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, 2016 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		arc_buf_destroy(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		arc_buf_destroy(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
487/*
488 * Determine if the zil is dirty in the specified txg. Callers wanting to
489 * ensure that the dirty state does not change must hold the itxg_lock for
490 * the specified txg. Holding the lock will ensure that the zil cannot be
491 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
492 * state.
493 */
494boolean_t
495zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
496{
497	dsl_pool_t *dp = zilog->zl_dmu_pool;
498
499	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
500		return (B_TRUE);
501	return (B_FALSE);
502}
503
504/*
505 * Determine if the zil is dirty. The zil is considered dirty if it has
506 * any pending itx records that have not been cleaned by zil_clean().
507 */
508boolean_t
509zilog_is_dirty(zilog_t *zilog)
510{
511	dsl_pool_t *dp = zilog->zl_dmu_pool;
512
513	for (int t = 0; t < TXG_SIZE; t++) {
514		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
515			return (B_TRUE);
516	}
517	return (B_FALSE);
518}
519
520/*
521 * Create an on-disk intent log.
522 */
523static lwb_t *
524zil_create(zilog_t *zilog)
525{
526	const zil_header_t *zh = zilog->zl_header;
527	lwb_t *lwb = NULL;
528	uint64_t txg = 0;
529	dmu_tx_t *tx = NULL;
530	blkptr_t blk;
531	int error = 0;
532
533	/*
534	 * Wait for any previous destroy to complete.
535	 */
536	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
537
538	ASSERT(zh->zh_claim_txg == 0);
539	ASSERT(zh->zh_replay_seq == 0);
540
541	blk = zh->zh_log;
542
543	/*
544	 * Allocate an initial log block if:
545	 *    - there isn't one already
546	 *    - the existing block is the wrong endianess
547	 */
548	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
549		tx = dmu_tx_create(zilog->zl_os);
550		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
551		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
552		txg = dmu_tx_get_txg(tx);
553
554		if (!BP_IS_HOLE(&blk)) {
555			zio_free_zil(zilog->zl_spa, txg, &blk);
556			BP_ZERO(&blk);
557		}
558
559		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
560		    ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
561
562		if (error == 0)
563			zil_init_log_chain(zilog, &blk);
564	}
565
566	/*
567	 * Allocate a log write buffer (lwb) for the first log block.
568	 */
569	if (error == 0)
570		lwb = zil_alloc_lwb(zilog, &blk, txg);
571
572	/*
573	 * If we just allocated the first log block, commit our transaction
574	 * and wait for zil_sync() to stuff the block poiner into zh_log.
575	 * (zh is part of the MOS, so we cannot modify it in open context.)
576	 */
577	if (tx != NULL) {
578		dmu_tx_commit(tx);
579		txg_wait_synced(zilog->zl_dmu_pool, txg);
580	}
581
582	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
583
584	return (lwb);
585}
586
587/*
588 * In one tx, free all log blocks and clear the log header.
589 * If keep_first is set, then we're replaying a log with no content.
590 * We want to keep the first block, however, so that the first
591 * synchronous transaction doesn't require a txg_wait_synced()
592 * in zil_create().  We don't need to txg_wait_synced() here either
593 * when keep_first is set, because both zil_create() and zil_destroy()
594 * will wait for any in-progress destroys to complete.
595 */
596void
597zil_destroy(zilog_t *zilog, boolean_t keep_first)
598{
599	const zil_header_t *zh = zilog->zl_header;
600	lwb_t *lwb;
601	dmu_tx_t *tx;
602	uint64_t txg;
603
604	/*
605	 * Wait for any previous destroy to complete.
606	 */
607	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
608
609	zilog->zl_old_header = *zh;		/* debugging aid */
610
611	if (BP_IS_HOLE(&zh->zh_log))
612		return;
613
614	tx = dmu_tx_create(zilog->zl_os);
615	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
616	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
617	txg = dmu_tx_get_txg(tx);
618
619	mutex_enter(&zilog->zl_lock);
620
621	ASSERT3U(zilog->zl_destroy_txg, <, txg);
622	zilog->zl_destroy_txg = txg;
623	zilog->zl_keep_first = keep_first;
624
625	if (!list_is_empty(&zilog->zl_lwb_list)) {
626		ASSERT(zh->zh_claim_txg == 0);
627		VERIFY(!keep_first);
628		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
629			list_remove(&zilog->zl_lwb_list, lwb);
630			if (lwb->lwb_buf != NULL)
631				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
632			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
633			kmem_cache_free(zil_lwb_cache, lwb);
634		}
635	} else if (!keep_first) {
636		zil_destroy_sync(zilog, tx);
637	}
638	mutex_exit(&zilog->zl_lock);
639
640	dmu_tx_commit(tx);
641}
642
643void
644zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
645{
646	ASSERT(list_is_empty(&zilog->zl_lwb_list));
647	(void) zil_parse(zilog, zil_free_log_block,
648	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
649}
650
651int
652zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
653{
654	dmu_tx_t *tx = txarg;
655	uint64_t first_txg = dmu_tx_get_txg(tx);
656	zilog_t *zilog;
657	zil_header_t *zh;
658	objset_t *os;
659	int error;
660
661	error = dmu_objset_own_obj(dp, ds->ds_object,
662	    DMU_OST_ANY, B_FALSE, FTAG, &os);
663	if (error != 0) {
664		/*
665		 * EBUSY indicates that the objset is inconsistent, in which
666		 * case it can not have a ZIL.
667		 */
668		if (error != EBUSY) {
669			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
670			    (unsigned long long)ds->ds_object, error);
671		}
672		return (0);
673	}
674
675	zilog = dmu_objset_zil(os);
676	zh = zil_header_in_syncing_context(zilog);
677
678	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
679		if (!BP_IS_HOLE(&zh->zh_log))
680			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
681		BP_ZERO(&zh->zh_log);
682		dsl_dataset_dirty(dmu_objset_ds(os), tx);
683		dmu_objset_disown(os, FTAG);
684		return (0);
685	}
686
687	/*
688	 * Claim all log blocks if we haven't already done so, and remember
689	 * the highest claimed sequence number.  This ensures that if we can
690	 * read only part of the log now (e.g. due to a missing device),
691	 * but we can read the entire log later, we will not try to replay
692	 * or destroy beyond the last block we successfully claimed.
693	 */
694	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
695	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
696		(void) zil_parse(zilog, zil_claim_log_block,
697		    zil_claim_log_record, tx, first_txg);
698		zh->zh_claim_txg = first_txg;
699		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
700		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
701		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
702			zh->zh_flags |= ZIL_REPLAY_NEEDED;
703		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
704		dsl_dataset_dirty(dmu_objset_ds(os), tx);
705	}
706
707	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
708	dmu_objset_disown(os, FTAG);
709	return (0);
710}
711
712/*
713 * Check the log by walking the log chain.
714 * Checksum errors are ok as they indicate the end of the chain.
715 * Any other error (no device or read failure) returns an error.
716 */
717/* ARGSUSED */
718int
719zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
720{
721	zilog_t *zilog;
722	objset_t *os;
723	blkptr_t *bp;
724	int error;
725
726	ASSERT(tx == NULL);
727
728	error = dmu_objset_from_ds(ds, &os);
729	if (error != 0) {
730		cmn_err(CE_WARN, "can't open objset %llu, error %d",
731		    (unsigned long long)ds->ds_object, error);
732		return (0);
733	}
734
735	zilog = dmu_objset_zil(os);
736	bp = (blkptr_t *)&zilog->zl_header->zh_log;
737
738	/*
739	 * Check the first block and determine if it's on a log device
740	 * which may have been removed or faulted prior to loading this
741	 * pool.  If so, there's no point in checking the rest of the log
742	 * as its content should have already been synced to the pool.
743	 */
744	if (!BP_IS_HOLE(bp)) {
745		vdev_t *vd;
746		boolean_t valid = B_TRUE;
747
748		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
749		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
750		if (vd->vdev_islog && vdev_is_dead(vd))
751			valid = vdev_log_state_valid(vd);
752		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
753
754		if (!valid)
755			return (0);
756	}
757
758	/*
759	 * Because tx == NULL, zil_claim_log_block() will not actually claim
760	 * any blocks, but just determine whether it is possible to do so.
761	 * In addition to checking the log chain, zil_claim_log_block()
762	 * will invoke zio_claim() with a done func of spa_claim_notify(),
763	 * which will update spa_max_claim_txg.  See spa_load() for details.
764	 */
765	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
766	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
767
768	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
769}
770
771static int
772zil_vdev_compare(const void *x1, const void *x2)
773{
774	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
775	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
776
777	if (v1 < v2)
778		return (-1);
779	if (v1 > v2)
780		return (1);
781
782	return (0);
783}
784
785void
786zil_add_block(zilog_t *zilog, const blkptr_t *bp)
787{
788	avl_tree_t *t = &zilog->zl_vdev_tree;
789	avl_index_t where;
790	zil_vdev_node_t *zv, zvsearch;
791	int ndvas = BP_GET_NDVAS(bp);
792	int i;
793
794	if (zfs_nocacheflush)
795		return;
796
797	ASSERT(zilog->zl_writer);
798
799	/*
800	 * Even though we're zl_writer, we still need a lock because the
801	 * zl_get_data() callbacks may have dmu_sync() done callbacks
802	 * that will run concurrently.
803	 */
804	mutex_enter(&zilog->zl_vdev_lock);
805	for (i = 0; i < ndvas; i++) {
806		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
807		if (avl_find(t, &zvsearch, &where) == NULL) {
808			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
809			zv->zv_vdev = zvsearch.zv_vdev;
810			avl_insert(t, zv, where);
811		}
812	}
813	mutex_exit(&zilog->zl_vdev_lock);
814}
815
816static void
817zil_flush_vdevs(zilog_t *zilog)
818{
819	spa_t *spa = zilog->zl_spa;
820	avl_tree_t *t = &zilog->zl_vdev_tree;
821	void *cookie = NULL;
822	zil_vdev_node_t *zv;
823	zio_t *zio;
824
825	ASSERT(zilog->zl_writer);
826
827	/*
828	 * We don't need zl_vdev_lock here because we're the zl_writer,
829	 * and all zl_get_data() callbacks are done.
830	 */
831	if (avl_numnodes(t) == 0)
832		return;
833
834	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
835
836	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
837
838	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
839		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
840		if (vd != NULL)
841			zio_flush(zio, vd);
842		kmem_free(zv, sizeof (*zv));
843	}
844
845	/*
846	 * Wait for all the flushes to complete.  Not all devices actually
847	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
848	 */
849	(void) zio_wait(zio);
850
851	spa_config_exit(spa, SCL_STATE, FTAG);
852}
853
854/*
855 * Function called when a log block write completes
856 */
857static void
858zil_lwb_write_done(zio_t *zio)
859{
860	lwb_t *lwb = zio->io_private;
861	zilog_t *zilog = lwb->lwb_zilog;
862	dmu_tx_t *tx = lwb->lwb_tx;
863
864	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
865	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
866	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
867	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
868	ASSERT(!BP_IS_GANG(zio->io_bp));
869	ASSERT(!BP_IS_HOLE(zio->io_bp));
870	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
871
872	/*
873	 * Ensure the lwb buffer pointer is cleared before releasing
874	 * the txg. If we have had an allocation failure and
875	 * the txg is waiting to sync then we want want zil_sync()
876	 * to remove the lwb so that it's not picked up as the next new
877	 * one in zil_commit_writer(). zil_sync() will only remove
878	 * the lwb if lwb_buf is null.
879	 */
880	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
881	mutex_enter(&zilog->zl_lock);
882	lwb->lwb_buf = NULL;
883	lwb->lwb_tx = NULL;
884	mutex_exit(&zilog->zl_lock);
885
886	/*
887	 * Now that we've written this log block, we have a stable pointer
888	 * to the next block in the chain, so it's OK to let the txg in
889	 * which we allocated the next block sync.
890	 */
891	dmu_tx_commit(tx);
892}
893
894/*
895 * Initialize the io for a log block.
896 */
897static void
898zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
899{
900	zbookmark_phys_t zb;
901
902	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
903	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
904	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
905
906	if (zilog->zl_root_zio == NULL) {
907		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
908		    ZIO_FLAG_CANFAIL);
909	}
910	if (lwb->lwb_zio == NULL) {
911		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
912		    0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
913		    zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
914		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
915	}
916}
917
918/*
919 * Define a limited set of intent log block sizes.
920 *
921 * These must be a multiple of 4KB. Note only the amount used (again
922 * aligned to 4KB) actually gets written. However, we can't always just
923 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
924 */
925uint64_t zil_block_buckets[] = {
926    4096,		/* non TX_WRITE */
927    8192+4096,		/* data base */
928    32*1024 + 4096, 	/* NFS writes */
929    UINT64_MAX
930};
931
932/*
933 * Use the slog as long as the logbias is 'latency' and the current commit size
934 * is less than the limit or the total list size is less than 2X the limit.
935 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
936 */
937uint64_t zil_slog_limit = 1024 * 1024;
938#define	USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
939	(((zilog)->zl_cur_used < zil_slog_limit) || \
940	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
941
942/*
943 * Start a log block write and advance to the next log block.
944 * Calls are serialized.
945 */
946static lwb_t *
947zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
948{
949	lwb_t *nlwb = NULL;
950	zil_chain_t *zilc;
951	spa_t *spa = zilog->zl_spa;
952	blkptr_t *bp;
953	dmu_tx_t *tx;
954	uint64_t txg;
955	uint64_t zil_blksz, wsz;
956	int i, error;
957
958	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
959		zilc = (zil_chain_t *)lwb->lwb_buf;
960		bp = &zilc->zc_next_blk;
961	} else {
962		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
963		bp = &zilc->zc_next_blk;
964	}
965
966	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
967
968	/*
969	 * Allocate the next block and save its address in this block
970	 * before writing it in order to establish the log chain.
971	 * Note that if the allocation of nlwb synced before we wrote
972	 * the block that points at it (lwb), we'd leak it if we crashed.
973	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
974	 * We dirty the dataset to ensure that zil_sync() will be called
975	 * to clean up in the event of allocation failure or I/O failure.
976	 */
977	tx = dmu_tx_create(zilog->zl_os);
978	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
979	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
980	txg = dmu_tx_get_txg(tx);
981
982	lwb->lwb_tx = tx;
983
984	/*
985	 * Log blocks are pre-allocated. Here we select the size of the next
986	 * block, based on size used in the last block.
987	 * - first find the smallest bucket that will fit the block from a
988	 *   limited set of block sizes. This is because it's faster to write
989	 *   blocks allocated from the same metaslab as they are adjacent or
990	 *   close.
991	 * - next find the maximum from the new suggested size and an array of
992	 *   previous sizes. This lessens a picket fence effect of wrongly
993	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
994	 *   requests.
995	 *
996	 * Note we only write what is used, but we can't just allocate
997	 * the maximum block size because we can exhaust the available
998	 * pool log space.
999	 */
1000	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1001	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1002		continue;
1003	zil_blksz = zil_block_buckets[i];
1004	if (zil_blksz == UINT64_MAX)
1005		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1006	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1007	for (i = 0; i < ZIL_PREV_BLKS; i++)
1008		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1009	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1010
1011	BP_ZERO(bp);
1012	/* pass the old blkptr in order to spread log blocks across devs */
1013	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
1014	    USE_SLOG(zilog));
1015	if (error == 0) {
1016		ASSERT3U(bp->blk_birth, ==, txg);
1017		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1018		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1019
1020		/*
1021		 * Allocate a new log write buffer (lwb).
1022		 */
1023		nlwb = zil_alloc_lwb(zilog, bp, txg);
1024
1025		/* Record the block for later vdev flushing */
1026		zil_add_block(zilog, &lwb->lwb_blk);
1027	}
1028
1029	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1030		/* For Slim ZIL only write what is used. */
1031		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1032		ASSERT3U(wsz, <=, lwb->lwb_sz);
1033		zio_shrink(lwb->lwb_zio, wsz);
1034
1035	} else {
1036		wsz = lwb->lwb_sz;
1037	}
1038
1039	zilc->zc_pad = 0;
1040	zilc->zc_nused = lwb->lwb_nused;
1041	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1042
1043	/*
1044	 * clear unused data for security
1045	 */
1046	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1047
1048	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1049
1050	/*
1051	 * If there was an allocation failure then nlwb will be null which
1052	 * forces a txg_wait_synced().
1053	 */
1054	return (nlwb);
1055}
1056
1057static lwb_t *
1058zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1059{
1060	lr_t *lrc = &itx->itx_lr; /* common log record */
1061	lr_write_t *lrw = (lr_write_t *)lrc;
1062	char *lr_buf;
1063	uint64_t txg = lrc->lrc_txg;
1064	uint64_t reclen = lrc->lrc_reclen;
1065	uint64_t dlen = 0;
1066
1067	if (lwb == NULL)
1068		return (NULL);
1069
1070	ASSERT(lwb->lwb_buf != NULL);
1071
1072	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1073		dlen = P2ROUNDUP_TYPED(
1074		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1075
1076	zilog->zl_cur_used += (reclen + dlen);
1077
1078	zil_lwb_write_init(zilog, lwb);
1079
1080	/*
1081	 * If this record won't fit in the current log block, start a new one.
1082	 */
1083	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1084		lwb = zil_lwb_write_start(zilog, lwb);
1085		if (lwb == NULL)
1086			return (NULL);
1087		zil_lwb_write_init(zilog, lwb);
1088		ASSERT(LWB_EMPTY(lwb));
1089		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1090			txg_wait_synced(zilog->zl_dmu_pool, txg);
1091			return (lwb);
1092		}
1093	}
1094
1095	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1096	bcopy(lrc, lr_buf, reclen);
1097	lrc = (lr_t *)lr_buf;
1098	lrw = (lr_write_t *)lrc;
1099
1100	/*
1101	 * If it's a write, fetch the data or get its blkptr as appropriate.
1102	 */
1103	if (lrc->lrc_txtype == TX_WRITE) {
1104		if (txg > spa_freeze_txg(zilog->zl_spa))
1105			txg_wait_synced(zilog->zl_dmu_pool, txg);
1106		if (itx->itx_wr_state != WR_COPIED) {
1107			char *dbuf;
1108			int error;
1109
1110			if (dlen) {
1111				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1112				dbuf = lr_buf + reclen;
1113				lrw->lr_common.lrc_reclen += dlen;
1114			} else {
1115				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1116				dbuf = NULL;
1117			}
1118			error = zilog->zl_get_data(
1119			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1120			if (error == EIO) {
1121				txg_wait_synced(zilog->zl_dmu_pool, txg);
1122				return (lwb);
1123			}
1124			if (error != 0) {
1125				ASSERT(error == ENOENT || error == EEXIST ||
1126				    error == EALREADY);
1127				return (lwb);
1128			}
1129		}
1130	}
1131
1132	/*
1133	 * We're actually making an entry, so update lrc_seq to be the
1134	 * log record sequence number.  Note that this is generally not
1135	 * equal to the itx sequence number because not all transactions
1136	 * are synchronous, and sometimes spa_sync() gets there first.
1137	 */
1138	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1139	lwb->lwb_nused += reclen + dlen;
1140	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1141	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1142	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1143
1144	return (lwb);
1145}
1146
1147itx_t *
1148zil_itx_create(uint64_t txtype, size_t lrsize)
1149{
1150	itx_t *itx;
1151
1152	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1153
1154	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1155	itx->itx_lr.lrc_txtype = txtype;
1156	itx->itx_lr.lrc_reclen = lrsize;
1157	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1158	itx->itx_lr.lrc_seq = 0;	/* defensive */
1159	itx->itx_sync = B_TRUE;		/* default is synchronous */
1160
1161	return (itx);
1162}
1163
1164void
1165zil_itx_destroy(itx_t *itx)
1166{
1167	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1168}
1169
1170/*
1171 * Free up the sync and async itxs. The itxs_t has already been detached
1172 * so no locks are needed.
1173 */
1174static void
1175zil_itxg_clean(itxs_t *itxs)
1176{
1177	itx_t *itx;
1178	list_t *list;
1179	avl_tree_t *t;
1180	void *cookie;
1181	itx_async_node_t *ian;
1182
1183	list = &itxs->i_sync_list;
1184	while ((itx = list_head(list)) != NULL) {
1185		list_remove(list, itx);
1186		kmem_free(itx, offsetof(itx_t, itx_lr) +
1187		    itx->itx_lr.lrc_reclen);
1188	}
1189
1190	cookie = NULL;
1191	t = &itxs->i_async_tree;
1192	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1193		list = &ian->ia_list;
1194		while ((itx = list_head(list)) != NULL) {
1195			list_remove(list, itx);
1196			kmem_free(itx, offsetof(itx_t, itx_lr) +
1197			    itx->itx_lr.lrc_reclen);
1198		}
1199		list_destroy(list);
1200		kmem_free(ian, sizeof (itx_async_node_t));
1201	}
1202	avl_destroy(t);
1203
1204	kmem_free(itxs, sizeof (itxs_t));
1205}
1206
1207static int
1208zil_aitx_compare(const void *x1, const void *x2)
1209{
1210	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1211	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1212
1213	if (o1 < o2)
1214		return (-1);
1215	if (o1 > o2)
1216		return (1);
1217
1218	return (0);
1219}
1220
1221/*
1222 * Remove all async itx with the given oid.
1223 */
1224static void
1225zil_remove_async(zilog_t *zilog, uint64_t oid)
1226{
1227	uint64_t otxg, txg;
1228	itx_async_node_t *ian;
1229	avl_tree_t *t;
1230	avl_index_t where;
1231	list_t clean_list;
1232	itx_t *itx;
1233
1234	ASSERT(oid != 0);
1235	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1236
1237	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1238		otxg = ZILTEST_TXG;
1239	else
1240		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1241
1242	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1243		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1244
1245		mutex_enter(&itxg->itxg_lock);
1246		if (itxg->itxg_txg != txg) {
1247			mutex_exit(&itxg->itxg_lock);
1248			continue;
1249		}
1250
1251		/*
1252		 * Locate the object node and append its list.
1253		 */
1254		t = &itxg->itxg_itxs->i_async_tree;
1255		ian = avl_find(t, &oid, &where);
1256		if (ian != NULL)
1257			list_move_tail(&clean_list, &ian->ia_list);
1258		mutex_exit(&itxg->itxg_lock);
1259	}
1260	while ((itx = list_head(&clean_list)) != NULL) {
1261		list_remove(&clean_list, itx);
1262		kmem_free(itx, offsetof(itx_t, itx_lr) +
1263		    itx->itx_lr.lrc_reclen);
1264	}
1265	list_destroy(&clean_list);
1266}
1267
1268void
1269zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1270{
1271	uint64_t txg;
1272	itxg_t *itxg;
1273	itxs_t *itxs, *clean = NULL;
1274
1275	/*
1276	 * Object ids can be re-instantiated in the next txg so
1277	 * remove any async transactions to avoid future leaks.
1278	 * This can happen if a fsync occurs on the re-instantiated
1279	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1280	 * the new file data and flushes a write record for the old object.
1281	 */
1282	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1283		zil_remove_async(zilog, itx->itx_oid);
1284
1285	/*
1286	 * Ensure the data of a renamed file is committed before the rename.
1287	 */
1288	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1289		zil_async_to_sync(zilog, itx->itx_oid);
1290
1291	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1292		txg = ZILTEST_TXG;
1293	else
1294		txg = dmu_tx_get_txg(tx);
1295
1296	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1297	mutex_enter(&itxg->itxg_lock);
1298	itxs = itxg->itxg_itxs;
1299	if (itxg->itxg_txg != txg) {
1300		if (itxs != NULL) {
1301			/*
1302			 * The zil_clean callback hasn't got around to cleaning
1303			 * this itxg. Save the itxs for release below.
1304			 * This should be rare.
1305			 */
1306			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1307			    "txg %llu", itxg->itxg_txg);
1308			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1309			itxg->itxg_sod = 0;
1310			clean = itxg->itxg_itxs;
1311		}
1312		ASSERT(itxg->itxg_sod == 0);
1313		itxg->itxg_txg = txg;
1314		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1315
1316		list_create(&itxs->i_sync_list, sizeof (itx_t),
1317		    offsetof(itx_t, itx_node));
1318		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1319		    sizeof (itx_async_node_t),
1320		    offsetof(itx_async_node_t, ia_node));
1321	}
1322	if (itx->itx_sync) {
1323		list_insert_tail(&itxs->i_sync_list, itx);
1324		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1325		itxg->itxg_sod += itx->itx_sod;
1326	} else {
1327		avl_tree_t *t = &itxs->i_async_tree;
1328		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1329		itx_async_node_t *ian;
1330		avl_index_t where;
1331
1332		ian = avl_find(t, &foid, &where);
1333		if (ian == NULL) {
1334			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1335			list_create(&ian->ia_list, sizeof (itx_t),
1336			    offsetof(itx_t, itx_node));
1337			ian->ia_foid = foid;
1338			avl_insert(t, ian, where);
1339		}
1340		list_insert_tail(&ian->ia_list, itx);
1341	}
1342
1343	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1344	zilog_dirty(zilog, txg);
1345	mutex_exit(&itxg->itxg_lock);
1346
1347	/* Release the old itxs now we've dropped the lock */
1348	if (clean != NULL)
1349		zil_itxg_clean(clean);
1350}
1351
1352/*
1353 * If there are any in-memory intent log transactions which have now been
1354 * synced then start up a taskq to free them. We should only do this after we
1355 * have written out the uberblocks (i.e. txg has been comitted) so that
1356 * don't inadvertently clean out in-memory log records that would be required
1357 * by zil_commit().
1358 */
1359void
1360zil_clean(zilog_t *zilog, uint64_t synced_txg)
1361{
1362	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1363	itxs_t *clean_me;
1364
1365	mutex_enter(&itxg->itxg_lock);
1366	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1367		mutex_exit(&itxg->itxg_lock);
1368		return;
1369	}
1370	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1371	ASSERT(itxg->itxg_txg != 0);
1372	ASSERT(zilog->zl_clean_taskq != NULL);
1373	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1374	itxg->itxg_sod = 0;
1375	clean_me = itxg->itxg_itxs;
1376	itxg->itxg_itxs = NULL;
1377	itxg->itxg_txg = 0;
1378	mutex_exit(&itxg->itxg_lock);
1379	/*
1380	 * Preferably start a task queue to free up the old itxs but
1381	 * if taskq_dispatch can't allocate resources to do that then
1382	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1383	 * created a bad performance problem.
1384	 */
1385	if (taskq_dispatch(zilog->zl_clean_taskq,
1386	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1387		zil_itxg_clean(clean_me);
1388}
1389
1390/*
1391 * Get the list of itxs to commit into zl_itx_commit_list.
1392 */
1393static void
1394zil_get_commit_list(zilog_t *zilog)
1395{
1396	uint64_t otxg, txg;
1397	list_t *commit_list = &zilog->zl_itx_commit_list;
1398	uint64_t push_sod = 0;
1399
1400	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1401		otxg = ZILTEST_TXG;
1402	else
1403		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1404
1405	/*
1406	 * This is inherently racy, since there is nothing to prevent
1407	 * the last synced txg from changing. That's okay since we'll
1408	 * only commit things in the future.
1409	 */
1410	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1411		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1412
1413		mutex_enter(&itxg->itxg_lock);
1414		if (itxg->itxg_txg != txg) {
1415			mutex_exit(&itxg->itxg_lock);
1416			continue;
1417		}
1418
1419		/*
1420		 * If we're adding itx records to the zl_itx_commit_list,
1421		 * then the zil better be dirty in this "txg". We can assert
1422		 * that here since we're holding the itxg_lock which will
1423		 * prevent spa_sync from cleaning it. Once we add the itxs
1424		 * to the zl_itx_commit_list we must commit it to disk even
1425		 * if it's unnecessary (i.e. the txg was synced).
1426		 */
1427		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1428		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1429		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1430		push_sod += itxg->itxg_sod;
1431		itxg->itxg_sod = 0;
1432
1433		mutex_exit(&itxg->itxg_lock);
1434	}
1435	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1436}
1437
1438/*
1439 * Move the async itxs for a specified object to commit into sync lists.
1440 */
1441static void
1442zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1443{
1444	uint64_t otxg, txg;
1445	itx_async_node_t *ian;
1446	avl_tree_t *t;
1447	avl_index_t where;
1448
1449	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1450		otxg = ZILTEST_TXG;
1451	else
1452		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1453
1454	/*
1455	 * This is inherently racy, since there is nothing to prevent
1456	 * the last synced txg from changing.
1457	 */
1458	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1459		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1460
1461		mutex_enter(&itxg->itxg_lock);
1462		if (itxg->itxg_txg != txg) {
1463			mutex_exit(&itxg->itxg_lock);
1464			continue;
1465		}
1466
1467		/*
1468		 * If a foid is specified then find that node and append its
1469		 * list. Otherwise walk the tree appending all the lists
1470		 * to the sync list. We add to the end rather than the
1471		 * beginning to ensure the create has happened.
1472		 */
1473		t = &itxg->itxg_itxs->i_async_tree;
1474		if (foid != 0) {
1475			ian = avl_find(t, &foid, &where);
1476			if (ian != NULL) {
1477				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1478				    &ian->ia_list);
1479			}
1480		} else {
1481			void *cookie = NULL;
1482
1483			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1484				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1485				    &ian->ia_list);
1486				list_destroy(&ian->ia_list);
1487				kmem_free(ian, sizeof (itx_async_node_t));
1488			}
1489		}
1490		mutex_exit(&itxg->itxg_lock);
1491	}
1492}
1493
1494static void
1495zil_commit_writer(zilog_t *zilog)
1496{
1497	uint64_t txg;
1498	itx_t *itx;
1499	lwb_t *lwb;
1500	spa_t *spa = zilog->zl_spa;
1501	int error = 0;
1502
1503	ASSERT(zilog->zl_root_zio == NULL);
1504
1505	mutex_exit(&zilog->zl_lock);
1506
1507	zil_get_commit_list(zilog);
1508
1509	/*
1510	 * Return if there's nothing to commit before we dirty the fs by
1511	 * calling zil_create().
1512	 */
1513	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1514		mutex_enter(&zilog->zl_lock);
1515		return;
1516	}
1517
1518	if (zilog->zl_suspend) {
1519		lwb = NULL;
1520	} else {
1521		lwb = list_tail(&zilog->zl_lwb_list);
1522		if (lwb == NULL)
1523			lwb = zil_create(zilog);
1524	}
1525
1526	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1527	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1528		txg = itx->itx_lr.lrc_txg;
1529		ASSERT3U(txg, !=, 0);
1530
1531		/*
1532		 * This is inherently racy and may result in us writing
1533		 * out a log block for a txg that was just synced. This is
1534		 * ok since we'll end cleaning up that log block the next
1535		 * time we call zil_sync().
1536		 */
1537		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1538			lwb = zil_lwb_commit(zilog, itx, lwb);
1539		list_remove(&zilog->zl_itx_commit_list, itx);
1540		kmem_free(itx, offsetof(itx_t, itx_lr)
1541		    + itx->itx_lr.lrc_reclen);
1542	}
1543	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1544
1545	/* write the last block out */
1546	if (lwb != NULL && lwb->lwb_zio != NULL)
1547		lwb = zil_lwb_write_start(zilog, lwb);
1548
1549	zilog->zl_cur_used = 0;
1550
1551	/*
1552	 * Wait if necessary for the log blocks to be on stable storage.
1553	 */
1554	if (zilog->zl_root_zio) {
1555		error = zio_wait(zilog->zl_root_zio);
1556		zilog->zl_root_zio = NULL;
1557		zil_flush_vdevs(zilog);
1558	}
1559
1560	if (error || lwb == NULL)
1561		txg_wait_synced(zilog->zl_dmu_pool, 0);
1562
1563	mutex_enter(&zilog->zl_lock);
1564
1565	/*
1566	 * Remember the highest committed log sequence number for ztest.
1567	 * We only update this value when all the log writes succeeded,
1568	 * because ztest wants to ASSERT that it got the whole log chain.
1569	 */
1570	if (error == 0 && lwb != NULL)
1571		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1572}
1573
1574/*
1575 * Commit zfs transactions to stable storage.
1576 * If foid is 0 push out all transactions, otherwise push only those
1577 * for that object or might reference that object.
1578 *
1579 * itxs are committed in batches. In a heavily stressed zil there will be
1580 * a commit writer thread who is writing out a bunch of itxs to the log
1581 * for a set of committing threads (cthreads) in the same batch as the writer.
1582 * Those cthreads are all waiting on the same cv for that batch.
1583 *
1584 * There will also be a different and growing batch of threads that are
1585 * waiting to commit (qthreads). When the committing batch completes
1586 * a transition occurs such that the cthreads exit and the qthreads become
1587 * cthreads. One of the new cthreads becomes the writer thread for the
1588 * batch. Any new threads arriving become new qthreads.
1589 *
1590 * Only 2 condition variables are needed and there's no transition
1591 * between the two cvs needed. They just flip-flop between qthreads
1592 * and cthreads.
1593 *
1594 * Using this scheme we can efficiently wakeup up only those threads
1595 * that have been committed.
1596 */
1597void
1598zil_commit(zilog_t *zilog, uint64_t foid)
1599{
1600	uint64_t mybatch;
1601
1602	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1603		return;
1604
1605	/* move the async itxs for the foid to the sync queues */
1606	zil_async_to_sync(zilog, foid);
1607
1608	mutex_enter(&zilog->zl_lock);
1609	mybatch = zilog->zl_next_batch;
1610	while (zilog->zl_writer) {
1611		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1612		if (mybatch <= zilog->zl_com_batch) {
1613			mutex_exit(&zilog->zl_lock);
1614			return;
1615		}
1616	}
1617
1618	zilog->zl_next_batch++;
1619	zilog->zl_writer = B_TRUE;
1620	zil_commit_writer(zilog);
1621	zilog->zl_com_batch = mybatch;
1622	zilog->zl_writer = B_FALSE;
1623	mutex_exit(&zilog->zl_lock);
1624
1625	/* wake up one thread to become the next writer */
1626	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1627
1628	/* wake up all threads waiting for this batch to be committed */
1629	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1630}
1631
1632/*
1633 * Called in syncing context to free committed log blocks and update log header.
1634 */
1635void
1636zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1637{
1638	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1639	uint64_t txg = dmu_tx_get_txg(tx);
1640	spa_t *spa = zilog->zl_spa;
1641	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1642	lwb_t *lwb;
1643
1644	/*
1645	 * We don't zero out zl_destroy_txg, so make sure we don't try
1646	 * to destroy it twice.
1647	 */
1648	if (spa_sync_pass(spa) != 1)
1649		return;
1650
1651	mutex_enter(&zilog->zl_lock);
1652
1653	ASSERT(zilog->zl_stop_sync == 0);
1654
1655	if (*replayed_seq != 0) {
1656		ASSERT(zh->zh_replay_seq < *replayed_seq);
1657		zh->zh_replay_seq = *replayed_seq;
1658		*replayed_seq = 0;
1659	}
1660
1661	if (zilog->zl_destroy_txg == txg) {
1662		blkptr_t blk = zh->zh_log;
1663
1664		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1665
1666		bzero(zh, sizeof (zil_header_t));
1667		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1668
1669		if (zilog->zl_keep_first) {
1670			/*
1671			 * If this block was part of log chain that couldn't
1672			 * be claimed because a device was missing during
1673			 * zil_claim(), but that device later returns,
1674			 * then this block could erroneously appear valid.
1675			 * To guard against this, assign a new GUID to the new
1676			 * log chain so it doesn't matter what blk points to.
1677			 */
1678			zil_init_log_chain(zilog, &blk);
1679			zh->zh_log = blk;
1680		}
1681	}
1682
1683	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1684		zh->zh_log = lwb->lwb_blk;
1685		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1686			break;
1687		list_remove(&zilog->zl_lwb_list, lwb);
1688		zio_free_zil(spa, txg, &lwb->lwb_blk);
1689		kmem_cache_free(zil_lwb_cache, lwb);
1690
1691		/*
1692		 * If we don't have anything left in the lwb list then
1693		 * we've had an allocation failure and we need to zero
1694		 * out the zil_header blkptr so that we don't end
1695		 * up freeing the same block twice.
1696		 */
1697		if (list_head(&zilog->zl_lwb_list) == NULL)
1698			BP_ZERO(&zh->zh_log);
1699	}
1700	mutex_exit(&zilog->zl_lock);
1701}
1702
1703void
1704zil_init(void)
1705{
1706	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1707	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1708}
1709
1710void
1711zil_fini(void)
1712{
1713	kmem_cache_destroy(zil_lwb_cache);
1714}
1715
1716void
1717zil_set_sync(zilog_t *zilog, uint64_t sync)
1718{
1719	zilog->zl_sync = sync;
1720}
1721
1722void
1723zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1724{
1725	zilog->zl_logbias = logbias;
1726}
1727
1728zilog_t *
1729zil_alloc(objset_t *os, zil_header_t *zh_phys)
1730{
1731	zilog_t *zilog;
1732
1733	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1734
1735	zilog->zl_header = zh_phys;
1736	zilog->zl_os = os;
1737	zilog->zl_spa = dmu_objset_spa(os);
1738	zilog->zl_dmu_pool = dmu_objset_pool(os);
1739	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1740	zilog->zl_logbias = dmu_objset_logbias(os);
1741	zilog->zl_sync = dmu_objset_syncprop(os);
1742	zilog->zl_next_batch = 1;
1743
1744	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1745
1746	for (int i = 0; i < TXG_SIZE; i++) {
1747		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1748		    MUTEX_DEFAULT, NULL);
1749	}
1750
1751	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1752	    offsetof(lwb_t, lwb_node));
1753
1754	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1755	    offsetof(itx_t, itx_node));
1756
1757	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1758
1759	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1760	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1761
1762	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1763	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1764	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1765	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1766
1767	return (zilog);
1768}
1769
1770void
1771zil_free(zilog_t *zilog)
1772{
1773	zilog->zl_stop_sync = 1;
1774
1775	ASSERT0(zilog->zl_suspend);
1776	ASSERT0(zilog->zl_suspending);
1777
1778	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1779	list_destroy(&zilog->zl_lwb_list);
1780
1781	avl_destroy(&zilog->zl_vdev_tree);
1782	mutex_destroy(&zilog->zl_vdev_lock);
1783
1784	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1785	list_destroy(&zilog->zl_itx_commit_list);
1786
1787	for (int i = 0; i < TXG_SIZE; i++) {
1788		/*
1789		 * It's possible for an itx to be generated that doesn't dirty
1790		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1791		 * callback to remove the entry. We remove those here.
1792		 *
1793		 * Also free up the ziltest itxs.
1794		 */
1795		if (zilog->zl_itxg[i].itxg_itxs)
1796			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1797		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1798	}
1799
1800	mutex_destroy(&zilog->zl_lock);
1801
1802	cv_destroy(&zilog->zl_cv_writer);
1803	cv_destroy(&zilog->zl_cv_suspend);
1804	cv_destroy(&zilog->zl_cv_batch[0]);
1805	cv_destroy(&zilog->zl_cv_batch[1]);
1806
1807	kmem_free(zilog, sizeof (zilog_t));
1808}
1809
1810/*
1811 * Open an intent log.
1812 */
1813zilog_t *
1814zil_open(objset_t *os, zil_get_data_t *get_data)
1815{
1816	zilog_t *zilog = dmu_objset_zil(os);
1817
1818	ASSERT(zilog->zl_clean_taskq == NULL);
1819	ASSERT(zilog->zl_get_data == NULL);
1820	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1821
1822	zilog->zl_get_data = get_data;
1823	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1824	    2, 2, TASKQ_PREPOPULATE);
1825
1826	return (zilog);
1827}
1828
1829/*
1830 * Close an intent log.
1831 */
1832void
1833zil_close(zilog_t *zilog)
1834{
1835	lwb_t *lwb;
1836	uint64_t txg = 0;
1837
1838	zil_commit(zilog, 0); /* commit all itx */
1839
1840	/*
1841	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1842	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1843	 * callbacks have occurred that may clean the zil.  Only then can we
1844	 * destroy the zl_clean_taskq.
1845	 */
1846	mutex_enter(&zilog->zl_lock);
1847	lwb = list_tail(&zilog->zl_lwb_list);
1848	if (lwb != NULL)
1849		txg = lwb->lwb_max_txg;
1850	mutex_exit(&zilog->zl_lock);
1851	if (txg)
1852		txg_wait_synced(zilog->zl_dmu_pool, txg);
1853
1854	if (zilog_is_dirty(zilog))
1855		zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1856	VERIFY(!zilog_is_dirty(zilog));
1857
1858	taskq_destroy(zilog->zl_clean_taskq);
1859	zilog->zl_clean_taskq = NULL;
1860	zilog->zl_get_data = NULL;
1861
1862	/*
1863	 * We should have only one LWB left on the list; remove it now.
1864	 */
1865	mutex_enter(&zilog->zl_lock);
1866	lwb = list_head(&zilog->zl_lwb_list);
1867	if (lwb != NULL) {
1868		ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1869		list_remove(&zilog->zl_lwb_list, lwb);
1870		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1871		kmem_cache_free(zil_lwb_cache, lwb);
1872	}
1873	mutex_exit(&zilog->zl_lock);
1874}
1875
1876static char *suspend_tag = "zil suspending";
1877
1878/*
1879 * Suspend an intent log.  While in suspended mode, we still honor
1880 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1881 * On old version pools, we suspend the log briefly when taking a
1882 * snapshot so that it will have an empty intent log.
1883 *
1884 * Long holds are not really intended to be used the way we do here --
1885 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
1886 * could fail.  Therefore we take pains to only put a long hold if it is
1887 * actually necessary.  Fortunately, it will only be necessary if the
1888 * objset is currently mounted (or the ZVOL equivalent).  In that case it
1889 * will already have a long hold, so we are not really making things any worse.
1890 *
1891 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1892 * zvol_state_t), and use their mechanism to prevent their hold from being
1893 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
1894 * very little gain.
1895 *
1896 * if cookiep == NULL, this does both the suspend & resume.
1897 * Otherwise, it returns with the dataset "long held", and the cookie
1898 * should be passed into zil_resume().
1899 */
1900int
1901zil_suspend(const char *osname, void **cookiep)
1902{
1903	objset_t *os;
1904	zilog_t *zilog;
1905	const zil_header_t *zh;
1906	int error;
1907
1908	error = dmu_objset_hold(osname, suspend_tag, &os);
1909	if (error != 0)
1910		return (error);
1911	zilog = dmu_objset_zil(os);
1912
1913	mutex_enter(&zilog->zl_lock);
1914	zh = zilog->zl_header;
1915
1916	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1917		mutex_exit(&zilog->zl_lock);
1918		dmu_objset_rele(os, suspend_tag);
1919		return (SET_ERROR(EBUSY));
1920	}
1921
1922	/*
1923	 * Don't put a long hold in the cases where we can avoid it.  This
1924	 * is when there is no cookie so we are doing a suspend & resume
1925	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1926	 * for the suspend because it's already suspended, or there's no ZIL.
1927	 */
1928	if (cookiep == NULL && !zilog->zl_suspending &&
1929	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1930		mutex_exit(&zilog->zl_lock);
1931		dmu_objset_rele(os, suspend_tag);
1932		return (0);
1933	}
1934
1935	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1936	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1937
1938	zilog->zl_suspend++;
1939
1940	if (zilog->zl_suspend > 1) {
1941		/*
1942		 * Someone else is already suspending it.
1943		 * Just wait for them to finish.
1944		 */
1945
1946		while (zilog->zl_suspending)
1947			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1948		mutex_exit(&zilog->zl_lock);
1949
1950		if (cookiep == NULL)
1951			zil_resume(os);
1952		else
1953			*cookiep = os;
1954		return (0);
1955	}
1956
1957	/*
1958	 * If there is no pointer to an on-disk block, this ZIL must not
1959	 * be active (e.g. filesystem not mounted), so there's nothing
1960	 * to clean up.
1961	 */
1962	if (BP_IS_HOLE(&zh->zh_log)) {
1963		ASSERT(cookiep != NULL); /* fast path already handled */
1964
1965		*cookiep = os;
1966		mutex_exit(&zilog->zl_lock);
1967		return (0);
1968	}
1969
1970	zilog->zl_suspending = B_TRUE;
1971	mutex_exit(&zilog->zl_lock);
1972
1973	zil_commit(zilog, 0);
1974
1975	zil_destroy(zilog, B_FALSE);
1976
1977	mutex_enter(&zilog->zl_lock);
1978	zilog->zl_suspending = B_FALSE;
1979	cv_broadcast(&zilog->zl_cv_suspend);
1980	mutex_exit(&zilog->zl_lock);
1981
1982	if (cookiep == NULL)
1983		zil_resume(os);
1984	else
1985		*cookiep = os;
1986	return (0);
1987}
1988
1989void
1990zil_resume(void *cookie)
1991{
1992	objset_t *os = cookie;
1993	zilog_t *zilog = dmu_objset_zil(os);
1994
1995	mutex_enter(&zilog->zl_lock);
1996	ASSERT(zilog->zl_suspend != 0);
1997	zilog->zl_suspend--;
1998	mutex_exit(&zilog->zl_lock);
1999	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2000	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2001}
2002
2003typedef struct zil_replay_arg {
2004	zil_replay_func_t **zr_replay;
2005	void		*zr_arg;
2006	boolean_t	zr_byteswap;
2007	char		*zr_lr;
2008} zil_replay_arg_t;
2009
2010static int
2011zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2012{
2013	char name[ZFS_MAX_DATASET_NAME_LEN];
2014
2015	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
2016
2017	dmu_objset_name(zilog->zl_os, name);
2018
2019	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2020	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2021	    (u_longlong_t)lr->lrc_seq,
2022	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2023	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
2024
2025	return (error);
2026}
2027
2028static int
2029zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2030{
2031	zil_replay_arg_t *zr = zra;
2032	const zil_header_t *zh = zilog->zl_header;
2033	uint64_t reclen = lr->lrc_reclen;
2034	uint64_t txtype = lr->lrc_txtype;
2035	int error = 0;
2036
2037	zilog->zl_replaying_seq = lr->lrc_seq;
2038
2039	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
2040		return (0);
2041
2042	if (lr->lrc_txg < claim_txg)		/* already committed */
2043		return (0);
2044
2045	/* Strip case-insensitive bit, still present in log record */
2046	txtype &= ~TX_CI;
2047
2048	if (txtype == 0 || txtype >= TX_MAX_TYPE)
2049		return (zil_replay_error(zilog, lr, EINVAL));
2050
2051	/*
2052	 * If this record type can be logged out of order, the object
2053	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
2054	 */
2055	if (TX_OOO(txtype)) {
2056		error = dmu_object_info(zilog->zl_os,
2057		    ((lr_ooo_t *)lr)->lr_foid, NULL);
2058		if (error == ENOENT || error == EEXIST)
2059			return (0);
2060	}
2061
2062	/*
2063	 * Make a copy of the data so we can revise and extend it.
2064	 */
2065	bcopy(lr, zr->zr_lr, reclen);
2066
2067	/*
2068	 * If this is a TX_WRITE with a blkptr, suck in the data.
2069	 */
2070	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2071		error = zil_read_log_data(zilog, (lr_write_t *)lr,
2072		    zr->zr_lr + reclen);
2073		if (error != 0)
2074			return (zil_replay_error(zilog, lr, error));
2075	}
2076
2077	/*
2078	 * The log block containing this lr may have been byteswapped
2079	 * so that we can easily examine common fields like lrc_txtype.
2080	 * However, the log is a mix of different record types, and only the
2081	 * replay vectors know how to byteswap their records.  Therefore, if
2082	 * the lr was byteswapped, undo it before invoking the replay vector.
2083	 */
2084	if (zr->zr_byteswap)
2085		byteswap_uint64_array(zr->zr_lr, reclen);
2086
2087	/*
2088	 * We must now do two things atomically: replay this log record,
2089	 * and update the log header sequence number to reflect the fact that
2090	 * we did so. At the end of each replay function the sequence number
2091	 * is updated if we are in replay mode.
2092	 */
2093	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2094	if (error != 0) {
2095		/*
2096		 * The DMU's dnode layer doesn't see removes until the txg
2097		 * commits, so a subsequent claim can spuriously fail with
2098		 * EEXIST. So if we receive any error we try syncing out
2099		 * any removes then retry the transaction.  Note that we
2100		 * specify B_FALSE for byteswap now, so we don't do it twice.
2101		 */
2102		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2103		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2104		if (error != 0)
2105			return (zil_replay_error(zilog, lr, error));
2106	}
2107	return (0);
2108}
2109
2110/* ARGSUSED */
2111static int
2112zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2113{
2114	zilog->zl_replay_blks++;
2115
2116	return (0);
2117}
2118
2119/*
2120 * If this dataset has a non-empty intent log, replay it and destroy it.
2121 */
2122void
2123zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2124{
2125	zilog_t *zilog = dmu_objset_zil(os);
2126	const zil_header_t *zh = zilog->zl_header;
2127	zil_replay_arg_t zr;
2128
2129	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2130		zil_destroy(zilog, B_TRUE);
2131		return;
2132	}
2133
2134	zr.zr_replay = replay_func;
2135	zr.zr_arg = arg;
2136	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2137	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2138
2139	/*
2140	 * Wait for in-progress removes to sync before starting replay.
2141	 */
2142	txg_wait_synced(zilog->zl_dmu_pool, 0);
2143
2144	zilog->zl_replay = B_TRUE;
2145	zilog->zl_replay_time = ddi_get_lbolt();
2146	ASSERT(zilog->zl_replay_blks == 0);
2147	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2148	    zh->zh_claim_txg);
2149	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2150
2151	zil_destroy(zilog, B_FALSE);
2152	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2153	zilog->zl_replay = B_FALSE;
2154}
2155
2156boolean_t
2157zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2158{
2159	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2160		return (B_TRUE);
2161
2162	if (zilog->zl_replay) {
2163		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2164		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2165		    zilog->zl_replaying_seq;
2166		return (B_TRUE);
2167	}
2168
2169	return (B_FALSE);
2170}
2171
2172/* ARGSUSED */
2173int
2174zil_vdev_offline(const char *osname, void *arg)
2175{
2176	int error;
2177
2178	error = zil_suspend(osname, NULL);
2179	if (error != 0)
2180		return (SET_ERROR(EEXIST));
2181	return (0);
2182}
2183