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