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