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, 2014 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
43/*
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
52 *
53 * There is one ZIL per file system. Its on-disk (pool) format consists
54 * of 3 parts:
55 *
56 * 	- ZIL header
57 * 	- ZIL blocks
58 * 	- ZIL records
59 *
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
67 */
68
69/*
70 * Disable intent logging replay.  This global ZIL switch affects all pools.
71 */
72int zil_replay_disable = 0;
73
74/*
75 * Tunable parameter for debugging or performance analysis.  Setting
76 * zfs_nocacheflush will cause corruption on power loss if a volatile
77 * out-of-order write cache is enabled.
78 */
79boolean_t zfs_nocacheflush = B_FALSE;
80
81static kmem_cache_t *zil_lwb_cache;
82
83static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
84
85#define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
86    sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
87
88
89/*
90 * ziltest is by and large an ugly hack, but very useful in
91 * checking replay without tedious work.
92 * When running ziltest we want to keep all itx's and so maintain
93 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
94 * We subtract TXG_CONCURRENT_STATES to allow for common code.
95 */
96#define	ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
97
98static int
99zil_bp_compare(const void *x1, const void *x2)
100{
101	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
102	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
103
104	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
105		return (-1);
106	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
107		return (1);
108
109	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
110		return (-1);
111	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
112		return (1);
113
114	return (0);
115}
116
117static void
118zil_bp_tree_init(zilog_t *zilog)
119{
120	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
121	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
122}
123
124static void
125zil_bp_tree_fini(zilog_t *zilog)
126{
127	avl_tree_t *t = &zilog->zl_bp_tree;
128	zil_bp_node_t *zn;
129	void *cookie = NULL;
130
131	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
132		kmem_free(zn, sizeof (zil_bp_node_t));
133
134	avl_destroy(t);
135}
136
137int
138zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
139{
140	avl_tree_t *t = &zilog->zl_bp_tree;
141	const dva_t *dva;
142	zil_bp_node_t *zn;
143	avl_index_t where;
144
145	if (BP_IS_EMBEDDED(bp))
146		return (0);
147
148	dva = BP_IDENTITY(bp);
149
150	if (avl_find(t, dva, &where) != NULL)
151		return (SET_ERROR(EEXIST));
152
153	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
154	zn->zn_dva = *dva;
155	avl_insert(t, zn, where);
156
157	return (0);
158}
159
160static zil_header_t *
161zil_header_in_syncing_context(zilog_t *zilog)
162{
163	return ((zil_header_t *)zilog->zl_header);
164}
165
166static void
167zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
168{
169	zio_cksum_t *zc = &bp->blk_cksum;
170
171	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
172	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
173	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
174	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
175}
176
177/*
178 * Read a log block and make sure it's valid.
179 */
180static int
181zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
182    char **end)
183{
184	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
185	arc_flags_t aflags = ARC_FLAG_WAIT;
186	arc_buf_t *abuf = NULL;
187	zbookmark_phys_t zb;
188	int error;
189
190	if (zilog->zl_header->zh_claim_txg == 0)
191		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
192
193	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
194		zio_flags |= ZIO_FLAG_SPECULATIVE;
195
196	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
197	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
198
199	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
200	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
201
202	if (error == 0) {
203		zio_cksum_t cksum = bp->blk_cksum;
204
205		/*
206		 * Validate the checksummed log block.
207		 *
208		 * Sequence numbers should be... sequential.  The checksum
209		 * verifier for the next block should be bp's checksum plus 1.
210		 *
211		 * Also check the log chain linkage and size used.
212		 */
213		cksum.zc_word[ZIL_ZC_SEQ]++;
214
215		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
216			zil_chain_t *zilc = abuf->b_data;
217			char *lr = (char *)(zilc + 1);
218			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
219
220			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
221			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
222				error = SET_ERROR(ECKSUM);
223			} else {
224				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
225				bcopy(lr, dst, len);
226				*end = (char *)dst + len;
227				*nbp = zilc->zc_next_blk;
228			}
229		} else {
230			char *lr = abuf->b_data;
231			uint64_t size = BP_GET_LSIZE(bp);
232			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
233
234			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
235			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
236			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
237				error = SET_ERROR(ECKSUM);
238			} else {
239				ASSERT3U(zilc->zc_nused, <=,
240				    SPA_OLD_MAXBLOCKSIZE);
241				bcopy(lr, dst, zilc->zc_nused);
242				*end = (char *)dst + zilc->zc_nused;
243				*nbp = zilc->zc_next_blk;
244			}
245		}
246
247		VERIFY(arc_buf_remove_ref(abuf, &abuf));
248	}
249
250	return (error);
251}
252
253/*
254 * Read a TX_WRITE log data block.
255 */
256static int
257zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
258{
259	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
260	const blkptr_t *bp = &lr->lr_blkptr;
261	arc_flags_t aflags = ARC_FLAG_WAIT;
262	arc_buf_t *abuf = NULL;
263	zbookmark_phys_t zb;
264	int error;
265
266	if (BP_IS_HOLE(bp)) {
267		if (wbuf != NULL)
268			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
269		return (0);
270	}
271
272	if (zilog->zl_header->zh_claim_txg == 0)
273		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
274
275	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
276	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
277
278	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
279	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
280
281	if (error == 0) {
282		if (wbuf != NULL)
283			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
284		(void) arc_buf_remove_ref(abuf, &abuf);
285	}
286
287	return (error);
288}
289
290/*
291 * Parse the intent log, and call parse_func for each valid record within.
292 */
293int
294zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
295    zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
296{
297	const zil_header_t *zh = zilog->zl_header;
298	boolean_t claimed = !!zh->zh_claim_txg;
299	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
300	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
301	uint64_t max_blk_seq = 0;
302	uint64_t max_lr_seq = 0;
303	uint64_t blk_count = 0;
304	uint64_t lr_count = 0;
305	blkptr_t blk, next_blk;
306	char *lrbuf, *lrp;
307	int error = 0;
308
309	/*
310	 * Old logs didn't record the maximum zh_claim_lr_seq.
311	 */
312	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
313		claim_lr_seq = UINT64_MAX;
314
315	/*
316	 * Starting at the block pointed to by zh_log we read the log chain.
317	 * For each block in the chain we strongly check that block to
318	 * ensure its validity.  We stop when an invalid block is found.
319	 * For each block pointer in the chain we call parse_blk_func().
320	 * For each record in each valid block we call parse_lr_func().
321	 * If the log has been claimed, stop if we encounter a sequence
322	 * number greater than the highest claimed sequence number.
323	 */
324	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
325	zil_bp_tree_init(zilog);
326
327	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
328		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
329		int reclen;
330		char *end;
331
332		if (blk_seq > claim_blk_seq)
333			break;
334		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
335			break;
336		ASSERT3U(max_blk_seq, <, blk_seq);
337		max_blk_seq = blk_seq;
338		blk_count++;
339
340		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
341			break;
342
343		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
344		if (error != 0)
345			break;
346
347		for (lrp = lrbuf; lrp < end; lrp += reclen) {
348			lr_t *lr = (lr_t *)lrp;
349			reclen = lr->lrc_reclen;
350			ASSERT3U(reclen, >=, sizeof (lr_t));
351			if (lr->lrc_seq > claim_lr_seq)
352				goto done;
353			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
354				goto done;
355			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
356			max_lr_seq = lr->lrc_seq;
357			lr_count++;
358		}
359	}
360done:
361	zilog->zl_parse_error = error;
362	zilog->zl_parse_blk_seq = max_blk_seq;
363	zilog->zl_parse_lr_seq = max_lr_seq;
364	zilog->zl_parse_blk_count = blk_count;
365	zilog->zl_parse_lr_count = lr_count;
366
367	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
368	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
369
370	zil_bp_tree_fini(zilog);
371	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
372
373	return (error);
374}
375
376static int
377zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
378{
379	/*
380	 * Claim log block if not already committed and not already claimed.
381	 * If tx == NULL, just verify that the block is claimable.
382	 */
383	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
384	    zil_bp_tree_add(zilog, bp) != 0)
385		return (0);
386
387	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
388	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
389	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
390}
391
392static int
393zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
394{
395	lr_write_t *lr = (lr_write_t *)lrc;
396	int error;
397
398	if (lrc->lrc_txtype != TX_WRITE)
399		return (0);
400
401	/*
402	 * If the block is not readable, don't claim it.  This can happen
403	 * in normal operation when a log block is written to disk before
404	 * some of the dmu_sync() blocks it points to.  In this case, the
405	 * transaction cannot have been committed to anyone (we would have
406	 * waited for all writes to be stable first), so it is semantically
407	 * correct to declare this the end of the log.
408	 */
409	if (lr->lr_blkptr.blk_birth >= first_txg &&
410	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
411		return (error);
412	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
413}
414
415/* ARGSUSED */
416static int
417zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
418{
419	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
420
421	return (0);
422}
423
424static int
425zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
426{
427	lr_write_t *lr = (lr_write_t *)lrc;
428	blkptr_t *bp = &lr->lr_blkptr;
429
430	/*
431	 * If we previously claimed it, we need to free it.
432	 */
433	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
434	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
435	    !BP_IS_HOLE(bp))
436		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
437
438	return (0);
439}
440
441static lwb_t *
442zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
443{
444	lwb_t *lwb;
445
446	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
447	lwb->lwb_zilog = zilog;
448	lwb->lwb_blk = *bp;
449	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
450	lwb->lwb_max_txg = txg;
451	lwb->lwb_zio = NULL;
452	lwb->lwb_tx = NULL;
453	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
454		lwb->lwb_nused = sizeof (zil_chain_t);
455		lwb->lwb_sz = BP_GET_LSIZE(bp);
456	} else {
457		lwb->lwb_nused = 0;
458		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
459	}
460
461	mutex_enter(&zilog->zl_lock);
462	list_insert_tail(&zilog->zl_lwb_list, lwb);
463	mutex_exit(&zilog->zl_lock);
464
465	return (lwb);
466}
467
468/*
469 * Called when we create in-memory log transactions so that we know
470 * to cleanup the itxs at the end of spa_sync().
471 */
472void
473zilog_dirty(zilog_t *zilog, uint64_t txg)
474{
475	dsl_pool_t *dp = zilog->zl_dmu_pool;
476	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
477
478	if (ds->ds_is_snapshot)
479		panic("dirtying snapshot!");
480
481	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
482		/* up the hold count until we can be written out */
483		dmu_buf_add_ref(ds->ds_dbuf, zilog);
484	}
485}
486
487boolean_t
488zilog_is_dirty(zilog_t *zilog)
489{
490	dsl_pool_t *dp = zilog->zl_dmu_pool;
491
492	for (int t = 0; t < TXG_SIZE; t++) {
493		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
494			return (B_TRUE);
495	}
496	return (B_FALSE);
497}
498
499/*
500 * Create an on-disk intent log.
501 */
502static lwb_t *
503zil_create(zilog_t *zilog)
504{
505	const zil_header_t *zh = zilog->zl_header;
506	lwb_t *lwb = NULL;
507	uint64_t txg = 0;
508	dmu_tx_t *tx = NULL;
509	blkptr_t blk;
510	int error = 0;
511
512	/*
513	 * Wait for any previous destroy to complete.
514	 */
515	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
516
517	ASSERT(zh->zh_claim_txg == 0);
518	ASSERT(zh->zh_replay_seq == 0);
519
520	blk = zh->zh_log;
521
522	/*
523	 * Allocate an initial log block if:
524	 *    - there isn't one already
525	 *    - the existing block is the wrong endianess
526	 */
527	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
528		tx = dmu_tx_create(zilog->zl_os);
529		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
530		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
531		txg = dmu_tx_get_txg(tx);
532
533		if (!BP_IS_HOLE(&blk)) {
534			zio_free_zil(zilog->zl_spa, txg, &blk);
535			BP_ZERO(&blk);
536		}
537
538		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
539		    ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
540
541		if (error == 0)
542			zil_init_log_chain(zilog, &blk);
543	}
544
545	/*
546	 * Allocate a log write buffer (lwb) for the first log block.
547	 */
548	if (error == 0)
549		lwb = zil_alloc_lwb(zilog, &blk, txg);
550
551	/*
552	 * If we just allocated the first log block, commit our transaction
553	 * and wait for zil_sync() to stuff the block poiner into zh_log.
554	 * (zh is part of the MOS, so we cannot modify it in open context.)
555	 */
556	if (tx != NULL) {
557		dmu_tx_commit(tx);
558		txg_wait_synced(zilog->zl_dmu_pool, txg);
559	}
560
561	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
562
563	return (lwb);
564}
565
566/*
567 * In one tx, free all log blocks and clear the log header.
568 * If keep_first is set, then we're replaying a log with no content.
569 * We want to keep the first block, however, so that the first
570 * synchronous transaction doesn't require a txg_wait_synced()
571 * in zil_create().  We don't need to txg_wait_synced() here either
572 * when keep_first is set, because both zil_create() and zil_destroy()
573 * will wait for any in-progress destroys to complete.
574 */
575void
576zil_destroy(zilog_t *zilog, boolean_t keep_first)
577{
578	const zil_header_t *zh = zilog->zl_header;
579	lwb_t *lwb;
580	dmu_tx_t *tx;
581	uint64_t txg;
582
583	/*
584	 * Wait for any previous destroy to complete.
585	 */
586	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
587
588	zilog->zl_old_header = *zh;		/* debugging aid */
589
590	if (BP_IS_HOLE(&zh->zh_log))
591		return;
592
593	tx = dmu_tx_create(zilog->zl_os);
594	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
595	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
596	txg = dmu_tx_get_txg(tx);
597
598	mutex_enter(&zilog->zl_lock);
599
600	ASSERT3U(zilog->zl_destroy_txg, <, txg);
601	zilog->zl_destroy_txg = txg;
602	zilog->zl_keep_first = keep_first;
603
604	if (!list_is_empty(&zilog->zl_lwb_list)) {
605		ASSERT(zh->zh_claim_txg == 0);
606		VERIFY(!keep_first);
607		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
608			list_remove(&zilog->zl_lwb_list, lwb);
609			if (lwb->lwb_buf != NULL)
610				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
611			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
612			kmem_cache_free(zil_lwb_cache, lwb);
613		}
614	} else if (!keep_first) {
615		zil_destroy_sync(zilog, tx);
616	}
617	mutex_exit(&zilog->zl_lock);
618
619	dmu_tx_commit(tx);
620}
621
622void
623zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
624{
625	ASSERT(list_is_empty(&zilog->zl_lwb_list));
626	(void) zil_parse(zilog, zil_free_log_block,
627	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
628}
629
630int
631zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
632{
633	dmu_tx_t *tx = txarg;
634	uint64_t first_txg = dmu_tx_get_txg(tx);
635	zilog_t *zilog;
636	zil_header_t *zh;
637	objset_t *os;
638	int error;
639
640	error = dmu_objset_own_obj(dp, ds->ds_object,
641	    DMU_OST_ANY, B_FALSE, FTAG, &os);
642	if (error != 0) {
643		/*
644		 * EBUSY indicates that the objset is inconsistent, in which
645		 * case it can not have a ZIL.
646		 */
647		if (error != EBUSY) {
648			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
649			    (unsigned long long)ds->ds_object, error);
650		}
651		return (0);
652	}
653
654	zilog = dmu_objset_zil(os);
655	zh = zil_header_in_syncing_context(zilog);
656
657	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
658		if (!BP_IS_HOLE(&zh->zh_log))
659			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
660		BP_ZERO(&zh->zh_log);
661		dsl_dataset_dirty(dmu_objset_ds(os), tx);
662		dmu_objset_disown(os, FTAG);
663		return (0);
664	}
665
666	/*
667	 * Claim all log blocks if we haven't already done so, and remember
668	 * the highest claimed sequence number.  This ensures that if we can
669	 * read only part of the log now (e.g. due to a missing device),
670	 * but we can read the entire log later, we will not try to replay
671	 * or destroy beyond the last block we successfully claimed.
672	 */
673	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
674	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
675		(void) zil_parse(zilog, zil_claim_log_block,
676		    zil_claim_log_record, tx, first_txg);
677		zh->zh_claim_txg = first_txg;
678		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
679		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
680		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
681			zh->zh_flags |= ZIL_REPLAY_NEEDED;
682		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
683		dsl_dataset_dirty(dmu_objset_ds(os), tx);
684	}
685
686	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
687	dmu_objset_disown(os, FTAG);
688	return (0);
689}
690
691/*
692 * Check the log by walking the log chain.
693 * Checksum errors are ok as they indicate the end of the chain.
694 * Any other error (no device or read failure) returns an error.
695 */
696/* ARGSUSED */
697int
698zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
699{
700	zilog_t *zilog;
701	objset_t *os;
702	blkptr_t *bp;
703	int error;
704
705	ASSERT(tx == NULL);
706
707	error = dmu_objset_from_ds(ds, &os);
708	if (error != 0) {
709		cmn_err(CE_WARN, "can't open objset %llu, error %d",
710		    (unsigned long long)ds->ds_object, error);
711		return (0);
712	}
713
714	zilog = dmu_objset_zil(os);
715	bp = (blkptr_t *)&zilog->zl_header->zh_log;
716
717	/*
718	 * Check the first block and determine if it's on a log device
719	 * which may have been removed or faulted prior to loading this
720	 * pool.  If so, there's no point in checking the rest of the log
721	 * as its content should have already been synced to the pool.
722	 */
723	if (!BP_IS_HOLE(bp)) {
724		vdev_t *vd;
725		boolean_t valid = B_TRUE;
726
727		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
728		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
729		if (vd->vdev_islog && vdev_is_dead(vd))
730			valid = vdev_log_state_valid(vd);
731		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
732
733		if (!valid)
734			return (0);
735	}
736
737	/*
738	 * Because tx == NULL, zil_claim_log_block() will not actually claim
739	 * any blocks, but just determine whether it is possible to do so.
740	 * In addition to checking the log chain, zil_claim_log_block()
741	 * will invoke zio_claim() with a done func of spa_claim_notify(),
742	 * which will update spa_max_claim_txg.  See spa_load() for details.
743	 */
744	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
745	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
746
747	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
748}
749
750static int
751zil_vdev_compare(const void *x1, const void *x2)
752{
753	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
754	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
755
756	if (v1 < v2)
757		return (-1);
758	if (v1 > v2)
759		return (1);
760
761	return (0);
762}
763
764void
765zil_add_block(zilog_t *zilog, const blkptr_t *bp)
766{
767	avl_tree_t *t = &zilog->zl_vdev_tree;
768	avl_index_t where;
769	zil_vdev_node_t *zv, zvsearch;
770	int ndvas = BP_GET_NDVAS(bp);
771	int i;
772
773	if (zfs_nocacheflush)
774		return;
775
776	ASSERT(zilog->zl_writer);
777
778	/*
779	 * Even though we're zl_writer, we still need a lock because the
780	 * zl_get_data() callbacks may have dmu_sync() done callbacks
781	 * that will run concurrently.
782	 */
783	mutex_enter(&zilog->zl_vdev_lock);
784	for (i = 0; i < ndvas; i++) {
785		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
786		if (avl_find(t, &zvsearch, &where) == NULL) {
787			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
788			zv->zv_vdev = zvsearch.zv_vdev;
789			avl_insert(t, zv, where);
790		}
791	}
792	mutex_exit(&zilog->zl_vdev_lock);
793}
794
795static void
796zil_flush_vdevs(zilog_t *zilog)
797{
798	spa_t *spa = zilog->zl_spa;
799	avl_tree_t *t = &zilog->zl_vdev_tree;
800	void *cookie = NULL;
801	zil_vdev_node_t *zv;
802	zio_t *zio;
803
804	ASSERT(zilog->zl_writer);
805
806	/*
807	 * We don't need zl_vdev_lock here because we're the zl_writer,
808	 * and all zl_get_data() callbacks are done.
809	 */
810	if (avl_numnodes(t) == 0)
811		return;
812
813	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
814
815	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
816
817	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
818		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
819		if (vd != NULL)
820			zio_flush(zio, vd);
821		kmem_free(zv, sizeof (*zv));
822	}
823
824	/*
825	 * Wait for all the flushes to complete.  Not all devices actually
826	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
827	 */
828	(void) zio_wait(zio);
829
830	spa_config_exit(spa, SCL_STATE, FTAG);
831}
832
833/*
834 * Function called when a log block write completes
835 */
836static void
837zil_lwb_write_done(zio_t *zio)
838{
839	lwb_t *lwb = zio->io_private;
840	zilog_t *zilog = lwb->lwb_zilog;
841	dmu_tx_t *tx = lwb->lwb_tx;
842
843	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
844	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
845	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
846	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
847	ASSERT(!BP_IS_GANG(zio->io_bp));
848	ASSERT(!BP_IS_HOLE(zio->io_bp));
849	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
850
851	/*
852	 * Ensure the lwb buffer pointer is cleared before releasing
853	 * the txg. If we have had an allocation failure and
854	 * the txg is waiting to sync then we want want zil_sync()
855	 * to remove the lwb so that it's not picked up as the next new
856	 * one in zil_commit_writer(). zil_sync() will only remove
857	 * the lwb if lwb_buf is null.
858	 */
859	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
860	mutex_enter(&zilog->zl_lock);
861	lwb->lwb_buf = NULL;
862	lwb->lwb_tx = NULL;
863	mutex_exit(&zilog->zl_lock);
864
865	/*
866	 * Now that we've written this log block, we have a stable pointer
867	 * to the next block in the chain, so it's OK to let the txg in
868	 * which we allocated the next block sync.
869	 */
870	dmu_tx_commit(tx);
871}
872
873/*
874 * Initialize the io for a log block.
875 */
876static void
877zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
878{
879	zbookmark_phys_t zb;
880
881	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
882	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
883	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
884
885	if (zilog->zl_root_zio == NULL) {
886		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
887		    ZIO_FLAG_CANFAIL);
888	}
889	if (lwb->lwb_zio == NULL) {
890		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
891		    0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
892		    zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
893		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
894	}
895}
896
897/*
898 * Define a limited set of intent log block sizes.
899 *
900 * These must be a multiple of 4KB. Note only the amount used (again
901 * aligned to 4KB) actually gets written. However, we can't always just
902 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
903 */
904uint64_t zil_block_buckets[] = {
905    4096,		/* non TX_WRITE */
906    8192+4096,		/* data base */
907    32*1024 + 4096, 	/* NFS writes */
908    UINT64_MAX
909};
910
911/*
912 * Use the slog as long as the logbias is 'latency' and the current commit size
913 * is less than the limit or the total list size is less than 2X the limit.
914 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
915 */
916uint64_t zil_slog_limit = 1024 * 1024;
917#define	USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
918	(((zilog)->zl_cur_used < zil_slog_limit) || \
919	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
920
921/*
922 * Start a log block write and advance to the next log block.
923 * Calls are serialized.
924 */
925static lwb_t *
926zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
927{
928	lwb_t *nlwb = NULL;
929	zil_chain_t *zilc;
930	spa_t *spa = zilog->zl_spa;
931	blkptr_t *bp;
932	dmu_tx_t *tx;
933	uint64_t txg;
934	uint64_t zil_blksz, wsz;
935	int i, error;
936
937	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
938		zilc = (zil_chain_t *)lwb->lwb_buf;
939		bp = &zilc->zc_next_blk;
940	} else {
941		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
942		bp = &zilc->zc_next_blk;
943	}
944
945	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
946
947	/*
948	 * Allocate the next block and save its address in this block
949	 * before writing it in order to establish the log chain.
950	 * Note that if the allocation of nlwb synced before we wrote
951	 * the block that points at it (lwb), we'd leak it if we crashed.
952	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
953	 * We dirty the dataset to ensure that zil_sync() will be called
954	 * to clean up in the event of allocation failure or I/O failure.
955	 */
956	tx = dmu_tx_create(zilog->zl_os);
957	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
958	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
959	txg = dmu_tx_get_txg(tx);
960
961	lwb->lwb_tx = tx;
962
963	/*
964	 * Log blocks are pre-allocated. Here we select the size of the next
965	 * block, based on size used in the last block.
966	 * - first find the smallest bucket that will fit the block from a
967	 *   limited set of block sizes. This is because it's faster to write
968	 *   blocks allocated from the same metaslab as they are adjacent or
969	 *   close.
970	 * - next find the maximum from the new suggested size and an array of
971	 *   previous sizes. This lessens a picket fence effect of wrongly
972	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
973	 *   requests.
974	 *
975	 * Note we only write what is used, but we can't just allocate
976	 * the maximum block size because we can exhaust the available
977	 * pool log space.
978	 */
979	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
980	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
981		continue;
982	zil_blksz = zil_block_buckets[i];
983	if (zil_blksz == UINT64_MAX)
984		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
985	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
986	for (i = 0; i < ZIL_PREV_BLKS; i++)
987		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
988	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
989
990	BP_ZERO(bp);
991	/* pass the old blkptr in order to spread log blocks across devs */
992	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
993	    USE_SLOG(zilog));
994	if (error == 0) {
995		ASSERT3U(bp->blk_birth, ==, txg);
996		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
997		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
998
999		/*
1000		 * Allocate a new log write buffer (lwb).
1001		 */
1002		nlwb = zil_alloc_lwb(zilog, bp, txg);
1003
1004		/* Record the block for later vdev flushing */
1005		zil_add_block(zilog, &lwb->lwb_blk);
1006	}
1007
1008	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1009		/* For Slim ZIL only write what is used. */
1010		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1011		ASSERT3U(wsz, <=, lwb->lwb_sz);
1012		zio_shrink(lwb->lwb_zio, wsz);
1013
1014	} else {
1015		wsz = lwb->lwb_sz;
1016	}
1017
1018	zilc->zc_pad = 0;
1019	zilc->zc_nused = lwb->lwb_nused;
1020	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1021
1022	/*
1023	 * clear unused data for security
1024	 */
1025	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1026
1027	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1028
1029	/*
1030	 * If there was an allocation failure then nlwb will be null which
1031	 * forces a txg_wait_synced().
1032	 */
1033	return (nlwb);
1034}
1035
1036static lwb_t *
1037zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1038{
1039	lr_t *lrc = &itx->itx_lr; /* common log record */
1040	lr_write_t *lrw = (lr_write_t *)lrc;
1041	char *lr_buf;
1042	uint64_t txg = lrc->lrc_txg;
1043	uint64_t reclen = lrc->lrc_reclen;
1044	uint64_t dlen = 0;
1045
1046	if (lwb == NULL)
1047		return (NULL);
1048
1049	ASSERT(lwb->lwb_buf != NULL);
1050	ASSERT(zilog_is_dirty(zilog) ||
1051	    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1052
1053	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1054		dlen = P2ROUNDUP_TYPED(
1055		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1056
1057	zilog->zl_cur_used += (reclen + dlen);
1058
1059	zil_lwb_write_init(zilog, lwb);
1060
1061	/*
1062	 * If this record won't fit in the current log block, start a new one.
1063	 */
1064	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1065		lwb = zil_lwb_write_start(zilog, lwb);
1066		if (lwb == NULL)
1067			return (NULL);
1068		zil_lwb_write_init(zilog, lwb);
1069		ASSERT(LWB_EMPTY(lwb));
1070		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1071			txg_wait_synced(zilog->zl_dmu_pool, txg);
1072			return (lwb);
1073		}
1074	}
1075
1076	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1077	bcopy(lrc, lr_buf, reclen);
1078	lrc = (lr_t *)lr_buf;
1079	lrw = (lr_write_t *)lrc;
1080
1081	/*
1082	 * If it's a write, fetch the data or get its blkptr as appropriate.
1083	 */
1084	if (lrc->lrc_txtype == TX_WRITE) {
1085		if (txg > spa_freeze_txg(zilog->zl_spa))
1086			txg_wait_synced(zilog->zl_dmu_pool, txg);
1087		if (itx->itx_wr_state != WR_COPIED) {
1088			char *dbuf;
1089			int error;
1090
1091			if (dlen) {
1092				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1093				dbuf = lr_buf + reclen;
1094				lrw->lr_common.lrc_reclen += dlen;
1095			} else {
1096				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1097				dbuf = NULL;
1098			}
1099			error = zilog->zl_get_data(
1100			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1101			if (error == EIO) {
1102				txg_wait_synced(zilog->zl_dmu_pool, txg);
1103				return (lwb);
1104			}
1105			if (error != 0) {
1106				ASSERT(error == ENOENT || error == EEXIST ||
1107				    error == EALREADY);
1108				return (lwb);
1109			}
1110		}
1111	}
1112
1113	/*
1114	 * We're actually making an entry, so update lrc_seq to be the
1115	 * log record sequence number.  Note that this is generally not
1116	 * equal to the itx sequence number because not all transactions
1117	 * are synchronous, and sometimes spa_sync() gets there first.
1118	 */
1119	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1120	lwb->lwb_nused += reclen + dlen;
1121	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1122	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1123	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1124
1125	return (lwb);
1126}
1127
1128itx_t *
1129zil_itx_create(uint64_t txtype, size_t lrsize)
1130{
1131	itx_t *itx;
1132
1133	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1134
1135	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1136	itx->itx_lr.lrc_txtype = txtype;
1137	itx->itx_lr.lrc_reclen = lrsize;
1138	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1139	itx->itx_lr.lrc_seq = 0;	/* defensive */
1140	itx->itx_sync = B_TRUE;		/* default is synchronous */
1141
1142	return (itx);
1143}
1144
1145void
1146zil_itx_destroy(itx_t *itx)
1147{
1148	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1149}
1150
1151/*
1152 * Free up the sync and async itxs. The itxs_t has already been detached
1153 * so no locks are needed.
1154 */
1155static void
1156zil_itxg_clean(itxs_t *itxs)
1157{
1158	itx_t *itx;
1159	list_t *list;
1160	avl_tree_t *t;
1161	void *cookie;
1162	itx_async_node_t *ian;
1163
1164	list = &itxs->i_sync_list;
1165	while ((itx = list_head(list)) != NULL) {
1166		list_remove(list, itx);
1167		kmem_free(itx, offsetof(itx_t, itx_lr) +
1168		    itx->itx_lr.lrc_reclen);
1169	}
1170
1171	cookie = NULL;
1172	t = &itxs->i_async_tree;
1173	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1174		list = &ian->ia_list;
1175		while ((itx = list_head(list)) != NULL) {
1176			list_remove(list, itx);
1177			kmem_free(itx, offsetof(itx_t, itx_lr) +
1178			    itx->itx_lr.lrc_reclen);
1179		}
1180		list_destroy(list);
1181		kmem_free(ian, sizeof (itx_async_node_t));
1182	}
1183	avl_destroy(t);
1184
1185	kmem_free(itxs, sizeof (itxs_t));
1186}
1187
1188static int
1189zil_aitx_compare(const void *x1, const void *x2)
1190{
1191	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1192	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1193
1194	if (o1 < o2)
1195		return (-1);
1196	if (o1 > o2)
1197		return (1);
1198
1199	return (0);
1200}
1201
1202/*
1203 * Remove all async itx with the given oid.
1204 */
1205static void
1206zil_remove_async(zilog_t *zilog, uint64_t oid)
1207{
1208	uint64_t otxg, txg;
1209	itx_async_node_t *ian;
1210	avl_tree_t *t;
1211	avl_index_t where;
1212	list_t clean_list;
1213	itx_t *itx;
1214
1215	ASSERT(oid != 0);
1216	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1217
1218	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1219		otxg = ZILTEST_TXG;
1220	else
1221		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1222
1223	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1224		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1225
1226		mutex_enter(&itxg->itxg_lock);
1227		if (itxg->itxg_txg != txg) {
1228			mutex_exit(&itxg->itxg_lock);
1229			continue;
1230		}
1231
1232		/*
1233		 * Locate the object node and append its list.
1234		 */
1235		t = &itxg->itxg_itxs->i_async_tree;
1236		ian = avl_find(t, &oid, &where);
1237		if (ian != NULL)
1238			list_move_tail(&clean_list, &ian->ia_list);
1239		mutex_exit(&itxg->itxg_lock);
1240	}
1241	while ((itx = list_head(&clean_list)) != NULL) {
1242		list_remove(&clean_list, itx);
1243		kmem_free(itx, offsetof(itx_t, itx_lr) +
1244		    itx->itx_lr.lrc_reclen);
1245	}
1246	list_destroy(&clean_list);
1247}
1248
1249void
1250zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1251{
1252	uint64_t txg;
1253	itxg_t *itxg;
1254	itxs_t *itxs, *clean = NULL;
1255
1256	/*
1257	 * Object ids can be re-instantiated in the next txg so
1258	 * remove any async transactions to avoid future leaks.
1259	 * This can happen if a fsync occurs on the re-instantiated
1260	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1261	 * the new file data and flushes a write record for the old object.
1262	 */
1263	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1264		zil_remove_async(zilog, itx->itx_oid);
1265
1266	/*
1267	 * Ensure the data of a renamed file is committed before the rename.
1268	 */
1269	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1270		zil_async_to_sync(zilog, itx->itx_oid);
1271
1272	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1273		txg = ZILTEST_TXG;
1274	else
1275		txg = dmu_tx_get_txg(tx);
1276
1277	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1278	mutex_enter(&itxg->itxg_lock);
1279	itxs = itxg->itxg_itxs;
1280	if (itxg->itxg_txg != txg) {
1281		if (itxs != NULL) {
1282			/*
1283			 * The zil_clean callback hasn't got around to cleaning
1284			 * this itxg. Save the itxs for release below.
1285			 * This should be rare.
1286			 */
1287			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1288			itxg->itxg_sod = 0;
1289			clean = itxg->itxg_itxs;
1290		}
1291		ASSERT(itxg->itxg_sod == 0);
1292		itxg->itxg_txg = txg;
1293		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1294
1295		list_create(&itxs->i_sync_list, sizeof (itx_t),
1296		    offsetof(itx_t, itx_node));
1297		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1298		    sizeof (itx_async_node_t),
1299		    offsetof(itx_async_node_t, ia_node));
1300	}
1301	if (itx->itx_sync) {
1302		list_insert_tail(&itxs->i_sync_list, itx);
1303		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1304		itxg->itxg_sod += itx->itx_sod;
1305	} else {
1306		avl_tree_t *t = &itxs->i_async_tree;
1307		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1308		itx_async_node_t *ian;
1309		avl_index_t where;
1310
1311		ian = avl_find(t, &foid, &where);
1312		if (ian == NULL) {
1313			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1314			list_create(&ian->ia_list, sizeof (itx_t),
1315			    offsetof(itx_t, itx_node));
1316			ian->ia_foid = foid;
1317			avl_insert(t, ian, where);
1318		}
1319		list_insert_tail(&ian->ia_list, itx);
1320	}
1321
1322	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1323	zilog_dirty(zilog, txg);
1324	mutex_exit(&itxg->itxg_lock);
1325
1326	/* Release the old itxs now we've dropped the lock */
1327	if (clean != NULL)
1328		zil_itxg_clean(clean);
1329}
1330
1331/*
1332 * If there are any in-memory intent log transactions which have now been
1333 * synced then start up a taskq to free them. We should only do this after we
1334 * have written out the uberblocks (i.e. txg has been comitted) so that
1335 * don't inadvertently clean out in-memory log records that would be required
1336 * by zil_commit().
1337 */
1338void
1339zil_clean(zilog_t *zilog, uint64_t synced_txg)
1340{
1341	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1342	itxs_t *clean_me;
1343
1344	mutex_enter(&itxg->itxg_lock);
1345	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1346		mutex_exit(&itxg->itxg_lock);
1347		return;
1348	}
1349	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1350	ASSERT(itxg->itxg_txg != 0);
1351	ASSERT(zilog->zl_clean_taskq != NULL);
1352	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1353	itxg->itxg_sod = 0;
1354	clean_me = itxg->itxg_itxs;
1355	itxg->itxg_itxs = NULL;
1356	itxg->itxg_txg = 0;
1357	mutex_exit(&itxg->itxg_lock);
1358	/*
1359	 * Preferably start a task queue to free up the old itxs but
1360	 * if taskq_dispatch can't allocate resources to do that then
1361	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1362	 * created a bad performance problem.
1363	 */
1364	if (taskq_dispatch(zilog->zl_clean_taskq,
1365	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1366		zil_itxg_clean(clean_me);
1367}
1368
1369/*
1370 * Get the list of itxs to commit into zl_itx_commit_list.
1371 */
1372static void
1373zil_get_commit_list(zilog_t *zilog)
1374{
1375	uint64_t otxg, txg;
1376	list_t *commit_list = &zilog->zl_itx_commit_list;
1377	uint64_t push_sod = 0;
1378
1379	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1380		otxg = ZILTEST_TXG;
1381	else
1382		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1383
1384	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1385		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1386
1387		mutex_enter(&itxg->itxg_lock);
1388		if (itxg->itxg_txg != txg) {
1389			mutex_exit(&itxg->itxg_lock);
1390			continue;
1391		}
1392
1393		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1394		push_sod += itxg->itxg_sod;
1395		itxg->itxg_sod = 0;
1396
1397		mutex_exit(&itxg->itxg_lock);
1398	}
1399	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1400}
1401
1402/*
1403 * Move the async itxs for a specified object to commit into sync lists.
1404 */
1405static void
1406zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1407{
1408	uint64_t otxg, txg;
1409	itx_async_node_t *ian;
1410	avl_tree_t *t;
1411	avl_index_t where;
1412
1413	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1414		otxg = ZILTEST_TXG;
1415	else
1416		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1417
1418	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1419		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1420
1421		mutex_enter(&itxg->itxg_lock);
1422		if (itxg->itxg_txg != txg) {
1423			mutex_exit(&itxg->itxg_lock);
1424			continue;
1425		}
1426
1427		/*
1428		 * If a foid is specified then find that node and append its
1429		 * list. Otherwise walk the tree appending all the lists
1430		 * to the sync list. We add to the end rather than the
1431		 * beginning to ensure the create has happened.
1432		 */
1433		t = &itxg->itxg_itxs->i_async_tree;
1434		if (foid != 0) {
1435			ian = avl_find(t, &foid, &where);
1436			if (ian != NULL) {
1437				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1438				    &ian->ia_list);
1439			}
1440		} else {
1441			void *cookie = NULL;
1442
1443			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1444				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1445				    &ian->ia_list);
1446				list_destroy(&ian->ia_list);
1447				kmem_free(ian, sizeof (itx_async_node_t));
1448			}
1449		}
1450		mutex_exit(&itxg->itxg_lock);
1451	}
1452}
1453
1454static void
1455zil_commit_writer(zilog_t *zilog)
1456{
1457	uint64_t txg;
1458	itx_t *itx;
1459	lwb_t *lwb;
1460	spa_t *spa = zilog->zl_spa;
1461	int error = 0;
1462
1463	ASSERT(zilog->zl_root_zio == NULL);
1464
1465	mutex_exit(&zilog->zl_lock);
1466
1467	zil_get_commit_list(zilog);
1468
1469	/*
1470	 * Return if there's nothing to commit before we dirty the fs by
1471	 * calling zil_create().
1472	 */
1473	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1474		mutex_enter(&zilog->zl_lock);
1475		return;
1476	}
1477
1478	if (zilog->zl_suspend) {
1479		lwb = NULL;
1480	} else {
1481		lwb = list_tail(&zilog->zl_lwb_list);
1482		if (lwb == NULL)
1483			lwb = zil_create(zilog);
1484	}
1485
1486	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1487	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1488		txg = itx->itx_lr.lrc_txg;
1489		ASSERT(txg);
1490
1491		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1492			lwb = zil_lwb_commit(zilog, itx, lwb);
1493		list_remove(&zilog->zl_itx_commit_list, itx);
1494		kmem_free(itx, offsetof(itx_t, itx_lr)
1495		    + itx->itx_lr.lrc_reclen);
1496	}
1497	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1498
1499	/* write the last block out */
1500	if (lwb != NULL && lwb->lwb_zio != NULL)
1501		lwb = zil_lwb_write_start(zilog, lwb);
1502
1503	zilog->zl_cur_used = 0;
1504
1505	/*
1506	 * Wait if necessary for the log blocks to be on stable storage.
1507	 */
1508	if (zilog->zl_root_zio) {
1509		error = zio_wait(zilog->zl_root_zio);
1510		zilog->zl_root_zio = NULL;
1511		zil_flush_vdevs(zilog);
1512	}
1513
1514	if (error || lwb == NULL)
1515		txg_wait_synced(zilog->zl_dmu_pool, 0);
1516
1517	mutex_enter(&zilog->zl_lock);
1518
1519	/*
1520	 * Remember the highest committed log sequence number for ztest.
1521	 * We only update this value when all the log writes succeeded,
1522	 * because ztest wants to ASSERT that it got the whole log chain.
1523	 */
1524	if (error == 0 && lwb != NULL)
1525		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1526}
1527
1528/*
1529 * Commit zfs transactions to stable storage.
1530 * If foid is 0 push out all transactions, otherwise push only those
1531 * for that object or might reference that object.
1532 *
1533 * itxs are committed in batches. In a heavily stressed zil there will be
1534 * a commit writer thread who is writing out a bunch of itxs to the log
1535 * for a set of committing threads (cthreads) in the same batch as the writer.
1536 * Those cthreads are all waiting on the same cv for that batch.
1537 *
1538 * There will also be a different and growing batch of threads that are
1539 * waiting to commit (qthreads). When the committing batch completes
1540 * a transition occurs such that the cthreads exit and the qthreads become
1541 * cthreads. One of the new cthreads becomes the writer thread for the
1542 * batch. Any new threads arriving become new qthreads.
1543 *
1544 * Only 2 condition variables are needed and there's no transition
1545 * between the two cvs needed. They just flip-flop between qthreads
1546 * and cthreads.
1547 *
1548 * Using this scheme we can efficiently wakeup up only those threads
1549 * that have been committed.
1550 */
1551void
1552zil_commit(zilog_t *zilog, uint64_t foid)
1553{
1554	uint64_t mybatch;
1555
1556	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1557		return;
1558
1559	/* move the async itxs for the foid to the sync queues */
1560	zil_async_to_sync(zilog, foid);
1561
1562	mutex_enter(&zilog->zl_lock);
1563	mybatch = zilog->zl_next_batch;
1564	while (zilog->zl_writer) {
1565		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1566		if (mybatch <= zilog->zl_com_batch) {
1567			mutex_exit(&zilog->zl_lock);
1568			return;
1569		}
1570	}
1571
1572	zilog->zl_next_batch++;
1573	zilog->zl_writer = B_TRUE;
1574	zil_commit_writer(zilog);
1575	zilog->zl_com_batch = mybatch;
1576	zilog->zl_writer = B_FALSE;
1577	mutex_exit(&zilog->zl_lock);
1578
1579	/* wake up one thread to become the next writer */
1580	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1581
1582	/* wake up all threads waiting for this batch to be committed */
1583	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1584}
1585
1586/*
1587 * Called in syncing context to free committed log blocks and update log header.
1588 */
1589void
1590zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1591{
1592	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1593	uint64_t txg = dmu_tx_get_txg(tx);
1594	spa_t *spa = zilog->zl_spa;
1595	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1596	lwb_t *lwb;
1597
1598	/*
1599	 * We don't zero out zl_destroy_txg, so make sure we don't try
1600	 * to destroy it twice.
1601	 */
1602	if (spa_sync_pass(spa) != 1)
1603		return;
1604
1605	mutex_enter(&zilog->zl_lock);
1606
1607	ASSERT(zilog->zl_stop_sync == 0);
1608
1609	if (*replayed_seq != 0) {
1610		ASSERT(zh->zh_replay_seq < *replayed_seq);
1611		zh->zh_replay_seq = *replayed_seq;
1612		*replayed_seq = 0;
1613	}
1614
1615	if (zilog->zl_destroy_txg == txg) {
1616		blkptr_t blk = zh->zh_log;
1617
1618		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1619
1620		bzero(zh, sizeof (zil_header_t));
1621		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1622
1623		if (zilog->zl_keep_first) {
1624			/*
1625			 * If this block was part of log chain that couldn't
1626			 * be claimed because a device was missing during
1627			 * zil_claim(), but that device later returns,
1628			 * then this block could erroneously appear valid.
1629			 * To guard against this, assign a new GUID to the new
1630			 * log chain so it doesn't matter what blk points to.
1631			 */
1632			zil_init_log_chain(zilog, &blk);
1633			zh->zh_log = blk;
1634		}
1635	}
1636
1637	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1638		zh->zh_log = lwb->lwb_blk;
1639		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1640			break;
1641		list_remove(&zilog->zl_lwb_list, lwb);
1642		zio_free_zil(spa, txg, &lwb->lwb_blk);
1643		kmem_cache_free(zil_lwb_cache, lwb);
1644
1645		/*
1646		 * If we don't have anything left in the lwb list then
1647		 * we've had an allocation failure and we need to zero
1648		 * out the zil_header blkptr so that we don't end
1649		 * up freeing the same block twice.
1650		 */
1651		if (list_head(&zilog->zl_lwb_list) == NULL)
1652			BP_ZERO(&zh->zh_log);
1653	}
1654	mutex_exit(&zilog->zl_lock);
1655}
1656
1657void
1658zil_init(void)
1659{
1660	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1661	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1662}
1663
1664void
1665zil_fini(void)
1666{
1667	kmem_cache_destroy(zil_lwb_cache);
1668}
1669
1670void
1671zil_set_sync(zilog_t *zilog, uint64_t sync)
1672{
1673	zilog->zl_sync = sync;
1674}
1675
1676void
1677zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1678{
1679	zilog->zl_logbias = logbias;
1680}
1681
1682zilog_t *
1683zil_alloc(objset_t *os, zil_header_t *zh_phys)
1684{
1685	zilog_t *zilog;
1686
1687	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1688
1689	zilog->zl_header = zh_phys;
1690	zilog->zl_os = os;
1691	zilog->zl_spa = dmu_objset_spa(os);
1692	zilog->zl_dmu_pool = dmu_objset_pool(os);
1693	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1694	zilog->zl_logbias = dmu_objset_logbias(os);
1695	zilog->zl_sync = dmu_objset_syncprop(os);
1696	zilog->zl_next_batch = 1;
1697
1698	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1699
1700	for (int i = 0; i < TXG_SIZE; i++) {
1701		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1702		    MUTEX_DEFAULT, NULL);
1703	}
1704
1705	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1706	    offsetof(lwb_t, lwb_node));
1707
1708	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1709	    offsetof(itx_t, itx_node));
1710
1711	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1712
1713	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1714	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1715
1716	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1717	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1718	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1719	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1720
1721	return (zilog);
1722}
1723
1724void
1725zil_free(zilog_t *zilog)
1726{
1727	zilog->zl_stop_sync = 1;
1728
1729	ASSERT0(zilog->zl_suspend);
1730	ASSERT0(zilog->zl_suspending);
1731
1732	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1733	list_destroy(&zilog->zl_lwb_list);
1734
1735	avl_destroy(&zilog->zl_vdev_tree);
1736	mutex_destroy(&zilog->zl_vdev_lock);
1737
1738	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1739	list_destroy(&zilog->zl_itx_commit_list);
1740
1741	for (int i = 0; i < TXG_SIZE; i++) {
1742		/*
1743		 * It's possible for an itx to be generated that doesn't dirty
1744		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1745		 * callback to remove the entry. We remove those here.
1746		 *
1747		 * Also free up the ziltest itxs.
1748		 */
1749		if (zilog->zl_itxg[i].itxg_itxs)
1750			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1751		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1752	}
1753
1754	mutex_destroy(&zilog->zl_lock);
1755
1756	cv_destroy(&zilog->zl_cv_writer);
1757	cv_destroy(&zilog->zl_cv_suspend);
1758	cv_destroy(&zilog->zl_cv_batch[0]);
1759	cv_destroy(&zilog->zl_cv_batch[1]);
1760
1761	kmem_free(zilog, sizeof (zilog_t));
1762}
1763
1764/*
1765 * Open an intent log.
1766 */
1767zilog_t *
1768zil_open(objset_t *os, zil_get_data_t *get_data)
1769{
1770	zilog_t *zilog = dmu_objset_zil(os);
1771
1772	ASSERT(zilog->zl_clean_taskq == NULL);
1773	ASSERT(zilog->zl_get_data == NULL);
1774	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1775
1776	zilog->zl_get_data = get_data;
1777	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1778	    2, 2, TASKQ_PREPOPULATE);
1779
1780	return (zilog);
1781}
1782
1783/*
1784 * Close an intent log.
1785 */
1786void
1787zil_close(zilog_t *zilog)
1788{
1789	lwb_t *lwb;
1790	uint64_t txg = 0;
1791
1792	zil_commit(zilog, 0); /* commit all itx */
1793
1794	/*
1795	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1796	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1797	 * callbacks have occurred that may clean the zil.  Only then can we
1798	 * destroy the zl_clean_taskq.
1799	 */
1800	mutex_enter(&zilog->zl_lock);
1801	lwb = list_tail(&zilog->zl_lwb_list);
1802	if (lwb != NULL)
1803		txg = lwb->lwb_max_txg;
1804	mutex_exit(&zilog->zl_lock);
1805	if (txg)
1806		txg_wait_synced(zilog->zl_dmu_pool, txg);
1807	ASSERT(!zilog_is_dirty(zilog));
1808
1809	taskq_destroy(zilog->zl_clean_taskq);
1810	zilog->zl_clean_taskq = NULL;
1811	zilog->zl_get_data = NULL;
1812
1813	/*
1814	 * We should have only one LWB left on the list; remove it now.
1815	 */
1816	mutex_enter(&zilog->zl_lock);
1817	lwb = list_head(&zilog->zl_lwb_list);
1818	if (lwb != NULL) {
1819		ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1820		list_remove(&zilog->zl_lwb_list, lwb);
1821		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1822		kmem_cache_free(zil_lwb_cache, lwb);
1823	}
1824	mutex_exit(&zilog->zl_lock);
1825}
1826
1827static char *suspend_tag = "zil suspending";
1828
1829/*
1830 * Suspend an intent log.  While in suspended mode, we still honor
1831 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1832 * On old version pools, we suspend the log briefly when taking a
1833 * snapshot so that it will have an empty intent log.
1834 *
1835 * Long holds are not really intended to be used the way we do here --
1836 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
1837 * could fail.  Therefore we take pains to only put a long hold if it is
1838 * actually necessary.  Fortunately, it will only be necessary if the
1839 * objset is currently mounted (or the ZVOL equivalent).  In that case it
1840 * will already have a long hold, so we are not really making things any worse.
1841 *
1842 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1843 * zvol_state_t), and use their mechanism to prevent their hold from being
1844 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
1845 * very little gain.
1846 *
1847 * if cookiep == NULL, this does both the suspend & resume.
1848 * Otherwise, it returns with the dataset "long held", and the cookie
1849 * should be passed into zil_resume().
1850 */
1851int
1852zil_suspend(const char *osname, void **cookiep)
1853{
1854	objset_t *os;
1855	zilog_t *zilog;
1856	const zil_header_t *zh;
1857	int error;
1858
1859	error = dmu_objset_hold(osname, suspend_tag, &os);
1860	if (error != 0)
1861		return (error);
1862	zilog = dmu_objset_zil(os);
1863
1864	mutex_enter(&zilog->zl_lock);
1865	zh = zilog->zl_header;
1866
1867	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1868		mutex_exit(&zilog->zl_lock);
1869		dmu_objset_rele(os, suspend_tag);
1870		return (SET_ERROR(EBUSY));
1871	}
1872
1873	/*
1874	 * Don't put a long hold in the cases where we can avoid it.  This
1875	 * is when there is no cookie so we are doing a suspend & resume
1876	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1877	 * for the suspend because it's already suspended, or there's no ZIL.
1878	 */
1879	if (cookiep == NULL && !zilog->zl_suspending &&
1880	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1881		mutex_exit(&zilog->zl_lock);
1882		dmu_objset_rele(os, suspend_tag);
1883		return (0);
1884	}
1885
1886	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1887	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1888
1889	zilog->zl_suspend++;
1890
1891	if (zilog->zl_suspend > 1) {
1892		/*
1893		 * Someone else is already suspending it.
1894		 * Just wait for them to finish.
1895		 */
1896
1897		while (zilog->zl_suspending)
1898			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1899		mutex_exit(&zilog->zl_lock);
1900
1901		if (cookiep == NULL)
1902			zil_resume(os);
1903		else
1904			*cookiep = os;
1905		return (0);
1906	}
1907
1908	/*
1909	 * If there is no pointer to an on-disk block, this ZIL must not
1910	 * be active (e.g. filesystem not mounted), so there's nothing
1911	 * to clean up.
1912	 */
1913	if (BP_IS_HOLE(&zh->zh_log)) {
1914		ASSERT(cookiep != NULL); /* fast path already handled */
1915
1916		*cookiep = os;
1917		mutex_exit(&zilog->zl_lock);
1918		return (0);
1919	}
1920
1921	zilog->zl_suspending = B_TRUE;
1922	mutex_exit(&zilog->zl_lock);
1923
1924	zil_commit(zilog, 0);
1925
1926	zil_destroy(zilog, B_FALSE);
1927
1928	mutex_enter(&zilog->zl_lock);
1929	zilog->zl_suspending = B_FALSE;
1930	cv_broadcast(&zilog->zl_cv_suspend);
1931	mutex_exit(&zilog->zl_lock);
1932
1933	if (cookiep == NULL)
1934		zil_resume(os);
1935	else
1936		*cookiep = os;
1937	return (0);
1938}
1939
1940void
1941zil_resume(void *cookie)
1942{
1943	objset_t *os = cookie;
1944	zilog_t *zilog = dmu_objset_zil(os);
1945
1946	mutex_enter(&zilog->zl_lock);
1947	ASSERT(zilog->zl_suspend != 0);
1948	zilog->zl_suspend--;
1949	mutex_exit(&zilog->zl_lock);
1950	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1951	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1952}
1953
1954typedef struct zil_replay_arg {
1955	zil_replay_func_t **zr_replay;
1956	void		*zr_arg;
1957	boolean_t	zr_byteswap;
1958	char		*zr_lr;
1959} zil_replay_arg_t;
1960
1961static int
1962zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1963{
1964	char name[MAXNAMELEN];
1965
1966	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
1967
1968	dmu_objset_name(zilog->zl_os, name);
1969
1970	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1971	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1972	    (u_longlong_t)lr->lrc_seq,
1973	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1974	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1975
1976	return (error);
1977}
1978
1979static int
1980zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1981{
1982	zil_replay_arg_t *zr = zra;
1983	const zil_header_t *zh = zilog->zl_header;
1984	uint64_t reclen = lr->lrc_reclen;
1985	uint64_t txtype = lr->lrc_txtype;
1986	int error = 0;
1987
1988	zilog->zl_replaying_seq = lr->lrc_seq;
1989
1990	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1991		return (0);
1992
1993	if (lr->lrc_txg < claim_txg)		/* already committed */
1994		return (0);
1995
1996	/* Strip case-insensitive bit, still present in log record */
1997	txtype &= ~TX_CI;
1998
1999	if (txtype == 0 || txtype >= TX_MAX_TYPE)
2000		return (zil_replay_error(zilog, lr, EINVAL));
2001
2002	/*
2003	 * If this record type can be logged out of order, the object
2004	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
2005	 */
2006	if (TX_OOO(txtype)) {
2007		error = dmu_object_info(zilog->zl_os,
2008		    ((lr_ooo_t *)lr)->lr_foid, NULL);
2009		if (error == ENOENT || error == EEXIST)
2010			return (0);
2011	}
2012
2013	/*
2014	 * Make a copy of the data so we can revise and extend it.
2015	 */
2016	bcopy(lr, zr->zr_lr, reclen);
2017
2018	/*
2019	 * If this is a TX_WRITE with a blkptr, suck in the data.
2020	 */
2021	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2022		error = zil_read_log_data(zilog, (lr_write_t *)lr,
2023		    zr->zr_lr + reclen);
2024		if (error != 0)
2025			return (zil_replay_error(zilog, lr, error));
2026	}
2027
2028	/*
2029	 * The log block containing this lr may have been byteswapped
2030	 * so that we can easily examine common fields like lrc_txtype.
2031	 * However, the log is a mix of different record types, and only the
2032	 * replay vectors know how to byteswap their records.  Therefore, if
2033	 * the lr was byteswapped, undo it before invoking the replay vector.
2034	 */
2035	if (zr->zr_byteswap)
2036		byteswap_uint64_array(zr->zr_lr, reclen);
2037
2038	/*
2039	 * We must now do two things atomically: replay this log record,
2040	 * and update the log header sequence number to reflect the fact that
2041	 * we did so. At the end of each replay function the sequence number
2042	 * is updated if we are in replay mode.
2043	 */
2044	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2045	if (error != 0) {
2046		/*
2047		 * The DMU's dnode layer doesn't see removes until the txg
2048		 * commits, so a subsequent claim can spuriously fail with
2049		 * EEXIST. So if we receive any error we try syncing out
2050		 * any removes then retry the transaction.  Note that we
2051		 * specify B_FALSE for byteswap now, so we don't do it twice.
2052		 */
2053		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2054		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2055		if (error != 0)
2056			return (zil_replay_error(zilog, lr, error));
2057	}
2058	return (0);
2059}
2060
2061/* ARGSUSED */
2062static int
2063zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2064{
2065	zilog->zl_replay_blks++;
2066
2067	return (0);
2068}
2069
2070/*
2071 * If this dataset has a non-empty intent log, replay it and destroy it.
2072 */
2073void
2074zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2075{
2076	zilog_t *zilog = dmu_objset_zil(os);
2077	const zil_header_t *zh = zilog->zl_header;
2078	zil_replay_arg_t zr;
2079
2080	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2081		zil_destroy(zilog, B_TRUE);
2082		return;
2083	}
2084
2085	zr.zr_replay = replay_func;
2086	zr.zr_arg = arg;
2087	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2088	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2089
2090	/*
2091	 * Wait for in-progress removes to sync before starting replay.
2092	 */
2093	txg_wait_synced(zilog->zl_dmu_pool, 0);
2094
2095	zilog->zl_replay = B_TRUE;
2096	zilog->zl_replay_time = ddi_get_lbolt();
2097	ASSERT(zilog->zl_replay_blks == 0);
2098	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2099	    zh->zh_claim_txg);
2100	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2101
2102	zil_destroy(zilog, B_FALSE);
2103	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2104	zilog->zl_replay = B_FALSE;
2105}
2106
2107boolean_t
2108zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2109{
2110	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2111		return (B_TRUE);
2112
2113	if (zilog->zl_replay) {
2114		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2115		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2116		    zilog->zl_replaying_seq;
2117		return (B_TRUE);
2118	}
2119
2120	return (B_FALSE);
2121}
2122
2123/* ARGSUSED */
2124int
2125zil_vdev_offline(const char *osname, void *arg)
2126{
2127	int error;
2128
2129	error = zil_suspend(osname, NULL);
2130	if (error != 0)
2131		return (SET_ERROR(EEXIST));
2132	return (0);
2133}
2134