1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 */
25
26/* Portions Copyright 2010 Robert Milkowski */
27
28#include <sys/zfs_context.h>
29#include <sys/spa.h>
30#include <sys/dmu.h>
31#include <sys/zap.h>
32#include <sys/arc.h>
33#include <sys/stat.h>
34#include <sys/resource.h>
35#include <sys/zil.h>
36#include <sys/zil_impl.h>
37#include <sys/dsl_dataset.h>
38#include <sys/vdev_impl.h>
39#include <sys/dmu_tx.h>
40#include <sys/dsl_pool.h>
41
42/*
43 * The zfs intent log (ZIL) saves transaction records of system calls
44 * that change the file system in memory with enough information
45 * to be able to replay them. These are stored in memory until
46 * either the DMU transaction group (txg) commits them to the stable pool
47 * and they can be discarded, or they are flushed to the stable log
48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49 * requirement. In the event of a panic or power fail then those log
50 * records (transactions) are replayed.
51 *
52 * There is one ZIL per file system. Its on-disk (pool) format consists
53 * of 3 parts:
54 *
55 * 	- ZIL header
56 * 	- ZIL blocks
57 * 	- ZIL records
58 *
59 * A log record holds a system call transaction. Log blocks can
60 * hold many log records and the blocks are chained together.
61 * Each ZIL block contains a block pointer (blkptr_t) to the next
62 * ZIL block in the chain. The ZIL header points to the first
63 * block in the chain. Note there is not a fixed place in the pool
64 * to hold blocks. They are dynamically allocated and freed as
65 * needed from the blocks available. Figure X shows the ZIL structure:
66 */
67
68/*
69 * Disable intent logging replay.  This global ZIL switch affects all pools.
70 */
71int zil_replay_disable = 0;
72
73/*
74 * Tunable parameter for debugging or performance analysis.  Setting
75 * zfs_nocacheflush will cause corruption on power loss if a volatile
76 * out-of-order write cache is enabled.
77 */
78boolean_t zfs_nocacheflush = B_FALSE;
79
80static kmem_cache_t *zil_lwb_cache;
81
82static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
83
84#define	LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
85    sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
86
87
88/*
89 * ziltest is by and large an ugly hack, but very useful in
90 * checking replay without tedious work.
91 * When running ziltest we want to keep all itx's and so maintain
92 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
93 * We subtract TXG_CONCURRENT_STATES to allow for common code.
94 */
95#define	ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
96
97static int
98zil_bp_compare(const void *x1, const void *x2)
99{
100	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
101	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
102
103	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
104		return (-1);
105	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
106		return (1);
107
108	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
109		return (-1);
110	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
111		return (1);
112
113	return (0);
114}
115
116static void
117zil_bp_tree_init(zilog_t *zilog)
118{
119	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
120	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
121}
122
123static void
124zil_bp_tree_fini(zilog_t *zilog)
125{
126	avl_tree_t *t = &zilog->zl_bp_tree;
127	zil_bp_node_t *zn;
128	void *cookie = NULL;
129
130	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
131		kmem_free(zn, sizeof (zil_bp_node_t));
132
133	avl_destroy(t);
134}
135
136int
137zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
138{
139	avl_tree_t *t = &zilog->zl_bp_tree;
140	const dva_t *dva;
141	zil_bp_node_t *zn;
142	avl_index_t where;
143
144	if (BP_IS_EMBEDDED(bp))
145		return (0);
146
147	dva = BP_IDENTITY(bp);
148
149	if (avl_find(t, dva, &where) != NULL)
150		return (SET_ERROR(EEXIST));
151
152	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
153	zn->zn_dva = *dva;
154	avl_insert(t, zn, where);
155
156	return (0);
157}
158
159static zil_header_t *
160zil_header_in_syncing_context(zilog_t *zilog)
161{
162	return ((zil_header_t *)zilog->zl_header);
163}
164
165static void
166zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
167{
168	zio_cksum_t *zc = &bp->blk_cksum;
169
170	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
171	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
172	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
173	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
174}
175
176/*
177 * Read a log block and make sure it's valid.
178 */
179static int
180zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
181    char **end)
182{
183	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
184	uint32_t aflags = ARC_WAIT;
185	arc_buf_t *abuf = NULL;
186	zbookmark_phys_t zb;
187	int error;
188
189	if (zilog->zl_header->zh_claim_txg == 0)
190		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
191
192	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
193		zio_flags |= ZIO_FLAG_SPECULATIVE;
194
195	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
196	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
197
198	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
199	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
200
201	if (error == 0) {
202		zio_cksum_t cksum = bp->blk_cksum;
203
204		/*
205		 * Validate the checksummed log block.
206		 *
207		 * Sequence numbers should be... sequential.  The checksum
208		 * verifier for the next block should be bp's checksum plus 1.
209		 *
210		 * Also check the log chain linkage and size used.
211		 */
212		cksum.zc_word[ZIL_ZC_SEQ]++;
213
214		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
215			zil_chain_t *zilc = abuf->b_data;
216			char *lr = (char *)(zilc + 1);
217			uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
218
219			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
220			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
221				error = SET_ERROR(ECKSUM);
222			} else {
223				bcopy(lr, dst, len);
224				*end = (char *)dst + len;
225				*nbp = zilc->zc_next_blk;
226			}
227		} else {
228			char *lr = abuf->b_data;
229			uint64_t size = BP_GET_LSIZE(bp);
230			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
231
232			if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
233			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
234			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
235				error = SET_ERROR(ECKSUM);
236			} else {
237				bcopy(lr, dst, zilc->zc_nused);
238				*end = (char *)dst + zilc->zc_nused;
239				*nbp = zilc->zc_next_blk;
240			}
241		}
242
243		VERIFY(arc_buf_remove_ref(abuf, &abuf));
244	}
245
246	return (error);
247}
248
249/*
250 * Read a TX_WRITE log data block.
251 */
252static int
253zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
254{
255	enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
256	const blkptr_t *bp = &lr->lr_blkptr;
257	uint32_t aflags = ARC_WAIT;
258	arc_buf_t *abuf = NULL;
259	zbookmark_phys_t zb;
260	int error;
261
262	if (BP_IS_HOLE(bp)) {
263		if (wbuf != NULL)
264			bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
265		return (0);
266	}
267
268	if (zilog->zl_header->zh_claim_txg == 0)
269		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
270
271	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
272	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
273
274	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
275	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
276
277	if (error == 0) {
278		if (wbuf != NULL)
279			bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
280		(void) arc_buf_remove_ref(abuf, &abuf);
281	}
282
283	return (error);
284}
285
286/*
287 * Parse the intent log, and call parse_func for each valid record within.
288 */
289int
290zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
291    zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
292{
293	const zil_header_t *zh = zilog->zl_header;
294	boolean_t claimed = !!zh->zh_claim_txg;
295	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
296	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
297	uint64_t max_blk_seq = 0;
298	uint64_t max_lr_seq = 0;
299	uint64_t blk_count = 0;
300	uint64_t lr_count = 0;
301	blkptr_t blk, next_blk;
302	char *lrbuf, *lrp;
303	int error = 0;
304
305	/*
306	 * Old logs didn't record the maximum zh_claim_lr_seq.
307	 */
308	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
309		claim_lr_seq = UINT64_MAX;
310
311	/*
312	 * Starting at the block pointed to by zh_log we read the log chain.
313	 * For each block in the chain we strongly check that block to
314	 * ensure its validity.  We stop when an invalid block is found.
315	 * For each block pointer in the chain we call parse_blk_func().
316	 * For each record in each valid block we call parse_lr_func().
317	 * If the log has been claimed, stop if we encounter a sequence
318	 * number greater than the highest claimed sequence number.
319	 */
320	lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
321	zil_bp_tree_init(zilog);
322
323	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
324		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
325		int reclen;
326		char *end;
327
328		if (blk_seq > claim_blk_seq)
329			break;
330		if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
331			break;
332		ASSERT3U(max_blk_seq, <, blk_seq);
333		max_blk_seq = blk_seq;
334		blk_count++;
335
336		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
337			break;
338
339		error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
340		if (error != 0)
341			break;
342
343		for (lrp = lrbuf; lrp < end; lrp += reclen) {
344			lr_t *lr = (lr_t *)lrp;
345			reclen = lr->lrc_reclen;
346			ASSERT3U(reclen, >=, sizeof (lr_t));
347			if (lr->lrc_seq > claim_lr_seq)
348				goto done;
349			if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
350				goto done;
351			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
352			max_lr_seq = lr->lrc_seq;
353			lr_count++;
354		}
355	}
356done:
357	zilog->zl_parse_error = error;
358	zilog->zl_parse_blk_seq = max_blk_seq;
359	zilog->zl_parse_lr_seq = max_lr_seq;
360	zilog->zl_parse_blk_count = blk_count;
361	zilog->zl_parse_lr_count = lr_count;
362
363	ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
364	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
365
366	zil_bp_tree_fini(zilog);
367	zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
368
369	return (error);
370}
371
372static int
373zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
374{
375	/*
376	 * Claim log block if not already committed and not already claimed.
377	 * If tx == NULL, just verify that the block is claimable.
378	 */
379	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
380	    zil_bp_tree_add(zilog, bp) != 0)
381		return (0);
382
383	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
384	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
385	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
386}
387
388static int
389zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
390{
391	lr_write_t *lr = (lr_write_t *)lrc;
392	int error;
393
394	if (lrc->lrc_txtype != TX_WRITE)
395		return (0);
396
397	/*
398	 * If the block is not readable, don't claim it.  This can happen
399	 * in normal operation when a log block is written to disk before
400	 * some of the dmu_sync() blocks it points to.  In this case, the
401	 * transaction cannot have been committed to anyone (we would have
402	 * waited for all writes to be stable first), so it is semantically
403	 * correct to declare this the end of the log.
404	 */
405	if (lr->lr_blkptr.blk_birth >= first_txg &&
406	    (error = zil_read_log_data(zilog, lr, NULL)) != 0)
407		return (error);
408	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
409}
410
411/* ARGSUSED */
412static int
413zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
414{
415	zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
416
417	return (0);
418}
419
420static int
421zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
422{
423	lr_write_t *lr = (lr_write_t *)lrc;
424	blkptr_t *bp = &lr->lr_blkptr;
425
426	/*
427	 * If we previously claimed it, we need to free it.
428	 */
429	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
430	    bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
431	    !BP_IS_HOLE(bp))
432		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
433
434	return (0);
435}
436
437static lwb_t *
438zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
439{
440	lwb_t *lwb;
441
442	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
443	lwb->lwb_zilog = zilog;
444	lwb->lwb_blk = *bp;
445	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
446	lwb->lwb_max_txg = txg;
447	lwb->lwb_zio = NULL;
448	lwb->lwb_tx = NULL;
449	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
450		lwb->lwb_nused = sizeof (zil_chain_t);
451		lwb->lwb_sz = BP_GET_LSIZE(bp);
452	} else {
453		lwb->lwb_nused = 0;
454		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
455	}
456
457	mutex_enter(&zilog->zl_lock);
458	list_insert_tail(&zilog->zl_lwb_list, lwb);
459	mutex_exit(&zilog->zl_lock);
460
461	return (lwb);
462}
463
464/*
465 * Called when we create in-memory log transactions so that we know
466 * to cleanup the itxs at the end of spa_sync().
467 */
468void
469zilog_dirty(zilog_t *zilog, uint64_t txg)
470{
471	dsl_pool_t *dp = zilog->zl_dmu_pool;
472	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
473
474	if (dsl_dataset_is_snapshot(ds))
475		panic("dirtying snapshot!");
476
477	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
478		/* up the hold count until we can be written out */
479		dmu_buf_add_ref(ds->ds_dbuf, zilog);
480	}
481}
482
483boolean_t
484zilog_is_dirty(zilog_t *zilog)
485{
486	dsl_pool_t *dp = zilog->zl_dmu_pool;
487
488	for (int t = 0; t < TXG_SIZE; t++) {
489		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
490			return (B_TRUE);
491	}
492	return (B_FALSE);
493}
494
495/*
496 * Create an on-disk intent log.
497 */
498static lwb_t *
499zil_create(zilog_t *zilog)
500{
501	const zil_header_t *zh = zilog->zl_header;
502	lwb_t *lwb = NULL;
503	uint64_t txg = 0;
504	dmu_tx_t *tx = NULL;
505	blkptr_t blk;
506	int error = 0;
507
508	/*
509	 * Wait for any previous destroy to complete.
510	 */
511	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
512
513	ASSERT(zh->zh_claim_txg == 0);
514	ASSERT(zh->zh_replay_seq == 0);
515
516	blk = zh->zh_log;
517
518	/*
519	 * Allocate an initial log block if:
520	 *    - there isn't one already
521	 *    - the existing block is the wrong endianess
522	 */
523	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
524		tx = dmu_tx_create(zilog->zl_os);
525		VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
526		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
527		txg = dmu_tx_get_txg(tx);
528
529		if (!BP_IS_HOLE(&blk)) {
530			zio_free_zil(zilog->zl_spa, txg, &blk);
531			BP_ZERO(&blk);
532		}
533
534		error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
535		    ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
536
537		if (error == 0)
538			zil_init_log_chain(zilog, &blk);
539	}
540
541	/*
542	 * Allocate a log write buffer (lwb) for the first log block.
543	 */
544	if (error == 0)
545		lwb = zil_alloc_lwb(zilog, &blk, txg);
546
547	/*
548	 * If we just allocated the first log block, commit our transaction
549	 * and wait for zil_sync() to stuff the block poiner into zh_log.
550	 * (zh is part of the MOS, so we cannot modify it in open context.)
551	 */
552	if (tx != NULL) {
553		dmu_tx_commit(tx);
554		txg_wait_synced(zilog->zl_dmu_pool, txg);
555	}
556
557	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
558
559	return (lwb);
560}
561
562/*
563 * In one tx, free all log blocks and clear the log header.
564 * If keep_first is set, then we're replaying a log with no content.
565 * We want to keep the first block, however, so that the first
566 * synchronous transaction doesn't require a txg_wait_synced()
567 * in zil_create().  We don't need to txg_wait_synced() here either
568 * when keep_first is set, because both zil_create() and zil_destroy()
569 * will wait for any in-progress destroys to complete.
570 */
571void
572zil_destroy(zilog_t *zilog, boolean_t keep_first)
573{
574	const zil_header_t *zh = zilog->zl_header;
575	lwb_t *lwb;
576	dmu_tx_t *tx;
577	uint64_t txg;
578
579	/*
580	 * Wait for any previous destroy to complete.
581	 */
582	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
583
584	zilog->zl_old_header = *zh;		/* debugging aid */
585
586	if (BP_IS_HOLE(&zh->zh_log))
587		return;
588
589	tx = dmu_tx_create(zilog->zl_os);
590	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
591	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
592	txg = dmu_tx_get_txg(tx);
593
594	mutex_enter(&zilog->zl_lock);
595
596	ASSERT3U(zilog->zl_destroy_txg, <, txg);
597	zilog->zl_destroy_txg = txg;
598	zilog->zl_keep_first = keep_first;
599
600	if (!list_is_empty(&zilog->zl_lwb_list)) {
601		ASSERT(zh->zh_claim_txg == 0);
602		VERIFY(!keep_first);
603		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
604			list_remove(&zilog->zl_lwb_list, lwb);
605			if (lwb->lwb_buf != NULL)
606				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
607			zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
608			kmem_cache_free(zil_lwb_cache, lwb);
609		}
610	} else if (!keep_first) {
611		zil_destroy_sync(zilog, tx);
612	}
613	mutex_exit(&zilog->zl_lock);
614
615	dmu_tx_commit(tx);
616}
617
618void
619zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
620{
621	ASSERT(list_is_empty(&zilog->zl_lwb_list));
622	(void) zil_parse(zilog, zil_free_log_block,
623	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
624}
625
626int
627zil_claim(const char *osname, void *txarg)
628{
629	dmu_tx_t *tx = txarg;
630	uint64_t first_txg = dmu_tx_get_txg(tx);
631	zilog_t *zilog;
632	zil_header_t *zh;
633	objset_t *os;
634	int error;
635
636	error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
637	if (error != 0) {
638		/*
639		 * EBUSY indicates that the objset is inconsistent, in which
640		 * case it can not have a ZIL.
641		 */
642		if (error != EBUSY) {
643			cmn_err(CE_WARN, "can't open objset for %s, error %u",
644			    osname, error);
645		}
646		return (0);
647	}
648
649	zilog = dmu_objset_zil(os);
650	zh = zil_header_in_syncing_context(zilog);
651
652	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
653		if (!BP_IS_HOLE(&zh->zh_log))
654			zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
655		BP_ZERO(&zh->zh_log);
656		dsl_dataset_dirty(dmu_objset_ds(os), tx);
657		dmu_objset_disown(os, FTAG);
658		return (0);
659	}
660
661	/*
662	 * Claim all log blocks if we haven't already done so, and remember
663	 * the highest claimed sequence number.  This ensures that if we can
664	 * read only part of the log now (e.g. due to a missing device),
665	 * but we can read the entire log later, we will not try to replay
666	 * or destroy beyond the last block we successfully claimed.
667	 */
668	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
669	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
670		(void) zil_parse(zilog, zil_claim_log_block,
671		    zil_claim_log_record, tx, first_txg);
672		zh->zh_claim_txg = first_txg;
673		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
674		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
675		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
676			zh->zh_flags |= ZIL_REPLAY_NEEDED;
677		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
678		dsl_dataset_dirty(dmu_objset_ds(os), tx);
679	}
680
681	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
682	dmu_objset_disown(os, FTAG);
683	return (0);
684}
685
686/*
687 * Check the log by walking the log chain.
688 * Checksum errors are ok as they indicate the end of the chain.
689 * Any other error (no device or read failure) returns an error.
690 */
691int
692zil_check_log_chain(const char *osname, void *tx)
693{
694	zilog_t *zilog;
695	objset_t *os;
696	blkptr_t *bp;
697	int error;
698
699	ASSERT(tx == NULL);
700
701	error = dmu_objset_hold(osname, FTAG, &os);
702	if (error != 0) {
703		cmn_err(CE_WARN, "can't open objset for %s", osname);
704		return (0);
705	}
706
707	zilog = dmu_objset_zil(os);
708	bp = (blkptr_t *)&zilog->zl_header->zh_log;
709
710	/*
711	 * Check the first block and determine if it's on a log device
712	 * which may have been removed or faulted prior to loading this
713	 * pool.  If so, there's no point in checking the rest of the log
714	 * as its content should have already been synced to the pool.
715	 */
716	if (!BP_IS_HOLE(bp)) {
717		vdev_t *vd;
718		boolean_t valid = B_TRUE;
719
720		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
721		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
722		if (vd->vdev_islog && vdev_is_dead(vd))
723			valid = vdev_log_state_valid(vd);
724		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
725
726		if (!valid) {
727			dmu_objset_rele(os, FTAG);
728			return (0);
729		}
730	}
731
732	/*
733	 * Because tx == NULL, zil_claim_log_block() will not actually claim
734	 * any blocks, but just determine whether it is possible to do so.
735	 * In addition to checking the log chain, zil_claim_log_block()
736	 * will invoke zio_claim() with a done func of spa_claim_notify(),
737	 * which will update spa_max_claim_txg.  See spa_load() for details.
738	 */
739	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
740	    zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
741
742	dmu_objset_rele(os, FTAG);
743
744	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
745}
746
747static int
748zil_vdev_compare(const void *x1, const void *x2)
749{
750	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
751	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
752
753	if (v1 < v2)
754		return (-1);
755	if (v1 > v2)
756		return (1);
757
758	return (0);
759}
760
761void
762zil_add_block(zilog_t *zilog, const blkptr_t *bp)
763{
764	avl_tree_t *t = &zilog->zl_vdev_tree;
765	avl_index_t where;
766	zil_vdev_node_t *zv, zvsearch;
767	int ndvas = BP_GET_NDVAS(bp);
768	int i;
769
770	if (zfs_nocacheflush)
771		return;
772
773	ASSERT(zilog->zl_writer);
774
775	/*
776	 * Even though we're zl_writer, we still need a lock because the
777	 * zl_get_data() callbacks may have dmu_sync() done callbacks
778	 * that will run concurrently.
779	 */
780	mutex_enter(&zilog->zl_vdev_lock);
781	for (i = 0; i < ndvas; i++) {
782		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
783		if (avl_find(t, &zvsearch, &where) == NULL) {
784			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
785			zv->zv_vdev = zvsearch.zv_vdev;
786			avl_insert(t, zv, where);
787		}
788	}
789	mutex_exit(&zilog->zl_vdev_lock);
790}
791
792static void
793zil_flush_vdevs(zilog_t *zilog)
794{
795	spa_t *spa = zilog->zl_spa;
796	avl_tree_t *t = &zilog->zl_vdev_tree;
797	void *cookie = NULL;
798	zil_vdev_node_t *zv;
799	zio_t *zio;
800
801	ASSERT(zilog->zl_writer);
802
803	/*
804	 * We don't need zl_vdev_lock here because we're the zl_writer,
805	 * and all zl_get_data() callbacks are done.
806	 */
807	if (avl_numnodes(t) == 0)
808		return;
809
810	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
811
812	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
813
814	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
815		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
816		if (vd != NULL)
817			zio_flush(zio, vd);
818		kmem_free(zv, sizeof (*zv));
819	}
820
821	/*
822	 * Wait for all the flushes to complete.  Not all devices actually
823	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
824	 */
825	(void) zio_wait(zio);
826
827	spa_config_exit(spa, SCL_STATE, FTAG);
828}
829
830/*
831 * Function called when a log block write completes
832 */
833static void
834zil_lwb_write_done(zio_t *zio)
835{
836	lwb_t *lwb = zio->io_private;
837	zilog_t *zilog = lwb->lwb_zilog;
838	dmu_tx_t *tx = lwb->lwb_tx;
839
840	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
841	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
842	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
843	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
844	ASSERT(!BP_IS_GANG(zio->io_bp));
845	ASSERT(!BP_IS_HOLE(zio->io_bp));
846	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
847
848	/*
849	 * Ensure the lwb buffer pointer is cleared before releasing
850	 * the txg. If we have had an allocation failure and
851	 * the txg is waiting to sync then we want want zil_sync()
852	 * to remove the lwb so that it's not picked up as the next new
853	 * one in zil_commit_writer(). zil_sync() will only remove
854	 * the lwb if lwb_buf is null.
855	 */
856	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
857	mutex_enter(&zilog->zl_lock);
858	lwb->lwb_buf = NULL;
859	lwb->lwb_tx = NULL;
860	mutex_exit(&zilog->zl_lock);
861
862	/*
863	 * Now that we've written this log block, we have a stable pointer
864	 * to the next block in the chain, so it's OK to let the txg in
865	 * which we allocated the next block sync.
866	 */
867	dmu_tx_commit(tx);
868}
869
870/*
871 * Initialize the io for a log block.
872 */
873static void
874zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
875{
876	zbookmark_phys_t zb;
877
878	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
879	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
880	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
881
882	if (zilog->zl_root_zio == NULL) {
883		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
884		    ZIO_FLAG_CANFAIL);
885	}
886	if (lwb->lwb_zio == NULL) {
887		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
888		    0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
889		    zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
890		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
891	}
892}
893
894/*
895 * Define a limited set of intent log block sizes.
896 *
897 * These must be a multiple of 4KB. Note only the amount used (again
898 * aligned to 4KB) actually gets written. However, we can't always just
899 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
900 */
901uint64_t zil_block_buckets[] = {
902    4096,		/* non TX_WRITE */
903    8192+4096,		/* data base */
904    32*1024 + 4096, 	/* NFS writes */
905    UINT64_MAX
906};
907
908/*
909 * Use the slog as long as the logbias is 'latency' and the current commit size
910 * is less than the limit or the total list size is less than 2X the limit.
911 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
912 */
913uint64_t zil_slog_limit = 1024 * 1024;
914#define	USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
915	(((zilog)->zl_cur_used < zil_slog_limit) || \
916	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
917
918/*
919 * Start a log block write and advance to the next log block.
920 * Calls are serialized.
921 */
922static lwb_t *
923zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
924{
925	lwb_t *nlwb = NULL;
926	zil_chain_t *zilc;
927	spa_t *spa = zilog->zl_spa;
928	blkptr_t *bp;
929	dmu_tx_t *tx;
930	uint64_t txg;
931	uint64_t zil_blksz, wsz;
932	int i, error;
933
934	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
935		zilc = (zil_chain_t *)lwb->lwb_buf;
936		bp = &zilc->zc_next_blk;
937	} else {
938		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
939		bp = &zilc->zc_next_blk;
940	}
941
942	ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
943
944	/*
945	 * Allocate the next block and save its address in this block
946	 * before writing it in order to establish the log chain.
947	 * Note that if the allocation of nlwb synced before we wrote
948	 * the block that points at it (lwb), we'd leak it if we crashed.
949	 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
950	 * We dirty the dataset to ensure that zil_sync() will be called
951	 * to clean up in the event of allocation failure or I/O failure.
952	 */
953	tx = dmu_tx_create(zilog->zl_os);
954	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
955	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
956	txg = dmu_tx_get_txg(tx);
957
958	lwb->lwb_tx = tx;
959
960	/*
961	 * Log blocks are pre-allocated. Here we select the size of the next
962	 * block, based on size used in the last block.
963	 * - first find the smallest bucket that will fit the block from a
964	 *   limited set of block sizes. This is because it's faster to write
965	 *   blocks allocated from the same metaslab as they are adjacent or
966	 *   close.
967	 * - next find the maximum from the new suggested size and an array of
968	 *   previous sizes. This lessens a picket fence effect of wrongly
969	 *   guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
970	 *   requests.
971	 *
972	 * Note we only write what is used, but we can't just allocate
973	 * the maximum block size because we can exhaust the available
974	 * pool log space.
975	 */
976	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
977	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
978		continue;
979	zil_blksz = zil_block_buckets[i];
980	if (zil_blksz == UINT64_MAX)
981		zil_blksz = SPA_MAXBLOCKSIZE;
982	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
983	for (i = 0; i < ZIL_PREV_BLKS; i++)
984		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
985	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
986
987	BP_ZERO(bp);
988	/* pass the old blkptr in order to spread log blocks across devs */
989	error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
990	    USE_SLOG(zilog));
991	if (error == 0) {
992		ASSERT3U(bp->blk_birth, ==, txg);
993		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
994		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
995
996		/*
997		 * Allocate a new log write buffer (lwb).
998		 */
999		nlwb = zil_alloc_lwb(zilog, bp, txg);
1000
1001		/* Record the block for later vdev flushing */
1002		zil_add_block(zilog, &lwb->lwb_blk);
1003	}
1004
1005	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1006		/* For Slim ZIL only write what is used. */
1007		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1008		ASSERT3U(wsz, <=, lwb->lwb_sz);
1009		zio_shrink(lwb->lwb_zio, wsz);
1010
1011	} else {
1012		wsz = lwb->lwb_sz;
1013	}
1014
1015	zilc->zc_pad = 0;
1016	zilc->zc_nused = lwb->lwb_nused;
1017	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1018
1019	/*
1020	 * clear unused data for security
1021	 */
1022	bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1023
1024	zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1025
1026	/*
1027	 * If there was an allocation failure then nlwb will be null which
1028	 * forces a txg_wait_synced().
1029	 */
1030	return (nlwb);
1031}
1032
1033static lwb_t *
1034zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1035{
1036	lr_t *lrc = &itx->itx_lr; /* common log record */
1037	lr_write_t *lrw = (lr_write_t *)lrc;
1038	char *lr_buf;
1039	uint64_t txg = lrc->lrc_txg;
1040	uint64_t reclen = lrc->lrc_reclen;
1041	uint64_t dlen = 0;
1042
1043	if (lwb == NULL)
1044		return (NULL);
1045
1046	ASSERT(lwb->lwb_buf != NULL);
1047	ASSERT(zilog_is_dirty(zilog) ||
1048	    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1049
1050	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1051		dlen = P2ROUNDUP_TYPED(
1052		    lrw->lr_length, sizeof (uint64_t), uint64_t);
1053
1054	zilog->zl_cur_used += (reclen + dlen);
1055
1056	zil_lwb_write_init(zilog, lwb);
1057
1058	/*
1059	 * If this record won't fit in the current log block, start a new one.
1060	 */
1061	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1062		lwb = zil_lwb_write_start(zilog, lwb);
1063		if (lwb == NULL)
1064			return (NULL);
1065		zil_lwb_write_init(zilog, lwb);
1066		ASSERT(LWB_EMPTY(lwb));
1067		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1068			txg_wait_synced(zilog->zl_dmu_pool, txg);
1069			return (lwb);
1070		}
1071	}
1072
1073	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1074	bcopy(lrc, lr_buf, reclen);
1075	lrc = (lr_t *)lr_buf;
1076	lrw = (lr_write_t *)lrc;
1077
1078	/*
1079	 * If it's a write, fetch the data or get its blkptr as appropriate.
1080	 */
1081	if (lrc->lrc_txtype == TX_WRITE) {
1082		if (txg > spa_freeze_txg(zilog->zl_spa))
1083			txg_wait_synced(zilog->zl_dmu_pool, txg);
1084		if (itx->itx_wr_state != WR_COPIED) {
1085			char *dbuf;
1086			int error;
1087
1088			if (dlen) {
1089				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1090				dbuf = lr_buf + reclen;
1091				lrw->lr_common.lrc_reclen += dlen;
1092			} else {
1093				ASSERT(itx->itx_wr_state == WR_INDIRECT);
1094				dbuf = NULL;
1095			}
1096			error = zilog->zl_get_data(
1097			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1098			if (error == EIO) {
1099				txg_wait_synced(zilog->zl_dmu_pool, txg);
1100				return (lwb);
1101			}
1102			if (error != 0) {
1103				ASSERT(error == ENOENT || error == EEXIST ||
1104				    error == EALREADY);
1105				return (lwb);
1106			}
1107		}
1108	}
1109
1110	/*
1111	 * We're actually making an entry, so update lrc_seq to be the
1112	 * log record sequence number.  Note that this is generally not
1113	 * equal to the itx sequence number because not all transactions
1114	 * are synchronous, and sometimes spa_sync() gets there first.
1115	 */
1116	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1117	lwb->lwb_nused += reclen + dlen;
1118	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1119	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1120	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1121
1122	return (lwb);
1123}
1124
1125itx_t *
1126zil_itx_create(uint64_t txtype, size_t lrsize)
1127{
1128	itx_t *itx;
1129
1130	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1131
1132	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1133	itx->itx_lr.lrc_txtype = txtype;
1134	itx->itx_lr.lrc_reclen = lrsize;
1135	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1136	itx->itx_lr.lrc_seq = 0;	/* defensive */
1137	itx->itx_sync = B_TRUE;		/* default is synchronous */
1138
1139	return (itx);
1140}
1141
1142void
1143zil_itx_destroy(itx_t *itx)
1144{
1145	kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1146}
1147
1148/*
1149 * Free up the sync and async itxs. The itxs_t has already been detached
1150 * so no locks are needed.
1151 */
1152static void
1153zil_itxg_clean(itxs_t *itxs)
1154{
1155	itx_t *itx;
1156	list_t *list;
1157	avl_tree_t *t;
1158	void *cookie;
1159	itx_async_node_t *ian;
1160
1161	list = &itxs->i_sync_list;
1162	while ((itx = list_head(list)) != NULL) {
1163		list_remove(list, itx);
1164		kmem_free(itx, offsetof(itx_t, itx_lr) +
1165		    itx->itx_lr.lrc_reclen);
1166	}
1167
1168	cookie = NULL;
1169	t = &itxs->i_async_tree;
1170	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1171		list = &ian->ia_list;
1172		while ((itx = list_head(list)) != NULL) {
1173			list_remove(list, itx);
1174			kmem_free(itx, offsetof(itx_t, itx_lr) +
1175			    itx->itx_lr.lrc_reclen);
1176		}
1177		list_destroy(list);
1178		kmem_free(ian, sizeof (itx_async_node_t));
1179	}
1180	avl_destroy(t);
1181
1182	kmem_free(itxs, sizeof (itxs_t));
1183}
1184
1185static int
1186zil_aitx_compare(const void *x1, const void *x2)
1187{
1188	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1189	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1190
1191	if (o1 < o2)
1192		return (-1);
1193	if (o1 > o2)
1194		return (1);
1195
1196	return (0);
1197}
1198
1199/*
1200 * Remove all async itx with the given oid.
1201 */
1202static void
1203zil_remove_async(zilog_t *zilog, uint64_t oid)
1204{
1205	uint64_t otxg, txg;
1206	itx_async_node_t *ian;
1207	avl_tree_t *t;
1208	avl_index_t where;
1209	list_t clean_list;
1210	itx_t *itx;
1211
1212	ASSERT(oid != 0);
1213	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1214
1215	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1216		otxg = ZILTEST_TXG;
1217	else
1218		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1219
1220	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1221		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1222
1223		mutex_enter(&itxg->itxg_lock);
1224		if (itxg->itxg_txg != txg) {
1225			mutex_exit(&itxg->itxg_lock);
1226			continue;
1227		}
1228
1229		/*
1230		 * Locate the object node and append its list.
1231		 */
1232		t = &itxg->itxg_itxs->i_async_tree;
1233		ian = avl_find(t, &oid, &where);
1234		if (ian != NULL)
1235			list_move_tail(&clean_list, &ian->ia_list);
1236		mutex_exit(&itxg->itxg_lock);
1237	}
1238	while ((itx = list_head(&clean_list)) != NULL) {
1239		list_remove(&clean_list, itx);
1240		kmem_free(itx, offsetof(itx_t, itx_lr) +
1241		    itx->itx_lr.lrc_reclen);
1242	}
1243	list_destroy(&clean_list);
1244}
1245
1246void
1247zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1248{
1249	uint64_t txg;
1250	itxg_t *itxg;
1251	itxs_t *itxs, *clean = NULL;
1252
1253	/*
1254	 * Object ids can be re-instantiated in the next txg so
1255	 * remove any async transactions to avoid future leaks.
1256	 * This can happen if a fsync occurs on the re-instantiated
1257	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1258	 * the new file data and flushes a write record for the old object.
1259	 */
1260	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1261		zil_remove_async(zilog, itx->itx_oid);
1262
1263	/*
1264	 * Ensure the data of a renamed file is committed before the rename.
1265	 */
1266	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1267		zil_async_to_sync(zilog, itx->itx_oid);
1268
1269	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1270		txg = ZILTEST_TXG;
1271	else
1272		txg = dmu_tx_get_txg(tx);
1273
1274	itxg = &zilog->zl_itxg[txg & TXG_MASK];
1275	mutex_enter(&itxg->itxg_lock);
1276	itxs = itxg->itxg_itxs;
1277	if (itxg->itxg_txg != txg) {
1278		if (itxs != NULL) {
1279			/*
1280			 * The zil_clean callback hasn't got around to cleaning
1281			 * this itxg. Save the itxs for release below.
1282			 * This should be rare.
1283			 */
1284			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1285			itxg->itxg_sod = 0;
1286			clean = itxg->itxg_itxs;
1287		}
1288		ASSERT(itxg->itxg_sod == 0);
1289		itxg->itxg_txg = txg;
1290		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1291
1292		list_create(&itxs->i_sync_list, sizeof (itx_t),
1293		    offsetof(itx_t, itx_node));
1294		avl_create(&itxs->i_async_tree, zil_aitx_compare,
1295		    sizeof (itx_async_node_t),
1296		    offsetof(itx_async_node_t, ia_node));
1297	}
1298	if (itx->itx_sync) {
1299		list_insert_tail(&itxs->i_sync_list, itx);
1300		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1301		itxg->itxg_sod += itx->itx_sod;
1302	} else {
1303		avl_tree_t *t = &itxs->i_async_tree;
1304		uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1305		itx_async_node_t *ian;
1306		avl_index_t where;
1307
1308		ian = avl_find(t, &foid, &where);
1309		if (ian == NULL) {
1310			ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1311			list_create(&ian->ia_list, sizeof (itx_t),
1312			    offsetof(itx_t, itx_node));
1313			ian->ia_foid = foid;
1314			avl_insert(t, ian, where);
1315		}
1316		list_insert_tail(&ian->ia_list, itx);
1317	}
1318
1319	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1320	zilog_dirty(zilog, txg);
1321	mutex_exit(&itxg->itxg_lock);
1322
1323	/* Release the old itxs now we've dropped the lock */
1324	if (clean != NULL)
1325		zil_itxg_clean(clean);
1326}
1327
1328/*
1329 * If there are any in-memory intent log transactions which have now been
1330 * synced then start up a taskq to free them. We should only do this after we
1331 * have written out the uberblocks (i.e. txg has been comitted) so that
1332 * don't inadvertently clean out in-memory log records that would be required
1333 * by zil_commit().
1334 */
1335void
1336zil_clean(zilog_t *zilog, uint64_t synced_txg)
1337{
1338	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1339	itxs_t *clean_me;
1340
1341	mutex_enter(&itxg->itxg_lock);
1342	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1343		mutex_exit(&itxg->itxg_lock);
1344		return;
1345	}
1346	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1347	ASSERT(itxg->itxg_txg != 0);
1348	ASSERT(zilog->zl_clean_taskq != NULL);
1349	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1350	itxg->itxg_sod = 0;
1351	clean_me = itxg->itxg_itxs;
1352	itxg->itxg_itxs = NULL;
1353	itxg->itxg_txg = 0;
1354	mutex_exit(&itxg->itxg_lock);
1355	/*
1356	 * Preferably start a task queue to free up the old itxs but
1357	 * if taskq_dispatch can't allocate resources to do that then
1358	 * free it in-line. This should be rare. Note, using TQ_SLEEP
1359	 * created a bad performance problem.
1360	 */
1361	if (taskq_dispatch(zilog->zl_clean_taskq,
1362	    (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1363		zil_itxg_clean(clean_me);
1364}
1365
1366/*
1367 * Get the list of itxs to commit into zl_itx_commit_list.
1368 */
1369static void
1370zil_get_commit_list(zilog_t *zilog)
1371{
1372	uint64_t otxg, txg;
1373	list_t *commit_list = &zilog->zl_itx_commit_list;
1374	uint64_t push_sod = 0;
1375
1376	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1377		otxg = ZILTEST_TXG;
1378	else
1379		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1380
1381	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1382		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1383
1384		mutex_enter(&itxg->itxg_lock);
1385		if (itxg->itxg_txg != txg) {
1386			mutex_exit(&itxg->itxg_lock);
1387			continue;
1388		}
1389
1390		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1391		push_sod += itxg->itxg_sod;
1392		itxg->itxg_sod = 0;
1393
1394		mutex_exit(&itxg->itxg_lock);
1395	}
1396	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1397}
1398
1399/*
1400 * Move the async itxs for a specified object to commit into sync lists.
1401 */
1402static void
1403zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1404{
1405	uint64_t otxg, txg;
1406	itx_async_node_t *ian;
1407	avl_tree_t *t;
1408	avl_index_t where;
1409
1410	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1411		otxg = ZILTEST_TXG;
1412	else
1413		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1414
1415	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1416		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1417
1418		mutex_enter(&itxg->itxg_lock);
1419		if (itxg->itxg_txg != txg) {
1420			mutex_exit(&itxg->itxg_lock);
1421			continue;
1422		}
1423
1424		/*
1425		 * If a foid is specified then find that node and append its
1426		 * list. Otherwise walk the tree appending all the lists
1427		 * to the sync list. We add to the end rather than the
1428		 * beginning to ensure the create has happened.
1429		 */
1430		t = &itxg->itxg_itxs->i_async_tree;
1431		if (foid != 0) {
1432			ian = avl_find(t, &foid, &where);
1433			if (ian != NULL) {
1434				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1435				    &ian->ia_list);
1436			}
1437		} else {
1438			void *cookie = NULL;
1439
1440			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1441				list_move_tail(&itxg->itxg_itxs->i_sync_list,
1442				    &ian->ia_list);
1443				list_destroy(&ian->ia_list);
1444				kmem_free(ian, sizeof (itx_async_node_t));
1445			}
1446		}
1447		mutex_exit(&itxg->itxg_lock);
1448	}
1449}
1450
1451static void
1452zil_commit_writer(zilog_t *zilog)
1453{
1454	uint64_t txg;
1455	itx_t *itx;
1456	lwb_t *lwb;
1457	spa_t *spa = zilog->zl_spa;
1458	int error = 0;
1459
1460	ASSERT(zilog->zl_root_zio == NULL);
1461
1462	mutex_exit(&zilog->zl_lock);
1463
1464	zil_get_commit_list(zilog);
1465
1466	/*
1467	 * Return if there's nothing to commit before we dirty the fs by
1468	 * calling zil_create().
1469	 */
1470	if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1471		mutex_enter(&zilog->zl_lock);
1472		return;
1473	}
1474
1475	if (zilog->zl_suspend) {
1476		lwb = NULL;
1477	} else {
1478		lwb = list_tail(&zilog->zl_lwb_list);
1479		if (lwb == NULL)
1480			lwb = zil_create(zilog);
1481	}
1482
1483	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1484	while (itx = list_head(&zilog->zl_itx_commit_list)) {
1485		txg = itx->itx_lr.lrc_txg;
1486		ASSERT(txg);
1487
1488		if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1489			lwb = zil_lwb_commit(zilog, itx, lwb);
1490		list_remove(&zilog->zl_itx_commit_list, itx);
1491		kmem_free(itx, offsetof(itx_t, itx_lr)
1492		    + itx->itx_lr.lrc_reclen);
1493	}
1494	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1495
1496	/* write the last block out */
1497	if (lwb != NULL && lwb->lwb_zio != NULL)
1498		lwb = zil_lwb_write_start(zilog, lwb);
1499
1500	zilog->zl_cur_used = 0;
1501
1502	/*
1503	 * Wait if necessary for the log blocks to be on stable storage.
1504	 */
1505	if (zilog->zl_root_zio) {
1506		error = zio_wait(zilog->zl_root_zio);
1507		zilog->zl_root_zio = NULL;
1508		zil_flush_vdevs(zilog);
1509	}
1510
1511	if (error || lwb == NULL)
1512		txg_wait_synced(zilog->zl_dmu_pool, 0);
1513
1514	mutex_enter(&zilog->zl_lock);
1515
1516	/*
1517	 * Remember the highest committed log sequence number for ztest.
1518	 * We only update this value when all the log writes succeeded,
1519	 * because ztest wants to ASSERT that it got the whole log chain.
1520	 */
1521	if (error == 0 && lwb != NULL)
1522		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1523}
1524
1525/*
1526 * Commit zfs transactions to stable storage.
1527 * If foid is 0 push out all transactions, otherwise push only those
1528 * for that object or might reference that object.
1529 *
1530 * itxs are committed in batches. In a heavily stressed zil there will be
1531 * a commit writer thread who is writing out a bunch of itxs to the log
1532 * for a set of committing threads (cthreads) in the same batch as the writer.
1533 * Those cthreads are all waiting on the same cv for that batch.
1534 *
1535 * There will also be a different and growing batch of threads that are
1536 * waiting to commit (qthreads). When the committing batch completes
1537 * a transition occurs such that the cthreads exit and the qthreads become
1538 * cthreads. One of the new cthreads becomes the writer thread for the
1539 * batch. Any new threads arriving become new qthreads.
1540 *
1541 * Only 2 condition variables are needed and there's no transition
1542 * between the two cvs needed. They just flip-flop between qthreads
1543 * and cthreads.
1544 *
1545 * Using this scheme we can efficiently wakeup up only those threads
1546 * that have been committed.
1547 */
1548void
1549zil_commit(zilog_t *zilog, uint64_t foid)
1550{
1551	uint64_t mybatch;
1552
1553	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1554		return;
1555
1556	/* move the async itxs for the foid to the sync queues */
1557	zil_async_to_sync(zilog, foid);
1558
1559	mutex_enter(&zilog->zl_lock);
1560	mybatch = zilog->zl_next_batch;
1561	while (zilog->zl_writer) {
1562		cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1563		if (mybatch <= zilog->zl_com_batch) {
1564			mutex_exit(&zilog->zl_lock);
1565			return;
1566		}
1567	}
1568
1569	zilog->zl_next_batch++;
1570	zilog->zl_writer = B_TRUE;
1571	zil_commit_writer(zilog);
1572	zilog->zl_com_batch = mybatch;
1573	zilog->zl_writer = B_FALSE;
1574	mutex_exit(&zilog->zl_lock);
1575
1576	/* wake up one thread to become the next writer */
1577	cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1578
1579	/* wake up all threads waiting for this batch to be committed */
1580	cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1581}
1582
1583/*
1584 * Called in syncing context to free committed log blocks and update log header.
1585 */
1586void
1587zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1588{
1589	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1590	uint64_t txg = dmu_tx_get_txg(tx);
1591	spa_t *spa = zilog->zl_spa;
1592	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1593	lwb_t *lwb;
1594
1595	/*
1596	 * We don't zero out zl_destroy_txg, so make sure we don't try
1597	 * to destroy it twice.
1598	 */
1599	if (spa_sync_pass(spa) != 1)
1600		return;
1601
1602	mutex_enter(&zilog->zl_lock);
1603
1604	ASSERT(zilog->zl_stop_sync == 0);
1605
1606	if (*replayed_seq != 0) {
1607		ASSERT(zh->zh_replay_seq < *replayed_seq);
1608		zh->zh_replay_seq = *replayed_seq;
1609		*replayed_seq = 0;
1610	}
1611
1612	if (zilog->zl_destroy_txg == txg) {
1613		blkptr_t blk = zh->zh_log;
1614
1615		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1616
1617		bzero(zh, sizeof (zil_header_t));
1618		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1619
1620		if (zilog->zl_keep_first) {
1621			/*
1622			 * If this block was part of log chain that couldn't
1623			 * be claimed because a device was missing during
1624			 * zil_claim(), but that device later returns,
1625			 * then this block could erroneously appear valid.
1626			 * To guard against this, assign a new GUID to the new
1627			 * log chain so it doesn't matter what blk points to.
1628			 */
1629			zil_init_log_chain(zilog, &blk);
1630			zh->zh_log = blk;
1631		}
1632	}
1633
1634	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1635		zh->zh_log = lwb->lwb_blk;
1636		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1637			break;
1638		list_remove(&zilog->zl_lwb_list, lwb);
1639		zio_free_zil(spa, txg, &lwb->lwb_blk);
1640		kmem_cache_free(zil_lwb_cache, lwb);
1641
1642		/*
1643		 * If we don't have anything left in the lwb list then
1644		 * we've had an allocation failure and we need to zero
1645		 * out the zil_header blkptr so that we don't end
1646		 * up freeing the same block twice.
1647		 */
1648		if (list_head(&zilog->zl_lwb_list) == NULL)
1649			BP_ZERO(&zh->zh_log);
1650	}
1651	mutex_exit(&zilog->zl_lock);
1652}
1653
1654void
1655zil_init(void)
1656{
1657	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1658	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1659}
1660
1661void
1662zil_fini(void)
1663{
1664	kmem_cache_destroy(zil_lwb_cache);
1665}
1666
1667void
1668zil_set_sync(zilog_t *zilog, uint64_t sync)
1669{
1670	zilog->zl_sync = sync;
1671}
1672
1673void
1674zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1675{
1676	zilog->zl_logbias = logbias;
1677}
1678
1679zilog_t *
1680zil_alloc(objset_t *os, zil_header_t *zh_phys)
1681{
1682	zilog_t *zilog;
1683
1684	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1685
1686	zilog->zl_header = zh_phys;
1687	zilog->zl_os = os;
1688	zilog->zl_spa = dmu_objset_spa(os);
1689	zilog->zl_dmu_pool = dmu_objset_pool(os);
1690	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1691	zilog->zl_logbias = dmu_objset_logbias(os);
1692	zilog->zl_sync = dmu_objset_syncprop(os);
1693	zilog->zl_next_batch = 1;
1694
1695	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1696
1697	for (int i = 0; i < TXG_SIZE; i++) {
1698		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1699		    MUTEX_DEFAULT, NULL);
1700	}
1701
1702	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1703	    offsetof(lwb_t, lwb_node));
1704
1705	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1706	    offsetof(itx_t, itx_node));
1707
1708	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1709
1710	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1711	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1712
1713	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1714	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1715	cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1716	cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1717
1718	return (zilog);
1719}
1720
1721void
1722zil_free(zilog_t *zilog)
1723{
1724	zilog->zl_stop_sync = 1;
1725
1726	ASSERT0(zilog->zl_suspend);
1727	ASSERT0(zilog->zl_suspending);
1728
1729	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1730	list_destroy(&zilog->zl_lwb_list);
1731
1732	avl_destroy(&zilog->zl_vdev_tree);
1733	mutex_destroy(&zilog->zl_vdev_lock);
1734
1735	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1736	list_destroy(&zilog->zl_itx_commit_list);
1737
1738	for (int i = 0; i < TXG_SIZE; i++) {
1739		/*
1740		 * It's possible for an itx to be generated that doesn't dirty
1741		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1742		 * callback to remove the entry. We remove those here.
1743		 *
1744		 * Also free up the ziltest itxs.
1745		 */
1746		if (zilog->zl_itxg[i].itxg_itxs)
1747			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1748		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1749	}
1750
1751	mutex_destroy(&zilog->zl_lock);
1752
1753	cv_destroy(&zilog->zl_cv_writer);
1754	cv_destroy(&zilog->zl_cv_suspend);
1755	cv_destroy(&zilog->zl_cv_batch[0]);
1756	cv_destroy(&zilog->zl_cv_batch[1]);
1757
1758	kmem_free(zilog, sizeof (zilog_t));
1759}
1760
1761/*
1762 * Open an intent log.
1763 */
1764zilog_t *
1765zil_open(objset_t *os, zil_get_data_t *get_data)
1766{
1767	zilog_t *zilog = dmu_objset_zil(os);
1768
1769	ASSERT(zilog->zl_clean_taskq == NULL);
1770	ASSERT(zilog->zl_get_data == NULL);
1771	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1772
1773	zilog->zl_get_data = get_data;
1774	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1775	    2, 2, TASKQ_PREPOPULATE);
1776
1777	return (zilog);
1778}
1779
1780/*
1781 * Close an intent log.
1782 */
1783void
1784zil_close(zilog_t *zilog)
1785{
1786	lwb_t *lwb;
1787	uint64_t txg = 0;
1788
1789	zil_commit(zilog, 0); /* commit all itx */
1790
1791	/*
1792	 * The lwb_max_txg for the stubby lwb will reflect the last activity
1793	 * for the zil.  After a txg_wait_synced() on the txg we know all the
1794	 * callbacks have occurred that may clean the zil.  Only then can we
1795	 * destroy the zl_clean_taskq.
1796	 */
1797	mutex_enter(&zilog->zl_lock);
1798	lwb = list_tail(&zilog->zl_lwb_list);
1799	if (lwb != NULL)
1800		txg = lwb->lwb_max_txg;
1801	mutex_exit(&zilog->zl_lock);
1802	if (txg)
1803		txg_wait_synced(zilog->zl_dmu_pool, txg);
1804	ASSERT(!zilog_is_dirty(zilog));
1805
1806	taskq_destroy(zilog->zl_clean_taskq);
1807	zilog->zl_clean_taskq = NULL;
1808	zilog->zl_get_data = NULL;
1809
1810	/*
1811	 * We should have only one LWB left on the list; remove it now.
1812	 */
1813	mutex_enter(&zilog->zl_lock);
1814	lwb = list_head(&zilog->zl_lwb_list);
1815	if (lwb != NULL) {
1816		ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1817		list_remove(&zilog->zl_lwb_list, lwb);
1818		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1819		kmem_cache_free(zil_lwb_cache, lwb);
1820	}
1821	mutex_exit(&zilog->zl_lock);
1822}
1823
1824static char *suspend_tag = "zil suspending";
1825
1826/*
1827 * Suspend an intent log.  While in suspended mode, we still honor
1828 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1829 * On old version pools, we suspend the log briefly when taking a
1830 * snapshot so that it will have an empty intent log.
1831 *
1832 * Long holds are not really intended to be used the way we do here --
1833 * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
1834 * could fail.  Therefore we take pains to only put a long hold if it is
1835 * actually necessary.  Fortunately, it will only be necessary if the
1836 * objset is currently mounted (or the ZVOL equivalent).  In that case it
1837 * will already have a long hold, so we are not really making things any worse.
1838 *
1839 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1840 * zvol_state_t), and use their mechanism to prevent their hold from being
1841 * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
1842 * very little gain.
1843 *
1844 * if cookiep == NULL, this does both the suspend & resume.
1845 * Otherwise, it returns with the dataset "long held", and the cookie
1846 * should be passed into zil_resume().
1847 */
1848int
1849zil_suspend(const char *osname, void **cookiep)
1850{
1851	objset_t *os;
1852	zilog_t *zilog;
1853	const zil_header_t *zh;
1854	int error;
1855
1856	error = dmu_objset_hold(osname, suspend_tag, &os);
1857	if (error != 0)
1858		return (error);
1859	zilog = dmu_objset_zil(os);
1860
1861	mutex_enter(&zilog->zl_lock);
1862	zh = zilog->zl_header;
1863
1864	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
1865		mutex_exit(&zilog->zl_lock);
1866		dmu_objset_rele(os, suspend_tag);
1867		return (SET_ERROR(EBUSY));
1868	}
1869
1870	/*
1871	 * Don't put a long hold in the cases where we can avoid it.  This
1872	 * is when there is no cookie so we are doing a suspend & resume
1873	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1874	 * for the suspend because it's already suspended, or there's no ZIL.
1875	 */
1876	if (cookiep == NULL && !zilog->zl_suspending &&
1877	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1878		mutex_exit(&zilog->zl_lock);
1879		dmu_objset_rele(os, suspend_tag);
1880		return (0);
1881	}
1882
1883	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1884	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1885
1886	zilog->zl_suspend++;
1887
1888	if (zilog->zl_suspend > 1) {
1889		/*
1890		 * Someone else is already suspending it.
1891		 * Just wait for them to finish.
1892		 */
1893
1894		while (zilog->zl_suspending)
1895			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1896		mutex_exit(&zilog->zl_lock);
1897
1898		if (cookiep == NULL)
1899			zil_resume(os);
1900		else
1901			*cookiep = os;
1902		return (0);
1903	}
1904
1905	/*
1906	 * If there is no pointer to an on-disk block, this ZIL must not
1907	 * be active (e.g. filesystem not mounted), so there's nothing
1908	 * to clean up.
1909	 */
1910	if (BP_IS_HOLE(&zh->zh_log)) {
1911		ASSERT(cookiep != NULL); /* fast path already handled */
1912
1913		*cookiep = os;
1914		mutex_exit(&zilog->zl_lock);
1915		return (0);
1916	}
1917
1918	zilog->zl_suspending = B_TRUE;
1919	mutex_exit(&zilog->zl_lock);
1920
1921	zil_commit(zilog, 0);
1922
1923	zil_destroy(zilog, B_FALSE);
1924
1925	mutex_enter(&zilog->zl_lock);
1926	zilog->zl_suspending = B_FALSE;
1927	cv_broadcast(&zilog->zl_cv_suspend);
1928	mutex_exit(&zilog->zl_lock);
1929
1930	if (cookiep == NULL)
1931		zil_resume(os);
1932	else
1933		*cookiep = os;
1934	return (0);
1935}
1936
1937void
1938zil_resume(void *cookie)
1939{
1940	objset_t *os = cookie;
1941	zilog_t *zilog = dmu_objset_zil(os);
1942
1943	mutex_enter(&zilog->zl_lock);
1944	ASSERT(zilog->zl_suspend != 0);
1945	zilog->zl_suspend--;
1946	mutex_exit(&zilog->zl_lock);
1947	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1948	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1949}
1950
1951typedef struct zil_replay_arg {
1952	zil_replay_func_t **zr_replay;
1953	void		*zr_arg;
1954	boolean_t	zr_byteswap;
1955	char		*zr_lr;
1956} zil_replay_arg_t;
1957
1958static int
1959zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1960{
1961	char name[MAXNAMELEN];
1962
1963	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
1964
1965	dmu_objset_name(zilog->zl_os, name);
1966
1967	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1968	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1969	    (u_longlong_t)lr->lrc_seq,
1970	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1971	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1972
1973	return (error);
1974}
1975
1976static int
1977zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1978{
1979	zil_replay_arg_t *zr = zra;
1980	const zil_header_t *zh = zilog->zl_header;
1981	uint64_t reclen = lr->lrc_reclen;
1982	uint64_t txtype = lr->lrc_txtype;
1983	int error = 0;
1984
1985	zilog->zl_replaying_seq = lr->lrc_seq;
1986
1987	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1988		return (0);
1989
1990	if (lr->lrc_txg < claim_txg)		/* already committed */
1991		return (0);
1992
1993	/* Strip case-insensitive bit, still present in log record */
1994	txtype &= ~TX_CI;
1995
1996	if (txtype == 0 || txtype >= TX_MAX_TYPE)
1997		return (zil_replay_error(zilog, lr, EINVAL));
1998
1999	/*
2000	 * If this record type can be logged out of order, the object
2001	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
2002	 */
2003	if (TX_OOO(txtype)) {
2004		error = dmu_object_info(zilog->zl_os,
2005		    ((lr_ooo_t *)lr)->lr_foid, NULL);
2006		if (error == ENOENT || error == EEXIST)
2007			return (0);
2008	}
2009
2010	/*
2011	 * Make a copy of the data so we can revise and extend it.
2012	 */
2013	bcopy(lr, zr->zr_lr, reclen);
2014
2015	/*
2016	 * If this is a TX_WRITE with a blkptr, suck in the data.
2017	 */
2018	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2019		error = zil_read_log_data(zilog, (lr_write_t *)lr,
2020		    zr->zr_lr + reclen);
2021		if (error != 0)
2022			return (zil_replay_error(zilog, lr, error));
2023	}
2024
2025	/*
2026	 * The log block containing this lr may have been byteswapped
2027	 * so that we can easily examine common fields like lrc_txtype.
2028	 * However, the log is a mix of different record types, and only the
2029	 * replay vectors know how to byteswap their records.  Therefore, if
2030	 * the lr was byteswapped, undo it before invoking the replay vector.
2031	 */
2032	if (zr->zr_byteswap)
2033		byteswap_uint64_array(zr->zr_lr, reclen);
2034
2035	/*
2036	 * We must now do two things atomically: replay this log record,
2037	 * and update the log header sequence number to reflect the fact that
2038	 * we did so. At the end of each replay function the sequence number
2039	 * is updated if we are in replay mode.
2040	 */
2041	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2042	if (error != 0) {
2043		/*
2044		 * The DMU's dnode layer doesn't see removes until the txg
2045		 * commits, so a subsequent claim can spuriously fail with
2046		 * EEXIST. So if we receive any error we try syncing out
2047		 * any removes then retry the transaction.  Note that we
2048		 * specify B_FALSE for byteswap now, so we don't do it twice.
2049		 */
2050		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2051		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2052		if (error != 0)
2053			return (zil_replay_error(zilog, lr, error));
2054	}
2055	return (0);
2056}
2057
2058/* ARGSUSED */
2059static int
2060zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2061{
2062	zilog->zl_replay_blks++;
2063
2064	return (0);
2065}
2066
2067/*
2068 * If this dataset has a non-empty intent log, replay it and destroy it.
2069 */
2070void
2071zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2072{
2073	zilog_t *zilog = dmu_objset_zil(os);
2074	const zil_header_t *zh = zilog->zl_header;
2075	zil_replay_arg_t zr;
2076
2077	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2078		zil_destroy(zilog, B_TRUE);
2079		return;
2080	}
2081
2082	zr.zr_replay = replay_func;
2083	zr.zr_arg = arg;
2084	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2085	zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2086
2087	/*
2088	 * Wait for in-progress removes to sync before starting replay.
2089	 */
2090	txg_wait_synced(zilog->zl_dmu_pool, 0);
2091
2092	zilog->zl_replay = B_TRUE;
2093	zilog->zl_replay_time = ddi_get_lbolt();
2094	ASSERT(zilog->zl_replay_blks == 0);
2095	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2096	    zh->zh_claim_txg);
2097	kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2098
2099	zil_destroy(zilog, B_FALSE);
2100	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2101	zilog->zl_replay = B_FALSE;
2102}
2103
2104boolean_t
2105zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2106{
2107	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2108		return (B_TRUE);
2109
2110	if (zilog->zl_replay) {
2111		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2112		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2113		    zilog->zl_replaying_seq;
2114		return (B_TRUE);
2115	}
2116
2117	return (B_FALSE);
2118}
2119
2120/* ARGSUSED */
2121int
2122zil_vdev_offline(const char *osname, void *arg)
2123{
2124	int error;
2125
2126	error = zil_suspend(osname, NULL);
2127	if (error != 0)
2128		return (SET_ERROR(EEXIST));
2129	return (0);
2130}
2131