1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 */
26
27/* Portions Copyright 2010 Robert Milkowski */
28
29#include <sys/zfs_context.h>
30#include <sys/spa.h>
31#include <sys/dmu.h>
32#include <sys/zap.h>
33#include <sys/arc.h>
34#include <sys/stat.h>
35#include <sys/resource.h>
36#include <sys/zil.h>
37#include <sys/zil_impl.h>
38#include <sys/dsl_dataset.h>
39#include <sys/vdev_impl.h>
40#include <sys/dmu_tx.h>
41#include <sys/dsl_pool.h>
42#include <sys/abd.h>
43
44/*
45 * The zfs intent log (ZIL) saves transaction records of system calls
46 * that change the file system in memory with enough information
47 * to be able to replay them. These are stored in memory until
48 * either the DMU transaction group (txg) commits them to the stable pool
49 * and they can be discarded, or they are flushed to the stable log
50 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
51 * requirement. In the event of a panic or power fail then those log
52 * records (transactions) are replayed.
53 *
54 * There is one ZIL per file system. Its on-disk (pool) format consists
55 * of 3 parts:
56 *
57 * 	- ZIL header
58 * 	- ZIL blocks
59 * 	- ZIL records
60 *
61 * A log record holds a system call transaction. Log blocks can
62 * hold many log records and the blocks are chained together.
63 * Each ZIL block contains a block pointer (blkptr_t) to the next
64 * ZIL block in the chain. The ZIL header points to the first
65 * block in the chain. Note there is not a fixed place in the pool
66 * to hold blocks. They are dynamically allocated and freed as
67 * needed from the blocks available. Figure X shows the ZIL structure:
68 */
69
70/*
71 * Disable intent logging replay.  This global ZIL switch affects all pools.
72 */
73int zil_replay_disable = 0;
74
75/*
76 * Tunable parameter for debugging or performance analysis.  Setting
77 * zfs_nocacheflush will cause corruption on power loss if a volatile
78 * out-of-order write cache is enabled.
79 */
80boolean_t zfs_nocacheflush = B_FALSE;
81
82/*
83 * Limit SLOG write size per commit executed with synchronous priority.
84 * Any writes above that will be executed with lower (asynchronous) priority
85 * to limit potential SLOG device abuse by single active ZIL writer.
86 */
87uint64_t zil_slog_bulk = 768 * 1024;
88
89static kmem_cache_t *zil_lwb_cache;
90
91static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
92
93#define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
94    sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
95
96static int
97zil_bp_compare(const void *x1, const void *x2)
98{
99	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
100	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
101
102	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
103		return (-1);
104	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
105		return (1);
106
107	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
108		return (-1);
109	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
110		return (1);
111
112	return (0);
113}
114
115static void
116zil_bp_tree_init(zilog_t *zilog)
117{
118	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
119	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
120}
121
122static void
123zil_bp_tree_fini(zilog_t *zilog)
124{
125	avl_tree_t *t = &zilog->zl_bp_tree;
126	zil_bp_node_t *zn;
127	void *cookie = NULL;
128
129	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
130		kmem_free(zn, sizeof (zil_bp_node_t));
131
132	avl_destroy(t);
133}
134
135int
136zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
137{
138	avl_tree_t *t = &zilog->zl_bp_tree;
139	const dva_t *dva;
140	zil_bp_node_t *zn;
141	avl_index_t where;
142
143	if (BP_IS_EMBEDDED(bp))
144		return (0);
145
146	dva = BP_IDENTITY(bp);
147
148	if (avl_find(t, dva, &where) != NULL)
149		return (SET_ERROR(EEXIST));
150
151	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
152	zn->zn_dva = *dva;
153	avl_insert(t, zn, where);
154
155	return (0);
156}
157
158static zil_header_t *
159zil_header_in_syncing_context(zilog_t *zilog)
160{
161	return ((zil_header_t *)zilog->zl_header);
162}
163
164static void
165zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
166{
167	zio_cksum_t *zc = &bp->blk_cksum;
168
169	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
170	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
171	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
172	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
173}
174
175/*
176 * Read a log block and make sure it's valid.
177 */
178static int
179zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
180    char **end)
181{
182	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
183	arc_flags_t aflags = ARC_FLAG_WAIT;
184	arc_buf_t *abuf = NULL;
185	zbookmark_phys_t zb;
186	int error;
187
188	if (zilog->zl_header->zh_claim_txg == 0)
189		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
190
191	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
192		zio_flags |= ZIO_FLAG_SPECULATIVE;
193
194	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
195	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
196
197	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
198	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
199
200	if (error == 0) {
201		zio_cksum_t cksum = bp->blk_cksum;
202
203		/*
204		 * Validate the checksummed log block.
205		 *
206		 * Sequence numbers should be... sequential.  The checksum
207		 * verifier for the next block should be bp's checksum plus 1.
208		 *
209		 * Also check the log chain linkage and size used.
210		 */
211		cksum.zc_word[ZIL_ZC_SEQ]++;
212
213		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
214			zil_chain_t *zilc = abuf->b_data;
215			char *lr = (char *)(zilc + 1);
216			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
217
218			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
219			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
220				error = SET_ERROR(ECKSUM);
221			} else {
222				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
223				bcopy(lr, dst, len);
224				*end = (char *)dst + len;
225				*nbp = zilc->zc_next_blk;
226			}
227		} else {
228			char *lr = abuf->b_data;
229			uint64_t size = BP_GET_LSIZE(bp);
230			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
231
232			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
233			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
234			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
235				error = SET_ERROR(ECKSUM);
236			} else {
237				ASSERT3U(zilc->zc_nused, <=,
238				    SPA_OLD_MAXBLOCKSIZE);
239				bcopy(lr, dst, zilc->zc_nused);
240				*end = (char *)dst + zilc->zc_nused;
241				*nbp = zilc->zc_next_blk;
242			}
243		}
244
245		arc_buf_destroy(abuf, &abuf);
246	}
247
248	return (error);
249}
250
251/*
252 * Read a TX_WRITE log data block.
253 */
254static int
255zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
256{
257	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
258	const blkptr_t *bp = &lr->lr_blkptr;
259	arc_flags_t aflags = ARC_FLAG_WAIT;
260	arc_buf_t *abuf = NULL;
261	zbookmark_phys_t zb;
262	int error;
263
264	if (BP_IS_HOLE(bp)) {
265		if (wbuf != NULL)
266			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
267		return (0);
268	}
269
270	if (zilog->zl_header->zh_claim_txg == 0)
271		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
272
273	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
274	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
275
276	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
277	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
278
279	if (error == 0) {
280		if (wbuf != NULL)
281			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
282		arc_buf_destroy(abuf, &abuf);
283	}
284
285	return (error);
286}
287
288/*
289 * Parse the intent log, and call parse_func for each valid record within.
290 */
291int
292zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
293    zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
294{
295	const zil_header_t *zh = zilog->zl_header;
296	boolean_t claimed = !!zh->zh_claim_txg;
297	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
298	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
299	uint64_t max_blk_seq = 0;
300	uint64_t max_lr_seq = 0;
301	uint64_t blk_count = 0;
302	uint64_t lr_count = 0;
303	blkptr_t blk, next_blk;
304	char *lrbuf, *lrp;
305	int error = 0;
306
307	/*
308	 * Old logs didn't record the maximum zh_claim_lr_seq.
309	 */
310	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
311		claim_lr_seq = UINT64_MAX;
312
313	/*
314	 * Starting at the block pointed to by zh_log we read the log chain.
315	 * For each block in the chain we strongly check that block to
316	 * ensure its validity.  We stop when an invalid block is found.
317	 * For each block pointer in the chain we call parse_blk_func().
318	 * For each record in each valid block we call parse_lr_func().
319	 * If the log has been claimed, stop if we encounter a sequence
320	 * number greater than the highest claimed sequence number.
321	 */
322	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
323	zil_bp_tree_init(zilog);
324
325	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
326		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
327		int reclen;
328		char *end;
329
330		if (blk_seq > claim_blk_seq)
331			break;
332		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
333			break;
334		ASSERT3U(max_blk_seq, <, blk_seq);
335		max_blk_seq = blk_seq;
336		blk_count++;
337
338		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
339			break;
340
341		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
342		if (error != 0)
343			break;
344
345		for (lrp = lrbuf; lrp < end; lrp += reclen) {
346			lr_t *lr = (lr_t *)lrp;
347			reclen = lr->lrc_reclen;
348			ASSERT3U(reclen, >=, sizeof (lr_t));
349			if (lr->lrc_seq > claim_lr_seq)
350				goto done;
351			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
352				goto done;
353			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
354			max_lr_seq = lr->lrc_seq;
355			lr_count++;
356		}
357	}
358done:
359	zilog->zl_parse_error = error;
360	zilog->zl_parse_blk_seq = max_blk_seq;
361	zilog->zl_parse_lr_seq = max_lr_seq;
362	zilog->zl_parse_blk_count = blk_count;
363	zilog->zl_parse_lr_count = lr_count;
364
365	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
366	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
367
368	zil_bp_tree_fini(zilog);
369	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
370
371	return (error);
372}
373
374static int
375zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
376{
377	/*
378	 * Claim log block if not already committed and not already claimed.
379	 * If tx == NULL, just verify that the block is claimable.
380	 */
381	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
382	    zil_bp_tree_add(zilog, bp) != 0)
383		return (0);
384
385	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
386	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
387	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
388}
389
390static int
391zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
392{
393	lr_write_t *lr = (lr_write_t *)lrc;
394	int error;
395
396	if (lrc->lrc_txtype != TX_WRITE)
397		return (0);
398
399	/*
400	 * If the block is not readable, don't claim it.  This can happen
401	 * in normal operation when a log block is written to disk before
402	 * some of the dmu_sync() blocks it points to.  In this case, the
403	 * transaction cannot have been committed to anyone (we would have
404	 * waited for all writes to be stable first), so it is semantically
405	 * correct to declare this the end of the log.
406	 */
407	if (lr->lr_blkptr.blk_birth >= first_txg &&
408	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
409		return (error);
410	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
411}
412
413/* ARGSUSED */
414static int
415zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
416{
417	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
418
419	return (0);
420}
421
422static int
423zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
424{
425	lr_write_t *lr = (lr_write_t *)lrc;
426	blkptr_t *bp = &lr->lr_blkptr;
427
428	/*
429	 * If we previously claimed it, we need to free it.
430	 */
431	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
432	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
433	    !BP_IS_HOLE(bp))
434		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
435
436	return (0);
437}
438
439static lwb_t *
440zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
441{
442	lwb_t *lwb;
443
444	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
445	lwb->lwb_zilog = zilog;
446	lwb->lwb_blk = *bp;
447	lwb->lwb_slog = slog;
448	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
449	lwb->lwb_max_txg = txg;
450	lwb->lwb_zio = NULL;
451	lwb->lwb_tx = NULL;
452	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
453		lwb->lwb_nused = sizeof (zil_chain_t);
454		lwb->lwb_sz = BP_GET_LSIZE(bp);
455	} else {
456		lwb->lwb_nused = 0;
457		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
458	}
459
460	mutex_enter(&zilog->zl_lock);
461	list_insert_tail(&zilog->zl_lwb_list, lwb);
462	mutex_exit(&zilog->zl_lock);
463
464	return (lwb);
465}
466
467/*
468 * Called when we create in-memory log transactions so that we know
469 * to cleanup the itxs at the end of spa_sync().
470 */
471void
472zilog_dirty(zilog_t *zilog, uint64_t txg)
473{
474	dsl_pool_t *dp = zilog->zl_dmu_pool;
475	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
476
477	if (ds->ds_is_snapshot)
478		panic("dirtying snapshot!");
479
480	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
481		/* up the hold count until we can be written out */
482		dmu_buf_add_ref(ds->ds_dbuf, zilog);
483	}
484}
485
486/*
487 * Determine if the zil is dirty in the specified txg. Callers wanting to
488 * ensure that the dirty state does not change must hold the itxg_lock for
489 * the specified txg. Holding the lock will ensure that the zil cannot be
490 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
491 * state.
492 */
493boolean_t
494zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
495{
496	dsl_pool_t *dp = zilog->zl_dmu_pool;
497
498	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
499		return (B_TRUE);
500	return (B_FALSE);
501}
502
503/*
504 * Determine if the zil is dirty. The zil is considered dirty if it has
505 * any pending itx records that have not been cleaned by zil_clean().
506 */
507boolean_t
508zilog_is_dirty(zilog_t *zilog)
509{
510	dsl_pool_t *dp = zilog->zl_dmu_pool;
511
512	for (int t = 0; t < TXG_SIZE; t++) {
513		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
514			return (B_TRUE);
515	}
516	return (B_FALSE);
517}
518
519/*
520 * Create an on-disk intent log.
521 */
522static lwb_t *
523zil_create(zilog_t *zilog)
524{
525	const zil_header_t *zh = zilog->zl_header;
526	lwb_t *lwb = NULL;
527	uint64_t txg = 0;
528	dmu_tx_t *tx = NULL;
529	blkptr_t blk;
530	int error = 0;
531	boolean_t slog = FALSE;
532
533	/*
534	 * Wait for any previous destroy to complete.
535	 */
536	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
537
538	ASSERT(zh->zh_claim_txg == 0);
539	ASSERT(zh->zh_replay_seq == 0);
540
541	blk = zh->zh_log;
542
543	/*
544	 * Allocate an initial log block if:
545	 *    - there isn't one already
546	 *    - the existing block is the wrong endianess
547	 */
548	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
549		tx = dmu_tx_create(zilog->zl_os);
550		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
551		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
552		txg = dmu_tx_get_txg(tx);
553
554		if (!BP_IS_HOLE(&blk)) {
555			zio_free_zil(zilog->zl_spa, txg, &blk);
556			BP_ZERO(&blk);
557		}
558
559		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
560		    ZIL_MIN_BLKSZ, &slog);
561
562		if (error == 0)
563			zil_init_log_chain(zilog, &blk);
564	}
565
566	/*
567	 * Allocate a log write buffer (lwb) for the first log block.
568	 */
569	if (error == 0)
570		lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
571
572	/*
573	 * If we just allocated the first log block, commit our transaction
574	 * and wait for zil_sync() to stuff the block poiner into zh_log.
575	 * (zh is part of the MOS, so we cannot modify it in open context.)
576	 */
577	if (tx != NULL) {
578		dmu_tx_commit(tx);
579		txg_wait_synced(zilog->zl_dmu_pool, txg);
580	}
581
582	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
583
584	return (lwb);
585}
586
587/*
588 * In one tx, free all log blocks and clear the log header.
589 * If keep_first is set, then we're replaying a log with no content.
590 * We want to keep the first block, however, so that the first
591 * synchronous transaction doesn't require a txg_wait_synced()
592 * in zil_create().  We don't need to txg_wait_synced() here either
593 * when keep_first is set, because both zil_create() and zil_destroy()
594 * will wait for any in-progress destroys to complete.
595 */
596void
597zil_destroy(zilog_t *zilog, boolean_t keep_first)
598{
599	const zil_header_t *zh = zilog->zl_header;
600	lwb_t *lwb;
601	dmu_tx_t *tx;
602	uint64_t txg;
603
604	/*
605	 * Wait for any previous destroy to complete.
606	 */
607	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
608
609	zilog->zl_old_header = *zh;		/* debugging aid */
610
611	if (BP_IS_HOLE(&zh->zh_log))
612		return;
613
614	tx = dmu_tx_create(zilog->zl_os);
615	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
616	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
617	txg = dmu_tx_get_txg(tx);
618
619	mutex_enter(&zilog->zl_lock);
620
621	ASSERT3U(zilog->zl_destroy_txg, <, txg);
622	zilog->zl_destroy_txg = txg;
623	zilog->zl_keep_first = keep_first;
624
625	if (!list_is_empty(&zilog->zl_lwb_list)) {
626		ASSERT(zh->zh_claim_txg == 0);
627		VERIFY(!keep_first);
628		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
629			list_remove(&zilog->zl_lwb_list, lwb);
630			if (lwb->lwb_buf != NULL)
631				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
632			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
633			kmem_cache_free(zil_lwb_cache, lwb);
634		}
635	} else if (!keep_first) {
636		zil_destroy_sync(zilog, tx);
637	}
638	mutex_exit(&zilog->zl_lock);
639
640	dmu_tx_commit(tx);
641}
642
643void
644zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
645{
646	ASSERT(list_is_empty(&zilog->zl_lwb_list));
647	(void) zil_parse(zilog, zil_free_log_block,
648	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
649}
650
651int
652zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
653{
654	dmu_tx_t *tx = txarg;
655	uint64_t first_txg = dmu_tx_get_txg(tx);
656	zilog_t *zilog;
657	zil_header_t *zh;
658	objset_t *os;
659	int error;
660
661	error = dmu_objset_own_obj(dp, ds->ds_object,
662	    DMU_OST_ANY, B_FALSE, FTAG, &os);
663	if (error != 0) {
664		/*
665		 * EBUSY indicates that the objset is inconsistent, in which
666		 * case it can not have a ZIL.
667		 */
668		if (error != EBUSY) {
669			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
670			    (unsigned long long)ds->ds_object, error);
671		}
672		return (0);
673	}
674
675	zilog = dmu_objset_zil(os);
676	zh = zil_header_in_syncing_context(zilog);
677
678	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
679		if (!BP_IS_HOLE(&zh->zh_log))
680			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
681		BP_ZERO(&zh->zh_log);
682		dsl_dataset_dirty(dmu_objset_ds(os), tx);
683		dmu_objset_disown(os, FTAG);
684		return (0);
685	}
686
687	/*
688	 * Claim all log blocks if we haven't already done so, and remember
689	 * the highest claimed sequence number.  This ensures that if we can
690	 * read only part of the log now (e.g. due to a missing device),
691	 * but we can read the entire log later, we will not try to replay
692	 * or destroy beyond the last block we successfully claimed.
693	 */
694	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
695	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
696		(void) zil_parse(zilog, zil_claim_log_block,
697		    zil_claim_log_record, tx, first_txg);
698		zh->zh_claim_txg = first_txg;
699		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
700		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
701		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
702			zh->zh_flags |= ZIL_REPLAY_NEEDED;
703		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
704		dsl_dataset_dirty(dmu_objset_ds(os), tx);
705	}
706
707	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
708	dmu_objset_disown(os, FTAG);
709	return (0);
710}
711
712/*
713 * Check the log by walking the log chain.
714 * Checksum errors are ok as they indicate the end of the chain.
715 * Any other error (no device or read failure) returns an error.
716 */
717/* ARGSUSED */
718int
719zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
720{
721	zilog_t *zilog;
722	objset_t *os;
723	blkptr_t *bp;
724	int error;
725
726	ASSERT(tx == NULL);
727
728	error = dmu_objset_from_ds(ds, &os);
729	if (error != 0) {
730		cmn_err(CE_WARN, "can't open objset %llu, error %d",
731		    (unsigned long long)ds->ds_object, error);
732		return (0);
733	}
734
735	zilog = dmu_objset_zil(os);
736	bp = (blkptr_t *)&zilog->zl_header->zh_log;
737
738	/*
739	 * Check the first block and determine if it's on a log device
740	 * which may have been removed or faulted prior to loading this
741	 * pool.  If so, there's no point in checking the rest of the log
742	 * as its content should have already been synced to the pool.
743	 */
744	if (!BP_IS_HOLE(bp)) {
745		vdev_t *vd;
746		boolean_t valid = B_TRUE;
747
748		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
749		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
750		if (vd->vdev_islog && vdev_is_dead(vd))
751			valid = vdev_log_state_valid(vd);
752		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
753
754		if (!valid)
755			return (0);
756	}
757
758	/*
759	 * Because tx == NULL, zil_claim_log_block() will not actually claim
760	 * any blocks, but just determine whether it is possible to do so.
761	 * In addition to checking the log chain, zil_claim_log_block()
762	 * will invoke zio_claim() with a done func of spa_claim_notify(),
763	 * which will update spa_max_claim_txg.  See spa_load() for details.
764	 */
765	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
766	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
767
768	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
769}
770
771static int
772zil_vdev_compare(const void *x1, const void *x2)
773{
774	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
775	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
776
777	if (v1 < v2)
778		return (-1);
779	if (v1 > v2)
780		return (1);
781
782	return (0);
783}
784
785void
786zil_add_block(zilog_t *zilog, const blkptr_t *bp)
787{
788	avl_tree_t *t = &zilog->zl_vdev_tree;
789	avl_index_t where;
790	zil_vdev_node_t *zv, zvsearch;
791	int ndvas = BP_GET_NDVAS(bp);
792	int i;
793
794	if (zfs_nocacheflush)
795		return;
796
797	ASSERT(zilog->zl_writer);
798
799	/*
800	 * Even though we're zl_writer, we still need a lock because the
801	 * zl_get_data() callbacks may have dmu_sync() done callbacks
802	 * that will run concurrently.
803	 */
804	mutex_enter(&zilog->zl_vdev_lock);
805	for (i = 0; i < ndvas; i++) {
806		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
807		if (avl_find(t, &zvsearch, &where) == NULL) {
808			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
809			zv->zv_vdev = zvsearch.zv_vdev;
810			avl_insert(t, zv, where);
811		}
812	}
813	mutex_exit(&zilog->zl_vdev_lock);
814}
815
816static void
817zil_flush_vdevs(zilog_t *zilog)
818{
819	spa_t *spa = zilog->zl_spa;
820	avl_tree_t *t = &zilog->zl_vdev_tree;
821	void *cookie = NULL;
822	zil_vdev_node_t *zv;
823	zio_t *zio;
824
825	ASSERT(zilog->zl_writer);
826
827	/*
828	 * We don't need zl_vdev_lock here because we're the zl_writer,
829	 * and all zl_get_data() callbacks are done.
830	 */
831	if (avl_numnodes(t) == 0)
832		return;
833
834	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
835
836	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
837
838	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
839		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
840		if (vd != NULL)
841			zio_flush(zio, vd);
842		kmem_free(zv, sizeof (*zv));
843	}
844
845	/*
846	 * Wait for all the flushes to complete.  Not all devices actually
847	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
848	 */
849	(void) zio_wait(zio);
850
851	spa_config_exit(spa, SCL_STATE, FTAG);
852}
853
854/*
855 * Function called when a log block write completes
856 */
857static void
858zil_lwb_write_done(zio_t *zio)
859{
860	lwb_t *lwb = zio->io_private;
861	zilog_t *zilog = lwb->lwb_zilog;
862	dmu_tx_t *tx = lwb->lwb_tx;
863
864	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
865	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
866	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
867	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
868	ASSERT(!BP_IS_GANG(zio->io_bp));
869	ASSERT(!BP_IS_HOLE(zio->io_bp));
870	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
871
872	/*
873	 * Ensure the lwb buffer pointer is cleared before releasing
874	 * the txg. If we have had an allocation failure and
875	 * the txg is waiting to sync then we want want zil_sync()
876	 * to remove the lwb so that it's not picked up as the next new
877	 * one in zil_commit_writer(). zil_sync() will only remove
878	 * the lwb if lwb_buf is null.
879	 */
880	abd_put(zio->io_abd);
881	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
882	mutex_enter(&zilog->zl_lock);
883	lwb->lwb_buf = NULL;
884	lwb->lwb_tx = NULL;
885	mutex_exit(&zilog->zl_lock);
886
887	/*
888	 * Now that we've written this log block, we have a stable pointer
889	 * to the next block in the chain, so it's OK to let the txg in
890	 * which we allocated the next block sync.
891	 */
892	dmu_tx_commit(tx);
893}
894
895/*
896 * Initialize the io for a log block.
897 */
898static void
899zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
900{
901	zbookmark_phys_t zb;
902	zio_priority_t prio;
903
904	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
905	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
906	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
907
908	if (zilog->zl_root_zio == NULL) {
909		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
910		    ZIO_FLAG_CANFAIL);
911	}
912	if (lwb->lwb_zio == NULL) {
913		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
914		    BP_GET_LSIZE(&lwb->lwb_blk));
915		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
916			prio = ZIO_PRIORITY_SYNC_WRITE;
917		else
918			prio = ZIO_PRIORITY_ASYNC_WRITE;
919		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
920		    0, &lwb->lwb_blk, lwb_abd, BP_GET_LSIZE(&lwb->lwb_blk),
921		    zil_lwb_write_done, lwb, prio,
922		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
923	}
924}
925
926/*
927 * Define a limited set of intent log block sizes.
928 *
929 * These must be a multiple of 4KB. Note only the amount used (again
930 * aligned to 4KB) actually gets written. However, we can't always just
931 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
932 */
933uint64_t zil_block_buckets[] = {
934    4096,		/* non TX_WRITE */
935    8192+4096,		/* data base */
936    32*1024 + 4096, 	/* NFS writes */
937    UINT64_MAX
938};
939
940/*
941 * Start a log block write and advance to the next log block.
942 * Calls are serialized.
943 */
944static lwb_t *
945zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
946{
947	lwb_t *nlwb = NULL;
948	zil_chain_t *zilc;
949	spa_t *spa = zilog->zl_spa;
950	blkptr_t *bp;
951	dmu_tx_t *tx;
952	uint64_t txg;
953	uint64_t zil_blksz, wsz;
954	int i, error;
955	boolean_t slog;
956
957	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
958		zilc = (zil_chain_t *)lwb->lwb_buf;
959		bp = &zilc->zc_next_blk;
960	} else {
961		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
962		bp = &zilc->zc_next_blk;
963	}
964
965	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
966
967	/*
968	 * Allocate the next block and save its address in this block
969	 * before writing it in order to establish the log chain.
970	 * Note that if the allocation of nlwb synced before we wrote
971	 * the block that points at it (lwb), we'd leak it if we crashed.
972	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
973	 * We dirty the dataset to ensure that zil_sync() will be called
974	 * to clean up in the event of allocation failure or I/O failure.
975	 */
976	tx = dmu_tx_create(zilog->zl_os);
977
978	/*
979	 * Since we are not going to create any new dirty data and we can even
980	 * help with clearing the existing dirty data, we should not be subject
981	 * to the dirty data based delays.
982	 * We (ab)use TXG_WAITED to bypass the delay mechanism.
983	 * One side effect from using TXG_WAITED is that dmu_tx_assign() can
984	 * fail if the pool is suspended.  Those are dramatic circumstances,
985	 * so we return NULL to signal that the normal ZIL processing is not
986	 * possible and txg_wait_synced() should be used to ensure that the data
987	 * is on disk.
988	 */
989	error = dmu_tx_assign(tx, TXG_WAITED);
990	if (error != 0) {
991		ASSERT3S(error, ==, EIO);
992		dmu_tx_abort(tx);
993		return (NULL);
994	}
995	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
996	txg = dmu_tx_get_txg(tx);
997
998	lwb->lwb_tx = tx;
999
1000	/*
1001	 * Log blocks are pre-allocated. Here we select the size of the next
1002	 * block, based on size used in the last block.
1003	 * - first find the smallest bucket that will fit the block from a
1004	 *   limited set of block sizes. This is because it's faster to write
1005	 *   blocks allocated from the same metaslab as they are adjacent or
1006	 *   close.
1007	 * - next find the maximum from the new suggested size and an array of
1008	 *   previous sizes. This lessens a picket fence effect of wrongly
1009	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1010	 *   requests.
1011	 *
1012	 * Note we only write what is used, but we can't just allocate
1013	 * the maximum block size because we can exhaust the available
1014	 * pool log space.
1015	 */
1016	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1017	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1018		continue;
1019	zil_blksz = zil_block_buckets[i];
1020	if (zil_blksz == UINT64_MAX)
1021		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1022	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1023	for (i = 0; i < ZIL_PREV_BLKS; i++)
1024		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1025	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1026
1027	BP_ZERO(bp);
1028	/* pass the old blkptr in order to spread log blocks across devs */
1029	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1030	if (error == 0) {
1031		ASSERT3U(bp->blk_birth, ==, txg);
1032		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1033		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1034
1035		/*
1036		 * Allocate a new log write buffer (lwb).
1037		 */
1038		nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1039
1040		/* Record the block for later vdev flushing */
1041		zil_add_block(zilog, &lwb->lwb_blk);
1042	}
1043
1044	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1045		/* For Slim ZIL only write what is used. */
1046		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1047		ASSERT3U(wsz, <=, lwb->lwb_sz);
1048		zio_shrink(lwb->lwb_zio, wsz);
1049
1050	} else {
1051		wsz = lwb->lwb_sz;
1052	}
1053
1054	zilc->zc_pad = 0;
1055	zilc->zc_nused = lwb->lwb_nused;
1056	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1057
1058	/*
1059	 * clear unused data for security
1060	 */
1061	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1062
1063	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1064
1065	/*
1066	 * If there was an allocation failure then nlwb will be null which
1067	 * forces a txg_wait_synced().
1068	 */
1069	return (nlwb);
1070}
1071
1072static lwb_t *
1073zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1074{
1075	lr_t *lrcb, *lrc;
1076	lr_write_t *lrwb, *lrw;
1077	char *lr_buf;
1078	uint64_t dlen, dnow, lwb_sp, reclen, txg;
1079
1080	if (lwb == NULL)
1081		return (NULL);
1082
1083	ASSERT(lwb->lwb_buf != NULL);
1084
1085	lrc = &itx->itx_lr;		/* Common log record inside itx. */
1086	lrw = (lr_write_t *)lrc;	/* Write log record inside itx. */
1087	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1088		dlen = P2ROUNDUP_TYPED(
1089		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1090	} else {
1091		dlen = 0;
1092	}
1093	reclen = lrc->lrc_reclen;
1094	zilog->zl_cur_used += (reclen + dlen);
1095	txg = lrc->lrc_txg;
1096
1097	zil_lwb_write_init(zilog, lwb);
1098
1099cont:
1100	/*
1101	 * If this record won't fit in the current log block, start a new one.
1102	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1103	 */
1104	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1105	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1106	    lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1107	    lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1108		lwb = zil_lwb_write_start(zilog, lwb);
1109		if (lwb == NULL)
1110			return (NULL);
1111		zil_lwb_write_init(zilog, lwb);
1112		ASSERT(LWB_EMPTY(lwb));
1113		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1114		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1115	}
1116
1117	dnow = MIN(dlen, lwb_sp - reclen);
1118	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1119	bcopy(lrc, lr_buf, reclen);
1120	lrcb = (lr_t *)lr_buf;		/* Like lrc, but inside lwb. */
1121	lrwb = (lr_write_t *)lrcb;	/* Like lrw, but inside lwb. */
1122
1123	/*
1124	 * If it's a write, fetch the data or get its blkptr as appropriate.
1125	 */
1126	if (lrc->lrc_txtype == TX_WRITE) {
1127		if (txg > spa_freeze_txg(zilog->zl_spa))
1128			txg_wait_synced(zilog->zl_dmu_pool, txg);
1129		if (itx->itx_wr_state != WR_COPIED) {
1130			char *dbuf;
1131			int error;
1132
1133			if (itx->itx_wr_state == WR_NEED_COPY) {
1134				dbuf = lr_buf + reclen;
1135				lrcb->lrc_reclen += dnow;
1136				if (lrwb->lr_length > dnow)
1137					lrwb->lr_length = dnow;
1138				lrw->lr_offset += dnow;
1139				lrw->lr_length -= dnow;
1140			} else {
1141				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1142				dbuf = NULL;
1143			}
1144			error = zilog->zl_get_data(
1145			    itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
1146			if (error == EIO) {
1147				txg_wait_synced(zilog->zl_dmu_pool, txg);
1148				return (lwb);
1149			}
1150			if (error != 0) {
1151				ASSERT(error == ENOENT || error == EEXIST ||
1152				    error == EALREADY);
1153				return (lwb);
1154			}
1155		}
1156	}
1157
1158	/*
1159	 * We're actually making an entry, so update lrc_seq to be the
1160	 * log record sequence number.  Note that this is generally not
1161	 * equal to the itx sequence number because not all transactions
1162	 * are synchronous, and sometimes spa_sync() gets there first.
1163	 */
1164	lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1165	lwb->lwb_nused += reclen + dnow;
1166	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1167	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1168	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1169
1170	dlen -= dnow;
1171	if (dlen > 0) {
1172		zilog->zl_cur_used += reclen;
1173		goto cont;
1174	}
1175
1176	return (lwb);
1177}
1178
1179itx_t *
1180zil_itx_create(uint64_t txtype, size_t lrsize)
1181{
1182	itx_t *itx;
1183
1184	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1185
1186	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1187	itx->itx_lr.lrc_txtype = txtype;
1188	itx->itx_lr.lrc_reclen = lrsize;
1189	itx->itx_lr.lrc_seq = 0;	/* defensive */
1190	itx->itx_sync = B_TRUE;		/* default is synchronous */
1191
1192	return (itx);
1193}
1194
1195void
1196zil_itx_destroy(itx_t *itx)
1197{
1198	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1199}
1200
1201/*
1202 * Free up the sync and async itxs. The itxs_t has already been detached
1203 * so no locks are needed.
1204 */
1205static void
1206zil_itxg_clean(itxs_t *itxs)
1207{
1208	itx_t *itx;
1209	list_t *list;
1210	avl_tree_t *t;
1211	void *cookie;
1212	itx_async_node_t *ian;
1213
1214	list = &itxs->i_sync_list;
1215	while ((itx = list_head(list)) != NULL) {
1216		list_remove(list, itx);
1217		kmem_free(itx, offsetof(itx_t, itx_lr) +
1218		    itx->itx_lr.lrc_reclen);
1219	}
1220
1221	cookie = NULL;
1222	t = &itxs->i_async_tree;
1223	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1224		list = &ian->ia_list;
1225		while ((itx = list_head(list)) != NULL) {
1226			list_remove(list, itx);
1227			kmem_free(itx, offsetof(itx_t, itx_lr) +
1228			    itx->itx_lr.lrc_reclen);
1229		}
1230		list_destroy(list);
1231		kmem_free(ian, sizeof (itx_async_node_t));
1232	}
1233	avl_destroy(t);
1234
1235	kmem_free(itxs, sizeof (itxs_t));
1236}
1237
1238static int
1239zil_aitx_compare(const void *x1, const void *x2)
1240{
1241	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1242	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1243
1244	if (o1 < o2)
1245		return (-1);
1246	if (o1 > o2)
1247		return (1);
1248
1249	return (0);
1250}
1251
1252/*
1253 * Remove all async itx with the given oid.
1254 */
1255static void
1256zil_remove_async(zilog_t *zilog, uint64_t oid)
1257{
1258	uint64_t otxg, txg;
1259	itx_async_node_t *ian;
1260	avl_tree_t *t;
1261	avl_index_t where;
1262	list_t clean_list;
1263	itx_t *itx;
1264
1265	ASSERT(oid != 0);
1266	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1267
1268	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1269		otxg = ZILTEST_TXG;
1270	else
1271		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1272
1273	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1274		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1275
1276		mutex_enter(&itxg->itxg_lock);
1277		if (itxg->itxg_txg != txg) {
1278			mutex_exit(&itxg->itxg_lock);
1279			continue;
1280		}
1281
1282		/*
1283		 * Locate the object node and append its list.
1284		 */
1285		t = &itxg->itxg_itxs->i_async_tree;
1286		ian = avl_find(t, &oid, &where);
1287		if (ian != NULL)
1288			list_move_tail(&clean_list, &ian->ia_list);
1289		mutex_exit(&itxg->itxg_lock);
1290	}
1291	while ((itx = list_head(&clean_list)) != NULL) {
1292		list_remove(&clean_list, itx);
1293		kmem_free(itx, offsetof(itx_t, itx_lr) +
1294		    itx->itx_lr.lrc_reclen);
1295	}
1296	list_destroy(&clean_list);
1297}
1298
1299void
1300zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1301{
1302	uint64_t txg;
1303	itxg_t *itxg;
1304	itxs_t *itxs, *clean = NULL;
1305
1306	/*
1307	 * Object ids can be re-instantiated in the next txg so
1308	 * remove any async transactions to avoid future leaks.
1309	 * This can happen if a fsync occurs on the re-instantiated
1310	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1311	 * the new file data and flushes a write record for the old object.
1312	 */
1313	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1314		zil_remove_async(zilog, itx->itx_oid);
1315
1316	/*
1317	 * Ensure the data of a renamed file is committed before the rename.
1318	 */
1319	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1320		zil_async_to_sync(zilog, itx->itx_oid);
1321
1322	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1323		txg = ZILTEST_TXG;
1324	else
1325		txg = dmu_tx_get_txg(tx);
1326
1327	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1328	mutex_enter(&itxg->itxg_lock);
1329	itxs = itxg->itxg_itxs;
1330	if (itxg->itxg_txg != txg) {
1331		if (itxs != NULL) {
1332			/*
1333			 * The zil_clean callback hasn't got around to cleaning
1334			 * this itxg. Save the itxs for release below.
1335			 * This should be rare.
1336			 */
1337			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1338			    "txg %llu", itxg->itxg_txg);
1339			clean = itxg->itxg_itxs;
1340		}
1341		itxg->itxg_txg = txg;
1342		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1343
1344		list_create(&itxs->i_sync_list, sizeof (itx_t),
1345		    offsetof(itx_t, itx_node));
1346		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1347		    sizeof (itx_async_node_t),
1348		    offsetof(itx_async_node_t, ia_node));
1349	}
1350	if (itx->itx_sync) {
1351		list_insert_tail(&itxs->i_sync_list, itx);
1352	} else {
1353		avl_tree_t *t = &itxs->i_async_tree;
1354		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1355		itx_async_node_t *ian;
1356		avl_index_t where;
1357
1358		ian = avl_find(t, &foid, &where);
1359		if (ian == NULL) {
1360			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1361			list_create(&ian->ia_list, sizeof (itx_t),
1362			    offsetof(itx_t, itx_node));
1363			ian->ia_foid = foid;
1364			avl_insert(t, ian, where);
1365		}
1366		list_insert_tail(&ian->ia_list, itx);
1367	}
1368
1369	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1370	zilog_dirty(zilog, txg);
1371	mutex_exit(&itxg->itxg_lock);
1372
1373	/* Release the old itxs now we've dropped the lock */
1374	if (clean != NULL)
1375		zil_itxg_clean(clean);
1376}
1377
1378/*
1379 * If there are any in-memory intent log transactions which have now been
1380 * synced then start up a taskq to free them. We should only do this after we
1381 * have written out the uberblocks (i.e. txg has been comitted) so that
1382 * don't inadvertently clean out in-memory log records that would be required
1383 * by zil_commit().
1384 */
1385void
1386zil_clean(zilog_t *zilog, uint64_t synced_txg)
1387{
1388	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1389	itxs_t *clean_me;
1390
1391	mutex_enter(&itxg->itxg_lock);
1392	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1393		mutex_exit(&itxg->itxg_lock);
1394		return;
1395	}
1396	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1397	ASSERT(itxg->itxg_txg != 0);
1398	ASSERT(zilog->zl_clean_taskq != NULL);
1399	clean_me = itxg->itxg_itxs;
1400	itxg->itxg_itxs = NULL;
1401	itxg->itxg_txg = 0;
1402	mutex_exit(&itxg->itxg_lock);
1403	/*
1404	 * Preferably start a task queue to free up the old itxs but
1405	 * if taskq_dispatch can't allocate resources to do that then
1406	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1407	 * created a bad performance problem.
1408	 */
1409	if (taskq_dispatch(zilog->zl_clean_taskq,
1410	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1411		zil_itxg_clean(clean_me);
1412}
1413
1414/*
1415 * Get the list of itxs to commit into zl_itx_commit_list.
1416 */
1417static void
1418zil_get_commit_list(zilog_t *zilog)
1419{
1420	uint64_t otxg, txg;
1421	list_t *commit_list = &zilog->zl_itx_commit_list;
1422
1423	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1424		otxg = ZILTEST_TXG;
1425	else
1426		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1427
1428	/*
1429	 * This is inherently racy, since there is nothing to prevent
1430	 * the last synced txg from changing. That's okay since we'll
1431	 * only commit things in the future.
1432	 */
1433	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1434		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1435
1436		mutex_enter(&itxg->itxg_lock);
1437		if (itxg->itxg_txg != txg) {
1438			mutex_exit(&itxg->itxg_lock);
1439			continue;
1440		}
1441
1442		/*
1443		 * If we're adding itx records to the zl_itx_commit_list,
1444		 * then the zil better be dirty in this "txg". We can assert
1445		 * that here since we're holding the itxg_lock which will
1446		 * prevent spa_sync from cleaning it. Once we add the itxs
1447		 * to the zl_itx_commit_list we must commit it to disk even
1448		 * if it's unnecessary (i.e. the txg was synced).
1449		 */
1450		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1451		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1452		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1453
1454		mutex_exit(&itxg->itxg_lock);
1455	}
1456}
1457
1458/*
1459 * Move the async itxs for a specified object to commit into sync lists.
1460 */
1461static void
1462zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1463{
1464	uint64_t otxg, txg;
1465	itx_async_node_t *ian;
1466	avl_tree_t *t;
1467	avl_index_t where;
1468
1469	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1470		otxg = ZILTEST_TXG;
1471	else
1472		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1473
1474	/*
1475	 * This is inherently racy, since there is nothing to prevent
1476	 * the last synced txg from changing.
1477	 */
1478	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1479		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1480
1481		mutex_enter(&itxg->itxg_lock);
1482		if (itxg->itxg_txg != txg) {
1483			mutex_exit(&itxg->itxg_lock);
1484			continue;
1485		}
1486
1487		/*
1488		 * If a foid is specified then find that node and append its
1489		 * list. Otherwise walk the tree appending all the lists
1490		 * to the sync list. We add to the end rather than the
1491		 * beginning to ensure the create has happened.
1492		 */
1493		t = &itxg->itxg_itxs->i_async_tree;
1494		if (foid != 0) {
1495			ian = avl_find(t, &foid, &where);
1496			if (ian != NULL) {
1497				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1498				    &ian->ia_list);
1499			}
1500		} else {
1501			void *cookie = NULL;
1502
1503			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1504				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1505				    &ian->ia_list);
1506				list_destroy(&ian->ia_list);
1507				kmem_free(ian, sizeof (itx_async_node_t));
1508			}
1509		}
1510		mutex_exit(&itxg->itxg_lock);
1511	}
1512}
1513
1514static void
1515zil_commit_writer(zilog_t *zilog)
1516{
1517	uint64_t txg;
1518	itx_t *itx;
1519	lwb_t *lwb;
1520	spa_t *spa = zilog->zl_spa;
1521	int error = 0;
1522
1523	ASSERT(zilog->zl_root_zio == NULL);
1524
1525	mutex_exit(&zilog->zl_lock);
1526
1527	zil_get_commit_list(zilog);
1528
1529	/*
1530	 * Return if there's nothing to commit before we dirty the fs by
1531	 * calling zil_create().
1532	 */
1533	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1534		mutex_enter(&zilog->zl_lock);
1535		return;
1536	}
1537
1538	if (zilog->zl_suspend) {
1539		lwb = NULL;
1540	} else {
1541		lwb = list_tail(&zilog->zl_lwb_list);
1542		if (lwb == NULL)
1543			lwb = zil_create(zilog);
1544	}
1545
1546	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1547	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1548		txg = itx->itx_lr.lrc_txg;
1549		ASSERT3U(txg, !=, 0);
1550
1551		/*
1552		 * This is inherently racy and may result in us writing
1553		 * out a log block for a txg that was just synced. This is
1554		 * ok since we'll end cleaning up that log block the next
1555		 * time we call zil_sync().
1556		 */
1557		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1558			lwb = zil_lwb_commit(zilog, itx, lwb);
1559		list_remove(&zilog->zl_itx_commit_list, itx);
1560		kmem_free(itx, offsetof(itx_t, itx_lr)
1561		    + itx->itx_lr.lrc_reclen);
1562	}
1563	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1564
1565	/* write the last block out */
1566	if (lwb != NULL && lwb->lwb_zio != NULL)
1567		lwb = zil_lwb_write_start(zilog, lwb);
1568
1569	zilog->zl_cur_used = 0;
1570
1571	/*
1572	 * Wait if necessary for the log blocks to be on stable storage.
1573	 */
1574	if (zilog->zl_root_zio) {
1575		error = zio_wait(zilog->zl_root_zio);
1576		zilog->zl_root_zio = NULL;
1577		zil_flush_vdevs(zilog);
1578	}
1579
1580	if (error || lwb == NULL)
1581		txg_wait_synced(zilog->zl_dmu_pool, 0);
1582
1583	mutex_enter(&zilog->zl_lock);
1584
1585	/*
1586	 * Remember the highest committed log sequence number for ztest.
1587	 * We only update this value when all the log writes succeeded,
1588	 * because ztest wants to ASSERT that it got the whole log chain.
1589	 */
1590	if (error == 0 && lwb != NULL)
1591		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1592}
1593
1594/*
1595 * Commit zfs transactions to stable storage.
1596 * If foid is 0 push out all transactions, otherwise push only those
1597 * for that object or might reference that object.
1598 *
1599 * itxs are committed in batches. In a heavily stressed zil there will be
1600 * a commit writer thread who is writing out a bunch of itxs to the log
1601 * for a set of committing threads (cthreads) in the same batch as the writer.
1602 * Those cthreads are all waiting on the same cv for that batch.
1603 *
1604 * There will also be a different and growing batch of threads that are
1605 * waiting to commit (qthreads). When the committing batch completes
1606 * a transition occurs such that the cthreads exit and the qthreads become
1607 * cthreads. One of the new cthreads becomes the writer thread for the
1608 * batch. Any new threads arriving become new qthreads.
1609 *
1610 * Only 2 condition variables are needed and there's no transition
1611 * between the two cvs needed. They just flip-flop between qthreads
1612 * and cthreads.
1613 *
1614 * Using this scheme we can efficiently wakeup up only those threads
1615 * that have been committed.
1616 */
1617void
1618zil_commit(zilog_t *zilog, uint64_t foid)
1619{
1620	uint64_t mybatch;
1621
1622	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1623		return;
1624
1625	/* move the async itxs for the foid to the sync queues */
1626	zil_async_to_sync(zilog, foid);
1627
1628	mutex_enter(&zilog->zl_lock);
1629	mybatch = zilog->zl_next_batch;
1630	while (zilog->zl_writer) {
1631		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1632		if (mybatch <= zilog->zl_com_batch) {
1633			mutex_exit(&zilog->zl_lock);
1634			return;
1635		}
1636	}
1637
1638	zilog->zl_next_batch++;
1639	zilog->zl_writer = B_TRUE;
1640	zil_commit_writer(zilog);
1641	zilog->zl_com_batch = mybatch;
1642	zilog->zl_writer = B_FALSE;
1643	mutex_exit(&zilog->zl_lock);
1644
1645	/* wake up one thread to become the next writer */
1646	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1647
1648	/* wake up all threads waiting for this batch to be committed */
1649	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1650}
1651
1652/*
1653 * Called in syncing context to free committed log blocks and update log header.
1654 */
1655void
1656zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1657{
1658	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1659	uint64_t txg = dmu_tx_get_txg(tx);
1660	spa_t *spa = zilog->zl_spa;
1661	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1662	lwb_t *lwb;
1663
1664	/*
1665	 * We don't zero out zl_destroy_txg, so make sure we don't try
1666	 * to destroy it twice.
1667	 */
1668	if (spa_sync_pass(spa) != 1)
1669		return;
1670
1671	mutex_enter(&zilog->zl_lock);
1672
1673	ASSERT(zilog->zl_stop_sync == 0);
1674
1675	if (*replayed_seq != 0) {
1676		ASSERT(zh->zh_replay_seq < *replayed_seq);
1677		zh->zh_replay_seq = *replayed_seq;
1678		*replayed_seq = 0;
1679	}
1680
1681	if (zilog->zl_destroy_txg == txg) {
1682		blkptr_t blk = zh->zh_log;
1683
1684		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1685
1686		bzero(zh, sizeof (zil_header_t));
1687		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1688
1689		if (zilog->zl_keep_first) {
1690			/*
1691			 * If this block was part of log chain that couldn't
1692			 * be claimed because a device was missing during
1693			 * zil_claim(), but that device later returns,
1694			 * then this block could erroneously appear valid.
1695			 * To guard against this, assign a new GUID to the new
1696			 * log chain so it doesn't matter what blk points to.
1697			 */
1698			zil_init_log_chain(zilog, &blk);
1699			zh->zh_log = blk;
1700		}
1701	}
1702
1703	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1704		zh->zh_log = lwb->lwb_blk;
1705		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1706			break;
1707		list_remove(&zilog->zl_lwb_list, lwb);
1708		zio_free_zil(spa, txg, &lwb->lwb_blk);
1709		kmem_cache_free(zil_lwb_cache, lwb);
1710
1711		/*
1712		 * If we don't have anything left in the lwb list then
1713		 * we've had an allocation failure and we need to zero
1714		 * out the zil_header blkptr so that we don't end
1715		 * up freeing the same block twice.
1716		 */
1717		if (list_head(&zilog->zl_lwb_list) == NULL)
1718			BP_ZERO(&zh->zh_log);
1719	}
1720	mutex_exit(&zilog->zl_lock);
1721}
1722
1723void
1724zil_init(void)
1725{
1726	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1727	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1728}
1729
1730void
1731zil_fini(void)
1732{
1733	kmem_cache_destroy(zil_lwb_cache);
1734}
1735
1736void
1737zil_set_sync(zilog_t *zilog, uint64_t sync)
1738{
1739	zilog->zl_sync = sync;
1740}
1741
1742void
1743zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1744{
1745	zilog->zl_logbias = logbias;
1746}
1747
1748zilog_t *
1749zil_alloc(objset_t *os, zil_header_t *zh_phys)
1750{
1751	zilog_t *zilog;
1752
1753	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1754
1755	zilog->zl_header = zh_phys;
1756	zilog->zl_os = os;
1757	zilog->zl_spa = dmu_objset_spa(os);
1758	zilog->zl_dmu_pool = dmu_objset_pool(os);
1759	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1760	zilog->zl_logbias = dmu_objset_logbias(os);
1761	zilog->zl_sync = dmu_objset_syncprop(os);
1762	zilog->zl_next_batch = 1;
1763
1764	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1765
1766	for (int i = 0; i < TXG_SIZE; i++) {
1767		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1768		    MUTEX_DEFAULT, NULL);
1769	}
1770
1771	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1772	    offsetof(lwb_t, lwb_node));
1773
1774	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1775	    offsetof(itx_t, itx_node));
1776
1777	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1778
1779	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1780	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1781
1782	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1783	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1784	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1785	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1786
1787	return (zilog);
1788}
1789
1790void
1791zil_free(zilog_t *zilog)
1792{
1793	zilog->zl_stop_sync = 1;
1794
1795	ASSERT0(zilog->zl_suspend);
1796	ASSERT0(zilog->zl_suspending);
1797
1798	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1799	list_destroy(&zilog->zl_lwb_list);
1800
1801	avl_destroy(&zilog->zl_vdev_tree);
1802	mutex_destroy(&zilog->zl_vdev_lock);
1803
1804	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1805	list_destroy(&zilog->zl_itx_commit_list);
1806
1807	for (int i = 0; i < TXG_SIZE; i++) {
1808		/*
1809		 * It's possible for an itx to be generated that doesn't dirty
1810		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1811		 * callback to remove the entry. We remove those here.
1812		 *
1813		 * Also free up the ziltest itxs.
1814		 */
1815		if (zilog->zl_itxg[i].itxg_itxs)
1816			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1817		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1818	}
1819
1820	mutex_destroy(&zilog->zl_lock);
1821
1822	cv_destroy(&zilog->zl_cv_writer);
1823	cv_destroy(&zilog->zl_cv_suspend);
1824	cv_destroy(&zilog->zl_cv_batch[0]);
1825	cv_destroy(&zilog->zl_cv_batch[1]);
1826
1827	kmem_free(zilog, sizeof (zilog_t));
1828}
1829
1830/*
1831 * Open an intent log.
1832 */
1833zilog_t *
1834zil_open(objset_t *os, zil_get_data_t *get_data)
1835{
1836	zilog_t *zilog = dmu_objset_zil(os);
1837
1838	ASSERT(zilog->zl_clean_taskq == NULL);
1839	ASSERT(zilog->zl_get_data == NULL);
1840	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1841
1842	zilog->zl_get_data = get_data;
1843	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1844	    2, 2, TASKQ_PREPOPULATE);
1845
1846	return (zilog);
1847}
1848
1849/*
1850 * Close an intent log.
1851 */
1852void
1853zil_close(zilog_t *zilog)
1854{
1855	lwb_t *lwb;
1856	uint64_t txg = 0;
1857
1858	zil_commit(zilog, 0); /* commit all itx */
1859
1860	/*
1861	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1862	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1863	 * callbacks have occurred that may clean the zil.  Only then can we
1864	 * destroy the zl_clean_taskq.
1865	 */
1866	mutex_enter(&zilog->zl_lock);
1867	lwb = list_tail(&zilog->zl_lwb_list);
1868	if (lwb != NULL)
1869		txg = lwb->lwb_max_txg;
1870	mutex_exit(&zilog->zl_lock);
1871	if (txg)
1872		txg_wait_synced(zilog->zl_dmu_pool, txg);
1873
1874	if (zilog_is_dirty(zilog))
1875		zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1876	VERIFY(!zilog_is_dirty(zilog));
1877
1878	taskq_destroy(zilog->zl_clean_taskq);
1879	zilog->zl_clean_taskq = NULL;
1880	zilog->zl_get_data = NULL;
1881
1882	/*
1883	 * We should have only one LWB left on the list; remove it now.
1884	 */
1885	mutex_enter(&zilog->zl_lock);
1886	lwb = list_head(&zilog->zl_lwb_list);
1887	if (lwb != NULL) {
1888		ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1889		list_remove(&zilog->zl_lwb_list, lwb);
1890		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1891		kmem_cache_free(zil_lwb_cache, lwb);
1892	}
1893	mutex_exit(&zilog->zl_lock);
1894}
1895
1896static char *suspend_tag = "zil suspending";
1897
1898/*
1899 * Suspend an intent log.  While in suspended mode, we still honor
1900 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1901 * On old version pools, we suspend the log briefly when taking a
1902 * snapshot so that it will have an empty intent log.
1903 *
1904 * Long holds are not really intended to be used the way we do here --
1905 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
1906 * could fail.  Therefore we take pains to only put a long hold if it is
1907 * actually necessary.  Fortunately, it will only be necessary if the
1908 * objset is currently mounted (or the ZVOL equivalent).  In that case it
1909 * will already have a long hold, so we are not really making things any worse.
1910 *
1911 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1912 * zvol_state_t), and use their mechanism to prevent their hold from being
1913 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
1914 * very little gain.
1915 *
1916 * if cookiep == NULL, this does both the suspend & resume.
1917 * Otherwise, it returns with the dataset "long held", and the cookie
1918 * should be passed into zil_resume().
1919 */
1920int
1921zil_suspend(const char *osname, void **cookiep)
1922{
1923	objset_t *os;
1924	zilog_t *zilog;
1925	const zil_header_t *zh;
1926	int error;
1927
1928	error = dmu_objset_hold(osname, suspend_tag, &os);
1929	if (error != 0)
1930		return (error);
1931	zilog = dmu_objset_zil(os);
1932
1933	mutex_enter(&zilog->zl_lock);
1934	zh = zilog->zl_header;
1935
1936	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1937		mutex_exit(&zilog->zl_lock);
1938		dmu_objset_rele(os, suspend_tag);
1939		return (SET_ERROR(EBUSY));
1940	}
1941
1942	/*
1943	 * Don't put a long hold in the cases where we can avoid it.  This
1944	 * is when there is no cookie so we are doing a suspend & resume
1945	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1946	 * for the suspend because it's already suspended, or there's no ZIL.
1947	 */
1948	if (cookiep == NULL && !zilog->zl_suspending &&
1949	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1950		mutex_exit(&zilog->zl_lock);
1951		dmu_objset_rele(os, suspend_tag);
1952		return (0);
1953	}
1954
1955	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1956	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1957
1958	zilog->zl_suspend++;
1959
1960	if (zilog->zl_suspend > 1) {
1961		/*
1962		 * Someone else is already suspending it.
1963		 * Just wait for them to finish.
1964		 */
1965
1966		while (zilog->zl_suspending)
1967			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1968		mutex_exit(&zilog->zl_lock);
1969
1970		if (cookiep == NULL)
1971			zil_resume(os);
1972		else
1973			*cookiep = os;
1974		return (0);
1975	}
1976
1977	/*
1978	 * If there is no pointer to an on-disk block, this ZIL must not
1979	 * be active (e.g. filesystem not mounted), so there's nothing
1980	 * to clean up.
1981	 */
1982	if (BP_IS_HOLE(&zh->zh_log)) {
1983		ASSERT(cookiep != NULL); /* fast path already handled */
1984
1985		*cookiep = os;
1986		mutex_exit(&zilog->zl_lock);
1987		return (0);
1988	}
1989
1990	zilog->zl_suspending = B_TRUE;
1991	mutex_exit(&zilog->zl_lock);
1992
1993	zil_commit(zilog, 0);
1994
1995	zil_destroy(zilog, B_FALSE);
1996
1997	mutex_enter(&zilog->zl_lock);
1998	zilog->zl_suspending = B_FALSE;
1999	cv_broadcast(&zilog->zl_cv_suspend);
2000	mutex_exit(&zilog->zl_lock);
2001
2002	if (cookiep == NULL)
2003		zil_resume(os);
2004	else
2005		*cookiep = os;
2006	return (0);
2007}
2008
2009void
2010zil_resume(void *cookie)
2011{
2012	objset_t *os = cookie;
2013	zilog_t *zilog = dmu_objset_zil(os);
2014
2015	mutex_enter(&zilog->zl_lock);
2016	ASSERT(zilog->zl_suspend != 0);
2017	zilog->zl_suspend--;
2018	mutex_exit(&zilog->zl_lock);
2019	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2020	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2021}
2022
2023typedef struct zil_replay_arg {
2024	zil_replay_func_t **zr_replay;
2025	void		*zr_arg;
2026	boolean_t	zr_byteswap;
2027	char		*zr_lr;
2028} zil_replay_arg_t;
2029
2030static int
2031zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2032{
2033	char name[ZFS_MAX_DATASET_NAME_LEN];
2034
2035	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
2036
2037	dmu_objset_name(zilog->zl_os, name);
2038
2039	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2040	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2041	    (u_longlong_t)lr->lrc_seq,
2042	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2043	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
2044
2045	return (error);
2046}
2047
2048static int
2049zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2050{
2051	zil_replay_arg_t *zr = zra;
2052	const zil_header_t *zh = zilog->zl_header;
2053	uint64_t reclen = lr->lrc_reclen;
2054	uint64_t txtype = lr->lrc_txtype;
2055	int error = 0;
2056
2057	zilog->zl_replaying_seq = lr->lrc_seq;
2058
2059	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
2060		return (0);
2061
2062	if (lr->lrc_txg < claim_txg)		/* already committed */
2063		return (0);
2064
2065	/* Strip case-insensitive bit, still present in log record */
2066	txtype &= ~TX_CI;
2067
2068	if (txtype == 0 || txtype >= TX_MAX_TYPE)
2069		return (zil_replay_error(zilog, lr, EINVAL));
2070
2071	/*
2072	 * If this record type can be logged out of order, the object
2073	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
2074	 */
2075	if (TX_OOO(txtype)) {
2076		error = dmu_object_info(zilog->zl_os,
2077		    ((lr_ooo_t *)lr)->lr_foid, NULL);
2078		if (error == ENOENT || error == EEXIST)
2079			return (0);
2080	}
2081
2082	/*
2083	 * Make a copy of the data so we can revise and extend it.
2084	 */
2085	bcopy(lr, zr->zr_lr, reclen);
2086
2087	/*
2088	 * If this is a TX_WRITE with a blkptr, suck in the data.
2089	 */
2090	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2091		error = zil_read_log_data(zilog, (lr_write_t *)lr,
2092		    zr->zr_lr + reclen);
2093		if (error != 0)
2094			return (zil_replay_error(zilog, lr, error));
2095	}
2096
2097	/*
2098	 * The log block containing this lr may have been byteswapped
2099	 * so that we can easily examine common fields like lrc_txtype.
2100	 * However, the log is a mix of different record types, and only the
2101	 * replay vectors know how to byteswap their records.  Therefore, if
2102	 * the lr was byteswapped, undo it before invoking the replay vector.
2103	 */
2104	if (zr->zr_byteswap)
2105		byteswap_uint64_array(zr->zr_lr, reclen);
2106
2107	/*
2108	 * We must now do two things atomically: replay this log record,
2109	 * and update the log header sequence number to reflect the fact that
2110	 * we did so. At the end of each replay function the sequence number
2111	 * is updated if we are in replay mode.
2112	 */
2113	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2114	if (error != 0) {
2115		/*
2116		 * The DMU's dnode layer doesn't see removes until the txg
2117		 * commits, so a subsequent claim can spuriously fail with
2118		 * EEXIST. So if we receive any error we try syncing out
2119		 * any removes then retry the transaction.  Note that we
2120		 * specify B_FALSE for byteswap now, so we don't do it twice.
2121		 */
2122		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2123		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2124		if (error != 0)
2125			return (zil_replay_error(zilog, lr, error));
2126	}
2127	return (0);
2128}
2129
2130/* ARGSUSED */
2131static int
2132zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2133{
2134	zilog->zl_replay_blks++;
2135
2136	return (0);
2137}
2138
2139/*
2140 * If this dataset has a non-empty intent log, replay it and destroy it.
2141 */
2142void
2143zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2144{
2145	zilog_t *zilog = dmu_objset_zil(os);
2146	const zil_header_t *zh = zilog->zl_header;
2147	zil_replay_arg_t zr;
2148
2149	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2150		zil_destroy(zilog, B_TRUE);
2151		return;
2152	}
2153
2154	zr.zr_replay = replay_func;
2155	zr.zr_arg = arg;
2156	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2157	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2158
2159	/*
2160	 * Wait for in-progress removes to sync before starting replay.
2161	 */
2162	txg_wait_synced(zilog->zl_dmu_pool, 0);
2163
2164	zilog->zl_replay = B_TRUE;
2165	zilog->zl_replay_time = ddi_get_lbolt();
2166	ASSERT(zilog->zl_replay_blks == 0);
2167	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2168	    zh->zh_claim_txg);
2169	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2170
2171	zil_destroy(zilog, B_FALSE);
2172	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2173	zilog->zl_replay = B_FALSE;
2174}
2175
2176boolean_t
2177zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2178{
2179	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2180		return (B_TRUE);
2181
2182	if (zilog->zl_replay) {
2183		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2184		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2185		    zilog->zl_replaying_seq;
2186		return (B_TRUE);
2187	}
2188
2189	return (B_FALSE);
2190}
2191
2192/* ARGSUSED */
2193int
2194zil_vdev_offline(const char *osname, void *arg)
2195{
2196	int error;
2197
2198	error = zil_suspend(osname, NULL);
2199	if (error != 0)
2200		return (SET_ERROR(EEXIST));
2201	return (0);
2202}
2203