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 */
25
26/* Portions Copyright 2010 Robert Milkowski */
27
28#include <sys/zfs_context.h>
29#include <sys/spa.h>
30#include <sys/dmu.h>
31#include <sys/zap.h>
32#include <sys/arc.h>
33#include <sys/stat.h>
34#include <sys/resource.h>
35#include <sys/zil.h>
36#include <sys/zil_impl.h>
37#include <sys/dsl_dataset.h>
38#include <sys/vdev_impl.h>
39#include <sys/dmu_tx.h>
40#include <sys/dsl_pool.h>
41
42/*
43 * The zfs intent log (ZIL) saves transaction records of system calls
44 * that change the file system in memory with enough information
45 * to be able to replay them. These are stored in memory until
46 * either the DMU transaction group (txg) commits them to the stable pool
47 * and they can be discarded, or they are flushed to the stable log
48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49 * requirement. In the event of a panic or power fail then those log
50 * records (transactions) are replayed.
51 *
52 * There is one ZIL per file system. Its on-disk (pool) format consists
53 * of 3 parts:
54 *
55 * 	- ZIL header
56 * 	- ZIL blocks
57 * 	- ZIL records
58 *
59 * A log record holds a system call transaction. Log blocks can
60 * hold many log records and the blocks are chained together.
61 * Each ZIL block contains a block pointer (blkptr_t) to the next
62 * ZIL block in the chain. The ZIL header points to the first
63 * block in the chain. Note there is not a fixed place in the pool
64 * to hold blocks. They are dynamically allocated and freed as
65 * needed from the blocks available. Figure X shows the ZIL structure:
66 */
67
68/*
69 * Disable intent logging replay.  This global ZIL switch affects all pools.
70 */
71int zil_replay_disable = 0;
72
73/*
74 * Tunable parameter for debugging or performance analysis.  Setting
75 * zfs_nocacheflush will cause corruption on power loss if a volatile
76 * out-of-order write cache is enabled.
77 */
78boolean_t zfs_nocacheflush = B_FALSE;
79
80static kmem_cache_t *zil_lwb_cache;
81
82static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
83
84#define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
85    sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
86
87
88/*
89 * ziltest is by and large an ugly hack, but very useful in
90 * checking replay without tedious work.
91 * When running ziltest we want to keep all itx's and so maintain
92 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
93 * We subtract TXG_CONCURRENT_STATES to allow for common code.
94 */
95#define	ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
96
97static int
98zil_bp_compare(const void *x1, const void *x2)
99{
100	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
101	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
102
103	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
104		return (-1);
105	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
106		return (1);
107
108	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
109		return (-1);
110	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
111		return (1);
112
113	return (0);
114}
115
116static void
117zil_bp_tree_init(zilog_t *zilog)
118{
119	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
120	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
121}
122
123static void
124zil_bp_tree_fini(zilog_t *zilog)
125{
126	avl_tree_t *t = &zilog->zl_bp_tree;
127	zil_bp_node_t *zn;
128	void *cookie = NULL;
129
130	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
131		kmem_free(zn, sizeof (zil_bp_node_t));
132
133	avl_destroy(t);
134}
135
136int
137zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
138{
139	avl_tree_t *t = &zilog->zl_bp_tree;
140	const dva_t *dva;
141	zil_bp_node_t *zn;
142	avl_index_t where;
143
144	if (BP_IS_EMBEDDED(bp))
145		return (0);
146
147	dva = BP_IDENTITY(bp);
148
149	if (avl_find(t, dva, &where) != NULL)
150		return (SET_ERROR(EEXIST));
151
152	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
153	zn->zn_dva = *dva;
154	avl_insert(t, zn, where);
155
156	return (0);
157}
158
159static zil_header_t *
160zil_header_in_syncing_context(zilog_t *zilog)
161{
162	return ((zil_header_t *)zilog->zl_header);
163}
164
165static void
166zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
167{
168	zio_cksum_t *zc = &bp->blk_cksum;
169
170	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
171	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
172	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
173	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
174}
175
176/*
177 * Read a log block and make sure it's valid.
178 */
179static int
180zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
181    char **end)
182{
183	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
184	arc_flags_t aflags = ARC_FLAG_WAIT;
185	arc_buf_t *abuf = NULL;
186	zbookmark_phys_t zb;
187	int error;
188
189	if (zilog->zl_header->zh_claim_txg == 0)
190		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
191
192	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
193		zio_flags |= ZIO_FLAG_SPECULATIVE;
194
195	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
196	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
197
198	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
199	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
200
201	if (error == 0) {
202		zio_cksum_t cksum = bp->blk_cksum;
203
204		/*
205		 * Validate the checksummed log block.
206		 *
207		 * Sequence numbers should be... sequential.  The checksum
208		 * verifier for the next block should be bp's checksum plus 1.
209		 *
210		 * Also check the log chain linkage and size used.
211		 */
212		cksum.zc_word[ZIL_ZC_SEQ]++;
213
214		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
215			zil_chain_t *zilc = abuf->b_data;
216			char *lr = (char *)(zilc + 1);
217			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
218
219			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
220			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
221				error = SET_ERROR(ECKSUM);
222			} else {
223				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
224				bcopy(lr, dst, len);
225				*end = (char *)dst + len;
226				*nbp = zilc->zc_next_blk;
227			}
228		} else {
229			char *lr = abuf->b_data;
230			uint64_t size = BP_GET_LSIZE(bp);
231			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
232
233			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
234			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
235			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
236				error = SET_ERROR(ECKSUM);
237			} else {
238				ASSERT3U(zilc->zc_nused, <=,
239				    SPA_OLD_MAXBLOCKSIZE);
240				bcopy(lr, dst, zilc->zc_nused);
241				*end = (char *)dst + zilc->zc_nused;
242				*nbp = zilc->zc_next_blk;
243			}
244		}
245
246		VERIFY(arc_buf_remove_ref(abuf, &abuf));
247	}
248
249	return (error);
250}
251
252/*
253 * Read a TX_WRITE log data block.
254 */
255static int
256zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
257{
258	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
259	const blkptr_t *bp = &lr->lr_blkptr;
260	arc_flags_t aflags = ARC_FLAG_WAIT;
261	arc_buf_t *abuf = NULL;
262	zbookmark_phys_t zb;
263	int error;
264
265	if (BP_IS_HOLE(bp)) {
266		if (wbuf != NULL)
267			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
268		return (0);
269	}
270
271	if (zilog->zl_header->zh_claim_txg == 0)
272		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
273
274	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
275	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
276
277	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
278	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
279
280	if (error == 0) {
281		if (wbuf != NULL)
282			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
283		(void) arc_buf_remove_ref(abuf, &abuf);
284	}
285
286	return (error);
287}
288
289/*
290 * Parse the intent log, and call parse_func for each valid record within.
291 */
292int
293zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
294    zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
295{
296	const zil_header_t *zh = zilog->zl_header;
297	boolean_t claimed = !!zh->zh_claim_txg;
298	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
299	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
300	uint64_t max_blk_seq = 0;
301	uint64_t max_lr_seq = 0;
302	uint64_t blk_count = 0;
303	uint64_t lr_count = 0;
304	blkptr_t blk, next_blk;
305	char *lrbuf, *lrp;
306	int error = 0;
307
308	/*
309	 * Old logs didn't record the maximum zh_claim_lr_seq.
310	 */
311	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
312		claim_lr_seq = UINT64_MAX;
313
314	/*
315	 * Starting at the block pointed to by zh_log we read the log chain.
316	 * For each block in the chain we strongly check that block to
317	 * ensure its validity.  We stop when an invalid block is found.
318	 * For each block pointer in the chain we call parse_blk_func().
319	 * For each record in each valid block we call parse_lr_func().
320	 * If the log has been claimed, stop if we encounter a sequence
321	 * number greater than the highest claimed sequence number.
322	 */
323	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
324	zil_bp_tree_init(zilog);
325
326	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
327		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
328		int reclen;
329		char *end;
330
331		if (blk_seq > claim_blk_seq)
332			break;
333		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
334			break;
335		ASSERT3U(max_blk_seq, <, blk_seq);
336		max_blk_seq = blk_seq;
337		blk_count++;
338
339		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
340			break;
341
342		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
343		if (error != 0)
344			break;
345
346		for (lrp = lrbuf; lrp < end; lrp += reclen) {
347			lr_t *lr = (lr_t *)lrp;
348			reclen = lr->lrc_reclen;
349			ASSERT3U(reclen, >=, sizeof (lr_t));
350			if (lr->lrc_seq > claim_lr_seq)
351				goto done;
352			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
353				goto done;
354			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
355			max_lr_seq = lr->lrc_seq;
356			lr_count++;
357		}
358	}
359done:
360	zilog->zl_parse_error = error;
361	zilog->zl_parse_blk_seq = max_blk_seq;
362	zilog->zl_parse_lr_seq = max_lr_seq;
363	zilog->zl_parse_blk_count = blk_count;
364	zilog->zl_parse_lr_count = lr_count;
365
366	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
367	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
368
369	zil_bp_tree_fini(zilog);
370	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
371
372	return (error);
373}
374
375static int
376zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
377{
378	/*
379	 * Claim log block if not already committed and not already claimed.
380	 * If tx == NULL, just verify that the block is claimable.
381	 */
382	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
383	    zil_bp_tree_add(zilog, bp) != 0)
384		return (0);
385
386	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
387	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
388	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
389}
390
391static int
392zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
393{
394	lr_write_t *lr = (lr_write_t *)lrc;
395	int error;
396
397	if (lrc->lrc_txtype != TX_WRITE)
398		return (0);
399
400	/*
401	 * If the block is not readable, don't claim it.  This can happen
402	 * in normal operation when a log block is written to disk before
403	 * some of the dmu_sync() blocks it points to.  In this case, the
404	 * transaction cannot have been committed to anyone (we would have
405	 * waited for all writes to be stable first), so it is semantically
406	 * correct to declare this the end of the log.
407	 */
408	if (lr->lr_blkptr.blk_birth >= first_txg &&
409	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
410		return (error);
411	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
412}
413
414/* ARGSUSED */
415static int
416zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
417{
418	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
419
420	return (0);
421}
422
423static int
424zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
425{
426	lr_write_t *lr = (lr_write_t *)lrc;
427	blkptr_t *bp = &lr->lr_blkptr;
428
429	/*
430	 * If we previously claimed it, we need to free it.
431	 */
432	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
433	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
434	    !BP_IS_HOLE(bp))
435		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
436
437	return (0);
438}
439
440static lwb_t *
441zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
442{
443	lwb_t *lwb;
444
445	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
446	lwb->lwb_zilog = zilog;
447	lwb->lwb_blk = *bp;
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
486boolean_t
487zilog_is_dirty(zilog_t *zilog)
488{
489	dsl_pool_t *dp = zilog->zl_dmu_pool;
490
491	for (int t = 0; t < TXG_SIZE; t++) {
492		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
493			return (B_TRUE);
494	}
495	return (B_FALSE);
496}
497
498/*
499 * Create an on-disk intent log.
500 */
501static lwb_t *
502zil_create(zilog_t *zilog)
503{
504	const zil_header_t *zh = zilog->zl_header;
505	lwb_t *lwb = NULL;
506	uint64_t txg = 0;
507	dmu_tx_t *tx = NULL;
508	blkptr_t blk;
509	int error = 0;
510
511	/*
512	 * Wait for any previous destroy to complete.
513	 */
514	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
515
516	ASSERT(zh->zh_claim_txg == 0);
517	ASSERT(zh->zh_replay_seq == 0);
518
519	blk = zh->zh_log;
520
521	/*
522	 * Allocate an initial log block if:
523	 *    - there isn't one already
524	 *    - the existing block is the wrong endianess
525	 */
526	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
527		tx = dmu_tx_create(zilog->zl_os);
528		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
529		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
530		txg = dmu_tx_get_txg(tx);
531
532		if (!BP_IS_HOLE(&blk)) {
533			zio_free_zil(zilog->zl_spa, txg, &blk);
534			BP_ZERO(&blk);
535		}
536
537		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
538		    ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
539
540		if (error == 0)
541			zil_init_log_chain(zilog, &blk);
542	}
543
544	/*
545	 * Allocate a log write buffer (lwb) for the first log block.
546	 */
547	if (error == 0)
548		lwb = zil_alloc_lwb(zilog, &blk, txg);
549
550	/*
551	 * If we just allocated the first log block, commit our transaction
552	 * and wait for zil_sync() to stuff the block poiner into zh_log.
553	 * (zh is part of the MOS, so we cannot modify it in open context.)
554	 */
555	if (tx != NULL) {
556		dmu_tx_commit(tx);
557		txg_wait_synced(zilog->zl_dmu_pool, txg);
558	}
559
560	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
561
562	return (lwb);
563}
564
565/*
566 * In one tx, free all log blocks and clear the log header.
567 * If keep_first is set, then we're replaying a log with no content.
568 * We want to keep the first block, however, so that the first
569 * synchronous transaction doesn't require a txg_wait_synced()
570 * in zil_create().  We don't need to txg_wait_synced() here either
571 * when keep_first is set, because both zil_create() and zil_destroy()
572 * will wait for any in-progress destroys to complete.
573 */
574void
575zil_destroy(zilog_t *zilog, boolean_t keep_first)
576{
577	const zil_header_t *zh = zilog->zl_header;
578	lwb_t *lwb;
579	dmu_tx_t *tx;
580	uint64_t txg;
581
582	/*
583	 * Wait for any previous destroy to complete.
584	 */
585	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
586
587	zilog->zl_old_header = *zh;		/* debugging aid */
588
589	if (BP_IS_HOLE(&zh->zh_log))
590		return;
591
592	tx = dmu_tx_create(zilog->zl_os);
593	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
594	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
595	txg = dmu_tx_get_txg(tx);
596
597	mutex_enter(&zilog->zl_lock);
598
599	ASSERT3U(zilog->zl_destroy_txg, <, txg);
600	zilog->zl_destroy_txg = txg;
601	zilog->zl_keep_first = keep_first;
602
603	if (!list_is_empty(&zilog->zl_lwb_list)) {
604		ASSERT(zh->zh_claim_txg == 0);
605		VERIFY(!keep_first);
606		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
607			list_remove(&zilog->zl_lwb_list, lwb);
608			if (lwb->lwb_buf != NULL)
609				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
610			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
611			kmem_cache_free(zil_lwb_cache, lwb);
612		}
613	} else if (!keep_first) {
614		zil_destroy_sync(zilog, tx);
615	}
616	mutex_exit(&zilog->zl_lock);
617
618	dmu_tx_commit(tx);
619}
620
621void
622zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
623{
624	ASSERT(list_is_empty(&zilog->zl_lwb_list));
625	(void) zil_parse(zilog, zil_free_log_block,
626	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
627}
628
629int
630zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
631{
632	dmu_tx_t *tx = txarg;
633	uint64_t first_txg = dmu_tx_get_txg(tx);
634	zilog_t *zilog;
635	zil_header_t *zh;
636	objset_t *os;
637	int error;
638
639	error = dmu_objset_own_obj(dp, ds->ds_object,
640	    DMU_OST_ANY, B_FALSE, FTAG, &os);
641	if (error != 0) {
642		/*
643		 * EBUSY indicates that the objset is inconsistent, in which
644		 * case it can not have a ZIL.
645		 */
646		if (error != EBUSY) {
647			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
648			    (unsigned long long)ds->ds_object, error);
649		}
650		return (0);
651	}
652
653	zilog = dmu_objset_zil(os);
654	zh = zil_header_in_syncing_context(zilog);
655
656	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
657		if (!BP_IS_HOLE(&zh->zh_log))
658			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
659		BP_ZERO(&zh->zh_log);
660		dsl_dataset_dirty(dmu_objset_ds(os), tx);
661		dmu_objset_disown(os, FTAG);
662		return (0);
663	}
664
665	/*
666	 * Claim all log blocks if we haven't already done so, and remember
667	 * the highest claimed sequence number.  This ensures that if we can
668	 * read only part of the log now (e.g. due to a missing device),
669	 * but we can read the entire log later, we will not try to replay
670	 * or destroy beyond the last block we successfully claimed.
671	 */
672	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
673	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
674		(void) zil_parse(zilog, zil_claim_log_block,
675		    zil_claim_log_record, tx, first_txg);
676		zh->zh_claim_txg = first_txg;
677		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
678		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
679		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
680			zh->zh_flags |= ZIL_REPLAY_NEEDED;
681		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
682		dsl_dataset_dirty(dmu_objset_ds(os), tx);
683	}
684
685	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
686	dmu_objset_disown(os, FTAG);
687	return (0);
688}
689
690/*
691 * Check the log by walking the log chain.
692 * Checksum errors are ok as they indicate the end of the chain.
693 * Any other error (no device or read failure) returns an error.
694 */
695/* ARGSUSED */
696int
697zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
698{
699	zilog_t *zilog;
700	objset_t *os;
701	blkptr_t *bp;
702	int error;
703
704	ASSERT(tx == NULL);
705
706	error = dmu_objset_from_ds(ds, &os);
707	if (error != 0) {
708		cmn_err(CE_WARN, "can't open objset %llu, error %d",
709		    (unsigned long long)ds->ds_object, error);
710		return (0);
711	}
712
713	zilog = dmu_objset_zil(os);
714	bp = (blkptr_t *)&zilog->zl_header->zh_log;
715
716	/*
717	 * Check the first block and determine if it's on a log device
718	 * which may have been removed or faulted prior to loading this
719	 * pool.  If so, there's no point in checking the rest of the log
720	 * as its content should have already been synced to the pool.
721	 */
722	if (!BP_IS_HOLE(bp)) {
723		vdev_t *vd;
724		boolean_t valid = B_TRUE;
725
726		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
727		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
728		if (vd->vdev_islog && vdev_is_dead(vd))
729			valid = vdev_log_state_valid(vd);
730		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
731
732		if (!valid)
733			return (0);
734	}
735
736	/*
737	 * Because tx == NULL, zil_claim_log_block() will not actually claim
738	 * any blocks, but just determine whether it is possible to do so.
739	 * In addition to checking the log chain, zil_claim_log_block()
740	 * will invoke zio_claim() with a done func of spa_claim_notify(),
741	 * which will update spa_max_claim_txg.  See spa_load() for details.
742	 */
743	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
744	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
745
746	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
747}
748
749static int
750zil_vdev_compare(const void *x1, const void *x2)
751{
752	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
753	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
754
755	if (v1 < v2)
756		return (-1);
757	if (v1 > v2)
758		return (1);
759
760	return (0);
761}
762
763void
764zil_add_block(zilog_t *zilog, const blkptr_t *bp)
765{
766	avl_tree_t *t = &zilog->zl_vdev_tree;
767	avl_index_t where;
768	zil_vdev_node_t *zv, zvsearch;
769	int ndvas = BP_GET_NDVAS(bp);
770	int i;
771
772	if (zfs_nocacheflush)
773		return;
774
775	ASSERT(zilog->zl_writer);
776
777	/*
778	 * Even though we're zl_writer, we still need a lock because the
779	 * zl_get_data() callbacks may have dmu_sync() done callbacks
780	 * that will run concurrently.
781	 */
782	mutex_enter(&zilog->zl_vdev_lock);
783	for (i = 0; i < ndvas; i++) {
784		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
785		if (avl_find(t, &zvsearch, &where) == NULL) {
786			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
787			zv->zv_vdev = zvsearch.zv_vdev;
788			avl_insert(t, zv, where);
789		}
790	}
791	mutex_exit(&zilog->zl_vdev_lock);
792}
793
794static void
795zil_flush_vdevs(zilog_t *zilog)
796{
797	spa_t *spa = zilog->zl_spa;
798	avl_tree_t *t = &zilog->zl_vdev_tree;
799	void *cookie = NULL;
800	zil_vdev_node_t *zv;
801	zio_t *zio;
802
803	ASSERT(zilog->zl_writer);
804
805	/*
806	 * We don't need zl_vdev_lock here because we're the zl_writer,
807	 * and all zl_get_data() callbacks are done.
808	 */
809	if (avl_numnodes(t) == 0)
810		return;
811
812	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
813
814	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
815
816	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
817		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
818		if (vd != NULL)
819			zio_flush(zio, vd);
820		kmem_free(zv, sizeof (*zv));
821	}
822
823	/*
824	 * Wait for all the flushes to complete.  Not all devices actually
825	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
826	 */
827	(void) zio_wait(zio);
828
829	spa_config_exit(spa, SCL_STATE, FTAG);
830}
831
832/*
833 * Function called when a log block write completes
834 */
835static void
836zil_lwb_write_done(zio_t *zio)
837{
838	lwb_t *lwb = zio->io_private;
839	zilog_t *zilog = lwb->lwb_zilog;
840	dmu_tx_t *tx = lwb->lwb_tx;
841
842	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
843	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
844	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
845	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
846	ASSERT(!BP_IS_GANG(zio->io_bp));
847	ASSERT(!BP_IS_HOLE(zio->io_bp));
848	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
849
850	/*
851	 * Ensure the lwb buffer pointer is cleared before releasing
852	 * the txg. If we have had an allocation failure and
853	 * the txg is waiting to sync then we want want zil_sync()
854	 * to remove the lwb so that it's not picked up as the next new
855	 * one in zil_commit_writer(). zil_sync() will only remove
856	 * the lwb if lwb_buf is null.
857	 */
858	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
859	mutex_enter(&zilog->zl_lock);
860	lwb->lwb_buf = NULL;
861	lwb->lwb_tx = NULL;
862	mutex_exit(&zilog->zl_lock);
863
864	/*
865	 * Now that we've written this log block, we have a stable pointer
866	 * to the next block in the chain, so it's OK to let the txg in
867	 * which we allocated the next block sync.
868	 */
869	dmu_tx_commit(tx);
870}
871
872/*
873 * Initialize the io for a log block.
874 */
875static void
876zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
877{
878	zbookmark_phys_t zb;
879
880	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
881	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
882	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
883
884	if (zilog->zl_root_zio == NULL) {
885		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
886		    ZIO_FLAG_CANFAIL);
887	}
888	if (lwb->lwb_zio == NULL) {
889		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
890		    0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
891		    zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
892		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
893	}
894}
895
896/*
897 * Define a limited set of intent log block sizes.
898 *
899 * These must be a multiple of 4KB. Note only the amount used (again
900 * aligned to 4KB) actually gets written. However, we can't always just
901 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
902 */
903uint64_t zil_block_buckets[] = {
904    4096,		/* non TX_WRITE */
905    8192+4096,		/* data base */
906    32*1024 + 4096, 	/* NFS writes */
907    UINT64_MAX
908};
909
910/*
911 * Use the slog as long as the logbias is 'latency' and the current commit size
912 * is less than the limit or the total list size is less than 2X the limit.
913 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
914 */
915uint64_t zil_slog_limit = 1024 * 1024;
916#define	USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
917	(((zilog)->zl_cur_used < zil_slog_limit) || \
918	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
919
920/*
921 * Start a log block write and advance to the next log block.
922 * Calls are serialized.
923 */
924static lwb_t *
925zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
926{
927	lwb_t *nlwb = NULL;
928	zil_chain_t *zilc;
929	spa_t *spa = zilog->zl_spa;
930	blkptr_t *bp;
931	dmu_tx_t *tx;
932	uint64_t txg;
933	uint64_t zil_blksz, wsz;
934	int i, error;
935
936	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
937		zilc = (zil_chain_t *)lwb->lwb_buf;
938		bp = &zilc->zc_next_blk;
939	} else {
940		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
941		bp = &zilc->zc_next_blk;
942	}
943
944	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
945
946	/*
947	 * Allocate the next block and save its address in this block
948	 * before writing it in order to establish the log chain.
949	 * Note that if the allocation of nlwb synced before we wrote
950	 * the block that points at it (lwb), we'd leak it if we crashed.
951	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
952	 * We dirty the dataset to ensure that zil_sync() will be called
953	 * to clean up in the event of allocation failure or I/O failure.
954	 */
955	tx = dmu_tx_create(zilog->zl_os);
956	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
957	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
958	txg = dmu_tx_get_txg(tx);
959
960	lwb->lwb_tx = tx;
961
962	/*
963	 * Log blocks are pre-allocated. Here we select the size of the next
964	 * block, based on size used in the last block.
965	 * - first find the smallest bucket that will fit the block from a
966	 *   limited set of block sizes. This is because it's faster to write
967	 *   blocks allocated from the same metaslab as they are adjacent or
968	 *   close.
969	 * - next find the maximum from the new suggested size and an array of
970	 *   previous sizes. This lessens a picket fence effect of wrongly
971	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
972	 *   requests.
973	 *
974	 * Note we only write what is used, but we can't just allocate
975	 * the maximum block size because we can exhaust the available
976	 * pool log space.
977	 */
978	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
979	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
980		continue;
981	zil_blksz = zil_block_buckets[i];
982	if (zil_blksz == UINT64_MAX)
983		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
984	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
985	for (i = 0; i < ZIL_PREV_BLKS; i++)
986		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
987	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
988
989	BP_ZERO(bp);
990	/* pass the old blkptr in order to spread log blocks across devs */
991	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
992	    USE_SLOG(zilog));
993	if (error == 0) {
994		ASSERT3U(bp->blk_birth, ==, txg);
995		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
996		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
997
998		/*
999		 * Allocate a new log write buffer (lwb).
1000		 */
1001		nlwb = zil_alloc_lwb(zilog, bp, txg);
1002
1003		/* Record the block for later vdev flushing */
1004		zil_add_block(zilog, &lwb->lwb_blk);
1005	}
1006
1007	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1008		/* For Slim ZIL only write what is used. */
1009		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1010		ASSERT3U(wsz, <=, lwb->lwb_sz);
1011		zio_shrink(lwb->lwb_zio, wsz);
1012
1013	} else {
1014		wsz = lwb->lwb_sz;
1015	}
1016
1017	zilc->zc_pad = 0;
1018	zilc->zc_nused = lwb->lwb_nused;
1019	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1020
1021	/*
1022	 * clear unused data for security
1023	 */
1024	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1025
1026	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1027
1028	/*
1029	 * If there was an allocation failure then nlwb will be null which
1030	 * forces a txg_wait_synced().
1031	 */
1032	return (nlwb);
1033}
1034
1035static lwb_t *
1036zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1037{
1038	lr_t *lrc = &itx->itx_lr; /* common log record */
1039	lr_write_t *lrw = (lr_write_t *)lrc;
1040	char *lr_buf;
1041	uint64_t txg = lrc->lrc_txg;
1042	uint64_t reclen = lrc->lrc_reclen;
1043	uint64_t dlen = 0;
1044
1045	if (lwb == NULL)
1046		return (NULL);
1047
1048	ASSERT(lwb->lwb_buf != NULL);
1049	ASSERT(zilog_is_dirty(zilog) ||
1050	    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1051
1052	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1053		dlen = P2ROUNDUP_TYPED(
1054		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1055
1056	zilog->zl_cur_used += (reclen + dlen);
1057
1058	zil_lwb_write_init(zilog, lwb);
1059
1060	/*
1061	 * If this record won't fit in the current log block, start a new one.
1062	 */
1063	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1064		lwb = zil_lwb_write_start(zilog, lwb);
1065		if (lwb == NULL)
1066			return (NULL);
1067		zil_lwb_write_init(zilog, lwb);
1068		ASSERT(LWB_EMPTY(lwb));
1069		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1070			txg_wait_synced(zilog->zl_dmu_pool, txg);
1071			return (lwb);
1072		}
1073	}
1074
1075	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1076	bcopy(lrc, lr_buf, reclen);
1077	lrc = (lr_t *)lr_buf;
1078	lrw = (lr_write_t *)lrc;
1079
1080	/*
1081	 * If it's a write, fetch the data or get its blkptr as appropriate.
1082	 */
1083	if (lrc->lrc_txtype == TX_WRITE) {
1084		if (txg > spa_freeze_txg(zilog->zl_spa))
1085			txg_wait_synced(zilog->zl_dmu_pool, txg);
1086		if (itx->itx_wr_state != WR_COPIED) {
1087			char *dbuf;
1088			int error;
1089
1090			if (dlen) {
1091				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1092				dbuf = lr_buf + reclen;
1093				lrw->lr_common.lrc_reclen += dlen;
1094			} else {
1095				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1096				dbuf = NULL;
1097			}
1098			error = zilog->zl_get_data(
1099			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1100			if (error == EIO) {
1101				txg_wait_synced(zilog->zl_dmu_pool, txg);
1102				return (lwb);
1103			}
1104			if (error != 0) {
1105				ASSERT(error == ENOENT || error == EEXIST ||
1106				    error == EALREADY);
1107				return (lwb);
1108			}
1109		}
1110	}
1111
1112	/*
1113	 * We're actually making an entry, so update lrc_seq to be the
1114	 * log record sequence number.  Note that this is generally not
1115	 * equal to the itx sequence number because not all transactions
1116	 * are synchronous, and sometimes spa_sync() gets there first.
1117	 */
1118	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1119	lwb->lwb_nused += reclen + dlen;
1120	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1121	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1122	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1123
1124	return (lwb);
1125}
1126
1127itx_t *
1128zil_itx_create(uint64_t txtype, size_t lrsize)
1129{
1130	itx_t *itx;
1131
1132	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1133
1134	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1135	itx->itx_lr.lrc_txtype = txtype;
1136	itx->itx_lr.lrc_reclen = lrsize;
1137	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1138	itx->itx_lr.lrc_seq = 0;	/* defensive */
1139	itx->itx_sync = B_TRUE;		/* default is synchronous */
1140
1141	return (itx);
1142}
1143
1144void
1145zil_itx_destroy(itx_t *itx)
1146{
1147	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1148}
1149
1150/*
1151 * Free up the sync and async itxs. The itxs_t has already been detached
1152 * so no locks are needed.
1153 */
1154static void
1155zil_itxg_clean(itxs_t *itxs)
1156{
1157	itx_t *itx;
1158	list_t *list;
1159	avl_tree_t *t;
1160	void *cookie;
1161	itx_async_node_t *ian;
1162
1163	list = &itxs->i_sync_list;
1164	while ((itx = list_head(list)) != NULL) {
1165		list_remove(list, itx);
1166		kmem_free(itx, offsetof(itx_t, itx_lr) +
1167		    itx->itx_lr.lrc_reclen);
1168	}
1169
1170	cookie = NULL;
1171	t = &itxs->i_async_tree;
1172	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1173		list = &ian->ia_list;
1174		while ((itx = list_head(list)) != NULL) {
1175			list_remove(list, itx);
1176			kmem_free(itx, offsetof(itx_t, itx_lr) +
1177			    itx->itx_lr.lrc_reclen);
1178		}
1179		list_destroy(list);
1180		kmem_free(ian, sizeof (itx_async_node_t));
1181	}
1182	avl_destroy(t);
1183
1184	kmem_free(itxs, sizeof (itxs_t));
1185}
1186
1187static int
1188zil_aitx_compare(const void *x1, const void *x2)
1189{
1190	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1191	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1192
1193	if (o1 < o2)
1194		return (-1);
1195	if (o1 > o2)
1196		return (1);
1197
1198	return (0);
1199}
1200
1201/*
1202 * Remove all async itx with the given oid.
1203 */
1204static void
1205zil_remove_async(zilog_t *zilog, uint64_t oid)
1206{
1207	uint64_t otxg, txg;
1208	itx_async_node_t *ian;
1209	avl_tree_t *t;
1210	avl_index_t where;
1211	list_t clean_list;
1212	itx_t *itx;
1213
1214	ASSERT(oid != 0);
1215	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1216
1217	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1218		otxg = ZILTEST_TXG;
1219	else
1220		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1221
1222	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1223		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1224
1225		mutex_enter(&itxg->itxg_lock);
1226		if (itxg->itxg_txg != txg) {
1227			mutex_exit(&itxg->itxg_lock);
1228			continue;
1229		}
1230
1231		/*
1232		 * Locate the object node and append its list.
1233		 */
1234		t = &itxg->itxg_itxs->i_async_tree;
1235		ian = avl_find(t, &oid, &where);
1236		if (ian != NULL)
1237			list_move_tail(&clean_list, &ian->ia_list);
1238		mutex_exit(&itxg->itxg_lock);
1239	}
1240	while ((itx = list_head(&clean_list)) != NULL) {
1241		list_remove(&clean_list, itx);
1242		kmem_free(itx, offsetof(itx_t, itx_lr) +
1243		    itx->itx_lr.lrc_reclen);
1244	}
1245	list_destroy(&clean_list);
1246}
1247
1248void
1249zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1250{
1251	uint64_t txg;
1252	itxg_t *itxg;
1253	itxs_t *itxs, *clean = NULL;
1254
1255	/*
1256	 * Object ids can be re-instantiated in the next txg so
1257	 * remove any async transactions to avoid future leaks.
1258	 * This can happen if a fsync occurs on the re-instantiated
1259	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1260	 * the new file data and flushes a write record for the old object.
1261	 */
1262	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1263		zil_remove_async(zilog, itx->itx_oid);
1264
1265	/*
1266	 * Ensure the data of a renamed file is committed before the rename.
1267	 */
1268	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1269		zil_async_to_sync(zilog, itx->itx_oid);
1270
1271	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1272		txg = ZILTEST_TXG;
1273	else
1274		txg = dmu_tx_get_txg(tx);
1275
1276	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1277	mutex_enter(&itxg->itxg_lock);
1278	itxs = itxg->itxg_itxs;
1279	if (itxg->itxg_txg != txg) {
1280		if (itxs != NULL) {
1281			/*
1282			 * The zil_clean callback hasn't got around to cleaning
1283			 * this itxg. Save the itxs for release below.
1284			 * This should be rare.
1285			 */
1286			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1287			itxg->itxg_sod = 0;
1288			clean = itxg->itxg_itxs;
1289		}
1290		ASSERT(itxg->itxg_sod == 0);
1291		itxg->itxg_txg = txg;
1292		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1293
1294		list_create(&itxs->i_sync_list, sizeof (itx_t),
1295		    offsetof(itx_t, itx_node));
1296		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1297		    sizeof (itx_async_node_t),
1298		    offsetof(itx_async_node_t, ia_node));
1299	}
1300	if (itx->itx_sync) {
1301		list_insert_tail(&itxs->i_sync_list, itx);
1302		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1303		itxg->itxg_sod += itx->itx_sod;
1304	} else {
1305		avl_tree_t *t = &itxs->i_async_tree;
1306		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1307		itx_async_node_t *ian;
1308		avl_index_t where;
1309
1310		ian = avl_find(t, &foid, &where);
1311		if (ian == NULL) {
1312			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1313			list_create(&ian->ia_list, sizeof (itx_t),
1314			    offsetof(itx_t, itx_node));
1315			ian->ia_foid = foid;
1316			avl_insert(t, ian, where);
1317		}
1318		list_insert_tail(&ian->ia_list, itx);
1319	}
1320
1321	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1322	zilog_dirty(zilog, txg);
1323	mutex_exit(&itxg->itxg_lock);
1324
1325	/* Release the old itxs now we've dropped the lock */
1326	if (clean != NULL)
1327		zil_itxg_clean(clean);
1328}
1329
1330/*
1331 * If there are any in-memory intent log transactions which have now been
1332 * synced then start up a taskq to free them. We should only do this after we
1333 * have written out the uberblocks (i.e. txg has been comitted) so that
1334 * don't inadvertently clean out in-memory log records that would be required
1335 * by zil_commit().
1336 */
1337void
1338zil_clean(zilog_t *zilog, uint64_t synced_txg)
1339{
1340	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1341	itxs_t *clean_me;
1342
1343	mutex_enter(&itxg->itxg_lock);
1344	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1345		mutex_exit(&itxg->itxg_lock);
1346		return;
1347	}
1348	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1349	ASSERT(itxg->itxg_txg != 0);
1350	ASSERT(zilog->zl_clean_taskq != NULL);
1351	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1352	itxg->itxg_sod = 0;
1353	clean_me = itxg->itxg_itxs;
1354	itxg->itxg_itxs = NULL;
1355	itxg->itxg_txg = 0;
1356	mutex_exit(&itxg->itxg_lock);
1357	/*
1358	 * Preferably start a task queue to free up the old itxs but
1359	 * if taskq_dispatch can't allocate resources to do that then
1360	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1361	 * created a bad performance problem.
1362	 */
1363	if (taskq_dispatch(zilog->zl_clean_taskq,
1364	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1365		zil_itxg_clean(clean_me);
1366}
1367
1368/*
1369 * Get the list of itxs to commit into zl_itx_commit_list.
1370 */
1371static void
1372zil_get_commit_list(zilog_t *zilog)
1373{
1374	uint64_t otxg, txg;
1375	list_t *commit_list = &zilog->zl_itx_commit_list;
1376	uint64_t push_sod = 0;
1377
1378	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1379		otxg = ZILTEST_TXG;
1380	else
1381		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1382
1383	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1384		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1385
1386		mutex_enter(&itxg->itxg_lock);
1387		if (itxg->itxg_txg != txg) {
1388			mutex_exit(&itxg->itxg_lock);
1389			continue;
1390		}
1391
1392		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1393		push_sod += itxg->itxg_sod;
1394		itxg->itxg_sod = 0;
1395
1396		mutex_exit(&itxg->itxg_lock);
1397	}
1398	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1399}
1400
1401/*
1402 * Move the async itxs for a specified object to commit into sync lists.
1403 */
1404static void
1405zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1406{
1407	uint64_t otxg, txg;
1408	itx_async_node_t *ian;
1409	avl_tree_t *t;
1410	avl_index_t where;
1411
1412	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1413		otxg = ZILTEST_TXG;
1414	else
1415		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1416
1417	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1418		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1419
1420		mutex_enter(&itxg->itxg_lock);
1421		if (itxg->itxg_txg != txg) {
1422			mutex_exit(&itxg->itxg_lock);
1423			continue;
1424		}
1425
1426		/*
1427		 * If a foid is specified then find that node and append its
1428		 * list. Otherwise walk the tree appending all the lists
1429		 * to the sync list. We add to the end rather than the
1430		 * beginning to ensure the create has happened.
1431		 */
1432		t = &itxg->itxg_itxs->i_async_tree;
1433		if (foid != 0) {
1434			ian = avl_find(t, &foid, &where);
1435			if (ian != NULL) {
1436				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1437				    &ian->ia_list);
1438			}
1439		} else {
1440			void *cookie = NULL;
1441
1442			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1443				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1444				    &ian->ia_list);
1445				list_destroy(&ian->ia_list);
1446				kmem_free(ian, sizeof (itx_async_node_t));
1447			}
1448		}
1449		mutex_exit(&itxg->itxg_lock);
1450	}
1451}
1452
1453static void
1454zil_commit_writer(zilog_t *zilog)
1455{
1456	uint64_t txg;
1457	itx_t *itx;
1458	lwb_t *lwb;
1459	spa_t *spa = zilog->zl_spa;
1460	int error = 0;
1461
1462	ASSERT(zilog->zl_root_zio == NULL);
1463
1464	mutex_exit(&zilog->zl_lock);
1465
1466	zil_get_commit_list(zilog);
1467
1468	/*
1469	 * Return if there's nothing to commit before we dirty the fs by
1470	 * calling zil_create().
1471	 */
1472	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1473		mutex_enter(&zilog->zl_lock);
1474		return;
1475	}
1476
1477	if (zilog->zl_suspend) {
1478		lwb = NULL;
1479	} else {
1480		lwb = list_tail(&zilog->zl_lwb_list);
1481		if (lwb == NULL)
1482			lwb = zil_create(zilog);
1483	}
1484
1485	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1486	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1487		txg = itx->itx_lr.lrc_txg;
1488		ASSERT(txg);
1489
1490		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1491			lwb = zil_lwb_commit(zilog, itx, lwb);
1492		list_remove(&zilog->zl_itx_commit_list, itx);
1493		kmem_free(itx, offsetof(itx_t, itx_lr)
1494		    + itx->itx_lr.lrc_reclen);
1495	}
1496	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1497
1498	/* write the last block out */
1499	if (lwb != NULL && lwb->lwb_zio != NULL)
1500		lwb = zil_lwb_write_start(zilog, lwb);
1501
1502	zilog->zl_cur_used = 0;
1503
1504	/*
1505	 * Wait if necessary for the log blocks to be on stable storage.
1506	 */
1507	if (zilog->zl_root_zio) {
1508		error = zio_wait(zilog->zl_root_zio);
1509		zilog->zl_root_zio = NULL;
1510		zil_flush_vdevs(zilog);
1511	}
1512
1513	if (error || lwb == NULL)
1514		txg_wait_synced(zilog->zl_dmu_pool, 0);
1515
1516	mutex_enter(&zilog->zl_lock);
1517
1518	/*
1519	 * Remember the highest committed log sequence number for ztest.
1520	 * We only update this value when all the log writes succeeded,
1521	 * because ztest wants to ASSERT that it got the whole log chain.
1522	 */
1523	if (error == 0 && lwb != NULL)
1524		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1525}
1526
1527/*
1528 * Commit zfs transactions to stable storage.
1529 * If foid is 0 push out all transactions, otherwise push only those
1530 * for that object or might reference that object.
1531 *
1532 * itxs are committed in batches. In a heavily stressed zil there will be
1533 * a commit writer thread who is writing out a bunch of itxs to the log
1534 * for a set of committing threads (cthreads) in the same batch as the writer.
1535 * Those cthreads are all waiting on the same cv for that batch.
1536 *
1537 * There will also be a different and growing batch of threads that are
1538 * waiting to commit (qthreads). When the committing batch completes
1539 * a transition occurs such that the cthreads exit and the qthreads become
1540 * cthreads. One of the new cthreads becomes the writer thread for the
1541 * batch. Any new threads arriving become new qthreads.
1542 *
1543 * Only 2 condition variables are needed and there's no transition
1544 * between the two cvs needed. They just flip-flop between qthreads
1545 * and cthreads.
1546 *
1547 * Using this scheme we can efficiently wakeup up only those threads
1548 * that have been committed.
1549 */
1550void
1551zil_commit(zilog_t *zilog, uint64_t foid)
1552{
1553	uint64_t mybatch;
1554
1555	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1556		return;
1557
1558	/* move the async itxs for the foid to the sync queues */
1559	zil_async_to_sync(zilog, foid);
1560
1561	mutex_enter(&zilog->zl_lock);
1562	mybatch = zilog->zl_next_batch;
1563	while (zilog->zl_writer) {
1564		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1565		if (mybatch <= zilog->zl_com_batch) {
1566			mutex_exit(&zilog->zl_lock);
1567			return;
1568		}
1569	}
1570
1571	zilog->zl_next_batch++;
1572	zilog->zl_writer = B_TRUE;
1573	zil_commit_writer(zilog);
1574	zilog->zl_com_batch = mybatch;
1575	zilog->zl_writer = B_FALSE;
1576	mutex_exit(&zilog->zl_lock);
1577
1578	/* wake up one thread to become the next writer */
1579	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1580
1581	/* wake up all threads waiting for this batch to be committed */
1582	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1583}
1584
1585/*
1586 * Called in syncing context to free committed log blocks and update log header.
1587 */
1588void
1589zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1590{
1591	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1592	uint64_t txg = dmu_tx_get_txg(tx);
1593	spa_t *spa = zilog->zl_spa;
1594	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1595	lwb_t *lwb;
1596
1597	/*
1598	 * We don't zero out zl_destroy_txg, so make sure we don't try
1599	 * to destroy it twice.
1600	 */
1601	if (spa_sync_pass(spa) != 1)
1602		return;
1603
1604	mutex_enter(&zilog->zl_lock);
1605
1606	ASSERT(zilog->zl_stop_sync == 0);
1607
1608	if (*replayed_seq != 0) {
1609		ASSERT(zh->zh_replay_seq < *replayed_seq);
1610		zh->zh_replay_seq = *replayed_seq;
1611		*replayed_seq = 0;
1612	}
1613
1614	if (zilog->zl_destroy_txg == txg) {
1615		blkptr_t blk = zh->zh_log;
1616
1617		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1618
1619		bzero(zh, sizeof (zil_header_t));
1620		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1621
1622		if (zilog->zl_keep_first) {
1623			/*
1624			 * If this block was part of log chain that couldn't
1625			 * be claimed because a device was missing during
1626			 * zil_claim(), but that device later returns,
1627			 * then this block could erroneously appear valid.
1628			 * To guard against this, assign a new GUID to the new
1629			 * log chain so it doesn't matter what blk points to.
1630			 */
1631			zil_init_log_chain(zilog, &blk);
1632			zh->zh_log = blk;
1633		}
1634	}
1635
1636	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1637		zh->zh_log = lwb->lwb_blk;
1638		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1639			break;
1640		list_remove(&zilog->zl_lwb_list, lwb);
1641		zio_free_zil(spa, txg, &lwb->lwb_blk);
1642		kmem_cache_free(zil_lwb_cache, lwb);
1643
1644		/*
1645		 * If we don't have anything left in the lwb list then
1646		 * we've had an allocation failure and we need to zero
1647		 * out the zil_header blkptr so that we don't end
1648		 * up freeing the same block twice.
1649		 */
1650		if (list_head(&zilog->zl_lwb_list) == NULL)
1651			BP_ZERO(&zh->zh_log);
1652	}
1653	mutex_exit(&zilog->zl_lock);
1654}
1655
1656void
1657zil_init(void)
1658{
1659	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1660	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1661}
1662
1663void
1664zil_fini(void)
1665{
1666	kmem_cache_destroy(zil_lwb_cache);
1667}
1668
1669void
1670zil_set_sync(zilog_t *zilog, uint64_t sync)
1671{
1672	zilog->zl_sync = sync;
1673}
1674
1675void
1676zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1677{
1678	zilog->zl_logbias = logbias;
1679}
1680
1681zilog_t *
1682zil_alloc(objset_t *os, zil_header_t *zh_phys)
1683{
1684	zilog_t *zilog;
1685
1686	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1687
1688	zilog->zl_header = zh_phys;
1689	zilog->zl_os = os;
1690	zilog->zl_spa = dmu_objset_spa(os);
1691	zilog->zl_dmu_pool = dmu_objset_pool(os);
1692	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1693	zilog->zl_logbias = dmu_objset_logbias(os);
1694	zilog->zl_sync = dmu_objset_syncprop(os);
1695	zilog->zl_next_batch = 1;
1696
1697	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1698
1699	for (int i = 0; i < TXG_SIZE; i++) {
1700		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1701		    MUTEX_DEFAULT, NULL);
1702	}
1703
1704	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1705	    offsetof(lwb_t, lwb_node));
1706
1707	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1708	    offsetof(itx_t, itx_node));
1709
1710	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1711
1712	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1713	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1714
1715	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1716	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1717	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1718	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1719
1720	return (zilog);
1721}
1722
1723void
1724zil_free(zilog_t *zilog)
1725{
1726	zilog->zl_stop_sync = 1;
1727
1728	ASSERT0(zilog->zl_suspend);
1729	ASSERT0(zilog->zl_suspending);
1730
1731	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1732	list_destroy(&zilog->zl_lwb_list);
1733
1734	avl_destroy(&zilog->zl_vdev_tree);
1735	mutex_destroy(&zilog->zl_vdev_lock);
1736
1737	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1738	list_destroy(&zilog->zl_itx_commit_list);
1739
1740	for (int i = 0; i < TXG_SIZE; i++) {
1741		/*
1742		 * It's possible for an itx to be generated that doesn't dirty
1743		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1744		 * callback to remove the entry. We remove those here.
1745		 *
1746		 * Also free up the ziltest itxs.
1747		 */
1748		if (zilog->zl_itxg[i].itxg_itxs)
1749			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1750		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1751	}
1752
1753	mutex_destroy(&zilog->zl_lock);
1754
1755	cv_destroy(&zilog->zl_cv_writer);
1756	cv_destroy(&zilog->zl_cv_suspend);
1757	cv_destroy(&zilog->zl_cv_batch[0]);
1758	cv_destroy(&zilog->zl_cv_batch[1]);
1759
1760	kmem_free(zilog, sizeof (zilog_t));
1761}
1762
1763/*
1764 * Open an intent log.
1765 */
1766zilog_t *
1767zil_open(objset_t *os, zil_get_data_t *get_data)
1768{
1769	zilog_t *zilog = dmu_objset_zil(os);
1770
1771	ASSERT(zilog->zl_clean_taskq == NULL);
1772	ASSERT(zilog->zl_get_data == NULL);
1773	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1774
1775	zilog->zl_get_data = get_data;
1776	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1777	    2, 2, TASKQ_PREPOPULATE);
1778
1779	return (zilog);
1780}
1781
1782/*
1783 * Close an intent log.
1784 */
1785void
1786zil_close(zilog_t *zilog)
1787{
1788	lwb_t *lwb;
1789	uint64_t txg = 0;
1790
1791	zil_commit(zilog, 0); /* commit all itx */
1792
1793	/*
1794	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1795	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1796	 * callbacks have occurred that may clean the zil.  Only then can we
1797	 * destroy the zl_clean_taskq.
1798	 */
1799	mutex_enter(&zilog->zl_lock);
1800	lwb = list_tail(&zilog->zl_lwb_list);
1801	if (lwb != NULL)
1802		txg = lwb->lwb_max_txg;
1803	mutex_exit(&zilog->zl_lock);
1804	if (txg)
1805		txg_wait_synced(zilog->zl_dmu_pool, txg);
1806	ASSERT(!zilog_is_dirty(zilog));
1807
1808	taskq_destroy(zilog->zl_clean_taskq);
1809	zilog->zl_clean_taskq = NULL;
1810	zilog->zl_get_data = NULL;
1811
1812	/*
1813	 * We should have only one LWB left on the list; remove it now.
1814	 */
1815	mutex_enter(&zilog->zl_lock);
1816	lwb = list_head(&zilog->zl_lwb_list);
1817	if (lwb != NULL) {
1818		ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1819		list_remove(&zilog->zl_lwb_list, lwb);
1820		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1821		kmem_cache_free(zil_lwb_cache, lwb);
1822	}
1823	mutex_exit(&zilog->zl_lock);
1824}
1825
1826static char *suspend_tag = "zil suspending";
1827
1828/*
1829 * Suspend an intent log.  While in suspended mode, we still honor
1830 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1831 * On old version pools, we suspend the log briefly when taking a
1832 * snapshot so that it will have an empty intent log.
1833 *
1834 * Long holds are not really intended to be used the way we do here --
1835 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
1836 * could fail.  Therefore we take pains to only put a long hold if it is
1837 * actually necessary.  Fortunately, it will only be necessary if the
1838 * objset is currently mounted (or the ZVOL equivalent).  In that case it
1839 * will already have a long hold, so we are not really making things any worse.
1840 *
1841 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1842 * zvol_state_t), and use their mechanism to prevent their hold from being
1843 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
1844 * very little gain.
1845 *
1846 * if cookiep == NULL, this does both the suspend & resume.
1847 * Otherwise, it returns with the dataset "long held", and the cookie
1848 * should be passed into zil_resume().
1849 */
1850int
1851zil_suspend(const char *osname, void **cookiep)
1852{
1853	objset_t *os;
1854	zilog_t *zilog;
1855	const zil_header_t *zh;
1856	int error;
1857
1858	error = dmu_objset_hold(osname, suspend_tag, &os);
1859	if (error != 0)
1860		return (error);
1861	zilog = dmu_objset_zil(os);
1862
1863	mutex_enter(&zilog->zl_lock);
1864	zh = zilog->zl_header;
1865
1866	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1867		mutex_exit(&zilog->zl_lock);
1868		dmu_objset_rele(os, suspend_tag);
1869		return (SET_ERROR(EBUSY));
1870	}
1871
1872	/*
1873	 * Don't put a long hold in the cases where we can avoid it.  This
1874	 * is when there is no cookie so we are doing a suspend & resume
1875	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1876	 * for the suspend because it's already suspended, or there's no ZIL.
1877	 */
1878	if (cookiep == NULL && !zilog->zl_suspending &&
1879	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1880		mutex_exit(&zilog->zl_lock);
1881		dmu_objset_rele(os, suspend_tag);
1882		return (0);
1883	}
1884
1885	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1886	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1887
1888	zilog->zl_suspend++;
1889
1890	if (zilog->zl_suspend > 1) {
1891		/*
1892		 * Someone else is already suspending it.
1893		 * Just wait for them to finish.
1894		 */
1895
1896		while (zilog->zl_suspending)
1897			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1898		mutex_exit(&zilog->zl_lock);
1899
1900		if (cookiep == NULL)
1901			zil_resume(os);
1902		else
1903			*cookiep = os;
1904		return (0);
1905	}
1906
1907	/*
1908	 * If there is no pointer to an on-disk block, this ZIL must not
1909	 * be active (e.g. filesystem not mounted), so there's nothing
1910	 * to clean up.
1911	 */
1912	if (BP_IS_HOLE(&zh->zh_log)) {
1913		ASSERT(cookiep != NULL); /* fast path already handled */
1914
1915		*cookiep = os;
1916		mutex_exit(&zilog->zl_lock);
1917		return (0);
1918	}
1919
1920	zilog->zl_suspending = B_TRUE;
1921	mutex_exit(&zilog->zl_lock);
1922
1923	zil_commit(zilog, 0);
1924
1925	zil_destroy(zilog, B_FALSE);
1926
1927	mutex_enter(&zilog->zl_lock);
1928	zilog->zl_suspending = B_FALSE;
1929	cv_broadcast(&zilog->zl_cv_suspend);
1930	mutex_exit(&zilog->zl_lock);
1931
1932	if (cookiep == NULL)
1933		zil_resume(os);
1934	else
1935		*cookiep = os;
1936	return (0);
1937}
1938
1939void
1940zil_resume(void *cookie)
1941{
1942	objset_t *os = cookie;
1943	zilog_t *zilog = dmu_objset_zil(os);
1944
1945	mutex_enter(&zilog->zl_lock);
1946	ASSERT(zilog->zl_suspend != 0);
1947	zilog->zl_suspend--;
1948	mutex_exit(&zilog->zl_lock);
1949	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1950	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1951}
1952
1953typedef struct zil_replay_arg {
1954	zil_replay_func_t **zr_replay;
1955	void		*zr_arg;
1956	boolean_t	zr_byteswap;
1957	char		*zr_lr;
1958} zil_replay_arg_t;
1959
1960static int
1961zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1962{
1963	char name[MAXNAMELEN];
1964
1965	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
1966
1967	dmu_objset_name(zilog->zl_os, name);
1968
1969	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1970	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1971	    (u_longlong_t)lr->lrc_seq,
1972	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1973	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1974
1975	return (error);
1976}
1977
1978static int
1979zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1980{
1981	zil_replay_arg_t *zr = zra;
1982	const zil_header_t *zh = zilog->zl_header;
1983	uint64_t reclen = lr->lrc_reclen;
1984	uint64_t txtype = lr->lrc_txtype;
1985	int error = 0;
1986
1987	zilog->zl_replaying_seq = lr->lrc_seq;
1988
1989	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1990		return (0);
1991
1992	if (lr->lrc_txg < claim_txg)		/* already committed */
1993		return (0);
1994
1995	/* Strip case-insensitive bit, still present in log record */
1996	txtype &= ~TX_CI;
1997
1998	if (txtype == 0 || txtype >= TX_MAX_TYPE)
1999		return (zil_replay_error(zilog, lr, EINVAL));
2000
2001	/*
2002	 * If this record type can be logged out of order, the object
2003	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
2004	 */
2005	if (TX_OOO(txtype)) {
2006		error = dmu_object_info(zilog->zl_os,
2007		    ((lr_ooo_t *)lr)->lr_foid, NULL);
2008		if (error == ENOENT || error == EEXIST)
2009			return (0);
2010	}
2011
2012	/*
2013	 * Make a copy of the data so we can revise and extend it.
2014	 */
2015	bcopy(lr, zr->zr_lr, reclen);
2016
2017	/*
2018	 * If this is a TX_WRITE with a blkptr, suck in the data.
2019	 */
2020	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2021		error = zil_read_log_data(zilog, (lr_write_t *)lr,
2022		    zr->zr_lr + reclen);
2023		if (error != 0)
2024			return (zil_replay_error(zilog, lr, error));
2025	}
2026
2027	/*
2028	 * The log block containing this lr may have been byteswapped
2029	 * so that we can easily examine common fields like lrc_txtype.
2030	 * However, the log is a mix of different record types, and only the
2031	 * replay vectors know how to byteswap their records.  Therefore, if
2032	 * the lr was byteswapped, undo it before invoking the replay vector.
2033	 */
2034	if (zr->zr_byteswap)
2035		byteswap_uint64_array(zr->zr_lr, reclen);
2036
2037	/*
2038	 * We must now do two things atomically: replay this log record,
2039	 * and update the log header sequence number to reflect the fact that
2040	 * we did so. At the end of each replay function the sequence number
2041	 * is updated if we are in replay mode.
2042	 */
2043	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2044	if (error != 0) {
2045		/*
2046		 * The DMU's dnode layer doesn't see removes until the txg
2047		 * commits, so a subsequent claim can spuriously fail with
2048		 * EEXIST. So if we receive any error we try syncing out
2049		 * any removes then retry the transaction.  Note that we
2050		 * specify B_FALSE for byteswap now, so we don't do it twice.
2051		 */
2052		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2053		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2054		if (error != 0)
2055			return (zil_replay_error(zilog, lr, error));
2056	}
2057	return (0);
2058}
2059
2060/* ARGSUSED */
2061static int
2062zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2063{
2064	zilog->zl_replay_blks++;
2065
2066	return (0);
2067}
2068
2069/*
2070 * If this dataset has a non-empty intent log, replay it and destroy it.
2071 */
2072void
2073zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2074{
2075	zilog_t *zilog = dmu_objset_zil(os);
2076	const zil_header_t *zh = zilog->zl_header;
2077	zil_replay_arg_t zr;
2078
2079	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2080		zil_destroy(zilog, B_TRUE);
2081		return;
2082	}
2083
2084	zr.zr_replay = replay_func;
2085	zr.zr_arg = arg;
2086	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2087	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2088
2089	/*
2090	 * Wait for in-progress removes to sync before starting replay.
2091	 */
2092	txg_wait_synced(zilog->zl_dmu_pool, 0);
2093
2094	zilog->zl_replay = B_TRUE;
2095	zilog->zl_replay_time = ddi_get_lbolt();
2096	ASSERT(zilog->zl_replay_blks == 0);
2097	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2098	    zh->zh_claim_txg);
2099	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2100
2101	zil_destroy(zilog, B_FALSE);
2102	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2103	zilog->zl_replay = B_FALSE;
2104}
2105
2106boolean_t
2107zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2108{
2109	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2110		return (B_TRUE);
2111
2112	if (zilog->zl_replay) {
2113		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2114		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2115		    zilog->zl_replaying_seq;
2116		return (B_TRUE);
2117	}
2118
2119	return (B_FALSE);
2120}
2121
2122/* ARGSUSED */
2123int
2124zil_vdev_offline(const char *osname, void *arg)
2125{
2126	int error;
2127
2128	error = zil_suspend(osname, NULL);
2129	if (error != 0)
2130		return (SET_ERROR(EEXIST));
2131	return (0);
2132}
2133