1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 */
26
27/* Portions Copyright 2010 Robert Milkowski */
28
29#include <sys/zfs_context.h>
30#include <sys/spa.h>
31#include <sys/spa_impl.h>
32#include <sys/dmu.h>
33#include <sys/zap.h>
34#include <sys/arc.h>
35#include <sys/stat.h>
36#include <sys/resource.h>
37#include <sys/zil.h>
38#include <sys/zil_impl.h>
39#include <sys/dsl_dataset.h>
40#include <sys/vdev_impl.h>
41#include <sys/dmu_tx.h>
42#include <sys/dsl_pool.h>
43#include <sys/abd.h>
44
45/*
46 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
47 * calls that change the file system. Each itx has enough information to
48 * be able to replay them after a system crash, power loss, or
49 * equivalent failure mode. These are stored in memory until either:
50 *
51 *   1. they are committed to the pool by the DMU transaction group
52 *      (txg), at which point they can be discarded; or
53 *   2. they are committed to the on-disk ZIL for the dataset being
54 *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
55 *      requirement).
56 *
57 * In the event of a crash or power loss, the itxs contained by each
58 * dataset's on-disk ZIL will be replayed when that dataset is first
59 * instantianted (e.g. if the dataset is a normal fileystem, when it is
60 * first mounted).
61 *
62 * As hinted at above, there is one ZIL per dataset (both the in-memory
63 * representation, and the on-disk representation). The on-disk format
64 * consists of 3 parts:
65 *
66 * 	- a single, per-dataset, ZIL header; which points to a chain of
67 * 	- zero or more ZIL blocks; each of which contains
68 * 	- zero or more ZIL records
69 *
70 * A ZIL record holds the information necessary to replay a single
71 * system call transaction. A ZIL block can hold many ZIL records, and
72 * the blocks are chained together, similarly to a singly linked list.
73 *
74 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
75 * block in the chain, and the ZIL header points to the first block in
76 * the chain.
77 *
78 * Note, there is not a fixed place in the pool to hold these ZIL
79 * blocks; they are dynamically allocated and freed as needed from the
80 * blocks available on the pool, though they can be preferentially
81 * allocated from a dedicated "log" vdev.
82 */
83
84/*
85 * This controls the amount of time that a ZIL block (lwb) will remain
86 * "open" when it isn't "full", and it has a thread waiting for it to be
87 * committed to stable storage. Please refer to the zil_commit_waiter()
88 * function (and the comments within it) for more details.
89 */
90int zfs_commit_timeout_pct = 5;
91
92/*
93 * Disable intent logging replay.  This global ZIL switch affects all pools.
94 */
95int zil_replay_disable = 0;
96
97/*
98 * Tunable parameter for debugging or performance analysis.  Setting
99 * zfs_nocacheflush will cause corruption on power loss if a volatile
100 * out-of-order write cache is enabled.
101 */
102boolean_t zfs_nocacheflush = B_FALSE;
103
104/*
105 * Limit SLOG write size per commit executed with synchronous priority.
106 * Any writes above that will be executed with lower (asynchronous) priority
107 * to limit potential SLOG device abuse by single active ZIL writer.
108 */
109uint64_t zil_slog_bulk = 768 * 1024;
110
111static kmem_cache_t *zil_lwb_cache;
112static kmem_cache_t *zil_zcw_cache;
113
114static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
115
116#define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
117    sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
118
119static int
120zil_bp_compare(const void *x1, const void *x2)
121{
122	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
123	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
124
125	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
126		return (-1);
127	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
128		return (1);
129
130	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
131		return (-1);
132	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
133		return (1);
134
135	return (0);
136}
137
138static void
139zil_bp_tree_init(zilog_t *zilog)
140{
141	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
142	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
143}
144
145static void
146zil_bp_tree_fini(zilog_t *zilog)
147{
148	avl_tree_t *t = &zilog->zl_bp_tree;
149	zil_bp_node_t *zn;
150	void *cookie = NULL;
151
152	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
153		kmem_free(zn, sizeof (zil_bp_node_t));
154
155	avl_destroy(t);
156}
157
158int
159zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
160{
161	avl_tree_t *t = &zilog->zl_bp_tree;
162	const dva_t *dva;
163	zil_bp_node_t *zn;
164	avl_index_t where;
165
166	if (BP_IS_EMBEDDED(bp))
167		return (0);
168
169	dva = BP_IDENTITY(bp);
170
171	if (avl_find(t, dva, &where) != NULL)
172		return (SET_ERROR(EEXIST));
173
174	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
175	zn->zn_dva = *dva;
176	avl_insert(t, zn, where);
177
178	return (0);
179}
180
181static zil_header_t *
182zil_header_in_syncing_context(zilog_t *zilog)
183{
184	return ((zil_header_t *)zilog->zl_header);
185}
186
187static void
188zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
189{
190	zio_cksum_t *zc = &bp->blk_cksum;
191
192	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
193	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
194	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
195	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
196}
197
198/*
199 * Read a log block and make sure it's valid.
200 */
201static int
202zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
203    char **end)
204{
205	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
206	arc_flags_t aflags = ARC_FLAG_WAIT;
207	arc_buf_t *abuf = NULL;
208	zbookmark_phys_t zb;
209	int error;
210
211	if (zilog->zl_header->zh_claim_txg == 0)
212		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
213
214	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
215		zio_flags |= ZIO_FLAG_SPECULATIVE;
216
217	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
218	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
219
220	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
221	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
222
223	if (error == 0) {
224		zio_cksum_t cksum = bp->blk_cksum;
225
226		/*
227		 * Validate the checksummed log block.
228		 *
229		 * Sequence numbers should be... sequential.  The checksum
230		 * verifier for the next block should be bp's checksum plus 1.
231		 *
232		 * Also check the log chain linkage and size used.
233		 */
234		cksum.zc_word[ZIL_ZC_SEQ]++;
235
236		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
237			zil_chain_t *zilc = abuf->b_data;
238			char *lr = (char *)(zilc + 1);
239			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
240
241			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
242			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
243				error = SET_ERROR(ECKSUM);
244			} else {
245				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
246				bcopy(lr, dst, len);
247				*end = (char *)dst + len;
248				*nbp = zilc->zc_next_blk;
249			}
250		} else {
251			char *lr = abuf->b_data;
252			uint64_t size = BP_GET_LSIZE(bp);
253			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
254
255			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
256			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
257			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
258				error = SET_ERROR(ECKSUM);
259			} else {
260				ASSERT3U(zilc->zc_nused, <=,
261				    SPA_OLD_MAXBLOCKSIZE);
262				bcopy(lr, dst, zilc->zc_nused);
263				*end = (char *)dst + zilc->zc_nused;
264				*nbp = zilc->zc_next_blk;
265			}
266		}
267
268		arc_buf_destroy(abuf, &abuf);
269	}
270
271	return (error);
272}
273
274/*
275 * Read a TX_WRITE log data block.
276 */
277static int
278zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
279{
280	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
281	const blkptr_t *bp = &lr->lr_blkptr;
282	arc_flags_t aflags = ARC_FLAG_WAIT;
283	arc_buf_t *abuf = NULL;
284	zbookmark_phys_t zb;
285	int error;
286
287	if (BP_IS_HOLE(bp)) {
288		if (wbuf != NULL)
289			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
290		return (0);
291	}
292
293	if (zilog->zl_header->zh_claim_txg == 0)
294		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
295
296	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
297	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
298
299	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
300	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
301
302	if (error == 0) {
303		if (wbuf != NULL)
304			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
305		arc_buf_destroy(abuf, &abuf);
306	}
307
308	return (error);
309}
310
311/*
312 * Parse the intent log, and call parse_func for each valid record within.
313 */
314int
315zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
316    zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
317{
318	const zil_header_t *zh = zilog->zl_header;
319	boolean_t claimed = !!zh->zh_claim_txg;
320	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
321	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
322	uint64_t max_blk_seq = 0;
323	uint64_t max_lr_seq = 0;
324	uint64_t blk_count = 0;
325	uint64_t lr_count = 0;
326	blkptr_t blk, next_blk;
327	char *lrbuf, *lrp;
328	int error = 0;
329
330	/*
331	 * Old logs didn't record the maximum zh_claim_lr_seq.
332	 */
333	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
334		claim_lr_seq = UINT64_MAX;
335
336	/*
337	 * Starting at the block pointed to by zh_log we read the log chain.
338	 * For each block in the chain we strongly check that block to
339	 * ensure its validity.  We stop when an invalid block is found.
340	 * For each block pointer in the chain we call parse_blk_func().
341	 * For each record in each valid block we call parse_lr_func().
342	 * If the log has been claimed, stop if we encounter a sequence
343	 * number greater than the highest claimed sequence number.
344	 */
345	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
346	zil_bp_tree_init(zilog);
347
348	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
349		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
350		int reclen;
351		char *end;
352
353		if (blk_seq > claim_blk_seq)
354			break;
355		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
356			break;
357		ASSERT3U(max_blk_seq, <, blk_seq);
358		max_blk_seq = blk_seq;
359		blk_count++;
360
361		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
362			break;
363
364		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
365		if (error != 0)
366			break;
367
368		for (lrp = lrbuf; lrp < end; lrp += reclen) {
369			lr_t *lr = (lr_t *)lrp;
370			reclen = lr->lrc_reclen;
371			ASSERT3U(reclen, >=, sizeof (lr_t));
372			if (lr->lrc_seq > claim_lr_seq)
373				goto done;
374			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
375				goto done;
376			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
377			max_lr_seq = lr->lrc_seq;
378			lr_count++;
379		}
380	}
381done:
382	zilog->zl_parse_error = error;
383	zilog->zl_parse_blk_seq = max_blk_seq;
384	zilog->zl_parse_lr_seq = max_lr_seq;
385	zilog->zl_parse_blk_count = blk_count;
386	zilog->zl_parse_lr_count = lr_count;
387
388	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
389	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
390
391	zil_bp_tree_fini(zilog);
392	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
393
394	return (error);
395}
396
397/* ARGSUSED */
398static int
399zil_clear_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
400{
401	ASSERT(!BP_IS_HOLE(bp));
402
403	/*
404	 * As we call this function from the context of a rewind to a
405	 * checkpoint, each ZIL block whose txg is later than the txg
406	 * that we rewind to is invalid. Thus, we return -1 so
407	 * zil_parse() doesn't attempt to read it.
408	 */
409	if (bp->blk_birth >= first_txg)
410		return (-1);
411
412	if (zil_bp_tree_add(zilog, bp) != 0)
413		return (0);
414
415	zio_free(zilog->zl_spa, first_txg, bp);
416	return (0);
417}
418
419/* ARGSUSED */
420static int
421zil_noop_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
422{
423	return (0);
424}
425
426static int
427zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
428{
429	/*
430	 * Claim log block if not already committed and not already claimed.
431	 * If tx == NULL, just verify that the block is claimable.
432	 */
433	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
434	    zil_bp_tree_add(zilog, bp) != 0)
435		return (0);
436
437	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
438	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
439	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
440}
441
442static int
443zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
444{
445	lr_write_t *lr = (lr_write_t *)lrc;
446	int error;
447
448	if (lrc->lrc_txtype != TX_WRITE)
449		return (0);
450
451	/*
452	 * If the block is not readable, don't claim it.  This can happen
453	 * in normal operation when a log block is written to disk before
454	 * some of the dmu_sync() blocks it points to.  In this case, the
455	 * transaction cannot have been committed to anyone (we would have
456	 * waited for all writes to be stable first), so it is semantically
457	 * correct to declare this the end of the log.
458	 */
459	if (lr->lr_blkptr.blk_birth >= first_txg &&
460	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
461		return (error);
462	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
463}
464
465/* ARGSUSED */
466static int
467zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
468{
469	zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
470
471	return (0);
472}
473
474static int
475zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
476{
477	lr_write_t *lr = (lr_write_t *)lrc;
478	blkptr_t *bp = &lr->lr_blkptr;
479
480	/*
481	 * If we previously claimed it, we need to free it.
482	 */
483	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
484	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
485	    !BP_IS_HOLE(bp))
486		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
487
488	return (0);
489}
490
491static int
492zil_lwb_vdev_compare(const void *x1, const void *x2)
493{
494	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
495	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
496
497	if (v1 < v2)
498		return (-1);
499	if (v1 > v2)
500		return (1);
501
502	return (0);
503}
504
505static lwb_t *
506zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
507{
508	lwb_t *lwb;
509
510	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
511	lwb->lwb_zilog = zilog;
512	lwb->lwb_blk = *bp;
513	lwb->lwb_slog = slog;
514	lwb->lwb_state = LWB_STATE_CLOSED;
515	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
516	lwb->lwb_max_txg = txg;
517	lwb->lwb_write_zio = NULL;
518	lwb->lwb_root_zio = NULL;
519	lwb->lwb_tx = NULL;
520	lwb->lwb_issued_timestamp = 0;
521	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
522		lwb->lwb_nused = sizeof (zil_chain_t);
523		lwb->lwb_sz = BP_GET_LSIZE(bp);
524	} else {
525		lwb->lwb_nused = 0;
526		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
527	}
528
529	mutex_enter(&zilog->zl_lock);
530	list_insert_tail(&zilog->zl_lwb_list, lwb);
531	mutex_exit(&zilog->zl_lock);
532
533	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
534	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
535	VERIFY(list_is_empty(&lwb->lwb_waiters));
536
537	return (lwb);
538}
539
540static void
541zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
542{
543	ASSERT(MUTEX_HELD(&zilog->zl_lock));
544	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
545	VERIFY(list_is_empty(&lwb->lwb_waiters));
546	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
547	ASSERT3P(lwb->lwb_write_zio, ==, NULL);
548	ASSERT3P(lwb->lwb_root_zio, ==, NULL);
549	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
550	ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
551	    lwb->lwb_state == LWB_STATE_DONE);
552
553	/*
554	 * Clear the zilog's field to indicate this lwb is no longer
555	 * valid, and prevent use-after-free errors.
556	 */
557	if (zilog->zl_last_lwb_opened == lwb)
558		zilog->zl_last_lwb_opened = NULL;
559
560	kmem_cache_free(zil_lwb_cache, lwb);
561}
562
563/*
564 * Called when we create in-memory log transactions so that we know
565 * to cleanup the itxs at the end of spa_sync().
566 */
567void
568zilog_dirty(zilog_t *zilog, uint64_t txg)
569{
570	dsl_pool_t *dp = zilog->zl_dmu_pool;
571	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
572
573	ASSERT(spa_writeable(zilog->zl_spa));
574
575	if (ds->ds_is_snapshot)
576		panic("dirtying snapshot!");
577
578	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
579		/* up the hold count until we can be written out */
580		dmu_buf_add_ref(ds->ds_dbuf, zilog);
581
582		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
583	}
584}
585
586/*
587 * Determine if the zil is dirty in the specified txg. Callers wanting to
588 * ensure that the dirty state does not change must hold the itxg_lock for
589 * the specified txg. Holding the lock will ensure that the zil cannot be
590 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
591 * state.
592 */
593boolean_t
594zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
595{
596	dsl_pool_t *dp = zilog->zl_dmu_pool;
597
598	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
599		return (B_TRUE);
600	return (B_FALSE);
601}
602
603/*
604 * Determine if the zil is dirty. The zil is considered dirty if it has
605 * any pending itx records that have not been cleaned by zil_clean().
606 */
607boolean_t
608zilog_is_dirty(zilog_t *zilog)
609{
610	dsl_pool_t *dp = zilog->zl_dmu_pool;
611
612	for (int t = 0; t < TXG_SIZE; t++) {
613		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
614			return (B_TRUE);
615	}
616	return (B_FALSE);
617}
618
619/*
620 * Create an on-disk intent log.
621 */
622static lwb_t *
623zil_create(zilog_t *zilog)
624{
625	const zil_header_t *zh = zilog->zl_header;
626	lwb_t *lwb = NULL;
627	uint64_t txg = 0;
628	dmu_tx_t *tx = NULL;
629	blkptr_t blk;
630	int error = 0;
631	boolean_t slog = FALSE;
632
633	/*
634	 * Wait for any previous destroy to complete.
635	 */
636	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
637
638	ASSERT(zh->zh_claim_txg == 0);
639	ASSERT(zh->zh_replay_seq == 0);
640
641	blk = zh->zh_log;
642
643	/*
644	 * Allocate an initial log block if:
645	 *    - there isn't one already
646	 *    - the existing block is the wrong endianess
647	 */
648	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
649		tx = dmu_tx_create(zilog->zl_os);
650		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
651		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
652		txg = dmu_tx_get_txg(tx);
653
654		if (!BP_IS_HOLE(&blk)) {
655			zio_free(zilog->zl_spa, txg, &blk);
656			BP_ZERO(&blk);
657		}
658
659		error = zio_alloc_zil(zilog->zl_spa,
660		    zilog->zl_os->os_dsl_dataset->ds_object, txg, &blk, NULL,
661		    ZIL_MIN_BLKSZ, &slog);
662
663		if (error == 0)
664			zil_init_log_chain(zilog, &blk);
665	}
666
667	/*
668	 * Allocate a log write block (lwb) for the first log block.
669	 */
670	if (error == 0)
671		lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
672
673	/*
674	 * If we just allocated the first log block, commit our transaction
675	 * and wait for zil_sync() to stuff the block poiner into zh_log.
676	 * (zh is part of the MOS, so we cannot modify it in open context.)
677	 */
678	if (tx != NULL) {
679		dmu_tx_commit(tx);
680		txg_wait_synced(zilog->zl_dmu_pool, txg);
681	}
682
683	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
684
685	return (lwb);
686}
687
688/*
689 * In one tx, free all log blocks and clear the log header. If keep_first
690 * is set, then we're replaying a log with no content. We want to keep the
691 * first block, however, so that the first synchronous transaction doesn't
692 * require a txg_wait_synced() in zil_create(). We don't need to
693 * txg_wait_synced() here either when keep_first is set, because both
694 * zil_create() and zil_destroy() will wait for any in-progress destroys
695 * to complete.
696 */
697void
698zil_destroy(zilog_t *zilog, boolean_t keep_first)
699{
700	const zil_header_t *zh = zilog->zl_header;
701	lwb_t *lwb;
702	dmu_tx_t *tx;
703	uint64_t txg;
704
705	/*
706	 * Wait for any previous destroy to complete.
707	 */
708	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
709
710	zilog->zl_old_header = *zh;		/* debugging aid */
711
712	if (BP_IS_HOLE(&zh->zh_log))
713		return;
714
715	tx = dmu_tx_create(zilog->zl_os);
716	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
717	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
718	txg = dmu_tx_get_txg(tx);
719
720	mutex_enter(&zilog->zl_lock);
721
722	ASSERT3U(zilog->zl_destroy_txg, <, txg);
723	zilog->zl_destroy_txg = txg;
724	zilog->zl_keep_first = keep_first;
725
726	if (!list_is_empty(&zilog->zl_lwb_list)) {
727		ASSERT(zh->zh_claim_txg == 0);
728		VERIFY(!keep_first);
729		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
730			list_remove(&zilog->zl_lwb_list, lwb);
731			if (lwb->lwb_buf != NULL)
732				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
733			zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
734			zil_free_lwb(zilog, lwb);
735		}
736	} else if (!keep_first) {
737		zil_destroy_sync(zilog, tx);
738	}
739	mutex_exit(&zilog->zl_lock);
740
741	dmu_tx_commit(tx);
742}
743
744void
745zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
746{
747	ASSERT(list_is_empty(&zilog->zl_lwb_list));
748	(void) zil_parse(zilog, zil_free_log_block,
749	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
750}
751
752int
753zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
754{
755	dmu_tx_t *tx = txarg;
756	zilog_t *zilog;
757	uint64_t first_txg;
758	zil_header_t *zh;
759	objset_t *os;
760	int error;
761
762	error = dmu_objset_own_obj(dp, ds->ds_object,
763	    DMU_OST_ANY, B_FALSE, FTAG, &os);
764	if (error != 0) {
765		/*
766		 * EBUSY indicates that the objset is inconsistent, in which
767		 * case it can not have a ZIL.
768		 */
769		if (error != EBUSY) {
770			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
771			    (unsigned long long)ds->ds_object, error);
772		}
773		return (0);
774	}
775
776	zilog = dmu_objset_zil(os);
777	zh = zil_header_in_syncing_context(zilog);
778	ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
779	first_txg = spa_min_claim_txg(zilog->zl_spa);
780
781	/*
782	 * If the spa_log_state is not set to be cleared, check whether
783	 * the current uberblock is a checkpoint one and if the current
784	 * header has been claimed before moving on.
785	 *
786	 * If the current uberblock is a checkpointed uberblock then
787	 * one of the following scenarios took place:
788	 *
789	 * 1] We are currently rewinding to the checkpoint of the pool.
790	 * 2] We crashed in the middle of a checkpoint rewind but we
791	 *    did manage to write the checkpointed uberblock to the
792	 *    vdev labels, so when we tried to import the pool again
793	 *    the checkpointed uberblock was selected from the import
794	 *    procedure.
795	 *
796	 * In both cases we want to zero out all the ZIL blocks, except
797	 * the ones that have been claimed at the time of the checkpoint
798	 * (their zh_claim_txg != 0). The reason is that these blocks
799	 * may be corrupted since we may have reused their locations on
800	 * disk after we took the checkpoint.
801	 *
802	 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
803	 * when we first figure out whether the current uberblock is
804	 * checkpointed or not. Unfortunately, that would discard all
805	 * the logs, including the ones that are claimed, and we would
806	 * leak space.
807	 */
808	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
809	    (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
810	    zh->zh_claim_txg == 0)) {
811		if (!BP_IS_HOLE(&zh->zh_log)) {
812			(void) zil_parse(zilog, zil_clear_log_block,
813			    zil_noop_log_record, tx, first_txg);
814		}
815		BP_ZERO(&zh->zh_log);
816		dsl_dataset_dirty(dmu_objset_ds(os), tx);
817		dmu_objset_disown(os, FTAG);
818		return (0);
819	}
820
821	/*
822	 * If we are not rewinding and opening the pool normally, then
823	 * the min_claim_txg should be equal to the first txg of the pool.
824	 */
825	ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
826
827	/*
828	 * Claim all log blocks if we haven't already done so, and remember
829	 * the highest claimed sequence number.  This ensures that if we can
830	 * read only part of the log now (e.g. due to a missing device),
831	 * but we can read the entire log later, we will not try to replay
832	 * or destroy beyond the last block we successfully claimed.
833	 */
834	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
835	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
836		(void) zil_parse(zilog, zil_claim_log_block,
837		    zil_claim_log_record, tx, first_txg);
838		zh->zh_claim_txg = first_txg;
839		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
840		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
841		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
842			zh->zh_flags |= ZIL_REPLAY_NEEDED;
843		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
844		dsl_dataset_dirty(dmu_objset_ds(os), tx);
845	}
846
847	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
848	dmu_objset_disown(os, FTAG);
849	return (0);
850}
851
852/*
853 * Check the log by walking the log chain.
854 * Checksum errors are ok as they indicate the end of the chain.
855 * Any other error (no device or read failure) returns an error.
856 */
857/* ARGSUSED */
858int
859zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
860{
861	zilog_t *zilog;
862	objset_t *os;
863	blkptr_t *bp;
864	int error;
865
866	ASSERT(tx == NULL);
867
868	error = dmu_objset_from_ds(ds, &os);
869	if (error != 0) {
870		cmn_err(CE_WARN, "can't open objset %llu, error %d",
871		    (unsigned long long)ds->ds_object, error);
872		return (0);
873	}
874
875	zilog = dmu_objset_zil(os);
876	bp = (blkptr_t *)&zilog->zl_header->zh_log;
877
878	if (!BP_IS_HOLE(bp)) {
879		vdev_t *vd;
880		boolean_t valid = B_TRUE;
881
882		/*
883		 * Check the first block and determine if it's on a log device
884		 * which may have been removed or faulted prior to loading this
885		 * pool.  If so, there's no point in checking the rest of the
886		 * log as its content should have already been synced to the
887		 * pool.
888		 */
889		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
890		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
891		if (vd->vdev_islog && vdev_is_dead(vd))
892			valid = vdev_log_state_valid(vd);
893		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
894
895		if (!valid)
896			return (0);
897
898		/*
899		 * Check whether the current uberblock is checkpointed (e.g.
900		 * we are rewinding) and whether the current header has been
901		 * claimed or not. If it hasn't then skip verifying it. We
902		 * do this because its ZIL blocks may be part of the pool's
903		 * state before the rewind, which is no longer valid.
904		 */
905		zil_header_t *zh = zil_header_in_syncing_context(zilog);
906		if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
907		    zh->zh_claim_txg == 0)
908			return (0);
909	}
910
911	/*
912	 * Because tx == NULL, zil_claim_log_block() will not actually claim
913	 * any blocks, but just determine whether it is possible to do so.
914	 * In addition to checking the log chain, zil_claim_log_block()
915	 * will invoke zio_claim() with a done func of spa_claim_notify(),
916	 * which will update spa_max_claim_txg.  See spa_load() for details.
917	 */
918	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
919	    zilog->zl_header->zh_claim_txg ? -1ULL :
920	    spa_min_claim_txg(os->os_spa));
921
922	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
923}
924
925/*
926 * When an itx is "skipped", this function is used to properly mark the
927 * waiter as "done, and signal any thread(s) waiting on it. An itx can
928 * be skipped (and not committed to an lwb) for a variety of reasons,
929 * one of them being that the itx was committed via spa_sync(), prior to
930 * it being committed to an lwb; this can happen if a thread calling
931 * zil_commit() is racing with spa_sync().
932 */
933static void
934zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
935{
936	mutex_enter(&zcw->zcw_lock);
937	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
938	zcw->zcw_done = B_TRUE;
939	cv_broadcast(&zcw->zcw_cv);
940	mutex_exit(&zcw->zcw_lock);
941}
942
943/*
944 * This function is used when the given waiter is to be linked into an
945 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
946 * At this point, the waiter will no longer be referenced by the itx,
947 * and instead, will be referenced by the lwb.
948 */
949static void
950zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
951{
952	/*
953	 * The lwb_waiters field of the lwb is protected by the zilog's
954	 * zl_lock, thus it must be held when calling this function.
955	 */
956	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
957
958	mutex_enter(&zcw->zcw_lock);
959	ASSERT(!list_link_active(&zcw->zcw_node));
960	ASSERT3P(zcw->zcw_lwb, ==, NULL);
961	ASSERT3P(lwb, !=, NULL);
962	ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
963	    lwb->lwb_state == LWB_STATE_ISSUED);
964
965	list_insert_tail(&lwb->lwb_waiters, zcw);
966	zcw->zcw_lwb = lwb;
967	mutex_exit(&zcw->zcw_lock);
968}
969
970/*
971 * This function is used when zio_alloc_zil() fails to allocate a ZIL
972 * block, and the given waiter must be linked to the "nolwb waiters"
973 * list inside of zil_process_commit_list().
974 */
975static void
976zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
977{
978	mutex_enter(&zcw->zcw_lock);
979	ASSERT(!list_link_active(&zcw->zcw_node));
980	ASSERT3P(zcw->zcw_lwb, ==, NULL);
981	list_insert_tail(nolwb, zcw);
982	mutex_exit(&zcw->zcw_lock);
983}
984
985void
986zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
987{
988	avl_tree_t *t = &lwb->lwb_vdev_tree;
989	avl_index_t where;
990	zil_vdev_node_t *zv, zvsearch;
991	int ndvas = BP_GET_NDVAS(bp);
992	int i;
993
994	if (zfs_nocacheflush)
995		return;
996
997	mutex_enter(&lwb->lwb_vdev_lock);
998	for (i = 0; i < ndvas; i++) {
999		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1000		if (avl_find(t, &zvsearch, &where) == NULL) {
1001			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1002			zv->zv_vdev = zvsearch.zv_vdev;
1003			avl_insert(t, zv, where);
1004		}
1005	}
1006	mutex_exit(&lwb->lwb_vdev_lock);
1007}
1008
1009void
1010zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1011{
1012	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1013}
1014
1015/*
1016 * This function is a called after all VDEVs associated with a given lwb
1017 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1018 * as the lwb write completes, if "zfs_nocacheflush" is set.
1019 *
1020 * The intention is for this function to be called as soon as the
1021 * contents of an lwb are considered "stable" on disk, and will survive
1022 * any sudden loss of power. At this point, any threads waiting for the
1023 * lwb to reach this state are signalled, and the "waiter" structures
1024 * are marked "done".
1025 */
1026static void
1027zil_lwb_flush_vdevs_done(zio_t *zio)
1028{
1029	lwb_t *lwb = zio->io_private;
1030	zilog_t *zilog = lwb->lwb_zilog;
1031	dmu_tx_t *tx = lwb->lwb_tx;
1032	zil_commit_waiter_t *zcw;
1033
1034	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1035
1036	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1037
1038	mutex_enter(&zilog->zl_lock);
1039
1040	/*
1041	 * Ensure the lwb buffer pointer is cleared before releasing the
1042	 * txg. If we have had an allocation failure and the txg is
1043	 * waiting to sync then we want zil_sync() to remove the lwb so
1044	 * that it's not picked up as the next new one in
1045	 * zil_process_commit_list(). zil_sync() will only remove the
1046	 * lwb if lwb_buf is null.
1047	 */
1048	lwb->lwb_buf = NULL;
1049	lwb->lwb_tx = NULL;
1050
1051	ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1052	zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1053
1054	lwb->lwb_root_zio = NULL;
1055	lwb->lwb_state = LWB_STATE_DONE;
1056
1057	if (zilog->zl_last_lwb_opened == lwb) {
1058		/*
1059		 * Remember the highest committed log sequence number
1060		 * for ztest. We only update this value when all the log
1061		 * writes succeeded, because ztest wants to ASSERT that
1062		 * it got the whole log chain.
1063		 */
1064		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1065	}
1066
1067	while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1068		mutex_enter(&zcw->zcw_lock);
1069
1070		ASSERT(list_link_active(&zcw->zcw_node));
1071		list_remove(&lwb->lwb_waiters, zcw);
1072
1073		ASSERT3P(zcw->zcw_lwb, ==, lwb);
1074		zcw->zcw_lwb = NULL;
1075
1076		zcw->zcw_zio_error = zio->io_error;
1077
1078		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1079		zcw->zcw_done = B_TRUE;
1080		cv_broadcast(&zcw->zcw_cv);
1081
1082		mutex_exit(&zcw->zcw_lock);
1083	}
1084
1085	mutex_exit(&zilog->zl_lock);
1086
1087	/*
1088	 * Now that we've written this log block, we have a stable pointer
1089	 * to the next block in the chain, so it's OK to let the txg in
1090	 * which we allocated the next block sync.
1091	 */
1092	dmu_tx_commit(tx);
1093}
1094
1095/*
1096 * This is called when an lwb write completes. This means, this specific
1097 * lwb was written to disk, and all dependent lwb have also been
1098 * written to disk.
1099 *
1100 * At this point, a DKIOCFLUSHWRITECACHE command hasn't been issued to
1101 * the VDEVs involved in writing out this specific lwb. The lwb will be
1102 * "done" once zil_lwb_flush_vdevs_done() is called, which occurs in the
1103 * zio completion callback for the lwb's root zio.
1104 */
1105static void
1106zil_lwb_write_done(zio_t *zio)
1107{
1108	lwb_t *lwb = zio->io_private;
1109	spa_t *spa = zio->io_spa;
1110	zilog_t *zilog = lwb->lwb_zilog;
1111	avl_tree_t *t = &lwb->lwb_vdev_tree;
1112	void *cookie = NULL;
1113	zil_vdev_node_t *zv;
1114
1115	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1116
1117	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1118	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1119	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1120	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1121	ASSERT(!BP_IS_GANG(zio->io_bp));
1122	ASSERT(!BP_IS_HOLE(zio->io_bp));
1123	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1124
1125	abd_put(zio->io_abd);
1126
1127	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1128
1129	mutex_enter(&zilog->zl_lock);
1130	lwb->lwb_write_zio = NULL;
1131	mutex_exit(&zilog->zl_lock);
1132
1133	if (avl_numnodes(t) == 0)
1134		return;
1135
1136	/*
1137	 * If there was an IO error, we're not going to call zio_flush()
1138	 * on these vdevs, so we simply empty the tree and free the
1139	 * nodes. We avoid calling zio_flush() since there isn't any
1140	 * good reason for doing so, after the lwb block failed to be
1141	 * written out.
1142	 */
1143	if (zio->io_error != 0) {
1144		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1145			kmem_free(zv, sizeof (*zv));
1146		return;
1147	}
1148
1149	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1150		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1151		if (vd != NULL)
1152			zio_flush(lwb->lwb_root_zio, vd);
1153		kmem_free(zv, sizeof (*zv));
1154	}
1155}
1156
1157/*
1158 * This function's purpose is to "open" an lwb such that it is ready to
1159 * accept new itxs being committed to it. To do this, the lwb's zio
1160 * structures are created, and linked to the lwb. This function is
1161 * idempotent; if the passed in lwb has already been opened, this
1162 * function is essentially a no-op.
1163 */
1164static void
1165zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1166{
1167	zbookmark_phys_t zb;
1168	zio_priority_t prio;
1169
1170	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1171	ASSERT3P(lwb, !=, NULL);
1172	EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1173	EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1174
1175	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1176	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1177	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1178
1179	if (lwb->lwb_root_zio == NULL) {
1180		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1181		    BP_GET_LSIZE(&lwb->lwb_blk));
1182
1183		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1184			prio = ZIO_PRIORITY_SYNC_WRITE;
1185		else
1186			prio = ZIO_PRIORITY_ASYNC_WRITE;
1187
1188		lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1189		    zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1190		ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1191
1192		lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1193		    zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1194		    BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1195		    prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
1196		ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1197
1198		lwb->lwb_state = LWB_STATE_OPENED;
1199
1200		mutex_enter(&zilog->zl_lock);
1201
1202		/*
1203		 * The zilog's "zl_last_lwb_opened" field is used to
1204		 * build the lwb/zio dependency chain, which is used to
1205		 * preserve the ordering of lwb completions that is
1206		 * required by the semantics of the ZIL. Each new lwb
1207		 * zio becomes a parent of the "previous" lwb zio, such
1208		 * that the new lwb's zio cannot complete until the
1209		 * "previous" lwb's zio completes.
1210		 *
1211		 * This is required by the semantics of zil_commit();
1212		 * the commit waiters attached to the lwbs will be woken
1213		 * in the lwb zio's completion callback, so this zio
1214		 * dependency graph ensures the waiters are woken in the
1215		 * correct order (the same order the lwbs were created).
1216		 */
1217		lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1218		if (last_lwb_opened != NULL &&
1219		    last_lwb_opened->lwb_state != LWB_STATE_DONE) {
1220			ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1221			    last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1222			ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1223			zio_add_child(lwb->lwb_root_zio,
1224			    last_lwb_opened->lwb_root_zio);
1225		}
1226		zilog->zl_last_lwb_opened = lwb;
1227
1228		mutex_exit(&zilog->zl_lock);
1229	}
1230
1231	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1232	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1233	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1234}
1235
1236/*
1237 * Define a limited set of intent log block sizes.
1238 *
1239 * These must be a multiple of 4KB. Note only the amount used (again
1240 * aligned to 4KB) actually gets written. However, we can't always just
1241 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1242 */
1243uint64_t zil_block_buckets[] = {
1244    4096,		/* non TX_WRITE */
1245    8192+4096,		/* data base */
1246    32*1024 + 4096, 	/* NFS writes */
1247    UINT64_MAX
1248};
1249
1250/*
1251 * Start a log block write and advance to the next log block.
1252 * Calls are serialized.
1253 */
1254static lwb_t *
1255zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1256{
1257	lwb_t *nlwb = NULL;
1258	zil_chain_t *zilc;
1259	spa_t *spa = zilog->zl_spa;
1260	blkptr_t *bp;
1261	dmu_tx_t *tx;
1262	uint64_t txg;
1263	uint64_t zil_blksz, wsz;
1264	int i, error;
1265	boolean_t slog;
1266
1267	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1268	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1269	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1270	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1271
1272	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1273		zilc = (zil_chain_t *)lwb->lwb_buf;
1274		bp = &zilc->zc_next_blk;
1275	} else {
1276		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1277		bp = &zilc->zc_next_blk;
1278	}
1279
1280	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1281
1282	/*
1283	 * Allocate the next block and save its address in this block
1284	 * before writing it in order to establish the log chain.
1285	 * Note that if the allocation of nlwb synced before we wrote
1286	 * the block that points at it (lwb), we'd leak it if we crashed.
1287	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1288	 * We dirty the dataset to ensure that zil_sync() will be called
1289	 * to clean up in the event of allocation failure or I/O failure.
1290	 */
1291
1292	tx = dmu_tx_create(zilog->zl_os);
1293
1294	/*
1295	 * Since we are not going to create any new dirty data, and we
1296	 * can even help with clearing the existing dirty data, we
1297	 * should not be subject to the dirty data based delays. We
1298	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1299	 */
1300	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1301
1302	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1303	txg = dmu_tx_get_txg(tx);
1304
1305	lwb->lwb_tx = tx;
1306
1307	/*
1308	 * Log blocks are pre-allocated. Here we select the size of the next
1309	 * block, based on size used in the last block.
1310	 * - first find the smallest bucket that will fit the block from a
1311	 *   limited set of block sizes. This is because it's faster to write
1312	 *   blocks allocated from the same metaslab as they are adjacent or
1313	 *   close.
1314	 * - next find the maximum from the new suggested size and an array of
1315	 *   previous sizes. This lessens a picket fence effect of wrongly
1316	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1317	 *   requests.
1318	 *
1319	 * Note we only write what is used, but we can't just allocate
1320	 * the maximum block size because we can exhaust the available
1321	 * pool log space.
1322	 */
1323	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1324	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1325		continue;
1326	zil_blksz = zil_block_buckets[i];
1327	if (zil_blksz == UINT64_MAX)
1328		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1329	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1330	for (i = 0; i < ZIL_PREV_BLKS; i++)
1331		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1332	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1333
1334	BP_ZERO(bp);
1335
1336	/* pass the old blkptr in order to spread log blocks across devs */
1337	error = zio_alloc_zil(spa, zilog->zl_os->os_dsl_dataset->ds_object,
1338	    txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1339	if (error == 0) {
1340		ASSERT3U(bp->blk_birth, ==, txg);
1341		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1342		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1343
1344		/*
1345		 * Allocate a new log write block (lwb).
1346		 */
1347		nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1348	}
1349
1350	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1351		/* For Slim ZIL only write what is used. */
1352		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1353		ASSERT3U(wsz, <=, lwb->lwb_sz);
1354		zio_shrink(lwb->lwb_write_zio, wsz);
1355
1356	} else {
1357		wsz = lwb->lwb_sz;
1358	}
1359
1360	zilc->zc_pad = 0;
1361	zilc->zc_nused = lwb->lwb_nused;
1362	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1363
1364	/*
1365	 * clear unused data for security
1366	 */
1367	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1368
1369	spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1370
1371	zil_lwb_add_block(lwb, &lwb->lwb_blk);
1372	lwb->lwb_issued_timestamp = gethrtime();
1373	lwb->lwb_state = LWB_STATE_ISSUED;
1374
1375	zio_nowait(lwb->lwb_root_zio);
1376	zio_nowait(lwb->lwb_write_zio);
1377
1378	/*
1379	 * If there was an allocation failure then nlwb will be null which
1380	 * forces a txg_wait_synced().
1381	 */
1382	return (nlwb);
1383}
1384
1385static lwb_t *
1386zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1387{
1388	lr_t *lrcb, *lrc;
1389	lr_write_t *lrwb, *lrw;
1390	char *lr_buf;
1391	uint64_t dlen, dnow, lwb_sp, reclen, txg;
1392
1393	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1394	ASSERT3P(lwb, !=, NULL);
1395	ASSERT3P(lwb->lwb_buf, !=, NULL);
1396
1397	zil_lwb_write_open(zilog, lwb);
1398
1399	lrc = &itx->itx_lr;
1400	lrw = (lr_write_t *)lrc;
1401
1402	/*
1403	 * A commit itx doesn't represent any on-disk state; instead
1404	 * it's simply used as a place holder on the commit list, and
1405	 * provides a mechanism for attaching a "commit waiter" onto the
1406	 * correct lwb (such that the waiter can be signalled upon
1407	 * completion of that lwb). Thus, we don't process this itx's
1408	 * log record if it's a commit itx (these itx's don't have log
1409	 * records), and instead link the itx's waiter onto the lwb's
1410	 * list of waiters.
1411	 *
1412	 * For more details, see the comment above zil_commit().
1413	 */
1414	if (lrc->lrc_txtype == TX_COMMIT) {
1415		mutex_enter(&zilog->zl_lock);
1416		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1417		itx->itx_private = NULL;
1418		mutex_exit(&zilog->zl_lock);
1419		return (lwb);
1420	}
1421
1422	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1423		dlen = P2ROUNDUP_TYPED(
1424		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1425	} else {
1426		dlen = 0;
1427	}
1428	reclen = lrc->lrc_reclen;
1429	zilog->zl_cur_used += (reclen + dlen);
1430	txg = lrc->lrc_txg;
1431
1432	ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1433
1434cont:
1435	/*
1436	 * If this record won't fit in the current log block, start a new one.
1437	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1438	 */
1439	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1440	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1441	    lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1442	    lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1443		lwb = zil_lwb_write_issue(zilog, lwb);
1444		if (lwb == NULL)
1445			return (NULL);
1446		zil_lwb_write_open(zilog, lwb);
1447		ASSERT(LWB_EMPTY(lwb));
1448		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1449		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1450	}
1451
1452	dnow = MIN(dlen, lwb_sp - reclen);
1453	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1454	bcopy(lrc, lr_buf, reclen);
1455	lrcb = (lr_t *)lr_buf;		/* Like lrc, but inside lwb. */
1456	lrwb = (lr_write_t *)lrcb;	/* Like lrw, but inside lwb. */
1457
1458	/*
1459	 * If it's a write, fetch the data or get its blkptr as appropriate.
1460	 */
1461	if (lrc->lrc_txtype == TX_WRITE) {
1462		if (txg > spa_freeze_txg(zilog->zl_spa))
1463			txg_wait_synced(zilog->zl_dmu_pool, txg);
1464		if (itx->itx_wr_state != WR_COPIED) {
1465			char *dbuf;
1466			int error;
1467
1468			if (itx->itx_wr_state == WR_NEED_COPY) {
1469				dbuf = lr_buf + reclen;
1470				lrcb->lrc_reclen += dnow;
1471				if (lrwb->lr_length > dnow)
1472					lrwb->lr_length = dnow;
1473				lrw->lr_offset += dnow;
1474				lrw->lr_length -= dnow;
1475			} else {
1476				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1477				dbuf = NULL;
1478			}
1479
1480			/*
1481			 * We pass in the "lwb_write_zio" rather than
1482			 * "lwb_root_zio" so that the "lwb_write_zio"
1483			 * becomes the parent of any zio's created by
1484			 * the "zl_get_data" callback. The vdevs are
1485			 * flushed after the "lwb_write_zio" completes,
1486			 * so we want to make sure that completion
1487			 * callback waits for these additional zio's,
1488			 * such that the vdevs used by those zio's will
1489			 * be included in the lwb's vdev tree, and those
1490			 * vdevs will be properly flushed. If we passed
1491			 * in "lwb_root_zio" here, then these additional
1492			 * vdevs may not be flushed; e.g. if these zio's
1493			 * completed after "lwb_write_zio" completed.
1494			 */
1495			error = zilog->zl_get_data(itx->itx_private,
1496			    lrwb, dbuf, lwb, lwb->lwb_write_zio);
1497
1498			if (error == EIO) {
1499				txg_wait_synced(zilog->zl_dmu_pool, txg);
1500				return (lwb);
1501			}
1502			if (error != 0) {
1503				ASSERT(error == ENOENT || error == EEXIST ||
1504				    error == EALREADY);
1505				return (lwb);
1506			}
1507		}
1508	}
1509
1510	/*
1511	 * We're actually making an entry, so update lrc_seq to be the
1512	 * log record sequence number.  Note that this is generally not
1513	 * equal to the itx sequence number because not all transactions
1514	 * are synchronous, and sometimes spa_sync() gets there first.
1515	 */
1516	lrcb->lrc_seq = ++zilog->zl_lr_seq;
1517	lwb->lwb_nused += reclen + dnow;
1518
1519	zil_lwb_add_txg(lwb, txg);
1520
1521	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1522	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1523
1524	dlen -= dnow;
1525	if (dlen > 0) {
1526		zilog->zl_cur_used += reclen;
1527		goto cont;
1528	}
1529
1530	return (lwb);
1531}
1532
1533itx_t *
1534zil_itx_create(uint64_t txtype, size_t lrsize)
1535{
1536	itx_t *itx;
1537
1538	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1539
1540	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1541	itx->itx_lr.lrc_txtype = txtype;
1542	itx->itx_lr.lrc_reclen = lrsize;
1543	itx->itx_lr.lrc_seq = 0;	/* defensive */
1544	itx->itx_sync = B_TRUE;		/* default is synchronous */
1545
1546	return (itx);
1547}
1548
1549void
1550zil_itx_destroy(itx_t *itx)
1551{
1552	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1553}
1554
1555/*
1556 * Free up the sync and async itxs. The itxs_t has already been detached
1557 * so no locks are needed.
1558 */
1559static void
1560zil_itxg_clean(itxs_t *itxs)
1561{
1562	itx_t *itx;
1563	list_t *list;
1564	avl_tree_t *t;
1565	void *cookie;
1566	itx_async_node_t *ian;
1567
1568	list = &itxs->i_sync_list;
1569	while ((itx = list_head(list)) != NULL) {
1570		/*
1571		 * In the general case, commit itxs will not be found
1572		 * here, as they'll be committed to an lwb via
1573		 * zil_lwb_commit(), and free'd in that function. Having
1574		 * said that, it is still possible for commit itxs to be
1575		 * found here, due to the following race:
1576		 *
1577		 *	- a thread calls zil_commit() which assigns the
1578		 *	  commit itx to a per-txg i_sync_list
1579		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
1580		 *	  while the waiter is still on the i_sync_list
1581		 *
1582		 * There's nothing to prevent syncing the txg while the
1583		 * waiter is on the i_sync_list. This normally doesn't
1584		 * happen because spa_sync() is slower than zil_commit(),
1585		 * but if zil_commit() calls txg_wait_synced() (e.g.
1586		 * because zil_create() or zil_commit_writer_stall() is
1587		 * called) we will hit this case.
1588		 */
1589		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1590			zil_commit_waiter_skip(itx->itx_private);
1591
1592		list_remove(list, itx);
1593		zil_itx_destroy(itx);
1594	}
1595
1596	cookie = NULL;
1597	t = &itxs->i_async_tree;
1598	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1599		list = &ian->ia_list;
1600		while ((itx = list_head(list)) != NULL) {
1601			list_remove(list, itx);
1602			/* commit itxs should never be on the async lists. */
1603			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1604			zil_itx_destroy(itx);
1605		}
1606		list_destroy(list);
1607		kmem_free(ian, sizeof (itx_async_node_t));
1608	}
1609	avl_destroy(t);
1610
1611	kmem_free(itxs, sizeof (itxs_t));
1612}
1613
1614static int
1615zil_aitx_compare(const void *x1, const void *x2)
1616{
1617	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1618	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1619
1620	if (o1 < o2)
1621		return (-1);
1622	if (o1 > o2)
1623		return (1);
1624
1625	return (0);
1626}
1627
1628/*
1629 * Remove all async itx with the given oid.
1630 */
1631static void
1632zil_remove_async(zilog_t *zilog, uint64_t oid)
1633{
1634	uint64_t otxg, txg;
1635	itx_async_node_t *ian;
1636	avl_tree_t *t;
1637	avl_index_t where;
1638	list_t clean_list;
1639	itx_t *itx;
1640
1641	ASSERT(oid != 0);
1642	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1643
1644	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1645		otxg = ZILTEST_TXG;
1646	else
1647		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1648
1649	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1650		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1651
1652		mutex_enter(&itxg->itxg_lock);
1653		if (itxg->itxg_txg != txg) {
1654			mutex_exit(&itxg->itxg_lock);
1655			continue;
1656		}
1657
1658		/*
1659		 * Locate the object node and append its list.
1660		 */
1661		t = &itxg->itxg_itxs->i_async_tree;
1662		ian = avl_find(t, &oid, &where);
1663		if (ian != NULL)
1664			list_move_tail(&clean_list, &ian->ia_list);
1665		mutex_exit(&itxg->itxg_lock);
1666	}
1667	while ((itx = list_head(&clean_list)) != NULL) {
1668		list_remove(&clean_list, itx);
1669		/* commit itxs should never be on the async lists. */
1670		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1671		zil_itx_destroy(itx);
1672	}
1673	list_destroy(&clean_list);
1674}
1675
1676void
1677zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1678{
1679	uint64_t txg;
1680	itxg_t *itxg;
1681	itxs_t *itxs, *clean = NULL;
1682
1683	/*
1684	 * Object ids can be re-instantiated in the next txg so
1685	 * remove any async transactions to avoid future leaks.
1686	 * This can happen if a fsync occurs on the re-instantiated
1687	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1688	 * the new file data and flushes a write record for the old object.
1689	 */
1690	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1691		zil_remove_async(zilog, itx->itx_oid);
1692
1693	/*
1694	 * Ensure the data of a renamed file is committed before the rename.
1695	 */
1696	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1697		zil_async_to_sync(zilog, itx->itx_oid);
1698
1699	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1700		txg = ZILTEST_TXG;
1701	else
1702		txg = dmu_tx_get_txg(tx);
1703
1704	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1705	mutex_enter(&itxg->itxg_lock);
1706	itxs = itxg->itxg_itxs;
1707	if (itxg->itxg_txg != txg) {
1708		if (itxs != NULL) {
1709			/*
1710			 * The zil_clean callback hasn't got around to cleaning
1711			 * this itxg. Save the itxs for release below.
1712			 * This should be rare.
1713			 */
1714			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1715			    "txg %llu", itxg->itxg_txg);
1716			clean = itxg->itxg_itxs;
1717		}
1718		itxg->itxg_txg = txg;
1719		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1720
1721		list_create(&itxs->i_sync_list, sizeof (itx_t),
1722		    offsetof(itx_t, itx_node));
1723		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1724		    sizeof (itx_async_node_t),
1725		    offsetof(itx_async_node_t, ia_node));
1726	}
1727	if (itx->itx_sync) {
1728		list_insert_tail(&itxs->i_sync_list, itx);
1729	} else {
1730		avl_tree_t *t = &itxs->i_async_tree;
1731		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1732		itx_async_node_t *ian;
1733		avl_index_t where;
1734
1735		ian = avl_find(t, &foid, &where);
1736		if (ian == NULL) {
1737			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1738			list_create(&ian->ia_list, sizeof (itx_t),
1739			    offsetof(itx_t, itx_node));
1740			ian->ia_foid = foid;
1741			avl_insert(t, ian, where);
1742		}
1743		list_insert_tail(&ian->ia_list, itx);
1744	}
1745
1746	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1747
1748	/*
1749	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1750	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1751	 * need to be careful to always dirty the ZIL using the "real"
1752	 * TXG (not itxg_txg) even when the SPA is frozen.
1753	 */
1754	zilog_dirty(zilog, dmu_tx_get_txg(tx));
1755	mutex_exit(&itxg->itxg_lock);
1756
1757	/* Release the old itxs now we've dropped the lock */
1758	if (clean != NULL)
1759		zil_itxg_clean(clean);
1760}
1761
1762/*
1763 * If there are any in-memory intent log transactions which have now been
1764 * synced then start up a taskq to free them. We should only do this after we
1765 * have written out the uberblocks (i.e. txg has been comitted) so that
1766 * don't inadvertently clean out in-memory log records that would be required
1767 * by zil_commit().
1768 */
1769void
1770zil_clean(zilog_t *zilog, uint64_t synced_txg)
1771{
1772	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1773	itxs_t *clean_me;
1774
1775	ASSERT3U(synced_txg, <, ZILTEST_TXG);
1776
1777	mutex_enter(&itxg->itxg_lock);
1778	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1779		mutex_exit(&itxg->itxg_lock);
1780		return;
1781	}
1782	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1783	ASSERT3U(itxg->itxg_txg, !=, 0);
1784	clean_me = itxg->itxg_itxs;
1785	itxg->itxg_itxs = NULL;
1786	itxg->itxg_txg = 0;
1787	mutex_exit(&itxg->itxg_lock);
1788	/*
1789	 * Preferably start a task queue to free up the old itxs but
1790	 * if taskq_dispatch can't allocate resources to do that then
1791	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1792	 * created a bad performance problem.
1793	 */
1794	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1795	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1796	if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1797	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1798		zil_itxg_clean(clean_me);
1799}
1800
1801/*
1802 * This function will traverse the queue of itxs that need to be
1803 * committed, and move them onto the ZIL's zl_itx_commit_list.
1804 */
1805static void
1806zil_get_commit_list(zilog_t *zilog)
1807{
1808	uint64_t otxg, txg;
1809	list_t *commit_list = &zilog->zl_itx_commit_list;
1810
1811	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1812
1813	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1814		otxg = ZILTEST_TXG;
1815	else
1816		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1817
1818	/*
1819	 * This is inherently racy, since there is nothing to prevent
1820	 * the last synced txg from changing. That's okay since we'll
1821	 * only commit things in the future.
1822	 */
1823	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1824		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1825
1826		mutex_enter(&itxg->itxg_lock);
1827		if (itxg->itxg_txg != txg) {
1828			mutex_exit(&itxg->itxg_lock);
1829			continue;
1830		}
1831
1832		/*
1833		 * If we're adding itx records to the zl_itx_commit_list,
1834		 * then the zil better be dirty in this "txg". We can assert
1835		 * that here since we're holding the itxg_lock which will
1836		 * prevent spa_sync from cleaning it. Once we add the itxs
1837		 * to the zl_itx_commit_list we must commit it to disk even
1838		 * if it's unnecessary (i.e. the txg was synced).
1839		 */
1840		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1841		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1842		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1843
1844		mutex_exit(&itxg->itxg_lock);
1845	}
1846}
1847
1848/*
1849 * Move the async itxs for a specified object to commit into sync lists.
1850 */
1851static void
1852zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1853{
1854	uint64_t otxg, txg;
1855	itx_async_node_t *ian;
1856	avl_tree_t *t;
1857	avl_index_t where;
1858
1859	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1860		otxg = ZILTEST_TXG;
1861	else
1862		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1863
1864	/*
1865	 * This is inherently racy, since there is nothing to prevent
1866	 * the last synced txg from changing.
1867	 */
1868	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1869		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1870
1871		mutex_enter(&itxg->itxg_lock);
1872		if (itxg->itxg_txg != txg) {
1873			mutex_exit(&itxg->itxg_lock);
1874			continue;
1875		}
1876
1877		/*
1878		 * If a foid is specified then find that node and append its
1879		 * list. Otherwise walk the tree appending all the lists
1880		 * to the sync list. We add to the end rather than the
1881		 * beginning to ensure the create has happened.
1882		 */
1883		t = &itxg->itxg_itxs->i_async_tree;
1884		if (foid != 0) {
1885			ian = avl_find(t, &foid, &where);
1886			if (ian != NULL) {
1887				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1888				    &ian->ia_list);
1889			}
1890		} else {
1891			void *cookie = NULL;
1892
1893			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1894				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1895				    &ian->ia_list);
1896				list_destroy(&ian->ia_list);
1897				kmem_free(ian, sizeof (itx_async_node_t));
1898			}
1899		}
1900		mutex_exit(&itxg->itxg_lock);
1901	}
1902}
1903
1904/*
1905 * This function will prune commit itxs that are at the head of the
1906 * commit list (it won't prune past the first non-commit itx), and
1907 * either: a) attach them to the last lwb that's still pending
1908 * completion, or b) skip them altogether.
1909 *
1910 * This is used as a performance optimization to prevent commit itxs
1911 * from generating new lwbs when it's unnecessary to do so.
1912 */
1913static void
1914zil_prune_commit_list(zilog_t *zilog)
1915{
1916	itx_t *itx;
1917
1918	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1919
1920	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1921		lr_t *lrc = &itx->itx_lr;
1922		if (lrc->lrc_txtype != TX_COMMIT)
1923			break;
1924
1925		mutex_enter(&zilog->zl_lock);
1926
1927		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
1928		if (last_lwb == NULL || last_lwb->lwb_state == LWB_STATE_DONE) {
1929			/*
1930			 * All of the itxs this waiter was waiting on
1931			 * must have already completed (or there were
1932			 * never any itx's for it to wait on), so it's
1933			 * safe to skip this waiter and mark it done.
1934			 */
1935			zil_commit_waiter_skip(itx->itx_private);
1936		} else {
1937			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
1938			itx->itx_private = NULL;
1939		}
1940
1941		mutex_exit(&zilog->zl_lock);
1942
1943		list_remove(&zilog->zl_itx_commit_list, itx);
1944		zil_itx_destroy(itx);
1945	}
1946
1947	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
1948}
1949
1950static void
1951zil_commit_writer_stall(zilog_t *zilog)
1952{
1953	/*
1954	 * When zio_alloc_zil() fails to allocate the next lwb block on
1955	 * disk, we must call txg_wait_synced() to ensure all of the
1956	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
1957	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
1958	 * to zil_process_commit_list()) will have to call zil_create(),
1959	 * and start a new ZIL chain.
1960	 *
1961	 * Since zil_alloc_zil() failed, the lwb that was previously
1962	 * issued does not have a pointer to the "next" lwb on disk.
1963	 * Thus, if another ZIL writer thread was to allocate the "next"
1964	 * on-disk lwb, that block could be leaked in the event of a
1965	 * crash (because the previous lwb on-disk would not point to
1966	 * it).
1967	 *
1968	 * We must hold the zilog's zl_issuer_lock while we do this, to
1969	 * ensure no new threads enter zil_process_commit_list() until
1970	 * all lwb's in the zl_lwb_list have been synced and freed
1971	 * (which is achieved via the txg_wait_synced() call).
1972	 */
1973	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1974	txg_wait_synced(zilog->zl_dmu_pool, 0);
1975	ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
1976}
1977
1978/*
1979 * This function will traverse the commit list, creating new lwbs as
1980 * needed, and committing the itxs from the commit list to these newly
1981 * created lwbs. Additionally, as a new lwb is created, the previous
1982 * lwb will be issued to the zio layer to be written to disk.
1983 */
1984static void
1985zil_process_commit_list(zilog_t *zilog)
1986{
1987	spa_t *spa = zilog->zl_spa;
1988	list_t nolwb_waiters;
1989	lwb_t *lwb;
1990	itx_t *itx;
1991
1992	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1993
1994	/*
1995	 * Return if there's nothing to commit before we dirty the fs by
1996	 * calling zil_create().
1997	 */
1998	if (list_head(&zilog->zl_itx_commit_list) == NULL)
1999		return;
2000
2001	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2002	    offsetof(zil_commit_waiter_t, zcw_node));
2003
2004	lwb = list_tail(&zilog->zl_lwb_list);
2005	if (lwb == NULL) {
2006		lwb = zil_create(zilog);
2007	} else {
2008		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2009		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2010	}
2011
2012	while (itx = list_head(&zilog->zl_itx_commit_list)) {
2013		lr_t *lrc = &itx->itx_lr;
2014		uint64_t txg = lrc->lrc_txg;
2015
2016		ASSERT3U(txg, !=, 0);
2017
2018		if (lrc->lrc_txtype == TX_COMMIT) {
2019			DTRACE_PROBE2(zil__process__commit__itx,
2020			    zilog_t *, zilog, itx_t *, itx);
2021		} else {
2022			DTRACE_PROBE2(zil__process__normal__itx,
2023			    zilog_t *, zilog, itx_t *, itx);
2024		}
2025
2026		boolean_t synced = txg <= spa_last_synced_txg(spa);
2027		boolean_t frozen = txg > spa_freeze_txg(spa);
2028
2029		/*
2030		 * If the txg of this itx has already been synced out, then
2031		 * we don't need to commit this itx to an lwb. This is
2032		 * because the data of this itx will have already been
2033		 * written to the main pool. This is inherently racy, and
2034		 * it's still ok to commit an itx whose txg has already
2035		 * been synced; this will result in a write that's
2036		 * unnecessary, but will do no harm.
2037		 *
2038		 * With that said, we always want to commit TX_COMMIT itxs
2039		 * to an lwb, regardless of whether or not that itx's txg
2040		 * has been synced out. We do this to ensure any OPENED lwb
2041		 * will always have at least one zil_commit_waiter_t linked
2042		 * to the lwb.
2043		 *
2044		 * As a counter-example, if we skipped TX_COMMIT itx's
2045		 * whose txg had already been synced, the following
2046		 * situation could occur if we happened to be racing with
2047		 * spa_sync:
2048		 *
2049		 * 1. we commit a non-TX_COMMIT itx to an lwb, where the
2050		 *    itx's txg is 10 and the last synced txg is 9.
2051		 * 2. spa_sync finishes syncing out txg 10.
2052		 * 3. we move to the next itx in the list, it's a TX_COMMIT
2053		 *    whose txg is 10, so we skip it rather than committing
2054		 *    it to the lwb used in (1).
2055		 *
2056		 * If the itx that is skipped in (3) is the last TX_COMMIT
2057		 * itx in the commit list, than it's possible for the lwb
2058		 * used in (1) to remain in the OPENED state indefinitely.
2059		 *
2060		 * To prevent the above scenario from occuring, ensuring
2061		 * that once an lwb is OPENED it will transition to ISSUED
2062		 * and eventually DONE, we always commit TX_COMMIT itx's to
2063		 * an lwb here, even if that itx's txg has already been
2064		 * synced.
2065		 *
2066		 * Finally, if the pool is frozen, we _always_ commit the
2067		 * itx.  The point of freezing the pool is to prevent data
2068		 * from being written to the main pool via spa_sync, and
2069		 * instead rely solely on the ZIL to persistently store the
2070		 * data; i.e.  when the pool is frozen, the last synced txg
2071		 * value can't be trusted.
2072		 */
2073		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2074			if (lwb != NULL) {
2075				lwb = zil_lwb_commit(zilog, itx, lwb);
2076			} else if (lrc->lrc_txtype == TX_COMMIT) {
2077				ASSERT3P(lwb, ==, NULL);
2078				zil_commit_waiter_link_nolwb(
2079				    itx->itx_private, &nolwb_waiters);
2080			}
2081		}
2082
2083		list_remove(&zilog->zl_itx_commit_list, itx);
2084		zil_itx_destroy(itx);
2085	}
2086
2087	if (lwb == NULL) {
2088		/*
2089		 * This indicates zio_alloc_zil() failed to allocate the
2090		 * "next" lwb on-disk. When this happens, we must stall
2091		 * the ZIL write pipeline; see the comment within
2092		 * zil_commit_writer_stall() for more details.
2093		 */
2094		zil_commit_writer_stall(zilog);
2095
2096		/*
2097		 * Additionally, we have to signal and mark the "nolwb"
2098		 * waiters as "done" here, since without an lwb, we
2099		 * can't do this via zil_lwb_flush_vdevs_done() like
2100		 * normal.
2101		 */
2102		zil_commit_waiter_t *zcw;
2103		while (zcw = list_head(&nolwb_waiters)) {
2104			zil_commit_waiter_skip(zcw);
2105			list_remove(&nolwb_waiters, zcw);
2106		}
2107	} else {
2108		ASSERT(list_is_empty(&nolwb_waiters));
2109		ASSERT3P(lwb, !=, NULL);
2110		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2111		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2112
2113		/*
2114		 * At this point, the ZIL block pointed at by the "lwb"
2115		 * variable is in one of the following states: "closed"
2116		 * or "open".
2117		 *
2118		 * If its "closed", then no itxs have been committed to
2119		 * it, so there's no point in issuing its zio (i.e.
2120		 * it's "empty").
2121		 *
2122		 * If its "open" state, then it contains one or more
2123		 * itxs that eventually need to be committed to stable
2124		 * storage. In this case we intentionally do not issue
2125		 * the lwb's zio to disk yet, and instead rely on one of
2126		 * the following two mechanisms for issuing the zio:
2127		 *
2128		 * 1. Ideally, there will be more ZIL activity occuring
2129		 * on the system, such that this function will be
2130		 * immediately called again (not necessarily by the same
2131		 * thread) and this lwb's zio will be issued via
2132		 * zil_lwb_commit(). This way, the lwb is guaranteed to
2133		 * be "full" when it is issued to disk, and we'll make
2134		 * use of the lwb's size the best we can.
2135		 *
2136		 * 2. If there isn't sufficient ZIL activity occuring on
2137		 * the system, such that this lwb's zio isn't issued via
2138		 * zil_lwb_commit(), zil_commit_waiter() will issue the
2139		 * lwb's zio. If this occurs, the lwb is not guaranteed
2140		 * to be "full" by the time its zio is issued, and means
2141		 * the size of the lwb was "too large" given the amount
2142		 * of ZIL activity occuring on the system at that time.
2143		 *
2144		 * We do this for a couple of reasons:
2145		 *
2146		 * 1. To try and reduce the number of IOPs needed to
2147		 * write the same number of itxs. If an lwb has space
2148		 * available in it's buffer for more itxs, and more itxs
2149		 * will be committed relatively soon (relative to the
2150		 * latency of performing a write), then it's beneficial
2151		 * to wait for these "next" itxs. This way, more itxs
2152		 * can be committed to stable storage with fewer writes.
2153		 *
2154		 * 2. To try and use the largest lwb block size that the
2155		 * incoming rate of itxs can support. Again, this is to
2156		 * try and pack as many itxs into as few lwbs as
2157		 * possible, without significantly impacting the latency
2158		 * of each individual itx.
2159		 */
2160	}
2161}
2162
2163/*
2164 * This function is responsible for ensuring the passed in commit waiter
2165 * (and associated commit itx) is committed to an lwb. If the waiter is
2166 * not already committed to an lwb, all itxs in the zilog's queue of
2167 * itxs will be processed. The assumption is the passed in waiter's
2168 * commit itx will found in the queue just like the other non-commit
2169 * itxs, such that when the entire queue is processed, the waiter will
2170 * have been commited to an lwb.
2171 *
2172 * The lwb associated with the passed in waiter is not guaranteed to
2173 * have been issued by the time this function completes. If the lwb is
2174 * not issued, we rely on future calls to zil_commit_writer() to issue
2175 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2176 */
2177static void
2178zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2179{
2180	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2181	ASSERT(spa_writeable(zilog->zl_spa));
2182
2183	mutex_enter(&zilog->zl_issuer_lock);
2184
2185	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2186		/*
2187		 * It's possible that, while we were waiting to acquire
2188		 * the "zl_issuer_lock", another thread committed this
2189		 * waiter to an lwb. If that occurs, we bail out early,
2190		 * without processing any of the zilog's queue of itxs.
2191		 *
2192		 * On certain workloads and system configurations, the
2193		 * "zl_issuer_lock" can become highly contended. In an
2194		 * attempt to reduce this contention, we immediately drop
2195		 * the lock if the waiter has already been processed.
2196		 *
2197		 * We've measured this optimization to reduce CPU spent
2198		 * contending on this lock by up to 5%, using a system
2199		 * with 32 CPUs, low latency storage (~50 usec writes),
2200		 * and 1024 threads performing sync writes.
2201		 */
2202		goto out;
2203	}
2204
2205	zil_get_commit_list(zilog);
2206	zil_prune_commit_list(zilog);
2207	zil_process_commit_list(zilog);
2208
2209out:
2210	mutex_exit(&zilog->zl_issuer_lock);
2211}
2212
2213static void
2214zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2215{
2216	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2217	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2218	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2219
2220	lwb_t *lwb = zcw->zcw_lwb;
2221	ASSERT3P(lwb, !=, NULL);
2222	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2223
2224	/*
2225	 * If the lwb has already been issued by another thread, we can
2226	 * immediately return since there's no work to be done (the
2227	 * point of this function is to issue the lwb). Additionally, we
2228	 * do this prior to acquiring the zl_issuer_lock, to avoid
2229	 * acquiring it when it's not necessary to do so.
2230	 */
2231	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2232	    lwb->lwb_state == LWB_STATE_DONE)
2233		return;
2234
2235	/*
2236	 * In order to call zil_lwb_write_issue() we must hold the
2237	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2238	 * since we're already holding the commit waiter's "zcw_lock",
2239	 * and those two locks are aquired in the opposite order
2240	 * elsewhere.
2241	 */
2242	mutex_exit(&zcw->zcw_lock);
2243	mutex_enter(&zilog->zl_issuer_lock);
2244	mutex_enter(&zcw->zcw_lock);
2245
2246	/*
2247	 * Since we just dropped and re-acquired the commit waiter's
2248	 * lock, we have to re-check to see if the waiter was marked
2249	 * "done" during that process. If the waiter was marked "done",
2250	 * the "lwb" pointer is no longer valid (it can be free'd after
2251	 * the waiter is marked "done"), so without this check we could
2252	 * wind up with a use-after-free error below.
2253	 */
2254	if (zcw->zcw_done)
2255		goto out;
2256
2257	ASSERT3P(lwb, ==, zcw->zcw_lwb);
2258
2259	/*
2260	 * We've already checked this above, but since we hadn't acquired
2261	 * the zilog's zl_issuer_lock, we have to perform this check a
2262	 * second time while holding the lock.
2263	 *
2264	 * We don't need to hold the zl_lock since the lwb cannot transition
2265	 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2266	 * _can_ transition from ISSUED to DONE, but it's OK to race with
2267	 * that transition since we treat the lwb the same, whether it's in
2268	 * the ISSUED or DONE states.
2269	 *
2270	 * The important thing, is we treat the lwb differently depending on
2271	 * if it's ISSUED or OPENED, and block any other threads that might
2272	 * attempt to issue this lwb. For that reason we hold the
2273	 * zl_issuer_lock when checking the lwb_state; we must not call
2274	 * zil_lwb_write_issue() if the lwb had already been issued.
2275	 *
2276	 * See the comment above the lwb_state_t structure definition for
2277	 * more details on the lwb states, and locking requirements.
2278	 */
2279	if (lwb->lwb_state == LWB_STATE_ISSUED ||
2280	    lwb->lwb_state == LWB_STATE_DONE)
2281		goto out;
2282
2283	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2284
2285	/*
2286	 * As described in the comments above zil_commit_waiter() and
2287	 * zil_process_commit_list(), we need to issue this lwb's zio
2288	 * since we've reached the commit waiter's timeout and it still
2289	 * hasn't been issued.
2290	 */
2291	lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2292
2293	IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2294
2295	/*
2296	 * Since the lwb's zio hadn't been issued by the time this thread
2297	 * reached its timeout, we reset the zilog's "zl_cur_used" field
2298	 * to influence the zil block size selection algorithm.
2299	 *
2300	 * By having to issue the lwb's zio here, it means the size of the
2301	 * lwb was too large, given the incoming throughput of itxs.  By
2302	 * setting "zl_cur_used" to zero, we communicate this fact to the
2303	 * block size selection algorithm, so it can take this informaiton
2304	 * into account, and potentially select a smaller size for the
2305	 * next lwb block that is allocated.
2306	 */
2307	zilog->zl_cur_used = 0;
2308
2309	if (nlwb == NULL) {
2310		/*
2311		 * When zil_lwb_write_issue() returns NULL, this
2312		 * indicates zio_alloc_zil() failed to allocate the
2313		 * "next" lwb on-disk. When this occurs, the ZIL write
2314		 * pipeline must be stalled; see the comment within the
2315		 * zil_commit_writer_stall() function for more details.
2316		 *
2317		 * We must drop the commit waiter's lock prior to
2318		 * calling zil_commit_writer_stall() or else we can wind
2319		 * up with the following deadlock:
2320		 *
2321		 * - This thread is waiting for the txg to sync while
2322		 *   holding the waiter's lock; txg_wait_synced() is
2323		 *   used within txg_commit_writer_stall().
2324		 *
2325		 * - The txg can't sync because it is waiting for this
2326		 *   lwb's zio callback to call dmu_tx_commit().
2327		 *
2328		 * - The lwb's zio callback can't call dmu_tx_commit()
2329		 *   because it's blocked trying to acquire the waiter's
2330		 *   lock, which occurs prior to calling dmu_tx_commit()
2331		 */
2332		mutex_exit(&zcw->zcw_lock);
2333		zil_commit_writer_stall(zilog);
2334		mutex_enter(&zcw->zcw_lock);
2335	}
2336
2337out:
2338	mutex_exit(&zilog->zl_issuer_lock);
2339	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2340}
2341
2342/*
2343 * This function is responsible for performing the following two tasks:
2344 *
2345 * 1. its primary responsibility is to block until the given "commit
2346 *    waiter" is considered "done".
2347 *
2348 * 2. its secondary responsibility is to issue the zio for the lwb that
2349 *    the given "commit waiter" is waiting on, if this function has
2350 *    waited "long enough" and the lwb is still in the "open" state.
2351 *
2352 * Given a sufficient amount of itxs being generated and written using
2353 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2354 * function. If this does not occur, this secondary responsibility will
2355 * ensure the lwb is issued even if there is not other synchronous
2356 * activity on the system.
2357 *
2358 * For more details, see zil_process_commit_list(); more specifically,
2359 * the comment at the bottom of that function.
2360 */
2361static void
2362zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2363{
2364	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2365	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2366	ASSERT(spa_writeable(zilog->zl_spa));
2367
2368	mutex_enter(&zcw->zcw_lock);
2369
2370	/*
2371	 * The timeout is scaled based on the lwb latency to avoid
2372	 * significantly impacting the latency of each individual itx.
2373	 * For more details, see the comment at the bottom of the
2374	 * zil_process_commit_list() function.
2375	 */
2376	int pct = MAX(zfs_commit_timeout_pct, 1);
2377	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2378	hrtime_t wakeup = gethrtime() + sleep;
2379	boolean_t timedout = B_FALSE;
2380
2381	while (!zcw->zcw_done) {
2382		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2383
2384		lwb_t *lwb = zcw->zcw_lwb;
2385
2386		/*
2387		 * Usually, the waiter will have a non-NULL lwb field here,
2388		 * but it's possible for it to be NULL as a result of
2389		 * zil_commit() racing with spa_sync().
2390		 *
2391		 * When zil_clean() is called, it's possible for the itxg
2392		 * list (which may be cleaned via a taskq) to contain
2393		 * commit itxs. When this occurs, the commit waiters linked
2394		 * off of these commit itxs will not be committed to an
2395		 * lwb.  Additionally, these commit waiters will not be
2396		 * marked done until zil_commit_waiter_skip() is called via
2397		 * zil_itxg_clean().
2398		 *
2399		 * Thus, it's possible for this commit waiter (i.e. the
2400		 * "zcw" variable) to be found in this "in between" state;
2401		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2402		 * been skipped, so it's "zcw_done" field is still B_FALSE.
2403		 */
2404		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2405
2406		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2407			ASSERT3B(timedout, ==, B_FALSE);
2408
2409			/*
2410			 * If the lwb hasn't been issued yet, then we
2411			 * need to wait with a timeout, in case this
2412			 * function needs to issue the lwb after the
2413			 * timeout is reached; responsibility (2) from
2414			 * the comment above this function.
2415			 */
2416			clock_t timeleft = cv_timedwait_hires(&zcw->zcw_cv,
2417			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2418			    CALLOUT_FLAG_ABSOLUTE);
2419
2420			if (timeleft >= 0 || zcw->zcw_done)
2421				continue;
2422
2423			timedout = B_TRUE;
2424			zil_commit_waiter_timeout(zilog, zcw);
2425
2426			if (!zcw->zcw_done) {
2427				/*
2428				 * If the commit waiter has already been
2429				 * marked "done", it's possible for the
2430				 * waiter's lwb structure to have already
2431				 * been freed.  Thus, we can only reliably
2432				 * make these assertions if the waiter
2433				 * isn't done.
2434				 */
2435				ASSERT3P(lwb, ==, zcw->zcw_lwb);
2436				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2437			}
2438		} else {
2439			/*
2440			 * If the lwb isn't open, then it must have already
2441			 * been issued. In that case, there's no need to
2442			 * use a timeout when waiting for the lwb to
2443			 * complete.
2444			 *
2445			 * Additionally, if the lwb is NULL, the waiter
2446			 * will soon be signalled and marked done via
2447			 * zil_clean() and zil_itxg_clean(), so no timeout
2448			 * is required.
2449			 */
2450
2451			IMPLY(lwb != NULL,
2452			    lwb->lwb_state == LWB_STATE_ISSUED ||
2453			    lwb->lwb_state == LWB_STATE_DONE);
2454			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2455		}
2456	}
2457
2458	mutex_exit(&zcw->zcw_lock);
2459}
2460
2461static zil_commit_waiter_t *
2462zil_alloc_commit_waiter()
2463{
2464	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2465
2466	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2467	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2468	list_link_init(&zcw->zcw_node);
2469	zcw->zcw_lwb = NULL;
2470	zcw->zcw_done = B_FALSE;
2471	zcw->zcw_zio_error = 0;
2472
2473	return (zcw);
2474}
2475
2476static void
2477zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2478{
2479	ASSERT(!list_link_active(&zcw->zcw_node));
2480	ASSERT3P(zcw->zcw_lwb, ==, NULL);
2481	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2482	mutex_destroy(&zcw->zcw_lock);
2483	cv_destroy(&zcw->zcw_cv);
2484	kmem_cache_free(zil_zcw_cache, zcw);
2485}
2486
2487/*
2488 * This function is used to create a TX_COMMIT itx and assign it. This
2489 * way, it will be linked into the ZIL's list of synchronous itxs, and
2490 * then later committed to an lwb (or skipped) when
2491 * zil_process_commit_list() is called.
2492 */
2493static void
2494zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2495{
2496	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2497	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2498
2499	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2500	itx->itx_sync = B_TRUE;
2501	itx->itx_private = zcw;
2502
2503	zil_itx_assign(zilog, itx, tx);
2504
2505	dmu_tx_commit(tx);
2506}
2507
2508/*
2509 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2510 *
2511 * When writing ZIL transactions to the on-disk representation of the
2512 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2513 * itxs can be committed to a single lwb. Once a lwb is written and
2514 * committed to stable storage (i.e. the lwb is written, and vdevs have
2515 * been flushed), each itx that was committed to that lwb is also
2516 * considered to be committed to stable storage.
2517 *
2518 * When an itx is committed to an lwb, the log record (lr_t) contained
2519 * by the itx is copied into the lwb's zio buffer, and once this buffer
2520 * is written to disk, it becomes an on-disk ZIL block.
2521 *
2522 * As itxs are generated, they're inserted into the ZIL's queue of
2523 * uncommitted itxs. The semantics of zil_commit() are such that it will
2524 * block until all itxs that were in the queue when it was called, are
2525 * committed to stable storage.
2526 *
2527 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2528 * itxs, for all objects in the dataset, will be committed to stable
2529 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2530 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2531 * that correspond to the foid passed in, will be committed to stable
2532 * storage prior to zil_commit() returning.
2533 *
2534 * Generally speaking, when zil_commit() is called, the consumer doesn't
2535 * actually care about _all_ of the uncommitted itxs. Instead, they're
2536 * simply trying to waiting for a specific itx to be committed to disk,
2537 * but the interface(s) for interacting with the ZIL don't allow such
2538 * fine-grained communication. A better interface would allow a consumer
2539 * to create and assign an itx, and then pass a reference to this itx to
2540 * zil_commit(); such that zil_commit() would return as soon as that
2541 * specific itx was committed to disk (instead of waiting for _all_
2542 * itxs to be committed).
2543 *
2544 * When a thread calls zil_commit() a special "commit itx" will be
2545 * generated, along with a corresponding "waiter" for this commit itx.
2546 * zil_commit() will wait on this waiter's CV, such that when the waiter
2547 * is marked done, and signalled, zil_commit() will return.
2548 *
2549 * This commit itx is inserted into the queue of uncommitted itxs. This
2550 * provides an easy mechanism for determining which itxs were in the
2551 * queue prior to zil_commit() having been called, and which itxs were
2552 * added after zil_commit() was called.
2553 *
2554 * The commit it is special; it doesn't have any on-disk representation.
2555 * When a commit itx is "committed" to an lwb, the waiter associated
2556 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2557 * completes, each waiter on the lwb's list is marked done and signalled
2558 * -- allowing the thread waiting on the waiter to return from zil_commit().
2559 *
2560 * It's important to point out a few critical factors that allow us
2561 * to make use of the commit itxs, commit waiters, per-lwb lists of
2562 * commit waiters, and zio completion callbacks like we're doing:
2563 *
2564 *   1. The list of waiters for each lwb is traversed, and each commit
2565 *      waiter is marked "done" and signalled, in the zio completion
2566 *      callback of the lwb's zio[*].
2567 *
2568 *      * Actually, the waiters are signalled in the zio completion
2569 *        callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2570 *        that are sent to the vdevs upon completion of the lwb zio.
2571 *
2572 *   2. When the itxs are inserted into the ZIL's queue of uncommitted
2573 *      itxs, the order in which they are inserted is preserved[*]; as
2574 *      itxs are added to the queue, they are added to the tail of
2575 *      in-memory linked lists.
2576 *
2577 *      When committing the itxs to lwbs (to be written to disk), they
2578 *      are committed in the same order in which the itxs were added to
2579 *      the uncommitted queue's linked list(s); i.e. the linked list of
2580 *      itxs to commit is traversed from head to tail, and each itx is
2581 *      committed to an lwb in that order.
2582 *
2583 *      * To clarify:
2584 *
2585 *        - the order of "sync" itxs is preserved w.r.t. other
2586 *          "sync" itxs, regardless of the corresponding objects.
2587 *        - the order of "async" itxs is preserved w.r.t. other
2588 *          "async" itxs corresponding to the same object.
2589 *        - the order of "async" itxs is *not* preserved w.r.t. other
2590 *          "async" itxs corresponding to different objects.
2591 *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
2592 *          versa) is *not* preserved, even for itxs that correspond
2593 *          to the same object.
2594 *
2595 *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
2596 *      zil_get_commit_list(), and zil_process_commit_list().
2597 *
2598 *   3. The lwbs represent a linked list of blocks on disk. Thus, any
2599 *      lwb cannot be considered committed to stable storage, until its
2600 *      "previous" lwb is also committed to stable storage. This fact,
2601 *      coupled with the fact described above, means that itxs are
2602 *      committed in (roughly) the order in which they were generated.
2603 *      This is essential because itxs are dependent on prior itxs.
2604 *      Thus, we *must not* deem an itx as being committed to stable
2605 *      storage, until *all* prior itxs have also been committed to
2606 *      stable storage.
2607 *
2608 *      To enforce this ordering of lwb zio's, while still leveraging as
2609 *      much of the underlying storage performance as possible, we rely
2610 *      on two fundamental concepts:
2611 *
2612 *          1. The creation and issuance of lwb zio's is protected by
2613 *             the zilog's "zl_issuer_lock", which ensures only a single
2614 *             thread is creating and/or issuing lwb's at a time
2615 *          2. The "previous" lwb is a child of the "current" lwb
2616 *             (leveraging the zio parent-child depenency graph)
2617 *
2618 *      By relying on this parent-child zio relationship, we can have
2619 *      many lwb zio's concurrently issued to the underlying storage,
2620 *      but the order in which they complete will be the same order in
2621 *      which they were created.
2622 */
2623void
2624zil_commit(zilog_t *zilog, uint64_t foid)
2625{
2626	/*
2627	 * We should never attempt to call zil_commit on a snapshot for
2628	 * a couple of reasons:
2629	 *
2630	 * 1. A snapshot may never be modified, thus it cannot have any
2631	 *    in-flight itxs that would have modified the dataset.
2632	 *
2633	 * 2. By design, when zil_commit() is called, a commit itx will
2634	 *    be assigned to this zilog; as a result, the zilog will be
2635	 *    dirtied. We must not dirty the zilog of a snapshot; there's
2636	 *    checks in the code that enforce this invariant, and will
2637	 *    cause a panic if it's not upheld.
2638	 */
2639	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2640
2641	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2642		return;
2643
2644	if (!spa_writeable(zilog->zl_spa)) {
2645		/*
2646		 * If the SPA is not writable, there should never be any
2647		 * pending itxs waiting to be committed to disk. If that
2648		 * weren't true, we'd skip writing those itxs out, and
2649		 * would break the sematics of zil_commit(); thus, we're
2650		 * verifying that truth before we return to the caller.
2651		 */
2652		ASSERT(list_is_empty(&zilog->zl_lwb_list));
2653		ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2654		for (int i = 0; i < TXG_SIZE; i++)
2655			ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2656		return;
2657	}
2658
2659	/*
2660	 * If the ZIL is suspended, we don't want to dirty it by calling
2661	 * zil_commit_itx_assign() below, nor can we write out
2662	 * lwbs like would be done in zil_commit_write(). Thus, we
2663	 * simply rely on txg_wait_synced() to maintain the necessary
2664	 * semantics, and avoid calling those functions altogether.
2665	 */
2666	if (zilog->zl_suspend > 0) {
2667		txg_wait_synced(zilog->zl_dmu_pool, 0);
2668		return;
2669	}
2670
2671	zil_commit_impl(zilog, foid);
2672}
2673
2674void
2675zil_commit_impl(zilog_t *zilog, uint64_t foid)
2676{
2677	/*
2678	 * Move the "async" itxs for the specified foid to the "sync"
2679	 * queues, such that they will be later committed (or skipped)
2680	 * to an lwb when zil_process_commit_list() is called.
2681	 *
2682	 * Since these "async" itxs must be committed prior to this
2683	 * call to zil_commit returning, we must perform this operation
2684	 * before we call zil_commit_itx_assign().
2685	 */
2686	zil_async_to_sync(zilog, foid);
2687
2688	/*
2689	 * We allocate a new "waiter" structure which will initially be
2690	 * linked to the commit itx using the itx's "itx_private" field.
2691	 * Since the commit itx doesn't represent any on-disk state,
2692	 * when it's committed to an lwb, rather than copying the its
2693	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2694	 * added to the lwb's list of waiters. Then, when the lwb is
2695	 * committed to stable storage, each waiter in the lwb's list of
2696	 * waiters will be marked "done", and signalled.
2697	 *
2698	 * We must create the waiter and assign the commit itx prior to
2699	 * calling zil_commit_writer(), or else our specific commit itx
2700	 * is not guaranteed to be committed to an lwb prior to calling
2701	 * zil_commit_waiter().
2702	 */
2703	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
2704	zil_commit_itx_assign(zilog, zcw);
2705
2706	zil_commit_writer(zilog, zcw);
2707	zil_commit_waiter(zilog, zcw);
2708
2709	if (zcw->zcw_zio_error != 0) {
2710		/*
2711		 * If there was an error writing out the ZIL blocks that
2712		 * this thread is waiting on, then we fallback to
2713		 * relying on spa_sync() to write out the data this
2714		 * thread is waiting on. Obviously this has performance
2715		 * implications, but the expectation is for this to be
2716		 * an exceptional case, and shouldn't occur often.
2717		 */
2718		DTRACE_PROBE2(zil__commit__io__error,
2719		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
2720		txg_wait_synced(zilog->zl_dmu_pool, 0);
2721	}
2722
2723	zil_free_commit_waiter(zcw);
2724}
2725
2726/*
2727 * Called in syncing context to free committed log blocks and update log header.
2728 */
2729void
2730zil_sync(zilog_t *zilog, dmu_tx_t *tx)
2731{
2732	zil_header_t *zh = zil_header_in_syncing_context(zilog);
2733	uint64_t txg = dmu_tx_get_txg(tx);
2734	spa_t *spa = zilog->zl_spa;
2735	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
2736	lwb_t *lwb;
2737
2738	/*
2739	 * We don't zero out zl_destroy_txg, so make sure we don't try
2740	 * to destroy it twice.
2741	 */
2742	if (spa_sync_pass(spa) != 1)
2743		return;
2744
2745	mutex_enter(&zilog->zl_lock);
2746
2747	ASSERT(zilog->zl_stop_sync == 0);
2748
2749	if (*replayed_seq != 0) {
2750		ASSERT(zh->zh_replay_seq < *replayed_seq);
2751		zh->zh_replay_seq = *replayed_seq;
2752		*replayed_seq = 0;
2753	}
2754
2755	if (zilog->zl_destroy_txg == txg) {
2756		blkptr_t blk = zh->zh_log;
2757
2758		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
2759
2760		bzero(zh, sizeof (zil_header_t));
2761		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
2762
2763		if (zilog->zl_keep_first) {
2764			/*
2765			 * If this block was part of log chain that couldn't
2766			 * be claimed because a device was missing during
2767			 * zil_claim(), but that device later returns,
2768			 * then this block could erroneously appear valid.
2769			 * To guard against this, assign a new GUID to the new
2770			 * log chain so it doesn't matter what blk points to.
2771			 */
2772			zil_init_log_chain(zilog, &blk);
2773			zh->zh_log = blk;
2774		}
2775	}
2776
2777	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
2778		zh->zh_log = lwb->lwb_blk;
2779		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
2780			break;
2781		list_remove(&zilog->zl_lwb_list, lwb);
2782		zio_free(spa, txg, &lwb->lwb_blk);
2783		zil_free_lwb(zilog, lwb);
2784
2785		/*
2786		 * If we don't have anything left in the lwb list then
2787		 * we've had an allocation failure and we need to zero
2788		 * out the zil_header blkptr so that we don't end
2789		 * up freeing the same block twice.
2790		 */
2791		if (list_head(&zilog->zl_lwb_list) == NULL)
2792			BP_ZERO(&zh->zh_log);
2793	}
2794	mutex_exit(&zilog->zl_lock);
2795}
2796
2797/* ARGSUSED */
2798static int
2799zil_lwb_cons(void *vbuf, void *unused, int kmflag)
2800{
2801	lwb_t *lwb = vbuf;
2802	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
2803	    offsetof(zil_commit_waiter_t, zcw_node));
2804	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
2805	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
2806	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
2807	return (0);
2808}
2809
2810/* ARGSUSED */
2811static void
2812zil_lwb_dest(void *vbuf, void *unused)
2813{
2814	lwb_t *lwb = vbuf;
2815	mutex_destroy(&lwb->lwb_vdev_lock);
2816	avl_destroy(&lwb->lwb_vdev_tree);
2817	list_destroy(&lwb->lwb_waiters);
2818}
2819
2820void
2821zil_init(void)
2822{
2823	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
2824	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
2825
2826	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
2827	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2828}
2829
2830void
2831zil_fini(void)
2832{
2833	kmem_cache_destroy(zil_zcw_cache);
2834	kmem_cache_destroy(zil_lwb_cache);
2835}
2836
2837void
2838zil_set_sync(zilog_t *zilog, uint64_t sync)
2839{
2840	zilog->zl_sync = sync;
2841}
2842
2843void
2844zil_set_logbias(zilog_t *zilog, uint64_t logbias)
2845{
2846	zilog->zl_logbias = logbias;
2847}
2848
2849zilog_t *
2850zil_alloc(objset_t *os, zil_header_t *zh_phys)
2851{
2852	zilog_t *zilog;
2853
2854	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
2855
2856	zilog->zl_header = zh_phys;
2857	zilog->zl_os = os;
2858	zilog->zl_spa = dmu_objset_spa(os);
2859	zilog->zl_dmu_pool = dmu_objset_pool(os);
2860	zilog->zl_destroy_txg = TXG_INITIAL - 1;
2861	zilog->zl_logbias = dmu_objset_logbias(os);
2862	zilog->zl_sync = dmu_objset_syncprop(os);
2863	zilog->zl_dirty_max_txg = 0;
2864	zilog->zl_last_lwb_opened = NULL;
2865	zilog->zl_last_lwb_latency = 0;
2866
2867	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
2868	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
2869
2870	for (int i = 0; i < TXG_SIZE; i++) {
2871		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
2872		    MUTEX_DEFAULT, NULL);
2873	}
2874
2875	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
2876	    offsetof(lwb_t, lwb_node));
2877
2878	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
2879	    offsetof(itx_t, itx_node));
2880
2881	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
2882
2883	return (zilog);
2884}
2885
2886void
2887zil_free(zilog_t *zilog)
2888{
2889	zilog->zl_stop_sync = 1;
2890
2891	ASSERT0(zilog->zl_suspend);
2892	ASSERT0(zilog->zl_suspending);
2893
2894	ASSERT(list_is_empty(&zilog->zl_lwb_list));
2895	list_destroy(&zilog->zl_lwb_list);
2896
2897	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
2898	list_destroy(&zilog->zl_itx_commit_list);
2899
2900	for (int i = 0; i < TXG_SIZE; i++) {
2901		/*
2902		 * It's possible for an itx to be generated that doesn't dirty
2903		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
2904		 * callback to remove the entry. We remove those here.
2905		 *
2906		 * Also free up the ziltest itxs.
2907		 */
2908		if (zilog->zl_itxg[i].itxg_itxs)
2909			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
2910		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
2911	}
2912
2913	mutex_destroy(&zilog->zl_issuer_lock);
2914	mutex_destroy(&zilog->zl_lock);
2915
2916	cv_destroy(&zilog->zl_cv_suspend);
2917
2918	kmem_free(zilog, sizeof (zilog_t));
2919}
2920
2921/*
2922 * Open an intent log.
2923 */
2924zilog_t *
2925zil_open(objset_t *os, zil_get_data_t *get_data)
2926{
2927	zilog_t *zilog = dmu_objset_zil(os);
2928
2929	ASSERT3P(zilog->zl_get_data, ==, NULL);
2930	ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2931	ASSERT(list_is_empty(&zilog->zl_lwb_list));
2932
2933	zilog->zl_get_data = get_data;
2934
2935	return (zilog);
2936}
2937
2938/*
2939 * Close an intent log.
2940 */
2941void
2942zil_close(zilog_t *zilog)
2943{
2944	lwb_t *lwb;
2945	uint64_t txg;
2946
2947	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
2948		zil_commit(zilog, 0);
2949	} else {
2950		ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2951		ASSERT0(zilog->zl_dirty_max_txg);
2952		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
2953	}
2954
2955	mutex_enter(&zilog->zl_lock);
2956	lwb = list_tail(&zilog->zl_lwb_list);
2957	if (lwb == NULL)
2958		txg = zilog->zl_dirty_max_txg;
2959	else
2960		txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
2961	mutex_exit(&zilog->zl_lock);
2962
2963	/*
2964	 * We need to use txg_wait_synced() to wait long enough for the
2965	 * ZIL to be clean, and to wait for all pending lwbs to be
2966	 * written out.
2967	 */
2968	if (txg != 0)
2969		txg_wait_synced(zilog->zl_dmu_pool, txg);
2970
2971	if (zilog_is_dirty(zilog))
2972		zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
2973	VERIFY(!zilog_is_dirty(zilog));
2974
2975	zilog->zl_get_data = NULL;
2976
2977	/*
2978	 * We should have only one lwb left on the list; remove it now.
2979	 */
2980	mutex_enter(&zilog->zl_lock);
2981	lwb = list_head(&zilog->zl_lwb_list);
2982	if (lwb != NULL) {
2983		ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
2984		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2985		list_remove(&zilog->zl_lwb_list, lwb);
2986		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
2987		zil_free_lwb(zilog, lwb);
2988	}
2989	mutex_exit(&zilog->zl_lock);
2990}
2991
2992static char *suspend_tag = "zil suspending";
2993
2994/*
2995 * Suspend an intent log.  While in suspended mode, we still honor
2996 * synchronous semantics, but we rely on txg_wait_synced() to do it.
2997 * On old version pools, we suspend the log briefly when taking a
2998 * snapshot so that it will have an empty intent log.
2999 *
3000 * Long holds are not really intended to be used the way we do here --
3001 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
3002 * could fail.  Therefore we take pains to only put a long hold if it is
3003 * actually necessary.  Fortunately, it will only be necessary if the
3004 * objset is currently mounted (or the ZVOL equivalent).  In that case it
3005 * will already have a long hold, so we are not really making things any worse.
3006 *
3007 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3008 * zvol_state_t), and use their mechanism to prevent their hold from being
3009 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
3010 * very little gain.
3011 *
3012 * if cookiep == NULL, this does both the suspend & resume.
3013 * Otherwise, it returns with the dataset "long held", and the cookie
3014 * should be passed into zil_resume().
3015 */
3016int
3017zil_suspend(const char *osname, void **cookiep)
3018{
3019	objset_t *os;
3020	zilog_t *zilog;
3021	const zil_header_t *zh;
3022	int error;
3023
3024	error = dmu_objset_hold(osname, suspend_tag, &os);
3025	if (error != 0)
3026		return (error);
3027	zilog = dmu_objset_zil(os);
3028
3029	mutex_enter(&zilog->zl_lock);
3030	zh = zilog->zl_header;
3031
3032	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
3033		mutex_exit(&zilog->zl_lock);
3034		dmu_objset_rele(os, suspend_tag);
3035		return (SET_ERROR(EBUSY));
3036	}
3037
3038	/*
3039	 * Don't put a long hold in the cases where we can avoid it.  This
3040	 * is when there is no cookie so we are doing a suspend & resume
3041	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3042	 * for the suspend because it's already suspended, or there's no ZIL.
3043	 */
3044	if (cookiep == NULL && !zilog->zl_suspending &&
3045	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3046		mutex_exit(&zilog->zl_lock);
3047		dmu_objset_rele(os, suspend_tag);
3048		return (0);
3049	}
3050
3051	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3052	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3053
3054	zilog->zl_suspend++;
3055
3056	if (zilog->zl_suspend > 1) {
3057		/*
3058		 * Someone else is already suspending it.
3059		 * Just wait for them to finish.
3060		 */
3061
3062		while (zilog->zl_suspending)
3063			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3064		mutex_exit(&zilog->zl_lock);
3065
3066		if (cookiep == NULL)
3067			zil_resume(os);
3068		else
3069			*cookiep = os;
3070		return (0);
3071	}
3072
3073	/*
3074	 * If there is no pointer to an on-disk block, this ZIL must not
3075	 * be active (e.g. filesystem not mounted), so there's nothing
3076	 * to clean up.
3077	 */
3078	if (BP_IS_HOLE(&zh->zh_log)) {
3079		ASSERT(cookiep != NULL); /* fast path already handled */
3080
3081		*cookiep = os;
3082		mutex_exit(&zilog->zl_lock);
3083		return (0);
3084	}
3085
3086	zilog->zl_suspending = B_TRUE;
3087	mutex_exit(&zilog->zl_lock);
3088
3089	/*
3090	 * We need to use zil_commit_impl to ensure we wait for all
3091	 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwb's to be committed
3092	 * to disk before proceeding. If we used zil_commit instead, it
3093	 * would just call txg_wait_synced(), because zl_suspend is set.
3094	 * txg_wait_synced() doesn't wait for these lwb's to be
3095	 * LWB_STATE_DONE before returning.
3096	 */
3097	zil_commit_impl(zilog, 0);
3098
3099	/*
3100	 * Now that we've ensured all lwb's are LWB_STATE_DONE, we use
3101	 * txg_wait_synced() to ensure the data from the zilog has
3102	 * migrated to the main pool before calling zil_destroy().
3103	 */
3104	txg_wait_synced(zilog->zl_dmu_pool, 0);
3105
3106	zil_destroy(zilog, B_FALSE);
3107
3108	mutex_enter(&zilog->zl_lock);
3109	zilog->zl_suspending = B_FALSE;
3110	cv_broadcast(&zilog->zl_cv_suspend);
3111	mutex_exit(&zilog->zl_lock);
3112
3113	if (cookiep == NULL)
3114		zil_resume(os);
3115	else
3116		*cookiep = os;
3117	return (0);
3118}
3119
3120void
3121zil_resume(void *cookie)
3122{
3123	objset_t *os = cookie;
3124	zilog_t *zilog = dmu_objset_zil(os);
3125
3126	mutex_enter(&zilog->zl_lock);
3127	ASSERT(zilog->zl_suspend != 0);
3128	zilog->zl_suspend--;
3129	mutex_exit(&zilog->zl_lock);
3130	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3131	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3132}
3133
3134typedef struct zil_replay_arg {
3135	zil_replay_func_t **zr_replay;
3136	void		*zr_arg;
3137	boolean_t	zr_byteswap;
3138	char		*zr_lr;
3139} zil_replay_arg_t;
3140
3141static int
3142zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
3143{
3144	char name[ZFS_MAX_DATASET_NAME_LEN];
3145
3146	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
3147
3148	dmu_objset_name(zilog->zl_os, name);
3149
3150	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3151	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3152	    (u_longlong_t)lr->lrc_seq,
3153	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3154	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
3155
3156	return (error);
3157}
3158
3159static int
3160zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
3161{
3162	zil_replay_arg_t *zr = zra;
3163	const zil_header_t *zh = zilog->zl_header;
3164	uint64_t reclen = lr->lrc_reclen;
3165	uint64_t txtype = lr->lrc_txtype;
3166	int error = 0;
3167
3168	zilog->zl_replaying_seq = lr->lrc_seq;
3169
3170	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
3171		return (0);
3172
3173	if (lr->lrc_txg < claim_txg)		/* already committed */
3174		return (0);
3175
3176	/* Strip case-insensitive bit, still present in log record */
3177	txtype &= ~TX_CI;
3178
3179	if (txtype == 0 || txtype >= TX_MAX_TYPE)
3180		return (zil_replay_error(zilog, lr, EINVAL));
3181
3182	/*
3183	 * If this record type can be logged out of order, the object
3184	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
3185	 */
3186	if (TX_OOO(txtype)) {
3187		error = dmu_object_info(zilog->zl_os,
3188		    ((lr_ooo_t *)lr)->lr_foid, NULL);
3189		if (error == ENOENT || error == EEXIST)
3190			return (0);
3191	}
3192
3193	/*
3194	 * Make a copy of the data so we can revise and extend it.
3195	 */
3196	bcopy(lr, zr->zr_lr, reclen);
3197
3198	/*
3199	 * If this is a TX_WRITE with a blkptr, suck in the data.
3200	 */
3201	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3202		error = zil_read_log_data(zilog, (lr_write_t *)lr,
3203		    zr->zr_lr + reclen);
3204		if (error != 0)
3205			return (zil_replay_error(zilog, lr, error));
3206	}
3207
3208	/*
3209	 * The log block containing this lr may have been byteswapped
3210	 * so that we can easily examine common fields like lrc_txtype.
3211	 * However, the log is a mix of different record types, and only the
3212	 * replay vectors know how to byteswap their records.  Therefore, if
3213	 * the lr was byteswapped, undo it before invoking the replay vector.
3214	 */
3215	if (zr->zr_byteswap)
3216		byteswap_uint64_array(zr->zr_lr, reclen);
3217
3218	/*
3219	 * We must now do two things atomically: replay this log record,
3220	 * and update the log header sequence number to reflect the fact that
3221	 * we did so. At the end of each replay function the sequence number
3222	 * is updated if we are in replay mode.
3223	 */
3224	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3225	if (error != 0) {
3226		/*
3227		 * The DMU's dnode layer doesn't see removes until the txg
3228		 * commits, so a subsequent claim can spuriously fail with
3229		 * EEXIST. So if we receive any error we try syncing out
3230		 * any removes then retry the transaction.  Note that we
3231		 * specify B_FALSE for byteswap now, so we don't do it twice.
3232		 */
3233		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3234		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3235		if (error != 0)
3236			return (zil_replay_error(zilog, lr, error));
3237	}
3238	return (0);
3239}
3240
3241/* ARGSUSED */
3242static int
3243zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
3244{
3245	zilog->zl_replay_blks++;
3246
3247	return (0);
3248}
3249
3250/*
3251 * If this dataset has a non-empty intent log, replay it and destroy it.
3252 */
3253void
3254zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
3255{
3256	zilog_t *zilog = dmu_objset_zil(os);
3257	const zil_header_t *zh = zilog->zl_header;
3258	zil_replay_arg_t zr;
3259
3260	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3261		zil_destroy(zilog, B_TRUE);
3262		return;
3263	}
3264
3265	zr.zr_replay = replay_func;
3266	zr.zr_arg = arg;
3267	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3268	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3269
3270	/*
3271	 * Wait for in-progress removes to sync before starting replay.
3272	 */
3273	txg_wait_synced(zilog->zl_dmu_pool, 0);
3274
3275	zilog->zl_replay = B_TRUE;
3276	zilog->zl_replay_time = ddi_get_lbolt();
3277	ASSERT(zilog->zl_replay_blks == 0);
3278	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3279	    zh->zh_claim_txg);
3280	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3281
3282	zil_destroy(zilog, B_FALSE);
3283	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3284	zilog->zl_replay = B_FALSE;
3285}
3286
3287boolean_t
3288zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3289{
3290	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3291		return (B_TRUE);
3292
3293	if (zilog->zl_replay) {
3294		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3295		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3296		    zilog->zl_replaying_seq;
3297		return (B_TRUE);
3298	}
3299
3300	return (B_FALSE);
3301}
3302
3303/* ARGSUSED */
3304int
3305zil_reset(const char *osname, void *arg)
3306{
3307	int error;
3308
3309	error = zil_suspend(osname, NULL);
3310	if (error != 0)
3311		return (SET_ERROR(EEXIST));
3312	return (0);
3313}
3314