMODE) ||
3968 	    spa_get_failmode(zilog->zl_spa) == ZIO_FAILURE_MODE_CONTINUE)
3969 		return (SET_ERROR(EIO));
3970 
3971 	/*
3972 	 * Block until the pool returns. We assume that the data will make
3973 	 * it out to disk in the end, and so return success.
3974 	 */
3975 	txg_wait_synced(zilog->zl_dmu_pool, 0);
3976 	return (0);
3977 }
3978 
3979 static int
3980 zil_commit_impl(zilog_t *zilog, uint64_t foid)
3981 {
3982 	ZIL_STAT_BUMP(zilog, zil_commit_count);
3983 
3984 	/*
3985 	 * Move the "async" itxs for the specified foid to the "sync"
3986 	 * queues, such that they will be later committed (or skipped)
3987 	 * to an lwb when zil_process_commit_list() is called.
3988 	 *
3989 	 * Since these "async" itxs must be committed prior to this
3990 	 * call to zil_commit returning, we must perform this operation
3991 	 * before we call zil_commit_itx_assign().
3992 	 */
3993 	zil_async_to_sync(zilog, foid);
3994 
3995 	/*
3996 	 * We allocate a new "waiter" structure which will initially be
3997 	 * linked to the commit itx using the itx's "itx_private" field.
3998 	 * Since the commit itx doesn't represent any on-disk state,
3999 	 * when it's committed to an lwb, rather than copying the its
4000 	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
4001 	 * added to the lwb's list of waiters. Then, when the lwb is
4002 	 * committed to stable storage, each waiter in the lwb's list of
4003 	 * waiters will be marked "done", and signalled.
4004 	 *
4005 	 * We must create the waiter and assign the commit itx prior to
4006 	 * calling zil_commit_writer(), or else our specific commit itx
4007 	 * is not guaranteed to be committed to an lwb prior to calling
4008 	 * zil_commit_waiter().
4009 	 */
4010 	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
4011 	zil_commit_itx_assign(zilog, zcw);
4012 
4013 	uint64_t wtxg = zil_commit_writer(zilog, zcw);
4014 	zil_commit_waiter(zilog, zcw);
4015 
4016 	int err = 0;
4017 	if (zcw->zcw_zio_error != 0) {
4018 		/*
4019 		 * If there was an error writing out the ZIL blocks that
4020 		 * this thread is waiting on, then we fallback to
4021 		 * relying on spa_sync() to write out the data this
4022 		 * thread is waiting on. Obviously this has performance
4023 		 * implications, but the expectation is for this to be
4024 		 * an exceptional case, and shouldn't occur often.
4025 		 */
4026 		ZIL_STAT_BUMP(zilog, zil_commit_error_count);
4027 		DTRACE_PROBE2(zil__commit__io__error,
4028 		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
4029 		err = txg_wait_synced_flags(zilog->zl_dmu_pool, 0,
4030 		    TXG_WAIT_SUSPEND);
4031 	} else if (wtxg != 0) {
4032 		ZIL_STAT_BUMP(zilog, zil_commit_suspend_count);
4033 		err = txg_wait_synced_flags(zilog->zl_dmu_pool, wtxg,
4034 		    TXG_WAIT_SUSPEND);
4035 	}
4036 
4037 	zil_free_commit_waiter(zcw);
4038 
4039 	if (err == 0)
4040 		return (0);
4041 
4042 	/*
4043 	 * ZIL write failed and pool failed in the fallback to
4044 	 * txg_wait_synced_flags(). Right now we have no idea if the data is on
4045 	 * disk and the pool is probably suspended so we have no idea when it's
4046 	 * coming back. All we can do is shut down and return error to the
4047 	 * caller.
4048 	 */
4049 	ASSERT3U(err, ==, ESHUTDOWN);
4050 	zil_crash(zilog);
4051 	return (err);
4052 }
4053 
4054 /*
4055  * Called in syncing context to free committed log blocks and update log header.
4056  */
4057 void
4058 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
4059 {
4060 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
4061 	uint64_t txg = dmu_tx_get_txg(tx);
4062 	spa_t *spa = zilog->zl_spa;
4063 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
4064 	lwb_t *lwb;
4065 
4066 	/*
4067 	 * We don't zero out zl_destroy_txg, so make sure we don't try
4068 	 * to destroy it twice.
4069 	 */
4070 	if (spa_sync_pass(spa) != 1)
4071 		return;
4072 
4073 	zil_lwb_flush_wait_all(zilog, txg);
4074 
4075 	mutex_enter(&zilog->zl_lock);
4076 
4077 	ASSERT0(zilog->zl_stop_sync);
4078 
4079 	if (*replayed_seq != 0) {
4080 		ASSERT(zh->zh_replay_seq < *replayed_seq);
4081 		zh->zh_replay_seq = *replayed_seq;
4082 		*replayed_seq = 0;
4083 	}
4084 
4085 	if (zilog->zl_destroy_txg == txg) {
4086 		blkptr_t blk = zh->zh_log;
4087 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
4088 
4089 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
4090 
4091 		memset(zh, 0, sizeof (zil_header_t));
4092 		memset(zilog->zl_replayed_seq, 0,
4093 		    sizeof (zilog->zl_replayed_seq));
4094 
4095 		if (zilog->zl_keep_first) {
4096 			/*
4097 			 * If this block was part of log chain that couldn't
4098 			 * be claimed because a device was missing during
4099 			 * zil_claim(), but that device later returns,
4100 			 * then this block could erroneously appear valid.
4101 			 * To guard against this, assign a new GUID to the new
4102 			 * log chain so it doesn't matter what blk points to.
4103 			 */
4104 			zil_init_log_chain(zilog, &blk);
4105 			zh->zh_log = blk;
4106 		} else {
4107 			/*
4108 			 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
4109 			 * records. So, deactivate the feature for this dataset.
4110 			 * We activate it again when we start a new ZIL chain.
4111 			 */
4112 			if (dsl_dataset_feature_is_active(ds,
4113 			    SPA_FEATURE_ZILSAXATTR))
4114 				dsl_dataset_deactivate_feature(ds,
4115 				    SPA_FEATURE_ZILSAXATTR, tx);
4116 		}
4117 	}
4118 
4119 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
4120 		zh->zh_log = lwb->lwb_blk;
4121 		if (lwb->lwb_state != LWB_STATE_FLUSH_DONE ||
4122 		    lwb->lwb_alloc_txg > txg || lwb->lwb_max_txg > txg)
4123 			break;
4124 		list_remove(&zilog->zl_lwb_list, lwb);
4125 		if (!BP_IS_HOLE(&lwb->lwb_blk))
4126 			zio_free(spa, txg, &lwb->lwb_blk);
4127 		zil_free_lwb(zilog, lwb);
4128 
4129 		/*
4130 		 * If we don't have anything left in the lwb list then
4131 		 * we've had an allocation failure and we need to zero
4132 		 * out the zil_header blkptr so that we don't end
4133 		 * up freeing the same block twice.
4134 		 */
4135 		if (list_is_empty(&zilog->zl_lwb_list))
4136 			BP_ZERO(&zh->zh_log);
4137 	}
4138 
4139 	mutex_exit(&zilog->zl_lock);
4140 }
4141 
4142 static int
4143 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
4144 {
4145 	(void) unused, (void) kmflag;
4146 	lwb_t *lwb = vbuf;
4147 	list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
4148 	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
4149 	    offsetof(zil_commit_waiter_t, zcw_node));
4150 	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
4151 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
4152 	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
4153 	return (0);
4154 }
4155 
4156 static void
4157 zil_lwb_dest(void *vbuf, void *unused)
4158 {
4159 	(void) unused;
4160 	lwb_t *lwb = vbuf;
4161 	mutex_destroy(&lwb->lwb_vdev_lock);
4162 	avl_destroy(&lwb->lwb_vdev_tree);
4163 	list_destroy(&lwb->lwb_waiters);
4164 	list_destroy(&lwb->lwb_itxs);
4165 }
4166 
4167 void
4168 zil_init(void)
4169 {
4170 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
4171 	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
4172 
4173 	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
4174 	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
4175 
4176 	zil_sums_init(&zil_sums_global);
4177 	zil_kstats_global = kstat_create("zfs", 0, "zil", "misc",
4178 	    KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
4179 	    KSTAT_FLAG_VIRTUAL);
4180 
4181 	if (zil_kstats_global != NULL) {
4182 		zil_kstats_global->ks_data = &zil_stats;
4183 		zil_kstats_global->ks_update = zil_kstats_global_update;
4184 		zil_kstats_global->ks_private = NULL;
4185 		kstat_install(zil_kstats_global);
4186 	}
4187 }
4188 
4189 void
4190 zil_fini(void)
4191 {
4192 	kmem_cache_destroy(zil_zcw_cache);
4193 	kmem_cache_destroy(zil_lwb_cache);
4194 
4195 	if (zil_kstats_global != NULL) {
4196 		kstat_delete(zil_kstats_global);
4197 		zil_kstats_global = NULL;
4198 	}
4199 
4200 	zil_sums_fini(&zil_sums_global);
4201 }
4202 
4203 void
4204 zil_set_sync(zilog_t *zilog, uint64_t sync)
4205 {
4206 	zilog->zl_sync = sync;
4207 }
4208 
4209 void
4210 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
4211 {
4212 	zilog->zl_logbias = logbias;
4213 }
4214 
4215 zilog_t *
4216 zil_alloc(objset_t *os, zil_header_t *zh_phys)
4217 {
4218 	zilog_t *zilog;
4219 
4220 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
4221 
4222 	zilog->zl_header = zh_phys;
4223 	zilog->zl_os = os;
4224 	zilog->zl_spa = dmu_objset_spa(os);
4225 	zilog->zl_dmu_pool = dmu_objset_pool(os);
4226 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
4227 	zilog->zl_logbias = dmu_objset_logbias(os);
4228 	zilog->zl_sync = dmu_objset_syncprop(os);
4229 	zilog->zl_dirty_max_txg = 0;
4230 	zilog->zl_last_lwb_opened = NULL;
4231 	zilog->zl_last_lwb_latency = 0;
4232 	zilog->zl_max_block_size = MIN(MAX(P2ALIGN_TYPED(zil_maxblocksize,
4233 	    ZIL_MIN_BLKSZ, uint64_t), ZIL_MIN_BLKSZ),
4234 	    spa_maxblocksize(dmu_objset_spa(os)));
4235 
4236 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
4237 	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
4238 	mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL);
4239 
4240 	for (int i = 0; i < TXG_SIZE; i++) {
4241 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
4242 		    MUTEX_DEFAULT, NULL);
4243 	}
4244 
4245 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
4246 	    offsetof(lwb_t, lwb_node));
4247 	list_create(&zilog->zl_lwb_crash_list, sizeof (lwb_t),
4248 	    offsetof(lwb_t, lwb_node));
4249 
4250 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
4251 	    offsetof(itx_t, itx_node));
4252 
4253 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
4254 	cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL);
4255 
4256 	for (int i = 0; i < ZIL_BURSTS; i++) {
4257 		zilog->zl_prev_opt[i] = zilog->zl_max_block_size -
4258 		    sizeof (zil_chain_t);
4259 	}
4260 
4261 	return (zilog);
4262 }
4263 
4264 void
4265 zil_free(zilog_t *zilog)
4266 {
4267 	int i;
4268 
4269 	zilog->zl_stop_sync = 1;
4270 
4271 	ASSERT0(zilog->zl_suspend);
4272 	ASSERT0(zilog->zl_suspending);
4273 	ASSERT0(zilog->zl_restart_txg);
4274 
4275 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
4276 	list_destroy(&zilog->zl_lwb_list);
4277 	ASSERT(list_is_empty(&zilog->zl_lwb_crash_list));
4278 	list_destroy(&zilog->zl_lwb_crash_list);
4279 
4280 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
4281 	list_destroy(&zilog->zl_itx_commit_list);
4282 
4283 	for (i = 0; i < TXG_SIZE; i++) {
4284 		/*
4285 		 * It's possible for an itx to be generated that doesn't dirty
4286 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
4287 		 * callback to remove the entry. We remove those here.
4288 		 *
4289 		 * Also free up the ziltest itxs.
4290 		 */
4291 		if (zilog->zl_itxg[i].itxg_itxs)
4292 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
4293 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
4294 	}
4295 
4296 	mutex_destroy(&zilog->zl_issuer_lock);
4297 	mutex_destroy(&zilog->zl_lock);
4298 	mutex_destroy(&zilog->zl_lwb_io_lock);
4299 
4300 	cv_destroy(&zilog->zl_cv_suspend);
4301 	cv_destroy(&zilog->zl_lwb_io_cv);
4302 
4303 	kmem_free(zilog, sizeof (zilog_t));
4304 }
4305 
4306 /*
4307  * Open an intent log.
4308  */
4309 zilog_t *
4310 zil_open(objset_t *os, zil_get_data_t *get_data, zil_sums_t *zil_sums)
4311 {
4312 	zilog_t *zilog = dmu_objset_zil(os);
4313 
4314 	ASSERT0P(zilog->zl_get_data);
4315 	ASSERT0P(zilog->zl_last_lwb_opened);
4316 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
4317 
4318 	zilog->zl_get_data = get_data;
4319 	zilog->zl_sums = zil_sums;
4320 
4321 	return (zilog);
4322 }
4323 
4324 /*
4325  * Close an intent log.
4326  */
4327 void
4328 zil_close(zilog_t *zilog)
4329 {
4330 	lwb_t *lwb;
4331 	uint64_t txg;
4332 
4333 	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
4334 		if (zil_commit_flags(zilog, 0, ZIL_COMMIT_NOW) != 0)
4335 			txg_wait_synced(zilog->zl_dmu_pool, 0);
4336 	} else {
4337 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
4338 		ASSERT0(zilog->zl_dirty_max_txg);
4339 		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
4340 	}
4341 
4342 	mutex_enter(&zilog->zl_lock);
4343 	txg = zilog->zl_dirty_max_txg;
4344 	lwb = list_tail(&zilog->zl_lwb_list);
4345 	if (lwb != NULL) {
4346 		txg = MAX(txg, lwb->lwb_alloc_txg);
4347 		txg = MAX(txg, lwb->lwb_max_txg);
4348 	}
4349 	mutex_exit(&zilog->zl_lock);
4350 
4351 	/*
4352 	 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
4353 	 * on the time when the dmu_tx transaction is assigned in
4354 	 * zil_lwb_write_issue().
4355 	 */
4356 	mutex_enter(&zilog->zl_lwb_io_lock);
4357 	txg = MAX(zilog->zl_lwb_max_issued_txg, txg);
4358 	mutex_exit(&zilog->zl_lwb_io_lock);
4359 
4360 	/*
4361 	 * We need to use txg_wait_synced() to wait until that txg is synced.
4362 	 * zil_sync() will guarantee all lwbs up to that txg have been
4363 	 * written out, flushed, and cleaned.
4364 	 */
4365 	if (txg != 0)
4366 		txg_wait_synced(zilog->zl_dmu_pool, txg);
4367 
4368 	if (zilog_is_dirty(zilog))
4369 		zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog,
4370 		    (u_longlong_t)txg);
4371 	if (txg < spa_freeze_txg(zilog->zl_spa))
4372 		VERIFY(!zilog_is_dirty(zilog));
4373 
4374 	zilog->zl_get_data = NULL;
4375 
4376 	/*
4377 	 * We should have only one lwb left on the list; remove it now.
4378 	 */
4379 	mutex_enter(&zilog->zl_lock);
4380 	lwb = list_remove_head(&zilog->zl_lwb_list);
4381 	if (lwb != NULL) {
4382 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
4383 		ASSERT3S(lwb->lwb_state, ==, LWB_STATE_NEW);
4384 		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
4385 		zil_free_lwb(zilog, lwb);
4386 	}
4387 	mutex_exit(&zilog->zl_lock);
4388 }
4389 
4390 static const char *suspend_tag = "zil suspending";
4391 
4392 /*
4393  * Suspend an intent log.  While in suspended mode, we still honor
4394  * synchronous semantics, but we rely on txg_wait_synced() to do it.
4395  * On old version pools, we suspend the log briefly when taking a
4396  * snapshot so that it will have an empty intent log.
4397  *
4398  * Long holds are not really intended to be used the way we do here --
4399  * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
4400  * could fail.  Therefore we take pains to only put a long hold if it is
4401  * actually necessary.  Fortunately, it will only be necessary if the
4402  * objset is currently mounted (or the ZVOL equivalent).  In that case it
4403  * will already have a long hold, so we are not really making things any worse.
4404  *
4405  * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
4406  * zvol_state_t), and use their mechanism to prevent their hold from being
4407  * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
4408  * very little gain.
4409  *
4410  * if cookiep == NULL, this does both the suspend & resume.
4411  * Otherwise, it returns with the dataset "long held", and the cookie
4412  * should be passed into zil_resume().
4413  */
4414 int
4415 zil_suspend(const char *osname, void **cookiep)
4416 {
4417 	objset_t *os;
4418 	zilog_t *zilog;
4419 	const zil_header_t *zh;
4420 	int error;
4421 
4422 	error = dmu_objset_hold(osname, suspend_tag, &os);
4423 	if (error != 0)
4424 		return (error);
4425 	zilog = dmu_objset_zil(os);
4426 
4427 	mutex_enter(&zilog->zl_lock);
4428 	zh = zilog->zl_header;
4429 
4430 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
4431 		mutex_exit(&zilog->zl_lock);
4432 		dmu_objset_rele(os, suspend_tag);
4433 		return (SET_ERROR(EBUSY));
4434 	}
4435 
4436 	if (zilog->zl_restart_txg > 0) {
4437 		/*
4438 		 * ZIL crashed. It effectively _is_ suspended, but callers
4439 		 * are usually trying to make sure it's empty on-disk, which
4440 		 * we can't guarantee right now.
4441 		 */
4442 		mutex_exit(&zilog->zl_lock);
4443 		dmu_objset_rele(os, suspend_tag);
4444 		return (SET_ERROR(EBUSY));
4445 	}
4446 
4447 	/*
4448 	 * Don't put a long hold in the cases where we can avoid it.  This
4449 	 * is when there is no cookie so we are doing a suspend & resume
4450 	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
4451 	 * for the suspend because it's already suspended, or there's no ZIL.
4452 	 */
4453 	if (cookiep == NULL && !zilog->zl_suspending &&
4454 	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
4455 		mutex_exit(&zilog->zl_lock);
4456 		dmu_objset_rele(os, suspend_tag);
4457 		return (0);
4458 	}
4459 
4460 	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
4461 	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
4462 
4463 	zilog->zl_suspend++;
4464 
4465 	if (zilog->zl_suspend > 1) {
4466 		/*
4467 		 * Someone else is already suspending it.
4468 		 * Just wait for them to finish.
4469 		 */
4470 
4471 		while (zilog->zl_suspending)
4472 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
4473 		mutex_exit(&zilog->zl_lock);
4474 
4475 		if (cookiep == NULL)
4476 			zil_resume(os);
4477 		else
4478 			*cookiep = os;
4479 
4480 		if (zilog->zl_restart_txg > 0)
4481 			/* ZIL crashed while we were waiting. */
4482 			return (SET_ERROR(EBUSY));
4483 
4484 		return (0);
4485 	}
4486 
4487 	/*
4488 	 * If there is no pointer to an on-disk block, this ZIL must not
4489 	 * be active (e.g. filesystem not mounted), so there's nothing
4490 	 * to clean up.
4491 	 */
4492 	if (BP_IS_HOLE(&zh->zh_log)) {
4493 		ASSERT(cookiep != NULL); /* fast path already handled */
4494 
4495 		*cookiep = os;
4496 		mutex_exit(&zilog->zl_lock);
4497 		return (0);
4498 	}
4499 
4500 	/*
4501 	 * The ZIL has work to do. Ensure that the associated encryption
4502 	 * key will remain mapped while we are committing the log by
4503 	 * grabbing a reference to it. If the key isn't loaded we have no
4504 	 * choice but to return an error until the wrapping key is loaded.
4505 	 */
4506 	if (os->os_encrypted &&
4507 	    dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
4508 		zilog->zl_suspend--;
4509 		mutex_exit(&zilog->zl_lock);
4510 		dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
4511 		dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
4512 		return (SET_ERROR(EACCES));
4513 	}
4514 
4515 	zilog->zl_suspending = B_TRUE;
4516 	mutex_exit(&zilog->zl_lock);
4517 
4518 	/*
4519 	 * We need to use zil_commit_impl to ensure we wait for all
4520 	 * LWB_STATE_OPENED, _CLOSED and _READY lwbs to be committed
4521 	 * to disk before proceeding. If we used zil_commit instead, it
4522 	 * would just call txg_wait_synced(), because zl_suspend is set.
4523 	 * txg_wait_synced() doesn't wait for these lwb's to be
4524 	 * LWB_STATE_FLUSH_DONE before returning.
4525 	 *
4526 	 * However, zil_commit_impl() itself can return an error if any of the
4527 	 * lwbs fail, or the pool suspends in the fallback
4528 	 * txg_wait_sync_flushed(), which affects what we do next, so we
4529 	 * capture that error.
4530 	 */
4531 	error = zil_commit_impl(zilog, 0);
4532 	if (error == ESHUTDOWN)
4533 		/* zil_commit_impl() has called zil_crash() already */
4534 		error = SET_ERROR(EBUSY);
4535 
4536 	/*
4537 	 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
4538 	 * use txg_wait_synced() to ensure the data from the zilog has
4539 	 * migrated to the main pool before calling zil_destroy().
4540 	 */
4541 	if (error == 0) {
4542 		error = txg_wait_synced_flags(zilog->zl_dmu_pool, 0,
4543 		    TXG_WAIT_SUSPEND);
4544 		if (error != 0) {
4545 			ASSERT3U(error, ==, ESHUTDOWN);
4546 			zil_crash(zilog);
4547 			error = SET_ERROR(EBUSY);
4548 		}
4549 	}
4550 
4551 	if (error == 0)
4552 		zil_destroy(zilog, B_FALSE);
4553 
4554 	mutex_enter(&zilog->zl_lock);
4555 	zilog->zl_suspending = B_FALSE;
4556 	cv_broadcast(&zilog->zl_cv_suspend);
4557 	mutex_exit(&zilog->zl_lock);
4558 
4559 	if (os->os_encrypted)
4560 		dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
4561 
4562 	if (cookiep == NULL)
4563 		zil_resume(os);
4564 	else
4565 		*cookiep = os;
4566 
4567 	return (error);
4568 }
4569 
4570 void
4571 zil_resume(void *cookie)
4572 {
4573 	objset_t *os = cookie;
4574 	zilog_t *zilog = dmu_objset_zil(os);
4575 
4576 	mutex_enter(&zilog->zl_lock);
4577 	ASSERT(zilog->zl_suspend != 0);
4578 	zilog->zl_suspend--;
4579 	mutex_exit(&zilog->zl_lock);
4580 	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
4581 	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
4582 }
4583 
4584 typedef struct zil_replay_arg {
4585 	zil_replay_func_t *const *zr_replay;
4586 	void		*zr_arg;
4587 	boolean_t	zr_byteswap;
4588 	char		*zr_lr;
4589 } zil_replay_arg_t;
4590 
4591 static int
4592 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
4593 {
4594 	char name[ZFS_MAX_DATASET_NAME_LEN];
4595 
4596 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
4597 
4598 	dmu_objset_name(zilog->zl_os, name);
4599 
4600 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
4601 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
4602 	    (u_longlong_t)lr->lrc_seq,
4603 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
4604 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
4605 
4606 	return (error);
4607 }
4608 
4609 static int
4610 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
4611     uint64_t claim_txg)
4612 {
4613 	zil_replay_arg_t *zr = zra;
4614 	const zil_header_t *zh = zilog->zl_header;
4615 	uint64_t reclen = lr->lrc_reclen;
4616 	uint64_t txtype = lr->lrc_txtype;
4617 	int error = 0;
4618 
4619 	zilog->zl_replaying_seq = lr->lrc_seq;
4620 
4621 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
4622 		return (0);
4623 
4624 	if (lr->lrc_txg < claim_txg)		/* already committed */
4625 		return (0);
4626 
4627 	/* Strip case-insensitive bit, still present in log record */
4628 	txtype &= ~TX_CI;
4629 
4630 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
4631 		return (zil_replay_error(zilog, lr, EINVAL));
4632 
4633 	/*
4634 	 * If this record type can be logged out of order, the object
4635 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
4636 	 */
4637 	if (TX_OOO(txtype)) {
4638 		error = dmu_object_info(zilog->zl_os,
4639 		    LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
4640 		if (error == ENOENT || error == EEXIST)
4641 			return (0);
4642 	}
4643 
4644 	/*
4645 	 * Make a copy of the data so we can revise and extend it.
4646 	 */
4647 	memcpy(zr->zr_lr, lr, reclen);
4648 
4649 	/*
4650 	 * If this is a TX_WRITE with a blkptr, suck in the data.
4651 	 */
4652 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
4653 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
4654 		    zr->zr_lr + reclen);
4655 		if (error != 0)
4656 			return (zil_replay_error(zilog, lr, error));
4657 	}
4658 
4659 	/*
4660 	 * The log block containing this lr may have been byteswapped
4661 	 * so that we can easily examine common fields like lrc_txtype.
4662 	 * However, the log is a mix of different record types, and only the
4663 	 * replay vectors know how to byteswap their records.  Therefore, if
4664 	 * the lr was byteswapped, undo it before invoking the replay vector.
4665 	 */
4666 	if (zr->zr_byteswap)
4667 		byteswap_uint64_array(zr->zr_lr, reclen);
4668 
4669 	/*
4670 	 * We must now do two things atomically: replay this log record,
4671 	 * and update the log header sequence number to reflect the fact that
4672 	 * we did so. At the end of each replay function the sequence number
4673 	 * is updated if we are in replay mode.
4674 	 */
4675 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
4676 	if (error != 0) {
4677 		/*
4678 		 * The DMU's dnode layer doesn't see removes until the txg
4679 		 * commits, so a subsequent claim can spuriously fail with
4680 		 * EEXIST. So if we receive any error we try syncing out
4681 		 * any removes then retry the transaction.  Note that we
4682 		 * specify B_FALSE for byteswap now, so we don't do it twice.
4683 		 */
4684 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
4685 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
4686 		if (error != 0)
4687 			return (zil_replay_error(zilog, lr, error));
4688 	}
4689 	return (0);
4690 }
4691 
4692 static int
4693 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
4694 {
4695 	(void) bp, (void) arg, (void) claim_txg;
4696 
4697 	zilog->zl_replay_blks++;
4698 
4699 	return (0);
4700 }
4701 
4702 /*
4703  * If this dataset has a non-empty intent log, replay it and destroy it.
4704  * Return B_TRUE if there were any entries to replay.
4705  */
4706 boolean_t
4707 zil_replay(objset_t *os, void *arg,
4708     zil_replay_func_t *const replay_func[TX_MAX_TYPE])
4709 {
4710 	zilog_t *zilog = dmu_objset_zil(os);
4711 	const zil_header_t *zh = zilog->zl_header;
4712 	zil_replay_arg_t zr;
4713 
4714 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
4715 		return (zil_destroy(zilog, B_TRUE));
4716 	}
4717 
4718 	zr.zr_replay = replay_func;
4719 	zr.zr_arg = arg;
4720 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
4721 	zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
4722 
4723 	/*
4724 	 * Wait for in-progress removes to sync before starting replay.
4725 	 */
4726 	txg_wait_synced(zilog->zl_dmu_pool, 0);
4727 
4728 	zilog->zl_replay = B_TRUE;
4729 	zilog->zl_replay_time = ddi_get_lbolt();
4730 	ASSERT0(zilog->zl_replay_blks);
4731 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
4732 	    zh->zh_claim_txg, B_TRUE);
4733 	vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
4734 
4735 	zil_destroy(zilog, B_FALSE);
4736 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
4737 	zilog->zl_replay = B_FALSE;
4738 
4739 	return (B_TRUE);
4740 }
4741 
4742 boolean_t
4743 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
4744 {
4745 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
4746 		return (B_TRUE);
4747 
4748 	if (zilog->zl_replay) {
4749 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
4750 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
4751 		    zilog->zl_replaying_seq;
4752 		return (B_TRUE);
4753 	}
4754 
4755 	return (B_FALSE);
4756 }
4757 
4758 int
4759 zil_reset(const char *osname, void *arg)
4760 {
4761 	(void) arg;
4762 
4763 	int error = zil_suspend(osname, NULL);
4764 	/* EACCES means crypto key not loaded */
4765 	if ((error == EACCES) || (error == EBUSY))
4766 		return (SET_ERROR(error));
4767 	if (error != 0)
4768 		return (SET_ERROR(EEXIST));
4769 	return (0);
4770 }
4771 
4772 EXPORT_SYMBOL(zil_alloc);
4773 EXPORT_SYMBOL(zil_free);
4774 EXPORT_SYMBOL(zil_open);
4775 EXPORT_SYMBOL(zil_close);
4776 EXPORT_SYMBOL(zil_replay);
4777 EXPORT_SYMBOL(zil_replaying);
4778 EXPORT_SYMBOL(zil_destroy);
4779 EXPORT_SYMBOL(zil_destroy_sync);
4780 EXPORT_SYMBOL(zil_itx_create);
4781 EXPORT_SYMBOL(zil_itx_destroy);
4782 EXPORT_SYMBOL(zil_itx_assign);
4783 EXPORT_SYMBOL(zil_commit);
4784 EXPORT_SYMBOL(zil_claim);
4785 EXPORT_SYMBOL(zil_check_log_chain);
4786 EXPORT_SYMBOL(zil_sync);
4787 EXPORT_SYMBOL(zil_clean);
4788 EXPORT_SYMBOL(zil_suspend);
4789 EXPORT_SYMBOL(zil_resume);
4790 EXPORT_SYMBOL(zil_lwb_add_block);
4791 EXPORT_SYMBOL(zil_bp_tree_add);
4792 EXPORT_SYMBOL(zil_set_sync);
4793 EXPORT_SYMBOL(zil_set_logbias);
4794 EXPORT_SYMBOL(zil_sums_init);
4795 EXPORT_SYMBOL(zil_sums_fini);
4796 EXPORT_SYMBOL(zil_kstat_values_update);
4797 
4798 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, UINT, ZMOD_RW,
4799 	"ZIL block open timeout percentage");
4800 
4801 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
4802 	"Disable intent logging replay");
4803 
4804 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
4805 	"Disable ZIL cache flushes");
4806 
4807 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, U64, ZMOD_RW,
4808 	"Limit in bytes slog sync writes per commit");
4809 
4810 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, UINT, ZMOD_RW,
4811 	"Limit in bytes of ZIL log block size");
4812 
4813 ZFS_MODULE_PARAM(zfs_zil, zil_, maxcopied, UINT, ZMOD_RW,
4814 	"Limit in bytes WR_COPIED size");
4815 
4816 ZFS_MODULE_PARAM(zfs, zfs_, immediate_write_sz, UINT, ZMOD_RW,
4817 	"Largest write size to store data into ZIL");
4818 
4819 ZFS_MODULE_PARAM(zfs_zil, zil_, special_is_slog, INT, ZMOD_RW,
4820 	"Treat special vdevs as SLOG");
4821