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) 2013 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 = BP_IDENTITY(bp);
141	zil_bp_node_t *zn;
142	avl_index_t where;
143
144	if (avl_find(t, dva, &where) != NULL)
145		return (SET_ERROR(EEXIST));
146
147	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
148	zn->zn_dva = *dva;
149	avl_insert(t, zn, where);
150
151	return (0);
152}
153
154static zil_header_t *
155zil_header_in_syncing_context(zilog_t *zilog)
156{
157	return ((zil_header_t *)zilog->zl_header);
158}
159
160static void
161zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
162{
163	zio_cksum_t *zc = &bp->blk_cksum;
164
165	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
166	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
167	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
168	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
169}
170
171/*
172 * Read a log block and make sure it's valid.
173 */
174static int
175zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
176    char **end)
177{
178	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
179	uint32_t aflags = ARC_WAIT;
180	arc_buf_t *abuf = NULL;
181	zbookmark_t zb;
182	int error;
183
184	if (zilog->zl_header->zh_claim_txg == 0)
185		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
186
187	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
188		zio_flags |= ZIO_FLAG_SPECULATIVE;
189
190	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
191	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
192
193	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
194	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
195
196	if (error == 0) {
197		zio_cksum_t cksum = bp->blk_cksum;
198
199		/*
200		 * Validate the checksummed log block.
201		 *
202		 * Sequence numbers should be... sequential.  The checksum
203		 * verifier for the next block should be bp's checksum plus 1.
204		 *
205		 * Also check the log chain linkage and size used.
206		 */
207		cksum.zc_word[ZIL_ZC_SEQ]++;
208
209		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
210			zil_chain_t *zilc = abuf->b_data;
211			char *lr = (char *)(zilc + 1);
212			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
213
214			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
215			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
216				error = SET_ERROR(ECKSUM);
217			} else {
218				bcopy(lr, dst, len);
219				*end = (char *)dst + len;
220				*nbp = zilc->zc_next_blk;
221			}
222		} else {
223			char *lr = abuf->b_data;
224			uint64_t size = BP_GET_LSIZE(bp);
225			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
226
227			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
228			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
229			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
230				error = SET_ERROR(ECKSUM);
231			} else {
232				bcopy(lr, dst, zilc->zc_nused);
233				*end = (char *)dst + zilc->zc_nused;
234				*nbp = zilc->zc_next_blk;
235			}
236		}
237
238		VERIFY(arc_buf_remove_ref(abuf, &abuf));
239	}
240
241	return (error);
242}
243
244/*
245 * Read a TX_WRITE log data block.
246 */
247static int
248zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
249{
250	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
251	const blkptr_t *bp = &lr->lr_blkptr;
252	uint32_t aflags = ARC_WAIT;
253	arc_buf_t *abuf = NULL;
254	zbookmark_t zb;
255	int error;
256
257	if (BP_IS_HOLE(bp)) {
258		if (wbuf != NULL)
259			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
260		return (0);
261	}
262
263	if (zilog->zl_header->zh_claim_txg == 0)
264		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
265
266	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
267	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
268
269	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
270	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
271
272	if (error == 0) {
273		if (wbuf != NULL)
274			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
275		(void) arc_buf_remove_ref(abuf, &abuf);
276	}
277
278	return (error);
279}
280
281/*
282 * Parse the intent log, and call parse_func for each valid record within.
283 */
284int
285zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
286    zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
287{
288	const zil_header_t *zh = zilog->zl_header;
289	boolean_t claimed = !!zh->zh_claim_txg;
290	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
291	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
292	uint64_t max_blk_seq = 0;
293	uint64_t max_lr_seq = 0;
294	uint64_t blk_count = 0;
295	uint64_t lr_count = 0;
296	blkptr_t blk, next_blk;
297	char *lrbuf, *lrp;
298	int error = 0;
299
300	/*
301	 * Old logs didn't record the maximum zh_claim_lr_seq.
302	 */
303	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
304		claim_lr_seq = UINT64_MAX;
305
306	/*
307	 * Starting at the block pointed to by zh_log we read the log chain.
308	 * For each block in the chain we strongly check that block to
309	 * ensure its validity.  We stop when an invalid block is found.
310	 * For each block pointer in the chain we call parse_blk_func().
311	 * For each record in each valid block we call parse_lr_func().
312	 * If the log has been claimed, stop if we encounter a sequence
313	 * number greater than the highest claimed sequence number.
314	 */
315	lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
316	zil_bp_tree_init(zilog);
317
318	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
319		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
320		int reclen;
321		char *end;
322
323		if (blk_seq > claim_blk_seq)
324			break;
325		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
326			break;
327		ASSERT3U(max_blk_seq, <, blk_seq);
328		max_blk_seq = blk_seq;
329		blk_count++;
330
331		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
332			break;
333
334		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
335		if (error != 0)
336			break;
337
338		for (lrp = lrbuf; lrp < end; lrp += reclen) {
339			lr_t *lr = (lr_t *)lrp;
340			reclen = lr->lrc_reclen;
341			ASSERT3U(reclen, >=, sizeof (lr_t));
342			if (lr->lrc_seq > claim_lr_seq)
343				goto done;
344			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
345				goto done;
346			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
347			max_lr_seq = lr->lrc_seq;
348			lr_count++;
349		}
350	}
351done:
352	zilog->zl_parse_error = error;
353	zilog->zl_parse_blk_seq = max_blk_seq;
354	zilog->zl_parse_lr_seq = max_lr_seq;
355	zilog->zl_parse_blk_count = blk_count;
356	zilog->zl_parse_lr_count = lr_count;
357
358	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
359	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
360
361	zil_bp_tree_fini(zilog);
362	zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
363
364	return (error);
365}
366
367static int
368zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
369{
370	/*
371	 * Claim log block if not already committed and not already claimed.
372	 * If tx == NULL, just verify that the block is claimable.
373	 */
374	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
375	    zil_bp_tree_add(zilog, bp) != 0)
376		return (0);
377
378	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
379	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
380	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
381}
382
383static int
384zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
385{
386	lr_write_t *lr = (lr_write_t *)lrc;
387	int error;
388
389	if (lrc->lrc_txtype != TX_WRITE)
390		return (0);
391
392	/*
393	 * If the block is not readable, don't claim it.  This can happen
394	 * in normal operation when a log block is written to disk before
395	 * some of the dmu_sync() blocks it points to.  In this case, the
396	 * transaction cannot have been committed to anyone (we would have
397	 * waited for all writes to be stable first), so it is semantically
398	 * correct to declare this the end of the log.
399	 */
400	if (lr->lr_blkptr.blk_birth >= first_txg &&
401	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
402		return (error);
403	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
404}
405
406/* ARGSUSED */
407static int
408zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
409{
410	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
411
412	return (0);
413}
414
415static int
416zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
417{
418	lr_write_t *lr = (lr_write_t *)lrc;
419	blkptr_t *bp = &lr->lr_blkptr;
420
421	/*
422	 * If we previously claimed it, we need to free it.
423	 */
424	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
425	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
426	    !BP_IS_HOLE(bp))
427		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
428
429	return (0);
430}
431
432static lwb_t *
433zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
434{
435	lwb_t *lwb;
436
437	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
438	lwb->lwb_zilog = zilog;
439	lwb->lwb_blk = *bp;
440	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
441	lwb->lwb_max_txg = txg;
442	lwb->lwb_zio = NULL;
443	lwb->lwb_tx = NULL;
444	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
445		lwb->lwb_nused = sizeof (zil_chain_t);
446		lwb->lwb_sz = BP_GET_LSIZE(bp);
447	} else {
448		lwb->lwb_nused = 0;
449		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
450	}
451
452	mutex_enter(&zilog->zl_lock);
453	list_insert_tail(&zilog->zl_lwb_list, lwb);
454	mutex_exit(&zilog->zl_lock);
455
456	return (lwb);
457}
458
459/*
460 * Called when we create in-memory log transactions so that we know
461 * to cleanup the itxs at the end of spa_sync().
462 */
463void
464zilog_dirty(zilog_t *zilog, uint64_t txg)
465{
466	dsl_pool_t *dp = zilog->zl_dmu_pool;
467	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
468
469	if (dsl_dataset_is_snapshot(ds))
470		panic("dirtying snapshot!");
471
472	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
473		/* up the hold count until we can be written out */
474		dmu_buf_add_ref(ds->ds_dbuf, zilog);
475	}
476}
477
478boolean_t
479zilog_is_dirty(zilog_t *zilog)
480{
481	dsl_pool_t *dp = zilog->zl_dmu_pool;
482
483	for (int t = 0; t < TXG_SIZE; t++) {
484		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
485			return (B_TRUE);
486	}
487	return (B_FALSE);
488}
489
490/*
491 * Create an on-disk intent log.
492 */
493static lwb_t *
494zil_create(zilog_t *zilog)
495{
496	const zil_header_t *zh = zilog->zl_header;
497	lwb_t *lwb = NULL;
498	uint64_t txg = 0;
499	dmu_tx_t *tx = NULL;
500	blkptr_t blk;
501	int error = 0;
502
503	/*
504	 * Wait for any previous destroy to complete.
505	 */
506	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
507
508	ASSERT(zh->zh_claim_txg == 0);
509	ASSERT(zh->zh_replay_seq == 0);
510
511	blk = zh->zh_log;
512
513	/*
514	 * Allocate an initial log block if:
515	 *    - there isn't one already
516	 *    - the existing block is the wrong endianess
517	 */
518	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
519		tx = dmu_tx_create(zilog->zl_os);
520		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
521		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
522		txg = dmu_tx_get_txg(tx);
523
524		if (!BP_IS_HOLE(&blk)) {
525			zio_free_zil(zilog->zl_spa, txg, &blk);
526			BP_ZERO(&blk);
527		}
528
529		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
530		    ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
531
532		if (error == 0)
533			zil_init_log_chain(zilog, &blk);
534	}
535
536	/*
537	 * Allocate a log write buffer (lwb) for the first log block.
538	 */
539	if (error == 0)
540		lwb = zil_alloc_lwb(zilog, &blk, txg);
541
542	/*
543	 * If we just allocated the first log block, commit our transaction
544	 * and wait for zil_sync() to stuff the block poiner into zh_log.
545	 * (zh is part of the MOS, so we cannot modify it in open context.)
546	 */
547	if (tx != NULL) {
548		dmu_tx_commit(tx);
549		txg_wait_synced(zilog->zl_dmu_pool, txg);
550	}
551
552	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
553
554	return (lwb);
555}
556
557/*
558 * In one tx, free all log blocks and clear the log header.
559 * If keep_first is set, then we're replaying a log with no content.
560 * We want to keep the first block, however, so that the first
561 * synchronous transaction doesn't require a txg_wait_synced()
562 * in zil_create().  We don't need to txg_wait_synced() here either
563 * when keep_first is set, because both zil_create() and zil_destroy()
564 * will wait for any in-progress destroys to complete.
565 */
566void
567zil_destroy(zilog_t *zilog, boolean_t keep_first)
568{
569	const zil_header_t *zh = zilog->zl_header;
570	lwb_t *lwb;
571	dmu_tx_t *tx;
572	uint64_t txg;
573
574	/*
575	 * Wait for any previous destroy to complete.
576	 */
577	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
578
579	zilog->zl_old_header = *zh;		/* debugging aid */
580
581	if (BP_IS_HOLE(&zh->zh_log))
582		return;
583
584	tx = dmu_tx_create(zilog->zl_os);
585	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
586	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
587	txg = dmu_tx_get_txg(tx);
588
589	mutex_enter(&zilog->zl_lock);
590
591	ASSERT3U(zilog->zl_destroy_txg, <, txg);
592	zilog->zl_destroy_txg = txg;
593	zilog->zl_keep_first = keep_first;
594
595	if (!list_is_empty(&zilog->zl_lwb_list)) {
596		ASSERT(zh->zh_claim_txg == 0);
597		VERIFY(!keep_first);
598		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
599			list_remove(&zilog->zl_lwb_list, lwb);
600			if (lwb->lwb_buf != NULL)
601				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
602			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
603			kmem_cache_free(zil_lwb_cache, lwb);
604		}
605	} else if (!keep_first) {
606		zil_destroy_sync(zilog, tx);
607	}
608	mutex_exit(&zilog->zl_lock);
609
610	dmu_tx_commit(tx);
611}
612
613void
614zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
615{
616	ASSERT(list_is_empty(&zilog->zl_lwb_list));
617	(void) zil_parse(zilog, zil_free_log_block,
618	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
619}
620
621int
622zil_claim(const char *osname, void *txarg)
623{
624	dmu_tx_t *tx = txarg;
625	uint64_t first_txg = dmu_tx_get_txg(tx);
626	zilog_t *zilog;
627	zil_header_t *zh;
628	objset_t *os;
629	int error;
630
631	error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
632	if (error != 0) {
633		cmn_err(CE_WARN, "can't open objset for %s", osname);
634		return (0);
635	}
636
637	zilog = dmu_objset_zil(os);
638	zh = zil_header_in_syncing_context(zilog);
639
640	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
641		if (!BP_IS_HOLE(&zh->zh_log))
642			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
643		BP_ZERO(&zh->zh_log);
644		dsl_dataset_dirty(dmu_objset_ds(os), tx);
645		dmu_objset_disown(os, FTAG);
646		return (0);
647	}
648
649	/*
650	 * Claim all log blocks if we haven't already done so, and remember
651	 * the highest claimed sequence number.  This ensures that if we can
652	 * read only part of the log now (e.g. due to a missing device),
653	 * but we can read the entire log later, we will not try to replay
654	 * or destroy beyond the last block we successfully claimed.
655	 */
656	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
657	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
658		(void) zil_parse(zilog, zil_claim_log_block,
659		    zil_claim_log_record, tx, first_txg);
660		zh->zh_claim_txg = first_txg;
661		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
662		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
663		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
664			zh->zh_flags |= ZIL_REPLAY_NEEDED;
665		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
666		dsl_dataset_dirty(dmu_objset_ds(os), tx);
667	}
668
669	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
670	dmu_objset_disown(os, FTAG);
671	return (0);
672}
673
674/*
675 * Check the log by walking the log chain.
676 * Checksum errors are ok as they indicate the end of the chain.
677 * Any other error (no device or read failure) returns an error.
678 */
679int
680zil_check_log_chain(const char *osname, void *tx)
681{
682	zilog_t *zilog;
683	objset_t *os;
684	blkptr_t *bp;
685	int error;
686
687	ASSERT(tx == NULL);
688
689	error = dmu_objset_hold(osname, FTAG, &os);
690	if (error != 0) {
691		cmn_err(CE_WARN, "can't open objset for %s", osname);
692		return (0);
693	}
694
695	zilog = dmu_objset_zil(os);
696	bp = (blkptr_t *)&zilog->zl_header->zh_log;
697
698	/*
699	 * Check the first block and determine if it's on a log device
700	 * which may have been removed or faulted prior to loading this
701	 * pool.  If so, there's no point in checking the rest of the log
702	 * as its content should have already been synced to the pool.
703	 */
704	if (!BP_IS_HOLE(bp)) {
705		vdev_t *vd;
706		boolean_t valid = B_TRUE;
707
708		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
709		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
710		if (vd->vdev_islog && vdev_is_dead(vd))
711			valid = vdev_log_state_valid(vd);
712		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
713
714		if (!valid) {
715			dmu_objset_rele(os, FTAG);
716			return (0);
717		}
718	}
719
720	/*
721	 * Because tx == NULL, zil_claim_log_block() will not actually claim
722	 * any blocks, but just determine whether it is possible to do so.
723	 * In addition to checking the log chain, zil_claim_log_block()
724	 * will invoke zio_claim() with a done func of spa_claim_notify(),
725	 * which will update spa_max_claim_txg.  See spa_load() for details.
726	 */
727	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
728	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
729
730	dmu_objset_rele(os, FTAG);
731
732	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
733}
734
735static int
736zil_vdev_compare(const void *x1, const void *x2)
737{
738	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
739	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
740
741	if (v1 < v2)
742		return (-1);
743	if (v1 > v2)
744		return (1);
745
746	return (0);
747}
748
749void
750zil_add_block(zilog_t *zilog, const blkptr_t *bp)
751{
752	avl_tree_t *t = &zilog->zl_vdev_tree;
753	avl_index_t where;
754	zil_vdev_node_t *zv, zvsearch;
755	int ndvas = BP_GET_NDVAS(bp);
756	int i;
757
758	if (zfs_nocacheflush)
759		return;
760
761	ASSERT(zilog->zl_writer);
762
763	/*
764	 * Even though we're zl_writer, we still need a lock because the
765	 * zl_get_data() callbacks may have dmu_sync() done callbacks
766	 * that will run concurrently.
767	 */
768	mutex_enter(&zilog->zl_vdev_lock);
769	for (i = 0; i < ndvas; i++) {
770		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
771		if (avl_find(t, &zvsearch, &where) == NULL) {
772			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
773			zv->zv_vdev = zvsearch.zv_vdev;
774			avl_insert(t, zv, where);
775		}
776	}
777	mutex_exit(&zilog->zl_vdev_lock);
778}
779
780static void
781zil_flush_vdevs(zilog_t *zilog)
782{
783	spa_t *spa = zilog->zl_spa;
784	avl_tree_t *t = &zilog->zl_vdev_tree;
785	void *cookie = NULL;
786	zil_vdev_node_t *zv;
787	zio_t *zio;
788
789	ASSERT(zilog->zl_writer);
790
791	/*
792	 * We don't need zl_vdev_lock here because we're the zl_writer,
793	 * and all zl_get_data() callbacks are done.
794	 */
795	if (avl_numnodes(t) == 0)
796		return;
797
798	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
799
800	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
801
802	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
803		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
804		if (vd != NULL)
805			zio_flush(zio, vd);
806		kmem_free(zv, sizeof (*zv));
807	}
808
809	/*
810	 * Wait for all the flushes to complete.  Not all devices actually
811	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
812	 */
813	(void) zio_wait(zio);
814
815	spa_config_exit(spa, SCL_STATE, FTAG);
816}
817
818/*
819 * Function called when a log block write completes
820 */
821static void
822zil_lwb_write_done(zio_t *zio)
823{
824	lwb_t *lwb = zio->io_private;
825	zilog_t *zilog = lwb->lwb_zilog;
826	dmu_tx_t *tx = lwb->lwb_tx;
827
828	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
829	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
830	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
831	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
832	ASSERT(!BP_IS_GANG(zio->io_bp));
833	ASSERT(!BP_IS_HOLE(zio->io_bp));
834	ASSERT(zio->io_bp->blk_fill == 0);
835
836	/*
837	 * Ensure the lwb buffer pointer is cleared before releasing
838	 * the txg. If we have had an allocation failure and
839	 * the txg is waiting to sync then we want want zil_sync()
840	 * to remove the lwb so that it's not picked up as the next new
841	 * one in zil_commit_writer(). zil_sync() will only remove
842	 * the lwb if lwb_buf is null.
843	 */
844	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
845	mutex_enter(&zilog->zl_lock);
846	lwb->lwb_buf = NULL;
847	lwb->lwb_tx = NULL;
848	mutex_exit(&zilog->zl_lock);
849
850	/*
851	 * Now that we've written this log block, we have a stable pointer
852	 * to the next block in the chain, so it's OK to let the txg in
853	 * which we allocated the next block sync.
854	 */
855	dmu_tx_commit(tx);
856}
857
858/*
859 * Initialize the io for a log block.
860 */
861static void
862zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
863{
864	zbookmark_t zb;
865
866	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
867	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
868	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
869
870	if (zilog->zl_root_zio == NULL) {
871		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
872		    ZIO_FLAG_CANFAIL);
873	}
874	if (lwb->lwb_zio == NULL) {
875		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
876		    0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
877		    zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
878		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
879	}
880}
881
882/*
883 * Define a limited set of intent log block sizes.
884 *
885 * These must be a multiple of 4KB. Note only the amount used (again
886 * aligned to 4KB) actually gets written. However, we can't always just
887 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
888 */
889uint64_t zil_block_buckets[] = {
890    4096,		/* non TX_WRITE */
891    8192+4096,		/* data base */
892    32*1024 + 4096, 	/* NFS writes */
893    UINT64_MAX
894};
895
896/*
897 * Use the slog as long as the logbias is 'latency' and the current commit size
898 * is less than the limit or the total list size is less than 2X the limit.
899 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
900 */
901uint64_t zil_slog_limit = 1024 * 1024;
902#define	USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
903	(((zilog)->zl_cur_used < zil_slog_limit) || \
904	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
905
906/*
907 * Start a log block write and advance to the next log block.
908 * Calls are serialized.
909 */
910static lwb_t *
911zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
912{
913	lwb_t *nlwb = NULL;
914	zil_chain_t *zilc;
915	spa_t *spa = zilog->zl_spa;
916	blkptr_t *bp;
917	dmu_tx_t *tx;
918	uint64_t txg;
919	uint64_t zil_blksz, wsz;
920	int i, error;
921
922	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
923		zilc = (zil_chain_t *)lwb->lwb_buf;
924		bp = &zilc->zc_next_blk;
925	} else {
926		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
927		bp = &zilc->zc_next_blk;
928	}
929
930	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
931
932	/*
933	 * Allocate the next block and save its address in this block
934	 * before writing it in order to establish the log chain.
935	 * Note that if the allocation of nlwb synced before we wrote
936	 * the block that points at it (lwb), we'd leak it if we crashed.
937	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
938	 * We dirty the dataset to ensure that zil_sync() will be called
939	 * to clean up in the event of allocation failure or I/O failure.
940	 */
941	tx = dmu_tx_create(zilog->zl_os);
942	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
943	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
944	txg = dmu_tx_get_txg(tx);
945
946	lwb->lwb_tx = tx;
947
948	/*
949	 * Log blocks are pre-allocated. Here we select the size of the next
950	 * block, based on size used in the last block.
951	 * - first find the smallest bucket that will fit the block from a
952	 *   limited set of block sizes. This is because it's faster to write
953	 *   blocks allocated from the same metaslab as they are adjacent or
954	 *   close.
955	 * - next find the maximum from the new suggested size and an array of
956	 *   previous sizes. This lessens a picket fence effect of wrongly
957	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
958	 *   requests.
959	 *
960	 * Note we only write what is used, but we can't just allocate
961	 * the maximum block size because we can exhaust the available
962	 * pool log space.
963	 */
964	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
965	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
966		continue;
967	zil_blksz = zil_block_buckets[i];
968	if (zil_blksz == UINT64_MAX)
969		zil_blksz = SPA_MAXBLOCKSIZE;
970	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
971	for (i = 0; i < ZIL_PREV_BLKS; i++)
972		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
973	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
974
975	BP_ZERO(bp);
976	/* pass the old blkptr in order to spread log blocks across devs */
977	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
978	    USE_SLOG(zilog));
979	if (error == 0) {
980		ASSERT3U(bp->blk_birth, ==, txg);
981		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
982		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
983
984		/*
985		 * Allocate a new log write buffer (lwb).
986		 */
987		nlwb = zil_alloc_lwb(zilog, bp, txg);
988
989		/* Record the block for later vdev flushing */
990		zil_add_block(zilog, &lwb->lwb_blk);
991	}
992
993	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
994		/* For Slim ZIL only write what is used. */
995		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
996		ASSERT3U(wsz, <=, lwb->lwb_sz);
997		zio_shrink(lwb->lwb_zio, wsz);
998
999	} else {
1000		wsz = lwb->lwb_sz;
1001	}
1002
1003	zilc->zc_pad = 0;
1004	zilc->zc_nused = lwb->lwb_nused;
1005	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1006
1007	/*
1008	 * clear unused data for security
1009	 */
1010	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1011
1012	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1013
1014	/*
1015	 * If there was an allocation failure then nlwb will be null which
1016	 * forces a txg_wait_synced().
1017	 */
1018	return (nlwb);
1019}
1020
1021static lwb_t *
1022zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1023{
1024	lr_t *lrc = &itx->itx_lr; /* common log record */
1025	lr_write_t *lrw = (lr_write_t *)lrc;
1026	char *lr_buf;
1027	uint64_t txg = lrc->lrc_txg;
1028	uint64_t reclen = lrc->lrc_reclen;
1029	uint64_t dlen = 0;
1030
1031	if (lwb == NULL)
1032		return (NULL);
1033
1034	ASSERT(lwb->lwb_buf != NULL);
1035	ASSERT(zilog_is_dirty(zilog) ||
1036	    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1037
1038	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1039		dlen = P2ROUNDUP_TYPED(
1040		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1041
1042	zilog->zl_cur_used += (reclen + dlen);
1043
1044	zil_lwb_write_init(zilog, lwb);
1045
1046	/*
1047	 * If this record won't fit in the current log block, start a new one.
1048	 */
1049	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1050		lwb = zil_lwb_write_start(zilog, lwb);
1051		if (lwb == NULL)
1052			return (NULL);
1053		zil_lwb_write_init(zilog, lwb);
1054		ASSERT(LWB_EMPTY(lwb));
1055		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1056			txg_wait_synced(zilog->zl_dmu_pool, txg);
1057			return (lwb);
1058		}
1059	}
1060
1061	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1062	bcopy(lrc, lr_buf, reclen);
1063	lrc = (lr_t *)lr_buf;
1064	lrw = (lr_write_t *)lrc;
1065
1066	/*
1067	 * If it's a write, fetch the data or get its blkptr as appropriate.
1068	 */
1069	if (lrc->lrc_txtype == TX_WRITE) {
1070		if (txg > spa_freeze_txg(zilog->zl_spa))
1071			txg_wait_synced(zilog->zl_dmu_pool, txg);
1072		if (itx->itx_wr_state != WR_COPIED) {
1073			char *dbuf;
1074			int error;
1075
1076			if (dlen) {
1077				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1078				dbuf = lr_buf + reclen;
1079				lrw->lr_common.lrc_reclen += dlen;
1080			} else {
1081				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1082				dbuf = NULL;
1083			}
1084			error = zilog->zl_get_data(
1085			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1086			if (error == EIO) {
1087				txg_wait_synced(zilog->zl_dmu_pool, txg);
1088				return (lwb);
1089			}
1090			if (error != 0) {
1091				ASSERT(error == ENOENT || error == EEXIST ||
1092				    error == EALREADY);
1093				return (lwb);
1094			}
1095		}
1096	}
1097
1098	/*
1099	 * We're actually making an entry, so update lrc_seq to be the
1100	 * log record sequence number.  Note that this is generally not
1101	 * equal to the itx sequence number because not all transactions
1102	 * are synchronous, and sometimes spa_sync() gets there first.
1103	 */
1104	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1105	lwb->lwb_nused += reclen + dlen;
1106	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1107	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1108	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1109
1110	return (lwb);
1111}
1112
1113itx_t *
1114zil_itx_create(uint64_t txtype, size_t lrsize)
1115{
1116	itx_t *itx;
1117
1118	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1119
1120	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1121	itx->itx_lr.lrc_txtype = txtype;
1122	itx->itx_lr.lrc_reclen = lrsize;
1123	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1124	itx->itx_lr.lrc_seq = 0;	/* defensive */
1125	itx->itx_sync = B_TRUE;		/* default is synchronous */
1126
1127	return (itx);
1128}
1129
1130void
1131zil_itx_destroy(itx_t *itx)
1132{
1133	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1134}
1135
1136/*
1137 * Free up the sync and async itxs. The itxs_t has already been detached
1138 * so no locks are needed.
1139 */
1140static void
1141zil_itxg_clean(itxs_t *itxs)
1142{
1143	itx_t *itx;
1144	list_t *list;
1145	avl_tree_t *t;
1146	void *cookie;
1147	itx_async_node_t *ian;
1148
1149	list = &itxs->i_sync_list;
1150	while ((itx = list_head(list)) != NULL) {
1151		list_remove(list, itx);
1152		kmem_free(itx, offsetof(itx_t, itx_lr) +
1153		    itx->itx_lr.lrc_reclen);
1154	}
1155
1156	cookie = NULL;
1157	t = &itxs->i_async_tree;
1158	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1159		list = &ian->ia_list;
1160		while ((itx = list_head(list)) != NULL) {
1161			list_remove(list, itx);
1162			kmem_free(itx, offsetof(itx_t, itx_lr) +
1163			    itx->itx_lr.lrc_reclen);
1164		}
1165		list_destroy(list);
1166		kmem_free(ian, sizeof (itx_async_node_t));
1167	}
1168	avl_destroy(t);
1169
1170	kmem_free(itxs, sizeof (itxs_t));
1171}
1172
1173static int
1174zil_aitx_compare(const void *x1, const void *x2)
1175{
1176	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1177	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1178
1179	if (o1 < o2)
1180		return (-1);
1181	if (o1 > o2)
1182		return (1);
1183
1184	return (0);
1185}
1186
1187/*
1188 * Remove all async itx with the given oid.
1189 */
1190static void
1191zil_remove_async(zilog_t *zilog, uint64_t oid)
1192{
1193	uint64_t otxg, txg;
1194	itx_async_node_t *ian;
1195	avl_tree_t *t;
1196	avl_index_t where;
1197	list_t clean_list;
1198	itx_t *itx;
1199
1200	ASSERT(oid != 0);
1201	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1202
1203	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1204		otxg = ZILTEST_TXG;
1205	else
1206		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1207
1208	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1209		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1210
1211		mutex_enter(&itxg->itxg_lock);
1212		if (itxg->itxg_txg != txg) {
1213			mutex_exit(&itxg->itxg_lock);
1214			continue;
1215		}
1216
1217		/*
1218		 * Locate the object node and append its list.
1219		 */
1220		t = &itxg->itxg_itxs->i_async_tree;
1221		ian = avl_find(t, &oid, &where);
1222		if (ian != NULL)
1223			list_move_tail(&clean_list, &ian->ia_list);
1224		mutex_exit(&itxg->itxg_lock);
1225	}
1226	while ((itx = list_head(&clean_list)) != NULL) {
1227		list_remove(&clean_list, itx);
1228		kmem_free(itx, offsetof(itx_t, itx_lr) +
1229		    itx->itx_lr.lrc_reclen);
1230	}
1231	list_destroy(&clean_list);
1232}
1233
1234void
1235zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1236{
1237	uint64_t txg;
1238	itxg_t *itxg;
1239	itxs_t *itxs, *clean = NULL;
1240
1241	/*
1242	 * Object ids can be re-instantiated in the next txg so
1243	 * remove any async transactions to avoid future leaks.
1244	 * This can happen if a fsync occurs on the re-instantiated
1245	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1246	 * the new file data and flushes a write record for the old object.
1247	 */
1248	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1249		zil_remove_async(zilog, itx->itx_oid);
1250
1251	/*
1252	 * Ensure the data of a renamed file is committed before the rename.
1253	 */
1254	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1255		zil_async_to_sync(zilog, itx->itx_oid);
1256
1257	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1258		txg = ZILTEST_TXG;
1259	else
1260		txg = dmu_tx_get_txg(tx);
1261
1262	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1263	mutex_enter(&itxg->itxg_lock);
1264	itxs = itxg->itxg_itxs;
1265	if (itxg->itxg_txg != txg) {
1266		if (itxs != NULL) {
1267			/*
1268			 * The zil_clean callback hasn't got around to cleaning
1269			 * this itxg. Save the itxs for release below.
1270			 * This should be rare.
1271			 */
1272			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1273			itxg->itxg_sod = 0;
1274			clean = itxg->itxg_itxs;
1275		}
1276		ASSERT(itxg->itxg_sod == 0);
1277		itxg->itxg_txg = txg;
1278		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1279
1280		list_create(&itxs->i_sync_list, sizeof (itx_t),
1281		    offsetof(itx_t, itx_node));
1282		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1283		    sizeof (itx_async_node_t),
1284		    offsetof(itx_async_node_t, ia_node));
1285	}
1286	if (itx->itx_sync) {
1287		list_insert_tail(&itxs->i_sync_list, itx);
1288		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1289		itxg->itxg_sod += itx->itx_sod;
1290	} else {
1291		avl_tree_t *t = &itxs->i_async_tree;
1292		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1293		itx_async_node_t *ian;
1294		avl_index_t where;
1295
1296		ian = avl_find(t, &foid, &where);
1297		if (ian == NULL) {
1298			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1299			list_create(&ian->ia_list, sizeof (itx_t),
1300			    offsetof(itx_t, itx_node));
1301			ian->ia_foid = foid;
1302			avl_insert(t, ian, where);
1303		}
1304		list_insert_tail(&ian->ia_list, itx);
1305	}
1306
1307	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1308	zilog_dirty(zilog, txg);
1309	mutex_exit(&itxg->itxg_lock);
1310
1311	/* Release the old itxs now we've dropped the lock */
1312	if (clean != NULL)
1313		zil_itxg_clean(clean);
1314}
1315
1316/*
1317 * If there are any in-memory intent log transactions which have now been
1318 * synced then start up a taskq to free them. We should only do this after we
1319 * have written out the uberblocks (i.e. txg has been comitted) so that
1320 * don't inadvertently clean out in-memory log records that would be required
1321 * by zil_commit().
1322 */
1323void
1324zil_clean(zilog_t *zilog, uint64_t synced_txg)
1325{
1326	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1327	itxs_t *clean_me;
1328
1329	mutex_enter(&itxg->itxg_lock);
1330	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1331		mutex_exit(&itxg->itxg_lock);
1332		return;
1333	}
1334	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1335	ASSERT(itxg->itxg_txg != 0);
1336	ASSERT(zilog->zl_clean_taskq != NULL);
1337	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1338	itxg->itxg_sod = 0;
1339	clean_me = itxg->itxg_itxs;
1340	itxg->itxg_itxs = NULL;
1341	itxg->itxg_txg = 0;
1342	mutex_exit(&itxg->itxg_lock);
1343	/*
1344	 * Preferably start a task queue to free up the old itxs but
1345	 * if taskq_dispatch can't allocate resources to do that then
1346	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1347	 * created a bad performance problem.
1348	 */
1349	if (taskq_dispatch(zilog->zl_clean_taskq,
1350	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1351		zil_itxg_clean(clean_me);
1352}
1353
1354/*
1355 * Get the list of itxs to commit into zl_itx_commit_list.
1356 */
1357static void
1358zil_get_commit_list(zilog_t *zilog)
1359{
1360	uint64_t otxg, txg;
1361	list_t *commit_list = &zilog->zl_itx_commit_list;
1362	uint64_t push_sod = 0;
1363
1364	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1365		otxg = ZILTEST_TXG;
1366	else
1367		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1368
1369	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1370		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1371
1372		mutex_enter(&itxg->itxg_lock);
1373		if (itxg->itxg_txg != txg) {
1374			mutex_exit(&itxg->itxg_lock);
1375			continue;
1376		}
1377
1378		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1379		push_sod += itxg->itxg_sod;
1380		itxg->itxg_sod = 0;
1381
1382		mutex_exit(&itxg->itxg_lock);
1383	}
1384	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1385}
1386
1387/*
1388 * Move the async itxs for a specified object to commit into sync lists.
1389 */
1390static void
1391zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1392{
1393	uint64_t otxg, txg;
1394	itx_async_node_t *ian;
1395	avl_tree_t *t;
1396	avl_index_t where;
1397
1398	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1399		otxg = ZILTEST_TXG;
1400	else
1401		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1402
1403	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1404		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1405
1406		mutex_enter(&itxg->itxg_lock);
1407		if (itxg->itxg_txg != txg) {
1408			mutex_exit(&itxg->itxg_lock);
1409			continue;
1410		}
1411
1412		/*
1413		 * If a foid is specified then find that node and append its
1414		 * list. Otherwise walk the tree appending all the lists
1415		 * to the sync list. We add to the end rather than the
1416		 * beginning to ensure the create has happened.
1417		 */
1418		t = &itxg->itxg_itxs->i_async_tree;
1419		if (foid != 0) {
1420			ian = avl_find(t, &foid, &where);
1421			if (ian != NULL) {
1422				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1423				    &ian->ia_list);
1424			}
1425		} else {
1426			void *cookie = NULL;
1427
1428			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1429				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1430				    &ian->ia_list);
1431				list_destroy(&ian->ia_list);
1432				kmem_free(ian, sizeof (itx_async_node_t));
1433			}
1434		}
1435		mutex_exit(&itxg->itxg_lock);
1436	}
1437}
1438
1439static void
1440zil_commit_writer(zilog_t *zilog)
1441{
1442	uint64_t txg;
1443	itx_t *itx;
1444	lwb_t *lwb;
1445	spa_t *spa = zilog->zl_spa;
1446	int error = 0;
1447
1448	ASSERT(zilog->zl_root_zio == NULL);
1449
1450	mutex_exit(&zilog->zl_lock);
1451
1452	zil_get_commit_list(zilog);
1453
1454	/*
1455	 * Return if there's nothing to commit before we dirty the fs by
1456	 * calling zil_create().
1457	 */
1458	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1459		mutex_enter(&zilog->zl_lock);
1460		return;
1461	}
1462
1463	if (zilog->zl_suspend) {
1464		lwb = NULL;
1465	} else {
1466		lwb = list_tail(&zilog->zl_lwb_list);
1467		if (lwb == NULL)
1468			lwb = zil_create(zilog);
1469	}
1470
1471	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1472	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1473		txg = itx->itx_lr.lrc_txg;
1474		ASSERT(txg);
1475
1476		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1477			lwb = zil_lwb_commit(zilog, itx, lwb);
1478		list_remove(&zilog->zl_itx_commit_list, itx);
1479		kmem_free(itx, offsetof(itx_t, itx_lr)
1480		    + itx->itx_lr.lrc_reclen);
1481	}
1482	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1483
1484	/* write the last block out */
1485	if (lwb != NULL && lwb->lwb_zio != NULL)
1486		lwb = zil_lwb_write_start(zilog, lwb);
1487
1488	zilog->zl_cur_used = 0;
1489
1490	/*
1491	 * Wait if necessary for the log blocks to be on stable storage.
1492	 */
1493	if (zilog->zl_root_zio) {
1494		error = zio_wait(zilog->zl_root_zio);
1495		zilog->zl_root_zio = NULL;
1496		zil_flush_vdevs(zilog);
1497	}
1498
1499	if (error || lwb == NULL)
1500		txg_wait_synced(zilog->zl_dmu_pool, 0);
1501
1502	mutex_enter(&zilog->zl_lock);
1503
1504	/*
1505	 * Remember the highest committed log sequence number for ztest.
1506	 * We only update this value when all the log writes succeeded,
1507	 * because ztest wants to ASSERT that it got the whole log chain.
1508	 */
1509	if (error == 0 && lwb != NULL)
1510		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1511}
1512
1513/*
1514 * Commit zfs transactions to stable storage.
1515 * If foid is 0 push out all transactions, otherwise push only those
1516 * for that object or might reference that object.
1517 *
1518 * itxs are committed in batches. In a heavily stressed zil there will be
1519 * a commit writer thread who is writing out a bunch of itxs to the log
1520 * for a set of committing threads (cthreads) in the same batch as the writer.
1521 * Those cthreads are all waiting on the same cv for that batch.
1522 *
1523 * There will also be a different and growing batch of threads that are
1524 * waiting to commit (qthreads). When the committing batch completes
1525 * a transition occurs such that the cthreads exit and the qthreads become
1526 * cthreads. One of the new cthreads becomes the writer thread for the
1527 * batch. Any new threads arriving become new qthreads.
1528 *
1529 * Only 2 condition variables are needed and there's no transition
1530 * between the two cvs needed. They just flip-flop between qthreads
1531 * and cthreads.
1532 *
1533 * Using this scheme we can efficiently wakeup up only those threads
1534 * that have been committed.
1535 */
1536void
1537zil_commit(zilog_t *zilog, uint64_t foid)
1538{
1539	uint64_t mybatch;
1540
1541	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1542		return;
1543
1544	/* move the async itxs for the foid to the sync queues */
1545	zil_async_to_sync(zilog, foid);
1546
1547	mutex_enter(&zilog->zl_lock);
1548	mybatch = zilog->zl_next_batch;
1549	while (zilog->zl_writer) {
1550		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1551		if (mybatch <= zilog->zl_com_batch) {
1552			mutex_exit(&zilog->zl_lock);
1553			return;
1554		}
1555	}
1556
1557	zilog->zl_next_batch++;
1558	zilog->zl_writer = B_TRUE;
1559	zil_commit_writer(zilog);
1560	zilog->zl_com_batch = mybatch;
1561	zilog->zl_writer = B_FALSE;
1562	mutex_exit(&zilog->zl_lock);
1563
1564	/* wake up one thread to become the next writer */
1565	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1566
1567	/* wake up all threads waiting for this batch to be committed */
1568	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1569}
1570
1571/*
1572 * Called in syncing context to free committed log blocks and update log header.
1573 */
1574void
1575zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1576{
1577	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1578	uint64_t txg = dmu_tx_get_txg(tx);
1579	spa_t *spa = zilog->zl_spa;
1580	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1581	lwb_t *lwb;
1582
1583	/*
1584	 * We don't zero out zl_destroy_txg, so make sure we don't try
1585	 * to destroy it twice.
1586	 */
1587	if (spa_sync_pass(spa) != 1)
1588		return;
1589
1590	mutex_enter(&zilog->zl_lock);
1591
1592	ASSERT(zilog->zl_stop_sync == 0);
1593
1594	if (*replayed_seq != 0) {
1595		ASSERT(zh->zh_replay_seq < *replayed_seq);
1596		zh->zh_replay_seq = *replayed_seq;
1597		*replayed_seq = 0;
1598	}
1599
1600	if (zilog->zl_destroy_txg == txg) {
1601		blkptr_t blk = zh->zh_log;
1602
1603		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1604
1605		bzero(zh, sizeof (zil_header_t));
1606		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1607
1608		if (zilog->zl_keep_first) {
1609			/*
1610			 * If this block was part of log chain that couldn't
1611			 * be claimed because a device was missing during
1612			 * zil_claim(), but that device later returns,
1613			 * then this block could erroneously appear valid.
1614			 * To guard against this, assign a new GUID to the new
1615			 * log chain so it doesn't matter what blk points to.
1616			 */
1617			zil_init_log_chain(zilog, &blk);
1618			zh->zh_log = blk;
1619		}
1620	}
1621
1622	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1623		zh->zh_log = lwb->lwb_blk;
1624		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1625			break;
1626		list_remove(&zilog->zl_lwb_list, lwb);
1627		zio_free_zil(spa, txg, &lwb->lwb_blk);
1628		kmem_cache_free(zil_lwb_cache, lwb);
1629
1630		/*
1631		 * If we don't have anything left in the lwb list then
1632		 * we've had an allocation failure and we need to zero
1633		 * out the zil_header blkptr so that we don't end
1634		 * up freeing the same block twice.
1635		 */
1636		if (list_head(&zilog->zl_lwb_list) == NULL)
1637			BP_ZERO(&zh->zh_log);
1638	}
1639	mutex_exit(&zilog->zl_lock);
1640}
1641
1642void
1643zil_init(void)
1644{
1645	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1646	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1647}
1648
1649void
1650zil_fini(void)
1651{
1652	kmem_cache_destroy(zil_lwb_cache);
1653}
1654
1655void
1656zil_set_sync(zilog_t *zilog, uint64_t sync)
1657{
1658	zilog->zl_sync = sync;
1659}
1660
1661void
1662zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1663{
1664	zilog->zl_logbias = logbias;
1665}
1666
1667zilog_t *
1668zil_alloc(objset_t *os, zil_header_t *zh_phys)
1669{
1670	zilog_t *zilog;
1671
1672	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1673
1674	zilog->zl_header = zh_phys;
1675	zilog->zl_os = os;
1676	zilog->zl_spa = dmu_objset_spa(os);
1677	zilog->zl_dmu_pool = dmu_objset_pool(os);
1678	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1679	zilog->zl_logbias = dmu_objset_logbias(os);
1680	zilog->zl_sync = dmu_objset_syncprop(os);
1681	zilog->zl_next_batch = 1;
1682
1683	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1684
1685	for (int i = 0; i < TXG_SIZE; i++) {
1686		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1687		    MUTEX_DEFAULT, NULL);
1688	}
1689
1690	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1691	    offsetof(lwb_t, lwb_node));
1692
1693	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1694	    offsetof(itx_t, itx_node));
1695
1696	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1697
1698	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1699	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1700
1701	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1702	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1703	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1704	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1705
1706	return (zilog);
1707}
1708
1709void
1710zil_free(zilog_t *zilog)
1711{
1712	zilog->zl_stop_sync = 1;
1713
1714	ASSERT0(zilog->zl_suspend);
1715	ASSERT0(zilog->zl_suspending);
1716
1717	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1718	list_destroy(&zilog->zl_lwb_list);
1719
1720	avl_destroy(&zilog->zl_vdev_tree);
1721	mutex_destroy(&zilog->zl_vdev_lock);
1722
1723	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1724	list_destroy(&zilog->zl_itx_commit_list);
1725
1726	for (int i = 0; i < TXG_SIZE; i++) {
1727		/*
1728		 * It's possible for an itx to be generated that doesn't dirty
1729		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1730		 * callback to remove the entry. We remove those here.
1731		 *
1732		 * Also free up the ziltest itxs.
1733		 */
1734		if (zilog->zl_itxg[i].itxg_itxs)
1735			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1736		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1737	}
1738
1739	mutex_destroy(&zilog->zl_lock);
1740
1741	cv_destroy(&zilog->zl_cv_writer);
1742	cv_destroy(&zilog->zl_cv_suspend);
1743	cv_destroy(&zilog->zl_cv_batch[0]);
1744	cv_destroy(&zilog->zl_cv_batch[1]);
1745
1746	kmem_free(zilog, sizeof (zilog_t));
1747}
1748
1749/*
1750 * Open an intent log.
1751 */
1752zilog_t *
1753zil_open(objset_t *os, zil_get_data_t *get_data)
1754{
1755	zilog_t *zilog = dmu_objset_zil(os);
1756
1757	ASSERT(zilog->zl_clean_taskq == NULL);
1758	ASSERT(zilog->zl_get_data == NULL);
1759	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1760
1761	zilog->zl_get_data = get_data;
1762	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1763	    2, 2, TASKQ_PREPOPULATE);
1764
1765	return (zilog);
1766}
1767
1768/*
1769 * Close an intent log.
1770 */
1771void
1772zil_close(zilog_t *zilog)
1773{
1774	lwb_t *lwb;
1775	uint64_t txg = 0;
1776
1777	zil_commit(zilog, 0); /* commit all itx */
1778
1779	/*
1780	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1781	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1782	 * callbacks have occurred that may clean the zil.  Only then can we
1783	 * destroy the zl_clean_taskq.
1784	 */
1785	mutex_enter(&zilog->zl_lock);
1786	lwb = list_tail(&zilog->zl_lwb_list);
1787	if (lwb != NULL)
1788		txg = lwb->lwb_max_txg;
1789	mutex_exit(&zilog->zl_lock);
1790	if (txg)
1791		txg_wait_synced(zilog->zl_dmu_pool, txg);
1792	ASSERT(!zilog_is_dirty(zilog));
1793
1794	taskq_destroy(zilog->zl_clean_taskq);
1795	zilog->zl_clean_taskq = NULL;
1796	zilog->zl_get_data = NULL;
1797
1798	/*
1799	 * We should have only one LWB left on the list; remove it now.
1800	 */
1801	mutex_enter(&zilog->zl_lock);
1802	lwb = list_head(&zilog->zl_lwb_list);
1803	if (lwb != NULL) {
1804		ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1805		list_remove(&zilog->zl_lwb_list, lwb);
1806		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1807		kmem_cache_free(zil_lwb_cache, lwb);
1808	}
1809	mutex_exit(&zilog->zl_lock);
1810}
1811
1812static char *suspend_tag = "zil suspending";
1813
1814/*
1815 * Suspend an intent log.  While in suspended mode, we still honor
1816 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1817 * On old version pools, we suspend the log briefly when taking a
1818 * snapshot so that it will have an empty intent log.
1819 *
1820 * Long holds are not really intended to be used the way we do here --
1821 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
1822 * could fail.  Therefore we take pains to only put a long hold if it is
1823 * actually necessary.  Fortunately, it will only be necessary if the
1824 * objset is currently mounted (or the ZVOL equivalent).  In that case it
1825 * will already have a long hold, so we are not really making things any worse.
1826 *
1827 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1828 * zvol_state_t), and use their mechanism to prevent their hold from being
1829 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
1830 * very little gain.
1831 *
1832 * if cookiep == NULL, this does both the suspend & resume.
1833 * Otherwise, it returns with the dataset "long held", and the cookie
1834 * should be passed into zil_resume().
1835 */
1836int
1837zil_suspend(const char *osname, void **cookiep)
1838{
1839	objset_t *os;
1840	zilog_t *zilog;
1841	const zil_header_t *zh;
1842	int error;
1843
1844	error = dmu_objset_hold(osname, suspend_tag, &os);
1845	if (error != 0)
1846		return (error);
1847	zilog = dmu_objset_zil(os);
1848
1849	mutex_enter(&zilog->zl_lock);
1850	zh = zilog->zl_header;
1851
1852	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1853		mutex_exit(&zilog->zl_lock);
1854		dmu_objset_rele(os, suspend_tag);
1855		return (SET_ERROR(EBUSY));
1856	}
1857
1858	/*
1859	 * Don't put a long hold in the cases where we can avoid it.  This
1860	 * is when there is no cookie so we are doing a suspend & resume
1861	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1862	 * for the suspend because it's already suspended, or there's no ZIL.
1863	 */
1864	if (cookiep == NULL && !zilog->zl_suspending &&
1865	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1866		mutex_exit(&zilog->zl_lock);
1867		dmu_objset_rele(os, suspend_tag);
1868		return (0);
1869	}
1870
1871	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1872	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1873
1874	zilog->zl_suspend++;
1875
1876	if (zilog->zl_suspend > 1) {
1877		/*
1878		 * Someone else is already suspending it.
1879		 * Just wait for them to finish.
1880		 */
1881
1882		while (zilog->zl_suspending)
1883			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1884		mutex_exit(&zilog->zl_lock);
1885
1886		if (cookiep == NULL)
1887			zil_resume(os);
1888		else
1889			*cookiep = os;
1890		return (0);
1891	}
1892
1893	/*
1894	 * If there is no pointer to an on-disk block, this ZIL must not
1895	 * be active (e.g. filesystem not mounted), so there's nothing
1896	 * to clean up.
1897	 */
1898	if (BP_IS_HOLE(&zh->zh_log)) {
1899		ASSERT(cookiep != NULL); /* fast path already handled */
1900
1901		*cookiep = os;
1902		mutex_exit(&zilog->zl_lock);
1903		return (0);
1904	}
1905
1906	zilog->zl_suspending = B_TRUE;
1907	mutex_exit(&zilog->zl_lock);
1908
1909	zil_commit(zilog, 0);
1910
1911	zil_destroy(zilog, B_FALSE);
1912
1913	mutex_enter(&zilog->zl_lock);
1914	zilog->zl_suspending = B_FALSE;
1915	cv_broadcast(&zilog->zl_cv_suspend);
1916	mutex_exit(&zilog->zl_lock);
1917
1918	if (cookiep == NULL)
1919		zil_resume(os);
1920	else
1921		*cookiep = os;
1922	return (0);
1923}
1924
1925void
1926zil_resume(void *cookie)
1927{
1928	objset_t *os = cookie;
1929	zilog_t *zilog = dmu_objset_zil(os);
1930
1931	mutex_enter(&zilog->zl_lock);
1932	ASSERT(zilog->zl_suspend != 0);
1933	zilog->zl_suspend--;
1934	mutex_exit(&zilog->zl_lock);
1935	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1936	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1937}
1938
1939typedef struct zil_replay_arg {
1940	zil_replay_func_t **zr_replay;
1941	void		*zr_arg;
1942	boolean_t	zr_byteswap;
1943	char		*zr_lr;
1944} zil_replay_arg_t;
1945
1946static int
1947zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1948{
1949	char name[MAXNAMELEN];
1950
1951	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
1952
1953	dmu_objset_name(zilog->zl_os, name);
1954
1955	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1956	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1957	    (u_longlong_t)lr->lrc_seq,
1958	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1959	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1960
1961	return (error);
1962}
1963
1964static int
1965zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1966{
1967	zil_replay_arg_t *zr = zra;
1968	const zil_header_t *zh = zilog->zl_header;
1969	uint64_t reclen = lr->lrc_reclen;
1970	uint64_t txtype = lr->lrc_txtype;
1971	int error = 0;
1972
1973	zilog->zl_replaying_seq = lr->lrc_seq;
1974
1975	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1976		return (0);
1977
1978	if (lr->lrc_txg < claim_txg)		/* already committed */
1979		return (0);
1980
1981	/* Strip case-insensitive bit, still present in log record */
1982	txtype &= ~TX_CI;
1983
1984	if (txtype == 0 || txtype >= TX_MAX_TYPE)
1985		return (zil_replay_error(zilog, lr, EINVAL));
1986
1987	/*
1988	 * If this record type can be logged out of order, the object
1989	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
1990	 */
1991	if (TX_OOO(txtype)) {
1992		error = dmu_object_info(zilog->zl_os,
1993		    ((lr_ooo_t *)lr)->lr_foid, NULL);
1994		if (error == ENOENT || error == EEXIST)
1995			return (0);
1996	}
1997
1998	/*
1999	 * Make a copy of the data so we can revise and extend it.
2000	 */
2001	bcopy(lr, zr->zr_lr, reclen);
2002
2003	/*
2004	 * If this is a TX_WRITE with a blkptr, suck in the data.
2005	 */
2006	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2007		error = zil_read_log_data(zilog, (lr_write_t *)lr,
2008		    zr->zr_lr + reclen);
2009		if (error != 0)
2010			return (zil_replay_error(zilog, lr, error));
2011	}
2012
2013	/*
2014	 * The log block containing this lr may have been byteswapped
2015	 * so that we can easily examine common fields like lrc_txtype.
2016	 * However, the log is a mix of different record types, and only the
2017	 * replay vectors know how to byteswap their records.  Therefore, if
2018	 * the lr was byteswapped, undo it before invoking the replay vector.
2019	 */
2020	if (zr->zr_byteswap)
2021		byteswap_uint64_array(zr->zr_lr, reclen);
2022
2023	/*
2024	 * We must now do two things atomically: replay this log record,
2025	 * and update the log header sequence number to reflect the fact that
2026	 * we did so. At the end of each replay function the sequence number
2027	 * is updated if we are in replay mode.
2028	 */
2029	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2030	if (error != 0) {
2031		/*
2032		 * The DMU's dnode layer doesn't see removes until the txg
2033		 * commits, so a subsequent claim can spuriously fail with
2034		 * EEXIST. So if we receive any error we try syncing out
2035		 * any removes then retry the transaction.  Note that we
2036		 * specify B_FALSE for byteswap now, so we don't do it twice.
2037		 */
2038		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2039		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2040		if (error != 0)
2041			return (zil_replay_error(zilog, lr, error));
2042	}
2043	return (0);
2044}
2045
2046/* ARGSUSED */
2047static int
2048zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2049{
2050	zilog->zl_replay_blks++;
2051
2052	return (0);
2053}
2054
2055/*
2056 * If this dataset has a non-empty intent log, replay it and destroy it.
2057 */
2058void
2059zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2060{
2061	zilog_t *zilog = dmu_objset_zil(os);
2062	const zil_header_t *zh = zilog->zl_header;
2063	zil_replay_arg_t zr;
2064
2065	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2066		zil_destroy(zilog, B_TRUE);
2067		return;
2068	}
2069
2070	zr.zr_replay = replay_func;
2071	zr.zr_arg = arg;
2072	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2073	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2074
2075	/*
2076	 * Wait for in-progress removes to sync before starting replay.
2077	 */
2078	txg_wait_synced(zilog->zl_dmu_pool, 0);
2079
2080	zilog->zl_replay = B_TRUE;
2081	zilog->zl_replay_time = ddi_get_lbolt();
2082	ASSERT(zilog->zl_replay_blks == 0);
2083	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2084	    zh->zh_claim_txg);
2085	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2086
2087	zil_destroy(zilog, B_FALSE);
2088	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2089	zilog->zl_replay = B_FALSE;
2090}
2091
2092boolean_t
2093zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2094{
2095	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2096		return (B_TRUE);
2097
2098	if (zilog->zl_replay) {
2099		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2100		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2101		    zilog->zl_replaying_seq;
2102		return (B_TRUE);
2103	}
2104
2105	return (B_FALSE);
2106}
2107
2108/* ARGSUSED */
2109int
2110zil_vdev_offline(const char *osname, void *arg)
2111{
2112	int error;
2113
2114	error = zil_suspend(osname, NULL);
2115	if (error != 0)
2116		return (SET_ERROR(EEXIST));
2117	return (0);
2118}
2119