xref: /illumos-gate/usr/src/uts/common/fs/zfs/zil.c (revision 745cd3c5371d020efae7a911c58c526aa1fd0dba)
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 #pragma ident	"%Z%%M%	%I%	%E% SMI"
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.h>
39 #include <sys/dmu_tx.h>
40 
41 /*
42  * The zfs intent log (ZIL) saves transaction records of system calls
43  * that change the file system in memory with enough information
44  * to be able to replay them. These are stored in memory until
45  * either the DMU transaction group (txg) commits them to the stable pool
46  * and they can be discarded, or they are flushed to the stable log
47  * (also in the pool) due to a fsync, O_DSYNC or other synchronous
48  * requirement. In the event of a panic or power fail then those log
49  * records (transactions) are replayed.
50  *
51  * There is one ZIL per file system. Its on-disk (pool) format consists
52  * of 3 parts:
53  *
54  * 	- ZIL header
55  * 	- ZIL blocks
56  * 	- ZIL records
57  *
58  * A log record holds a system call transaction. Log blocks can
59  * hold many log records and the blocks are chained together.
60  * Each ZIL block contains a block pointer (blkptr_t) to the next
61  * ZIL block in the chain. The ZIL header points to the first
62  * block in the chain. Note there is not a fixed place in the pool
63  * to hold blocks. They are dynamically allocated and freed as
64  * needed from the blocks available. Figure X shows the ZIL structure:
65  */
66 
67 /*
68  * This global ZIL switch affects all pools
69  */
70 int zil_disable = 0;	/* disable intent logging */
71 
72 /*
73  * Tunable parameter for debugging or performance analysis.  Setting
74  * zfs_nocacheflush will cause corruption on power loss if a volatile
75  * out-of-order write cache is enabled.
76  */
77 boolean_t zfs_nocacheflush = B_FALSE;
78 
79 static kmem_cache_t *zil_lwb_cache;
80 
81 static int
82 zil_dva_compare(const void *x1, const void *x2)
83 {
84 	const dva_t *dva1 = x1;
85 	const dva_t *dva2 = x2;
86 
87 	if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
88 		return (-1);
89 	if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
90 		return (1);
91 
92 	if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
93 		return (-1);
94 	if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
95 		return (1);
96 
97 	return (0);
98 }
99 
100 static void
101 zil_dva_tree_init(avl_tree_t *t)
102 {
103 	avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t),
104 	    offsetof(zil_dva_node_t, zn_node));
105 }
106 
107 static void
108 zil_dva_tree_fini(avl_tree_t *t)
109 {
110 	zil_dva_node_t *zn;
111 	void *cookie = NULL;
112 
113 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
114 		kmem_free(zn, sizeof (zil_dva_node_t));
115 
116 	avl_destroy(t);
117 }
118 
119 static int
120 zil_dva_tree_add(avl_tree_t *t, dva_t *dva)
121 {
122 	zil_dva_node_t *zn;
123 	avl_index_t where;
124 
125 	if (avl_find(t, dva, &where) != NULL)
126 		return (EEXIST);
127 
128 	zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP);
129 	zn->zn_dva = *dva;
130 	avl_insert(t, zn, where);
131 
132 	return (0);
133 }
134 
135 static zil_header_t *
136 zil_header_in_syncing_context(zilog_t *zilog)
137 {
138 	return ((zil_header_t *)zilog->zl_header);
139 }
140 
141 static void
142 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
143 {
144 	zio_cksum_t *zc = &bp->blk_cksum;
145 
146 	zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
147 	zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
148 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
149 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
150 }
151 
152 /*
153  * Read a log block, make sure it's valid, and byteswap it if necessary.
154  */
155 static int
156 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp)
157 {
158 	blkptr_t blk = *bp;
159 	zbookmark_t zb;
160 	uint32_t aflags = ARC_WAIT;
161 	int error;
162 
163 	zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET];
164 	zb.zb_object = 0;
165 	zb.zb_level = -1;
166 	zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
167 
168 	*abufpp = NULL;
169 
170 	error = arc_read(NULL, zilog->zl_spa, &blk, byteswap_uint64_array,
171 	    arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL |
172 	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb);
173 
174 	if (error == 0) {
175 		char *data = (*abufpp)->b_data;
176 		uint64_t blksz = BP_GET_LSIZE(bp);
177 		zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1;
178 		zio_cksum_t cksum = bp->blk_cksum;
179 
180 		/*
181 		 * Sequence numbers should be... sequential.  The checksum
182 		 * verifier for the next block should be bp's checksum plus 1.
183 		 */
184 		cksum.zc_word[ZIL_ZC_SEQ]++;
185 
186 		if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, sizeof (cksum)))
187 			error = ESTALE;
188 		else if (BP_IS_HOLE(&ztp->zit_next_blk))
189 			error = ENOENT;
190 		else if (ztp->zit_nused > (blksz - sizeof (zil_trailer_t)))
191 			error = EOVERFLOW;
192 
193 		if (error) {
194 			VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1);
195 			*abufpp = NULL;
196 		}
197 	}
198 
199 	dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid);
200 
201 	return (error);
202 }
203 
204 /*
205  * Parse the intent log, and call parse_func for each valid record within.
206  * Return the highest sequence number.
207  */
208 uint64_t
209 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
210     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
211 {
212 	const zil_header_t *zh = zilog->zl_header;
213 	uint64_t claim_seq = zh->zh_claim_seq;
214 	uint64_t seq = 0;
215 	uint64_t max_seq = 0;
216 	blkptr_t blk = zh->zh_log;
217 	arc_buf_t *abuf;
218 	char *lrbuf, *lrp;
219 	zil_trailer_t *ztp;
220 	int reclen, error;
221 
222 	if (BP_IS_HOLE(&blk))
223 		return (max_seq);
224 
225 	/*
226 	 * Starting at the block pointed to by zh_log we read the log chain.
227 	 * For each block in the chain we strongly check that block to
228 	 * ensure its validity.  We stop when an invalid block is found.
229 	 * For each block pointer in the chain we call parse_blk_func().
230 	 * For each record in each valid block we call parse_lr_func().
231 	 * If the log has been claimed, stop if we encounter a sequence
232 	 * number greater than the highest claimed sequence number.
233 	 */
234 	zil_dva_tree_init(&zilog->zl_dva_tree);
235 	for (;;) {
236 		seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
237 
238 		if (claim_seq != 0 && seq > claim_seq)
239 			break;
240 
241 		ASSERT(max_seq < seq);
242 		max_seq = seq;
243 
244 		error = zil_read_log_block(zilog, &blk, &abuf);
245 
246 		if (parse_blk_func != NULL)
247 			parse_blk_func(zilog, &blk, arg, txg);
248 
249 		if (error)
250 			break;
251 
252 		lrbuf = abuf->b_data;
253 		ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1;
254 		blk = ztp->zit_next_blk;
255 
256 		if (parse_lr_func == NULL) {
257 			VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
258 			continue;
259 		}
260 
261 		for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) {
262 			lr_t *lr = (lr_t *)lrp;
263 			reclen = lr->lrc_reclen;
264 			ASSERT3U(reclen, >=, sizeof (lr_t));
265 			parse_lr_func(zilog, lr, arg, txg);
266 		}
267 		VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
268 	}
269 	zil_dva_tree_fini(&zilog->zl_dva_tree);
270 
271 	return (max_seq);
272 }
273 
274 /* ARGSUSED */
275 static void
276 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
277 {
278 	spa_t *spa = zilog->zl_spa;
279 	int err;
280 
281 	/*
282 	 * Claim log block if not already committed and not already claimed.
283 	 */
284 	if (bp->blk_birth >= first_txg &&
285 	    zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) {
286 		err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL));
287 		ASSERT(err == 0);
288 	}
289 }
290 
291 static void
292 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
293 {
294 	if (lrc->lrc_txtype == TX_WRITE) {
295 		lr_write_t *lr = (lr_write_t *)lrc;
296 		zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg);
297 	}
298 }
299 
300 /* ARGSUSED */
301 static void
302 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
303 {
304 	zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx));
305 }
306 
307 static void
308 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
309 {
310 	/*
311 	 * If we previously claimed it, we need to free it.
312 	 */
313 	if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) {
314 		lr_write_t *lr = (lr_write_t *)lrc;
315 		blkptr_t *bp = &lr->lr_blkptr;
316 		if (bp->blk_birth >= claim_txg &&
317 		    !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) {
318 			(void) arc_free(NULL, zilog->zl_spa,
319 			    dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT);
320 		}
321 	}
322 }
323 
324 /*
325  * Create an on-disk intent log.
326  */
327 static void
328 zil_create(zilog_t *zilog)
329 {
330 	const zil_header_t *zh = zilog->zl_header;
331 	lwb_t *lwb;
332 	uint64_t txg = 0;
333 	dmu_tx_t *tx = NULL;
334 	blkptr_t blk;
335 	int error = 0;
336 
337 	/*
338 	 * Wait for any previous destroy to complete.
339 	 */
340 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
341 
342 	ASSERT(zh->zh_claim_txg == 0);
343 	ASSERT(zh->zh_replay_seq == 0);
344 
345 	blk = zh->zh_log;
346 
347 	/*
348 	 * If we don't already have an initial log block, allocate one now.
349 	 */
350 	if (BP_IS_HOLE(&blk)) {
351 		tx = dmu_tx_create(zilog->zl_os);
352 		(void) dmu_tx_assign(tx, TXG_WAIT);
353 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
354 		txg = dmu_tx_get_txg(tx);
355 
356 		error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk,
357 		    NULL, txg);
358 
359 		if (error == 0)
360 			zil_init_log_chain(zilog, &blk);
361 	}
362 
363 	/*
364 	 * Allocate a log write buffer (lwb) for the first log block.
365 	 */
366 	if (error == 0) {
367 		lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
368 		lwb->lwb_zilog = zilog;
369 		lwb->lwb_blk = blk;
370 		lwb->lwb_nused = 0;
371 		lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk);
372 		lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz);
373 		lwb->lwb_max_txg = txg;
374 		lwb->lwb_zio = NULL;
375 
376 		mutex_enter(&zilog->zl_lock);
377 		list_insert_tail(&zilog->zl_lwb_list, lwb);
378 		mutex_exit(&zilog->zl_lock);
379 	}
380 
381 	/*
382 	 * If we just allocated the first log block, commit our transaction
383 	 * and wait for zil_sync() to stuff the block poiner into zh_log.
384 	 * (zh is part of the MOS, so we cannot modify it in open context.)
385 	 */
386 	if (tx != NULL) {
387 		dmu_tx_commit(tx);
388 		txg_wait_synced(zilog->zl_dmu_pool, txg);
389 	}
390 
391 	ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
392 }
393 
394 /*
395  * In one tx, free all log blocks and clear the log header.
396  * If keep_first is set, then we're replaying a log with no content.
397  * We want to keep the first block, however, so that the first
398  * synchronous transaction doesn't require a txg_wait_synced()
399  * in zil_create().  We don't need to txg_wait_synced() here either
400  * when keep_first is set, because both zil_create() and zil_destroy()
401  * will wait for any in-progress destroys to complete.
402  */
403 void
404 zil_destroy(zilog_t *zilog, boolean_t keep_first)
405 {
406 	const zil_header_t *zh = zilog->zl_header;
407 	lwb_t *lwb;
408 	dmu_tx_t *tx;
409 	uint64_t txg;
410 
411 	/*
412 	 * Wait for any previous destroy to complete.
413 	 */
414 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
415 
416 	if (BP_IS_HOLE(&zh->zh_log))
417 		return;
418 
419 	tx = dmu_tx_create(zilog->zl_os);
420 	(void) dmu_tx_assign(tx, TXG_WAIT);
421 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
422 	txg = dmu_tx_get_txg(tx);
423 
424 	mutex_enter(&zilog->zl_lock);
425 
426 	/*
427 	 * It is possible for the ZIL to get the previously mounted zilog
428 	 * structure of the same dataset if quickly remounted and the dbuf
429 	 * eviction has not completed. In this case we can see a non
430 	 * empty lwb list and keep_first will be set. We fix this by
431 	 * clearing the keep_first. This will be slower but it's very rare.
432 	 */
433 	if (!list_is_empty(&zilog->zl_lwb_list) && keep_first)
434 		keep_first = B_FALSE;
435 
436 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
437 	zilog->zl_destroy_txg = txg;
438 	zilog->zl_keep_first = keep_first;
439 
440 	if (!list_is_empty(&zilog->zl_lwb_list)) {
441 		ASSERT(zh->zh_claim_txg == 0);
442 		ASSERT(!keep_first);
443 		while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
444 			list_remove(&zilog->zl_lwb_list, lwb);
445 			if (lwb->lwb_buf != NULL)
446 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
447 			zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg);
448 			kmem_cache_free(zil_lwb_cache, lwb);
449 		}
450 	} else {
451 		if (!keep_first) {
452 			(void) zil_parse(zilog, zil_free_log_block,
453 			    zil_free_log_record, tx, zh->zh_claim_txg);
454 		}
455 	}
456 	mutex_exit(&zilog->zl_lock);
457 
458 	dmu_tx_commit(tx);
459 }
460 
461 /*
462  * zil_rollback_destroy() is only called by the rollback code.
463  * We already have a syncing tx. Rollback has exclusive access to the
464  * dataset, so we don't have to worry about concurrent zil access.
465  * The actual freeing of any log blocks occurs in zil_sync() later in
466  * this txg syncing phase.
467  */
468 void
469 zil_rollback_destroy(zilog_t *zilog, dmu_tx_t *tx)
470 {
471 	const zil_header_t *zh = zilog->zl_header;
472 	uint64_t txg;
473 
474 	if (BP_IS_HOLE(&zh->zh_log))
475 		return;
476 
477 	txg = dmu_tx_get_txg(tx);
478 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
479 	zilog->zl_destroy_txg = txg;
480 	zilog->zl_keep_first = B_FALSE;
481 
482 	/*
483 	 * Ensure there's no outstanding ZIL IO.  No lwbs or just the
484 	 * unused one that allocated in advance is ok.
485 	 */
486 	ASSERT(zilog->zl_lwb_list.list_head.list_next ==
487 	    zilog->zl_lwb_list.list_head.list_prev);
488 	(void) zil_parse(zilog, zil_free_log_block, zil_free_log_record,
489 	    tx, zh->zh_claim_txg);
490 }
491 
492 int
493 zil_claim(char *osname, void *txarg)
494 {
495 	dmu_tx_t *tx = txarg;
496 	uint64_t first_txg = dmu_tx_get_txg(tx);
497 	zilog_t *zilog;
498 	zil_header_t *zh;
499 	objset_t *os;
500 	int error;
501 
502 	error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
503 	if (error) {
504 		cmn_err(CE_WARN, "can't process intent log for %s", osname);
505 		return (0);
506 	}
507 
508 	zilog = dmu_objset_zil(os);
509 	zh = zil_header_in_syncing_context(zilog);
510 
511 	/*
512 	 * Claim all log blocks if we haven't already done so, and remember
513 	 * the highest claimed sequence number.  This ensures that if we can
514 	 * read only part of the log now (e.g. due to a missing device),
515 	 * but we can read the entire log later, we will not try to replay
516 	 * or destroy beyond the last block we successfully claimed.
517 	 */
518 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
519 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
520 		zh->zh_claim_txg = first_txg;
521 		zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block,
522 		    zil_claim_log_record, tx, first_txg);
523 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
524 	}
525 
526 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
527 	dmu_objset_close(os);
528 	return (0);
529 }
530 
531 static int
532 zil_vdev_compare(const void *x1, const void *x2)
533 {
534 	uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
535 	uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
536 
537 	if (v1 < v2)
538 		return (-1);
539 	if (v1 > v2)
540 		return (1);
541 
542 	return (0);
543 }
544 
545 void
546 zil_add_block(zilog_t *zilog, blkptr_t *bp)
547 {
548 	avl_tree_t *t = &zilog->zl_vdev_tree;
549 	avl_index_t where;
550 	zil_vdev_node_t *zv, zvsearch;
551 	int ndvas = BP_GET_NDVAS(bp);
552 	int i;
553 
554 	if (zfs_nocacheflush)
555 		return;
556 
557 	ASSERT(zilog->zl_writer);
558 
559 	/*
560 	 * Even though we're zl_writer, we still need a lock because the
561 	 * zl_get_data() callbacks may have dmu_sync() done callbacks
562 	 * that will run concurrently.
563 	 */
564 	mutex_enter(&zilog->zl_vdev_lock);
565 	for (i = 0; i < ndvas; i++) {
566 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
567 		if (avl_find(t, &zvsearch, &where) == NULL) {
568 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
569 			zv->zv_vdev = zvsearch.zv_vdev;
570 			avl_insert(t, zv, where);
571 		}
572 	}
573 	mutex_exit(&zilog->zl_vdev_lock);
574 }
575 
576 void
577 zil_flush_vdevs(zilog_t *zilog)
578 {
579 	spa_t *spa = zilog->zl_spa;
580 	avl_tree_t *t = &zilog->zl_vdev_tree;
581 	void *cookie = NULL;
582 	zil_vdev_node_t *zv;
583 	zio_t *zio;
584 
585 	ASSERT(zilog->zl_writer);
586 
587 	/*
588 	 * We don't need zl_vdev_lock here because we're the zl_writer,
589 	 * and all zl_get_data() callbacks are done.
590 	 */
591 	if (avl_numnodes(t) == 0)
592 		return;
593 
594 	spa_config_enter(spa, RW_READER, FTAG);
595 
596 	zio = zio_root(spa, NULL, NULL,
597 	    ZIO_FLAG_CONFIG_HELD | ZIO_FLAG_CANFAIL);
598 
599 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
600 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
601 		if (vd != NULL)
602 			zio_flush(zio, vd);
603 		kmem_free(zv, sizeof (*zv));
604 	}
605 
606 	/*
607 	 * Wait for all the flushes to complete.  Not all devices actually
608 	 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
609 	 */
610 	(void) zio_wait(zio);
611 
612 	spa_config_exit(spa, FTAG);
613 }
614 
615 /*
616  * Function called when a log block write completes
617  */
618 static void
619 zil_lwb_write_done(zio_t *zio)
620 {
621 	lwb_t *lwb = zio->io_private;
622 	zilog_t *zilog = lwb->lwb_zilog;
623 
624 	/*
625 	 * Now that we've written this log block, we have a stable pointer
626 	 * to the next block in the chain, so it's OK to let the txg in
627 	 * which we allocated the next block sync.
628 	 */
629 	txg_rele_to_sync(&lwb->lwb_txgh);
630 
631 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
632 	mutex_enter(&zilog->zl_lock);
633 	lwb->lwb_buf = NULL;
634 	if (zio->io_error)
635 		zilog->zl_log_error = B_TRUE;
636 	mutex_exit(&zilog->zl_lock);
637 }
638 
639 /*
640  * Initialize the io for a log block.
641  *
642  * Note, we should not initialize the IO until we are about
643  * to use it, since zio_rewrite() does a spa_config_enter().
644  */
645 static void
646 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
647 {
648 	zbookmark_t zb;
649 
650 	zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET];
651 	zb.zb_object = 0;
652 	zb.zb_level = -1;
653 	zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
654 
655 	if (zilog->zl_root_zio == NULL) {
656 		zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
657 		    ZIO_FLAG_CANFAIL);
658 	}
659 	if (lwb->lwb_zio == NULL) {
660 		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
661 		    ZIO_CHECKSUM_ZILOG, 0, &lwb->lwb_blk, lwb->lwb_buf,
662 		    lwb->lwb_sz, zil_lwb_write_done, lwb,
663 		    ZIO_PRIORITY_LOG_WRITE, ZIO_FLAG_CANFAIL, &zb);
664 	}
665 }
666 
667 /*
668  * Start a log block write and advance to the next log block.
669  * Calls are serialized.
670  */
671 static lwb_t *
672 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
673 {
674 	lwb_t *nlwb;
675 	zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1;
676 	spa_t *spa = zilog->zl_spa;
677 	blkptr_t *bp = &ztp->zit_next_blk;
678 	uint64_t txg;
679 	uint64_t zil_blksz;
680 	int error;
681 
682 	ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb));
683 
684 	/*
685 	 * Allocate the next block and save its address in this block
686 	 * before writing it in order to establish the log chain.
687 	 * Note that if the allocation of nlwb synced before we wrote
688 	 * the block that points at it (lwb), we'd leak it if we crashed.
689 	 * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done().
690 	 */
691 	txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh);
692 	txg_rele_to_quiesce(&lwb->lwb_txgh);
693 
694 	/*
695 	 * Pick a ZIL blocksize. We request a size that is the
696 	 * maximum of the previous used size, the current used size and
697 	 * the amount waiting in the queue.
698 	 */
699 	zil_blksz = MAX(zilog->zl_prev_used,
700 	    zilog->zl_cur_used + sizeof (*ztp));
701 	zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp));
702 	zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t);
703 	if (zil_blksz > ZIL_MAX_BLKSZ)
704 		zil_blksz = ZIL_MAX_BLKSZ;
705 
706 	BP_ZERO(bp);
707 	/* pass the old blkptr in order to spread log blocks across devs */
708 	error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg);
709 	if (error) {
710 		dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg);
711 
712 		/*
713 		 * We dirty the dataset to ensure that zil_sync() will
714 		 * be called to remove this lwb from our zl_lwb_list.
715 		 * Failing to do so, may leave an lwb with a NULL lwb_buf
716 		 * hanging around on the zl_lwb_list.
717 		 */
718 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
719 		dmu_tx_commit(tx);
720 
721 		/*
722 		 * Since we've just experienced an allocation failure so we
723 		 * terminate the current lwb and send it on its way.
724 		 */
725 		ztp->zit_pad = 0;
726 		ztp->zit_nused = lwb->lwb_nused;
727 		ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
728 		zio_nowait(lwb->lwb_zio);
729 
730 		/*
731 		 * By returning NULL the caller will call tx_wait_synced()
732 		 */
733 		return (NULL);
734 	}
735 
736 	ASSERT3U(bp->blk_birth, ==, txg);
737 	ztp->zit_pad = 0;
738 	ztp->zit_nused = lwb->lwb_nused;
739 	ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
740 	bp->blk_cksum = lwb->lwb_blk.blk_cksum;
741 	bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
742 
743 	/*
744 	 * Allocate a new log write buffer (lwb).
745 	 */
746 	nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
747 
748 	nlwb->lwb_zilog = zilog;
749 	nlwb->lwb_blk = *bp;
750 	nlwb->lwb_nused = 0;
751 	nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk);
752 	nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz);
753 	nlwb->lwb_max_txg = txg;
754 	nlwb->lwb_zio = NULL;
755 
756 	/*
757 	 * Put new lwb at the end of the log chain
758 	 */
759 	mutex_enter(&zilog->zl_lock);
760 	list_insert_tail(&zilog->zl_lwb_list, nlwb);
761 	mutex_exit(&zilog->zl_lock);
762 
763 	/* Record the block for later vdev flushing */
764 	zil_add_block(zilog, &lwb->lwb_blk);
765 
766 	/*
767 	 * kick off the write for the old log block
768 	 */
769 	dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg);
770 	ASSERT(lwb->lwb_zio);
771 	zio_nowait(lwb->lwb_zio);
772 
773 	return (nlwb);
774 }
775 
776 static lwb_t *
777 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
778 {
779 	lr_t *lrc = &itx->itx_lr; /* common log record */
780 	lr_write_t *lr = (lr_write_t *)lrc;
781 	uint64_t txg = lrc->lrc_txg;
782 	uint64_t reclen = lrc->lrc_reclen;
783 	uint64_t dlen;
784 
785 	if (lwb == NULL)
786 		return (NULL);
787 	ASSERT(lwb->lwb_buf != NULL);
788 
789 	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
790 		dlen = P2ROUNDUP_TYPED(
791 		    lr->lr_length, sizeof (uint64_t), uint64_t);
792 	else
793 		dlen = 0;
794 
795 	zilog->zl_cur_used += (reclen + dlen);
796 
797 	zil_lwb_write_init(zilog, lwb);
798 
799 	/*
800 	 * If this record won't fit in the current log block, start a new one.
801 	 */
802 	if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
803 		lwb = zil_lwb_write_start(zilog, lwb);
804 		if (lwb == NULL)
805 			return (NULL);
806 		zil_lwb_write_init(zilog, lwb);
807 		ASSERT(lwb->lwb_nused == 0);
808 		if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
809 			txg_wait_synced(zilog->zl_dmu_pool, txg);
810 			return (lwb);
811 		}
812 	}
813 
814 	/*
815 	 * Update the lrc_seq, to be log record sequence number. See zil.h
816 	 * Then copy the record to the log buffer.
817 	 */
818 	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
819 	bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen);
820 
821 	/*
822 	 * If it's a write, fetch the data or get its blkptr as appropriate.
823 	 */
824 	if (lrc->lrc_txtype == TX_WRITE) {
825 		if (txg > spa_freeze_txg(zilog->zl_spa))
826 			txg_wait_synced(zilog->zl_dmu_pool, txg);
827 		if (itx->itx_wr_state != WR_COPIED) {
828 			char *dbuf;
829 			int error;
830 
831 			/* alignment is guaranteed */
832 			lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused);
833 			if (dlen) {
834 				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
835 				dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen;
836 				lr->lr_common.lrc_reclen += dlen;
837 			} else {
838 				ASSERT(itx->itx_wr_state == WR_INDIRECT);
839 				dbuf = NULL;
840 			}
841 			error = zilog->zl_get_data(
842 			    itx->itx_private, lr, dbuf, lwb->lwb_zio);
843 			if (error) {
844 				ASSERT(error == ENOENT || error == EEXIST ||
845 				    error == EALREADY);
846 				return (lwb);
847 			}
848 		}
849 	}
850 
851 	lwb->lwb_nused += reclen + dlen;
852 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
853 	ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb));
854 	ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
855 
856 	return (lwb);
857 }
858 
859 itx_t *
860 zil_itx_create(uint64_t txtype, size_t lrsize)
861 {
862 	itx_t *itx;
863 
864 	lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
865 
866 	itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
867 	itx->itx_lr.lrc_txtype = txtype;
868 	itx->itx_lr.lrc_reclen = lrsize;
869 	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
870 	itx->itx_lr.lrc_seq = 0;	/* defensive */
871 
872 	return (itx);
873 }
874 
875 uint64_t
876 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
877 {
878 	uint64_t seq;
879 
880 	ASSERT(itx->itx_lr.lrc_seq == 0);
881 
882 	mutex_enter(&zilog->zl_lock);
883 	list_insert_tail(&zilog->zl_itx_list, itx);
884 	zilog->zl_itx_list_sz += itx->itx_sod;
885 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
886 	itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq;
887 	mutex_exit(&zilog->zl_lock);
888 
889 	return (seq);
890 }
891 
892 /*
893  * Free up all in-memory intent log transactions that have now been synced.
894  */
895 static void
896 zil_itx_clean(zilog_t *zilog)
897 {
898 	uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa);
899 	uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa);
900 	list_t clean_list;
901 	itx_t *itx;
902 
903 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
904 
905 	mutex_enter(&zilog->zl_lock);
906 	/* wait for a log writer to finish walking list */
907 	while (zilog->zl_writer) {
908 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
909 	}
910 
911 	/*
912 	 * Move the sync'd log transactions to a separate list so we can call
913 	 * kmem_free without holding the zl_lock.
914 	 *
915 	 * There is no need to set zl_writer as we don't drop zl_lock here
916 	 */
917 	while ((itx = list_head(&zilog->zl_itx_list)) != NULL &&
918 	    itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) {
919 		list_remove(&zilog->zl_itx_list, itx);
920 		zilog->zl_itx_list_sz -= itx->itx_sod;
921 		list_insert_tail(&clean_list, itx);
922 	}
923 	cv_broadcast(&zilog->zl_cv_writer);
924 	mutex_exit(&zilog->zl_lock);
925 
926 	/* destroy sync'd log transactions */
927 	while ((itx = list_head(&clean_list)) != NULL) {
928 		list_remove(&clean_list, itx);
929 		kmem_free(itx, offsetof(itx_t, itx_lr)
930 		    + itx->itx_lr.lrc_reclen);
931 	}
932 	list_destroy(&clean_list);
933 }
934 
935 /*
936  * If there are any in-memory intent log transactions which have now been
937  * synced then start up a taskq to free them.
938  */
939 void
940 zil_clean(zilog_t *zilog)
941 {
942 	itx_t *itx;
943 
944 	mutex_enter(&zilog->zl_lock);
945 	itx = list_head(&zilog->zl_itx_list);
946 	if ((itx != NULL) &&
947 	    (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) {
948 		(void) taskq_dispatch(zilog->zl_clean_taskq,
949 		    (void (*)(void *))zil_itx_clean, zilog, TQ_NOSLEEP);
950 	}
951 	mutex_exit(&zilog->zl_lock);
952 }
953 
954 void
955 zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid)
956 {
957 	uint64_t txg;
958 	uint64_t commit_seq = 0;
959 	itx_t *itx, *itx_next = (itx_t *)-1;
960 	lwb_t *lwb;
961 	spa_t *spa;
962 
963 	zilog->zl_writer = B_TRUE;
964 	zilog->zl_root_zio = NULL;
965 	spa = zilog->zl_spa;
966 
967 	if (zilog->zl_suspend) {
968 		lwb = NULL;
969 	} else {
970 		lwb = list_tail(&zilog->zl_lwb_list);
971 		if (lwb == NULL) {
972 			/*
973 			 * Return if there's nothing to flush before we
974 			 * dirty the fs by calling zil_create()
975 			 */
976 			if (list_is_empty(&zilog->zl_itx_list)) {
977 				zilog->zl_writer = B_FALSE;
978 				return;
979 			}
980 			mutex_exit(&zilog->zl_lock);
981 			zil_create(zilog);
982 			mutex_enter(&zilog->zl_lock);
983 			lwb = list_tail(&zilog->zl_lwb_list);
984 		}
985 	}
986 
987 	/* Loop through in-memory log transactions filling log blocks. */
988 	DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
989 	for (;;) {
990 		/*
991 		 * Find the next itx to push:
992 		 * Push all transactions related to specified foid and all
993 		 * other transactions except TX_WRITE, TX_TRUNCATE,
994 		 * TX_SETATTR and TX_ACL for all other files.
995 		 */
996 		if (itx_next != (itx_t *)-1)
997 			itx = itx_next;
998 		else
999 			itx = list_head(&zilog->zl_itx_list);
1000 		for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) {
1001 			if (foid == 0) /* push all foids? */
1002 				break;
1003 			if (itx->itx_sync) /* push all O_[D]SYNC */
1004 				break;
1005 			switch (itx->itx_lr.lrc_txtype) {
1006 			case TX_SETATTR:
1007 			case TX_WRITE:
1008 			case TX_TRUNCATE:
1009 			case TX_ACL:
1010 				/* lr_foid is same offset for these records */
1011 				if (((lr_write_t *)&itx->itx_lr)->lr_foid
1012 				    != foid) {
1013 					continue; /* skip this record */
1014 				}
1015 			}
1016 			break;
1017 		}
1018 		if (itx == NULL)
1019 			break;
1020 
1021 		if ((itx->itx_lr.lrc_seq > seq) &&
1022 		    ((lwb == NULL) || (lwb->lwb_nused == 0) ||
1023 		    (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) {
1024 			break;
1025 		}
1026 
1027 		/*
1028 		 * Save the next pointer.  Even though we soon drop
1029 		 * zl_lock all threads that may change the list
1030 		 * (another writer or zil_itx_clean) can't do so until
1031 		 * they have zl_writer.
1032 		 */
1033 		itx_next = list_next(&zilog->zl_itx_list, itx);
1034 		list_remove(&zilog->zl_itx_list, itx);
1035 		zilog->zl_itx_list_sz -= itx->itx_sod;
1036 		mutex_exit(&zilog->zl_lock);
1037 		txg = itx->itx_lr.lrc_txg;
1038 		ASSERT(txg);
1039 
1040 		if (txg > spa_last_synced_txg(spa) ||
1041 		    txg > spa_freeze_txg(spa))
1042 			lwb = zil_lwb_commit(zilog, itx, lwb);
1043 		kmem_free(itx, offsetof(itx_t, itx_lr)
1044 		    + itx->itx_lr.lrc_reclen);
1045 		mutex_enter(&zilog->zl_lock);
1046 	}
1047 	DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1048 	/* determine commit sequence number */
1049 	itx = list_head(&zilog->zl_itx_list);
1050 	if (itx)
1051 		commit_seq = itx->itx_lr.lrc_seq;
1052 	else
1053 		commit_seq = zilog->zl_itx_seq;
1054 	mutex_exit(&zilog->zl_lock);
1055 
1056 	/* write the last block out */
1057 	if (lwb != NULL && lwb->lwb_zio != NULL)
1058 		lwb = zil_lwb_write_start(zilog, lwb);
1059 
1060 	zilog->zl_prev_used = zilog->zl_cur_used;
1061 	zilog->zl_cur_used = 0;
1062 
1063 	/*
1064 	 * Wait if necessary for the log blocks to be on stable storage.
1065 	 */
1066 	if (zilog->zl_root_zio) {
1067 		DTRACE_PROBE1(zil__cw3, zilog_t *, zilog);
1068 		(void) zio_wait(zilog->zl_root_zio);
1069 		DTRACE_PROBE1(zil__cw4, zilog_t *, zilog);
1070 		zil_flush_vdevs(zilog);
1071 	}
1072 
1073 	if (zilog->zl_log_error || lwb == NULL) {
1074 		zilog->zl_log_error = 0;
1075 		txg_wait_synced(zilog->zl_dmu_pool, 0);
1076 	}
1077 
1078 	mutex_enter(&zilog->zl_lock);
1079 	zilog->zl_writer = B_FALSE;
1080 
1081 	ASSERT3U(commit_seq, >=, zilog->zl_commit_seq);
1082 	zilog->zl_commit_seq = commit_seq;
1083 }
1084 
1085 /*
1086  * Push zfs transactions to stable storage up to the supplied sequence number.
1087  * If foid is 0 push out all transactions, otherwise push only those
1088  * for that file or might have been used to create that file.
1089  */
1090 void
1091 zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid)
1092 {
1093 	if (zilog == NULL || seq == 0)
1094 		return;
1095 
1096 	mutex_enter(&zilog->zl_lock);
1097 
1098 	seq = MIN(seq, zilog->zl_itx_seq);	/* cap seq at largest itx seq */
1099 
1100 	while (zilog->zl_writer) {
1101 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1102 		if (seq < zilog->zl_commit_seq) {
1103 			mutex_exit(&zilog->zl_lock);
1104 			return;
1105 		}
1106 	}
1107 	zil_commit_writer(zilog, seq, foid); /* drops zl_lock */
1108 	/* wake up others waiting on the commit */
1109 	cv_broadcast(&zilog->zl_cv_writer);
1110 	mutex_exit(&zilog->zl_lock);
1111 }
1112 
1113 /*
1114  * Called in syncing context to free committed log blocks and update log header.
1115  */
1116 void
1117 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1118 {
1119 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
1120 	uint64_t txg = dmu_tx_get_txg(tx);
1121 	spa_t *spa = zilog->zl_spa;
1122 	lwb_t *lwb;
1123 
1124 	mutex_enter(&zilog->zl_lock);
1125 
1126 	ASSERT(zilog->zl_stop_sync == 0);
1127 
1128 	zh->zh_replay_seq = zilog->zl_replay_seq[txg & TXG_MASK];
1129 
1130 	if (zilog->zl_destroy_txg == txg) {
1131 		blkptr_t blk = zh->zh_log;
1132 
1133 		ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1134 		ASSERT(spa_sync_pass(spa) == 1);
1135 
1136 		bzero(zh, sizeof (zil_header_t));
1137 		bzero(zilog->zl_replay_seq, sizeof (zilog->zl_replay_seq));
1138 
1139 		if (zilog->zl_keep_first) {
1140 			/*
1141 			 * If this block was part of log chain that couldn't
1142 			 * be claimed because a device was missing during
1143 			 * zil_claim(), but that device later returns,
1144 			 * then this block could erroneously appear valid.
1145 			 * To guard against this, assign a new GUID to the new
1146 			 * log chain so it doesn't matter what blk points to.
1147 			 */
1148 			zil_init_log_chain(zilog, &blk);
1149 			zh->zh_log = blk;
1150 		}
1151 	}
1152 
1153 	for (;;) {
1154 		lwb = list_head(&zilog->zl_lwb_list);
1155 		if (lwb == NULL) {
1156 			mutex_exit(&zilog->zl_lock);
1157 			return;
1158 		}
1159 		zh->zh_log = lwb->lwb_blk;
1160 		if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1161 			break;
1162 		list_remove(&zilog->zl_lwb_list, lwb);
1163 		zio_free_blk(spa, &lwb->lwb_blk, txg);
1164 		kmem_cache_free(zil_lwb_cache, lwb);
1165 
1166 		/*
1167 		 * If we don't have anything left in the lwb list then
1168 		 * we've had an allocation failure and we need to zero
1169 		 * out the zil_header blkptr so that we don't end
1170 		 * up freeing the same block twice.
1171 		 */
1172 		if (list_head(&zilog->zl_lwb_list) == NULL)
1173 			BP_ZERO(&zh->zh_log);
1174 	}
1175 	mutex_exit(&zilog->zl_lock);
1176 }
1177 
1178 void
1179 zil_init(void)
1180 {
1181 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1182 	    sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1183 }
1184 
1185 void
1186 zil_fini(void)
1187 {
1188 	kmem_cache_destroy(zil_lwb_cache);
1189 }
1190 
1191 zilog_t *
1192 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1193 {
1194 	zilog_t *zilog;
1195 
1196 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1197 
1198 	zilog->zl_header = zh_phys;
1199 	zilog->zl_os = os;
1200 	zilog->zl_spa = dmu_objset_spa(os);
1201 	zilog->zl_dmu_pool = dmu_objset_pool(os);
1202 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
1203 
1204 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1205 
1206 	list_create(&zilog->zl_itx_list, sizeof (itx_t),
1207 	    offsetof(itx_t, itx_node));
1208 
1209 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1210 	    offsetof(lwb_t, lwb_node));
1211 
1212 	mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1213 
1214 	avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1215 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1216 
1217 	cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1218 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1219 
1220 	return (zilog);
1221 }
1222 
1223 void
1224 zil_free(zilog_t *zilog)
1225 {
1226 	lwb_t *lwb;
1227 
1228 	zilog->zl_stop_sync = 1;
1229 
1230 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1231 		list_remove(&zilog->zl_lwb_list, lwb);
1232 		if (lwb->lwb_buf != NULL)
1233 			zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1234 		kmem_cache_free(zil_lwb_cache, lwb);
1235 	}
1236 	list_destroy(&zilog->zl_lwb_list);
1237 
1238 	avl_destroy(&zilog->zl_vdev_tree);
1239 	mutex_destroy(&zilog->zl_vdev_lock);
1240 
1241 	ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1242 	list_destroy(&zilog->zl_itx_list);
1243 	mutex_destroy(&zilog->zl_lock);
1244 
1245 	cv_destroy(&zilog->zl_cv_writer);
1246 	cv_destroy(&zilog->zl_cv_suspend);
1247 
1248 	kmem_free(zilog, sizeof (zilog_t));
1249 }
1250 
1251 /*
1252  * return true if the initial log block is not valid
1253  */
1254 static int
1255 zil_empty(zilog_t *zilog)
1256 {
1257 	const zil_header_t *zh = zilog->zl_header;
1258 	arc_buf_t *abuf = NULL;
1259 
1260 	if (BP_IS_HOLE(&zh->zh_log))
1261 		return (1);
1262 
1263 	if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0)
1264 		return (1);
1265 
1266 	VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1267 	return (0);
1268 }
1269 
1270 /*
1271  * Open an intent log.
1272  */
1273 zilog_t *
1274 zil_open(objset_t *os, zil_get_data_t *get_data)
1275 {
1276 	zilog_t *zilog = dmu_objset_zil(os);
1277 
1278 	zilog->zl_get_data = get_data;
1279 	zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1280 	    2, 2, TASKQ_PREPOPULATE);
1281 
1282 	return (zilog);
1283 }
1284 
1285 /*
1286  * Close an intent log.
1287  */
1288 void
1289 zil_close(zilog_t *zilog)
1290 {
1291 	/*
1292 	 * If the log isn't already committed, mark the objset dirty
1293 	 * (so zil_sync() will be called) and wait for that txg to sync.
1294 	 */
1295 	if (!zil_is_committed(zilog)) {
1296 		uint64_t txg;
1297 		dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1298 		(void) dmu_tx_assign(tx, TXG_WAIT);
1299 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1300 		txg = dmu_tx_get_txg(tx);
1301 		dmu_tx_commit(tx);
1302 		txg_wait_synced(zilog->zl_dmu_pool, txg);
1303 	}
1304 
1305 	taskq_destroy(zilog->zl_clean_taskq);
1306 	zilog->zl_clean_taskq = NULL;
1307 	zilog->zl_get_data = NULL;
1308 
1309 	zil_itx_clean(zilog);
1310 	ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1311 }
1312 
1313 /*
1314  * Suspend an intent log.  While in suspended mode, we still honor
1315  * synchronous semantics, but we rely on txg_wait_synced() to do it.
1316  * We suspend the log briefly when taking a snapshot so that the snapshot
1317  * contains all the data it's supposed to, and has an empty intent log.
1318  */
1319 int
1320 zil_suspend(zilog_t *zilog)
1321 {
1322 	const zil_header_t *zh = zilog->zl_header;
1323 
1324 	mutex_enter(&zilog->zl_lock);
1325 	if (zh->zh_claim_txg != 0) {		/* unplayed log */
1326 		mutex_exit(&zilog->zl_lock);
1327 		return (EBUSY);
1328 	}
1329 	if (zilog->zl_suspend++ != 0) {
1330 		/*
1331 		 * Someone else already began a suspend.
1332 		 * Just wait for them to finish.
1333 		 */
1334 		while (zilog->zl_suspending)
1335 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1336 		ASSERT(BP_IS_HOLE(&zh->zh_log));
1337 		mutex_exit(&zilog->zl_lock);
1338 		return (0);
1339 	}
1340 	zilog->zl_suspending = B_TRUE;
1341 	mutex_exit(&zilog->zl_lock);
1342 
1343 	zil_commit(zilog, UINT64_MAX, 0);
1344 
1345 	/*
1346 	 * Wait for any in-flight log writes to complete.
1347 	 */
1348 	mutex_enter(&zilog->zl_lock);
1349 	while (zilog->zl_writer)
1350 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1351 	mutex_exit(&zilog->zl_lock);
1352 
1353 	zil_destroy(zilog, B_FALSE);
1354 
1355 	mutex_enter(&zilog->zl_lock);
1356 	zilog->zl_suspending = B_FALSE;
1357 	cv_broadcast(&zilog->zl_cv_suspend);
1358 	mutex_exit(&zilog->zl_lock);
1359 
1360 	return (0);
1361 }
1362 
1363 void
1364 zil_resume(zilog_t *zilog)
1365 {
1366 	mutex_enter(&zilog->zl_lock);
1367 	ASSERT(zilog->zl_suspend != 0);
1368 	zilog->zl_suspend--;
1369 	mutex_exit(&zilog->zl_lock);
1370 }
1371 
1372 typedef struct zil_replay_arg {
1373 	objset_t	*zr_os;
1374 	zil_replay_func_t **zr_replay;
1375 	void		*zr_arg;
1376 	uint64_t	*zr_txgp;
1377 	boolean_t	zr_byteswap;
1378 	char		*zr_lrbuf;
1379 } zil_replay_arg_t;
1380 
1381 static void
1382 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1383 {
1384 	zil_replay_arg_t *zr = zra;
1385 	const zil_header_t *zh = zilog->zl_header;
1386 	uint64_t reclen = lr->lrc_reclen;
1387 	uint64_t txtype = lr->lrc_txtype;
1388 	char *name;
1389 	int pass, error, sunk;
1390 
1391 	if (zilog->zl_stop_replay)
1392 		return;
1393 
1394 	if (lr->lrc_txg < claim_txg)		/* already committed */
1395 		return;
1396 
1397 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
1398 		return;
1399 
1400 	/* Strip case-insensitive bit, still present in log record */
1401 	txtype &= ~TX_CI;
1402 
1403 	/*
1404 	 * Make a copy of the data so we can revise and extend it.
1405 	 */
1406 	bcopy(lr, zr->zr_lrbuf, reclen);
1407 
1408 	/*
1409 	 * The log block containing this lr may have been byteswapped
1410 	 * so that we can easily examine common fields like lrc_txtype.
1411 	 * However, the log is a mix of different data types, and only the
1412 	 * replay vectors know how to byteswap their records.  Therefore, if
1413 	 * the lr was byteswapped, undo it before invoking the replay vector.
1414 	 */
1415 	if (zr->zr_byteswap)
1416 		byteswap_uint64_array(zr->zr_lrbuf, reclen);
1417 
1418 	/*
1419 	 * If this is a TX_WRITE with a blkptr, suck in the data.
1420 	 */
1421 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1422 		lr_write_t *lrw = (lr_write_t *)lr;
1423 		blkptr_t *wbp = &lrw->lr_blkptr;
1424 		uint64_t wlen = lrw->lr_length;
1425 		char *wbuf = zr->zr_lrbuf + reclen;
1426 
1427 		if (BP_IS_HOLE(wbp)) {	/* compressed to a hole */
1428 			bzero(wbuf, wlen);
1429 		} else {
1430 			/*
1431 			 * A subsequent write may have overwritten this block,
1432 			 * in which case wbp may have been been freed and
1433 			 * reallocated, and our read of wbp may fail with a
1434 			 * checksum error.  We can safely ignore this because
1435 			 * the later write will provide the correct data.
1436 			 */
1437 			zbookmark_t zb;
1438 
1439 			zb.zb_objset = dmu_objset_id(zilog->zl_os);
1440 			zb.zb_object = lrw->lr_foid;
1441 			zb.zb_level = -1;
1442 			zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp);
1443 
1444 			(void) zio_wait(zio_read(NULL, zilog->zl_spa,
1445 			    wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL,
1446 			    ZIO_PRIORITY_SYNC_READ,
1447 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb));
1448 			(void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen);
1449 		}
1450 	}
1451 
1452 	/*
1453 	 * We must now do two things atomically: replay this log record,
1454 	 * and update the log header to reflect the fact that we did so.
1455 	 * We use the DMU's ability to assign into a specific txg to do this.
1456 	 */
1457 	for (pass = 1, sunk = B_FALSE; /* CONSTANTCONDITION */; pass++) {
1458 		uint64_t replay_txg;
1459 		dmu_tx_t *replay_tx;
1460 
1461 		replay_tx = dmu_tx_create(zr->zr_os);
1462 		error = dmu_tx_assign(replay_tx, TXG_WAIT);
1463 		if (error) {
1464 			dmu_tx_abort(replay_tx);
1465 			break;
1466 		}
1467 
1468 		replay_txg = dmu_tx_get_txg(replay_tx);
1469 
1470 		if (txtype == 0 || txtype >= TX_MAX_TYPE) {
1471 			error = EINVAL;
1472 		} else {
1473 			/*
1474 			 * On the first pass, arrange for the replay vector
1475 			 * to fail its dmu_tx_assign().  That's the only way
1476 			 * to ensure that those code paths remain well tested.
1477 			 *
1478 			 * Only byteswap (if needed) on the 1st pass.
1479 			 */
1480 			*zr->zr_txgp = replay_txg - (pass == 1);
1481 			error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf,
1482 			    zr->zr_byteswap && pass == 1);
1483 			*zr->zr_txgp = TXG_NOWAIT;
1484 		}
1485 
1486 		if (error == 0) {
1487 			dsl_dataset_dirty(dmu_objset_ds(zr->zr_os), replay_tx);
1488 			zilog->zl_replay_seq[replay_txg & TXG_MASK] =
1489 			    lr->lrc_seq;
1490 		}
1491 
1492 		dmu_tx_commit(replay_tx);
1493 
1494 		if (!error)
1495 			return;
1496 
1497 		/*
1498 		 * The DMU's dnode layer doesn't see removes until the txg
1499 		 * commits, so a subsequent claim can spuriously fail with
1500 		 * EEXIST. So if we receive any error other than ERESTART
1501 		 * we try syncing out any removes then retrying the
1502 		 * transaction.
1503 		 */
1504 		if (error != ERESTART && !sunk) {
1505 			txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1506 			sunk = B_TRUE;
1507 			continue; /* retry */
1508 		}
1509 
1510 		if (error != ERESTART)
1511 			break;
1512 
1513 		if (pass != 1)
1514 			txg_wait_open(spa_get_dsl(zilog->zl_spa),
1515 			    replay_txg + 1);
1516 
1517 		dprintf("pass %d, retrying\n", pass);
1518 	}
1519 
1520 	ASSERT(error && error != ERESTART);
1521 	name = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1522 	dmu_objset_name(zr->zr_os, name);
1523 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1524 	    "dataset %s, seq 0x%llx, txtype %llu %s\n",
1525 	    error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype,
1526 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
1527 	zilog->zl_stop_replay = 1;
1528 	kmem_free(name, MAXNAMELEN);
1529 }
1530 
1531 /* ARGSUSED */
1532 static void
1533 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1534 {
1535 	zilog->zl_replay_blks++;
1536 }
1537 
1538 /*
1539  * If this dataset has a non-empty intent log, replay it and destroy it.
1540  */
1541 void
1542 zil_replay(objset_t *os, void *arg, uint64_t *txgp,
1543 	zil_replay_func_t *replay_func[TX_MAX_TYPE])
1544 {
1545 	zilog_t *zilog = dmu_objset_zil(os);
1546 	const zil_header_t *zh = zilog->zl_header;
1547 	zil_replay_arg_t zr;
1548 
1549 	if (zil_empty(zilog)) {
1550 		zil_destroy(zilog, B_TRUE);
1551 		return;
1552 	}
1553 
1554 	zr.zr_os = os;
1555 	zr.zr_replay = replay_func;
1556 	zr.zr_arg = arg;
1557 	zr.zr_txgp = txgp;
1558 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1559 	zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1560 
1561 	/*
1562 	 * Wait for in-progress removes to sync before starting replay.
1563 	 */
1564 	txg_wait_synced(zilog->zl_dmu_pool, 0);
1565 
1566 	zilog->zl_stop_replay = 0;
1567 	zilog->zl_replay_time = lbolt;
1568 	ASSERT(zilog->zl_replay_blks == 0);
1569 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1570 	    zh->zh_claim_txg);
1571 	kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE);
1572 
1573 	zil_destroy(zilog, B_FALSE);
1574 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1575 }
1576 
1577 /*
1578  * Report whether all transactions are committed
1579  */
1580 int
1581 zil_is_committed(zilog_t *zilog)
1582 {
1583 	lwb_t *lwb;
1584 	int ret;
1585 
1586 	mutex_enter(&zilog->zl_lock);
1587 	while (zilog->zl_writer)
1588 		cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1589 
1590 	/* recent unpushed intent log transactions? */
1591 	if (!list_is_empty(&zilog->zl_itx_list)) {
1592 		ret = B_FALSE;
1593 		goto out;
1594 	}
1595 
1596 	/* intent log never used? */
1597 	lwb = list_head(&zilog->zl_lwb_list);
1598 	if (lwb == NULL) {
1599 		ret = B_TRUE;
1600 		goto out;
1601 	}
1602 
1603 	/*
1604 	 * more than 1 log buffer means zil_sync() hasn't yet freed
1605 	 * entries after a txg has committed
1606 	 */
1607 	if (list_next(&zilog->zl_lwb_list, lwb)) {
1608 		ret = B_FALSE;
1609 		goto out;
1610 	}
1611 
1612 	ASSERT(zil_empty(zilog));
1613 	ret = B_TRUE;
1614 out:
1615 	cv_broadcast(&zilog->zl_cv_writer);
1616 	mutex_exit(&zilog->zl_lock);
1617 	return (ret);
1618 }
1619