xref: /titanic_41/usr/src/uts/common/fs/zfs/zil.c (revision 24a1f0af9f770e0e795ef1fa1c6dece8dd8dc959)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/spa.h>
28 #include <sys/dmu.h>
29 #include <sys/zap.h>
30 #include <sys/arc.h>
31 #include <sys/stat.h>
32 #include <sys/resource.h>
33 #include <sys/zil.h>
34 #include <sys/zil_impl.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/vdev.h>
37 #include <sys/dmu_tx.h>
38 
39 /*
40  * The zfs intent log (ZIL) saves transaction records of system calls
41  * that change the file system in memory with enough information
42  * to be able to replay them. These are stored in memory until
43  * either the DMU transaction group (txg) commits them to the stable pool
44  * and they can be discarded, or they are flushed to the stable log
45  * (also in the pool) due to a fsync, O_DSYNC or other synchronous
46  * requirement. In the event of a panic or power fail then those log
47  * records (transactions) are replayed.
48  *
49  * There is one ZIL per file system. Its on-disk (pool) format consists
50  * of 3 parts:
51  *
52  * 	- ZIL header
53  * 	- ZIL blocks
54  * 	- ZIL records
55  *
56  * A log record holds a system call transaction. Log blocks can
57  * hold many log records and the blocks are chained together.
58  * Each ZIL block contains a block pointer (blkptr_t) to the next
59  * ZIL block in the chain. The ZIL header points to the first
60  * block in the chain. Note there is not a fixed place in the pool
61  * to hold blocks. They are dynamically allocated and freed as
62  * needed from the blocks available. Figure X shows the ZIL structure:
63  */
64 
65 /*
66  * This global ZIL switch affects all pools
67  */
68 int zil_disable = 0;	/* disable intent logging */
69 
70 /*
71  * Tunable parameter for debugging or performance analysis.  Setting
72  * zfs_nocacheflush will cause corruption on power loss if a volatile
73  * out-of-order write cache is enabled.
74  */
75 boolean_t zfs_nocacheflush = B_FALSE;
76 
77 static kmem_cache_t *zil_lwb_cache;
78 
79 static int
80 zil_dva_compare(const void *x1, const void *x2)
81 {
82 	const dva_t *dva1 = x1;
83 	const dva_t *dva2 = x2;
84 
85 	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
86 		return (-1);
87 	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
88 		return (1);
89 
90 	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
91 		return (-1);
92 	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
93 		return (1);
94 
95 	return (0);
96 }
97 
98 static void
99 zil_dva_tree_init(avl_tree_t *t)
100 {
101 	avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t),
102 	    offsetof(zil_dva_node_t, zn_node));
103 }
104 
105 static void
106 zil_dva_tree_fini(avl_tree_t *t)
107 {
108 	zil_dva_node_t *zn;
109 	void *cookie = NULL;
110 
111 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
112 		kmem_free(zn, sizeof (zil_dva_node_t));
113 
114 	avl_destroy(t);
115 }
116 
117 static int
118 zil_dva_tree_add(avl_tree_t *t, dva_t *dva)
119 {
120 	zil_dva_node_t *zn;
121 	avl_index_t where;
122 
123 	if (avl_find(t, dva, &where) != NULL)
124 		return (EEXIST);
125 
126 	zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP);
127 	zn->zn_dva = *dva;
128 	avl_insert(t, zn, where);
129 
130 	return (0);
131 }
132 
133 static zil_header_t *
134 zil_header_in_syncing_context(zilog_t *zilog)
135 {
136 	return ((zil_header_t *)zilog->zl_header);
137 }
138 
139 static void
140 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
141 {
142 	zio_cksum_t *zc = &bp->blk_cksum;
143 
144 	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
145 	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
146 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
147 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
148 }
149 
150 /*
151  * Read a log block, make sure it's valid, and byteswap it if necessary.
152  */
153 static int
154 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp)
155 {
156 	blkptr_t blk = *bp;
157 	zbookmark_t zb;
158 	uint32_t aflags = ARC_WAIT;
159 	int error;
160 
161 	zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET];
162 	zb.zb_object = 0;
163 	zb.zb_level = -1;
164 	zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
165 
166 	*abufpp = NULL;
167 
168 	/*
169 	 * We shouldn't be doing any scrubbing while we're doing log
170 	 * replay, it's OK to not lock.
171 	 */
172 	error = arc_read_nolock(NULL, zilog->zl_spa, &blk,
173 	    arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL |
174 	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb);
175 
176 	if (error == 0) {
177 		char *data = (*abufpp)->b_data;
178 		uint64_t blksz = BP_GET_LSIZE(bp);
179 		zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1;
180 		zio_cksum_t cksum = bp->blk_cksum;
181 
182 		/*
183 		 * Validate the checksummed log block.
184 		 *
185 		 * Sequence numbers should be... sequential.  The checksum
186 		 * verifier for the next block should be bp's checksum plus 1.
187 		 *
188 		 * Also check the log chain linkage and size used.
189 		 */
190 		cksum.zc_word[ZIL_ZC_SEQ]++;
191 
192 		if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum,
193 		    sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) ||
194 		    (ztp->zit_nused > (blksz - sizeof (zil_trailer_t)))) {
195 			error = ECKSUM;
196 		}
197 
198 		if (error) {
199 			VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1);
200 			*abufpp = NULL;
201 		}
202 	}
203 
204 	dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid);
205 
206 	return (error);
207 }
208 
209 /*
210  * Parse the intent log, and call parse_func for each valid record within.
211  * Return the highest sequence number.
212  */
213 uint64_t
214 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
215     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
216 {
217 	const zil_header_t *zh = zilog->zl_header;
218 	uint64_t claim_seq = zh->zh_claim_seq;
219 	uint64_t seq = 0;
220 	uint64_t max_seq = 0;
221 	blkptr_t blk = zh->zh_log;
222 	arc_buf_t *abuf;
223 	char *lrbuf, *lrp;
224 	zil_trailer_t *ztp;
225 	int reclen, error;
226 
227 	if (BP_IS_HOLE(&blk))
228 		return (max_seq);
229 
230 	/*
231 	 * Starting at the block pointed to by zh_log we read the log chain.
232 	 * For each block in the chain we strongly check that block to
233 	 * ensure its validity.  We stop when an invalid block is found.
234 	 * For each block pointer in the chain we call parse_blk_func().
235 	 * For each record in each valid block we call parse_lr_func().
236 	 * If the log has been claimed, stop if we encounter a sequence
237 	 * number greater than the highest claimed sequence number.
238 	 */
239 	zil_dva_tree_init(&zilog->zl_dva_tree);
240 	for (;;) {
241 		seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
242 
243 		if (claim_seq != 0 && seq > claim_seq)
244 			break;
245 
246 		ASSERT(max_seq < seq);
247 		max_seq = seq;
248 
249 		error = zil_read_log_block(zilog, &blk, &abuf);
250 
251 		if (parse_blk_func != NULL)
252 			parse_blk_func(zilog, &blk, arg, txg);
253 
254 		if (error)
255 			break;
256 
257 		lrbuf = abuf->b_data;
258 		ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1;
259 		blk = ztp->zit_next_blk;
260 
261 		if (parse_lr_func == NULL) {
262 			VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
263 			continue;
264 		}
265 
266 		for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) {
267 			lr_t *lr = (lr_t *)lrp;
268 			reclen = lr->lrc_reclen;
269 			ASSERT3U(reclen, >=, sizeof (lr_t));
270 			parse_lr_func(zilog, lr, arg, txg);
271 		}
272 		VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
273 	}
274 	zil_dva_tree_fini(&zilog->zl_dva_tree);
275 
276 	return (max_seq);
277 }
278 
279 /* ARGSUSED */
280 static void
281 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
282 {
283 	spa_t *spa = zilog->zl_spa;
284 	int err;
285 
286 	/*
287 	 * Claim log block if not already committed and not already claimed.
288 	 */
289 	if (bp->blk_birth >= first_txg &&
290 	    zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) {
291 		err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL,
292 		    ZIO_FLAG_MUSTSUCCEED));
293 		ASSERT(err == 0);
294 	}
295 }
296 
297 static void
298 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
299 {
300 	if (lrc->lrc_txtype == TX_WRITE) {
301 		lr_write_t *lr = (lr_write_t *)lrc;
302 		zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg);
303 	}
304 }
305 
306 /* ARGSUSED */
307 static void
308 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
309 {
310 	zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx));
311 }
312 
313 static void
314 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
315 {
316 	/*
317 	 * If we previously claimed it, we need to free it.
318 	 */
319 	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) {
320 		lr_write_t *lr = (lr_write_t *)lrc;
321 		blkptr_t *bp = &lr->lr_blkptr;
322 		if (bp->blk_birth >= claim_txg &&
323 		    !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) {
324 			(void) arc_free(NULL, zilog->zl_spa,
325 			    dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT);
326 		}
327 	}
328 }
329 
330 /*
331  * Create an on-disk intent log.
332  */
333 static void
334 zil_create(zilog_t *zilog)
335 {
336 	const zil_header_t *zh = zilog->zl_header;
337 	lwb_t *lwb;
338 	uint64_t txg = 0;
339 	dmu_tx_t *tx = NULL;
340 	blkptr_t blk;
341 	int error = 0;
342 
343 	/*
344 	 * Wait for any previous destroy to complete.
345 	 */
346 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
347 
348 	ASSERT(zh->zh_claim_txg == 0);
349 	ASSERT(zh->zh_replay_seq == 0);
350 
351 	blk = zh->zh_log;
352 
353 	/*
354 	 * If we don't already have an initial log block or we have one
355 	 * but it's the wrong endianness then allocate one.
356 	 */
357 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
358 		tx = dmu_tx_create(zilog->zl_os);
359 		(void) dmu_tx_assign(tx, TXG_WAIT);
360 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
361 		txg = dmu_tx_get_txg(tx);
362 
363 		if (!BP_IS_HOLE(&blk)) {
364 			zio_free_blk(zilog->zl_spa, &blk, txg);
365 			BP_ZERO(&blk);
366 		}
367 
368 		error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk,
369 		    NULL, txg);
370 
371 		if (error == 0)
372 			zil_init_log_chain(zilog, &blk);
373 	}
374 
375 	/*
376 	 * Allocate a log write buffer (lwb) for the first log block.
377 	 */
378 	if (error == 0) {
379 		lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
380 		lwb->lwb_zilog = zilog;
381 		lwb->lwb_blk = blk;
382 		lwb->lwb_nused = 0;
383 		lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk);
384 		lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz);
385 		lwb->lwb_max_txg = txg;
386 		lwb->lwb_zio = NULL;
387 
388 		mutex_enter(&zilog->zl_lock);
389 		list_insert_tail(&zilog->zl_lwb_list, lwb);
390 		mutex_exit(&zilog->zl_lock);
391 	}
392 
393 	/*
394 	 * If we just allocated the first log block, commit our transaction
395 	 * and wait for zil_sync() to stuff the block poiner into zh_log.
396 	 * (zh is part of the MOS, so we cannot modify it in open context.)
397 	 */
398 	if (tx != NULL) {
399 		dmu_tx_commit(tx);
400 		txg_wait_synced(zilog->zl_dmu_pool, txg);
401 	}
402 
403 	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
404 }
405 
406 /*
407  * In one tx, free all log blocks and clear the log header.
408  * If keep_first is set, then we're replaying a log with no content.
409  * We want to keep the first block, however, so that the first
410  * synchronous transaction doesn't require a txg_wait_synced()
411  * in zil_create().  We don't need to txg_wait_synced() here either
412  * when keep_first is set, because both zil_create() and zil_destroy()
413  * will wait for any in-progress destroys to complete.
414  */
415 void
416 zil_destroy(zilog_t *zilog, boolean_t keep_first)
417 {
418 	const zil_header_t *zh = zilog->zl_header;
419 	lwb_t *lwb;
420 	dmu_tx_t *tx;
421 	uint64_t txg;
422 
423 	/*
424 	 * Wait for any previous destroy to complete.
425 	 */
426 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
427 
428 	if (BP_IS_HOLE(&zh->zh_log))
429 		return;
430 
431 	tx = dmu_tx_create(zilog->zl_os);
432 	(void) dmu_tx_assign(tx, TXG_WAIT);
433 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
434 	txg = dmu_tx_get_txg(tx);
435 
436 	mutex_enter(&zilog->zl_lock);
437 
438 	/*
439 	 * It is possible for the ZIL to get the previously mounted zilog
440 	 * structure of the same dataset if quickly remounted and the dbuf
441 	 * eviction has not completed. In this case we can see a non
442 	 * empty lwb list and keep_first will be set. We fix this by
443 	 * clearing the keep_first. This will be slower but it's very rare.
444 	 */
445 	if (!list_is_empty(&zilog->zl_lwb_list) && keep_first)
446 		keep_first = B_FALSE;
447 
448 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
449 	zilog->zl_destroy_txg = txg;
450 	zilog->zl_keep_first = keep_first;
451 
452 	if (!list_is_empty(&zilog->zl_lwb_list)) {
453 		ASSERT(zh->zh_claim_txg == 0);
454 		ASSERT(!keep_first);
455 		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
456 			list_remove(&zilog->zl_lwb_list, lwb);
457 			if (lwb->lwb_buf != NULL)
458 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
459 			zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg);
460 			kmem_cache_free(zil_lwb_cache, lwb);
461 		}
462 	} else {
463 		if (!keep_first) {
464 			(void) zil_parse(zilog, zil_free_log_block,
465 			    zil_free_log_record, tx, zh->zh_claim_txg);
466 		}
467 	}
468 	mutex_exit(&zilog->zl_lock);
469 
470 	dmu_tx_commit(tx);
471 }
472 
473 /*
474  * zil_rollback_destroy() is only called by the rollback code.
475  * We already have a syncing tx. Rollback has exclusive access to the
476  * dataset, so we don't have to worry about concurrent zil access.
477  * The actual freeing of any log blocks occurs in zil_sync() later in
478  * this txg syncing phase.
479  */
480 void
481 zil_rollback_destroy(zilog_t *zilog, dmu_tx_t *tx)
482 {
483 	const zil_header_t *zh = zilog->zl_header;
484 	uint64_t txg;
485 
486 	if (BP_IS_HOLE(&zh->zh_log))
487 		return;
488 
489 	txg = dmu_tx_get_txg(tx);
490 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
491 	zilog->zl_destroy_txg = txg;
492 	zilog->zl_keep_first = B_FALSE;
493 
494 	/*
495 	 * Ensure there's no outstanding ZIL IO.  No lwbs or just the
496 	 * unused one that allocated in advance is ok.
497 	 */
498 	ASSERT(zilog->zl_lwb_list.list_head.list_next ==
499 	    zilog->zl_lwb_list.list_head.list_prev);
500 	(void) zil_parse(zilog, zil_free_log_block, zil_free_log_record,
501 	    tx, zh->zh_claim_txg);
502 }
503 
504 int
505 zil_claim(char *osname, void *txarg)
506 {
507 	dmu_tx_t *tx = txarg;
508 	uint64_t first_txg = dmu_tx_get_txg(tx);
509 	zilog_t *zilog;
510 	zil_header_t *zh;
511 	objset_t *os;
512 	int error;
513 
514 	error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
515 	if (error) {
516 		cmn_err(CE_WARN, "can't open objset for %s", osname);
517 		return (0);
518 	}
519 
520 	zilog = dmu_objset_zil(os);
521 	zh = zil_header_in_syncing_context(zilog);
522 
523 	/*
524 	 * Claim all log blocks if we haven't already done so, and remember
525 	 * the highest claimed sequence number.  This ensures that if we can
526 	 * read only part of the log now (e.g. due to a missing device),
527 	 * but we can read the entire log later, we will not try to replay
528 	 * or destroy beyond the last block we successfully claimed.
529 	 */
530 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
531 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
532 		zh->zh_claim_txg = first_txg;
533 		zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block,
534 		    zil_claim_log_record, tx, first_txg);
535 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
536 	}
537 
538 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
539 	dmu_objset_close(os);
540 	return (0);
541 }
542 
543 /*
544  * Check the log by walking the log chain.
545  * Checksum errors are ok as they indicate the end of the chain.
546  * Any other error (no device or read failure) returns an error.
547  */
548 /* ARGSUSED */
549 int
550 zil_check_log_chain(char *osname, void *txarg)
551 {
552 	zilog_t *zilog;
553 	zil_header_t *zh;
554 	blkptr_t blk;
555 	arc_buf_t *abuf;
556 	objset_t *os;
557 	char *lrbuf;
558 	zil_trailer_t *ztp;
559 	int error;
560 
561 	error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
562 	if (error) {
563 		cmn_err(CE_WARN, "can't open objset for %s", osname);
564 		return (0);
565 	}
566 
567 	zilog = dmu_objset_zil(os);
568 	zh = zil_header_in_syncing_context(zilog);
569 	blk = zh->zh_log;
570 	if (BP_IS_HOLE(&blk)) {
571 		dmu_objset_close(os);
572 		return (0); /* no chain */
573 	}
574 
575 	for (;;) {
576 		error = zil_read_log_block(zilog, &blk, &abuf);
577 		if (error)
578 			break;
579 		lrbuf = abuf->b_data;
580 		ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1;
581 		blk = ztp->zit_next_blk;
582 		VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
583 	}
584 	dmu_objset_close(os);
585 	if (error == ECKSUM)
586 		return (0); /* normal end of chain */
587 	return (error);
588 }
589 
590 /*
591  * Clear a log chain
592  */
593 /* ARGSUSED */
594 int
595 zil_clear_log_chain(char *osname, void *txarg)
596 {
597 	zilog_t *zilog;
598 	zil_header_t *zh;
599 	objset_t *os;
600 	dmu_tx_t *tx;
601 	int error;
602 
603 	error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
604 	if (error) {
605 		cmn_err(CE_WARN, "can't open objset for %s", osname);
606 		return (0);
607 	}
608 
609 	zilog = dmu_objset_zil(os);
610 	tx = dmu_tx_create(zilog->zl_os);
611 	(void) dmu_tx_assign(tx, TXG_WAIT);
612 	zh = zil_header_in_syncing_context(zilog);
613 	BP_ZERO(&zh->zh_log);
614 	dsl_dataset_dirty(dmu_objset_ds(os), tx);
615 	dmu_tx_commit(tx);
616 	dmu_objset_close(os);
617 	return (0);
618 }
619 
620 static int
621 zil_vdev_compare(const void *x1, const void *x2)
622 {
623 	uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
624 	uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
625 
626 	if (v1 < v2)
627 		return (-1);
628 	if (v1 > v2)
629 		return (1);
630 
631 	return (0);
632 }
633 
634 void
635 zil_add_block(zilog_t *zilog, blkptr_t *bp)
636 {
637 	avl_tree_t *t = &zilog->zl_vdev_tree;
638 	avl_index_t where;
639 	zil_vdev_node_t *zv, zvsearch;
640 	int ndvas = BP_GET_NDVAS(bp);
641 	int i;
642 
643 	if (zfs_nocacheflush)
644 		return;
645 
646 	ASSERT(zilog->zl_writer);
647 
648 	/*
649 	 * Even though we're zl_writer, we still need a lock because the
650 	 * zl_get_data() callbacks may have dmu_sync() done callbacks
651 	 * that will run concurrently.
652 	 */
653 	mutex_enter(&zilog->zl_vdev_lock);
654 	for (i = 0; i < ndvas; i++) {
655 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
656 		if (avl_find(t, &zvsearch, &where) == NULL) {
657 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
658 			zv->zv_vdev = zvsearch.zv_vdev;
659 			avl_insert(t, zv, where);
660 		}
661 	}
662 	mutex_exit(&zilog->zl_vdev_lock);
663 }
664 
665 void
666 zil_flush_vdevs(zilog_t *zilog)
667 {
668 	spa_t *spa = zilog->zl_spa;
669 	avl_tree_t *t = &zilog->zl_vdev_tree;
670 	void *cookie = NULL;
671 	zil_vdev_node_t *zv;
672 	zio_t *zio;
673 
674 	ASSERT(zilog->zl_writer);
675 
676 	/*
677 	 * We don't need zl_vdev_lock here because we're the zl_writer,
678 	 * and all zl_get_data() callbacks are done.
679 	 */
680 	if (avl_numnodes(t) == 0)
681 		return;
682 
683 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
684 
685 	zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
686 
687 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
688 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
689 		if (vd != NULL)
690 			zio_flush(zio, vd);
691 		kmem_free(zv, sizeof (*zv));
692 	}
693 
694 	/*
695 	 * Wait for all the flushes to complete.  Not all devices actually
696 	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
697 	 */
698 	(void) zio_wait(zio);
699 
700 	spa_config_exit(spa, SCL_STATE, FTAG);
701 }
702 
703 /*
704  * Function called when a log block write completes
705  */
706 static void
707 zil_lwb_write_done(zio_t *zio)
708 {
709 	lwb_t *lwb = zio->io_private;
710 	zilog_t *zilog = lwb->lwb_zilog;
711 
712 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
713 	ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG);
714 	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
715 	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
716 	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
717 	ASSERT(!BP_IS_GANG(zio->io_bp));
718 	ASSERT(!BP_IS_HOLE(zio->io_bp));
719 	ASSERT(zio->io_bp->blk_fill == 0);
720 
721 	/*
722 	 * Now that we've written this log block, we have a stable pointer
723 	 * to the next block in the chain, so it's OK to let the txg in
724 	 * which we allocated the next block sync.
725 	 */
726 	txg_rele_to_sync(&lwb->lwb_txgh);
727 
728 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
729 	mutex_enter(&zilog->zl_lock);
730 	lwb->lwb_buf = NULL;
731 	if (zio->io_error)
732 		zilog->zl_log_error = B_TRUE;
733 	mutex_exit(&zilog->zl_lock);
734 }
735 
736 /*
737  * Initialize the io for a log block.
738  */
739 static void
740 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
741 {
742 	zbookmark_t zb;
743 
744 	zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET];
745 	zb.zb_object = 0;
746 	zb.zb_level = -1;
747 	zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
748 
749 	if (zilog->zl_root_zio == NULL) {
750 		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
751 		    ZIO_FLAG_CANFAIL);
752 	}
753 	if (lwb->lwb_zio == NULL) {
754 		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
755 		    0, &lwb->lwb_blk, lwb->lwb_buf,
756 		    lwb->lwb_sz, zil_lwb_write_done, lwb,
757 		    ZIO_PRIORITY_LOG_WRITE, ZIO_FLAG_CANFAIL, &zb);
758 	}
759 }
760 
761 /*
762  * Start a log block write and advance to the next log block.
763  * Calls are serialized.
764  */
765 static lwb_t *
766 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
767 {
768 	lwb_t *nlwb;
769 	zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1;
770 	spa_t *spa = zilog->zl_spa;
771 	blkptr_t *bp = &ztp->zit_next_blk;
772 	uint64_t txg;
773 	uint64_t zil_blksz;
774 	int error;
775 
776 	ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb));
777 
778 	/*
779 	 * Allocate the next block and save its address in this block
780 	 * before writing it in order to establish the log chain.
781 	 * Note that if the allocation of nlwb synced before we wrote
782 	 * the block that points at it (lwb), we'd leak it if we crashed.
783 	 * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done().
784 	 */
785 	txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh);
786 	txg_rele_to_quiesce(&lwb->lwb_txgh);
787 
788 	/*
789 	 * Pick a ZIL blocksize. We request a size that is the
790 	 * maximum of the previous used size, the current used size and
791 	 * the amount waiting in the queue.
792 	 */
793 	zil_blksz = MAX(zilog->zl_prev_used,
794 	    zilog->zl_cur_used + sizeof (*ztp));
795 	zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp));
796 	zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t);
797 	if (zil_blksz > ZIL_MAX_BLKSZ)
798 		zil_blksz = ZIL_MAX_BLKSZ;
799 
800 	BP_ZERO(bp);
801 	/* pass the old blkptr in order to spread log blocks across devs */
802 	error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg);
803 	if (error) {
804 		dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg);
805 
806 		/*
807 		 * We dirty the dataset to ensure that zil_sync() will
808 		 * be called to remove this lwb from our zl_lwb_list.
809 		 * Failing to do so, may leave an lwb with a NULL lwb_buf
810 		 * hanging around on the zl_lwb_list.
811 		 */
812 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
813 		dmu_tx_commit(tx);
814 
815 		/*
816 		 * Since we've just experienced an allocation failure so we
817 		 * terminate the current lwb and send it on its way.
818 		 */
819 		ztp->zit_pad = 0;
820 		ztp->zit_nused = lwb->lwb_nused;
821 		ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
822 		zio_nowait(lwb->lwb_zio);
823 
824 		/*
825 		 * By returning NULL the caller will call tx_wait_synced()
826 		 */
827 		return (NULL);
828 	}
829 
830 	ASSERT3U(bp->blk_birth, ==, txg);
831 	ztp->zit_pad = 0;
832 	ztp->zit_nused = lwb->lwb_nused;
833 	ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
834 	bp->blk_cksum = lwb->lwb_blk.blk_cksum;
835 	bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
836 
837 	/*
838 	 * Allocate a new log write buffer (lwb).
839 	 */
840 	nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
841 
842 	nlwb->lwb_zilog = zilog;
843 	nlwb->lwb_blk = *bp;
844 	nlwb->lwb_nused = 0;
845 	nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk);
846 	nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz);
847 	nlwb->lwb_max_txg = txg;
848 	nlwb->lwb_zio = NULL;
849 
850 	/*
851 	 * Put new lwb at the end of the log chain
852 	 */
853 	mutex_enter(&zilog->zl_lock);
854 	list_insert_tail(&zilog->zl_lwb_list, nlwb);
855 	mutex_exit(&zilog->zl_lock);
856 
857 	/* Record the block for later vdev flushing */
858 	zil_add_block(zilog, &lwb->lwb_blk);
859 
860 	/*
861 	 * kick off the write for the old log block
862 	 */
863 	dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg);
864 	ASSERT(lwb->lwb_zio);
865 	zio_nowait(lwb->lwb_zio);
866 
867 	return (nlwb);
868 }
869 
870 static lwb_t *
871 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
872 {
873 	lr_t *lrc = &itx->itx_lr; /* common log record */
874 	lr_write_t *lr = (lr_write_t *)lrc;
875 	uint64_t txg = lrc->lrc_txg;
876 	uint64_t reclen = lrc->lrc_reclen;
877 	uint64_t dlen;
878 
879 	if (lwb == NULL)
880 		return (NULL);
881 	ASSERT(lwb->lwb_buf != NULL);
882 
883 	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
884 		dlen = P2ROUNDUP_TYPED(
885 		    lr->lr_length, sizeof (uint64_t), uint64_t);
886 	else
887 		dlen = 0;
888 
889 	zilog->zl_cur_used += (reclen + dlen);
890 
891 	zil_lwb_write_init(zilog, lwb);
892 
893 	/*
894 	 * If this record won't fit in the current log block, start a new one.
895 	 */
896 	if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
897 		lwb = zil_lwb_write_start(zilog, lwb);
898 		if (lwb == NULL)
899 			return (NULL);
900 		zil_lwb_write_init(zilog, lwb);
901 		ASSERT(lwb->lwb_nused == 0);
902 		if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
903 			txg_wait_synced(zilog->zl_dmu_pool, txg);
904 			return (lwb);
905 		}
906 	}
907 
908 	/*
909 	 * Update the lrc_seq, to be log record sequence number. See zil.h
910 	 * Then copy the record to the log buffer.
911 	 */
912 	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
913 	bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen);
914 
915 	/*
916 	 * If it's a write, fetch the data or get its blkptr as appropriate.
917 	 */
918 	if (lrc->lrc_txtype == TX_WRITE) {
919 		if (txg > spa_freeze_txg(zilog->zl_spa))
920 			txg_wait_synced(zilog->zl_dmu_pool, txg);
921 		if (itx->itx_wr_state != WR_COPIED) {
922 			char *dbuf;
923 			int error;
924 
925 			/* alignment is guaranteed */
926 			lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused);
927 			if (dlen) {
928 				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
929 				dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen;
930 				lr->lr_common.lrc_reclen += dlen;
931 			} else {
932 				ASSERT(itx->itx_wr_state == WR_INDIRECT);
933 				dbuf = NULL;
934 			}
935 			error = zilog->zl_get_data(
936 			    itx->itx_private, lr, dbuf, lwb->lwb_zio);
937 			if (error) {
938 				ASSERT(error == ENOENT || error == EEXIST ||
939 				    error == EALREADY);
940 				return (lwb);
941 			}
942 		}
943 	}
944 
945 	lwb->lwb_nused += reclen + dlen;
946 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
947 	ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb));
948 	ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
949 
950 	return (lwb);
951 }
952 
953 itx_t *
954 zil_itx_create(uint64_t txtype, size_t lrsize)
955 {
956 	itx_t *itx;
957 
958 	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
959 
960 	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
961 	itx->itx_lr.lrc_txtype = txtype;
962 	itx->itx_lr.lrc_reclen = lrsize;
963 	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
964 	itx->itx_lr.lrc_seq = 0;	/* defensive */
965 
966 	return (itx);
967 }
968 
969 uint64_t
970 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
971 {
972 	uint64_t seq;
973 
974 	ASSERT(itx->itx_lr.lrc_seq == 0);
975 
976 	mutex_enter(&zilog->zl_lock);
977 	list_insert_tail(&zilog->zl_itx_list, itx);
978 	zilog->zl_itx_list_sz += itx->itx_sod;
979 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
980 	itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq;
981 	mutex_exit(&zilog->zl_lock);
982 
983 	return (seq);
984 }
985 
986 /*
987  * Free up all in-memory intent log transactions that have now been synced.
988  */
989 static void
990 zil_itx_clean(zilog_t *zilog)
991 {
992 	uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa);
993 	uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa);
994 	list_t clean_list;
995 	itx_t *itx;
996 
997 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
998 
999 	mutex_enter(&zilog->zl_lock);
1000 	/* wait for a log writer to finish walking list */
1001 	while (zilog->zl_writer) {
1002 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1003 	}
1004 
1005 	/*
1006 	 * Move the sync'd log transactions to a separate list so we can call
1007 	 * kmem_free without holding the zl_lock.
1008 	 *
1009 	 * There is no need to set zl_writer as we don't drop zl_lock here
1010 	 */
1011 	while ((itx = list_head(&zilog->zl_itx_list)) != NULL &&
1012 	    itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) {
1013 		list_remove(&zilog->zl_itx_list, itx);
1014 		zilog->zl_itx_list_sz -= itx->itx_sod;
1015 		list_insert_tail(&clean_list, itx);
1016 	}
1017 	cv_broadcast(&zilog->zl_cv_writer);
1018 	mutex_exit(&zilog->zl_lock);
1019 
1020 	/* destroy sync'd log transactions */
1021 	while ((itx = list_head(&clean_list)) != NULL) {
1022 		list_remove(&clean_list, itx);
1023 		kmem_free(itx, offsetof(itx_t, itx_lr)
1024 		    + itx->itx_lr.lrc_reclen);
1025 	}
1026 	list_destroy(&clean_list);
1027 }
1028 
1029 /*
1030  * If there are any in-memory intent log transactions which have now been
1031  * synced then start up a taskq to free them.
1032  */
1033 void
1034 zil_clean(zilog_t *zilog)
1035 {
1036 	itx_t *itx;
1037 
1038 	mutex_enter(&zilog->zl_lock);
1039 	itx = list_head(&zilog->zl_itx_list);
1040 	if ((itx != NULL) &&
1041 	    (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) {
1042 		(void) taskq_dispatch(zilog->zl_clean_taskq,
1043 		    (void (*)(void *))zil_itx_clean, zilog, TQ_NOSLEEP);
1044 	}
1045 	mutex_exit(&zilog->zl_lock);
1046 }
1047 
1048 static void
1049 zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid)
1050 {
1051 	uint64_t txg;
1052 	uint64_t commit_seq = 0;
1053 	itx_t *itx, *itx_next = (itx_t *)-1;
1054 	lwb_t *lwb;
1055 	spa_t *spa;
1056 
1057 	zilog->zl_writer = B_TRUE;
1058 	ASSERT(zilog->zl_root_zio == NULL);
1059 	spa = zilog->zl_spa;
1060 
1061 	if (zilog->zl_suspend) {
1062 		lwb = NULL;
1063 	} else {
1064 		lwb = list_tail(&zilog->zl_lwb_list);
1065 		if (lwb == NULL) {
1066 			/*
1067 			 * Return if there's nothing to flush before we
1068 			 * dirty the fs by calling zil_create()
1069 			 */
1070 			if (list_is_empty(&zilog->zl_itx_list)) {
1071 				zilog->zl_writer = B_FALSE;
1072 				return;
1073 			}
1074 			mutex_exit(&zilog->zl_lock);
1075 			zil_create(zilog);
1076 			mutex_enter(&zilog->zl_lock);
1077 			lwb = list_tail(&zilog->zl_lwb_list);
1078 		}
1079 	}
1080 
1081 	/* Loop through in-memory log transactions filling log blocks. */
1082 	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1083 	for (;;) {
1084 		/*
1085 		 * Find the next itx to push:
1086 		 * Push all transactions related to specified foid and all
1087 		 * other transactions except TX_WRITE, TX_TRUNCATE,
1088 		 * TX_SETATTR and TX_ACL for all other files.
1089 		 */
1090 		if (itx_next != (itx_t *)-1)
1091 			itx = itx_next;
1092 		else
1093 			itx = list_head(&zilog->zl_itx_list);
1094 		for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) {
1095 			if (foid == 0) /* push all foids? */
1096 				break;
1097 			if (itx->itx_sync) /* push all O_[D]SYNC */
1098 				break;
1099 			switch (itx->itx_lr.lrc_txtype) {
1100 			case TX_SETATTR:
1101 			case TX_WRITE:
1102 			case TX_TRUNCATE:
1103 			case TX_ACL:
1104 				/* lr_foid is same offset for these records */
1105 				if (((lr_write_t *)&itx->itx_lr)->lr_foid
1106 				    != foid) {
1107 					continue; /* skip this record */
1108 				}
1109 			}
1110 			break;
1111 		}
1112 		if (itx == NULL)
1113 			break;
1114 
1115 		if ((itx->itx_lr.lrc_seq > seq) &&
1116 		    ((lwb == NULL) || (lwb->lwb_nused == 0) ||
1117 		    (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) {
1118 			break;
1119 		}
1120 
1121 		/*
1122 		 * Save the next pointer.  Even though we soon drop
1123 		 * zl_lock all threads that may change the list
1124 		 * (another writer or zil_itx_clean) can't do so until
1125 		 * they have zl_writer.
1126 		 */
1127 		itx_next = list_next(&zilog->zl_itx_list, itx);
1128 		list_remove(&zilog->zl_itx_list, itx);
1129 		zilog->zl_itx_list_sz -= itx->itx_sod;
1130 		mutex_exit(&zilog->zl_lock);
1131 		txg = itx->itx_lr.lrc_txg;
1132 		ASSERT(txg);
1133 
1134 		if (txg > spa_last_synced_txg(spa) ||
1135 		    txg > spa_freeze_txg(spa))
1136 			lwb = zil_lwb_commit(zilog, itx, lwb);
1137 		kmem_free(itx, offsetof(itx_t, itx_lr)
1138 		    + itx->itx_lr.lrc_reclen);
1139 		mutex_enter(&zilog->zl_lock);
1140 	}
1141 	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1142 	/* determine commit sequence number */
1143 	itx = list_head(&zilog->zl_itx_list);
1144 	if (itx)
1145 		commit_seq = itx->itx_lr.lrc_seq;
1146 	else
1147 		commit_seq = zilog->zl_itx_seq;
1148 	mutex_exit(&zilog->zl_lock);
1149 
1150 	/* write the last block out */
1151 	if (lwb != NULL && lwb->lwb_zio != NULL)
1152 		lwb = zil_lwb_write_start(zilog, lwb);
1153 
1154 	zilog->zl_prev_used = zilog->zl_cur_used;
1155 	zilog->zl_cur_used = 0;
1156 
1157 	/*
1158 	 * Wait if necessary for the log blocks to be on stable storage.
1159 	 */
1160 	if (zilog->zl_root_zio) {
1161 		DTRACE_PROBE1(zil__cw3, zilog_t *, zilog);
1162 		(void) zio_wait(zilog->zl_root_zio);
1163 		zilog->zl_root_zio = NULL;
1164 		DTRACE_PROBE1(zil__cw4, zilog_t *, zilog);
1165 		zil_flush_vdevs(zilog);
1166 	}
1167 
1168 	if (zilog->zl_log_error || lwb == NULL) {
1169 		zilog->zl_log_error = 0;
1170 		txg_wait_synced(zilog->zl_dmu_pool, 0);
1171 	}
1172 
1173 	mutex_enter(&zilog->zl_lock);
1174 	zilog->zl_writer = B_FALSE;
1175 
1176 	ASSERT3U(commit_seq, >=, zilog->zl_commit_seq);
1177 	zilog->zl_commit_seq = commit_seq;
1178 }
1179 
1180 /*
1181  * Push zfs transactions to stable storage up to the supplied sequence number.
1182  * If foid is 0 push out all transactions, otherwise push only those
1183  * for that file or might have been used to create that file.
1184  */
1185 void
1186 zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid)
1187 {
1188 	if (zilog == NULL || seq == 0)
1189 		return;
1190 
1191 	mutex_enter(&zilog->zl_lock);
1192 
1193 	seq = MIN(seq, zilog->zl_itx_seq);	/* cap seq at largest itx seq */
1194 
1195 	while (zilog->zl_writer) {
1196 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1197 		if (seq < zilog->zl_commit_seq) {
1198 			mutex_exit(&zilog->zl_lock);
1199 			return;
1200 		}
1201 	}
1202 	zil_commit_writer(zilog, seq, foid); /* drops zl_lock */
1203 	/* wake up others waiting on the commit */
1204 	cv_broadcast(&zilog->zl_cv_writer);
1205 	mutex_exit(&zilog->zl_lock);
1206 }
1207 
1208 /*
1209  * Called in syncing context to free committed log blocks and update log header.
1210  */
1211 void
1212 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1213 {
1214 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1215 	uint64_t txg = dmu_tx_get_txg(tx);
1216 	spa_t *spa = zilog->zl_spa;
1217 	lwb_t *lwb;
1218 
1219 	mutex_enter(&zilog->zl_lock);
1220 
1221 	ASSERT(zilog->zl_stop_sync == 0);
1222 
1223 	zh->zh_replay_seq = zilog->zl_replayed_seq[txg & TXG_MASK];
1224 
1225 	if (zilog->zl_destroy_txg == txg) {
1226 		blkptr_t blk = zh->zh_log;
1227 
1228 		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1229 		ASSERT(spa_sync_pass(spa) == 1);
1230 
1231 		bzero(zh, sizeof (zil_header_t));
1232 		bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1233 
1234 		if (zilog->zl_keep_first) {
1235 			/*
1236 			 * If this block was part of log chain that couldn't
1237 			 * be claimed because a device was missing during
1238 			 * zil_claim(), but that device later returns,
1239 			 * then this block could erroneously appear valid.
1240 			 * To guard against this, assign a new GUID to the new
1241 			 * log chain so it doesn't matter what blk points to.
1242 			 */
1243 			zil_init_log_chain(zilog, &blk);
1244 			zh->zh_log = blk;
1245 		}
1246 	}
1247 
1248 	for (;;) {
1249 		lwb = list_head(&zilog->zl_lwb_list);
1250 		if (lwb == NULL) {
1251 			mutex_exit(&zilog->zl_lock);
1252 			return;
1253 		}
1254 		zh->zh_log = lwb->lwb_blk;
1255 		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1256 			break;
1257 		list_remove(&zilog->zl_lwb_list, lwb);
1258 		zio_free_blk(spa, &lwb->lwb_blk, txg);
1259 		kmem_cache_free(zil_lwb_cache, lwb);
1260 
1261 		/*
1262 		 * If we don't have anything left in the lwb list then
1263 		 * we've had an allocation failure and we need to zero
1264 		 * out the zil_header blkptr so that we don't end
1265 		 * up freeing the same block twice.
1266 		 */
1267 		if (list_head(&zilog->zl_lwb_list) == NULL)
1268 			BP_ZERO(&zh->zh_log);
1269 	}
1270 	mutex_exit(&zilog->zl_lock);
1271 }
1272 
1273 void
1274 zil_init(void)
1275 {
1276 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1277 	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1278 }
1279 
1280 void
1281 zil_fini(void)
1282 {
1283 	kmem_cache_destroy(zil_lwb_cache);
1284 }
1285 
1286 zilog_t *
1287 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1288 {
1289 	zilog_t *zilog;
1290 
1291 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1292 
1293 	zilog->zl_header = zh_phys;
1294 	zilog->zl_os = os;
1295 	zilog->zl_spa = dmu_objset_spa(os);
1296 	zilog->zl_dmu_pool = dmu_objset_pool(os);
1297 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1298 
1299 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1300 
1301 	list_create(&zilog->zl_itx_list, sizeof (itx_t),
1302 	    offsetof(itx_t, itx_node));
1303 
1304 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1305 	    offsetof(lwb_t, lwb_node));
1306 
1307 	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1308 
1309 	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1310 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1311 
1312 	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1313 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1314 
1315 	return (zilog);
1316 }
1317 
1318 void
1319 zil_free(zilog_t *zilog)
1320 {
1321 	lwb_t *lwb;
1322 
1323 	zilog->zl_stop_sync = 1;
1324 
1325 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1326 		list_remove(&zilog->zl_lwb_list, lwb);
1327 		if (lwb->lwb_buf != NULL)
1328 			zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1329 		kmem_cache_free(zil_lwb_cache, lwb);
1330 	}
1331 	list_destroy(&zilog->zl_lwb_list);
1332 
1333 	avl_destroy(&zilog->zl_vdev_tree);
1334 	mutex_destroy(&zilog->zl_vdev_lock);
1335 
1336 	ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1337 	list_destroy(&zilog->zl_itx_list);
1338 	mutex_destroy(&zilog->zl_lock);
1339 
1340 	cv_destroy(&zilog->zl_cv_writer);
1341 	cv_destroy(&zilog->zl_cv_suspend);
1342 
1343 	kmem_free(zilog, sizeof (zilog_t));
1344 }
1345 
1346 /*
1347  * return true if the initial log block is not valid
1348  */
1349 static boolean_t
1350 zil_empty(zilog_t *zilog)
1351 {
1352 	const zil_header_t *zh = zilog->zl_header;
1353 	arc_buf_t *abuf = NULL;
1354 
1355 	if (BP_IS_HOLE(&zh->zh_log))
1356 		return (B_TRUE);
1357 
1358 	if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0)
1359 		return (B_TRUE);
1360 
1361 	VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1362 	return (B_FALSE);
1363 }
1364 
1365 /*
1366  * Open an intent log.
1367  */
1368 zilog_t *
1369 zil_open(objset_t *os, zil_get_data_t *get_data)
1370 {
1371 	zilog_t *zilog = dmu_objset_zil(os);
1372 
1373 	zilog->zl_get_data = get_data;
1374 	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1375 	    2, 2, TASKQ_PREPOPULATE);
1376 
1377 	return (zilog);
1378 }
1379 
1380 /*
1381  * Close an intent log.
1382  */
1383 void
1384 zil_close(zilog_t *zilog)
1385 {
1386 	/*
1387 	 * If the log isn't already committed, mark the objset dirty
1388 	 * (so zil_sync() will be called) and wait for that txg to sync.
1389 	 */
1390 	if (!zil_is_committed(zilog)) {
1391 		uint64_t txg;
1392 		dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1393 		(void) dmu_tx_assign(tx, TXG_WAIT);
1394 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1395 		txg = dmu_tx_get_txg(tx);
1396 		dmu_tx_commit(tx);
1397 		txg_wait_synced(zilog->zl_dmu_pool, txg);
1398 	}
1399 
1400 	taskq_destroy(zilog->zl_clean_taskq);
1401 	zilog->zl_clean_taskq = NULL;
1402 	zilog->zl_get_data = NULL;
1403 
1404 	zil_itx_clean(zilog);
1405 	ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1406 }
1407 
1408 /*
1409  * Suspend an intent log.  While in suspended mode, we still honor
1410  * synchronous semantics, but we rely on txg_wait_synced() to do it.
1411  * We suspend the log briefly when taking a snapshot so that the snapshot
1412  * contains all the data it's supposed to, and has an empty intent log.
1413  */
1414 int
1415 zil_suspend(zilog_t *zilog)
1416 {
1417 	const zil_header_t *zh = zilog->zl_header;
1418 
1419 	mutex_enter(&zilog->zl_lock);
1420 	if (zh->zh_claim_txg != 0) {		/* unplayed log */
1421 		mutex_exit(&zilog->zl_lock);
1422 		return (EBUSY);
1423 	}
1424 	if (zilog->zl_suspend++ != 0) {
1425 		/*
1426 		 * Someone else already began a suspend.
1427 		 * Just wait for them to finish.
1428 		 */
1429 		while (zilog->zl_suspending)
1430 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1431 		mutex_exit(&zilog->zl_lock);
1432 		return (0);
1433 	}
1434 	zilog->zl_suspending = B_TRUE;
1435 	mutex_exit(&zilog->zl_lock);
1436 
1437 	zil_commit(zilog, UINT64_MAX, 0);
1438 
1439 	/*
1440 	 * Wait for any in-flight log writes to complete.
1441 	 */
1442 	mutex_enter(&zilog->zl_lock);
1443 	while (zilog->zl_writer)
1444 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1445 	mutex_exit(&zilog->zl_lock);
1446 
1447 	zil_destroy(zilog, B_FALSE);
1448 
1449 	mutex_enter(&zilog->zl_lock);
1450 	zilog->zl_suspending = B_FALSE;
1451 	cv_broadcast(&zilog->zl_cv_suspend);
1452 	mutex_exit(&zilog->zl_lock);
1453 
1454 	return (0);
1455 }
1456 
1457 void
1458 zil_resume(zilog_t *zilog)
1459 {
1460 	mutex_enter(&zilog->zl_lock);
1461 	ASSERT(zilog->zl_suspend != 0);
1462 	zilog->zl_suspend--;
1463 	mutex_exit(&zilog->zl_lock);
1464 }
1465 
1466 typedef struct zil_replay_arg {
1467 	objset_t	*zr_os;
1468 	zil_replay_func_t **zr_replay;
1469 	void		*zr_arg;
1470 	boolean_t	zr_byteswap;
1471 	char		*zr_lrbuf;
1472 } zil_replay_arg_t;
1473 
1474 static void
1475 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1476 {
1477 	zil_replay_arg_t *zr = zra;
1478 	const zil_header_t *zh = zilog->zl_header;
1479 	uint64_t reclen = lr->lrc_reclen;
1480 	uint64_t txtype = lr->lrc_txtype;
1481 	char *name;
1482 	int pass, error;
1483 
1484 	if (!zilog->zl_replay)			/* giving up */
1485 		return;
1486 
1487 	if (lr->lrc_txg < claim_txg)		/* already committed */
1488 		return;
1489 
1490 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1491 		return;
1492 
1493 	/* Strip case-insensitive bit, still present in log record */
1494 	txtype &= ~TX_CI;
1495 
1496 	if (txtype == 0 || txtype >= TX_MAX_TYPE) {
1497 		error = EINVAL;
1498 		goto bad;
1499 	}
1500 
1501 	/*
1502 	 * Make a copy of the data so we can revise and extend it.
1503 	 */
1504 	bcopy(lr, zr->zr_lrbuf, reclen);
1505 
1506 	/*
1507 	 * The log block containing this lr may have been byteswapped
1508 	 * so that we can easily examine common fields like lrc_txtype.
1509 	 * However, the log is a mix of different data types, and only the
1510 	 * replay vectors know how to byteswap their records.  Therefore, if
1511 	 * the lr was byteswapped, undo it before invoking the replay vector.
1512 	 */
1513 	if (zr->zr_byteswap)
1514 		byteswap_uint64_array(zr->zr_lrbuf, reclen);
1515 
1516 	/*
1517 	 * If this is a TX_WRITE with a blkptr, suck in the data.
1518 	 */
1519 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1520 		lr_write_t *lrw = (lr_write_t *)lr;
1521 		blkptr_t *wbp = &lrw->lr_blkptr;
1522 		uint64_t wlen = lrw->lr_length;
1523 		char *wbuf = zr->zr_lrbuf + reclen;
1524 
1525 		if (BP_IS_HOLE(wbp)) {	/* compressed to a hole */
1526 			bzero(wbuf, wlen);
1527 		} else {
1528 			/*
1529 			 * A subsequent write may have overwritten this block,
1530 			 * in which case wbp may have been been freed and
1531 			 * reallocated, and our read of wbp may fail with a
1532 			 * checksum error.  We can safely ignore this because
1533 			 * the later write will provide the correct data.
1534 			 */
1535 			zbookmark_t zb;
1536 
1537 			zb.zb_objset = dmu_objset_id(zilog->zl_os);
1538 			zb.zb_object = lrw->lr_foid;
1539 			zb.zb_level = -1;
1540 			zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp);
1541 
1542 			(void) zio_wait(zio_read(NULL, zilog->zl_spa,
1543 			    wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL,
1544 			    ZIO_PRIORITY_SYNC_READ,
1545 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb));
1546 			(void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen);
1547 		}
1548 	}
1549 
1550 	/*
1551 	 * We must now do two things atomically: replay this log record,
1552 	 * and update the log header sequence number to reflect the fact that
1553 	 * we did so. At the end of each replay function the sequence number
1554 	 * is updated if we are in replay mode.
1555 	 */
1556 	for (pass = 1; pass <= 2; pass++) {
1557 		zilog->zl_replaying_seq = lr->lrc_seq;
1558 		/* Only byteswap (if needed) on the 1st pass.  */
1559 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf,
1560 		    zr->zr_byteswap && pass == 1);
1561 
1562 		if (!error)
1563 			return;
1564 
1565 		/*
1566 		 * The DMU's dnode layer doesn't see removes until the txg
1567 		 * commits, so a subsequent claim can spuriously fail with
1568 		 * EEXIST. So if we receive any error we try syncing out
1569 		 * any removes then retry the transaction.
1570 		 */
1571 		if (pass == 1)
1572 			txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1573 	}
1574 
1575 bad:
1576 	ASSERT(error);
1577 	name = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1578 	dmu_objset_name(zr->zr_os, name);
1579 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1580 	    "dataset %s, seq 0x%llx, txtype %llu %s\n",
1581 	    error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype,
1582 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1583 	zilog->zl_replay = B_FALSE;
1584 	kmem_free(name, MAXNAMELEN);
1585 }
1586 
1587 /* ARGSUSED */
1588 static void
1589 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1590 {
1591 	zilog->zl_replay_blks++;
1592 }
1593 
1594 /*
1595  * If this dataset has a non-empty intent log, replay it and destroy it.
1596  */
1597 void
1598 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1599 {
1600 	zilog_t *zilog = dmu_objset_zil(os);
1601 	const zil_header_t *zh = zilog->zl_header;
1602 	zil_replay_arg_t zr;
1603 
1604 	if (zil_empty(zilog)) {
1605 		zil_destroy(zilog, B_TRUE);
1606 		return;
1607 	}
1608 
1609 	zr.zr_os = os;
1610 	zr.zr_replay = replay_func;
1611 	zr.zr_arg = arg;
1612 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1613 	zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1614 
1615 	/*
1616 	 * Wait for in-progress removes to sync before starting replay.
1617 	 */
1618 	txg_wait_synced(zilog->zl_dmu_pool, 0);
1619 
1620 	zilog->zl_replay = B_TRUE;
1621 	zilog->zl_replay_time = lbolt;
1622 	ASSERT(zilog->zl_replay_blks == 0);
1623 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1624 	    zh->zh_claim_txg);
1625 	kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE);
1626 
1627 	zil_destroy(zilog, B_FALSE);
1628 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1629 	zilog->zl_replay = B_FALSE;
1630 }
1631 
1632 /*
1633  * Report whether all transactions are committed
1634  */
1635 int
1636 zil_is_committed(zilog_t *zilog)
1637 {
1638 	lwb_t *lwb;
1639 	int ret;
1640 
1641 	mutex_enter(&zilog->zl_lock);
1642 	while (zilog->zl_writer)
1643 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1644 
1645 	/* recent unpushed intent log transactions? */
1646 	if (!list_is_empty(&zilog->zl_itx_list)) {
1647 		ret = B_FALSE;
1648 		goto out;
1649 	}
1650 
1651 	/* intent log never used? */
1652 	lwb = list_head(&zilog->zl_lwb_list);
1653 	if (lwb == NULL) {
1654 		ret = B_TRUE;
1655 		goto out;
1656 	}
1657 
1658 	/*
1659 	 * more than 1 log buffer means zil_sync() hasn't yet freed
1660 	 * entries after a txg has committed
1661 	 */
1662 	if (list_next(&zilog->zl_lwb_list, lwb)) {
1663 		ret = B_FALSE;
1664 		goto out;
1665 	}
1666 
1667 	ASSERT(zil_empty(zilog));
1668 	ret = B_TRUE;
1669 out:
1670 	cv_broadcast(&zilog->zl_cv_writer);
1671 	mutex_exit(&zilog->zl_lock);
1672 	return (ret);
1673 }
1674