1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 */
26
27/* Portions Copyright 2010 Robert Milkowski */
28
29#include <sys/zfs_context.h>
30#include <sys/spa.h>
31#include <sys/dmu.h>
32#include <sys/zap.h>
33#include <sys/arc.h>
34#include <sys/stat.h>
35#include <sys/resource.h>
36#include <sys/zil.h>
37#include <sys/zil_impl.h>
38#include <sys/dsl_dataset.h>
39#include <sys/vdev_impl.h>
40#include <sys/dmu_tx.h>
41#include <sys/dsl_pool.h>
42#include <sys/abd.h>
43
44/*
45 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
46 * calls that change the file system. Each itx has enough information to
47 * be able to replay them after a system crash, power loss, or
48 * equivalent failure mode. These are stored in memory until either:
49 *
50 *   1. they are committed to the pool by the DMU transaction group
51 *      (txg), at which point they can be discarded; or
52 *   2. they are committed to the on-disk ZIL for the dataset being
53 *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
54 *      requirement).
55 *
56 * In the event of a crash or power loss, the itxs contained by each
57 * dataset's on-disk ZIL will be replayed when that dataset is first
58 * instantianted (e.g. if the dataset is a normal fileystem, when it is
59 * first mounted).
60 *
61 * As hinted at above, there is one ZIL per dataset (both the in-memory
62 * representation, and the on-disk representation). The on-disk format
63 * consists of 3 parts:
64 *
65 * 	- a single, per-dataset, ZIL header; which points to a chain of
66 * 	- zero or more ZIL blocks; each of which contains
67 * 	- zero or more ZIL records
68 *
69 * A ZIL record holds the information necessary to replay a single
70 * system call transaction. A ZIL block can hold many ZIL records, and
71 * the blocks are chained together, similarly to a singly linked list.
72 *
73 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
74 * block in the chain, and the ZIL header points to the first block in
75 * the chain.
76 *
77 * Note, there is not a fixed place in the pool to hold these ZIL
78 * blocks; they are dynamically allocated and freed as needed from the
79 * blocks available on the pool, though they can be preferentially
80 * allocated from a dedicated "log" vdev.
81 */
82
83/*
84 * This controls the amount of time that a ZIL block (lwb) will remain
85 * "open" when it isn't "full", and it has a thread waiting for it to be
86 * committed to stable storage. Please refer to the zil_commit_waiter()
87 * function (and the comments within it) for more details.
88 */
89int zfs_commit_timeout_pct = 5;
90
91/*
92 * Disable intent logging replay.  This global ZIL switch affects all pools.
93 */
94int zil_replay_disable = 0;
95
96/*
97 * Tunable parameter for debugging or performance analysis.  Setting
98 * zfs_nocacheflush will cause corruption on power loss if a volatile
99 * out-of-order write cache is enabled.
100 */
101boolean_t zfs_nocacheflush = B_FALSE;
102
103/*
104 * Limit SLOG write size per commit executed with synchronous priority.
105 * Any writes above that will be executed with lower (asynchronous) priority
106 * to limit potential SLOG device abuse by single active ZIL writer.
107 */
108uint64_t zil_slog_bulk = 768 * 1024;
109
110static kmem_cache_t *zil_lwb_cache;
111static kmem_cache_t *zil_zcw_cache;
112
113static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
114
115#define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
116    sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
117
118static int
119zil_bp_compare(const void *x1, const void *x2)
120{
121	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
122	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
123
124	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
125		return (-1);
126	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
127		return (1);
128
129	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
130		return (-1);
131	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
132		return (1);
133
134	return (0);
135}
136
137static void
138zil_bp_tree_init(zilog_t *zilog)
139{
140	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
141	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
142}
143
144static void
145zil_bp_tree_fini(zilog_t *zilog)
146{
147	avl_tree_t *t = &zilog->zl_bp_tree;
148	zil_bp_node_t *zn;
149	void *cookie = NULL;
150
151	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
152		kmem_free(zn, sizeof (zil_bp_node_t));
153
154	avl_destroy(t);
155}
156
157int
158zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
159{
160	avl_tree_t *t = &zilog->zl_bp_tree;
161	const dva_t *dva;
162	zil_bp_node_t *zn;
163	avl_index_t where;
164
165	if (BP_IS_EMBEDDED(bp))
166		return (0);
167
168	dva = BP_IDENTITY(bp);
169
170	if (avl_find(t, dva, &where) != NULL)
171		return (SET_ERROR(EEXIST));
172
173	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
174	zn->zn_dva = *dva;
175	avl_insert(t, zn, where);
176
177	return (0);
178}
179
180static zil_header_t *
181zil_header_in_syncing_context(zilog_t *zilog)
182{
183	return ((zil_header_t *)zilog->zl_header);
184}
185
186static void
187zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
188{
189	zio_cksum_t *zc = &bp->blk_cksum;
190
191	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
192	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
193	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
194	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
195}
196
197/*
198 * Read a log block and make sure it's valid.
199 */
200static int
201zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
202    char **end)
203{
204	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
205	arc_flags_t aflags = ARC_FLAG_WAIT;
206	arc_buf_t *abuf = NULL;
207	zbookmark_phys_t zb;
208	int error;
209
210	if (zilog->zl_header->zh_claim_txg == 0)
211		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
212
213	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
214		zio_flags |= ZIO_FLAG_SPECULATIVE;
215
216	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
217	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
218
219	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
220	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
221
222	if (error == 0) {
223		zio_cksum_t cksum = bp->blk_cksum;
224
225		/*
226		 * Validate the checksummed log block.
227		 *
228		 * Sequence numbers should be... sequential.  The checksum
229		 * verifier for the next block should be bp's checksum plus 1.
230		 *
231		 * Also check the log chain linkage and size used.
232		 */
233		cksum.zc_word[ZIL_ZC_SEQ]++;
234
235		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
236			zil_chain_t *zilc = abuf->b_data;
237			char *lr = (char *)(zilc + 1);
238			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
239
240			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
241			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
242				error = SET_ERROR(ECKSUM);
243			} else {
244				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
245				bcopy(lr, dst, len);
246				*end = (char *)dst + len;
247				*nbp = zilc->zc_next_blk;
248			}
249		} else {
250			char *lr = abuf->b_data;
251			uint64_t size = BP_GET_LSIZE(bp);
252			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
253
254			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
255			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
256			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
257				error = SET_ERROR(ECKSUM);
258			} else {
259				ASSERT3U(zilc->zc_nused, <=,
260				    SPA_OLD_MAXBLOCKSIZE);
261				bcopy(lr, dst, zilc->zc_nused);
262				*end = (char *)dst + zilc->zc_nused;
263				*nbp = zilc->zc_next_blk;
264			}
265		}
266
267		arc_buf_destroy(abuf, &abuf);
268	}
269
270	return (error);
271}
272
273/*
274 * Read a TX_WRITE log data block.
275 */
276static int
277zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
278{
279	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
280	const blkptr_t *bp = &lr->lr_blkptr;
281	arc_flags_t aflags = ARC_FLAG_WAIT;
282	arc_buf_t *abuf = NULL;
283	zbookmark_phys_t zb;
284	int error;
285
286	if (BP_IS_HOLE(bp)) {
287		if (wbuf != NULL)
288			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
289		return (0);
290	}
291
292	if (zilog->zl_header->zh_claim_txg == 0)
293		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
294
295	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
296	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
297
298	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
299	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
300
301	if (error == 0) {
302		if (wbuf != NULL)
303			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
304		arc_buf_destroy(abuf, &abuf);
305	}
306
307	return (error);
308}
309
310/*
311 * Parse the intent log, and call parse_func for each valid record within.
312 */
313int
314zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
315    zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
316{
317	const zil_header_t *zh = zilog->zl_header;
318	boolean_t claimed = !!zh->zh_claim_txg;
319	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
320	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
321	uint64_t max_blk_seq = 0;
322	uint64_t max_lr_seq = 0;
323	uint64_t blk_count = 0;
324	uint64_t lr_count = 0;
325	blkptr_t blk, next_blk;
326	char *lrbuf, *lrp;
327	int error = 0;
328
329	/*
330	 * Old logs didn't record the maximum zh_claim_lr_seq.
331	 */
332	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
333		claim_lr_seq = UINT64_MAX;
334
335	/*
336	 * Starting at the block pointed to by zh_log we read the log chain.
337	 * For each block in the chain we strongly check that block to
338	 * ensure its validity.  We stop when an invalid block is found.
339	 * For each block pointer in the chain we call parse_blk_func().
340	 * For each record in each valid block we call parse_lr_func().
341	 * If the log has been claimed, stop if we encounter a sequence
342	 * number greater than the highest claimed sequence number.
343	 */
344	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
345	zil_bp_tree_init(zilog);
346
347	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
348		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
349		int reclen;
350		char *end;
351
352		if (blk_seq > claim_blk_seq)
353			break;
354		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
355			break;
356		ASSERT3U(max_blk_seq, <, blk_seq);
357		max_blk_seq = blk_seq;
358		blk_count++;
359
360		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
361			break;
362
363		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
364		if (error != 0)
365			break;
366
367		for (lrp = lrbuf; lrp < end; lrp += reclen) {
368			lr_t *lr = (lr_t *)lrp;
369			reclen = lr->lrc_reclen;
370			ASSERT3U(reclen, >=, sizeof (lr_t));
371			if (lr->lrc_seq > claim_lr_seq)
372				goto done;
373			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
374				goto done;
375			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
376			max_lr_seq = lr->lrc_seq;
377			lr_count++;
378		}
379	}
380done:
381	zilog->zl_parse_error = error;
382	zilog->zl_parse_blk_seq = max_blk_seq;
383	zilog->zl_parse_lr_seq = max_lr_seq;
384	zilog->zl_parse_blk_count = blk_count;
385	zilog->zl_parse_lr_count = lr_count;
386
387	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
388	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
389
390	zil_bp_tree_fini(zilog);
391	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
392
393	return (error);
394}
395
396static int
397zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
398{
399	/*
400	 * Claim log block if not already committed and not already claimed.
401	 * If tx == NULL, just verify that the block is claimable.
402	 */
403	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
404	    zil_bp_tree_add(zilog, bp) != 0)
405		return (0);
406
407	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
408	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
409	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
410}
411
412static int
413zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
414{
415	lr_write_t *lr = (lr_write_t *)lrc;
416	int error;
417
418	if (lrc->lrc_txtype != TX_WRITE)
419		return (0);
420
421	/*
422	 * If the block is not readable, don't claim it.  This can happen
423	 * in normal operation when a log block is written to disk before
424	 * some of the dmu_sync() blocks it points to.  In this case, the
425	 * transaction cannot have been committed to anyone (we would have
426	 * waited for all writes to be stable first), so it is semantically
427	 * correct to declare this the end of the log.
428	 */
429	if (lr->lr_blkptr.blk_birth >= first_txg &&
430	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
431		return (error);
432	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
433}
434
435/* ARGSUSED */
436static int
437zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
438{
439	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
440
441	return (0);
442}
443
444static int
445zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
446{
447	lr_write_t *lr = (lr_write_t *)lrc;
448	blkptr_t *bp = &lr->lr_blkptr;
449
450	/*
451	 * If we previously claimed it, we need to free it.
452	 */
453	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
454	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
455	    !BP_IS_HOLE(bp))
456		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
457
458	return (0);
459}
460
461static int
462zil_lwb_vdev_compare(const void *x1, const void *x2)
463{
464	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
465	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
466
467	if (v1 < v2)
468		return (-1);
469	if (v1 > v2)
470		return (1);
471
472	return (0);
473}
474
475static lwb_t *
476zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
477{
478	lwb_t *lwb;
479
480	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
481	lwb->lwb_zilog = zilog;
482	lwb->lwb_blk = *bp;
483	lwb->lwb_slog = slog;
484	lwb->lwb_state = LWB_STATE_CLOSED;
485	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
486	lwb->lwb_max_txg = txg;
487	lwb->lwb_write_zio = NULL;
488	lwb->lwb_root_zio = NULL;
489	lwb->lwb_tx = NULL;
490	lwb->lwb_issued_timestamp = 0;
491	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
492		lwb->lwb_nused = sizeof (zil_chain_t);
493		lwb->lwb_sz = BP_GET_LSIZE(bp);
494	} else {
495		lwb->lwb_nused = 0;
496		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
497	}
498
499	mutex_enter(&zilog->zl_lock);
500	list_insert_tail(&zilog->zl_lwb_list, lwb);
501	mutex_exit(&zilog->zl_lock);
502
503	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
504	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
505	VERIFY(list_is_empty(&lwb->lwb_waiters));
506
507	return (lwb);
508}
509
510static void
511zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
512{
513	ASSERT(MUTEX_HELD(&zilog->zl_lock));
514	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
515	VERIFY(list_is_empty(&lwb->lwb_waiters));
516	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
517	ASSERT3P(lwb->lwb_write_zio, ==, NULL);
518	ASSERT3P(lwb->lwb_root_zio, ==, NULL);
519	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
520	ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
521	    lwb->lwb_state == LWB_STATE_DONE);
522
523	/*
524	 * Clear the zilog's field to indicate this lwb is no longer
525	 * valid, and prevent use-after-free errors.
526	 */
527	if (zilog->zl_last_lwb_opened == lwb)
528		zilog->zl_last_lwb_opened = NULL;
529
530	kmem_cache_free(zil_lwb_cache, lwb);
531}
532
533/*
534 * Called when we create in-memory log transactions so that we know
535 * to cleanup the itxs at the end of spa_sync().
536 */
537void
538zilog_dirty(zilog_t *zilog, uint64_t txg)
539{
540	dsl_pool_t *dp = zilog->zl_dmu_pool;
541	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
542
543	ASSERT(spa_writeable(zilog->zl_spa));
544
545	if (ds->ds_is_snapshot)
546		panic("dirtying snapshot!");
547
548	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
549		/* up the hold count until we can be written out */
550		dmu_buf_add_ref(ds->ds_dbuf, zilog);
551
552		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
553	}
554}
555
556/*
557 * Determine if the zil is dirty in the specified txg. Callers wanting to
558 * ensure that the dirty state does not change must hold the itxg_lock for
559 * the specified txg. Holding the lock will ensure that the zil cannot be
560 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
561 * state.
562 */
563boolean_t
564zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
565{
566	dsl_pool_t *dp = zilog->zl_dmu_pool;
567
568	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
569		return (B_TRUE);
570	return (B_FALSE);
571}
572
573/*
574 * Determine if the zil is dirty. The zil is considered dirty if it has
575 * any pending itx records that have not been cleaned by zil_clean().
576 */
577boolean_t
578zilog_is_dirty(zilog_t *zilog)
579{
580	dsl_pool_t *dp = zilog->zl_dmu_pool;
581
582	for (int t = 0; t < TXG_SIZE; t++) {
583		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
584			return (B_TRUE);
585	}
586	return (B_FALSE);
587}
588
589/*
590 * Create an on-disk intent log.
591 */
592static lwb_t *
593zil_create(zilog_t *zilog)
594{
595	const zil_header_t *zh = zilog->zl_header;
596	lwb_t *lwb = NULL;
597	uint64_t txg = 0;
598	dmu_tx_t *tx = NULL;
599	blkptr_t blk;
600	int error = 0;
601	boolean_t slog = FALSE;
602
603	/*
604	 * Wait for any previous destroy to complete.
605	 */
606	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
607
608	ASSERT(zh->zh_claim_txg == 0);
609	ASSERT(zh->zh_replay_seq == 0);
610
611	blk = zh->zh_log;
612
613	/*
614	 * Allocate an initial log block if:
615	 *    - there isn't one already
616	 *    - the existing block is the wrong endianess
617	 */
618	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
619		tx = dmu_tx_create(zilog->zl_os);
620		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
621		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
622		txg = dmu_tx_get_txg(tx);
623
624		if (!BP_IS_HOLE(&blk)) {
625			zio_free_zil(zilog->zl_spa, txg, &blk);
626			BP_ZERO(&blk);
627		}
628
629		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
630		    ZIL_MIN_BLKSZ, &slog);
631
632		if (error == 0)
633			zil_init_log_chain(zilog, &blk);
634	}
635
636	/*
637	 * Allocate a log write block (lwb) for the first log block.
638	 */
639	if (error == 0)
640		lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
641
642	/*
643	 * If we just allocated the first log block, commit our transaction
644	 * and wait for zil_sync() to stuff the block poiner into zh_log.
645	 * (zh is part of the MOS, so we cannot modify it in open context.)
646	 */
647	if (tx != NULL) {
648		dmu_tx_commit(tx);
649		txg_wait_synced(zilog->zl_dmu_pool, txg);
650	}
651
652	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
653
654	return (lwb);
655}
656
657/*
658 * In one tx, free all log blocks and clear the log header. If keep_first
659 * is set, then we're replaying a log with no content. We want to keep the
660 * first block, however, so that the first synchronous transaction doesn't
661 * require a txg_wait_synced() in zil_create(). We don't need to
662 * txg_wait_synced() here either when keep_first is set, because both
663 * zil_create() and zil_destroy() will wait for any in-progress destroys
664 * to complete.
665 */
666void
667zil_destroy(zilog_t *zilog, boolean_t keep_first)
668{
669	const zil_header_t *zh = zilog->zl_header;
670	lwb_t *lwb;
671	dmu_tx_t *tx;
672	uint64_t txg;
673
674	/*
675	 * Wait for any previous destroy to complete.
676	 */
677	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
678
679	zilog->zl_old_header = *zh;		/* debugging aid */
680
681	if (BP_IS_HOLE(&zh->zh_log))
682		return;
683
684	tx = dmu_tx_create(zilog->zl_os);
685	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
686	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
687	txg = dmu_tx_get_txg(tx);
688
689	mutex_enter(&zilog->zl_lock);
690
691	ASSERT3U(zilog->zl_destroy_txg, <, txg);
692	zilog->zl_destroy_txg = txg;
693	zilog->zl_keep_first = keep_first;
694
695	if (!list_is_empty(&zilog->zl_lwb_list)) {
696		ASSERT(zh->zh_claim_txg == 0);
697		VERIFY(!keep_first);
698		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
699			list_remove(&zilog->zl_lwb_list, lwb);
700			if (lwb->lwb_buf != NULL)
701				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
702			zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
703			zil_free_lwb(zilog, lwb);
704		}
705	} else if (!keep_first) {
706		zil_destroy_sync(zilog, tx);
707	}
708	mutex_exit(&zilog->zl_lock);
709
710	dmu_tx_commit(tx);
711}
712
713void
714zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
715{
716	ASSERT(list_is_empty(&zilog->zl_lwb_list));
717	(void) zil_parse(zilog, zil_free_log_block,
718	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
719}
720
721int
722zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
723{
724	dmu_tx_t *tx = txarg;
725	uint64_t first_txg = dmu_tx_get_txg(tx);
726	zilog_t *zilog;
727	zil_header_t *zh;
728	objset_t *os;
729	int error;
730
731	error = dmu_objset_own_obj(dp, ds->ds_object,
732	    DMU_OST_ANY, B_FALSE, FTAG, &os);
733	if (error != 0) {
734		/*
735		 * EBUSY indicates that the objset is inconsistent, in which
736		 * case it can not have a ZIL.
737		 */
738		if (error != EBUSY) {
739			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
740			    (unsigned long long)ds->ds_object, error);
741		}
742		return (0);
743	}
744
745	zilog = dmu_objset_zil(os);
746	zh = zil_header_in_syncing_context(zilog);
747
748	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
749		if (!BP_IS_HOLE(&zh->zh_log))
750			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
751		BP_ZERO(&zh->zh_log);
752		dsl_dataset_dirty(dmu_objset_ds(os), tx);
753		dmu_objset_disown(os, FTAG);
754		return (0);
755	}
756
757	/*
758	 * Claim all log blocks if we haven't already done so, and remember
759	 * the highest claimed sequence number.  This ensures that if we can
760	 * read only part of the log now (e.g. due to a missing device),
761	 * but we can read the entire log later, we will not try to replay
762	 * or destroy beyond the last block we successfully claimed.
763	 */
764	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
765	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
766		(void) zil_parse(zilog, zil_claim_log_block,
767		    zil_claim_log_record, tx, first_txg);
768		zh->zh_claim_txg = first_txg;
769		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
770		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
771		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
772			zh->zh_flags |= ZIL_REPLAY_NEEDED;
773		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
774		dsl_dataset_dirty(dmu_objset_ds(os), tx);
775	}
776
777	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
778	dmu_objset_disown(os, FTAG);
779	return (0);
780}
781
782/*
783 * Check the log by walking the log chain.
784 * Checksum errors are ok as they indicate the end of the chain.
785 * Any other error (no device or read failure) returns an error.
786 */
787/* ARGSUSED */
788int
789zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
790{
791	zilog_t *zilog;
792	objset_t *os;
793	blkptr_t *bp;
794	int error;
795
796	ASSERT(tx == NULL);
797
798	error = dmu_objset_from_ds(ds, &os);
799	if (error != 0) {
800		cmn_err(CE_WARN, "can't open objset %llu, error %d",
801		    (unsigned long long)ds->ds_object, error);
802		return (0);
803	}
804
805	zilog = dmu_objset_zil(os);
806	bp = (blkptr_t *)&zilog->zl_header->zh_log;
807
808	/*
809	 * Check the first block and determine if it's on a log device
810	 * which may have been removed or faulted prior to loading this
811	 * pool.  If so, there's no point in checking the rest of the log
812	 * as its content should have already been synced to the pool.
813	 */
814	if (!BP_IS_HOLE(bp)) {
815		vdev_t *vd;
816		boolean_t valid = B_TRUE;
817
818		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
819		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
820		if (vd->vdev_islog && vdev_is_dead(vd))
821			valid = vdev_log_state_valid(vd);
822		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
823
824		if (!valid)
825			return (0);
826	}
827
828	/*
829	 * Because tx == NULL, zil_claim_log_block() will not actually claim
830	 * any blocks, but just determine whether it is possible to do so.
831	 * In addition to checking the log chain, zil_claim_log_block()
832	 * will invoke zio_claim() with a done func of spa_claim_notify(),
833	 * which will update spa_max_claim_txg.  See spa_load() for details.
834	 */
835	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
836	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
837
838	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
839}
840
841/*
842 * When an itx is "skipped", this function is used to properly mark the
843 * waiter as "done, and signal any thread(s) waiting on it. An itx can
844 * be skipped (and not committed to an lwb) for a variety of reasons,
845 * one of them being that the itx was committed via spa_sync(), prior to
846 * it being committed to an lwb; this can happen if a thread calling
847 * zil_commit() is racing with spa_sync().
848 */
849static void
850zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
851{
852	mutex_enter(&zcw->zcw_lock);
853	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
854	zcw->zcw_done = B_TRUE;
855	cv_broadcast(&zcw->zcw_cv);
856	mutex_exit(&zcw->zcw_lock);
857}
858
859/*
860 * This function is used when the given waiter is to be linked into an
861 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
862 * At this point, the waiter will no longer be referenced by the itx,
863 * and instead, will be referenced by the lwb.
864 */
865static void
866zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
867{
868	/*
869	 * The lwb_waiters field of the lwb is protected by the zilog's
870	 * zl_lock, thus it must be held when calling this function.
871	 */
872	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
873
874	mutex_enter(&zcw->zcw_lock);
875	ASSERT(!list_link_active(&zcw->zcw_node));
876	ASSERT3P(zcw->zcw_lwb, ==, NULL);
877	ASSERT3P(lwb, !=, NULL);
878	ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
879	    lwb->lwb_state == LWB_STATE_ISSUED);
880
881	list_insert_tail(&lwb->lwb_waiters, zcw);
882	zcw->zcw_lwb = lwb;
883	mutex_exit(&zcw->zcw_lock);
884}
885
886/*
887 * This function is used when zio_alloc_zil() fails to allocate a ZIL
888 * block, and the given waiter must be linked to the "nolwb waiters"
889 * list inside of zil_process_commit_list().
890 */
891static void
892zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
893{
894	mutex_enter(&zcw->zcw_lock);
895	ASSERT(!list_link_active(&zcw->zcw_node));
896	ASSERT3P(zcw->zcw_lwb, ==, NULL);
897	list_insert_tail(nolwb, zcw);
898	mutex_exit(&zcw->zcw_lock);
899}
900
901void
902zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
903{
904	avl_tree_t *t = &lwb->lwb_vdev_tree;
905	avl_index_t where;
906	zil_vdev_node_t *zv, zvsearch;
907	int ndvas = BP_GET_NDVAS(bp);
908	int i;
909
910	if (zfs_nocacheflush)
911		return;
912
913	mutex_enter(&lwb->lwb_vdev_lock);
914	for (i = 0; i < ndvas; i++) {
915		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
916		if (avl_find(t, &zvsearch, &where) == NULL) {
917			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
918			zv->zv_vdev = zvsearch.zv_vdev;
919			avl_insert(t, zv, where);
920		}
921	}
922	mutex_exit(&lwb->lwb_vdev_lock);
923}
924
925void
926zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
927{
928	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
929}
930
931/*
932 * This function is a called after all VDEVs associated with a given lwb
933 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
934 * as the lwb write completes, if "zfs_nocacheflush" is set.
935 *
936 * The intention is for this function to be called as soon as the
937 * contents of an lwb are considered "stable" on disk, and will survive
938 * any sudden loss of power. At this point, any threads waiting for the
939 * lwb to reach this state are signalled, and the "waiter" structures
940 * are marked "done".
941 */
942static void
943zil_lwb_flush_vdevs_done(zio_t *zio)
944{
945	lwb_t *lwb = zio->io_private;
946	zilog_t *zilog = lwb->lwb_zilog;
947	dmu_tx_t *tx = lwb->lwb_tx;
948	zil_commit_waiter_t *zcw;
949
950	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
951
952	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
953
954	mutex_enter(&zilog->zl_lock);
955
956	/*
957	 * Ensure the lwb buffer pointer is cleared before releasing the
958	 * txg. If we have had an allocation failure and the txg is
959	 * waiting to sync then we want zil_sync() to remove the lwb so
960	 * that it's not picked up as the next new one in
961	 * zil_process_commit_list(). zil_sync() will only remove the
962	 * lwb if lwb_buf is null.
963	 */
964	lwb->lwb_buf = NULL;
965	lwb->lwb_tx = NULL;
966
967	ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
968	zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
969
970	lwb->lwb_root_zio = NULL;
971	lwb->lwb_state = LWB_STATE_DONE;
972
973	if (zilog->zl_last_lwb_opened == lwb) {
974		/*
975		 * Remember the highest committed log sequence number
976		 * for ztest. We only update this value when all the log
977		 * writes succeeded, because ztest wants to ASSERT that
978		 * it got the whole log chain.
979		 */
980		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
981	}
982
983	while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
984		mutex_enter(&zcw->zcw_lock);
985
986		ASSERT(list_link_active(&zcw->zcw_node));
987		list_remove(&lwb->lwb_waiters, zcw);
988
989		ASSERT3P(zcw->zcw_lwb, ==, lwb);
990		zcw->zcw_lwb = NULL;
991
992		zcw->zcw_zio_error = zio->io_error;
993
994		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
995		zcw->zcw_done = B_TRUE;
996		cv_broadcast(&zcw->zcw_cv);
997
998		mutex_exit(&zcw->zcw_lock);
999	}
1000
1001	mutex_exit(&zilog->zl_lock);
1002
1003	/*
1004	 * Now that we've written this log block, we have a stable pointer
1005	 * to the next block in the chain, so it's OK to let the txg in
1006	 * which we allocated the next block sync.
1007	 */
1008	dmu_tx_commit(tx);
1009}
1010
1011/*
1012 * This is called when an lwb write completes. This means, this specific
1013 * lwb was written to disk, and all dependent lwb have also been
1014 * written to disk.
1015 *
1016 * At this point, a DKIOCFLUSHWRITECACHE command hasn't been issued to
1017 * the VDEVs involved in writing out this specific lwb. The lwb will be
1018 * "done" once zil_lwb_flush_vdevs_done() is called, which occurs in the
1019 * zio completion callback for the lwb's root zio.
1020 */
1021static void
1022zil_lwb_write_done(zio_t *zio)
1023{
1024	lwb_t *lwb = zio->io_private;
1025	spa_t *spa = zio->io_spa;
1026	zilog_t *zilog = lwb->lwb_zilog;
1027	avl_tree_t *t = &lwb->lwb_vdev_tree;
1028	void *cookie = NULL;
1029	zil_vdev_node_t *zv;
1030
1031	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1032
1033	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1034	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1035	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1036	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1037	ASSERT(!BP_IS_GANG(zio->io_bp));
1038	ASSERT(!BP_IS_HOLE(zio->io_bp));
1039	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1040
1041	abd_put(zio->io_abd);
1042
1043	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1044
1045	mutex_enter(&zilog->zl_lock);
1046	lwb->lwb_write_zio = NULL;
1047	mutex_exit(&zilog->zl_lock);
1048
1049	if (avl_numnodes(t) == 0)
1050		return;
1051
1052	/*
1053	 * If there was an IO error, we're not going to call zio_flush()
1054	 * on these vdevs, so we simply empty the tree and free the
1055	 * nodes. We avoid calling zio_flush() since there isn't any
1056	 * good reason for doing so, after the lwb block failed to be
1057	 * written out.
1058	 */
1059	if (zio->io_error != 0) {
1060		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1061			kmem_free(zv, sizeof (*zv));
1062		return;
1063	}
1064
1065	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1066		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1067		if (vd != NULL)
1068			zio_flush(lwb->lwb_root_zio, vd);
1069		kmem_free(zv, sizeof (*zv));
1070	}
1071}
1072
1073/*
1074 * This function's purpose is to "open" an lwb such that it is ready to
1075 * accept new itxs being committed to it. To do this, the lwb's zio
1076 * structures are created, and linked to the lwb. This function is
1077 * idempotent; if the passed in lwb has already been opened, this
1078 * function is essentially a no-op.
1079 */
1080static void
1081zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1082{
1083	zbookmark_phys_t zb;
1084	zio_priority_t prio;
1085
1086	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1087	ASSERT3P(lwb, !=, NULL);
1088	EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1089	EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1090
1091	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1092	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1093	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1094
1095	if (lwb->lwb_root_zio == NULL) {
1096		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1097		    BP_GET_LSIZE(&lwb->lwb_blk));
1098
1099		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1100			prio = ZIO_PRIORITY_SYNC_WRITE;
1101		else
1102			prio = ZIO_PRIORITY_ASYNC_WRITE;
1103
1104		lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1105		    zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1106		ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1107
1108		lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1109		    zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1110		    BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1111		    prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
1112		ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1113
1114		lwb->lwb_state = LWB_STATE_OPENED;
1115
1116		mutex_enter(&zilog->zl_lock);
1117
1118		/*
1119		 * The zilog's "zl_last_lwb_opened" field is used to
1120		 * build the lwb/zio dependency chain, which is used to
1121		 * preserve the ordering of lwb completions that is
1122		 * required by the semantics of the ZIL. Each new lwb
1123		 * zio becomes a parent of the "previous" lwb zio, such
1124		 * that the new lwb's zio cannot complete until the
1125		 * "previous" lwb's zio completes.
1126		 *
1127		 * This is required by the semantics of zil_commit();
1128		 * the commit waiters attached to the lwbs will be woken
1129		 * in the lwb zio's completion callback, so this zio
1130		 * dependency graph ensures the waiters are woken in the
1131		 * correct order (the same order the lwbs were created).
1132		 */
1133		lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1134		if (last_lwb_opened != NULL &&
1135		    last_lwb_opened->lwb_state != LWB_STATE_DONE) {
1136			ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1137			    last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1138			ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1139			zio_add_child(lwb->lwb_root_zio,
1140			    last_lwb_opened->lwb_root_zio);
1141		}
1142		zilog->zl_last_lwb_opened = lwb;
1143
1144		mutex_exit(&zilog->zl_lock);
1145	}
1146
1147	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1148	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1149	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1150}
1151
1152/*
1153 * Define a limited set of intent log block sizes.
1154 *
1155 * These must be a multiple of 4KB. Note only the amount used (again
1156 * aligned to 4KB) actually gets written. However, we can't always just
1157 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1158 */
1159uint64_t zil_block_buckets[] = {
1160    4096,		/* non TX_WRITE */
1161    8192+4096,		/* data base */
1162    32*1024 + 4096, 	/* NFS writes */
1163    UINT64_MAX
1164};
1165
1166/*
1167 * Start a log block write and advance to the next log block.
1168 * Calls are serialized.
1169 */
1170static lwb_t *
1171zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1172{
1173	lwb_t *nlwb = NULL;
1174	zil_chain_t *zilc;
1175	spa_t *spa = zilog->zl_spa;
1176	blkptr_t *bp;
1177	dmu_tx_t *tx;
1178	uint64_t txg;
1179	uint64_t zil_blksz, wsz;
1180	int i, error;
1181	boolean_t slog;
1182
1183	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1184	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1185	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1186	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1187
1188	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1189		zilc = (zil_chain_t *)lwb->lwb_buf;
1190		bp = &zilc->zc_next_blk;
1191	} else {
1192		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1193		bp = &zilc->zc_next_blk;
1194	}
1195
1196	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1197
1198	/*
1199	 * Allocate the next block and save its address in this block
1200	 * before writing it in order to establish the log chain.
1201	 * Note that if the allocation of nlwb synced before we wrote
1202	 * the block that points at it (lwb), we'd leak it if we crashed.
1203	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1204	 * We dirty the dataset to ensure that zil_sync() will be called
1205	 * to clean up in the event of allocation failure or I/O failure.
1206	 */
1207
1208	tx = dmu_tx_create(zilog->zl_os);
1209
1210	/*
1211	 * Since we are not going to create any new dirty data, and we
1212	 * can even help with clearing the existing dirty data, we
1213	 * should not be subject to the dirty data based delays. We
1214	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1215	 */
1216	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1217
1218	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1219	txg = dmu_tx_get_txg(tx);
1220
1221	lwb->lwb_tx = tx;
1222
1223	/*
1224	 * Log blocks are pre-allocated. Here we select the size of the next
1225	 * block, based on size used in the last block.
1226	 * - first find the smallest bucket that will fit the block from a
1227	 *   limited set of block sizes. This is because it's faster to write
1228	 *   blocks allocated from the same metaslab as they are adjacent or
1229	 *   close.
1230	 * - next find the maximum from the new suggested size and an array of
1231	 *   previous sizes. This lessens a picket fence effect of wrongly
1232	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1233	 *   requests.
1234	 *
1235	 * Note we only write what is used, but we can't just allocate
1236	 * the maximum block size because we can exhaust the available
1237	 * pool log space.
1238	 */
1239	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1240	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1241		continue;
1242	zil_blksz = zil_block_buckets[i];
1243	if (zil_blksz == UINT64_MAX)
1244		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1245	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1246	for (i = 0; i < ZIL_PREV_BLKS; i++)
1247		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1248	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1249
1250	BP_ZERO(bp);
1251
1252	/* pass the old blkptr in order to spread log blocks across devs */
1253	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1254	if (error == 0) {
1255		ASSERT3U(bp->blk_birth, ==, txg);
1256		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1257		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1258
1259		/*
1260		 * Allocate a new log write block (lwb).
1261		 */
1262		nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1263	}
1264
1265	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1266		/* For Slim ZIL only write what is used. */
1267		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1268		ASSERT3U(wsz, <=, lwb->lwb_sz);
1269		zio_shrink(lwb->lwb_write_zio, wsz);
1270
1271	} else {
1272		wsz = lwb->lwb_sz;
1273	}
1274
1275	zilc->zc_pad = 0;
1276	zilc->zc_nused = lwb->lwb_nused;
1277	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1278
1279	/*
1280	 * clear unused data for security
1281	 */
1282	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1283
1284	spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1285
1286	zil_lwb_add_block(lwb, &lwb->lwb_blk);
1287	lwb->lwb_issued_timestamp = gethrtime();
1288	lwb->lwb_state = LWB_STATE_ISSUED;
1289
1290	zio_nowait(lwb->lwb_root_zio);
1291	zio_nowait(lwb->lwb_write_zio);
1292
1293	/*
1294	 * If there was an allocation failure then nlwb will be null which
1295	 * forces a txg_wait_synced().
1296	 */
1297	return (nlwb);
1298}
1299
1300static lwb_t *
1301zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1302{
1303	lr_t *lrcb, *lrc;
1304	lr_write_t *lrwb, *lrw;
1305	char *lr_buf;
1306	uint64_t dlen, dnow, lwb_sp, reclen, txg;
1307
1308	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1309	ASSERT3P(lwb, !=, NULL);
1310	ASSERT3P(lwb->lwb_buf, !=, NULL);
1311
1312	zil_lwb_write_open(zilog, lwb);
1313
1314	lrc = &itx->itx_lr;
1315	lrw = (lr_write_t *)lrc;
1316
1317	/*
1318	 * A commit itx doesn't represent any on-disk state; instead
1319	 * it's simply used as a place holder on the commit list, and
1320	 * provides a mechanism for attaching a "commit waiter" onto the
1321	 * correct lwb (such that the waiter can be signalled upon
1322	 * completion of that lwb). Thus, we don't process this itx's
1323	 * log record if it's a commit itx (these itx's don't have log
1324	 * records), and instead link the itx's waiter onto the lwb's
1325	 * list of waiters.
1326	 *
1327	 * For more details, see the comment above zil_commit().
1328	 */
1329	if (lrc->lrc_txtype == TX_COMMIT) {
1330		mutex_enter(&zilog->zl_lock);
1331		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1332		itx->itx_private = NULL;
1333		mutex_exit(&zilog->zl_lock);
1334		return (lwb);
1335	}
1336
1337	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1338		dlen = P2ROUNDUP_TYPED(
1339		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1340	} else {
1341		dlen = 0;
1342	}
1343	reclen = lrc->lrc_reclen;
1344	zilog->zl_cur_used += (reclen + dlen);
1345	txg = lrc->lrc_txg;
1346
1347	ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1348
1349cont:
1350	/*
1351	 * If this record won't fit in the current log block, start a new one.
1352	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1353	 */
1354	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1355	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1356	    lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1357	    lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1358		lwb = zil_lwb_write_issue(zilog, lwb);
1359		if (lwb == NULL)
1360			return (NULL);
1361		zil_lwb_write_open(zilog, lwb);
1362		ASSERT(LWB_EMPTY(lwb));
1363		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1364		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1365	}
1366
1367	dnow = MIN(dlen, lwb_sp - reclen);
1368	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1369	bcopy(lrc, lr_buf, reclen);
1370	lrcb = (lr_t *)lr_buf;		/* Like lrc, but inside lwb. */
1371	lrwb = (lr_write_t *)lrcb;	/* Like lrw, but inside lwb. */
1372
1373	/*
1374	 * If it's a write, fetch the data or get its blkptr as appropriate.
1375	 */
1376	if (lrc->lrc_txtype == TX_WRITE) {
1377		if (txg > spa_freeze_txg(zilog->zl_spa))
1378			txg_wait_synced(zilog->zl_dmu_pool, txg);
1379		if (itx->itx_wr_state != WR_COPIED) {
1380			char *dbuf;
1381			int error;
1382
1383			if (itx->itx_wr_state == WR_NEED_COPY) {
1384				dbuf = lr_buf + reclen;
1385				lrcb->lrc_reclen += dnow;
1386				if (lrwb->lr_length > dnow)
1387					lrwb->lr_length = dnow;
1388				lrw->lr_offset += dnow;
1389				lrw->lr_length -= dnow;
1390			} else {
1391				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1392				dbuf = NULL;
1393			}
1394
1395			/*
1396			 * We pass in the "lwb_write_zio" rather than
1397			 * "lwb_root_zio" so that the "lwb_write_zio"
1398			 * becomes the parent of any zio's created by
1399			 * the "zl_get_data" callback. The vdevs are
1400			 * flushed after the "lwb_write_zio" completes,
1401			 * so we want to make sure that completion
1402			 * callback waits for these additional zio's,
1403			 * such that the vdevs used by those zio's will
1404			 * be included in the lwb's vdev tree, and those
1405			 * vdevs will be properly flushed. If we passed
1406			 * in "lwb_root_zio" here, then these additional
1407			 * vdevs may not be flushed; e.g. if these zio's
1408			 * completed after "lwb_write_zio" completed.
1409			 */
1410			error = zilog->zl_get_data(itx->itx_private,
1411			    lrwb, dbuf, lwb, lwb->lwb_write_zio);
1412
1413			if (error == EIO) {
1414				txg_wait_synced(zilog->zl_dmu_pool, txg);
1415				return (lwb);
1416			}
1417			if (error != 0) {
1418				ASSERT(error == ENOENT || error == EEXIST ||
1419				    error == EALREADY);
1420				return (lwb);
1421			}
1422		}
1423	}
1424
1425	/*
1426	 * We're actually making an entry, so update lrc_seq to be the
1427	 * log record sequence number.  Note that this is generally not
1428	 * equal to the itx sequence number because not all transactions
1429	 * are synchronous, and sometimes spa_sync() gets there first.
1430	 */
1431	lrcb->lrc_seq = ++zilog->zl_lr_seq;
1432	lwb->lwb_nused += reclen + dnow;
1433
1434	zil_lwb_add_txg(lwb, txg);
1435
1436	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1437	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1438
1439	dlen -= dnow;
1440	if (dlen > 0) {
1441		zilog->zl_cur_used += reclen;
1442		goto cont;
1443	}
1444
1445	return (lwb);
1446}
1447
1448itx_t *
1449zil_itx_create(uint64_t txtype, size_t lrsize)
1450{
1451	itx_t *itx;
1452
1453	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1454
1455	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1456	itx->itx_lr.lrc_txtype = txtype;
1457	itx->itx_lr.lrc_reclen = lrsize;
1458	itx->itx_lr.lrc_seq = 0;	/* defensive */
1459	itx->itx_sync = B_TRUE;		/* default is synchronous */
1460
1461	return (itx);
1462}
1463
1464void
1465zil_itx_destroy(itx_t *itx)
1466{
1467	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1468}
1469
1470/*
1471 * Free up the sync and async itxs. The itxs_t has already been detached
1472 * so no locks are needed.
1473 */
1474static void
1475zil_itxg_clean(itxs_t *itxs)
1476{
1477	itx_t *itx;
1478	list_t *list;
1479	avl_tree_t *t;
1480	void *cookie;
1481	itx_async_node_t *ian;
1482
1483	list = &itxs->i_sync_list;
1484	while ((itx = list_head(list)) != NULL) {
1485		/*
1486		 * In the general case, commit itxs will not be found
1487		 * here, as they'll be committed to an lwb via
1488		 * zil_lwb_commit(), and free'd in that function. Having
1489		 * said that, it is still possible for commit itxs to be
1490		 * found here, due to the following race:
1491		 *
1492		 *	- a thread calls zil_commit() which assigns the
1493		 *	  commit itx to a per-txg i_sync_list
1494		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
1495		 *	  while the waiter is still on the i_sync_list
1496		 *
1497		 * There's nothing to prevent syncing the txg while the
1498		 * waiter is on the i_sync_list. This normally doesn't
1499		 * happen because spa_sync() is slower than zil_commit(),
1500		 * but if zil_commit() calls txg_wait_synced() (e.g.
1501		 * because zil_create() or zil_commit_writer_stall() is
1502		 * called) we will hit this case.
1503		 */
1504		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1505			zil_commit_waiter_skip(itx->itx_private);
1506
1507		list_remove(list, itx);
1508		zil_itx_destroy(itx);
1509	}
1510
1511	cookie = NULL;
1512	t = &itxs->i_async_tree;
1513	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1514		list = &ian->ia_list;
1515		while ((itx = list_head(list)) != NULL) {
1516			list_remove(list, itx);
1517			/* commit itxs should never be on the async lists. */
1518			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1519			zil_itx_destroy(itx);
1520		}
1521		list_destroy(list);
1522		kmem_free(ian, sizeof (itx_async_node_t));
1523	}
1524	avl_destroy(t);
1525
1526	kmem_free(itxs, sizeof (itxs_t));
1527}
1528
1529static int
1530zil_aitx_compare(const void *x1, const void *x2)
1531{
1532	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1533	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1534
1535	if (o1 < o2)
1536		return (-1);
1537	if (o1 > o2)
1538		return (1);
1539
1540	return (0);
1541}
1542
1543/*
1544 * Remove all async itx with the given oid.
1545 */
1546static void
1547zil_remove_async(zilog_t *zilog, uint64_t oid)
1548{
1549	uint64_t otxg, txg;
1550	itx_async_node_t *ian;
1551	avl_tree_t *t;
1552	avl_index_t where;
1553	list_t clean_list;
1554	itx_t *itx;
1555
1556	ASSERT(oid != 0);
1557	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1558
1559	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1560		otxg = ZILTEST_TXG;
1561	else
1562		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1563
1564	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1565		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1566
1567		mutex_enter(&itxg->itxg_lock);
1568		if (itxg->itxg_txg != txg) {
1569			mutex_exit(&itxg->itxg_lock);
1570			continue;
1571		}
1572
1573		/*
1574		 * Locate the object node and append its list.
1575		 */
1576		t = &itxg->itxg_itxs->i_async_tree;
1577		ian = avl_find(t, &oid, &where);
1578		if (ian != NULL)
1579			list_move_tail(&clean_list, &ian->ia_list);
1580		mutex_exit(&itxg->itxg_lock);
1581	}
1582	while ((itx = list_head(&clean_list)) != NULL) {
1583		list_remove(&clean_list, itx);
1584		/* commit itxs should never be on the async lists. */
1585		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1586		zil_itx_destroy(itx);
1587	}
1588	list_destroy(&clean_list);
1589}
1590
1591void
1592zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1593{
1594	uint64_t txg;
1595	itxg_t *itxg;
1596	itxs_t *itxs, *clean = NULL;
1597
1598	/*
1599	 * Object ids can be re-instantiated in the next txg so
1600	 * remove any async transactions to avoid future leaks.
1601	 * This can happen if a fsync occurs on the re-instantiated
1602	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1603	 * the new file data and flushes a write record for the old object.
1604	 */
1605	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1606		zil_remove_async(zilog, itx->itx_oid);
1607
1608	/*
1609	 * Ensure the data of a renamed file is committed before the rename.
1610	 */
1611	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1612		zil_async_to_sync(zilog, itx->itx_oid);
1613
1614	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1615		txg = ZILTEST_TXG;
1616	else
1617		txg = dmu_tx_get_txg(tx);
1618
1619	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1620	mutex_enter(&itxg->itxg_lock);
1621	itxs = itxg->itxg_itxs;
1622	if (itxg->itxg_txg != txg) {
1623		if (itxs != NULL) {
1624			/*
1625			 * The zil_clean callback hasn't got around to cleaning
1626			 * this itxg. Save the itxs for release below.
1627			 * This should be rare.
1628			 */
1629			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1630			    "txg %llu", itxg->itxg_txg);
1631			clean = itxg->itxg_itxs;
1632		}
1633		itxg->itxg_txg = txg;
1634		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1635
1636		list_create(&itxs->i_sync_list, sizeof (itx_t),
1637		    offsetof(itx_t, itx_node));
1638		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1639		    sizeof (itx_async_node_t),
1640		    offsetof(itx_async_node_t, ia_node));
1641	}
1642	if (itx->itx_sync) {
1643		list_insert_tail(&itxs->i_sync_list, itx);
1644	} else {
1645		avl_tree_t *t = &itxs->i_async_tree;
1646		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1647		itx_async_node_t *ian;
1648		avl_index_t where;
1649
1650		ian = avl_find(t, &foid, &where);
1651		if (ian == NULL) {
1652			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1653			list_create(&ian->ia_list, sizeof (itx_t),
1654			    offsetof(itx_t, itx_node));
1655			ian->ia_foid = foid;
1656			avl_insert(t, ian, where);
1657		}
1658		list_insert_tail(&ian->ia_list, itx);
1659	}
1660
1661	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1662
1663	/*
1664	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1665	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1666	 * need to be careful to always dirty the ZIL using the "real"
1667	 * TXG (not itxg_txg) even when the SPA is frozen.
1668	 */
1669	zilog_dirty(zilog, dmu_tx_get_txg(tx));
1670	mutex_exit(&itxg->itxg_lock);
1671
1672	/* Release the old itxs now we've dropped the lock */
1673	if (clean != NULL)
1674		zil_itxg_clean(clean);
1675}
1676
1677/*
1678 * If there are any in-memory intent log transactions which have now been
1679 * synced then start up a taskq to free them. We should only do this after we
1680 * have written out the uberblocks (i.e. txg has been comitted) so that
1681 * don't inadvertently clean out in-memory log records that would be required
1682 * by zil_commit().
1683 */
1684void
1685zil_clean(zilog_t *zilog, uint64_t synced_txg)
1686{
1687	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1688	itxs_t *clean_me;
1689
1690	ASSERT3U(synced_txg, <, ZILTEST_TXG);
1691
1692	mutex_enter(&itxg->itxg_lock);
1693	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1694		mutex_exit(&itxg->itxg_lock);
1695		return;
1696	}
1697	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1698	ASSERT3U(itxg->itxg_txg, !=, 0);
1699	clean_me = itxg->itxg_itxs;
1700	itxg->itxg_itxs = NULL;
1701	itxg->itxg_txg = 0;
1702	mutex_exit(&itxg->itxg_lock);
1703	/*
1704	 * Preferably start a task queue to free up the old itxs but
1705	 * if taskq_dispatch can't allocate resources to do that then
1706	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1707	 * created a bad performance problem.
1708	 */
1709	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1710	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1711	if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1712	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1713		zil_itxg_clean(clean_me);
1714}
1715
1716/*
1717 * This function will traverse the queue of itxs that need to be
1718 * committed, and move them onto the ZIL's zl_itx_commit_list.
1719 */
1720static void
1721zil_get_commit_list(zilog_t *zilog)
1722{
1723	uint64_t otxg, txg;
1724	list_t *commit_list = &zilog->zl_itx_commit_list;
1725
1726	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1727
1728	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1729		otxg = ZILTEST_TXG;
1730	else
1731		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1732
1733	/*
1734	 * This is inherently racy, since there is nothing to prevent
1735	 * the last synced txg from changing. That's okay since we'll
1736	 * only commit things in the future.
1737	 */
1738	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1739		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1740
1741		mutex_enter(&itxg->itxg_lock);
1742		if (itxg->itxg_txg != txg) {
1743			mutex_exit(&itxg->itxg_lock);
1744			continue;
1745		}
1746
1747		/*
1748		 * If we're adding itx records to the zl_itx_commit_list,
1749		 * then the zil better be dirty in this "txg". We can assert
1750		 * that here since we're holding the itxg_lock which will
1751		 * prevent spa_sync from cleaning it. Once we add the itxs
1752		 * to the zl_itx_commit_list we must commit it to disk even
1753		 * if it's unnecessary (i.e. the txg was synced).
1754		 */
1755		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1756		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1757		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1758
1759		mutex_exit(&itxg->itxg_lock);
1760	}
1761}
1762
1763/*
1764 * Move the async itxs for a specified object to commit into sync lists.
1765 */
1766static void
1767zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1768{
1769	uint64_t otxg, txg;
1770	itx_async_node_t *ian;
1771	avl_tree_t *t;
1772	avl_index_t where;
1773
1774	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1775		otxg = ZILTEST_TXG;
1776	else
1777		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1778
1779	/*
1780	 * This is inherently racy, since there is nothing to prevent
1781	 * the last synced txg from changing.
1782	 */
1783	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1784		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1785
1786		mutex_enter(&itxg->itxg_lock);
1787		if (itxg->itxg_txg != txg) {
1788			mutex_exit(&itxg->itxg_lock);
1789			continue;
1790		}
1791
1792		/*
1793		 * If a foid is specified then find that node and append its
1794		 * list. Otherwise walk the tree appending all the lists
1795		 * to the sync list. We add to the end rather than the
1796		 * beginning to ensure the create has happened.
1797		 */
1798		t = &itxg->itxg_itxs->i_async_tree;
1799		if (foid != 0) {
1800			ian = avl_find(t, &foid, &where);
1801			if (ian != NULL) {
1802				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1803				    &ian->ia_list);
1804			}
1805		} else {
1806			void *cookie = NULL;
1807
1808			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1809				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1810				    &ian->ia_list);
1811				list_destroy(&ian->ia_list);
1812				kmem_free(ian, sizeof (itx_async_node_t));
1813			}
1814		}
1815		mutex_exit(&itxg->itxg_lock);
1816	}
1817}
1818
1819/*
1820 * This function will prune commit itxs that are at the head of the
1821 * commit list (it won't prune past the first non-commit itx), and
1822 * either: a) attach them to the last lwb that's still pending
1823 * completion, or b) skip them altogether.
1824 *
1825 * This is used as a performance optimization to prevent commit itxs
1826 * from generating new lwbs when it's unnecessary to do so.
1827 */
1828static void
1829zil_prune_commit_list(zilog_t *zilog)
1830{
1831	itx_t *itx;
1832
1833	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1834
1835	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1836		lr_t *lrc = &itx->itx_lr;
1837		if (lrc->lrc_txtype != TX_COMMIT)
1838			break;
1839
1840		mutex_enter(&zilog->zl_lock);
1841
1842		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
1843		if (last_lwb == NULL || last_lwb->lwb_state == LWB_STATE_DONE) {
1844			/*
1845			 * All of the itxs this waiter was waiting on
1846			 * must have already completed (or there were
1847			 * never any itx's for it to wait on), so it's
1848			 * safe to skip this waiter and mark it done.
1849			 */
1850			zil_commit_waiter_skip(itx->itx_private);
1851		} else {
1852			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
1853			itx->itx_private = NULL;
1854		}
1855
1856		mutex_exit(&zilog->zl_lock);
1857
1858		list_remove(&zilog->zl_itx_commit_list, itx);
1859		zil_itx_destroy(itx);
1860	}
1861
1862	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
1863}
1864
1865static void
1866zil_commit_writer_stall(zilog_t *zilog)
1867{
1868	/*
1869	 * When zio_alloc_zil() fails to allocate the next lwb block on
1870	 * disk, we must call txg_wait_synced() to ensure all of the
1871	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
1872	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
1873	 * to zil_process_commit_list()) will have to call zil_create(),
1874	 * and start a new ZIL chain.
1875	 *
1876	 * Since zil_alloc_zil() failed, the lwb that was previously
1877	 * issued does not have a pointer to the "next" lwb on disk.
1878	 * Thus, if another ZIL writer thread was to allocate the "next"
1879	 * on-disk lwb, that block could be leaked in the event of a
1880	 * crash (because the previous lwb on-disk would not point to
1881	 * it).
1882	 *
1883	 * We must hold the zilog's zl_issuer_lock while we do this, to
1884	 * ensure no new threads enter zil_process_commit_list() until
1885	 * all lwb's in the zl_lwb_list have been synced and freed
1886	 * (which is achieved via the txg_wait_synced() call).
1887	 */
1888	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1889	txg_wait_synced(zilog->zl_dmu_pool, 0);
1890	ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
1891}
1892
1893/*
1894 * This function will traverse the commit list, creating new lwbs as
1895 * needed, and committing the itxs from the commit list to these newly
1896 * created lwbs. Additionally, as a new lwb is created, the previous
1897 * lwb will be issued to the zio layer to be written to disk.
1898 */
1899static void
1900zil_process_commit_list(zilog_t *zilog)
1901{
1902	spa_t *spa = zilog->zl_spa;
1903	list_t nolwb_waiters;
1904	lwb_t *lwb;
1905	itx_t *itx;
1906
1907	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1908
1909	/*
1910	 * Return if there's nothing to commit before we dirty the fs by
1911	 * calling zil_create().
1912	 */
1913	if (list_head(&zilog->zl_itx_commit_list) == NULL)
1914		return;
1915
1916	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
1917	    offsetof(zil_commit_waiter_t, zcw_node));
1918
1919	lwb = list_tail(&zilog->zl_lwb_list);
1920	if (lwb == NULL) {
1921		lwb = zil_create(zilog);
1922	} else {
1923		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
1924		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
1925	}
1926
1927	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1928		lr_t *lrc = &itx->itx_lr;
1929		uint64_t txg = lrc->lrc_txg;
1930
1931		ASSERT3U(txg, !=, 0);
1932
1933		if (lrc->lrc_txtype == TX_COMMIT) {
1934			DTRACE_PROBE2(zil__process__commit__itx,
1935			    zilog_t *, zilog, itx_t *, itx);
1936		} else {
1937			DTRACE_PROBE2(zil__process__normal__itx,
1938			    zilog_t *, zilog, itx_t *, itx);
1939		}
1940
1941		boolean_t synced = txg <= spa_last_synced_txg(spa);
1942		boolean_t frozen = txg > spa_freeze_txg(spa);
1943
1944		/*
1945		 * If the txg of this itx has already been synced out, then
1946		 * we don't need to commit this itx to an lwb. This is
1947		 * because the data of this itx will have already been
1948		 * written to the main pool. This is inherently racy, and
1949		 * it's still ok to commit an itx whose txg has already
1950		 * been synced; this will result in a write that's
1951		 * unnecessary, but will do no harm.
1952		 *
1953		 * With that said, we always want to commit TX_COMMIT itxs
1954		 * to an lwb, regardless of whether or not that itx's txg
1955		 * has been synced out. We do this to ensure any OPENED lwb
1956		 * will always have at least one zil_commit_waiter_t linked
1957		 * to the lwb.
1958		 *
1959		 * As a counter-example, if we skipped TX_COMMIT itx's
1960		 * whose txg had already been synced, the following
1961		 * situation could occur if we happened to be racing with
1962		 * spa_sync:
1963		 *
1964		 * 1. we commit a non-TX_COMMIT itx to an lwb, where the
1965		 *    itx's txg is 10 and the last synced txg is 9.
1966		 * 2. spa_sync finishes syncing out txg 10.
1967		 * 3. we move to the next itx in the list, it's a TX_COMMIT
1968		 *    whose txg is 10, so we skip it rather than committing
1969		 *    it to the lwb used in (1).
1970		 *
1971		 * If the itx that is skipped in (3) is the last TX_COMMIT
1972		 * itx in the commit list, than it's possible for the lwb
1973		 * used in (1) to remain in the OPENED state indefinitely.
1974		 *
1975		 * To prevent the above scenario from occuring, ensuring
1976		 * that once an lwb is OPENED it will transition to ISSUED
1977		 * and eventually DONE, we always commit TX_COMMIT itx's to
1978		 * an lwb here, even if that itx's txg has already been
1979		 * synced.
1980		 *
1981		 * Finally, if the pool is frozen, we _always_ commit the
1982		 * itx.  The point of freezing the pool is to prevent data
1983		 * from being written to the main pool via spa_sync, and
1984		 * instead rely solely on the ZIL to persistently store the
1985		 * data; i.e.  when the pool is frozen, the last synced txg
1986		 * value can't be trusted.
1987		 */
1988		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
1989			if (lwb != NULL) {
1990				lwb = zil_lwb_commit(zilog, itx, lwb);
1991			} else if (lrc->lrc_txtype == TX_COMMIT) {
1992				ASSERT3P(lwb, ==, NULL);
1993				zil_commit_waiter_link_nolwb(
1994				    itx->itx_private, &nolwb_waiters);
1995			}
1996		}
1997
1998		list_remove(&zilog->zl_itx_commit_list, itx);
1999		zil_itx_destroy(itx);
2000	}
2001
2002	if (lwb == NULL) {
2003		/*
2004		 * This indicates zio_alloc_zil() failed to allocate the
2005		 * "next" lwb on-disk. When this happens, we must stall
2006		 * the ZIL write pipeline; see the comment within
2007		 * zil_commit_writer_stall() for more details.
2008		 */
2009		zil_commit_writer_stall(zilog);
2010
2011		/*
2012		 * Additionally, we have to signal and mark the "nolwb"
2013		 * waiters as "done" here, since without an lwb, we
2014		 * can't do this via zil_lwb_flush_vdevs_done() like
2015		 * normal.
2016		 */
2017		zil_commit_waiter_t *zcw;
2018		while (zcw = list_head(&nolwb_waiters)) {
2019			zil_commit_waiter_skip(zcw);
2020			list_remove(&nolwb_waiters, zcw);
2021		}
2022	} else {
2023		ASSERT(list_is_empty(&nolwb_waiters));
2024		ASSERT3P(lwb, !=, NULL);
2025		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2026		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2027
2028		/*
2029		 * At this point, the ZIL block pointed at by the "lwb"
2030		 * variable is in one of the following states: "closed"
2031		 * or "open".
2032		 *
2033		 * If its "closed", then no itxs have been committed to
2034		 * it, so there's no point in issuing its zio (i.e.
2035		 * it's "empty").
2036		 *
2037		 * If its "open" state, then it contains one or more
2038		 * itxs that eventually need to be committed to stable
2039		 * storage. In this case we intentionally do not issue
2040		 * the lwb's zio to disk yet, and instead rely on one of
2041		 * the following two mechanisms for issuing the zio:
2042		 *
2043		 * 1. Ideally, there will be more ZIL activity occuring
2044		 * on the system, such that this function will be
2045		 * immediately called again (not necessarily by the same
2046		 * thread) and this lwb's zio will be issued via
2047		 * zil_lwb_commit(). This way, the lwb is guaranteed to
2048		 * be "full" when it is issued to disk, and we'll make
2049		 * use of the lwb's size the best we can.
2050		 *
2051		 * 2. If there isn't sufficient ZIL activity occuring on
2052		 * the system, such that this lwb's zio isn't issued via
2053		 * zil_lwb_commit(), zil_commit_waiter() will issue the
2054		 * lwb's zio. If this occurs, the lwb is not guaranteed
2055		 * to be "full" by the time its zio is issued, and means
2056		 * the size of the lwb was "too large" given the amount
2057		 * of ZIL activity occuring on the system at that time.
2058		 *
2059		 * We do this for a couple of reasons:
2060		 *
2061		 * 1. To try and reduce the number of IOPs needed to
2062		 * write the same number of itxs. If an lwb has space
2063		 * available in it's buffer for more itxs, and more itxs
2064		 * will be committed relatively soon (relative to the
2065		 * latency of performing a write), then it's beneficial
2066		 * to wait for these "next" itxs. This way, more itxs
2067		 * can be committed to stable storage with fewer writes.
2068		 *
2069		 * 2. To try and use the largest lwb block size that the
2070		 * incoming rate of itxs can support. Again, this is to
2071		 * try and pack as many itxs into as few lwbs as
2072		 * possible, without significantly impacting the latency
2073		 * of each individual itx.
2074		 */
2075	}
2076}
2077
2078/*
2079 * This function is responsible for ensuring the passed in commit waiter
2080 * (and associated commit itx) is committed to an lwb. If the waiter is
2081 * not already committed to an lwb, all itxs in the zilog's queue of
2082 * itxs will be processed. The assumption is the passed in waiter's
2083 * commit itx will found in the queue just like the other non-commit
2084 * itxs, such that when the entire queue is processed, the waiter will
2085 * have been commited to an lwb.
2086 *
2087 * The lwb associated with the passed in waiter is not guaranteed to
2088 * have been issued by the time this function completes. If the lwb is
2089 * not issued, we rely on future calls to zil_commit_writer() to issue
2090 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2091 */
2092static void
2093zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2094{
2095	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2096	ASSERT(spa_writeable(zilog->zl_spa));
2097
2098	mutex_enter(&zilog->zl_issuer_lock);
2099
2100	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2101		/*
2102		 * It's possible that, while we were waiting to acquire
2103		 * the "zl_issuer_lock", another thread committed this
2104		 * waiter to an lwb. If that occurs, we bail out early,
2105		 * without processing any of the zilog's queue of itxs.
2106		 *
2107		 * On certain workloads and system configurations, the
2108		 * "zl_issuer_lock" can become highly contended. In an
2109		 * attempt to reduce this contention, we immediately drop
2110		 * the lock if the waiter has already been processed.
2111		 *
2112		 * We've measured this optimization to reduce CPU spent
2113		 * contending on this lock by up to 5%, using a system
2114		 * with 32 CPUs, low latency storage (~50 usec writes),
2115		 * and 1024 threads performing sync writes.
2116		 */
2117		goto out;
2118	}
2119
2120	zil_get_commit_list(zilog);
2121	zil_prune_commit_list(zilog);
2122	zil_process_commit_list(zilog);
2123
2124out:
2125	mutex_exit(&zilog->zl_issuer_lock);
2126}
2127
2128static void
2129zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2130{
2131	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2132	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2133	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2134
2135	lwb_t *lwb = zcw->zcw_lwb;
2136	ASSERT3P(lwb, !=, NULL);
2137	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2138
2139	/*
2140	 * If the lwb has already been issued by another thread, we can
2141	 * immediately return since there's no work to be done (the
2142	 * point of this function is to issue the lwb). Additionally, we
2143	 * do this prior to acquiring the zl_issuer_lock, to avoid
2144	 * acquiring it when it's not necessary to do so.
2145	 */
2146	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2147	    lwb->lwb_state == LWB_STATE_DONE)
2148		return;
2149
2150	/*
2151	 * In order to call zil_lwb_write_issue() we must hold the
2152	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2153	 * since we're already holding the commit waiter's "zcw_lock",
2154	 * and those two locks are aquired in the opposite order
2155	 * elsewhere.
2156	 */
2157	mutex_exit(&zcw->zcw_lock);
2158	mutex_enter(&zilog->zl_issuer_lock);
2159	mutex_enter(&zcw->zcw_lock);
2160
2161	/*
2162	 * Since we just dropped and re-acquired the commit waiter's
2163	 * lock, we have to re-check to see if the waiter was marked
2164	 * "done" during that process. If the waiter was marked "done",
2165	 * the "lwb" pointer is no longer valid (it can be free'd after
2166	 * the waiter is marked "done"), so without this check we could
2167	 * wind up with a use-after-free error below.
2168	 */
2169	if (zcw->zcw_done)
2170		goto out;
2171
2172	ASSERT3P(lwb, ==, zcw->zcw_lwb);
2173
2174	/*
2175	 * We've already checked this above, but since we hadn't acquired
2176	 * the zilog's zl_issuer_lock, we have to perform this check a
2177	 * second time while holding the lock.
2178	 *
2179	 * We don't need to hold the zl_lock since the lwb cannot transition
2180	 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2181	 * _can_ transition from ISSUED to DONE, but it's OK to race with
2182	 * that transition since we treat the lwb the same, whether it's in
2183	 * the ISSUED or DONE states.
2184	 *
2185	 * The important thing, is we treat the lwb differently depending on
2186	 * if it's ISSUED or OPENED, and block any other threads that might
2187	 * attempt to issue this lwb. For that reason we hold the
2188	 * zl_issuer_lock when checking the lwb_state; we must not call
2189	 * zil_lwb_write_issue() if the lwb had already been issued.
2190	 *
2191	 * See the comment above the lwb_state_t structure definition for
2192	 * more details on the lwb states, and locking requirements.
2193	 */
2194	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2195	    lwb->lwb_state == LWB_STATE_DONE)
2196		goto out;
2197
2198	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2199
2200	/*
2201	 * As described in the comments above zil_commit_waiter() and
2202	 * zil_process_commit_list(), we need to issue this lwb's zio
2203	 * since we've reached the commit waiter's timeout and it still
2204	 * hasn't been issued.
2205	 */
2206	lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2207
2208	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2209
2210	/*
2211	 * Since the lwb's zio hadn't been issued by the time this thread
2212	 * reached its timeout, we reset the zilog's "zl_cur_used" field
2213	 * to influence the zil block size selection algorithm.
2214	 *
2215	 * By having to issue the lwb's zio here, it means the size of the
2216	 * lwb was too large, given the incoming throughput of itxs.  By
2217	 * setting "zl_cur_used" to zero, we communicate this fact to the
2218	 * block size selection algorithm, so it can take this informaiton
2219	 * into account, and potentially select a smaller size for the
2220	 * next lwb block that is allocated.
2221	 */
2222	zilog->zl_cur_used = 0;
2223
2224	if (nlwb == NULL) {
2225		/*
2226		 * When zil_lwb_write_issue() returns NULL, this
2227		 * indicates zio_alloc_zil() failed to allocate the
2228		 * "next" lwb on-disk. When this occurs, the ZIL write
2229		 * pipeline must be stalled; see the comment within the
2230		 * zil_commit_writer_stall() function for more details.
2231		 *
2232		 * We must drop the commit waiter's lock prior to
2233		 * calling zil_commit_writer_stall() or else we can wind
2234		 * up with the following deadlock:
2235		 *
2236		 * - This thread is waiting for the txg to sync while
2237		 *   holding the waiter's lock; txg_wait_synced() is
2238		 *   used within txg_commit_writer_stall().
2239		 *
2240		 * - The txg can't sync because it is waiting for this
2241		 *   lwb's zio callback to call dmu_tx_commit().
2242		 *
2243		 * - The lwb's zio callback can't call dmu_tx_commit()
2244		 *   because it's blocked trying to acquire the waiter's
2245		 *   lock, which occurs prior to calling dmu_tx_commit()
2246		 */
2247		mutex_exit(&zcw->zcw_lock);
2248		zil_commit_writer_stall(zilog);
2249		mutex_enter(&zcw->zcw_lock);
2250	}
2251
2252out:
2253	mutex_exit(&zilog->zl_issuer_lock);
2254	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2255}
2256
2257/*
2258 * This function is responsible for performing the following two tasks:
2259 *
2260 * 1. its primary responsibility is to block until the given "commit
2261 *    waiter" is considered "done".
2262 *
2263 * 2. its secondary responsibility is to issue the zio for the lwb that
2264 *    the given "commit waiter" is waiting on, if this function has
2265 *    waited "long enough" and the lwb is still in the "open" state.
2266 *
2267 * Given a sufficient amount of itxs being generated and written using
2268 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2269 * function. If this does not occur, this secondary responsibility will
2270 * ensure the lwb is issued even if there is not other synchronous
2271 * activity on the system.
2272 *
2273 * For more details, see zil_process_commit_list(); more specifically,
2274 * the comment at the bottom of that function.
2275 */
2276static void
2277zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2278{
2279	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2280	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2281	ASSERT(spa_writeable(zilog->zl_spa));
2282
2283	mutex_enter(&zcw->zcw_lock);
2284
2285	/*
2286	 * The timeout is scaled based on the lwb latency to avoid
2287	 * significantly impacting the latency of each individual itx.
2288	 * For more details, see the comment at the bottom of the
2289	 * zil_process_commit_list() function.
2290	 */
2291	int pct = MAX(zfs_commit_timeout_pct, 1);
2292	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2293	hrtime_t wakeup = gethrtime() + sleep;
2294	boolean_t timedout = B_FALSE;
2295
2296	while (!zcw->zcw_done) {
2297		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2298
2299		lwb_t *lwb = zcw->zcw_lwb;
2300
2301		/*
2302		 * Usually, the waiter will have a non-NULL lwb field here,
2303		 * but it's possible for it to be NULL as a result of
2304		 * zil_commit() racing with spa_sync().
2305		 *
2306		 * When zil_clean() is called, it's possible for the itxg
2307		 * list (which may be cleaned via a taskq) to contain
2308		 * commit itxs. When this occurs, the commit waiters linked
2309		 * off of these commit itxs will not be committed to an
2310		 * lwb.  Additionally, these commit waiters will not be
2311		 * marked done until zil_commit_waiter_skip() is called via
2312		 * zil_itxg_clean().
2313		 *
2314		 * Thus, it's possible for this commit waiter (i.e. the
2315		 * "zcw" variable) to be found in this "in between" state;
2316		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2317		 * been skipped, so it's "zcw_done" field is still B_FALSE.
2318		 */
2319		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2320
2321		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2322			ASSERT3B(timedout, ==, B_FALSE);
2323
2324			/*
2325			 * If the lwb hasn't been issued yet, then we
2326			 * need to wait with a timeout, in case this
2327			 * function needs to issue the lwb after the
2328			 * timeout is reached; responsibility (2) from
2329			 * the comment above this function.
2330			 */
2331			clock_t timeleft = cv_timedwait_hires(&zcw->zcw_cv,
2332			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2333			    CALLOUT_FLAG_ABSOLUTE);
2334
2335			if (timeleft >= 0 || zcw->zcw_done)
2336				continue;
2337
2338			timedout = B_TRUE;
2339			zil_commit_waiter_timeout(zilog, zcw);
2340
2341			if (!zcw->zcw_done) {
2342				/*
2343				 * If the commit waiter has already been
2344				 * marked "done", it's possible for the
2345				 * waiter's lwb structure to have already
2346				 * been freed.  Thus, we can only reliably
2347				 * make these assertions if the waiter
2348				 * isn't done.
2349				 */
2350				ASSERT3P(lwb, ==, zcw->zcw_lwb);
2351				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2352			}
2353		} else {
2354			/*
2355			 * If the lwb isn't open, then it must have already
2356			 * been issued. In that case, there's no need to
2357			 * use a timeout when waiting for the lwb to
2358			 * complete.
2359			 *
2360			 * Additionally, if the lwb is NULL, the waiter
2361			 * will soon be signalled and marked done via
2362			 * zil_clean() and zil_itxg_clean(), so no timeout
2363			 * is required.
2364			 */
2365
2366			IMPLY(lwb != NULL,
2367			    lwb->lwb_state == LWB_STATE_ISSUED ||
2368			    lwb->lwb_state == LWB_STATE_DONE);
2369			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2370		}
2371	}
2372
2373	mutex_exit(&zcw->zcw_lock);
2374}
2375
2376static zil_commit_waiter_t *
2377zil_alloc_commit_waiter()
2378{
2379	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2380
2381	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2382	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2383	list_link_init(&zcw->zcw_node);
2384	zcw->zcw_lwb = NULL;
2385	zcw->zcw_done = B_FALSE;
2386	zcw->zcw_zio_error = 0;
2387
2388	return (zcw);
2389}
2390
2391static void
2392zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2393{
2394	ASSERT(!list_link_active(&zcw->zcw_node));
2395	ASSERT3P(zcw->zcw_lwb, ==, NULL);
2396	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2397	mutex_destroy(&zcw->zcw_lock);
2398	cv_destroy(&zcw->zcw_cv);
2399	kmem_cache_free(zil_zcw_cache, zcw);
2400}
2401
2402/*
2403 * This function is used to create a TX_COMMIT itx and assign it. This
2404 * way, it will be linked into the ZIL's list of synchronous itxs, and
2405 * then later committed to an lwb (or skipped) when
2406 * zil_process_commit_list() is called.
2407 */
2408static void
2409zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2410{
2411	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2412	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2413
2414	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2415	itx->itx_sync = B_TRUE;
2416	itx->itx_private = zcw;
2417
2418	zil_itx_assign(zilog, itx, tx);
2419
2420	dmu_tx_commit(tx);
2421}
2422
2423/*
2424 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2425 *
2426 * When writing ZIL transactions to the on-disk representation of the
2427 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2428 * itxs can be committed to a single lwb. Once a lwb is written and
2429 * committed to stable storage (i.e. the lwb is written, and vdevs have
2430 * been flushed), each itx that was committed to that lwb is also
2431 * considered to be committed to stable storage.
2432 *
2433 * When an itx is committed to an lwb, the log record (lr_t) contained
2434 * by the itx is copied into the lwb's zio buffer, and once this buffer
2435 * is written to disk, it becomes an on-disk ZIL block.
2436 *
2437 * As itxs are generated, they're inserted into the ZIL's queue of
2438 * uncommitted itxs. The semantics of zil_commit() are such that it will
2439 * block until all itxs that were in the queue when it was called, are
2440 * committed to stable storage.
2441 *
2442 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2443 * itxs, for all objects in the dataset, will be committed to stable
2444 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2445 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2446 * that correspond to the foid passed in, will be committed to stable
2447 * storage prior to zil_commit() returning.
2448 *
2449 * Generally speaking, when zil_commit() is called, the consumer doesn't
2450 * actually care about _all_ of the uncommitted itxs. Instead, they're
2451 * simply trying to waiting for a specific itx to be committed to disk,
2452 * but the interface(s) for interacting with the ZIL don't allow such
2453 * fine-grained communication. A better interface would allow a consumer
2454 * to create and assign an itx, and then pass a reference to this itx to
2455 * zil_commit(); such that zil_commit() would return as soon as that
2456 * specific itx was committed to disk (instead of waiting for _all_
2457 * itxs to be committed).
2458 *
2459 * When a thread calls zil_commit() a special "commit itx" will be
2460 * generated, along with a corresponding "waiter" for this commit itx.
2461 * zil_commit() will wait on this waiter's CV, such that when the waiter
2462 * is marked done, and signalled, zil_commit() will return.
2463 *
2464 * This commit itx is inserted into the queue of uncommitted itxs. This
2465 * provides an easy mechanism for determining which itxs were in the
2466 * queue prior to zil_commit() having been called, and which itxs were
2467 * added after zil_commit() was called.
2468 *
2469 * The commit it is special; it doesn't have any on-disk representation.
2470 * When a commit itx is "committed" to an lwb, the waiter associated
2471 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2472 * completes, each waiter on the lwb's list is marked done and signalled
2473 * -- allowing the thread waiting on the waiter to return from zil_commit().
2474 *
2475 * It's important to point out a few critical factors that allow us
2476 * to make use of the commit itxs, commit waiters, per-lwb lists of
2477 * commit waiters, and zio completion callbacks like we're doing:
2478 *
2479 *   1. The list of waiters for each lwb is traversed, and each commit
2480 *      waiter is marked "done" and signalled, in the zio completion
2481 *      callback of the lwb's zio[*].
2482 *
2483 *      * Actually, the waiters are signalled in the zio completion
2484 *        callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2485 *        that are sent to the vdevs upon completion of the lwb zio.
2486 *
2487 *   2. When the itxs are inserted into the ZIL's queue of uncommitted
2488 *      itxs, the order in which they are inserted is preserved[*]; as
2489 *      itxs are added to the queue, they are added to the tail of
2490 *      in-memory linked lists.
2491 *
2492 *      When committing the itxs to lwbs (to be written to disk), they
2493 *      are committed in the same order in which the itxs were added to
2494 *      the uncommitted queue's linked list(s); i.e. the linked list of
2495 *      itxs to commit is traversed from head to tail, and each itx is
2496 *      committed to an lwb in that order.
2497 *
2498 *      * To clarify:
2499 *
2500 *        - the order of "sync" itxs is preserved w.r.t. other
2501 *          "sync" itxs, regardless of the corresponding objects.
2502 *        - the order of "async" itxs is preserved w.r.t. other
2503 *          "async" itxs corresponding to the same object.
2504 *        - the order of "async" itxs is *not* preserved w.r.t. other
2505 *          "async" itxs corresponding to different objects.
2506 *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
2507 *          versa) is *not* preserved, even for itxs that correspond
2508 *          to the same object.
2509 *
2510 *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
2511 *      zil_get_commit_list(), and zil_process_commit_list().
2512 *
2513 *   3. The lwbs represent a linked list of blocks on disk. Thus, any
2514 *      lwb cannot be considered committed to stable storage, until its
2515 *      "previous" lwb is also committed to stable storage. This fact,
2516 *      coupled with the fact described above, means that itxs are
2517 *      committed in (roughly) the order in which they were generated.
2518 *      This is essential because itxs are dependent on prior itxs.
2519 *      Thus, we *must not* deem an itx as being committed to stable
2520 *      storage, until *all* prior itxs have also been committed to
2521 *      stable storage.
2522 *
2523 *      To enforce this ordering of lwb zio's, while still leveraging as
2524 *      much of the underlying storage performance as possible, we rely
2525 *      on two fundamental concepts:
2526 *
2527 *          1. The creation and issuance of lwb zio's is protected by
2528 *             the zilog's "zl_issuer_lock", which ensures only a single
2529 *             thread is creating and/or issuing lwb's at a time
2530 *          2. The "previous" lwb is a child of the "current" lwb
2531 *             (leveraging the zio parent-child depenency graph)
2532 *
2533 *      By relying on this parent-child zio relationship, we can have
2534 *      many lwb zio's concurrently issued to the underlying storage,
2535 *      but the order in which they complete will be the same order in
2536 *      which they were created.
2537 */
2538void
2539zil_commit(zilog_t *zilog, uint64_t foid)
2540{
2541	/*
2542	 * We should never attempt to call zil_commit on a snapshot for
2543	 * a couple of reasons:
2544	 *
2545	 * 1. A snapshot may never be modified, thus it cannot have any
2546	 *    in-flight itxs that would have modified the dataset.
2547	 *
2548	 * 2. By design, when zil_commit() is called, a commit itx will
2549	 *    be assigned to this zilog; as a result, the zilog will be
2550	 *    dirtied. We must not dirty the zilog of a snapshot; there's
2551	 *    checks in the code that enforce this invariant, and will
2552	 *    cause a panic if it's not upheld.
2553	 */
2554	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2555
2556	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2557		return;
2558
2559	if (!spa_writeable(zilog->zl_spa)) {
2560		/*
2561		 * If the SPA is not writable, there should never be any
2562		 * pending itxs waiting to be committed to disk. If that
2563		 * weren't true, we'd skip writing those itxs out, and
2564		 * would break the sematics of zil_commit(); thus, we're
2565		 * verifying that truth before we return to the caller.
2566		 */
2567		ASSERT(list_is_empty(&zilog->zl_lwb_list));
2568		ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2569		for (int i = 0; i < TXG_SIZE; i++)
2570			ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2571		return;
2572	}
2573
2574	/*
2575	 * If the ZIL is suspended, we don't want to dirty it by calling
2576	 * zil_commit_itx_assign() below, nor can we write out
2577	 * lwbs like would be done in zil_commit_write(). Thus, we
2578	 * simply rely on txg_wait_synced() to maintain the necessary
2579	 * semantics, and avoid calling those functions altogether.
2580	 */
2581	if (zilog->zl_suspend > 0) {
2582		txg_wait_synced(zilog->zl_dmu_pool, 0);
2583		return;
2584	}
2585
2586	zil_commit_impl(zilog, foid);
2587}
2588
2589void
2590zil_commit_impl(zilog_t *zilog, uint64_t foid)
2591{
2592	/*
2593	 * Move the "async" itxs for the specified foid to the "sync"
2594	 * queues, such that they will be later committed (or skipped)
2595	 * to an lwb when zil_process_commit_list() is called.
2596	 *
2597	 * Since these "async" itxs must be committed prior to this
2598	 * call to zil_commit returning, we must perform this operation
2599	 * before we call zil_commit_itx_assign().
2600	 */
2601	zil_async_to_sync(zilog, foid);
2602
2603	/*
2604	 * We allocate a new "waiter" structure which will initially be
2605	 * linked to the commit itx using the itx's "itx_private" field.
2606	 * Since the commit itx doesn't represent any on-disk state,
2607	 * when it's committed to an lwb, rather than copying the its
2608	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2609	 * added to the lwb's list of waiters. Then, when the lwb is
2610	 * committed to stable storage, each waiter in the lwb's list of
2611	 * waiters will be marked "done", and signalled.
2612	 *
2613	 * We must create the waiter and assign the commit itx prior to
2614	 * calling zil_commit_writer(), or else our specific commit itx
2615	 * is not guaranteed to be committed to an lwb prior to calling
2616	 * zil_commit_waiter().
2617	 */
2618	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
2619	zil_commit_itx_assign(zilog, zcw);
2620
2621	zil_commit_writer(zilog, zcw);
2622	zil_commit_waiter(zilog, zcw);
2623
2624	if (zcw->zcw_zio_error != 0) {
2625		/*
2626		 * If there was an error writing out the ZIL blocks that
2627		 * this thread is waiting on, then we fallback to
2628		 * relying on spa_sync() to write out the data this
2629		 * thread is waiting on. Obviously this has performance
2630		 * implications, but the expectation is for this to be
2631		 * an exceptional case, and shouldn't occur often.
2632		 */
2633		DTRACE_PROBE2(zil__commit__io__error,
2634		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
2635		txg_wait_synced(zilog->zl_dmu_pool, 0);
2636	}
2637
2638	zil_free_commit_waiter(zcw);
2639}
2640
2641/*
2642 * Called in syncing context to free committed log blocks and update log header.
2643 */
2644void
2645zil_sync(zilog_t *zilog, dmu_tx_t *tx)
2646{
2647	zil_header_t *zh = zil_header_in_syncing_context(zilog);
2648	uint64_t txg = dmu_tx_get_txg(tx);
2649	spa_t *spa = zilog->zl_spa;
2650	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
2651	lwb_t *lwb;
2652
2653	/*
2654	 * We don't zero out zl_destroy_txg, so make sure we don't try
2655	 * to destroy it twice.
2656	 */
2657	if (spa_sync_pass(spa) != 1)
2658		return;
2659
2660	mutex_enter(&zilog->zl_lock);
2661
2662	ASSERT(zilog->zl_stop_sync == 0);
2663
2664	if (*replayed_seq != 0) {
2665		ASSERT(zh->zh_replay_seq < *replayed_seq);
2666		zh->zh_replay_seq = *replayed_seq;
2667		*replayed_seq = 0;
2668	}
2669
2670	if (zilog->zl_destroy_txg == txg) {
2671		blkptr_t blk = zh->zh_log;
2672
2673		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
2674
2675		bzero(zh, sizeof (zil_header_t));
2676		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
2677
2678		if (zilog->zl_keep_first) {
2679			/*
2680			 * If this block was part of log chain that couldn't
2681			 * be claimed because a device was missing during
2682			 * zil_claim(), but that device later returns,
2683			 * then this block could erroneously appear valid.
2684			 * To guard against this, assign a new GUID to the new
2685			 * log chain so it doesn't matter what blk points to.
2686			 */
2687			zil_init_log_chain(zilog, &blk);
2688			zh->zh_log = blk;
2689		}
2690	}
2691
2692	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
2693		zh->zh_log = lwb->lwb_blk;
2694		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
2695			break;
2696		list_remove(&zilog->zl_lwb_list, lwb);
2697		zio_free(spa, txg, &lwb->lwb_blk);
2698		zil_free_lwb(zilog, lwb);
2699
2700		/*
2701		 * If we don't have anything left in the lwb list then
2702		 * we've had an allocation failure and we need to zero
2703		 * out the zil_header blkptr so that we don't end
2704		 * up freeing the same block twice.
2705		 */
2706		if (list_head(&zilog->zl_lwb_list) == NULL)
2707			BP_ZERO(&zh->zh_log);
2708	}
2709	mutex_exit(&zilog->zl_lock);
2710}
2711
2712/* ARGSUSED */
2713static int
2714zil_lwb_cons(void *vbuf, void *unused, int kmflag)
2715{
2716	lwb_t *lwb = vbuf;
2717	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
2718	    offsetof(zil_commit_waiter_t, zcw_node));
2719	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
2720	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
2721	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
2722	return (0);
2723}
2724
2725/* ARGSUSED */
2726static void
2727zil_lwb_dest(void *vbuf, void *unused)
2728{
2729	lwb_t *lwb = vbuf;
2730	mutex_destroy(&lwb->lwb_vdev_lock);
2731	avl_destroy(&lwb->lwb_vdev_tree);
2732	list_destroy(&lwb->lwb_waiters);
2733}
2734
2735void
2736zil_init(void)
2737{
2738	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
2739	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
2740
2741	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
2742	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2743}
2744
2745void
2746zil_fini(void)
2747{
2748	kmem_cache_destroy(zil_zcw_cache);
2749	kmem_cache_destroy(zil_lwb_cache);
2750}
2751
2752void
2753zil_set_sync(zilog_t *zilog, uint64_t sync)
2754{
2755	zilog->zl_sync = sync;
2756}
2757
2758void
2759zil_set_logbias(zilog_t *zilog, uint64_t logbias)
2760{
2761	zilog->zl_logbias = logbias;
2762}
2763
2764zilog_t *
2765zil_alloc(objset_t *os, zil_header_t *zh_phys)
2766{
2767	zilog_t *zilog;
2768
2769	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
2770
2771	zilog->zl_header = zh_phys;
2772	zilog->zl_os = os;
2773	zilog->zl_spa = dmu_objset_spa(os);
2774	zilog->zl_dmu_pool = dmu_objset_pool(os);
2775	zilog->zl_destroy_txg = TXG_INITIAL - 1;
2776	zilog->zl_logbias = dmu_objset_logbias(os);
2777	zilog->zl_sync = dmu_objset_syncprop(os);
2778	zilog->zl_dirty_max_txg = 0;
2779	zilog->zl_last_lwb_opened = NULL;
2780	zilog->zl_last_lwb_latency = 0;
2781
2782	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
2783	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
2784
2785	for (int i = 0; i < TXG_SIZE; i++) {
2786		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
2787		    MUTEX_DEFAULT, NULL);
2788	}
2789
2790	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
2791	    offsetof(lwb_t, lwb_node));
2792
2793	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
2794	    offsetof(itx_t, itx_node));
2795
2796	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
2797
2798	return (zilog);
2799}
2800
2801void
2802zil_free(zilog_t *zilog)
2803{
2804	zilog->zl_stop_sync = 1;
2805
2806	ASSERT0(zilog->zl_suspend);
2807	ASSERT0(zilog->zl_suspending);
2808
2809	ASSERT(list_is_empty(&zilog->zl_lwb_list));
2810	list_destroy(&zilog->zl_lwb_list);
2811
2812	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
2813	list_destroy(&zilog->zl_itx_commit_list);
2814
2815	for (int i = 0; i < TXG_SIZE; i++) {
2816		/*
2817		 * It's possible for an itx to be generated that doesn't dirty
2818		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
2819		 * callback to remove the entry. We remove those here.
2820		 *
2821		 * Also free up the ziltest itxs.
2822		 */
2823		if (zilog->zl_itxg[i].itxg_itxs)
2824			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
2825		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
2826	}
2827
2828	mutex_destroy(&zilog->zl_issuer_lock);
2829	mutex_destroy(&zilog->zl_lock);
2830
2831	cv_destroy(&zilog->zl_cv_suspend);
2832
2833	kmem_free(zilog, sizeof (zilog_t));
2834}
2835
2836/*
2837 * Open an intent log.
2838 */
2839zilog_t *
2840zil_open(objset_t *os, zil_get_data_t *get_data)
2841{
2842	zilog_t *zilog = dmu_objset_zil(os);
2843
2844	ASSERT3P(zilog->zl_get_data, ==, NULL);
2845	ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2846	ASSERT(list_is_empty(&zilog->zl_lwb_list));
2847
2848	zilog->zl_get_data = get_data;
2849
2850	return (zilog);
2851}
2852
2853/*
2854 * Close an intent log.
2855 */
2856void
2857zil_close(zilog_t *zilog)
2858{
2859	lwb_t *lwb;
2860	uint64_t txg;
2861
2862	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
2863		zil_commit(zilog, 0);
2864	} else {
2865		ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2866		ASSERT0(zilog->zl_dirty_max_txg);
2867		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
2868	}
2869
2870	mutex_enter(&zilog->zl_lock);
2871	lwb = list_tail(&zilog->zl_lwb_list);
2872	if (lwb == NULL)
2873		txg = zilog->zl_dirty_max_txg;
2874	else
2875		txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
2876	mutex_exit(&zilog->zl_lock);
2877
2878	/*
2879	 * We need to use txg_wait_synced() to wait long enough for the
2880	 * ZIL to be clean, and to wait for all pending lwbs to be
2881	 * written out.
2882	 */
2883	if (txg != 0)
2884		txg_wait_synced(zilog->zl_dmu_pool, txg);
2885
2886	if (zilog_is_dirty(zilog))
2887		zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
2888	VERIFY(!zilog_is_dirty(zilog));
2889
2890	zilog->zl_get_data = NULL;
2891
2892	/*
2893	 * We should have only one lwb left on the list; remove it now.
2894	 */
2895	mutex_enter(&zilog->zl_lock);
2896	lwb = list_head(&zilog->zl_lwb_list);
2897	if (lwb != NULL) {
2898		ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
2899		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2900		list_remove(&zilog->zl_lwb_list, lwb);
2901		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
2902		zil_free_lwb(zilog, lwb);
2903	}
2904	mutex_exit(&zilog->zl_lock);
2905}
2906
2907static char *suspend_tag = "zil suspending";
2908
2909/*
2910 * Suspend an intent log.  While in suspended mode, we still honor
2911 * synchronous semantics, but we rely on txg_wait_synced() to do it.
2912 * On old version pools, we suspend the log briefly when taking a
2913 * snapshot so that it will have an empty intent log.
2914 *
2915 * Long holds are not really intended to be used the way we do here --
2916 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
2917 * could fail.  Therefore we take pains to only put a long hold if it is
2918 * actually necessary.  Fortunately, it will only be necessary if the
2919 * objset is currently mounted (or the ZVOL equivalent).  In that case it
2920 * will already have a long hold, so we are not really making things any worse.
2921 *
2922 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
2923 * zvol_state_t), and use their mechanism to prevent their hold from being
2924 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
2925 * very little gain.
2926 *
2927 * if cookiep == NULL, this does both the suspend & resume.
2928 * Otherwise, it returns with the dataset "long held", and the cookie
2929 * should be passed into zil_resume().
2930 */
2931int
2932zil_suspend(const char *osname, void **cookiep)
2933{
2934	objset_t *os;
2935	zilog_t *zilog;
2936	const zil_header_t *zh;
2937	int error;
2938
2939	error = dmu_objset_hold(osname, suspend_tag, &os);
2940	if (error != 0)
2941		return (error);
2942	zilog = dmu_objset_zil(os);
2943
2944	mutex_enter(&zilog->zl_lock);
2945	zh = zilog->zl_header;
2946
2947	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
2948		mutex_exit(&zilog->zl_lock);
2949		dmu_objset_rele(os, suspend_tag);
2950		return (SET_ERROR(EBUSY));
2951	}
2952
2953	/*
2954	 * Don't put a long hold in the cases where we can avoid it.  This
2955	 * is when there is no cookie so we are doing a suspend & resume
2956	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
2957	 * for the suspend because it's already suspended, or there's no ZIL.
2958	 */
2959	if (cookiep == NULL && !zilog->zl_suspending &&
2960	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
2961		mutex_exit(&zilog->zl_lock);
2962		dmu_objset_rele(os, suspend_tag);
2963		return (0);
2964	}
2965
2966	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
2967	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
2968
2969	zilog->zl_suspend++;
2970
2971	if (zilog->zl_suspend > 1) {
2972		/*
2973		 * Someone else is already suspending it.
2974		 * Just wait for them to finish.
2975		 */
2976
2977		while (zilog->zl_suspending)
2978			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
2979		mutex_exit(&zilog->zl_lock);
2980
2981		if (cookiep == NULL)
2982			zil_resume(os);
2983		else
2984			*cookiep = os;
2985		return (0);
2986	}
2987
2988	/*
2989	 * If there is no pointer to an on-disk block, this ZIL must not
2990	 * be active (e.g. filesystem not mounted), so there's nothing
2991	 * to clean up.
2992	 */
2993	if (BP_IS_HOLE(&zh->zh_log)) {
2994		ASSERT(cookiep != NULL); /* fast path already handled */
2995
2996		*cookiep = os;
2997		mutex_exit(&zilog->zl_lock);
2998		return (0);
2999	}
3000
3001	zilog->zl_suspending = B_TRUE;
3002	mutex_exit(&zilog->zl_lock);
3003
3004	/*
3005	 * We need to use zil_commit_impl to ensure we wait for all
3006	 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwb's to be committed
3007	 * to disk before proceeding. If we used zil_commit instead, it
3008	 * would just call txg_wait_synced(), because zl_suspend is set.
3009	 * txg_wait_synced() doesn't wait for these lwb's to be
3010	 * LWB_STATE_DONE before returning.
3011	 */
3012	zil_commit_impl(zilog, 0);
3013
3014	/*
3015	 * Now that we've ensured all lwb's are LWB_STATE_DONE, we use
3016	 * txg_wait_synced() to ensure the data from the zilog has
3017	 * migrated to the main pool before calling zil_destroy().
3018	 */
3019	txg_wait_synced(zilog->zl_dmu_pool, 0);
3020
3021	zil_destroy(zilog, B_FALSE);
3022
3023	mutex_enter(&zilog->zl_lock);
3024	zilog->zl_suspending = B_FALSE;
3025	cv_broadcast(&zilog->zl_cv_suspend);
3026	mutex_exit(&zilog->zl_lock);
3027
3028	if (cookiep == NULL)
3029		zil_resume(os);
3030	else
3031		*cookiep = os;
3032	return (0);
3033}
3034
3035void
3036zil_resume(void *cookie)
3037{
3038	objset_t *os = cookie;
3039	zilog_t *zilog = dmu_objset_zil(os);
3040
3041	mutex_enter(&zilog->zl_lock);
3042	ASSERT(zilog->zl_suspend != 0);
3043	zilog->zl_suspend--;
3044	mutex_exit(&zilog->zl_lock);
3045	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3046	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3047}
3048
3049typedef struct zil_replay_arg {
3050	zil_replay_func_t **zr_replay;
3051	void		*zr_arg;
3052	boolean_t	zr_byteswap;
3053	char		*zr_lr;
3054} zil_replay_arg_t;
3055
3056static int
3057zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
3058{
3059	char name[ZFS_MAX_DATASET_NAME_LEN];
3060
3061	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
3062
3063	dmu_objset_name(zilog->zl_os, name);
3064
3065	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3066	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3067	    (u_longlong_t)lr->lrc_seq,
3068	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3069	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
3070
3071	return (error);
3072}
3073
3074static int
3075zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
3076{
3077	zil_replay_arg_t *zr = zra;
3078	const zil_header_t *zh = zilog->zl_header;
3079	uint64_t reclen = lr->lrc_reclen;
3080	uint64_t txtype = lr->lrc_txtype;
3081	int error = 0;
3082
3083	zilog->zl_replaying_seq = lr->lrc_seq;
3084
3085	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
3086		return (0);
3087
3088	if (lr->lrc_txg < claim_txg)		/* already committed */
3089		return (0);
3090
3091	/* Strip case-insensitive bit, still present in log record */
3092	txtype &= ~TX_CI;
3093
3094	if (txtype == 0 || txtype >= TX_MAX_TYPE)
3095		return (zil_replay_error(zilog, lr, EINVAL));
3096
3097	/*
3098	 * If this record type can be logged out of order, the object
3099	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
3100	 */
3101	if (TX_OOO(txtype)) {
3102		error = dmu_object_info(zilog->zl_os,
3103		    ((lr_ooo_t *)lr)->lr_foid, NULL);
3104		if (error == ENOENT || error == EEXIST)
3105			return (0);
3106	}
3107
3108	/*
3109	 * Make a copy of the data so we can revise and extend it.
3110	 */
3111	bcopy(lr, zr->zr_lr, reclen);
3112
3113	/*
3114	 * If this is a TX_WRITE with a blkptr, suck in the data.
3115	 */
3116	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3117		error = zil_read_log_data(zilog, (lr_write_t *)lr,
3118		    zr->zr_lr + reclen);
3119		if (error != 0)
3120			return (zil_replay_error(zilog, lr, error));
3121	}
3122
3123	/*
3124	 * The log block containing this lr may have been byteswapped
3125	 * so that we can easily examine common fields like lrc_txtype.
3126	 * However, the log is a mix of different record types, and only the
3127	 * replay vectors know how to byteswap their records.  Therefore, if
3128	 * the lr was byteswapped, undo it before invoking the replay vector.
3129	 */
3130	if (zr->zr_byteswap)
3131		byteswap_uint64_array(zr->zr_lr, reclen);
3132
3133	/*
3134	 * We must now do two things atomically: replay this log record,
3135	 * and update the log header sequence number to reflect the fact that
3136	 * we did so. At the end of each replay function the sequence number
3137	 * is updated if we are in replay mode.
3138	 */
3139	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3140	if (error != 0) {
3141		/*
3142		 * The DMU's dnode layer doesn't see removes until the txg
3143		 * commits, so a subsequent claim can spuriously fail with
3144		 * EEXIST. So if we receive any error we try syncing out
3145		 * any removes then retry the transaction.  Note that we
3146		 * specify B_FALSE for byteswap now, so we don't do it twice.
3147		 */
3148		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3149		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3150		if (error != 0)
3151			return (zil_replay_error(zilog, lr, error));
3152	}
3153	return (0);
3154}
3155
3156/* ARGSUSED */
3157static int
3158zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
3159{
3160	zilog->zl_replay_blks++;
3161
3162	return (0);
3163}
3164
3165/*
3166 * If this dataset has a non-empty intent log, replay it and destroy it.
3167 */
3168void
3169zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
3170{
3171	zilog_t *zilog = dmu_objset_zil(os);
3172	const zil_header_t *zh = zilog->zl_header;
3173	zil_replay_arg_t zr;
3174
3175	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3176		zil_destroy(zilog, B_TRUE);
3177		return;
3178	}
3179
3180	zr.zr_replay = replay_func;
3181	zr.zr_arg = arg;
3182	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3183	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3184
3185	/*
3186	 * Wait for in-progress removes to sync before starting replay.
3187	 */
3188	txg_wait_synced(zilog->zl_dmu_pool, 0);
3189
3190	zilog->zl_replay = B_TRUE;
3191	zilog->zl_replay_time = ddi_get_lbolt();
3192	ASSERT(zilog->zl_replay_blks == 0);
3193	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3194	    zh->zh_claim_txg);
3195	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3196
3197	zil_destroy(zilog, B_FALSE);
3198	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3199	zilog->zl_replay = B_FALSE;
3200}
3201
3202boolean_t
3203zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3204{
3205	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3206		return (B_TRUE);
3207
3208	if (zilog->zl_replay) {
3209		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3210		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3211		    zilog->zl_replaying_seq;
3212		return (B_TRUE);
3213	}
3214
3215	return (B_FALSE);
3216}
3217
3218/* ARGSUSED */
3219int
3220zil_reset(const char *osname, void *arg)
3221{
3222	int error;
3223
3224	error = zil_suspend(osname, NULL);
3225	if (error != 0)
3226		return (SET_ERROR(EEXIST));
3227	return (0);
3228}
3229